An Introduction To Non Linearity in Control System

An Introduction to Nonlinearity in Control Systems
Derek Atherton
Derek Atherton
An Introduction to Nonlinearity in Control Systems 2011 Derek Atherton & Ventus Publishing ApS ISBN 978-87-7681-790-9
Contents
Contents
Preface 1 1.1 1.2 1.3 1.4 1.5 2 2.1 2.2 2.2.1 2.2.2 2.2.3 2.3 2.3.1 2.3.2 2.3.3 Introduction What is nonlinearity? Forms of nonlinearity Structure and Behaviour Overview of contents References The Phase Plane Method Introductio n Basic Principles The Linear Case The Nonlinear Case An Example The Phase Plane for Systems with Linear Segmented Nonlinearities Example 1 Nonlinear Output Derivative Feedback Example 2 Relay Position Control Example 3 Position Control with Torque Saturation 10 12 12 13 16 17 18 19 19 19 20 23 24 26 27 28 34
WHATS MISSING IN THIS EQUATION?
Please click the advert
You could be one of our future talents
MAERSK INTERNATIONAL TECHNOLOGY & SCIENCE PROGRAMME Are you about to graduate as an engineer or geoscientist? Or have you already graduated? If so, there may be an exciting future for you with A.P. Moller - Maersk.
www.maersk.com/mitas
An Introduction to Nonlinearity in Control Systems 2. 4 2.5 2.6 3 3.1 3.2 3.3 3.4 3.4.1 3.4.2 3.4.3 3.5 3.6 3.7 3.8 3.9 3.10 3.11 4 4.1 Conclusions References Bibliography The Describing Function Introductio n The Sinusoidal Describing Functio n Some Properties of the DF The Evaluatio n of some DFs Cubic Nonlinearity Saturation Nonlinearity Relay with Dead Zone and Hysteresis Nonlinear Models and DFs Harmonic Outputs Sine plus Bias DF and the IDF Conclusions References Bibliography Appendix -Tables of Describing Functio ns Stability and Limit Cycles using the DF Introductio n 35 35 36 37 37 37 39 41 41 41 42 44 45 46 47 47 47 47 51 51
Contents
www.job.oticon.dk
An Introduction to Nonlinearity in Control Systems 4.2 4.3 4.4 4.5 4.5.1 4.5.2 4.5.3 4.5.4 4.6 4.7 4.7.1 4.7.2 4.8 4.9 5 5.1 5.2 5.3 5.3.1 5.3.2 5.3.3 Limit Cycle Evaluatio n Stability of a Predicted Limit Cycle DF Accuracy Some Examples of Limit Cycle Evaluation The Van der Pol Equation Feedback Loop Containing a Relay with Dead Zo ne Feedback Loop with On Off Relay Use of the IDF More than one Nonlinear Element Applicatio ns of SBDF to find Limit Cycles Asymmetrical no nlinearity Effect of Bias with Odd Symmetrical Nonlinearity Conclusions Bibliography The SSDF and Harmonically Forced Systems Introductio n The Cubic Nonlinearity with Two Sinusoidal Inputs Modified Nonlinearities and the SSDF Power Law Characteristic Harmonic Nonlinearity Ideal Relay 52 53 55 55 55 56 58 59 61 62 63 65 67 67 68 68 69 70 72 72 73
Contents
Join Vestas Winnovation 2012

Submit your solution before 18 November 2011 and win a trip around the world Read more at vestas.com/winnovation
An Introduction to Nonlinearity in Control Systems 5.4 5.5 5.6 5.7 5.7.1 5.7.2 5.8 5.9 6 6.1 6.2 6.3 6.4 6.4.1 6.4.2 6.5 6.6 6.6.1 6.6.2 6.6.3 6.6.4 The IDF for Related Signals More Accurate Determinatio n of Limit Cycles Closed Loop Frequency Response Jump Resonance Jump resonance region- use of the IDF Some Examples of Jump Resonance Conclusions References Limit cycles in relay systems Introductio n The Frequency Domain Approach Properties and Evaluation of A loci Solving for Limit Cycles Relay with no Dead Zone Relay with Dead Zone Limit Cycle Stability Some Interesting Limit Cycle Problems Example 1 Invalid Continuity Condition Example 2 Multipulse Limit Cycles Example 3 Limit Cycle with a Sliding Mode Example 4 Multiple Limit Cycle Solutions 74 76 79 83 85 85 95 96 97 97 98 101 102 102 104 105 106 106 107 108 111
Contents
An Introduction to Nonlinearity in Control Systems 6.6.5 6.6.6 6.7 6.8 6.9 6.10 7 7.1 7.2 7.2.1 7.2.2 7.3 7.3.1 7.4 7.4.1 7.5 7.6 8 8.1 8.2 Example 5 Chaotic Motion Example 6 Unstable Limit Cycles and Simulatio n Forced oscillations Conclusions References Appendix Controller Tuning from Relay Produced Limit Cycles Introductio n Knowledge from the Limit Cycle Relay Autotuning Plant Parameter Identifica tion Tuning the Co ntroller An Example Autotuning using the Relay in Parallel with the Controller An Example Conclusions References Absolute Stability Results Introductio n Lyapuno vs Method 111 114 114 114 114 116 125 125 127 127 129 130 133 133 136 137 137 138 138 138
Contents
678'< )25 <285 0$67(56 '(*5((

&KDOPHUV 8QLYHUVLW\ RI 7HFKQRORJ\ FRQGXFWV UHVHDUFK DQG HGXFDWLRQ LQ HQJLQHHU LQJ DQG QDWXUDO VFLHQFHV DUFKL WHFWXUH WHFKQRORJ\ UHODWHG PDWKHPD WLFDO VFLHQFHV DQG QDXWLFDO VFLHQFHV %HKLQG DOO WKDW &KDOPHUV DFFRPSOLVKHV WKH DLP SHUVLVWV IRU FRQWULEXWLQJ WR D VXVWDLQDEOH IXWXUH ERWK QDWLRQDOO\ DQG JOREDOO\ 9LVLW XV RQ &KDOPHUV VH RU 1H[W 6WRS &KDOPHUV RQ IDFHERRN
An Introduction to Nonlinearity in Control Systems 8.2.1 8.2.2 8.2.3 8.3 8.3.1 8.3.2 8.4 8.4.2 8.5 8.5.1 8.5.2 8.5.3 8.6 8.7 8.8 8.9 9 9.1 9.2 9.3 9.4 9.5 9.6 9.7 9.7.1 9.7.2 9.8 9.9 9.10 9.10.1 9.10.2 9.10.3 9.10.4 9.10.5 9.11 9.12 Definitio ns of Stability Positive Definite Functio ns Lyapuno vs Theorems Applicatio n of Lyapunovs Method Linear System Nonlinear System Definitio ns and Loop Transformations Loop Transformations Frequency Domain Criteria Popov Theorem The Circle Criterion The Off-axis Circle Criterion Examples Conclusions References Bibliography Design of Nonlinear Control Systems Introductio n Linearizatio n Frequency Response Shaping Nonlinear Compensation Compensation using DF Models Time Optimal Co ntrol Sliding Mode Control A Second Order System Further Comments Fuzzy Logic Neural Networks Exact Linearizatio n Relative Degree Input-Outp ut Linearizatio n Exact State Linearization Examples Further Comments General Comments References 139 140 140 141 141 142 142 143 144 144 145 147 148 150 150 150 151 151 152 153 154 155 155 157 157 162 163 163 164 166 167 169 169 174 175 175
Contents
Preface
Preface
The book is intended to provide an introduction to the effects of nonlinear elements in feedback control systems. A central topic is the use of the Describing Function (DF) method since in comb ination with simula tion it provides an excellent approach for the practicing engineer and follows on logically from a first course in classical control, such as the co m p a nio n volum e in this series. So m e of the ba sic m aterial o n the to pic ca n b e f o und in m y ea rlier b o ok which is frequently referenced throug hout the text. The first cha pter pro vid es a n i ntro d uctio n to no nlinearity fro m the b asic def initio n to a discussio n of the p o ssible effects it ca n ha ve o n a system a nd the diff erent b eha vio ur tha t m ig ht b e fo und , in p a rticula r whe n it o ccurs within a casca d e system of elem ents o r in a f eedb a ck lo o p. T he f ina l pa rt of the chap ter gives a b rief o verview of the co ntents of the b oo k. Phase plane methods for second order systems are covered in the second chapter. Many systems, particularly electromechanical ones, ca n be app roximated by seco nd order models so the co ncept ca n b e particularly useful in pra ctice. The second order linear system is first considered as surprisingly it is rarely covered in linear control system texts. The m etho d ha s the big a d va nta ge in tha t the eff ects of m o re tha n o ne no nlinear e lem ent m a y b e co nsid ered . The stud y is supported by several simulatio ns done in Simulink including one where sliding motion ta kes place. The third cha p ter is the first of three d evoted to the stud y of f eed b a ck loo ps using DF m etho d s. Altho ug h the m etho d is a n ap p ro xim ate techniq ue its value, a nd lim itatio ns, are sup p o rted b y a la rg e nu m b er of exa m ples co ntaining a nalytical results a nd sim ulatio ns, including the estima tio n of limit cycles a nd lo op sta bility. Ma ny early pa p ers o n DFs sho wing ho w theories co uld b e used to predict sp ecific p heno m ena were supp orted b y sim ulatio ns do ne o n a nalog ue co m p uters. Here results from digital simulations using Simulink are presented and this allows much more control of initial conditions to show how different modes may exist dependent on the initial conditions. In particular, Chapter 5 contains some more ad va nced wo rk including so m e new results o n jum p reso na nce, so so m e read ers ma y wish to o mit this cha pter o n a first reading. A rela y is a uniq ue no nlinea r elem ent in tha t its o utp ut do es not d ep end up o n the inp ut a t all tim es b ut is d etermined b y when the input passes through the relay switching levels. It is this feature which allows the exact determina tion of limit cycles a nd their sta bility in a f eed ba ck lo op . The b asic theory is p resented a nd so m e sim ple exa m ples co vered. It is the n sho wn in sectio n 6.6 ho w the a p pro a ch ca n be used with co m p uta tio nal sup p o rt to a na lyse q uite co m plicated p erio dic m o d es in rela y system s. Am o ng the to pics co vered fo r the first tim e in a textbo o k a re the eva luatio n of lim it cycles with multiple pulses per cycle, as found in a satellite attitude control system, the determination of a limit cycle with sliding and other more advanced aspects which some readers may wish to omit. A practical method developed in recent years for finding suitable parameters, or tuning, for a controller based on the so called lo o p cycling m etho d o f Ziegler a nd Nichols ha s used rela y p rod uced limit cycles. Cha p ter 7 co vers these id eas using b oth a p pro xim ate a nalysis b ased o n the DF a nd exa ct a nalysis b ased o n the rela y m ethod s of the p revio us chap ter.
10
Preface
Cha pter 8 co vers the to pic of a bsolute stability na m ely trying to o btain necessa ry a nd sufficient co nd itio ns f or the sta bility of a f eed ba ck lo op with a single no nlinear elem ent. This p ro blems ha s exercised the minds of theorists fo r nearly a centu ry but a solution seems no nearer! Several necessary but not sufficient results are presented, which dependent on ones viewpoint ma y be regarded as conservative or robust. The former is usually the case when one has a mathematically defined nonlinearity a nd t he latter ma y be used beca use the result gives stability for a ny no nlinearity with certain properties, for example, lying within a sector. The final chapter, chapter 9, discusses q uite briefly various methods which ca n be used for the desig n of nonlinear systems. The intent ha s b een to pro vid e sufficient inf orm a tio n o n the m ethod s a nd their po ssible a d va ntag es a nd disad va ntag es. Several of them have co mplete books written o n the topic and more d etail co uld not have been given witho ut the coverage of more specialised mathema tics. Finally my tha nks to the University of Sussex for the use of an office and access to the computing facilities during my retirem ent; to m y g oo d f riend Keith Go df rey a nd his student R ola nd f or the co m p uta tio na l inp ut o n jum p reso na nc e a nd their companio nship at the races and to my wife, Constance, for her love and support. Derek P Atherto n University of Sussex Brighto n May 2011.
11
Introduction
1 Introduction
1.1 What is nonlinearity?
In order to a nalyse the b eha vio ur of engineering systems ma thematical mo dels are req uired fo r the vario us co m po nents. It is co m m o n pra ctice to try a nd ob ta in linear m o dels a s a rich m a them atical theo ry exists f or linea r system s. These m o dels will alwa ys be app roximate, althoug h possibly q uite a ccurate for defined rang es of system variables, b ut inevitably no nlinea r effects will eventually be found for large excursions of system variables. Linear systems have the important property that they satisf y the sup erp ositio n principle. This lea d s to m a ny im p orta nt a d va nta g es in m etho d s fo r their a na lysis. Fo r example, in circuit theory when an RLC circuit has both d.c and a.c input voltages, the voltage or current elsewhere in the circuit ca n b e f o und b y sum ming the results of sepa ra te a na lyses f or the d .c. a nd a .c. inp uts ta ken individ ually, also if the m a g nitud e of the a.c. volta g e is d o ubled then the a .c. volta g es a nd currents elsewhere in the circuit will be do ub led . Thus, mathematically a linear system with inp ut x (t) and output y (t) satisfies the property that the output for an input ax 1 (t) + bx2 (t) is ay1 (t) + by2 (t) , if
y1 (t) and y2 (t) are the outputs in response to the inputs x1 (t) and x2 (t)
, respectively, and a and b are constants. A further point is that when the a.c. voltage is sinusoidal, which is normally assumed in using the term a .c., then all the other voltages a nd currents in the circuit are sinusoidal with the sam e freq uenc y as the input. A no nlinea r system is d efined a s o ne which d o es no t satisf y the sup erp ositio n pro p erty. The sim p lest f o rm of no nline ar system is the static no nlinearity where the o utp ut d ep end s o nly o n the current value of inp ut b ut in a no nlinea r m a nne r, for example
y (t) = cx (t) + dx3 (t)
(1.1)
where the o utp ut is the sum m a tio n of a linea r a nd a no nlinear, cubic, term . This d escriptio n of no nlinea rity is given b y a sim ple m athem a tical expressio n. Pra ctical no nlinearities which o ccur in co ntrol eng ineering , ho wever, ca n of ten not b e so ea sily d escribed ; so that a p pro xim a ting them m a y req uire m o re co m plex m a them atical m o dels, such a s a hig h o rd er power series or a linear segmented approximation. More commonly a nonlinear differential eq uation, for example
2
(1.2)
will describe the behavior. From an engineering viewpoint it ma y be desirable to think of this equation in terms of a block diagram, as shown in Figure 1.1, consisting of linear dynamic elements and a static nonlinearity, in this case a
3
nonlinear, unlike linear, differential equations. The major point about no nlinear systems, however, is that their response is a m plitud e d ep end ent so tha t if a pa rticula r fo rm of resp o nse, o r so m e m easure of it, o ccurs f or o ne inp ut m a g nitud e it may no t result for some other input mag nitude.
12
d y (t) /d (t) 2 + c (dy (t) /dt) 3 + y (t) = x ((t)
cubic, which with inp ut dy (t) /dt gives an outp ut c (dy (t) /dt) . U nfortunately there is no general approa ch to solving
Introduction
Figure 1.1 Block diagram in Simulink for equation 1.2
A further and very important point, is that unlike a linear operation, a nonlinear operation on a sinusoid of frequency, f, will not prod uce a n o utp ut a t f req uency f, alo ne. Fo r exa m ple, if such a sinusoid is ap plied to the static no nlinearity o f eq uatio n (1.1) it is ea sy to sho w f ro m substituting x (t) = a cos (~t) where ~ = 2rf tha t there are o utp uts a t freq uencies
3 3
is that, as will be shown later, they exhibit several forms of unique behavio ur which are not possible in linear systems.
1.2 Forms of nonlinearity

All p ra ctical system s a re no nlinear a nd in this sectio n a b rief o verview is given of so m e no nlinea r eff ects that often o ccur. In initial desig ns it ma y b e p ossible to a pp ro xim ate the no nlinear effects b y linea r m od els b ut invariably it will be necessary to finally check their effects on the system performance either in simula tion or/and the real hardware. Todays digital sim ula tio n la ng ua g es are very g oo d b ut to use them efficiently f o r investiga ting the eff ects of no nlinea rity, or to a ssist in the design of a nonlinear system, requires a knowledge of the supporting theoretical methods presented in this book. In typical control engineering problems nonlinearity may occur in the dynamics of the plant to be controlled or in the co m po nents used to implem ent the co ntrol. In the latter ca se, fo r exa m ple, a valve a ctuator ma y ha ve a d ead zo ne d ue to frictio n eff ects a nd will certainly saturate f or la rg e inp uts, so this m a y be ref erred to as a n inherent no nlinea rity, b eca use it exists a ltho ug h o ne m ig ht p ossibly p re fer this no t to b e the case. Alterna tively o ne m a y ha ve intentio nal no nlinea rities which ha ve been p urposely desig ned into the system to impro ve the system specifica tions, either for technical or eco nomic rea so ns. A g o od exa m ple of this is the o n -o ff co ntro l used in m a ny tem p era ture co ntro l system s, where the o bjective is to have the temperature oscillate about the required value. Identifying the p recise form of a nonlinearity ma y not b e easy a nd like all mod eling exercises the golden rule is to be a ware of the appro ximations in a no nlinear mod el and the co nditio ns for its validity. It might b e arg ued that linear systems theory is no t a p plica ble to p ra ctical co ntrol engineering p ro blem s beca use they a re alwa ys no nlinea r. This is a n o versta tem ent, of course, but a valid reminder. All systems have actua tor saturation and in some cases it might occur for relatively low error sig nals, f o r exa m ple in ro tary p ositio n co ntrol it is not unusua l f or a step inp ut o f sa y 10, or even less, to pro d uce the m a xim um m o to r d rive torq ue. It sim p ly is the result of g oo d eco no m ical d esig n as to pro d uce linea r o p era tio n to a hig her to rq ue level wo uld req uire a la rg er a nd m ore exp ensive m otor. Valves used to co ntrol fluid o r ga s f lo w, a p art f ro m having the no nlinear effects mentioned above, ca n have a slig htly different beha vior when opening compared with closing
due to the unidirectional pressure of the fluid.
13
f a nd 3f of mag nitud es ca + 3da /a and da /4 , respectively. Perhaps the most interesting aspect of no nlinear systems
Introduction
Frictio n a lwa ys o ccurs in m echa nical system s a nd is very d ifficult to mo d el, with m a ny q uite sop histica ted mo d els ha vin g b een p resented in the literature. The sim plest is to a ssum e the three co m p o nents, illustrated in Fig ure 1.2, of stictio n, an ab b revia tio n fo r sta tic frictio n, Co ulo m b f rictio n a nd visco us frictio n. As its na m e im plies stictio n is a ssum ed to exist o nly at zero differential speed between the two contact surfaces. Coulomb frictio n with a value less tha n stiction is assumed to be constant at all speeds, and viscous friction is a linear effect being directly pr oportional to speed. In practice there is often a term pro po rtio nal to a hig her p o wer of sp eed , a nd this is also the situa tio n fo r m a ny sha ft lo a ds, fo r exa m ple a fan for which the drive torque typically increases as a power of speed.
Fri cti on Fo r ce V is co u s S ticti on Co ulom b
S pee d
Figure 1.2 Three basic friction components
A mathematical expression sometimes used to approximate friction is
n (v) = (a - be- c | v | + d | v |) sgn (v)
(1.3)
Here the pa ra m eters a, b, c a nd d are cho sen to pro vid e the d esired sha p e a g ainst speed which initia lly decreases fro m the value at zero speed and then starts to increase. Many mecha nical loads a re driven throug h g earing rather tha n directly. Altho ug h geared drives, like all areas of technolog y, have improved throug h the years they always have some small backlash. This may be avoided by using anti -backlash gears, which are only available for low torques. Backlash, which is a very complicated phenomeno n, involving impacts between surfaces is often modeled in a very simplistic manner. For example, the simple approach used in some digital sim ulatio n la ng ua g es, such as Sim ulink , the sim ulatio n co m po nent of the well -k no wn sof twa re pa cka g e, Ma tla b , co nsists of a n inp ut -o utp ut po sitio n cha ra cteristic of two p a rallel straig ht lines with p ossible ho rizo ntal m o vem ent b etween them . This ma k es two m ajor assum ptio ns, first tha t the loa d shaf t f rictio n is hig h eno ug h f or co nta ct to b e m aintained with the drive side of the backlash when the drive slows down to rest. Secondly when the drive reverses the backlash is crossed a nd the new drive sid e of the g ear picks up the lo ad insta nta neo usly with no lo ss of energ y in the im pa ct a nd bo th then m o ve at the d rive sha ft sp ee d. Clea rly bo th these assum p tio ns a re never true in p ra ctice b ut no check s exist in Sim ulink
to d eterm ine ho w g o od they are o r, ind eed, when they a re co m pletely invalid . Better f a cilities f or m o d eling p heno m ena such as b a cklash ca n be fo und in sim ula tio n la ng ua ges such a s 20 -Sim or Dynast, which do no t use the blo ck co ncept of Simulink. A good discussion of friction and backlash can be found in reference1.1.
14
Introduction
Probably, the most wid ely used intentio nal nonlinearity is the relay. The on -off typ e, which ca n b e described mathem atically b y the sig num f unctio n, tha t is switches o n if its i np ut exceeds zero a nd o ff if it g oes b elo w zero , is wid ely used no rm ally with so m e hysteresis b etween the switching levels. U se of this a p proa ch p ro vid es a co ntro l stra teg y where the co ntrolled varia ble o scillates a bo ut the d esired level. The switching m ec ha nism va ries sig nifica ntly a cco rd ing to the ap plica tio n fro m electro m echa nical rela ys at lo w sp eed to fa st electro nic switches em plo ying tra nsistors o r thyristors. A co m m o n usa ge of the rela y is in the tem p erature co ntrol of b uilding s, where typically the switching is p ro vid ed f ro m a tem p erature senso r ha ving a po ol of m ercury o n a m etal expa nsio n coil. As the tem pera ture d ro ps the coil co ntra cts a nd this ca uses a cha ng e in a ngle of the mercury capsule so that eventually the mercury mo ves and closes a co ntact. When the temperature increases the coil expands causing a change in ang le for the mercury to flo w and break the contact. Electronic switching controllers are being used in many modern electric motor drive systems, for example, to regulate phase currents in stepping motors and switched reluctance motors and to control currents in vector control drives for induction motors. Relays with a dead zone, that is, three position relays giving positive, nega tive and a zero output are also used. When used in a p ositio n co ntrol system the zero o utp ut allo ws f or a stea d y state p ositio n within the d ea d zo ne but this affects the resulting steady state control accuracy.
Hysteresis eff ects in m ag netic ma terials so m etim es ha ve to b e m od eled . This is of ten do ne b y a ssum ing a hysteresis lo o p of the form of a B-H loop for a magnetic material typically obtained for a sinusoidal input. Ho wever the shape usually varies with the amplitude and frequency of the input and does not in fact remain consta nt with a random excitat io n. Pro vid ed the inp ut is sinusoid al a nd the sha p e do es rem ain reaso na bly co nsta nt then a no nlinea r f unctio n of the f o rm n (x, x)o m a y pro vide a reaso na ble m od el a s the p ath tak en a ro und the hysteresis d epend s up o n whether the inp ut, x, is increasing or decreasing, i.e the derivative of x, xo .
its an interesting world
Get under the skin of it.

Graduate opportunities Che lte nham | 24,945 + be nef its
One of the UKs intelligence services, GCHQs role is two-fold: to gather and analyse intelligence which helps shape Britains response to global events, and, to provide technical advice for the protection of Government communication and information systems. In doing so, our specialists in IT, internet, engineering, languages, information assurance, mathematics and intelligence get well beneath the surface of global affairs. If you thought the world was an interesting place, you really ought to explore our world of work.
G O V E R N M E N T E M P L O YER
TOP
15
ww w.ca reers inbri tishi ntell igen ce.c o.uk

Appli can ts must be British citizens . GCHQ values diversit y and welcome s applicant s from all sections of the com mun ity. We want our workforc e to reflect the diversit y of our work.
Introduction
1.3 Structure and Behaviour

From a control engineering viewpoint there are two major reasons why one needs to kno w about no nlinearity. Firstly with respect to obtaining mathematical models of devices, particularly if identification techniques are being used, and secondly for ensuring the design meets the desired specifications when the control system is no nlinear. To appreciate these asp ects it is a p pro pria te to discuss very briefly in the intro d uctio n a f ew a sp ects of b eha vio ur d ue to the presence of no nlinearity. These a re d ep end ent o n the structure o f the no nlinear system a nd the releva nce of this is explained b y again considering a sinusoidal input.
3
If they are placed in cascade and a sinusoidal input applied the outp ut will be a determinis tic waveform containing two frequency components one at the same frequency as the inp ut and the other at three times that frequency. Any casca d e co m b inatio n of linea r a nd no nlinea r elem ents will a lwa ys p rod uce a d eterministic o utp ut f or a ny given discrete inp ut freq uency spectrum , which in p rincip le ca n b e evaluated . New freq uency co m p o nents ca n o nly be created b y the nonlinearities and the linear elements simple alter the relative magnitudes and phases of these components. Fo r exa m ple, if the ab o ve no nlinea rity is pla ced b ef ore a nd after the linear tra nsf er f unctio n a nd a sinusoid of freq uenc y, f, is a p plied at the inp ut, then the inp ut to the seco nd no nlinea rity will co nsist of the f und a m ental, f, plus third ha rm o nic, 3f, with m a g nitud es a nd p hases d epend ent o n both the inp ut sinusoidal ma g nitud e a nd f req uency. These two freq uencies ap plied to the seco nd no nlinearity will prod uce a n o utp ut co nta ining the freq uency co m p o nents, f, 3f, 5f, 7f, a nd 9f. One co uld d efine a freq uency resp o nse fo r such a cascad e structure of linear a nd no nlinea r elements as the ra tio of the o utp ut at the fundamental frequency, f, to the input sinusoid at this frequency. The result, as for a linear system, would be a m a g nitud e a nd p ha se, which varies with, f, b ut beca use of the no nlinearities it wo uld also vary with the a m plitud e of the inp ut sinusoid . Thus a n a p pro xim a te f req uency respo nse m o d el fo r the co m bina tio n co uld be p ortra yed g ra p hically b y a set of freq uency respo nse plots fo r diff erent inp ut a m p litud es, o r g ain a nd p hase plots a g ainst a m plitud e fo r diff erent frequencies. Fo r m a ny p ro blem s enco u ntered in co ntro l eng ineering this m a y pro ve to b e a reaso na bly g oo d m o del since m a ny o f the linea r d yna m ic elem ents, lik e the o ne given, ha ve lo w p a s s d yna mics so that the f req uency, f, will p red o mina te at the o utp ut. With no linear d yna m ic elem ents in the co m bina tio n then these la tter plo ts wo uld b e the sa m e fo r all freq uencies a nd the a pp ro xim ate, first harm o nic, or q uasi -linea rized , m od el wo uld b e gain a nd p ha se curves as a f unctio n of the inp ut a mplitud e. This rep resentatio n of a no nlinea r elem ent is k no wn as a d escribing f unctio n (DF), which is the to pic of several cha pters. So m etim es the a bo ve m o del f o r the co m bina tio n b ased o n the f und a m ental f r eq uency, f, o nly is ref erred to as an amplitude and frequency dependent describing function.
16
Consider a nonlinearity x + dx , and a dynamic elem ent with tra nsfer f unctio n K /s (s + b) , where b a nd d are co nsta nts.
Introduction
If alterna tively we assum e the system structure to co nsist of a f eed ba ck co mb ina tio n of the linear a nd no nlinear elem ents,
3
neg ative g ain then, a very diff erent situatio n is p o ssib le f o r the respo nse to a sinusoid al inp ut. Dep end ent o n the values of b, d and the amplitude and frequency of the input, some possibilities for the output are that it is (a) approximately sinusoidal with the same frequency as the input, similar to the aforementioned cascade connection; (b) approximately sinusoid al with a f req uency related to tha t of the inp ut; (c) a co m bina tio n of p rim a rily the inp ut freq uency a nd a no ther freq uency o r (d ) a wa vef orm k no wn as cha otic, which is no t d efina ble m a them atically b ut co m pletely repeata ble f o r the same initial co nditio ns. These behaviours and others are unique to nonlinear feedback systems, aspects which make such systems extremely interesting. However, it has meant that no general analytical method is available for predicting their behavior. Several approaches will be considered in this book all of which will be restricted in their applicability or, put alterna tively, the situatio ns which they ca n address. Thus the importance of simulatio n studies for investigating no nlinear systems in asso ciatio n with a nalytical m etho ds ca nnot b e und erestim ated . Much of the sup po rt fo r the theoretical ma terial presented in the early chapters, particularly in the 50s to 60s, was done using analogue simulation. Today simula tions are done dig ita lly a nd several a re included using Sim ulink in the f ollo wing cha pters to illustrate the co ncepts a nd pro vid e solutio ns for specific problems. Care has to be taken in simulating nonlinear systems particularly those with linear segmented characteristics because of the discontinuities. Some comments are made on the simulatio ns where appropriate.
1.4 Overview of contents

The phase plane approach discussed in chapter 2 is very useful for step response and stability studies but is basically restricted to second order systems. However, many engineering systems, particularly in the mecha tronics field, may be approximated by a second order differential equation so the results are still of value. It also provides a simple basis for und ersta nding so me of the more ad vanced topics, such as optim um control and sliding mode co ntrol, co vered in chapter 9. Sta bility of a f eedb a ck lo o p is of m a jo r im p orta nce so tha t it is no t surprising that m uch ea rly wo rk wa s co ncerned with this top ic. During the 1940 s engineers in several co untries d evelop ed wha t ha s b eco m e k no wn as the describing functio n m ethod where a no nlinea rity is repla ced b y a n a m plitud e dep end ent gain to a sinuso id , k no wn as the d escrib ing f unctio n, DF. Cha pter 3 intro d uces the DF, sho ws ho w its value ca n be calcula ted fo r va rio us no nlinea rities a nd includ es a tab le of results. The next two chapters, 4 and 5, deal with applicatio ns of the describing function for estima ting limit cycles and the sta bility of a no nlinea r f eed b a ck loo p . It is a lso sho wn ho w the d escribing f unctio n ca n b e evalua ted f or other tha n a single sinusoidal input and applications of describing functions for a bias plus sinusoid and two sinusoids are given. These include areas such as limit cycle stability, jump resonance and subharmonic oscillatio ns. The rela y is a special fo rm of no nlinear elem ent where the o utp ut is no t co ntinuo usly d epend ent o n the inp ut. This allo ws sp ecial results to b e d evelo p ed f or rela y systems a to pic co vered in cha pter 6. Cha pter 7 d eals with a d esig n ap proa ch tha t has received m uch a ttentio n in recent years, na m ely using the info rm atio n co nta ined in rela y p ro d uced lim it cycles to set
the p ara m eters of a co ntroller, typically k no wn as rela y a utotuning . The topic of a bsolute sta bility, na m ely the develop m ent of exact criteria for guaranteeing the stability of a feedback loop with a single nonlinearity and linear transfer functio n, is co vered in cha pter 8. Dep end ent o n the viewp oint these results m a y b e reg ard ed a s rob ust, a s they p ro ve sta bility for with the tra nsfer function K /s (s + b) in the forward loop a nd the feedba ck loop co ntaining x + dx fed back thro ug h a variations in the nonlinearity, or conservative if one is considering stability for a specific nonlinearity.
17
Introduction
The coverage to this point, apart from chapter 7, has like many textbooks on linear control, been primarily concerned a nd other techniq ues m a y b e used f or no nlinear co ntrol system desig n. The co vera g e of these topics is q uite brief , so m e necessarily so b eca use of the a d ditio nal theo retical co ncep ts which wo uld ha ve to b e introd uced to g o into them m o re d eepl y. Ho pef ully, sufficient inf o rm atio n, tog ether with the ref erences, is given fo r the read er to und ersta nd the co ncepts involved, their possible relevance for particular applications and how they might be applied.
with intro d ucing a na lytical to ols. Cha pter 9, ho wever, f o cuses mo re o n d esig n a nd loo ks at ho w so m e of the id ea s co vered
1.5 References
1.1 Friedland, B, 1996 Ad vanced control system design, Chapter 7 Prentice Hall.
In Paris or Online
International prog rams taught by professors an d professi onals from all ov er the w orld BBA in G lobal Business M BA in International M anagement / International M arketing DBA in International Business / Interna tional M anagement MA in International E ducation MA in C ross-C ultural C ommunication MA in F oreign Languages Innov ative P ractical F lexible A ffordable Visit: www.HorizonsU niversity.org Write: A dmissions@horizo nsuniv ersity .org C all: 01.42.77.20.66
www.HorizonsUniversity.org
18
The Phase Plane Method
2 The Phase Plane Method

2.1 Introduction
A sig nifica nt a m o unt of the research into no nlinea r diff erential eq uatio ns in the nineteenth a nd ea rly twentieth centuries do ne b y m athem a ticia ns a nd p hysicists was d evoted to seco nd order diff erential eq uatio ns. There were two m ajor reaso ns for this, na m ely tha t the d yna mics of ma ny p ro blems of pra ctical interest co uld be a p pro xim ated b y these eq uatio ns a nd seco ndly the p ha se pla ne a p proa ch allo wed a grap hical exa mina tio n of their solutio ns. The system s of interest in the late nineteenth a nd early twentieth centuries were f o und in f ields such a s celestial m echa nics, no nlinea r m echa nical system s and electronic oscillations. This section will introduce the basic concepts of the phase plane approach and then give a brief overview of how the method has been further developed for use in control system analysis and design. A sig nifica nt a m o unt of the early d evelo p m ent in co ntrol theo ry f ro m 1930 wa s d riven b y two a rea s, na m ely to a chieve better co ntrol of ind ustrial p ro cesse s a nd to a chieve b etter p erfo rma nce in fire co ntrol p ro blem s. The problem s in the latter area received significantly more attention in the war years after 1939, where the requirement in many cases was related to the position co ntrol of radar antennas and guns in both stationary and mo ving situatio ns. The dynamic equatio ns representing many of these position co ntrol systems co uld be rep resented reasonably a ccurately b y seco nd order no nlinear differential equa tions. It was therefore not surprising that much o f the early work on no nlinear control used the phase plane approach. Control engineers did make significant contrib utio ns to this field since, whereas the earlier work had typically assum ed no nlinea rities d efined b y co ntinuo us m athem a tica l f unctio ns, f o r c o ntro l system a nalysis it wa s of ten m ore a pp ro p ria te to a pp ro xim a te intrinsic no nlinearity, such as frictio n, or intentio nally introd uced no nlinearity, such as a rela y, b y linear seg m ented cha ra cteristics. The a pp ro a ch is still usef ul to da y b eca use of the f unda m ental insig ht it pro vides into aspects of no nlinea r system b eha vio ur; the fa ct there a re still ma ny co ntrol p ro blems which ca n be ap pro xim ated b y second order dynamics, and also because more than one nonlinearity can be considered.
2.2 Basic Principles

The formulation used in early work on second order systems was to assume a representation in ter ms of the two firs t order equations xo1 = P (x1, x2) xo2 = Q (x1, x2) Equilibrium , or sing ular, points represent a stationary system for the d ynamics and occur when xo 1 = xo 2 = 0 The slope of any solution curve, or trajectory, in the x1 - x2 state plane is dx 2 dx 1 =2 (2.1)
Q (x1, x2) P (x1, x2)
(2.2)
Before considering the no nlinear case further it is important to first examine the linear case, which is rarely discussed in
d etail in texts o n linear co ntrol. Mo st, usually, just restrict the co ntrib utio n to co nsid ering the step resp o nse of a seco nd order system.
19
xo x o1
2.2.1 The Linear Case

The equations for the linear situatio n may be written xo1 = ax1 + bx2 xo2 = cx1 + dx2 or in matrix form xo = Ax ab where A = c c d m. (2.4) (2.3)
Eliminating x 2 fro m eq uatio ns (2.3) b y differentiating the first eq ua tio n a nd substituting fo r xo 2 f ro m the seco nd eq ua tio n and x2 from the first yields xp1 - (a + d) xo1 + (ad - bc) x1 = 0 which after taking Laplace transforms has the characteristic equa tion s - (a + d) s + ad - bc = 0 which is the sa m e as the eigenva lue eq ua tio n of the ma trix A, with typically repla cing s when discussed in m athem atical texts. The ro ots, o r eig envalues, of the eq uatio n d ep end o n the values of the elem ents of A. If the eq uatio n is written in the f orm m + Bm + C = 0 then the two eigenvalues are given by m1 = - B + (B - 4C)
2 1/2 2 2
/2 and m2 = - B - (B - 4C)
1/2
/2
The solutio n to the differential equa tion is of the form x1 (t) = K 1 e x1 (t) = K 1 e
m1 t m1 t
+ K2 e + tK 2 e
m2t m2t
for m1 ! m2 for m1
= m2
For the linear system there is only one singular point at the origin x1 = 0, x2 = 0 of the state plane and the beha viour of the motio n near to the sing ular point depends on the eigenvalues.
20
Trajectories ca nnot intersect and ca n o nly meet at a singular point. The following fo ur cases can occur for the sing ular point. (i) m1 and m2 are both real and have the same sign (either positive or negative). (ii) m1 and m2 are both real and have opposite signs. (iii) m1 and m2 are complex conjugates with no n zero real parts. (iv) m1 and m2 are imaginary, that is complex conjugates with zero real parts. Case (i) co rresp o nds to a sing ula r point k no wn as a no de, which is sta ble when the eig envalues are neg ative a nd unsta ble when they are po sitive. The fo rm of the trajecto ries near to the sing ular point is sho wn in the sim ulatio n results of Fig ure
2
of (1,-2)T and (1,-0.5)T. Arro ws are typically pla ced o n the trajectories sho wing the directio n of mo tio n which will b e to wa rd s the sing ular point for a stable nod e, as sho wn in the fig ure, a nd a wa y fro m it fo r a n unstab le o ne. Trajectories sta rting o n a n eig envecto r rem ain o n it as is clearly seen in the f ig ure. For the stable no de a ll tra jectories, a pa rt fro m tha t starting o n the first eig envect o r tend to wa rd s the no de alo ng the seco nd eig envecto r, i.e the o ne with the sm allest slo p e. The plo t sho wing trajecto ry m otio ns from different initial cond itio ns is known as a state or phase portrait, although the latter na me is often reserved for the special case where xo1 = xo2 , that is a = 0 and b = 1 in equation (2.3)
Figure 2.1 Phase Portrait for a Node
The sing ular point in case (ii) is a saddle point which is o nly theoretically reachable from initial co nditions on the eigenv ector of the stable solutio n asso ciated with the nega tive eig envalue. Any sm all perturba tio n will ca use the trajecto ries to diverg e as shown by the phase portrait of Figure 2.2 and move outwards as shown by the arrows.
21
2.1, for the transfer functio n 1/ (s + 2.5s + 1) which has eigenvalues of -0.5 a nd -2, with correspo nding eigenvectors
Figure 2.2 Phase Portrait for a Saddle Point
A p ha se p ortrait fo r ca se (iii) is sho wn in Fig ure 2.3 where the sing ula r p oint is a fo cus. The f o cus is sta ble whe n the real part is negative, so that trajectories spiral towards it as marked in the figure, and unstable when the real part is positive. Four trajectories are shown in the fig ure.
Figure 2.3 Phase Portrait for a Focus Showing Four Trajectories
In the final ca se (iv) the sing ular p oint is a centre, with the oscilla to ry m otio n pro d ucing co ncentric trajectories, a s sho wn in Fig ure 2.4. Their size depends on the initial conditio ns, which in a physical situatio n, such as an ideal oscillating pendulum, is defined by the initial inp ut energy.
22
Figure 2.4 Phase Portrait for a Centre
2.2.2 The Nonlinear Case

In g eneral it m a y no t b e po ssib le to ob tain a nalytical solutio ns to even seco nd ord er no nlinea r diff erential eq uatio ns so the va lue o f the p ha se p la ne a p pro a ch in early work wa s to allo w ap p ro xim ate solutio ns to b e o btained using gra p hica l techniques for sketching phase plane trajectories. The basic approach used was to determine the slope at a sufficient num b er of p oints in the state pla ne to allo w a picture of the m otio n to b e ob tained s ta rting fro m a ny initial co nditio ns in the x1 - x2 state plane. From equation (2.2) a trajectory will have a slope r when it crosses the curve rP (x1, x2) + Q (x1, x2) = 0 (2.5)
Destination MMU
MMU is proud to be one of the most popular universities in the UK. Some 34,000 students from all parts of the globe select from its curricula of over 1, 000 courses and qualifications. Please click the advert We are based in the dynamic yet conveniently compact city of Manchester, located at the heart of a sophisticated transport network including a major international airport on the outskirts. Parts of the campus are acclaimed for their architectural style and date back over 150 years, in direct contrast to our teaching style which is thoroughly modern, innovative and forward-thinking. MMU offers undergraduate and postgraduate courses in the following subject areas: Art, Design & Performance Computing, Engineering & Technology Business & Management Scienc e, Envir onm ent al St udies & Geogr aphy Law, Education & Psychology Food, Hospitality, Tourism & Leisure Studies Humanities & Social Science 23
For more details or an application form please contact MMU International. email: international@mmu.ac.uk telephone: +44 (0)161 247 1022 www.mmu.ac.uk/international
By selecting a ra nge of values of r, dra wing the co rresp o nding curves a nd m a rking the slop e r with a sho rt a rro w a t which a trajectory crosses, allows phase portraits to be sketched. This is usually known as the method of isoclines. Sometimes the lengths of the a rro ws are d ra wn in p ro p ortio n to the velo city at the point a nd the plot is then called a vecto r field plo t. Typically a second order nonlinear differential equation representing a control system with smooth nonlinearity can be written as xp + f (x, xo) = 0 a nd if this is rearra ng ed a s two first ord er eq ua tio ns, cho osing the p ha se va ria bles, tha t is the o utp ut a nd its d erivative, as the state variables, then x1 = x, x2 = xo and the equation ca n be written as xo1 = x2 xo2 = - f (x1, x2) (2.6)
which is a special case of equation (2.2). It may have multip le singular points but they will all lie on the x1 axis, that is for x2 = 0. A p hase p o rtrait ca n theref ore be ob ta ined f or a ny f unctio n, f, using the m etho d of iso clines, o r p ossib ly o ther m ethod s which can be found in the literature [2.1]. Digital simulatio n methods now enable phase plane trajectories to be easily displayed and they can often provide a clearer picture of the system behavio ur than time response plots for x1 and x2. Ma ny situatio ns in science a nd eng ineering m a y be m od eled a pp ro xim ately b y eq uatio n (2.1) or the sp ecial case eq ua tio n (2.6). Some good examples can be found in the recent book given in reference [2.2] whilst here just one well known example is taken.
2.2.3 An Example
Many investigations using the p hase plane technique were concerned with the possibility of a nonlinear differential eq uation having a limit cycle solutio n, the strict name for an oscillation in a nonlinear system. When a limit cycle exists it results in a clo sed trajectory in the p ha se pla ne a nd typical of such investiga tio ns wa s the work of Va n d er Pol o n o scillato rs. He considered the equation
2
(2.7)
where n is a positive constant. The phase plane form of this equation can be written as xo1 = x2
2
The slope of a trajectory in the phase plane is dx 2 = dx 1

2
x2
(2.8)
24
xp - n (1 - x ) xo + x = 0
xo 2 = - f (x 1 , x 2 ) = n (1 - x 1 ) x 2- x
x o2 x o1
n (1 - x1 )2x - x1
It ha s o ne sing ula r p oint at the origin (0, 0) a bo ut which the linea rised eq ua tio n is xp - nx o + x = 0 , so tha t the sing ular point depends upon the va lue of . It is an unstable focus for < 2 and an unstable node for > 2. All phase plane trajectories have a slope of r when they intersect the curve
2
(2.9)
Fig ure 2.5 sho ws p ha se po rtraits o btained using iso cline sk etching a nd Fig ure 2.6 sho ws sim ula tio n results o btained fro m the initial co nd itio ns of (-3, 3) a nd (-0.1, 0.1) f o r = 0.1 a nd 1. It ca n b e seen, a s exp ected , tha t th e resulting lim it cycles are d ep end ent o n with that f or the sm all va lue of b eing alm o st elliptica l. They a re rea ched f ro m o utsid e fo r the first initial condition and from inside for the second initial condition. A limit cycle has this feature that it i s reached from different initial conditions and in this case it is stable as trajectories tend to it from both without and within. In contrast if = 0 then one has an oscillation, no t a limit cycle, and the magnitude of the oscilla tion is determined by the initial conditio ns. Since the system has no damping there is no loss of energy and its value is set by the initial conditions.
Figure 2.5 Phase portraits of the Van der Pol equation from isocline sketching
25
rx2 = n (1 - x1 ) x2 - x1
Figure 2.6 Phase plane plots of the Van der Pol equation for two values of
2.3 The Phase Plane for Systems with Linear Segmented Nonlinearities
It is a n ad va nta g e when using the p ha se pla ne ap p ro a ch to ha ve no nlinearities which are d escrib ed b y linea r seg m ented chara cteristics. This is b eca use it results in a p hase pla ne which ca n be divid ed up into diff erent reg io ns with diff erent linear differential eq ua tio ns d escribing the m otio n in ea ch regio n. Altho ug h the d ifferential eq ua tio ns are linea r the m athema tical phase plane equations are only simple enough for calcula ting analytical solutio ns by hand in a few cases. To illustrate these ba sic co ncepts three exa m ples a re co nsid ered in the next sectio n. Sim ple d yna m ics a re used in so m e cases to allo w analytical solutio ns for the motion b ut also shown are some simulations with other dyna mics.
Develop the tools we need for Life Science

Masters Degree in Bioinformatics
Bioinformatics is the exciting eld where biology, computer science, and mathematics meet. We solve problems from biology and medicine using methods and tools from computer science and mathematics.
Read more about this and our other international masters degree programmes at www.uu.se/master
26
2.3.1 Example 1 Nonlinear Output Derivative Feedback

The Simulink diagram for this example is given in Fig ure 2.7, where the only no nlinearity is in the derivative of output feed b a ck pa th. To plo t the p ha se pla ne using Sim ulink the X -Y o scillo sco p e ca n b e used o r alternatively the o utp ut a n d its derivative can be fed to simout blocks for plotting after the simulation run has finished. The simout blocks must be p ut in the arra y m o d e a nd the tim e vecto r is a vaila ble in to ut. Of ten using the d ef a ult integ ratio n alg orithm in Sim ulin k wa vef orm s do not ap p ea r sm o oth so o ne m a y ha ve to use a n alg orithm with a sm all fixed step size or lim it the step size in a n a lg o rithm . R esults fro m the X -Y sco pe ca n b e rea d to the displa y a ccura cy. Mo re a ccurate results ca n b e o btained fro m the simo ut blo cks b ut plotting the d ata ma y not b e straig htf orwa rd a s o ne m a y ha ve uneq ual vecto r lengths of da ta stored when a default integration algorithm is not used. If the nonlinearity is replaced by a gain K then the characteristic equatio n is
2
(2.10)
Figure 2.7 Simulink diagram for system with dead zone feedback
It is clear that for K < 0.6 the linear system is unstable and stable for K > 0.6. For the dead zone the gain is zero for a sm all sinusoida l sig nal b ut a pp ro a ches unity f or a la rge sinusoid al sig na l. T hus, o ne expects the eff ect of the d ea d zo ne no nlinea rity to cha ng e the da m p ing in so m e wa y such tha t the system is unsta ble fo r sm all sig nals a nd stab le f o r la rg e sig nals. This sug g ests the existence of a sta ble limit cycle. Fig ure 2.8 sho ws p ha se pla ne p lo ts sta rting f ro m (-1, 0) f o r the linear second order system with characteristic equa tion
2
(2.11)
for various values of g , the damping ratio. For the unstable case with g = -0.05 the trajectory slowly spirals outwards and for the stable responses they converge to the origin in a spiral manner for g < 1, when the origin is a focus and directly with no o vershoo t fo r g 1, when the o rig in is a nod e. The m athem a tica l expressi o ns f o r these curves are q uite co mplica ted , so to sk etch them o n a p hase pla ne the m ethod of iso clines o r so m e special m etho ds ma y b e used [1]. The d ea d zo ne is set a t levels of 0.5, so tha t the tra jectories d escribing the m otio n are tho se of a linear sec o nd o rd er system with a d a mping ra tio of -0.3 f or | x2| < 0.5 a nd 0.2 fo r | x 2| > 0.5. Pha se p la ne plots ha ve been o btained b y sim ula tio n a nd are shown from initial co nditio ns of (-8, 0) and (0.1, 0) in Figure 2.9. The resulting limit cycle is clearly shown in the figur e.
27
s + (K - 0.6) s + 1 = 0
s + 2gs + 1 = 0
Figure 2.8 Phase plane plots for second order system
Figure 2.9 Phase plane plots for system of Figure 2.7
2.3.2 Example 2 Relay Position Control

Consider a basic relay position control system with no nlinear velocity feedback having the Simulink diagram shown in Fig ure 2.10. The two rela ys in p a rallel sim ulate a rela y with dea d zo ne a nd hysteresis which has the chara cteristic sho wn in Figure 2.11, where for positive inp uts the switch on level is d + 9 and the switch off level is d - 9 . The relay parameters are set to give an output, h, of 1, and 0.
28
Figure 2.10 Simulink diagram of the relay control system for example 2
Figure 2.11 Relay with dead zone and hysteresis
It is a ssum ed initially that the hysteresis in the rela y is negligible (i.e. 9 = 0 ) a nd that there is no sa tura tio n in the velo city feedb a ck p ath.. Denoting the system po sitio n o utp ut b y x 1 a nd its deriva tive xo 1 b y x 2 then the f eed ba ck to the rela y inp u t is assum ed to b e - x 1 - m x 2 . The rela y o utp ut of 1 or 0 is eq ual to xp 1 /K , where in the Sim ulink dia gra m b oth K a nd are chosen eq ual to 0.5. Taking the d ea d zo ne of the rela y ! d to be eq ual to 1, the motio n of the system is d escrib ed b y K if - x1 - mx2 > 1 x = *0 if |- x1 - mx2 | < 1 (2.12) - K if - x1 - mx2 < 1 Thus the eq uatio n of m otio n in the p ha se pla ne cha nges at the lines x 1 + m x 2 = ! 1 a nd these ca n be dra wn o n the plot, as sho wn in Fig ure 2.12, to d ivide up the p hase p la ne into the three regio ns where the m o tio n is describ ed b y the ab o ve three simple linear second-order differential equatio ns.
29
Figure 2.12 Switching lines in the phase plane
The solutio n of xp1 = K (2.13)
WELCOME TO OUR WORLD OF TEACHING!

INNOVATION, FLAT HIERARCHIES AND OPEN-MINDED PROFESSO RS
STUDY IN SWEDEN HOME OF THE NOBEL PRIZE

CLOSE COLLABORATION WITH FUTURE EMPLOYERS SUCH AS ABB AND ERICSSO N
SASH A SHAHB AZI LEFT IRAN FOR A MASTER S IN PRODUCT AND PROCES S DEVELOPMENT AND LOTS OF INNEB AND Y
HELL TELL YOU ALL ABOUT IT AND ANSWER YOU R QUESTIONS AT MDUSTUDENT.COM
www. mdh.se
30
An Introduction to Nonlinearity in Control Systems is obtained by writing it in the form x2 dx 2 dx 1 =K
(2.14)
and integrating with the initial cond itio ns for x1 and x2 of x10 and x20, respectively, to give
2 2
(2.15)
where x10 and x20 are the initial values of x1 and x2 . Since equatio n (2.15) describes a parabola, which for the special case of K = 0 has the solution x2 = x20 , it is easy to calculate the systems response from any initial co nditio n (x10, x20) in the phase plane. Figure 2.13 shows the response from (-2, 0) with = K = 0.5 as given in Figure 2.7. The initial parabola, with K = 0.5, starts from A and meets the first switching boundary x1 + 0.5x2 = - 1 at B; the ensuing motion is horizontal, tha t is, at constant velocity, until the second switching boundary x1 + 0.5x2 = 1 is reached at C. The ensuing parabola, with K = 0.5, m eets the sa m e switching bo und a ry a gain at D. The next m otio n is a ga in at co nsta nt velo city until the p oint E is rea ched o n the first switching b o und a ry. At this po int the fo llo wing p ara bo lic m otio n, with K = 0.5, bring s the trajecto ry straig ht b ack to the switching bo und ary. The situa tio n has theref ore arisen where the m otio n at either side of the switching line is directed back to it. Thus, the resulting motio n, which is known as a sliding mode, is directed along the switching line to come to rest at (-1, 0). In theory switching takes place at an infinite rate but in practice the relay will have some hysteresis, which will decrease the switching rate. By solving the relevant equatio ns it can be shown that B = (-1.39, 0.781), C = (0.610, 0.781), D = (1.14, -0.281) and E = (-0.86,-0.281). If the first integrator is replaced by the transfer function 1/(s + a) to provide some damping then the a nalytica l so lutio n is not a s easy. The m otio n in the p ha se pla ne whe n the rela y o utp ut is zero is still linea r, b ut no w with a slo p e of -1/a, a nd whe n the rela y o utp ut is 1 it is no lo ng er p ara b olic. Sim ulatio n resp o nses are a lso sho wn in Fig ure 2.13 fo r va lues of a eq ua l to 0.1 a nd 0.3. R esp o nses fro m a ny o ther initial co nditio ns a re o b vio usly ea sy to f ind, b ut, fro m the resp o nses sho wn, several a spects of the system s b eha vio r a re rea dily ap p arent. In pa rticular the system is seen to be sta ble since all resp o nses will m o ve inwa rd , p ossibly with severa l o versho o ts a nd und ersho o ts, a nd will f inally slid e d o wn a switching boundary to 1. Thus a steady-state error of unit magnitude will result from any motion. Finally a word of warning tha t difficulties may arise in simulating systems with sliding, since if a variable step length integratio n algorithm is used it may reduce the step length to zer o and stop in the sliding motion. Many simulation languag es have special routines for taking care of the disco ntinuo us sliding motion, b ut good results ca n usually be obtained by using a fixed step length integratio n algorithm or by incorporating a small hysteresis in the relay.
31
x2 - x20 = 2K (x1 - x10)
Figure 2.13 Initial condition response for system with
W hen the velo city f eedb a ck sig na l sa turates, tha t is, when | m x 2 | > h , the inp ut sig nal to the rela y is - x 1 ! h . Thus, if h is chosen equal to 0.5 when | x 2 | > 1 the switching bo undaries become vertical lines as sho wn dotted in Figure 2.12, altho ugh the eq uatio ns d escribing the m o tio n b etween the b o und a ries rem ain una ltered . Theref ore f or a larg e step inp ut of r, the equivalent of starting from the initial condi tion (-r, 0), the response will become more oscillatory when the velocity saturates as illustra ted in Fig ure 2.14 f or values o f r eq ua l to 6 a nd 3 a nd saturatio n of the velo city f eed ba ck sig nal at 0.5 for inp uts g reater tha n 0.5. The p a ra b olic curves a re clea rly seen, since the da m p ing has b een tak en a s zero, as also is the fact that the switching lines have become vertical for | x2 | > 1 .
Figure 2.14 Phase plane plots for step inputs of 6 and 3 with velocity saturation but no hysteresis or damping.
Again if there is no velocity saturation in the feedback path but the relay has a finite hysteresis, D , then the switching bo und a ries a re a g ain straig ht lines of slo p e -1/ b ut diff erent a ccording to whether x 2 is po sitive o r neg ative.In the general case they are x 1 + m x 2 = - d + D a nd x1 + mx2 = - d - D and x1 + mx2 = d + D for x2 positive, and x1 + mx2 = d - D for x2 negative. (2.16)
32
It is easily shown from symmetr y considerations, when the above lines are drawn on a phase plane, that a limit cycle exists for the system with no damping with a maximum value of x2 = D/m . Figure 2.15 shows simulatio n results for d = 1 and m = D = 0.5 . The responses from -1.51 and -4 are shown converging to the limit cycle from within and witho ut. If the system were sta bilized b y repla cing the first integrato r b y the tra nsfer f unctio n 1/( s + a) then the m o tio n could come to rest, dependent on the input step magnitude, with the relay inp ut in the range 1.5.
Figure 2.15 Initial condition responses resulting in a limit cycle
33
2.3.3 Example 3 Position Control with Torque Saturation

Fig ure 2.16 sho ws a Sim ulink diag ra m f or a p ositio n co ntrol system with no visco us d a m ping a nd with no nlinea r eff ects due to torque saturation and Coulo mb frictio n.
Figure 2.16 Simulink diagram of control system for Example 3
The differentia l equation of mo tion in phase variable form has xo2 = fs (- x1) - 0.5sgn (x2) (2.17)
where fs denotes the saturation no nlinearity and sgn the signum functio n, which is +1 for x2 > 0 and 1 for x2 < 0 . There are six linear differential equatio ns describing the motion in different regions of the phase plane. For x2 positive, equation (2.17) ca n be written xp1 + fs (x1) + 1/2 = 0 so that for (a) (b) (c) (2.18)
o o
x2 + ve, x1 < - 2, one has x1 = x2, x2 = 3/2, a parabola in the phase plane. o
x2 + ve, - 2 # x1 # 2, one has x1 = x2, x2 + x1 + 1/2 = 0, a circle in the phase plane. o
x2 + ve, x1 > 2, one has x1 = x2, x2 = - 5/2, a parabola in the phase plane.
Similarly for x2 negative, (d) (e) (f)
o o o
x2 - ve, x1 < - 2, one has x1 = xo x2 = 5/2, a parabola in the phase plane. 2, x2 - ve, - 2 # x1 # 2, one has x1 = x2, x2 + x1 - 1/2 = 0, a circle in the phase plane. o
x2 - ve, x1 > 2, one has x1 = x2, x2 = - 3/2, a parabola in the phase plane.
34
Beca use a ll the p hase pla ne trajectories are d escrib ed b y sim ple m athem atical expressio ns, it is straig htfo rward to calculate specific phase plane trajectories.
Figure 2.17 Response from initial condition (-6, 0) for Example 3
Caref ul exa mina tio n of Fig ure 2.17, which sho ws the sim ulatio n result f o r a respo nse f ro m ( -6, 0), reveals the cha ng es in the trajectory shape when crossing the lines x1 = 2.
2. 4 Conclusions
This chapter has covered the basic concepts in the analysis of second order systems using the phase plane approach. Since m a ny sim ple co ntrol systems ca n be a pp ro xim ated b y no nlinear seco nd o rd er diff erentia l eq uatio ns the a p pro a ch is po werf ul. It ca n b e used when mo re tha n o ne no nlinea r element exists a nd is p articularly usef ul when the no nlinea rity ca n b e ap p ro xima ted b y linear seg m ented cha ra cteristics. E xa m ples ha ve b een given using a do ub le integ rator pla nt tra nsf er functio n with a co nsta nt inp ut, which results in sim ple ma them atical expressio ns fo r the p ha se pla ne trajecto ries, so that certain features can be easily illustrated. Mathema tical expressions for the trajectories can always be obtained, since they are given by the solutions of linear diff erential eq ua tio ns whe n the no nlinea rities a re linear seg m ented cha ra cteristics, b ut they a re of ten q uite co m plica ted . Pha se p la ne plo ts ca n ea sily b e o b ta ined with m o d ern sim ulatio n la ng ua g es b ut the f a cility to sk etch a p ha se p ortrait is a n im po rta nt aid to und ersta nding the system b eha vio ur. The results ca n also b e usef ul f or co m pa ring results ob tained b y the approximate DF method, to be introduced in the next chapters, when it is applied to second order systems.
2.5 References
2.1 Atherton, DP 1975, Nonlinear Control Engineering: Describing Function Analysis and Design, Chap ter 2, Van Nostrand Reinhold,London. 2.2 Astrom, KJ & Murray, RM 2008, Feedback Systems: An Introductio n for Scientists and Engineers, Chapters 2, 3 &4Princeto n University Press.
35
2.6 Bibliography
Blaquiere, A 1966, Nonlinear Systems Analysis, Academic Press, New York. Cosgriff, RL 1958, Nonlinear Control Systems, McGraw Hill, New York. Cunningham, WJ 1958, Introductio n to Nonlinear Analysis, McGraw Hill, New York. Gelb, A & Vander Velde, WE 1968, Multiple-Inp ut Describing Functio ns and Nonlinear System Desig n, McGraw Hill, New York. Gibson, JE 1963, Nonlinear Auto matic Co ntrol, McGraw Hill, New York. Graham, D & McRuer, D 1961, Analysis of Nonlinear Control Sys tems, Wiley, New York. Hayashi, C 1964, Nonlinear Oscillatio ns in Physical Systems, McGraw Hill, New York. Kalman, RE 1954, Phase plane analysis of automatic co ntrol systems with nonlinear gain elements. Trans. AIEE, Vol 73(II), pp 383-390. Kalman, RE 1955, Analysis and design principles of second and higher order saturating servomechanis ms. Trans. AIEE, Vol74(II), pp 294-308. Minorsky, N 1962, Nonlinear Oscillations, Van Nostrand, New York. Slotine, JJE & Li, W 1991, Applied Nonlinear Control, Prentice Hall, New Jersey. Struble, RA 1962, Nonlinear Differential Equations, McGraw Hill, New Yor k. Thaler, GJ & Pastel, MP 1962, Analysis and Design of Nonlinear Feedback Control Systems, McGraw Hill, New York. Van der Pol, B 1934, Nonlinear theory of electric oscillatio ns. Proc. IRE, Vol 22, pp 1051-1086. West, JC. 1960, Analytical Techniques for Nonlinear Control Systems, EUP, London. West, JC, Do uce, JL & Na ylor, R 1954, The eff ects of so m e no nlinea r elem ents o n the tra nsient p erf orm a nce of a sim ple r.p.c. system possessing torque limitatio n. Proc. IEE, Vol 101, pp 156-165.
Brain power
By 2020, w ind could prov ide one-tenth of our p lanets electricity needs. A lready today , SKF s innov ativ e know how is crucial to ru nning a large propo rtion of the w orlds w ind turbines. U p to 25 % of the generating costs relate to maintenance. These can be reduced d ramatically thanks to our sy stems for on-line conditio n monitoring and automati c lubricati on. We help make it more economical to create cleaner, cheaper ene rgy out of thin air. By sharing our expe rience, expertise, and creativ ity , industrie s can boost performance bey ond expectat ions. Therefore w e need the best employ ees w ho can meet this challenge !
The Power of Knowledge Engineering
Plug into The Power of K nowledge E ngineering. Visit us at www.s kf.com/kno wledge
36
The Describing Function
3 The Describing Function

3.1 Introduction
The d escribing f unctio n, which will b e a bb revia ted DF, m ethod was d evelo ped sim ulta neo usly in several co untries d uring the 1940s. E ngineers f o und tha t co ntrol systems which were b eing used in m a ny ap plicatio ns, fo r exa m ple g un p ointin g and antenna control, could exhibit limit cycles under certain conditions rather tha n move to a static equilibrium. They realized this instab ility wa s d ue to no nlinea rities, such a s b a cklash in the gea rs of the co ntrol system , a nd they wished to obtain a d esig n m ethod which co uld ensure the resulting systems were f ree fro m limit cycle o pera tio n. They ob served that when limit cycles occurred the wa vefo rms at the s ystem o utp ut were of ten ap p ro xim ately sinuso id al a nd this ind ica ted to them a p o ssible a nalytical ap pro a ch, na m ely to a ssum e that the sig na l at the inp ut to the no nlinea r elem ent in the lo o p was a sinuso id . Since then there ha ve b een m a ny develo p m ents in t erm s of b oth using the DF co ncep t f o r other typ es o f sig na ls a nd the p ro blem s, or p heno m ena , which they ca n b e used to stud y. Mo re will b e said o n these a sp ects later b ut we begin b y co nsid ering the initial p ro blem of investiga ting the po ssibility of a lim it cycle in a f eed b a ck system using the DF or S (sinusoidal) DF as it is often named. Co nsid er the f eed b a ck system sho wn in Fig ure 3.1 co ntaining a single static no nlinea rity n(x) a nd linea r d yna m ics given b y the tra nsf er f unctio n G (s) = G c (s) G 1 (s) . If a limit cycle exists in the a uto no m o us system , tha t is with r(t ) = 0, with the o utp ut c(t ) a pp ro xim ately sinuso id al, then the inp ut x(t ) to the no nlinea rity mig ht also b e expected to b e near sinusoida l. If this a ssum ptio n is m a de the f unda m ental o utp ut of the no nlinearity ca n b e calculated a nd co nditio ns fo r the sinusoidal self-oscillatio n found, if the hig her harmonics generated at the nonlinearity o utput are neglected. This is the concept of harmonic balance, in this case balancing the first harmo nic only. This approach had previously b een used b y p hysicists to investiga te such asp ects a s the g enera tio n of oscilla tio ns in electro nic circuits; the Va n d er Pol oscillato r m entio ned in the previo us cha p ter b eing a n exa m ple. The DF of a no nlinearity is the ref ore d efined a s its g ain to a sinusoid , tha t is the ratio of the f und a m ental of the o utp ut to the a mp litud e of the sinusoidal inp ut. Since the o utp ut fundamental may not be in phase with the sinusoidal input the DF may be complex.
Figure 3.1 A simple nonlinear feedback system
3.2 The Sinusoidal Describing Function

Assume that in Figure 3.1 x (t) = a cos i, where i = ~t and n (x) , where and n(x) is a symmetrical odd nonlinearity, then the output y(t) will be given by the Fo urier series.
37
An Introduction to Nonlinearity in Control Systems 3 y (i) = / an cos ni + bn sin ni n=0 where an = bn = 0 for n even, and in particular a1 = (1/r) and b1 = (1/r)
(3.1)
#0 #
2r
y (i) cos i di
(3.2)
2r 0
y (i) sin i di
(3.3)
The funda mental outp ut from the nonlinearity is a1 cos i + b1 sin i , so that the DF is given by N (a) = (a1 - jb1) /a which ma y be written N (a) = N p (a) + jNq (a) where N p (a) = a1 /a and Nq (a) = - b1 /a Alter natively, in polar co-ordinates N (a) = M (a) e where M (a) = (a 1 + b 1 ) and } (a) = - tan- (b1 /a1) . If n (x) is single valued it is ea sily sho wn tha t b 1 = 0 a nd f ro m sym m etry co nsidera tio ns a s n (x) is o d d, tha t a1 = (4/r) giving N (a) = (a 1 /a) = (4/ar)
1 2 2 1/ 2 j}(a)
(3.4)
(3.5)
(3.6)
(3.7)
/a
(3.8)
(3.9)
# y (i) co s idi
0
r/2
(3.10)
# y (i) co s idi
0
r/2
(3.11)
38
Altho ug h eq uatio ns (3.2) a nd (3.3) a re a n ob vio us ap proa ch to the evaluatio n of the f unda m ental o utp ut of a no nlinearity, they are somewhat indirect, in that one must first determine the output waveform y (i) from the known nonlinear chara cteristic a nd sinusoidal input wa veform . This is a voided if the substitution i = cos - (x/a) is made; in which case, after some simple manipulations, it can be shown that a1 = (2/a) b1 = (2/ar)
1
-a
# xn (x) p (x) dx
p
(3.12) (3.13)
-a
# n (x) dx
p
The functio n p(x) is the amplitude probability density function of the input sinusoidal signal and is given by p (x) = (1/r) (a - x ) 2 2 1/2
(3.14)
and the nonlinear characteristics n p (x) and nq (x) , called the in-phase and quadrature nonlinearities, are defined by n p (x) = [n1 (x) + n2 (x)] /2 and nq (x) = [n2 (x) - n1 (x)] /2 (3.16) (3.15)
where n1 (x) and n2 (x) are the portions of a double valued characteristic traversed by the input for xo > 0 and xo < 0 respectively. For a single-valued characteristic, n 1 (x) = n2 (x) , so that np (x) = n (x) and nq (x) = 0 and integrating equation (3.12) by parts gives a 1 = (4/r) n (0 + ) + (4/ar)
# n' (x) (a - x )
a 2 0
2 1/ 2
dx
(3.17)
where n' (x) = dn (x) /dx a nd n (0 +) = lim fn 0(f) ; a useful expression for obtaining DFs for linear segmented cha racteristics. "
3.3 Some Properties of the DF

An ad va ntag e of the f orm ula tio n of eq uatio ns (3.12) a nd (3.13) is that they easily yield proofs of so m e interesting pro p erties of the DF for symmetrical odd nonlinearities. These include the follo wing: 1. For a double-valued no nlinearity the quadrature component Nq (a) is proportional to the area of the nonlinearity loop, that of no nlinearity lo op ) N q (a) = - (1/a r) (a rea is:
2
2. For two single-valued nonlinearities na (x) and nb (x) , with na (x) < nb (x) for all 0 < x < b, then Na (a) < Nb (a) for input amplitudes a less than b.
39
3. For the sector bounded single-valued no nlinearity that is k1 x < n (x) < k2 (x) for all 0 < x < b then a double-valued nonlinearity if N(a) is replaced by M(a). When the no nlinearity is single valued, it also follows directly from the properties of Fourier series that the DF, N(a), may also be defined as: 1. The variable gain, K, having the same sinusoidal inp ut as the no nlinearity, which minimizes the mea n squared value of the error between the output from the nonlinearity and that from the variable gain. 2. The covariance of the inp ut sinusoid and the nonlinearity output divided by the variance of the inp ut. It can also be shown tha t the gain K is the best transfer function model in the mea n squared error sense. That is the optimum G(s) is K.
k1 < N (a) < k2 for input amplitudes a less than b. This is the sector property of the DF and it also applies fo r
Trust and responsibility

NNE and Pharmaplan have joined forces to create NNE Pharmaplan, the worlds leading engineering and consultancy company focused entirely on the pharma and biotech industries. Ins Arizaga Esteva (Spain), 25 years old Education: Chemical Engineer You have to be proactive and open-minded as a newcomer and make it clear to your colleagues what you are able to cope. The pharmaceutical eld is new to me. But busy as they are, most of my colleagues nd the time to teach me, and they also trust me. Even though it was a bit hard at rst, I can feel over time that I am beginning to be taken seriously and that my contribution is appreciated.
NNE Pharmaplan is t he worlds leading engineering a nd cons ultancy co mpany focused e ntirely on the pharma a nd biotech industries. We employ more tha n 1500 people worldwide and offer global reach a nd local kno wledge along with our all-encompassi ng list of s ervices. nnepharma plan.com
40
3.4 The Evaluation of some DFs

To illustrate the evaluation of the DF for some specific nonlinearities a few simple examples are considered below.
3.4.1 Cubic Nonlinearity

For this nonlinearity n (x) = x a nd using equatio n (3.10) o ne has, a1 = (4/r) = (4/r) a = (4/r) a
3
#
0 3 3
r/2
(a cos i) cos idi

r/2
# #
0 0
cos idi
(3.18)
3
r/2
` 38 + cos22i + co s84i j di = 3a /4 giving

N (a) = 3a /4
2
Alter natively from equatio n (3.12) and using the fact that n(x) is odd, one has a1 = (4/a)
# x p (x) dx
a 4 0
The integral nn =
# x p (x) dx is kno wn as the n

3 n -3 2 2 - 1/2
th
moment of the probability d ensity f unction and for the sinusoidal
distributio n with p (x) = (1/r) (a - x ) 0 nn = *a n (n - 1) (n(n 3) 2) n - -
, n n , has the value
for n odd .... 1 for n even 2
2 1 Therefore N (a) = (4/a ) . 3 . 1 a 4 = 3a 2 /4 a s b ef ore. 242
3.4.2 Saturation Nonlinearity

The DF ca n also b e f o und b y ta king the no nlinearity inp ut as a sin i , in which ca se f or the id eal sa tura tio n cha ra cteristic shown in Figure 3.2, with unit slope in the linear regime and saturation for an input above d > 0.5 , the nonlinearity output waveform y (i) is as shown in the same figure.
Figure 3.2 Saturation nonlinearity, input sinusoidal and output waveforms
41
Again, because of the symmetry of the nonlinearity the funda mental of the outp ut ca n be evaluated from the integr al over a quarter period, so that for a linear regime slope of m and saturation at gives N (a) = 4 ar
# y (i) sin idi

r/2 0
which for a > d gives N (a) = ar ; 4
# ma sin idi + # md sin idi E

a 2 r/2 0 a 1
with a = sin - d/a . E valuatio n of the integ rals gives
N (a) = (4m/r)8 2 - sin42a + d cos aB a which on substituting for d gives the result N (a) = (m/r) (2a + sin 2a) . Since for a < d the characteristic is linear giving N (a) = m , the DF for ideal saturation is mNS (d/a) where 1 m N s (d/a) = for a < d (3.19)
(1/r)62a + sin 2a@ for a > d
Alterna tively o ne ca n evalua te, N(a), f ro m eq uatio n (3.17), which yields, a s the no nlinea rity slo pe is zero f or inp uts greater than N (a) = a1 /a = (4/a r)
2
# m (a
d 0
-x)
2 1 /2
dx
Using the substitution x = a sin i , this gives N (a) = (4m/r)
#
0
cos2 idi = (m/r) (2a + sin 2a) as before.
3.4.3 Relay with Dead Zone and Hysteresis

The characteristic of a relay with dead zone and hysteresis, for d = 1 and D = 0.5 , is shown in Figure 3.3 together with the corresponding input, assumed equal to a cos i , and output waveforms.
42
Figure 3.3 Relay with dead zone and hysteresis, and input and output waveforms
Using equations (3.2) and (3.3) over the interval - r/2 to r/2 and assuming that the input amplitude a is greater than d + D , gives a1 = (2/r)
# h cos idi = (2h/r) (sin b + sin a)

-a 1 1
where a = cos - (d - D) /a and b = cos - (d + D) /a , and b 1 = (2/r)
#
-a
h sin idi = (2h/r)`

2
(d + D) (d - D) a a
2 1/2
j = 4hD/ar . Thus
j4hD ar
2
N (a) = 22h "6a ar
- (d + D) @
+ 6a 2 - (d - D) 2 @1 /2 , -
(3.20)
43
Fo r the alterna tive a pp roa ch o ne m ust first o btain the in -p ha se a nd q uad rature no nlinea rities which a re sho wn in Fig ure 3.4. Using equations (3.12) and (3.13) one obtains a1 = (4/a) x (h/2) p (x) dx +
d-D 2
d+D
d +D
# xhp (x) dx
+ 6a2 - ^d - Dh2@ 1/2,
= 2h "6a ar b1 = (4/ar)
- (d + D) @
d+D
2 1/2
# (h/2) dx = 4hD/ar = (Area of no nlinea rity lo op ) /ar.

d-D
The DFs for other relay characteristics can easily be found from this result or worked out directly. The latter is very straightforward when the relay has no d ead zo ne a s its o utput is a sq uare wa ve of amplitude ! h , which has a fundamental co m po nent of a m plitud e 4 h/r , either in p hase with the inp ut f or D = 0 , or lag ging when is finite. U sing eq ua tio n 3.20 for no hysteresis, D = 0 ; for no dead zone, d = 0 ; and for an ideal relay, D = d = 0 .
Figure 3.4 In phase and quadrature nonlinearities
It is possible to model a typical do uble valued nonlinearity like the abo ve as a nonlinear f unctio n n (x, xo ) a nd the d escribing functio n theory based on this can be found in reference 3.1. Unfortuna tely this description is not correct for any inp ut signal, x, and it has been used to obtain erroneous results for x a random signal. Two tables of DFs for a sinusoidal inp ut are given in the Ap p endix 3.11, o ne fo r sing le valued no nlinea rities a nd the other f or do ub le va lued no nlinea rities. Extensive use is made in the tables of the DF for saturation, Ns (d/a) , defined in equa tion (3.19)
3.5 Nonlinear Models and DFs

No nlinear models, which co nsist of linear segments with m ultiple break points, ca n often be modeled from the no nlinearities in (b ) a nd (c) ab o ve a nd linea r g ains, usua lly co nnected in p a rallel. Fo r exa m ple, the d ea d zo ne cha ra cteristic which f o r a po sitive inp ut ha s a zero o utp ut until a value of a nd then a slo p e of unity is easily sho wn to b e m o deled b y a unit g ain can be modeled from relay characteristics with dead zone and no hysteresis in parallel.
in parallel with a saturation characteristic with slope m = -1, as illustrated in Fig ure 3.5. Also a quantized characteristic
44
Figure 3.5 Modelling of an ideal dead zone characteristic from a unit gain and ideal saturation
Since it is easily shown that the DF of two nonlinearities in parallel is equal to the sum of their individual DFs, the DFs of linear segmented characteristics with multiple break points can easily be written down from the DFs of simpler cha ra cteristics. Several p ro ced ures a re a vaila ble f or ob ta ining a p pro xim a tio ns f or the DF of a g iven no nlinea rity either by num erical integratio n or b y evalua tio n of the DF of a n a pp ro xi m ating no nlinear chara cteristic d ef ined , f or exa m ple, b y a quantized characteristic, linear segmented characteristic or Fourier series.
3.6 Harmonic Outputs

It is sometimes of value to kno w the mag nitudes of the other harmonic outp uts from a nonlinearity with a sinusoidal input. These can be obtained from the usual Fo urier series results for equation (3.1) of an = (1/r) and b n = (1/ r)
2r
# n (x) cos nidi # n (x) sin nidi

2r
(3.21)
(3.22)
In terms of the a m plitud e prob ability d ensity f unctio n fo r a sinusoid it ca n b e ea sily sho wn that these expressio ns b eco m e an = 2 and bn = 2
# n (x) T (x/ a) p (x) dx

a n
(3.23)
-a
# n (x) D (x/a) p (x) dx

a n
(3.24)
where T n (x/a) = cos 6 n cos - (x/a) @ a nd D n (x/a) = sin 6n 1cos - (x/a) @ 1
-a
45
Since f or single valued no nlinearities b n = 0 the m o re im p orta nt result is f or a n . Here T n (x/a) is the Cheb yshev p olyno mial of order n and its value for any n can easily be found from the recursion relationship Tn + 1 (x) = 2xTn (x) - Tn - 1 (x) with T0 (x) = 1 and T1 (x) = x (3.25)
3.7 Sine plus Bias DF and the IDF

So far only DFs for nonlinearities with odd symmetry have been co nsidered. For many control engineering situa tions this m a y b e sa tisf a ctory, as f or exa m ple in the co nsid era tio n of torq ue sa turatio n in a m oto r. Ma n y situa tio ns d o o ccur, ho wever, where no nlinearities do not ha ve o d d sym m etry a nd also even when they d o the o p era tio n m a y no t b e a bo u t the odd symmetric axis. Thus, to handle the general situatio n for a single valued nonlinearity with characteristic n(x) o ne needs to co nsid er a n inp ut co nsisting of a sinusoid p lus bia s, tha t is c + a co s i a nd evaluate b oth the sinusoid al a nd bias output of the nonlinearity. The describing function for the nonlinearity, kno wn as the SBDF, has two components N c (a, c) a nd N a (a, c) the g ains to the b ia s a nd sinusoid , resp ectively. It is easily sho wn f ollo wing the proced ures of sectio n 3.2 that N c (a, c) = (a 0 /c) = (1/c) and N a (a, c) = (a 1 /a) = (2/a)
-a
# n (x + c) p (x) dx # n (x + c) (x/a) p (x) dx

a
(3.26)
(3.27)
-a
Thus fo r the inp ut c + a co s i the o utp ut is cN c (c, a) + N a (c, a) a co s i . The im p orta nt p oint a b o ut the SBDF m o del is the intera ctio n b etween the sig nals ca used b y the no nlinea rity, so that the g ain to b oth sig nals is aff ected b y a cha nge in o ne of them . If a is co nsta nt a nd varied the bia s o utp ut of eq uatio n (3.26) ca n b e plotted a g ainst to give a new no nlinear chara cteristic. This is often called a mo dified no nlinearity a nd denoted b y n (a, c) rather tha n a 0 . The slo p e of the mo dified
nonlinearity at any inp ut value gives the gain g (a, c) to any small unrela ted signal in the presence of a and . That is g (a, c) = d
dc
-a
# n (x + c) p (x) dx
(3.28)
In particular, the gain to a small unrelated signal in the presence of x alone is g (a, 0) , usually called the incremental describing function (IDF) and deno ted by Nic (a) , and is given by N ic (a) = g (a, 0) =
-a
# n' (x) p (x) dx
(3.29)
where n' (x) = dn (x) /dx . E q uatio n (3.29) is the bia s o utp ut fro m the no nlinearity n' (x) with inp ut x. For a ny single valued nonlinearity with n(0+) = 0, equation (3.17) gives a N (a) = (2/r)
2
-a
# n' (x) (a
-x)
2 1 /2
dx
46
An Introduction to Nonlinearity in Control Systems which on differentiating both sides with respect to a gives 2aN (a) + a dN (a) /da = 2
2
-a
# an' (x) p (x) dx
The left ha nd side is 2aNic (a) , so that Nic (a) = N (a) + (a/2) N' (a) where N' (a) = dN (a) /da. (3.30)
3.8 Conclusions
This chapter has covered the definition, some properties, and approaches to evaluating the DF of a nonlinearity for a sinusoidal input. The approach has then been extended to cover the situa tion of a sinusoidal plus bias input. The next chapter will sho w how these results can be used in the approximate analysis of a simple nonlinear feedback system.
3.9 References
3.1 Atherto n, DP 1975, No nlinear Co ntrol E ngineering : Describing Fu nctio n Ana lysis a nd Desig n, Va n No stra nd R einhold , London. 3.2 Gelb, A & Vander Velde, WE 1968, Multiple-Input Describing Functio ns and Nonlinear System Desig n, McGraw Hill, New York.
3.10 Bibliography
Atherton, DP 1975, Nonlinear Control Engineering: Describing Function Analysis and Design, Van Nostrand Reinhold , London. Atherton, DP 1981, Stability of Nonlinear Systems, Research Studies Press, John Wiley, Chiches ter. Cook, PA 1994, Nonlinear Dynamical Systems, 2 nd ed. Prentice-Hall Inter national, Lo ndon. Cosgriff, RL 1958, Nonlinear Control Systems, McGraw Hill, New York. Gelb, A & Vander Velde, WE 1996, Multiple-Inp ut Describing Functio ns and Nonlinear System Desig n. McGraw Hill, New York Gibson, JE 1963, Nonlinear Auto matic Co ntrol, McGraw Hill, New York. Graham, D & McRuer, D 1961, Analysis of Nonlinear Control Sys tems, Wiley, New York. Slotine JJE & Li W 1991, Applied Nonlinear Control, Prentice -Hall Internatio nal, New Jersey.
3.11 Appendix -Tables of Describing Functions

Two tables of DFs, or SDFs, are given below, one for single valued nonlinearities and the other for double valued nonlinearities. Extensive use is made in the tables of the describing functio n for saturation, Ns (d/a) , defined in eqn. (3.19). More comprehensive tables are given in other books [3.1, 3.2] for other describing functions, such as the SBDF and IDF of section 3.7, and for other inputs.
47
48
Table 3.1 Describing Functions of Single Valued Nonlinearities
49
Table 3.2 Describing Functions of Double Valued Nonlinearities
50
Stability and Limit Cycles using the DF
4 Stability and Limit Cycles using the DF

4.1 Introduction
The problem of determining whether a linear feedback system is stable is simply one of determination of the roots of its cha ra cteristic eq ua tio n. This ca n b e d o ne in several wa ys with a p articula rly eleg a nt a nd co nvenient eng ineering o ne, since it can be done from kno wledge of the open loop frequency response locus, being due to Nyquist. On the other ha nd the pro blem of d eterm ining p recisely the sta bility of a n a uto no m o us no nlinea r f eedb a ck system , even o ne sim ply co ntaining o ne linear d yna mic elem ent a nd o ne no nlinearity, has no t b een solved , even tho ug h it ha s exercised the mind s of mathematicians and engineers for ma ny years. The DF pro vides a n approximate met hod for a nswering the problem a nd will be a dd ressed in this cha pter, whilst so m e exa ct a p p ro a ches, k no wn a s ab solute sta bility m etho ds, will b e co nsid ered in cha pter 8. The discussio ns will b e restricted to the f eed b a ck loo p structure of Fig ure 3.1. Since o ur initial investiga tio ns are concerned with the stability of the autono mous system, that is r(t) = 0, the two linear dynamic blocks are in series a nd ca n b e rep resented b y the sing le tra nsf er f unctio n G(s). Further the p ositio n of the single static no nlinea rity altho ug h assumed in the forward path could equally well be in the feedback path.
51
An early co ntrib ution to the problem was a co njecture b y Aizerma nn. He co njectured that if a symmetrical odd nonlinearity was co nfined within a secto r def ined b y stra ig ht lines of slo p e k 1 a nd k 2 then a ny no nlinear system with a no nlinea rity lyin g entirely within the secto r wo uld b e sta ble pro vid ed the linear system was sta ble f o r gains b etween. k 1 a nd k2 . U nf ortunately the conjecture is not true, but i t is if it is modified so that the last phrase reads would be stable or possess a limit cycle provided the linear system was stable for gains between. k1 and k2 Thus, since use of the DF provides an approximate m etho d f o r the d etermina tio n of limit cycles, it also p ro vid es a n a p pro xim a te sta bility test f o r o ur sim ple feed b a ck lo op .
4.2 Limit Cycle Evaluation

by its DF N(a). Thus, the open loop gain to a sinusoid is N (a) G (j~) and a limit cycle will exist if N (a) G (j~) = - 1 (4.1)
To stud y the p ossibility of limit cycles in the a uto no mo us closed loo p system of Fig ure 3.1, the no nlinearity n(x) is repla ced
where G (j ~) = G C (j ~ ) G 1 (j ~ ) . This co nditio n m ea ns tha t the f irst ha rm o nic is b ala nced aro und the clo sed lo op assuming its passage through the nonlinearity is accurately described by N(a). Since G (j~) is a complex functio n of ~ and N (a) may be a complex function of a, a solution to equation (4.1) will yield both the frequency ~ and amplitude a of an assumed sinusoidal limit cycle. Vario us a p pro a ches ca n b e used to exa mine eq uatio n (4.1) with the choice aff ected to so m e extent b y the p roblem . Fo r exa m p le, releva nt fa cto rs m ig ht b e whether the no nlinea rity is single or d o uble valued o r whether G (j ~ ) is a vailable fro m a transfer functio n G (s) or as measured frequency response data. Typically the functions G (j~) and N (a) are plotted sep a ra tely o n Bo d e, Nyq uist, o r Nicho ls dia gra m s. Alterna tively, sta bility criteria such a s the Hurw itz-R o uth o r ro ot lo cus plots may be used for the characteristic equation 1 + N (a) G (s) = 0 althoug h here it should be remembered that the equation is appropriate only for s . j~ . Figure 4.1 illustrates the procedure on a Nyquist diagram, where the G (j~) and C (a) = - 1/N (a) loci are plotted and sho wn intersecting at P f o r a = a 0 a nd ~ = ~ 0 . The DF m etho d theref ore ind ica tes tha t the system ha s a lim it cycle with the inp ut sinusoid to the no nlinearity, x, equal to a0 sin (~0 t + z) , where z depends on the initial conditio ns and time orig in. In pra ctice the lim it cycle will no t b e sinusoida l a nd a 0 is a n a pp ro xim a tio n fo r the a m plitud e o f the f und a m ental co m po nent of the limit cycle. Thus if o ne wishes to estima te the a ccura cy of the DF p redictio n f or a limit cycle o ne sho uld m ea sure the a m p litud e of the f und a m enta l no t the p eak a m p litud e of the wa vef orm as is of ten do ne fo r co nvenience . When the G (j~) and C (a) loci do not intersect, the DF method predicts that no limit cycle will exis t if the Nyquist stability criterion is satisfied for G (j~) with respect to any point on the C (a) locus. Obviously, if the nonlinearity has unit gain for small inp uts, the point (4.2)
(- 1, j0) will lie on C (a) and it may then be used as the critical point, analogous
to the situatio n for a linear system.
52
Figure 4.1 Limit cycle found from intersection of G (j~) and C (a) on a Nyquist plot
When the analysis indicates the system is stable its relative stability may be indica ted by evaluating its gain and phase margin. These ca n be fo und for every amplitude a o n the C (a) locus, so it is usually approp riate to use the minim um values
4.3 Stability of a Predicted Limit Cycle

A further point of interest when a solution to equation (4.1) exists is whether the predicted limit cycle is stable. This is obviously importa nt if the control system is d esigned to have a limit cycle operation, as in the case of a n o n -off temperature co ntrol syste m , b ut it m a y also b e im po rta nt in o ther system s. If , f or exa m ple, a n unstable limit cycle co nditio n is rea che d the signal amplitudes will not become bounded but continue to grow. The stability of a limit cycle, provided only one solutio n is p redicted , ca n in m ost cases b e a ssessed b y a pplying the Nyq uist sta bility criterio n to p oints o n the C (a) lo cus at both sides of the solution point. If the stability criterion indicates instability (stability) for a point on C (a) with a < a0 and stability (instability) for a point on C(a) with a > a0, then the limit cycle is stable (unstable). This result is kno wn a s the Loeb intersection criterio n and in terms of the approximate DF analysis, is a necessary, but not a sufficient criterion for a stable limit cycle. To investig ate the sta bility of a lim it cycle o ne ha s to exa mine the p erturb a tio n eq ua tio n. Co nsid er the tra nsf er f unctio n to be the ratio of two polynomials so that it can be written as G (s) = p (s) /q (s) Then if D denotes the differential operator the differential equatio n for the a uto nomo us system ca n be written q (D) x (t) + p (D) n (x (t)) = 0 If the system has a limit cycle solution x * (t ) a nd this is perturbed by a small amo unt q (D)6 x * (t) + dx (t)@ + p (D) n (x * (t) + dx (t)) = 0 (4.4) (4.3)
dx (t) then
53
An Introduction to Nonlinearity in Control Systems This can be written as q (D)6 x * (t) + dx (t)@ + p (D) n (x * (t)) + p (D) n' (x * (t) dx (t)) = 0 Thus on subtracting equatio n (4.4) one has the perturbatio n equatio n q (D) dx (t) + p (D) n' (x * (t)) dx (t)) = 0
(4.5)
This is a tim e va rying eq uatio n f or the perturb atio n, dx (t ) , which m ust b e stable f or the lim it cycle to b e sta ble. It ca n be sho wn that a necessa ry co nditio n f or this eq uatio n to b e sta ble is tha t the tim e a vera ged eq uatio n m ust b e sta ble. Then if o ne uses the ap proximate d escribing f unctio n solutio n fo r x * (t) , the tim e a verag ed value of the no nlinear term is the IDF of eq ua tio n (3.29), so tha t a necessary co nd itio n fo r stability of the lim it cycle is that the eq uatio n q (D) + p (D) N ic (a) = 0 is stable. In terms of the system transfer functio n this is the characteristic equa tion 1 + Nic (a) G (s) = 0 (4.6)
54
4.4 DF Accuracy
A sp ecif ica tio n fo r m a ny co ntrol system s is that when p erturb ed b y either a step inp ut or disturb a nce they sho uld return to eq uilibrium . For a no nlinear system this m ea ns checking that (a ) the resp o nse will no t g o unb o und ed a nd (b ) no lim it cycle will o ccur. Co nditio n (a ) ca n b e check ed f ro m linear m etho ds using the secto r b o unds of the no nlinea rity pro vided a ny inp ut d o es not ca use a bias at the no nlinearity inp ut a nd (b ) b y the DF m etho d . The DF m etho d, ho w ever, ca n o nly d o this a p pro xim a tely since it a ssum es a ny limit cycle will b e sinuso ida l at the inp ut to the no nlinea rity, which will never b e q uite true in pra ctice. This m ea ns tha t the DF m etho d co uld ind ica te inco rrectly either the existence o r no n -existence of a limit cycle. It is therefore important to have some idea of the validity and accuracy of any DF result. If the DF m etho d predicts a limit cycle then its validity ca n norm ally b e checked b y a ssum ing this sinusoidal sig nal a s the no nlinearity inp ut a nd evalua ting the sig nal f ed ba ck to the no nlinearity, assum ing the lo o p op en at the no nlinea rity inp ut. If the p ercentag e distortio n in this sig nal is less tha n 5% then the DF p red ictio n sho uld b e va lid . It ha s also b een sho w n theoretically that the estima te for the limit cycle frequency, ~ , is more accurate than for the amplitude, a. Further, as is often not clearly p ointed o ut, a is the estima te f or the f und a mental co m po nent in the limit cycle not the p eak a m plitud e. When the DF m etho d d oes no t predict a limit cycle it will b e correct if the C (a) a nd G (j ~ ) lo ci a re not near to intersecting . If this is not the case, however, one can assume a possible limit cycle at the nonlinearity inp ut with amplitude a and freq uenc y ~ corresp o nd ing to the near intersectio n values of C (a) = G (j ~ ) . U sing the d isto rtio n criteria given a bo ve a very lo w value m ust exist fo r the predictio n to be valid when the lo ci a re very clo se. Mo re d eta ils o n these co ncepts are presented in section 5.5.
4.5 Some Examples of Limit Cycle Evaluation

Some examples are given in the follo wing sectio ns to illustrate limit cycle determina tion using the DF method for a section 2.2, altho ugh a first harmonic balance approach rather than the DF is used as the nonlinear term involves two inp uts , x a nd xo . It is po ssible, ho wever, to use DF theory fo r a no nlinea r element d efined b y n (x, xo ) a nd the interested reader is referred to the first source in the bibliography. The second and third examples consider limit cycles in relay systems. It is shown in chapter 6 ho w limit cycles in relay systems can be calculated accurately so that comparisons of the results allo ws so me co m m ents to b e m ad e o n DF a ccura cy. The final exa m ple is intro d uced to sho w the value of the IDF in determining limit cycle stability.
feedback loop containing a single nonlinear element. The first example taken is the Van der Pol equatio n discussed i n
4.5.1 The Van der Pol Equation

If a harm o nic ba la nce a p pro a ch is used f or eq ua tio n (2.4) o ne m a y a ssum e x 1 = a sin ~t a nd x 2 = x o 1 = a~ cos ~t a nd the equation can be written
2
Substituting for x1 and x2 and neglecting frequencies higher tha n ~ in the right ha nd side gives
2 3
55
xo 2 + x1 = n (1 - x1 ) x 2
- a~ sin ~ t + a sin ~t = n (a~ cos ~t - a ~ cos ~t ) .

3
For this to be true req uires ~ = 1 for the left ha nd side a nd a~ = a ~ /a , giving a = 2 , for the right ha nd side. Thus, a limit cycle solutio n, which is ind ep end ent of n , with freq uency 1 ra d /s a nd a m plitud es of 2 f or x 1 a nd x 2 , is p redicted . The result is therefore only reasonably accurate for small values of n , where the phase plane plot is almost elliptical.
4.5.2 Feedback Loop Containing a Relay with Dead Zone

For this example the feedback loop of Fig ure 4.1 is co nsidered with n(x) a rela y with dead zo ne and G (s) = 2/s (s + 1) . The DF for this relay is obtained from equation (3.20) with D = 0 and is given by N (a) = 4h (a - d )
2 2 1/ 2 2
/a r for a > d,
which is real because the nonlinearity is single valued. A graph of N(a) against a is given in Figure 4.2.
Figure 4.2 Relay with dead zone and its DF
It shows that N (a) starts at zero, when a = d , increases to a maximum, with a value of 2h/rd at a = d 2 , then decreases toward zero for larger inputs. The C(a) locus, shown in Figure 4.3, lies on the negative real axis starting at 3 a nd returning there after rea at the point ( -1, 0) va freq uency of ta - ~ 45c that is ~ = unctio n a (j ~ ) cro sses real axis, as shown in Figure 4.3, ching a m a xim umat alue of - rd/2h .nThe=given tra nsf er f1 rad /sec,G nd therefore the neg ativ e
1
cuts the C(a) lo cus twice a t the point P. The two p ossib le limit cycle a mp litud es a t this f req uency ca n b e fo und b y solving ar
2
4h (a - d )
2 1/ 2
= G (j~) | ~ = 1 = 1
which assuming d = 1 and h = r gives a = 1.04 and 3.86. Using either the Loeb or the IDF criterion shows the smaller amplitude limit cycle to be unstable and the larger one to be stable. If a condition similar to the lower amplitude limit cycle is excited in the system, an oscillation will build up and stabilize at the higher amplitude limit cycle.
56
Figure 4.3 Solutions for the two limit cycles at P one unstable the other stable
The exact frequencies of the limit cycles, obtained using the approach given in Chapter 6, for the smaller and larger a m plitud es a re 0.709 a nd 0.989 resp ectively. Altho ug h the tra nsf er f unctio n is a g oo d lo w p ass filter, the freq uency of the sma ller a m p litud e is hig ble) limit cycle is tra nsfer function ccura tely b eca + 1) , then no limit m the rela y is a narro w pulses, which(unstahly distorted. If the no t p redicted a of G (s) is K /s (suse the o utp ut f ro cycle will exist in wa vef orm with
2
the feedback loop, and it will be stable according to the DF method if ~ (1 + ~ ) K

2 ~= 1
< rd 2h
(4.7)
Sharp Minds - Bright Ideas!

E mp lo y e es a t F O S S A n a ly ti c a l A / S a re l iv in g p ro o f o f t h e c o mp a ny v a lu e - Fi rst - u s in g
Th e F a m ily o w n ed F O S S g r o u p is th e w o r ld le ad er as s u p p lier o f d ed icat ed , h ig h - tech an aly tic al s o lu tio n s w h ich m e as u r e an d co n tr o l th e q u ality an d p r o d u ctio n of ag r icu ltu r al, f o o d , p h ar m a ceu t ica l an d ch e m ic al p r o d ucts . M ain a ctiv iti es ar e in i tiat ed fr o m D en m ar k , S w ed en an d U S A w ith h e ad q u ar ter s d o m i ciled in H ill er d , D K . Th e p r o d u cts ar e m ar k e ted g lo b ally b y 2 3 s al es co m p an ies an d an ex ten s iv e n et of d is tr ib u to r s . I n lin e w i th th e co r ev alu e to b e F ir s t , th e co m p an y in ten d s to ex p an d its m ar k et p o s itio n .
n e w in v en ti o ns to m ak e d e di c a t ed s ol ut io ns fo r ou r cu st o m e rs . W it h sh a rp m in ds a n d c ros s fu n ct i on a l t e a m w o rk, w e c on st a n tl y s t ri v e to d e v el op n e w u ni qu e p rod u ct s W o ul d y ou l ik e t o j oi n ou r t e a m ? F O S S w o rks di li g en tl y w i th i nn ov a ti on an d d e v e lo p m en t a s b a si s fo r i t s g ro w th . It i s re fl e ct e d i n th e fa ct th a t m o re th a n 2 0 0 o f t h e 1 20 0 e m pl oy e e s in F O S S w o rk w i th R e s e a rch & D e v el op m e nt i n S c a nd in a vi a an d U S A . E ng in e e rs at F O S S w o rk i n p rod u ct i on, d ev e lo p m en t a nd m a rk et i ng, w it hi n a w i d e ra ng e o f di ffe re nt fi e ld s, i. e. Ch e m is t ry , El e c t ro ni c s, M e c h an i cs , S o ft w a re , O p ti c s, M i c ro b io lo gy, Ch e m o m et ri cs . We offer A ch al l en gi n g j ob i n a n in t er na ti o na l a nd i nn o va ti v e co m p an y t ha t i s l ea di ng i n i t s f i el d. Y ou wi ll g e t t h e opportunity to work with the most advanced technology together with highly skilled colleagues. R ea d m or e a bo ut F O S S at w w w .f os s. d k - or g o d ir e c tl y t o ou r s tu d en t si t e w w w. f os s.d k /s ha r p m in ds w h e r e you can learn more about your possibilities of working together with us on projects, your thesis etc.
Dedicated Analytical Solutions

FOSS Slangerupgade 69 3400 Hillerd Tel. +45 70103370
www.foss.dk
57
that is, K < rd/h . Again, if d = 1 and h = r , then the system is stable if K < 1, which may be compared with the exact result for stability of K < 0.96.
4.5.3 Feedback Loop with On Off Relay

Consider the feedback loop of Figure 3.1 with n(x) an ideal relay, Gc (s) = 1 , and the plant with a transfer functio n
3
G 1 (s) = 10/ (s + 1) . E q uatio n (3.20) with d = D = 0 gives N (a) = 4h/ar . Thus the C(a) lo cus, - 1/N (a) = - ar/4h and lies on the negative real axis starting from the origin with a = 0. The Nyquist locus G (j~) intersects it where it has a phase of -180. The values of a and ~ at the intersection can be calculated from - ar/4h = arg and 10 ar = 2 3/2 . 4h (1 + ~ ) The solutio n for ~0 from equation (4.8) is tan- ~0 = 60c , giving w0 = sm aller the closer the o scilla tio n is to a sinuso id. The exa ct f req uenc y of o scilla tio n is 1.708 rad s/s so that ~ 0 is in error b y 1.41%. A symmetrical square wave has a harmonic content which is 1/nth that of the funda mental for n odd. It is thus easy to ca lcula te the ha rmo nic co mp o nents at the o utp ut o f G (s) a nd the third a nd fifth a re 5.40% a nd 1.21% of the f unda m ental. Solving equa tion (4.9) gives the amplitude of the assumed sinusoidal limit cycle a as 5h/r . 4h (a - D ) If the relay has an hys teresis of D , then putting d = 0 in equatio n (3.20) gives ar N (a) = 1/ - j 4h2 D from which C (a) = - 1 = 2- 2r 6(a - D ) a r + jD@. N (a) 4h D = 1 and h = r/4 giving C (a) = - (a - 1) Thus on cycle solution is C(a) by line parallel-to= real axis at a distance rD/4h below it, as res within n exa ct 4.4 for j the the limit the Nyquist plot, givenis a - (a - 1) where ~ = 1.266 , which co m pa shown a Figure (1 + j~) 2 1/ 2 - j . If the same transfer function as above is used for the plant, then 2 1/ 2 10
3 2 2 1/ 2 2 2 1
10 (1 + j~) 10 (1 + j~)
which can be written

3
= 180c
(4.8)
(4.9) 3 = 1.732 . Because the DF solution is
ap p ro xim ate, the a ctual m easured freq uency o f o scilla tio n will diff er fro m this value b y a n a m o unt which will b e
solutio n value of 1.254, a nd a = 1.91.
58
Figure 4.4 Nyquist plot for 10/(1+s)3 and DF for relay with hysteresis
For the ideal relay, equation (3.30) gives Nic (a) = 2h/ar for the IDF so tha t the roots of the characteristic equa tion 1 + N i c (a) G (s) = 0 sho w the limit cycle is sta ble. This a grees with the p erturba tio n ap p ro a ch which a lso sho ws that th e limit cycle is stable when the relay has hysteresis.
4.5.4 Use of the IDF

Consider an unstable plant with the transfer function G (s) = K/ (s - b) (s + 2s + 10)
2
(4.10)
in a f eed b a ck lo o p with a n id eal rela y with o utp ut 1. The cha ra cteristic eq uatio n f o r the loo p with the rela y g ain ta k en as its DF gives s + s (2 - b) + s (10 - 2b) + k - 10b = 0
3 2
(4.11)
where K N(a) is repla ced b y k. If a third o rd er chara cteristic eq uatio n ha s a limit cycle a t a f req uency , then its f orm will be (s + ~ ) (s + c) = 0
2 2
(4.12)
which gives (2 - b) (10 - 2b) = k - 10b so that k = 20 - 4b + 2b . The characteristic eq uatio n is stable for b < 2 a nd the freq uency
2
of the limit cycle is (10 - 2b)
1/ 2
. However, when the limit cycle is established the gain throug h the relay to any unrelated
signal is the IDF, which for the ideal relay is easily shown from equation (3.30) to be N(a)/2, tha t is it is halved and the s + s (2 - b) + s (10 - b) + 10 - 2b + b - 10b = 0 new characteristic equatio n is
3 2 2
(4.13)
59
which is stable for b < 0.90. Thus, the odd symmetrical limit cycle canno t exist for b > 0.9. Simulation results showed an odd symmetrical limit cycle to around b = 0.4 and an asymmetrical limit cycle for values of b > 0.4 up to around b = 0.9, although the limit cycle was dependent on the initial conditions. For b > 0.9 the relay lo ck ed at either 1 so tha t the o utp ut increased ind efinitely. A g oo d expla na tio n f or the p heno m eno n ca n b e seen if th e root locus for the transfer functio n G(s) is plotted, as done in Figure 4.5 for K = 1 and b = 1.
Figure 4.5 Root locus plot for transfer function of equation (4.7)
+LZPNU `V\Y V^U M\[\YL H[ 4(5 +PLZLS

^^^ THUKPLZLS JVT
60
It ca n b e seen fro m the gains m ark ed tha t the co m p lex poles cross the im ag inary a xis f or a ga in of 18 a nd f or tha t valu e of g ain the real pole is a t -1. Ho wever f o r a g ain of 9 (half o f 18) the real p ole will b e in the rhs o f the s pla ne so tha t the system is unsta ble. If the plot is d o ne f or b = 0.9 then the g ain when the co m p lex ro o ts cro ss the a xis will b e twice th at for the real ro ot cro ssing the a xis. The difficulty of a chieving a n o d d sym m etrical limit cycle fo r b slig htly less tha n 0.9 is presumably d ue to errors in the DF a nd the small ra nge of attractio n of the limit cycle, which is a closed curve in three spa ce.
4.6 More than one Nonlinear Element

When the closed loop system has more than one nonlinear element it is possible to write down no nlinear algebraic eq ua tio ns b ased o n the DF m etho d to give a ny p ossible lim it cycle solutio ns. The difficulty with this a pp ro a ch, ho wever, is that satisfa ctory a ccura cy ca n o nly be o btained if , when a lim it cycle o ccurs, it is rea so nable to assum e the inp uts at all t he no nlinea r elem ents are nearly sinuso ida l. In the cla ssica l m ultivaria ble versio n of Fig ure 3.1 where G (s) is a tra nsf er m a tri x a nd n(x) a ma trix with no nlinear elements o n the dia go nal this ma y b e reaso na ble since typ ical tra nsf er f unctio n elem ents of G(s) will be low pass, which filter the higher harmonic content. Ho wever, one has to assume that the limit cycle is at a single freq uency, which is norm ally the ca se in a m ultivaria ble co ntrol system f eed ba ck lo o p . A sim ple co unterexa m ple is easily o btained b y lig htly co upling two single lo o ps which ha ve limit cycles at two d iff erent f req uencies. Dep end ent on va rio us f a cto rs the loo p s m a y co ntinue to m aintain bo th f req uencies o r lo ck to a lim it cycle with two related , sa y 1 to 3, dominant frequencies. In a single lo op co nfig uratio n the assum ptio n of sinusoida l sig nals into all the no nlinear elem ents is usually no t justifia ble. If, for example, another nonlinear element is included in Fig ure 3.1 prior to Gc(s) then it may be reasonable to assume this has a sinusoidal input but since most controller transfer functio ns Gc(s) are not low pass filters the signal into the nonlinearity before the plant may be quite highly distorted. Thus, if a DF analysis is done the results may be incorrect. A well kno wn example of this is shown in the Simulink diagram of Figure 4.6 where both nonlinear elements are ideal saturations with limit levels 1 and the linear elements are a phase lead with a transfer functio n (1 + s) / (1 + 0.1s)
2
sa turatio n cha ra cteristic b ef ore the p ha se lea d a re b oth sta ble fo r K < 11. Ho wever the system of Fig ure 4.6 p ossesses a limit cycle for large initial conditio ns on the integrator.
Figure 4.6 Simulink diagram of system with two ideal saturations
61
and a pla nt with a tra nsfer f unction of K/s (1 + s) with K = 9. The linear system a nd the nonlinear system witho ut the
Figure 4.7 Initial condition responses of system of Figure 4.6
Figure 4.7 shows the responses of the system of Figure 4.6 from initial conditions on the output integrator of 4 and 1 respectively. A limit cycle results from the former whereas for the seco nd case the respo nse is stable althoug h very oscillatory. The third response shown on the fig ure labeled no sat is again for an initial condition of 4 but without the saturatio n before the lead network and is again very oscillatory. The limit cycle results because althoug h it is almost sinusoidal at the integrator output it is highly distorted at the input to the second saturation, with the result that the fundamental co m po nent under go es a p ha se shift when p assing thro ug h the seco nd sa turatio n. Fig ure 4.8 sho ws the inp ut a nd o utp ut waveforms for the limit cycle at the second saturatio n.
Figure 4.8 Waveforms of limit cycle at the input and output of the second saturation
If with more tha n o ne no nlinearity in the loop it is reaso nable to assum e a ny limit cycle in the system wo uld be appro ximately sinusoidal at the plant inp ut then, as mentioned in section 1.3, if the plant consists of linear and nonlinear elements a freq uency resp o nse m o d el N(a,) co uld be used to m od el it. This co uld b e d o ne either f ro m m a them atical m od eling o r exp erim ental d ata , where if the disto rtio n is sm all the o utp ut is ta ken as the sig nal a t the sa m e freq uency as the sinusoid al inp ut. This f req uency d ep end ent DF rep resentatio n m a y b e used in a nalysis a nd d esig n, a topic discussed in sectio n 9.5.
4.7 Applications of SBDF to find Limit Cycles

In this section some simple examples are considered to show the use of the SBDF in evaluating limit cycles. The first exa m ple co nsiders a system with a n asym m etrical no nlinea rity in a n a uto no m o us f eed b a ck lo o p a nd the seco nd a n o d d symmetrical nonlinearity in a loop with an input which causes a biased operation on the no nlinearity.
62
The p ro ced ure is to set up two eq ua tio ns, o ne f o r the first harm o nic a nd the other fo r the b ia s, or d c, sig nal. The f orm er is always the same, namely Na (a, c) G (j~) = - 1 (4.14)
but the latter depends upon where a ny bias signal enters the loop a nd the value of G (0). Since for a single valued no nlinearity Na (a, c) is real, the limit cycle frequency is given by arg G (j~) = - 180c which if this frequency is c, gives Na (a, c) = 1/ | G (j~c) (4.15)
(4.16)
The f req uenc y of the limit cycle g iven b y the SBDF so lutio n is thus, unlik e the a m plitud e, una ff ected b y the b ia s sig na l, which will no t normally be true in practice as the harmonic distortion will be affected by
4.7.1 Asymmetrical nonlinearity

The f eedb a ck system of Fig ure 3.1 is co nsid ered with no inp ut G (s) = 2.3/s (s + 1) a nd the no nlinearity sho wn in Fig ure 4.9. It is defined b y x + 0.5x + 0.25x f or | x| < 1, 1 fo r x 1 a nd -1 fo r x -1. Fig ure 4.10 sho ws the resulting limit cycle at the output of G(s) from an output initial condition of 0.8 which since it lies in the range |x| < 1, it may b e evaluated using its SBDF for this range.
2 3 2
Figure 4.9 Nonlinear characteristic for Example of section 4.7.2
63
Figure 4.10 Limit cycle in Example of section 4.7.1

n
For the polynomial nonlinearity n (x) = x the co mponents of the SBDF from equatio ns (3.26) a nd (3.27) are N c (a, c) = (1/c) N a (a, c) = (2/a)
-a
-a
# (x + c) # (x + c)
a
p (x) dx (x/a) p (x) dx
respectively. After expa nding (x + c) it can be shown that Nc (a, c) = / n Ck nk c - k =0 2 n n nk1
(4.16)
nk
Na (a, c) = (2/a ) / n Ck nk+ 1 c k =0
(4.17)
where the binomial coefficient

n k
C =
n! (n - k) !k!
2 3
(4.18)
Thus for the nonlinearity x + 0.5x - 0.25x the values are Nc (a, c) = 1 + 0.5c + (a 2/4c) - (3a /8) - (c /4)
2 2
(4.19) (4.20)
Na (a, c) = 1 + c - (0.75a /4) - 0.75c
The linear transfer functio n has a p hase shift of -180 when = 1 with a co rrespo nding gain of 1.15, so that using eq uation (4.16) gives Na (a, c) = 1/1.15 = 0.87 and since in the stead y state the average value of the input to the integrator must be zero the bias equatio n is cNc (a, c) = 0 (4.22) (4.21)
64
Eq ua tio ns (4.19) to (4.22) ca n b e solved to ob ta in a a nd . A q uick estim ate ca n b e fo und b y neglecting term s involving p o wers o f to g ive = -0.0836 a nd a = 0.497. The exa ct solutio ns a re = -0.0763 a nd a = 0.513, which ca n b e seen to be in good agreement with the simulatio n result of Figure 4.10.
4.7.2 Effect of Bias with Odd Symmetrical Nonlinearity

The f eedb a ck system in the Sim ulink dia gra m of Fig ure 4.11 is co nsid ered . The o dd sym m etrical no nlinear cha ra cteristic has slo linear 1 f or inp a re 1.1/s (s + 1) nd a 2 nstant value of 1.5 1 b etween 2 a nd 4 nd then zero , a ut. The 4.12.aThe p e of d ynamics uts b etween 0 aand 1, co b etween 1 a nd 2, enters the loop at theano nlinearity o utpnd is sho wn in Fig ure
2
system is stab le when the co nsta nt value is zero b ut its effect is to bias the no nlinearity o utp ut so that the a verag e value of a ny limit cycle a t the no nlinea rity o utp ut ha s a value of 1.5. This is d ue to the integ rato r in the lo op , the inp ut of which must ha ve an average value of zero in the steady state.
Figure 4.11 Simulink diagram of Example in section 4.7.2
Student Discounts
Student Events
Money Saving Advice
Happy Days!
2009
65
Figure 4.12 Nonlinearity in the Simulink diagram of Figure 4.11
The resulting limit cycle waveforms at the input and output of the no nlinearity, when an initial value of 2 is applied at the system output, are shown in Figure 4.13. To calculate the simulatio n result using SBDF analysis is not easy because of the linea r seg m ented no nlinear cha ra cteristic. It ca n b e sim p lified using the k no wn sim ulatio n solutio n f ro m which it ca n b e seen that the a m plitud e of the limit cycle at the no nlinearity inp ut is within the ra ng e -1 to 2 so tha t the SBDF just needs to be worked out for this range, where the characteristic is a gain of 2 in parallel with minus ideal saturation. The bias inp ut is such tha t it lies o n the saturated level of the sa tura tio n cha ra cteristic, thus > 1. The bias a nd f und a m ental outputs can be calculated to be
2 1 /2
and fund = (2a/r)6(r/4) - ({/2) - (sin 2{/4)@
where sin { = 6(c - 1) /a@. Thus for the no nlinearity the SBDF ter ms are
2 1/2
(4.23)
and Na (a, c) = 2 - (2/r)6(r/4) - ({/2) - (sin 2{/4)@ G(j) has a phase shift of 180 for = 1, with a corresponding gain of 0.55, so that equatio n (4.16) gives Na (a, c) = 1/0.55 and the bias equation is cNc = (a, c) = 1.5 The so lutio n to eq uatio ns (4.23) to (4.26) g ives = 1.20 a nd a = 0.379 which co m pa res reaso na ble well with the sim ulatio n result shown in Fig ure 4.13 for the nonlinearity input waveform. (4.24)
66
bias = 1 - (a/r) (1 - 6(c - 1) /a@ )
N c (a, c) = 2 - (1/c) + (a/rc) (1 - 6 (c - 1) /a @ )
Figure 4.13 Simulation result for the example of section 4.7.2
4.8 Conclusions
This section has sho wn how limit cycles ca n be estimated in no nlinear systems using describing f unctio n methods. Systems where bias sig nals m a y exist in the lo o p ha ve b een evalua ted using the SBDF. It ha s f urther b een sho wn ho w the stability of a ny lim it cycles ca n b e assessed. Further, it ha s b een d em o nstra ted that when single valued no nlinea r elem ents exist in a f eed ba ck system f ailure of the DF m ethod to p redict a ny limit cycles is d ue to their wa vefo rm not b eing sufficiently sinusoidal at a no nlinearity inp ut so that the fundamental sig nal undergoes a phase shift in passing thro ugh the nonlinearit y which is not p redicted b y the DF. Ma ny textbo oks co ver so m e ba sic uses of the DF usually f or evaluating o dd sym m etrica l limit cycles a nd the sta bility o f limit cycles is norm a lly co m m ented o n in term s of the Lo eb criterio n. So m e of the m ore detailed topics in this chapter can be found in the books given in the bibliography.
4.9 Bibliography
Atherton, DP 1975, Nonlinear Control Engineering: Describing Function Analysis and Design, Van Nostrand Reinhold , London. Atherton DP 1982, Nonlinear Control Engineering. Student edition, Van Nostrand Reinhold, London . Gelb, A & Vander Velde, WE 1968, Multiple-Inp ut Describing Functio ns and Nonlinear System Desig n, McGraw Hill, New York.
67
The SSDF and Harmoni cally Forced Systems
5 The SSDF and Harmonically Forced Systems

5.1 Introduction
Describing functio ns have so far been considered for a single sinusoid and a sinusoid plus bias. To examine various p heno m ena which ma y o ccur in no nlinear systems it is ho wever req uired to evaluate d escribing f unctio ns f or o ther inp uts, in particular two sinusoidal inp uts. The evalua tion of the SSDF for any no nlinearity is considered after introduction of the top ic b y use o f a cubic no nlinea rity. A g eneral theo ry is presented b ut the difficult alg ebra fo r sp ecific no nlinea rities m ak es evalua tio n difficult fo r m a ny no nlinea rities, specifically linea r seg m ented o nes, so tha t num erical a pp roa ches ha ve usua lly to b e ad o pted to o btai n solutio ns. The problem is sho wn to be even m ore co mplicated if the sinusoida l sig nals a re rela ted in freq uency, a situa tio n which of ten needs to b e co nsid ered in pra ctice. The co ncep t of m o dified no nlinea rities, as a m ea ns of und ersta nd ing the evalua tio n of the SSDF, is then co vered a nd exa m ples given fo r sp ecif ic no nlinea rities. So m etim es it is necessa ry to co nsid er a perturb atio n which is related in f req uency to a n existing sinuso id so the evaluatio n of this IDF is considered next. This is followed by a practical example of the use of the SSDF to evaluate a limit cycle using a two harmonic balance.
68
The closed lo o p freq uency respo nse of a no nlinear system is then co nsid ered. The m ost usual situa tio n is f or the system to ha ve a n o utp ut with a d o mina nt f unda m ental co m po nent of the sa m e f req uency as the inp ut. This situa tio n is exa mined using the DF and it is shown tha t over some ranges of amplitude and frequency of the inp ut three analytical solutions ra ther tha n o ne to the no nlinea r eq ua tio ns m a y b e ob tained . In pra ctice this results in the so called jum p p heno m en a or jum p reso na nce in the freq uency resp o nse. Finally it is sho wn ho w its existence ca n b e investig ated using the rela ted freq uency IDF. Vario us other p ossib le b eha vio urs tha t m a y o ccur, such a s the excitatio n of oscillatio ns a nd chao tic m otio n, are discussed b a sed o n a n und ersta nd ing of the SSDF a nd so m e sim ula tio n exa m ples given. Brief m entio n is a lso m a d e of the possible behaviours when a harmonic input is applied to an autonomous system which has a limit cycle. Analysis of a sp ecific system using the methods presented in this chapter is not easy unless the techniques are implemented in appropriate numerical programs, however, presentation of the theoretical background should allo w the reader to understand the mecha nisms that cause the aforementio ned uniq ue behavio urs of nonlinear feedback systems.
5.2 The Cubic Nonlinearity with Two Sinusoidal Inputs

Consider the two sinusoidal inputs x = a cos ia and y = b cos (ib + {) , where ia = ~a t and i b = ~b t into a cubic nonlinearity, then the output, f(t), is given by f (t ) = 6a cos i a + b cos (i b + {)@
3
(5.1)
Expanding the + 3a b cos iyields(ib + {) + 3ab cos ia cos (i + {) + b cos (ib + {) f (t) = a cos ia expression a cos
3 3 2 2 2 2 3 3
f (t) = (3a/4) (a + 2b ) cos ia + (3b/4) (b may cos (ib + which on expanding the powers of cos i+ 2a )be written{) + (3a b/2) cos 2ia cos (ib + {) + (3ab /2) cos ia cos 2 (ib + {) + (a /4) cos 3ia + (b /4) cos 3 (ib + {)
2 2 2 2 2 2 3 3
f (t) = (3a/4) (a + 2b ) cos ia + (3b/4) (b + 2a ) cos (ib + {) + (3a b/4) and on using 2 cos A cos B = cos (A + B) + cos (A - B) gives
6cos (2ia
3
2 +2 +2 {) cos (2ia b - {)@ + 2(3ab2 /4)6cos 2 a (i + (a /4) cos 3ia + (b /4) cos 3 (ib + {) +i -i + 2i b 3
+ 2{) + cos (ia + 2i
+ 2{) @ (5.2)
If the frequencies a and b are unrelated then the first two terms produce the output at the same frequencies as the inputs and the SSDF has the components Na (a, b) = (3/4) (a + 2b ) Nb (a, b) = (3/4) (b + 2a )
2 2 2 2
(5.3) (5.4)
69
Ho wever, if there is a n integ er rela tio nship b etween the two f req uencies then it is p o ssible tha t so m e of the other term s in f(t) can produce outputs at the same frequency as the input. For example, if sinusoidalthen the Further component of the third term is (3a b/4) cos (- ia - {) , which has the same frequency as the ~b = 3~a input x. second this
2
output component + (3a b/4) cos ina phase that the DF for the inp ut x, which awill bethe total output at this frequency is (3a/4) (a + 2b ) cos ia is no longer (i + {) so with the input at frequency and denoted N a (a | b) to indicate
2 2 2
that the signals a and b are related, is Na (a | b) = (3/4) (a + 2b + abe )

2 2 j{
(5.5)
Thus unlik e the unrela ted case the ab o ve DF value dep ends no t o nly o n the a m plitud es of the two sinusoids b ut also their relative phase. It is the phase shift produced throug h a nonlinearity to an harmonic input related to another harmonic input tha t results in ma ny of the interesting pheno mena in nonlinear feedback loops. Further when the nonlinearity is not a sim ple ma them atical f unctio n, such as the cubic of the a b o ve exa m ple, it b eco m es very diff icult to g et the o utp ut b y the sim ple trig o no m etric a pp roa ch used ab o ve. Fo rtuna tely, ho wever, it ca n b e sho w n that f or a ny no nlinea rity n(x) the output f(t) ca n be written f (t) = /3
j=0
/ f j fk a jk cos ji cos k (ib + {) 3

k =0
(5.6)
where a jk =
-a
# # n (x + y) T (x/a) T (y/b) p (x) p (y) dxdy

a b j -b k a b
(5.7)
and the Neumann factor fn is 1 for n = 0 and 2 for n 1. Comparing equation (5.6) with the result of equation (5.2) for the cubic it ca n be seen for the cubic tha t a10 = (3a/8) (a + 2b ), a21 = 3a b/8, a30 = a /8
2 2 2 3
a nd with n (x f y) = , 12 a ca e o also b e f y reversing jk = 0 o These > 3. (5.7)the values +or a 01 (xa +, y)03a ndnitb will btained bo und that aa a nd fb. r j + k values ca n b e check ed to result fro m eq uatio n
3
Thus even thoug h the expression of equation (5.6) can in principle be evaluated for any no nlinear characteristic it ca n co nsist of a larg e, a nd p ossibly infinite num b er of term s. Also fo r related inp ut f req uencies there m a y b e a larg e num b er of these term s that p rod uce a n o utp ut at o ne of the inp ut freq uencies a ll of which in g eneral will ha ve a diff erent p ha se.
5.3 Modified Nonlinearities and the SSDF

The evaluation of a jk from equation (5.7) can be regarded as the two stage process n (a, c) j =
b -a
# n (x + c) T (x/a) p (x) dx
a j a jk b
(5.8) (5.9)
a jk =
# n (a, c) T (y/b) p (y) dy

-b
70
An Introduction to Nonlinearity in Control Systems Or, alternatively n (b, c) k =

b
-a
# n (y + c) T (y/b) p (y) dy
k b kj a
(5.10) (5.11)
a jk =
# n (b, c) T (x/a) p (x) dx

-b
The term s n (a, c) j a nd n (b, c) k ma y b e reg arded a s g eneralized m odified no nlinearities with the o ne with subscrip t zero , already having been introduced in sectio n 3.7, using a notatio n omitting the zero subscript, which will continue to be used . Assum ing , where unless m entio ned to the co ntra ry the two sinuso id s a re no t rela ted , then the SSDF co m p o nen ts are N a (a, b) = 2a 10 /a a nd N b (a, b) = 2a 01 /b . An a d va nta g e of lo oking at the d erivatio n in this wa y is tha t if o ne wishes to g et the SSDF co m p o nent f o r x, then o ne calculates the DF f o r the origina l no nlinea rity m o dified b y y, na m ely the DF
a
of n (b, c) , or alternatively one calac ulates n ( a, c) 1 =

-a
# n (x + y) (x /a) p (x) dx , whic h is 0.5a times the DF for x in the
presence of a bias and then averages this with respect to b, as given b y eq uatio n (5.9) with k = 0. The evaluation of modified no nlinearities for linear segmented characteristics is quite difficult and therefore it becomes even more difficult to find the SSDF, so that numerical approaches ha ve to be used . Ho wever m any no nlinear chara cteristics ca n b e a pp ro xim ated o ver a given ra nge b y either a po wer series or Fo urier series exp a nsio n, so results are d erived belo w for a power law and harmonic nonlinearity, as well as for an ideal relay.

71
5.3.1 Power Law Characteristic

Consider the no nlinearity n (x) = x . Then one has n (a, c) k =
m
-a
# (x + y)
a
Tk (x/a) pa (x) dx
(5.12)
which on expanding gives n (a, c) k = n/= 0

m
-a
#C
m
xc
nm-n
Tk (x/a) pa (x) dx
(5.13)
where the bino mial co eff icient m C n = m !/ (m - n) !n! Thus the integ ral ca n b e co m p uted f o r a ny cho sen value of k, since after sub stituting f o r Tk (x/a), the integ ral will invo lve m o m ents of the distrib utio n of pa (x). In p articular f or k = 0, o ne ha s n (a, c) = /
m n =0 -a a
Cn x c
m n m-n
pa (x) dx = / mCn n (a) c - m n

m n =0
(5.14)
and for k = 1 n (a, c) 1 =
/#
-a n =0
a mn
c Cx
n+ 1 m-n
a pa (x) dx =
-1
/
m m n =0
Cn a- nn + 1 (a) c
1 m- n
(5.15)
Using equation (5.14) in equation (5.09) and multiplying by 2/b to get Nb (b, c) from a01 then gives for the SSDF to the signal y N b (a, b) = (2/b)
-n
#
0
/ mCn n (a) y
m m n= n =0
m-n -1
b ypb (y) dy = (2/b 2) / mCn n (a) nm - n + 1 (b) n (5.16)
Or, alternatively, equation (5.15) with b replacing a can be used in equatio n (5.11) and again multiplied by 2/b to get Nb (b, c) from a01 to give N b (a, b) = (2/b)
n= 0
-a a m
#/
m
Cn b- nn + 1 (b) x
1 m-n
pa (x) dx = (2/b )
2
/
m m n =0
Cn nn + 1 (b) nm - n (a)
(5.17)
Since m Cn = mCm- n these expressions are easily shown to be identical. Na (a, b) can be similarly found or simply obtained from Nb (a, b) by interchang ing a and b.
5.3.2 Harmonic Nonlinearity

sin m (x + c) T (x/a) p (x) dx For the harmonic nonlinearity n (x) = sin mx
a
n (a, c) k =
-a
which on expanding sin m (x + c) and substituting x = a cos i gives n (a, c) k = (1/r))
cos mc sin (m a cos i) cos kidi +

0
# sin my cos (ma cos i) cos kidi 3 .
72
An Introduction to Nonlinearity in Control Systems Substituting the expressions cos (z co s i) = J 0 (z) + 2 / (- 1) J 2i (z) co s 2ii a nd
i=1 i 3 i
sin (z co s i) = - 2 / (- 1) J 2p - 1 (z) co s (2i - 1) i a nd integrating gives

i=1
n (a, c) k = (- 1) J k (ma) sin "mc - (kr/2) , In particular for k = 0 the modified nonlinearity is n (a, c) = J 0 (ma) sin mc which differs only in amplitude by a factor J 0 (ma) from the original nonlinearity. Also n (a, c) 1 = J 1 (ma) cos mc It is easily shown, again using the above expansions, tha t Nb (a, b) = (2/b) J 0 (ma) J 1 (mb)
(5.18)
(5.19)
(5.20)
(5.21)
5.3.3 Ideal Relay

The two previous characteristics considered were continuous functio ns of the inp ut variable, whereas this is not the case for many approximations to nonlinearities found in co ntrol systems, such as the relay or ideal saturation. These create difficulties for obtaining general analytical results for the SSDF because the modified no nlinearity will possess disco ntinuities d ep end ent o n the rela tive inp ut m a g nitud es o f the sig nal a nd the bias. This is illustrated f or the id eal rela y with outp ut level h for which n (a, c) k =
-c
-a
- hTk (x/ a) p a (x) dx +

-c
# hT (x/a) p (x) dx
a k a
which gives for k 1 n (a, c) k = (2h/kr) Dk (- c/a) for c # a n (a, c) k = 0 for c > a and for k = 0 n (a, c) = (2h/r) sin- (c/a) for c # a
1
(5.22)
(5.23)
n (a, c) = hsgn c for c > a
73
If it is assumed that b a then to find Nb(a,b) only n(a,) is required for | | a and N b (a, b) = (2/b)
-b
# n (a, y) (y/b) p (y) dy = (4h/r b ) # y sin

b b 22 -b 2 2 1/ 2
-1
(y/a) (b - y ) -
1/2
dy
Integrating b y parts and using the fa ct that ( b - y ) Nb (a, b) = (4h/r b )

22
is zero at both limits gives
-b
(b - y ) 22 2 2 1/ 2 dy = (8h/r k a) (a - y )
2 1/ 2
"E (k) - (1 - k2) K (k) ,
(5.24)
where k = b/a < 1 and E(k) and K(k) are elliptic integrals. Similarly, to find Na(a,b) only o ne integral is required for b a if n(a,)1 is used and the solution is Na (a, b) = (4h/ar )
2
-b
# 6(1b--(yy/a)) @ dy = (8h/ar ) E (k)

2 2 2 1/2
(5.25)
Thus even for this very simple break point type no nlinearity the SSDF expressions are rather complicated.
5.4 The IDF for Related Signals

The IDF, the gain to an unrelated signal in the presence of a known other one, was considered in section 3.7 for the known signal being a sinusoid. Also the perturbatio n equatio n of (4.5) to assess stability was derived assuming the perturbatio n was unrelated to the existing signal. Here the IDF is investigated when the small signal is assumed to be related to the sinusoid. Consider a nonlinearity with inputs x (t) = a cos ia and a small signal (t). The nonlinearity output, making use of a Taylor series expansion, as used to derive equation (4.5), may be written as n" x (t) + d (t), . n (x (t)) + d (t) n' 6 x (t) @ where n' (x) = dn (x) /dx. The incremental o utp ut, y(t) d ue to (t ) is thus d (t ) n' 6 x (t) @ the o utp ut of the nonlinearity n' (x) with input x(t). This output may be written for any no nlinearity in the Fourier series. n'"(x (t), = / kc= 0cos (kia + {k) k where x(t) has a frequency of ~a = ia t . If (t) is the small sinusoid d cos (ib + z) , with ~b = ib t , then dy (t) = d
3
/ c cos (ki
k k= 0 3
+ {k) cos (ib + z) which can be written
dy (t) = d k/= (ck /2)"cos (ki a + {k + i b + z) + cos (kia + {k - ib - z), 0 The only output at ~b if ~b and ~a are not related is dc0 cos (ib + z) but if ia /ib = ~a /~b = m/n then the output at frequency ~b is d"c0 cos (ib + z) + (ck /2) cos (ib + {k - z) | k = 2n/m , which can be written (5.26)
74
dc0 cos (ib + z) + (dck /2)"cos ({k - 2z) cos (ib + z) - sin ({k - 2{) sin (ib + z), | k = 2n/m , Thus, dividing b y , gives the IDF f or situa tio ns where the freq uencies a re related such that if a n integer k = 2n/m exists, then it is given by N id (a) = c 0 + (ck / 2) e
j({k - 2z)
| k = 2n/m
(5.27)
where the subscript has been used to distinguis h it from the unrelated IDF Ni(a). Since in general the coefficients ck no rm ally d ecrea se with increa sing k, the m o st im p orta nt o nes a re fo r lo w values of k. Those fo r k = 1 a nd 2, in pa rticular, will b e of so m e interest la ter. Fro m eq uatio n (5.27) it ca n b e seen that N i (a) is co m plex a nd its lo cus is a circle centred at circle with centre at - c0 /"c0 + (ck /2) , a nd radius (ck /2)"c 0 + (ck /2) ,. c0 with radius ck/2. If one wis hes to plot Cid (a) = - 1/Nid (a) on a Nyquist plot then it can be shown tha t Cid (a) is also a
2 2 2 2
Also f or a sing le6(2x /a ) - 1@ p (x) rity=with N = 2, which is the situa tio n f or a perturba tio n of the sa m e freq uency, { k = 0 a nd n' (x) valued no nlinea dx 26 k
a
c2 = 2
-a
2 2
ic
(a) - N (a)@
so that
Nid (a) = N (a) + (a/2) N' (a) (1 + e- )

j2z
(5.28)
This is the equatio n of a circle with centre at (Nic (a), 0) and radius (a/2) N' (a) .
www.job.oticon.dk
75
5.5 More Accurate Determination of Limit Cycles

The classica l a pp ro a ch is to ba la nce m o re ha rm o nics no t just the f und a m ental. Altho ug h stra ig htf orwa rd in p rinciple it is not easy ainp uts. To illustrate the pro ced urecomp uting the o utput the simplenlinear element subject to two= x more related harmo nic nalytically due to the difficulty of a feedback loop with from a no polynomial nonlinearity n (x) or - (x /6)
3
and a loop transfer functio n G (s) = 2 (1 - s) /s (1 + s) is considered. The nonlinearity inp ut is taken as x = a cos ~t + b cos (3~t + {) Computing the outp ut from the nonlinearity at frequencies and 3, gives the terms (5.29)
6a - (a/8) (a2 + 2b2 + ab cos {@ cos ~t + (a2 b/8) sin { sin ~t

and
(5.30)
6b - (b/8) (b2 + 2a2) - (a3 /24) cos {@ cos (3t + {) - (a3 /24) sin { sin (3~t + {)
Thus the magnitude and phase components for the SSDF are M a (a b) = "6 1 - (1/8) (a + 2b 2 ab co s {)2 + 6 (1/8) ab2sin { @ , + @ 2 1/ 2 M b (a b) = (1/b)"6 b - (b/8) (b + 2a 2 + (a3 /24) co s2{ @ + 6 (a /24) 1/2 { @ , ) 2 3 2 sin za (a b) = tan- 6- ab sin {/ (8 - 2a - 2b - ab cos {)@ 1 2 zb (a b) = tan- 6a 3sin {/ (24b - 3b - 6a b - a cos {)@ 1 3 2 3
(5.31)
(5.32) (5.33) (5.34) (5.35)
If the gain a nd p hase of the tra nsf er f unctio n a re deno ted b y g a nd then the harm o nic b ala nce lo op eq ua tio ns fo r the two frequencies are M a (a b) g (~) = 1 M b (a b) g (3~) = 1 za (a b) + i (~) = - 180c zb (a b) + i (3~) = (2j + 1) 180c (5.36) (5.37) (5.38) (5.39)
76
An estimate for the solution can be obtained by using the DF solutio n of a = 2 and = 1, together with equatio ns (5.37) and (5.39) with (5.33) and (5.35) substituted. This gives b = 0.222 and = -53 or 127. Solution of the four nonlinear equations (5.36) to (5.39) with the values from equatio ns (5.32) to (5.35) substituted gives two solutions of sta ble limit cycle sho wn in the sim ula tio n result o f Fig ure 5.1. Also sho wn a re the sim ulatio n results f or the o utp ut of the nonlinearity and the calculated result for the stable limit cycle marked theoretical. Note the phase origin is arbitrary so the wa veform has been plotted just to sho w the good waveform agreement.
a = 2.19, b = 0.298, = 138, = 0.882 and a = 1.88, b = 0.166, = -52, = 1.06. The former solution is that of the
Figure 5.1 Comparison of theoretical and simulation limit cycle results.
It is rela tively ea sy to write num erical ro utines to co m p ute m o re precisely a ny lim it cycles which m a y exist in a feedb a ck loo p. Ref erences 5.1 to 5.3 give d etails of so me p ap ers which discuss po ssible ap proa ches using prog ra m s which m ak e use of the fa st Fo urier tra nsfo rm a nd intera ctive grap hics. Beca use a la rg e num b er of no nlinear alg eb raic eq uatio ns need to b e so lve a n initia l g uess f o r the lim it cycle is req uired . This ca n b e o btained using the DF, if a predicted so lutio n exists, or if not b y a ssum ing a solutio n using sa y a n a m plitud e a nd f req uency nea r to a n intersectio n of the Nyq uist plo t a nd the C(a) lo cus. The a ssum ed lim it cycle solutio n, x 1 (t ), is then used a s inp ut to the no nlinearity with the loo p assum ed op en a n d the sig nal fed b a ck to this point x2 (t ) calcula ted . This is d o ne b y discretising the sig nal, f inding the co rrespo nd ing o utp ut from the nonlinearity, evalua ting its Fourier series, passing these components through G(s) and recombining them to pro d uce the no nlinea rity inp ut. Gra p hs of x1 (t ) a nd x 2 (t ) a re then plo tted f or co m pa ra tive p urp oses. Further iteratio n ca n be d o ne to see if the wa vefo rms co nverg e, typ ically they ma y beco me q uite similar b ut beca use the a ssum ed freq uency is incorrect appear to have a phase difference.
77
The f req uenc y ca n then b e cha ng ed to see if co nverg ence o ccurs. Alternatively at a ny p oint in the pro cess a ha rm o nic bala nce ma y be co mp uted using a limited num b er of harm o nics in x 1 (t ) b y taking it eq ual to the last value o f x2 (t). Using these techniques exa mples are given in the papers referenced of deriving limit cycles in a large number of feedback systems so m e with m ultiple no nlinear a nd linea r elem ents. Ca ses are co nsid ered when the DF metho d d oes a nd do es no t predict a lim it cycle. An interesting feature of the a pp roa ch is tha t b eca use it is num erica l it ca n o btain unstab le limit cycle solutio ns which ca nno t b e f o und b y sim ulatio n. One exa m ple which wa s co nsid ered is the f eed ba ck loo p o rig ina lly investigated b y Fitts with a cubic no nlinearity and linear dynamics
2 4 3 2
(5.40)
This tra nsfer f unctio n, which is not typica l of o ne fo und f or a co ntrol system with the d o uble d erivative in the num erator, has two pairs of lightly damped poles near to frequencies of 0.9 and 1.1 rads/s. and its Nyquist plot is shown in Figure 5.2. The feature of this Nyquist plot, which causes the DF method to fail, is tha t it has related frequencies on different sides of the C(a) plot, which for a cubic lies on the negative real axis. The Nyquist plot starts from the origin and loops in the positive part of the plane to the real axis which it reaches at a frequency of 1.005 before looping in the negative half plane to finish at the origin.

78
G (s) = 10s / (s + 0.04s + 2.0206s + 0.0404s + 0.9803) .
Figure 5.2 Nyquist plot for transfer function of equation (5.40)
In cha p ter 6 it is sho wn ho w lim it cycles ca n be calculated exa ctly in f eed b a ck loo ps with a rela y no nlinea r elem ent. An y no nlinea r cha ra cteristic ca n be a p pro xim ated b y a q ua ntiza tio n cha ra cteristic , the m ore q ua ntizatio n levels used the m ore accurate the a pp ro xim atio n. Since a q ua ntizatio n cha ra cteristic ca n b e obtained b y rela ys in pa rallel the m etho d f or finding limit cycles in relay systems can be used for any nonlinearity using a quantized approximation. The problem with the m etho d is tha t the num b er of no nlinear eq uatio ns which ha ve to b e solved increases with the nu m b er of q ua ntizatio n levels used. Ho wever, b y slo wly increasing the num b er of q ua ntizatio n levels go o d initial estim ates ca n b e o b tained f or the no nlinear equations and good results are reported in reference 5.4 using this approach.
5.6 Closed Loop Frequency Response

W hen a sinusoid al inp ut is a pp lied to a f eed ba ck loo p which is linea r then p ro vid ed it is sta ble the o utp ut will also b e a sinusoid with a m plitud e a nd p ha se dep endent o n the inp ut freq uency. Further if the inp ut a mp litud e is d o ubled then th is just doubles the output amplitude. Obviously if the system is nonlinear then the frequency response will be amplitude dep endent b ut a n ob vio us q uestio n is will the o utp ut b e ro ug hly sinusoidal a nd a t the sa me f req uency as the inp ut? This asp ect will b e exa mined in the next sectio n, 5.7,b y a ssum ing that it will b e the ca se b ut b efo re d oing so it is ap p rop riate to comment on some other possibilities suggested by the previously derived DF results. R esults f or the SSDF ha ve sho wn that the b oth the gain a nd p hase shift which a sinusoid und erg oes in p assing thro ug h a nonlinear element can be affected by another signal. If the o ther signal is unrelated to a n existing o ne then o nly the gain to the existing o ne will b e aff ected Fo r a limit cycle to b uild up in a f eed b a ck lo op then the lo op m ust b e unstable to a small sig nal, clea rly a situa tio n which mig ht arise when a n external si g nal ca uses the g ain thro ug h a no nlinea r element to a sm all
3 2
79
signal to cha nge app ropriately. Fig ure 5.3 sho ws a simple feedba ck loop with a no nlinear element n (x) = x for |x| 1, 1 for |x| > 1 a nd -1 for |x| < -1, dynamics of 2/s (s + 1) a nd a sinusoidal inp ut of amplitude 1 at a freq uency of 4.1 rad /s.
Figure 5.3 Simple feedback loop with cubsat nonlinearity
Altho ug h the system is sta ble with no inp ut, a s indicated b y the DF m etho d, o r a s ca n b e a scertained b y sim ula tio n, when subject to the a b o ve sinuso idal inp ut the g ain thro ug h the no nlinea rity is increa sed suf ficiently to ca use a lim it cycle to o ccur at a n a p pro xim ate freq uency of 1 rad /s, the 180 p ha se shif t f req uency o f G (s). The o utp ut of the system sho win g the dominant limit cycle frequency, which is unrela ted to that of the inp ut, is shown in Fig ure 5.4
Figure 5.4 Output of System of Figure 5.3
Also the value of N id (a) fo r k = 1, which is a chieved with 2 n = m , ca n ha ve a p ha se shif t. This ind icates the p o ssib ility of a limit cycle o ccurring which is at o ne half of the f req uency o f the a pp lied inp ut, na m ely a seco nd sub ha rm o nic. Clea rly also other sub harm o nics mig ht exist if they ca n so m eho w b e excited since a p ha se cha ng e is p ossible f or finite a m plitud es of other related frequencies, such as the ratio 3:1. A f urther beha vio ur tha t is po ssible is f or the o utp ut to ha ve a cha o tic m o tio n, that is a wa vefo rm which is d eterm inistic, in the sense that if the experim ent is rep eated the sa m e wa vef o rm will result, b ut it ca nno t be describ ed a nalytically. This is
3 3
5.8 sho w the o utp ut respo nse a s the a m plitud e of the inp ut sinusoid is increased . In Fig ure 5.6 f or a n inp ut a m plitude of 1.1 the output, unlike for smaller input amplitudes, is noticeably distorted from a sinusoid.
80
demonstrated for the system of Fig ure 5.5 with an unusual no nlinearity n (x) = x - (x /6) , which cha nges from a positive to a negative slope. The loop d yna mics are 2/ (1 + s) a nd the inp ut freq uency of the sinusoid is 1 rad/s. Figures 5.6 to
Figure 5.5 System which demonstrates chaotic motion.
Figure 5.6 output for input amplitude of 1.1.
81
With an increase in amplitude to 1.3, shown in Fig ure 5.7, the output has reached a bias level and shows considerable third ha rm o nic disto rtio n. Finally in Fig ure 5.8 fo r a n inp ut a m plitud e of 1.4 the o utp ut sho ws a cha o tic m otio n a nd f o r an input amplitude of around 1.5 the response goes unbounded.
Figure 5.7 Output for input amplitude of 1.3
Figure 5.8 Chaotic motion for an input amplitude of 1.4.
Ob vio usly a no ther p ossib ility is that o ne ma y wish to a pply a sinusoid al inp ut to a system tha t alrea d y has a lim it cycle, a situatio n which o nce m ore mea ns two sinusoid al inp uts to a no nlinearit y need to b e co nsid ered . Ag ain b ased o n the SSDF three possibilities can be expected namely (i) the limit cycle and forcing signal will coexist, (ii) the forcing signal may quench the limit cycle or (iii) the forcing signal a nd limit cycle may lock to related frequencies, kno wn as synchro nization.
82
Clea rly to investiga te theoretically a ny of these va ried b eha vio urs which m a y exist in a no nlinear system f or a sinusoid al tha t the b eha vio ur is sta ble. Af firm a tive solutio ns to these two q uestio ns then m ea ns, a ssum ing the ap p ro xim a te a nalysis is sufficiently a ccura te, tha t the b eha vio ur will p ro ba bly exist. Ho wever the a na lysis is b ased o n stea d y state assum p tio ns a nd whether a sp ecific stea d y state is rea ched is a d yna m ic p ro blem which is difficult to solve. Lik e sing ular p oints m o re tha n o ne beha vio ur m a y be rea ched d ep end ent o n the initial co nd itio ns. Luck ily m ost of the a b o ve b eha vio urs are rare in a feedback control system since to exist they require features such as a lightly damped feedback loop, or unusual nonlinearities, for example possessing both positive a nd negative slopes. In the next section the normal frequency respo nse is examined such as that found for the system of Fig ure 5.5 for input amplitudes less than unity.
input one has to apply the appropriate analysis to see if such a behaviour might exis t. Secondly one has then to sho w
5.7 Jump Resonance

When the closed loo p system of Fig ure 3.1 has a sinusoid al inp ut r (t ) = R sin (~t + b) , it is po ssible to evaluate the closed loop frequency response using the DF. If the feedback loop has no limit cycle when r (t ) = 0 a nd, in additio n, the sinusoidal inp ut r (t ) d o es not ind uce a lim it cycle, then, pro vid ed tha t G c (s) G 1 (s) gives g o o d f iltering , x (t ) , the no nlinea rity inp ut may be taken eq ual to the sinusoid a sin ~t . Thus with the nonlinear element replaced b y its DF, equatio ns can be obtained for the loop operation. For example balancing the co mponents of frequency ~ around the loop gives Rgc sin (~t + b + ic) - aM (a) g1 gc sin (~t + z (a) + i1 + ic) = a sin ~t where G c (j~) = g c e
jic ji
(5.41)
algebraic equations, by resolving in the directio ns of ~t and cos ~t , can be solved for the two unknowns a and for the given sinusoidal inp ut with m ag nitud e, R, a nd f req uency, . The f und a m enta l o utp ut sig na l ca n then b e fo und f ro m c (t) = aM (a) g1 sin (~t + z (a) + i1) (5.42)
Various graphical procedures [5.5-5.7] have been proposed for solving the two nonlinear algebraic equations resulting from equation (5.41) or for the loop balance results written in a slightly different form. If the system is lig htly damped, the no nlinear equations may have more than one solution, indicating tha t the frequency response of the system ma y ha ve a jum p reso na nce. This p heno m eno n of a no nlinea r system ha s b een stud ied b y m a ny a utho rs, b o th theoretica lly a nd fo r p o ssib le pra ctical a p plicatio ns. The m ajo r a d va nta g e o f a g rap hical a p pro a ch to ob tain the solutio n, ra ther tha n a co m p utatio na l a p pro a ch to solving the no nlinea r alg ebraic eq uatio ns of (5.41), is the und ersta nd ing it p ro vid es of the jum p reso na nce p heno m eno n. O ne a pp ro a ch is b riefly o utlined which is a p plicable f or a single valued no nlinearity, tha t is for z (a) = 0 . From equation (5.41), on separating the sin ~t and cos ~t terms, one obtains Rg c cos (b + i c) - aM (a) g 1 g c cos (i 1 + i c) = a and Rgc sin (b + ic) = aM (a) g1 gc sin (i 1 + ic) 83
( 5 . 4 3 )
( 5 . 4 4 )
and G 1 (j~) = g 1 e . Then, in principle, equa tion (5.41) which can be written as two nonlinear
1
Denoting the normalized inp ut magnitude to the nonlinearity a/R by x, the normalized outp ut aM(a)/R by y and eliminating gives x = g c 6 1 - y g 1 sin (i 1 + i c ) @
22 2 1/2
- yg1 gc cos (i1 + i c)
(5.45)
This is the linea r freq uency d ep end ent relatio nship between the no nlinearity inp ut a nd o utp ut which need s to b e so lved together with the DF relationship y = xM (xR) (5.46)
which need s to b e plo tted o n the sa m e x -y p lo t a s eq ua tio n (5.44) fo r the given va lue of R to g et p ossible solutio ns. In cases where jump resonance occurs there are typically three solutions rather tha n one solutio n [5.8 -5.10]. In this case it is possible using the stability criterion of the next sectio n to show that two solutio ns will be stable and one unstable. An interesting p heno m eno n called isla nd -typ e jum p reso na nce is d escrib ed in reference 5.11. Here o ne sta ble solutio n is iso la ted , so that it m a y o nly b e rea ched fro m the ap plica tio n of suita ble initial co nditio ns, a nd there f ore ma y no t b e seen in a standard frequency response test.
678'< )25 <285 0$67(56 '(*5((

&KDOPHUV 8QLYHUVLW\ RI 7HFKQRORJ\ FRQGXFWV UHVHDUFK DQG HGXFDWLRQ LQ HQJLQHHU LQJ DQG QDWXUDO VFLHQFHV DUFKLWHFWXUH WHFKQRO RJ\ UHODWHG PDWKHPD WLFDO VFLHQFHV DQG QDXWLFDO VFLHQFHV %HKLQG DOO WKDW &KDOPHUV DFFRPSOLVKHV WKH DLP SHUVLVWV IRU FRQWULEXWLQJ WR D VXVWDLQDEOH IXWXUH ERWK QDWLRQDOO\ DQG JOREDOO\ 9LVLW XV RQ &KDOPHUV VH RU 1H[W 6WRS &KDOPHUV RQ IDFHERRN
84
5.7.1 Jump resonance region- use of the IDF

Determ ining whether a sp ecific system ca n ha ve a jum p reso na nce in its freq uency resp o nse ca n b e q uite lab orio us usin g the m etho d of the p revio us sectio n. Based o n a p ha so r diag ra m f or the DF linearised clo sed loo p system it ca n b e arg ued that for a constant frequency input if its magnitude R is increased then a should also increase. Thus Hatanaka [5.12] obtained the criterio n tha t for a system to possess jump resonance u uR u ~const ua ~const facts but all are consistent with the fact that jump reso na nce will occur if a solution exists for the IDF characteristic eq u ation 1 + Nid (a) G (j~) = 0 where the IDF is for a perturbation of the same frequency as the input as given in equatio n (5.28). For a specific solution amplitude a this equation is easily checked but of course this requires the solutio n for a to be available. By plotting Ci d (a) = - 1/Nid (a) , which ha ve been shown to be circles, for all possible values of a then no intersectio n of G (j~) with these loci ensures that jump resonance will not exist. From equa tion (5.28) the maximum phase shift produced for a given a is a where sin a = (a/2) N' (a) N (a) + (a/2) N' (a) (5.47) (5.46)
If am is the maximum value of a for all a then no jump resonance will exist if G (j~) does not enter the sector on the
n
can be shown that

1 1/ 2
(5.48)
which is independent of a so that am = a for all a. In particular for a cubic n = 3 and a = 30c .
5.7.2 Some Examples of Jump Resonance

As stated ab o ve the DF a nalysis presented in the p revio us sectio ns is b ased o n stea d y state co nditio ns. Whether a sp ecific steady state is rea chable is a d ynamic problem . The existence of jump reso nance b eha vio ur was reported in the 50 and 60s fro m so m e studies, a m o ngst others, of no nlinea r electrical circuits a nd a nalog ue co m p uter sim ulatio ns. The experim ents were usually done using sinusoidal signal generators with the amplitude and frequency controlled manually. Typically the a mplitud e f or the sinusoid al sig na l g enerator was set a t a fixed value a nd the f req uency va ried slo wly b y ha nd turning of the frequency setting dial, first in an increasing directio n to a certain value and then decreasing back to the starting freq uency. Clea rly there was no ca ref ul co ntrol of the d yna mics of the pro ced ure b ut two diff erent levels of o utp ut were fo und o ver the jum p ra ng e , o ne f or increasing a nd the o ther fo r decreasing the freq uency. Digita l sim ulatio n allo ws mo re precise control of the dynamics of a n experiment so for investigating the jump p henomenon questions arise as to ho w the sweep should be generated and how fast it should be to claim the steady state has been reached.
85
` aR j
< 0 a nd the jump resona nce points will be given b y
`j
=0. Further results ha ve b een found using these
Nyq uist pla ne b o und ed b y the ra dial lines 180 ! a m . Interestingly f or a p o wer la w no nlinearity n (x) = c x sgn (x) it
a = tan- 6(1 - n) /2n @
I a m ind ebted to H.K. Wo ng of the U niversity of Wa rwick f o r the m ajority of the results in this sectio n. To sta rt the stud y a sim ple f eedb a ck lo o p ha ving a tra nsf er f unctio n G (s) = 1/s (s + c) in the f o rward pa th a nd a cubic no nlinea rity in th e feedback path is considered. A simulink diagram is shown in Figure 5.9 with c = 0.4.
Figure 5.9 Simulink Block Diagram for Simulation Study
Deno ting the inp ut to the NL b y a co s ~t , which is the system o utp ut, a nd the inp ut to the system b y R co s (~t + b) , the loop balance equation is Rg cos (~t + b + i) - aM (a) g cos (~t + i) = a cos ~t giving Rg cos (b + i) - aM (a) g cos i = a and - R sin (b + i) + aM (a) sin i = 0. where G (j~) = g (~) e
ji (~)
(5.49)
(5.50)
(5.51)
and the dependence of g and on has been omitted.
Denoting aM(a)/R as y and a/R as x and eliminating from the above equatio ns gives
2 2 2
(5.52)
To sho w g rap hically the solutio ns this x-y relatio nship f or the f req uency respo nse of G is p lo tted f o r diff erent values of and the x-y relatio nship for the cub ic is then superimposed on it. The results are shown in Figure 5.10 for a range of
3
the relations hip in this way the possible solutions are easily seen for a fixed R and if R is changed only the sing le curve of the nonlinear relationship needs to be changed.
86
y + (x /g ) - (2xy cos i/g) = 1
together with the cubic relatio nship for R=1, which is simply y = x * M (x) and is y = 0.75x for the cubic. By plotting
Figure 5.10 Ellipses for G(j) with c = 0.4 and cubic nonlinearity for system of Figure 5.9
The intersections between each ellipse for a particular frequency with the nonlinearity inp ut -output (x-y) plot mark the amplitude solutions for the fundamental in this x-y spa ce. The coordinates of the intersections ca n then be used to comp ute the ma g nitud e o r p ha se resp o nse of the system . Jum p p heno m ena ma y o nly o ccur in the freq uency ra ng e where there is more than one intersection of the nonlinear characteristic with an ellipse for that frequency Substituting M (a) = 3a /4 for the cubic in eq uations (5.50) a nd (5.51) gives
2
Rg cos (b + i) - 0.75a g cos i = a

3
(5.53)
and - R sin (b + i) + 0.75a sin i = 0

3
(5.54)
87
These equations can be solved numerically for a and for given R and avoiding the graphical approach. However, initial g uesses fo r the solutio n values ha ve to be given to a no nlinea r algeb raic eq ua tio n solver, which is grea tly fa cilitated frequency response predictions are given in Figure 5.11 for c = 0.2 in the transfer function G and inp ut R = 1.
by kno wledge from the graphical approach. From software written using the graphical approach [5.13] the theoretical
Figure 5.11 Solutions from the graphical method for the frequency response with c = 0.2
A software tool [5.14] was used to obtain sweep frequency response tests in simulink. The tool was modified so that it could perform bi-directio nal sweeps necessary for the jump p henomeno n investigatio n. The input sweep generator cha nges
1 1
fro m ~ n to ~ 1 so that ~ k - 1 = r~ k , r < 1 fo r sweep d o wn. The freq uency increm ent is o nly a pp lied after the o utp ut sa tisf ies a stea d y sta te test. This is d o ne b y check ing o ver the la st n cycles of the o utp ut a nd co nfirming that the p ha se cha ng e is less tha n a specified tolera nce, . Avera ging is a lso p erfo rm ed f or m sp ecified cycles of the o utp ut to co m p ute the result. To o btain the results n a nd m were set to 3, to 10-5 a nd r to 0.995. Whe n the increm ent in f req uency is p erf o rm ed the system state conditions are preserved between frequencies, thus ensuring tha t conditions are similar to the previously m entio ned a nalo g ue co m p uter a p pro a ch. The m easured o utp ut is the f und a m enta l a m plitud e of the o utp ut wa vefo rm determined by Fo urier analysis, which should give best agreement with the DF analysis.
88
the freq uency of the sinusoidal inp ut from ~ 1 to ~ n rads/s progressively so that ~ k = r - ~k - 1, r - > 1 for sweep up , a nd
The results o b tained f or the cases of c = 0.4 a nd 0.2 with R = 1 a re sho wn in Fig ures 5.12 a nd 5.13, sup erim p osed o n th e theoretical curves. It ca n be seen tha t there is go od a greement b etween the theo retical a nd experim ental results particularly around the jump frequencies.
Figure 5.12 Swept frequency response test for c = 0.4 compared with the theory.
89
Figure 5.13 Swept frequency response test for c = 0.2 compared with the theory.
In a d ditio n sp ot tests at vario us freq uencies within the p redicted jum p ra ng e were und erta ken. This was do ne in sim ulink b y def ining the inp ut as a unit a m plitud e sine wa ve with p ha se zero a nd sim ula ting G (s) as in Fig ure 5.9, which allo ws initia l conditio ns to be applied to the integrator. For c = 0.4 two almost sinusoidal o utp ut amplitude solutions in good ag reement with the theory were obtained in the frequency range 1.38 to 1.49 as shown in Figure 5.14 for a frequency of 1.45. The
smaller amplitude solution was obtained with a zero initial condition and the larger one with an initial co nditio n of -2.
90
Figure 5.14 Plots of the two outputs for c = 0.4 and input frequency of 1.45 rads/s.
For c = 0.2 the larger output amplitude solution could only be found with an initial cond itio n on the integrator for frequencies above 1.54 up to 2.1 rads/s. With a zero initial condition on the integrator the output was nearly sinusoidal at the inp ut frequency until a frequency just exceeding 1.66 rad/s was reached when the lower amplitude contained a significant third harmonic which remained until a frequency of around 1.75 rads/s. Between approximately 1.75 rads/s and 1.80 rads/s the third sub harmo nic disappeared a nd the o utp ut again became almost sinusoidal at the inp ut freq uency. The third sub ha rm o nic co m p o nent ag ain returned f ro m just ab o ve 1.80 ra d s/s until it finally d isa pp ea red at 2.05 ra d s/s.

Gr aduat e oppo rtunities Ch eltenh a m | 24,945 + ben efits
One of the UKs int ellig ence services, GCHQs role is two-fold: to gather and analyse intelligence which helps shape Br itains response to global events, and, to p rovide technical advice for the protection of Government communication and infor mation systems. In doing so, our specialists in IT, internet , engineering, languages, information assurance, mathematics and int elligence get well beneath the surface of global affairs. If you thought the world was an interesting p lace, you really ought to exp lore our world of work.
TOP GOV ERNMENT EMPLOYER
www .careersinbritishi ntelligen ce.co.uk

Appli cants must be Br iti sh cit i zens. GCHQ values d iversi ty and wel comes app li cants from all sect ion s of the commu nity. We want our workf orce to reflec t the diver sity of our work .
91
Further investigation for frequencies between between 1.66 rads/s and 2.10 rads/s using different initial conditions on the integrator revealed three stable modes according to the initial conditions. These were two approximately sinusoidal a mplitud es a t the inp ut freq uency in g oo d a greem ent with the theory a nd a third with a sig nif ica nt sub harm o nic co ntent. Figure 5.15 shows these three outputs for an input frequency of 1.9 rads/s.
Figure 5.15 Plots of the three outputs for c = 0.2, R = 1 and an input frequency of 1.9 rads/s.
This example clearly illustrates that use of the approximate DF method can determine steady state frequency response solutions but canno t predict wha t can occur in a particular situa tion, as this is determined by the transient dyna mics when m o re tha n o ne sta ble state exists. The p ossibility of the third solutio n existing , na m ely the sub harm o nic o scillatio n can also be predicted by the DF method as is shown in the following. If the inp ut to the f eed b a ck loo p is ag ain ta k en a s R co s (~t + b) , then the inp ut to the cubic no nlinearity, which is also the output, may be taken as a cos ~t + b cos ((~t/3) + {) to investigate the possibility of a third subharmonic. The fundamental output signal from the cubic is aMa (a b) cos (~t + za (a b)) and the loop equation for the fundamental, with the dependence of g and , on shown explicitly in the equatio n, is Rg (~) cos (~t + b + i (~)) - aM a (a b) g (~) cos (~t + i (~) + za (a b)) = a cos ~t For the third subharmonic the corresponding loop equation is - bM b (a b) g (~/3) cos ((~t/3) + i (~/3) + zb (a b) + {) = b cos ((~t/3) {) (5.56) (5.55)
These two eq ua tio ns pro vid e f o ur no nlinea r eq ua tio ns which ca n b e solved f o r the unk no w ns a, b, a nd to ob tain a p ossible sub ha rm o nic solutio n when R a nd are given. A q uick test f o r a p o ssible sub ha rm o nic ca n b e o btained f ro m the phase shift requirement of equation (5.56), namely i (~/3) + zb (a b) = - 180c (5.57)
92
An Introduction to Nonlinearity in Control Systems From equation (5.2) it can be shown that
1 2 2
(5.58)
2 2
2 1
Thus a sub ha rmo nic ca n o nly o ccur pro vid ed the p ha se lag of G (j ) is grea ter tha n 159. The lo west freq uency f or which this occurs is approximately 0.52 rads/s, so that subharmonic oscillations might be possible when the input frequency exceeds 1.56 rads/s which is in reasonable agreement with the experimental results. The subharmonic only lasts for a sma ll f req uency ra ng e as the loo p g ain to the sub harm o nic of M b (a b) g (~ /3) d rop s b elo w unity. It ca n b e sho wn that the inverse of the d escribing f unctio n f or the sub ha rm o nic is a set o f circles d ep end ent o n within the sector of a ng le ! 20.7c a b o ut the nega tive real a xis of the Nyq uist p la ne. Thus f o r a sub ha rm o nic to exist fo r the cubic no nlinea rity in a feedback loop the Nyquist plot of G(j) must enter this sector. The disagreement in Figures 5.12 and 5.13 at low frequencies is due to the poor filtering of the harmonics provided by G (j ). If the cubic is m o ved to the fo rward lo op in fro nt of G (s) the d iscrep a ncy a t lo w freq uencies is less a s the sinusoid al loo p inp ut sig nal p asses straig ht to the cub ic no nlinearity. This is illustra ted b y the f req uency resp o nse sho wn in Fig ure 5.16 for the case o f c = 0.4 a nd R = 1. Here, ho wever, a nother situatio n arises in that theo retically there a re o nly two solutio ns ab o ve a f req uency of a ro und 2.25 rad s/s. a nd thus there is no jum p d o wn o n the up wa rd freq uency sweep as co nfirm ed b y the sweep f req uency test, d ue to the lo o p ha ving a negligible feed b ack sig nal. In p ra ctice, of co urse the increasing g ain of the cubic is not po ssible a nd a jum p d o wn will result. Spo t testing sho ws tha t if no initial co nditio ns are p la ced o n the integ ra to r the so lutio ns of the d ecreasing freq uency sweep a re o btained . Ho wever ab o ve the freq uency of 2.25 ra ds/s the high amplitude stable solution can be obtained by putting initial conditions on the integrator.
93
zb (a b) = - tan- 6ab sin 3{/ (b + 2a + ab cos 3{)@ Differentiation of the exp ression in the brackets sho ws that it is a ma ximum when cos 3 { = - ab/ (b + 2a ) which gives the m a xim um va lue fo r z b (a b) of ! sin - 6 ab/ (b + 2a ) @ . Deno ting the ratio o f b/a b y gives sin - 6t / (t + 2) @ , whic h on differentia tion gives a max im um when t = 2 and a max im um v alue of sin - ( 2 /4) = 20.7c .
Figure 5.16 Frequency response for cubic in the forward path with c = 0.4.
In Paris or Online
International prog rams taught by professors an d professi onals from all ov er the w orld BBA in G lobal Business M BA in International M anagement / International M arketing DBA in International Business / Interna tional M anagement MA in International E ducation MA in C ross-C ultural C ommunication MA in F oreign Languages Innov ative P ractical F lexible A ffordable Visit: www.HorizonsU niversity.org Write: A dmissions@horizo nsuniv ersity .org C all: 01.42.77.20.66
94
Results are given in Fig ure 5.17 for an ideal saturation characteristic in the forward path before the same G(s) with c = 0.2 a nd R = 1. In this case it is seen that the theoretical reso na nt peak b ends in a back ward direction so that the smaller jump occurs on the upward frequency sweep.
Figure 5.17 Frequency response results for saturation in the forward path, c = 0.2 and R = 1.
5.8 Conclusions
In this chap ter the respo nse of a no nlinearity to two sinuso id al inp uts wa s first discussed a nd altho ug h it wa s sho wn tha t a m a them a tical f orm ulatio n is straig htf o rward ob taining a nalytical so lutio ns f o r o ther tha n sim ply d efined m a them a tical characteristics is not. Thus numerical comp uta tion approaches have to be adopted for most practical situations. The co ncep ts were used , ho wever, to sho w ho w va rio us p rop erties of f eed ba ck systems co uld be investig ated b y co nsidering two freq uencies present in the lo o p. The a nalytical wo rk ha s b een sup p o rted b y sim ulatio n results a nd the cha pter wa s concluded by a detailed studied of jump resonance.
95
5.9 References
5.1 McNamara, OP & Atherto n, DP 1985, An Iterative Method for Limit Cycle Determinatio n, Control 85, Cambridge, pp 252-257. 5.2 Atherto n, DP, Mc Na m a ra , OP, Wa d ey, MD & G o ucem A 1985, SU NS: the Sussex U niversity No nlinear Co ntrol System s Software. 3rd IFAC/ICAD Symposium on Computer Aided Design in Co ntrol Engineering Systems, Copenhagen, pp 173-178. 5.3 McNamara, OP & Atherto n, DP, 1987, Limit Cycle Predictio n in Free Structured Nonlinear Systems, IFAC Congress, Munich, Volume 8, pp 23-28. 5.4 Atherton, DP & Ramani, N 1975, A Method for the Evaluation of Limit Cycles. Internatio nal Journal of Control, Vol 21, pp 375-384 5.5 Hammo nd, PH 1958, Feedback Theory and its Application, Chap ter 11, EUP, London. 5.6 West, JC and Douce, JL 1952, The Frequency Response of a Certain Class of Nonlinear Feedback Systems, British Journal of Applied Physics Vol 5, pp.204-210. 5.7 Sing h, YP 1964, G ra p hical Metho d f or find ing the Clo sed Lo op Freq uency R esp o nse of No nlinea r Feed b a ck Co ntrol Systems, Proc IEE, Vol 112, pp 2167-2170. 5.8 West, JC & Nikiforuk, P 1954, The Behaviour of a Remote Position Control Servomecha nism with Hard Spring Nonlinear Characteristics, Proc IEE, Vol 101, part II, pp 481-487. 5.9 West, JC, Jayawant, BV & Rea, DP 1967, Transition Characteristics of the Jump Phenomenon in Nonlinear Resonant Circuits, Proc IEE, Vol 114, pp 381-392. 5.10 La m b a , SS & Ka va na g h, R J 1971, The p heno m eno n of Isolated Jum p R eso na nce a nd its Ap plica tio ns, Pro c IE E , Vo l 118, pp 1047-1050. 5.11 Hirai, K &Sawai, N 1978, A General Criterion for Jump Resonance of Nonlinear Control Systems, IEEE Trans. on Auto. Control, pp. 896-901. 5.12 Ha tanaka , H 1963, The Frequency Respo nse a nd Jump Reso nance Phenomena of No nlinear Feedback Control Systems, Trans ASME J Basic Eng, D-85, pp236-242. 5.13 Wong, HK 2010, Priva te communicatio ns 5.14 Royalty, J 2007, Dynamic Signal Analyzer , MATLAB Central [online], http://www.mathworks.com/ma tlabcentral/fileexchange/14611
96
Limit cycles in relay systems
6 Limit cycles in relay systems

6.1 Introduction
In the ea rly da ys of co ntrol eng ineering , when p o wer electro nics was p rim itive co m pa red with the situa tio n tod a y, rela ys were co m m o nly used as they p ro vid ed a chea p f o rm of p o wer a m plificatio n. They a re still used in tem p erature co ntro l, where the temp erature is co ntrolled to limit cycle a ro und the d esired value, a nd switching techniq ues using m od ern po wer electro nics are also freq uently used . The distinctio n b etween switching no nlinearities a nd other no nlinea rities is tha t their inp ut d oes no t co ntro l the o utp ut co ntinuo usly b ut o nly d eterm ines the insta nt of switching . The inp ut has no co ntrol o f the o utp ut b etween switching insta nts This f eature, that the inp ut has no co ntrol of the o utp ut b etween switching insta nts, allows a unique theory to be developed for analysing limit cycles in feedback loops with relay elements. The sta rting point for the analysis is to assume a sp ecific o utp ut wa veform from the rela y a nd d evelop equatio ns which m ust be satisfied f or the a ssum ed wa vefo rm to exist. The m ost co m mo n situatio n is the case of a rela y with no d ea d zo ne where the o utp ut will typically be assum ed to b e a sq uare wa vef orm w ith a o ne to o ne m ark spa ce ra tio . Mo re co m plex m o d es with multiple pulses per cycle can exist in some systems and their prediction is still possible, however more nonlinea r equatio ns need to b e satisfied, thus typically req uiring a comp utational app roach. The a nalysis starts b y assuming a periodic o utp ut wa vef orm fro m the rela y a nd then the req uirem ent is to o btain the stea d y state periodic o utp ut this p ro d uces af ter passage through the transfer function G(s). This is also a problem of interest in circuit theory [6.1].
Destination MMU
MMU is proud to be one of the most popular universities in the UK. Some 34,000 students from all par ts of the globe select from its curricula of over 1,000 courses and qualifications. We are based in the dynamic yet conveniently compact city of Manchester, located at the heart of a sophisticated transport networ k including a major international airport on the outskir ts. Parts of the campus are acclaimed for their architectural style and date back over 150 years, in direct contrast to our teaching style which is thoroughly modern, innovative and forward-thinking. MMU offers under graduate and postgraduate courses in the following subject areas: Art, Design & Performance Computing, Engineering & Technology Business & Management Science, Environmental Studies & Geography Law, Education & Psychology F ood, H ospit ali ty , T ouri sm & L ei sure St udi es Humanities & Social Science
97
For more details or an application form please contact MMU International. email: international @mmu.a c.uk telephone: +44 (0)161 247 1022 www.mmu.ac.uk/inter national
Fo r a p plicatio n to the rela y f eed ba ck p rob lem three diff erent m etho d s ha ve b een p ro p o sed f or do ing the a nalysis all of which pro d uce id entical results. A tim e do m ain a pp ro a ch wa s d evelo ped b y Ha m el [6.2], a freq uency d o ma in a p pro a ch by Tsypkin [6.3], and a time domain approach using a state space representation by Chung [6.4, 6.5]. The frequency do m ain ap proa ch uses a sum m ed freq uency lo cus which m ak es it ea sy to co mp are with the DF m ethod so that this will be the method considered below. The formulation is done in ter ms of A loci which are more general than the Tsypkin loci, and this is followed with details on the properties of A loci and their evaluation. Solutio ns for limit cycles of the f und a m ental fo rm are d iscussed . In sectio n 6.5 it is sho wn tha t a n exa ct co nditio n c a n b e f o und giving necessa ry a n d sufficient co nditio ns f or the sta bility of a ny p redicted limit cycle, which is a n extrem ely usef ul result. This ena bles the stud y in sectio n 6.6 of so m e very interesting p ro blem s rela ting to lim it cycles in rela y system s wh ich ha ve b een o btained fro m sof twa re im plem entatio ns of the theo ry presented . To pics discussed includ e limit cycles with m ultiple p ulses per p erio d , limit cycles with sliding a nd co ncepts fo r the predictio n of cha o tic m otio n. Finally in sectio n 6.7 the ha r m o nica lly forced response of relay systems is considered.
6.2 The Frequency Domain Approach

Co nsider a p eriod ic p ulse wa vefo rm fo r which a o ne p erio d exa m ple is illustrated in Fig ure 6.1, where it is a ssum ed to be od d sym m etrical a nd to ha ve eig ht p ulses in a p erio d . If a ny o ne of the eig ht p ulses is d enoted b y y i (t ) which is a ssum e d to be of width Dti and to start at time ti then it is straightforward to show that it has the Fo urier series
3 n =1
(6.1)
If y i (t) is the inp ut to a linear system with the tra nsfer f unctio n G (s) , then the Fo urier series fo r the o utp ut c i (t ) , assuming lim s " 3 G (s) = 0 and G (s) is finite, is given by ci (t) = (hDti G (0) /T) + (h/r)
3 n =1
(6.2)
where the frequency response of G (s) is denoted by G (jn~) = gn e

j{ n
= UG (n~) + jVG (n~)
(6.3)
Figure 6.1 One period of an odd symmetrical multipulse waveform.
98
y i (t ) = (hDt i /T) + (h/r) / (1/n) 6sin n~ Dt i cos n~ (t - t i ) + (1 - cos n~ Dt i ) sin n~ (t - t i ) @
/ (g n /n)"sin n~Dt i cos 6n~ (t - t i ) + { n@ + (1 - cos n~Dt i) sin 6 n~ (t - t i ) + { n @,
If equatio n (6.2) is differentiated then the Fo urier series for coi (t) , provided lim s " 3 sG (s) = 0 , is coi (t) = (~h/r)
3 n =1 n
/ (g /n) " - sin n~ Dt i sin 6 n~ (t - t i ) + { n @ + (1 - cos n~ Dt i ) cos 6 n~ (t - t i ) + { n @,
(6.4)
When lim s " 3 sG (s) ! 0, ci (t) , will have discontinuities at the switching instants ti and ti + Dti . Since the Fourier series tends to the mea n value of the functio n at a discontinuity at time, t, then a term 0.5" yi (t +) - yi (t -), lim s " 3 sG (s) with appropriate sign must be included in equatio n (6.4) to obtain the correct value of ci (t) at these two time ins tants. It is co nvenient to exp ress c (t ) a nd co (t ) in term s of the A lo cus [6.6, 6.7], A (i, ~ ) , where i = ~ t , a nd is d efined fo r a tra nsf er functio n G(s) by AG (i, ~) = Re AG (i, ~) + j Im AG (i, ~) with Re AG (i, ~) = / VG (i, n~) sin ni + UG (i, n~) cos ni
3 n =1
(6.5)
(6.6)
and Im A G (i, ~ ) = / 3 (1/n) 6 V G (i, n~ ) cos ni + U G (i, n~ ) sin ni @

n =1
(6.7)
AG (i, ~ ) is a generalised summed frequency locus with its real a nd imaginary values at a particular freq uency, , dep ending on weighted values, according to the choice of, , of the real and imaginary values of G(j) at the frequencies n fo r n = 1,2,.. In particular for = 0 the real (imaginary) part of AG (i, ~) depends only upon the real (imaginary) part of G(j) at frequencies n. It is possible to evaluate Re AG (i, ~) and Im AG (i, ~) in closed form for specific transfer f unctio ns f ro m k no w n infinite series, as is discussed in the next sectio n, o r a p pro xim a tely b y co m p uta tio nally sum m in g to a finite number, n, of terms. Using the above relationships in equa tions (6.2) and (6.4) gives ci (t) = (hDti G (0) /T) + (h/r)6Im AG (- ~t + ~ti, ~) - Im AG (- ~t + ~ti + ~Dti, ~)@ and coi (t) = (~h/r)6Re AG (- ~t + ~ti, ~) - Re AG (- ~t + ~ti + ~Dti, ~)@ (6.9) (6.8)
Eq uatio n (6.9) m a y also be seen to b e ob ta ina ble directly fro m eq uatio n (6.8) since f ro m the definitio ns of eq uatio ns (6.6) and (6.7) it can be seen that dG d Im 6 A i(i, ~)@ = - Re AG (i, ~) (6.10)
99
Any p erio dic p ulse wa vef orm ca n b e co nstructed f ro m a sum m a tio n o f wa vef orm s y i (t ) with diff erent o rigins a nd p ulse widths. For example, the symmetrical square wave, y (t) , of period T is given by y (t) = /=yi (t) i 1
2
(6.11)
with y 1 (t ) switching po sitive a t tim e t 1 , y 2 (t ) , switching neg a tive a t tim e t 2 ; a nd with t 2 = t 1 + 0.5T a nd Dt 1 = Dt 2 = 0.5 T . Choosing the time origin t 1 = 0 , then the outp ut c(t) becomes c (t) = (2h/r)6Im AG (- ~t, ~) - Im AG (- ~t + 0.5~T, ~)@ Since 0.5~T = r , it is easily shown tha t this gives c (t) = (4h/r)6Im A G (- ~t , ~ )@ o (6.13) (6.12)
and, similarly, co (t) = (4~h/r)6Re A G 0 ~t, ~ ) @ ((6.14)
Here the A loci A G are again defined by eq uations (6.5) to (6.7) b ut with the summatio ns including odd terms only, that
o
is n = 1, 3, 5.. Bef ore pro ceeding to sho w ho w limit cycles ca n b e fo und using these exp ressio ns it is a p pro priate to summarise some properties of the A loci and show how they ca n be obtained for specific transfer functions.

100
6.3 Properties and Evaluation of A loci

Some further properties of the A loci, in addition to equation (6.10), which are easily found from the definitions are: (1) The A locus fo r = 0 is id entical with the Tsypkin lo cus [6.3], K ( ~ ) , if lim constant factor. The relationship is K (~) = (4h/r) A (0, ~) The A locus ca n thus be rega rded as a generalised Tsypkin locus. (2) The A locus satisfies the superposition property, that is if the linear plant G (s) = G1 (s) + G2 (s) , then AG (i, ~) = AG (i, ~) + AG (i, ~)
1 2
s"3
sG (s) = 0 , ap art fro m a
(6.15)
(6.16)
(3) If G1 (s) = G (s) e- , then AG (i, ~) = AG (i + ~x, ~)

1
sx
(6.17)
(4) The loci are periodic in , with period 2, that is AG (i, ~) = AG (i + 2r, ~) (5) For Ac the periodicity is odd, that is AG (i, ~) = - AG (- i, ~)
o o
(6.18)
(6.19)
Since a ny pla nt tra nsf er f unctio n ca n b e written in term s of a sum m a tio n o f tra nsf er f unctio ns ha ving a real o r co m p lex pa ir of p oles, use o f eq uatio n (6.16) allo ws A lo ci f or m o st tra nsf er f unctio ns to b e o btained in term s o f the A lo ci of the few basic transfer functions given in Table 6.1 in the Appendix 6.10. In addition A loci for transfer functions with time delays can a nd 6.3 respectively, for the A nd A loci. For the real pole situa loci the loci are expressible in of Table in Tables 6.2 be obtained using equationa(6.17). Expressions for the A tionsof the transfer functionsterms of 6.1 are given
o
well known sine and cosine series.
101
6.4 Solving for Limit Cycles

The results of the p revio us sectio n ha ve sho wn ho w a closed f o rm exp ressio n ca n be ob ta ined f or the o utp ut of a linea r tra nsf er f unctio n with a p eriod ic p ulse wa vefo rm inp ut. Thus the sig nal f ed b a ck to the rela y inp ut ca n b e f o und . If this pulse type limit cycle is to exist then the follo wing three conditions need to be satisfied: (1) The input mus t pass through the relay switching values in the correct direction at the appropriate times to provide the assumed periodic output. The number of these switching conditions to be satisfied will depend on the form of periodic mode being investigated. The simplest case is discussed in the next section for a relay with no dead zone with a symmetrical square wave outp ut, which yields one switch ing condition and thus the req uirem ent to solve o ne no nlinea r a lg eb ra ic eq ua tio n. For co m plex wa vef o rm s there m a y b e m a ny switching conditio ns resulting in the requirement to solve many no nlinear algebraic equations (2) Once the switching co nditio ns ha ve been f o und the rela y inp ut wa vef o rm ca n b e co m p uted . This m ust satisf y the continuity co nditio ns, that is it must no t pass through the relay switching levels at times which would cause the relay outp ut to be different from that assumed. (3) The limit cycle must be stable. Altho ug h a nalytical solutio ns ca n b e o b tained f or so m e sim ple cases of o ne or two switching s p er lim it cycle p eriod , the best approach is to use software with graphics [6.8]. This then allo ws the above three conditions to be accomplished , na m ely solutio n of the no nlinea r alg ebraic eq uatio ns of (1), disp la y of solutio n wa vef orm s to check (2), a nd the sta bility computa tions, discussed in section 6.5, to be carried out for (3)
6.4.1 Relay with no Dead Zone

For a rela y with no dea d zo ne the sim plest assum ptio n fo r the p eriod ic o utp ut is a sq uare wa ve with 1:1 m ark sp a ce ratio . Equations (6.13) and (6.14) give expressions for the waveform and its derivative produced at the outp ut of the linear transfer function G, where the positive switching of the square wave was assumed to be at t = 0. Taking the inp ut to the relay, x (t) = - c (t) , then the required switching condition, assuming the relay to have a hysteresis of ! D , that is it switches from h to +h for an input of and from +h to -h for an input of -, is x (0-) = D and xo (0-) > 0 (6.20)
The is used to sho w tha t the tim e sho uld b e tak en b ef ore a ny p ossible d isco ntinuity which m a y b e created d ue to the switching. Thus using (6.13) and (6.14) one obtains Im A Go (0, ~) = - rD/4h and Re A Go (0, ~) < 0 provided lim s " 3 sG (s) = 0 . (6.21)
102
For a given G(s) the solution for the unknown frequency of the limit cycle, , ca n be found by solving the equality of eq uatio n (6.21) a nd check ing the ineq uality, or do ne gra p hically f ro m a plo t of A Go (0, ~ ) o n a Nyq uist d ia g ra m to o btain where it intersects the line - rD/4h . This approach allows easy comparison with the DF method where the solution is the intersectio n of G (j~) with the same line. Note from the definitio n A Go (0, ~) has real and imaginary parts given by Re A Go (0, ~) = and Im A Go (0, ~) =
3 3 n = 1 ( 2) n = 1 ( 2) G 3 n = 1 (2)
UG (0, n~) =
3 n = 1 (2)
(6.22)
(6.23)
Thus the locus ca n be obtained app roximately from these expressions by taking a small value for n, a nd exa ctly b y sum ming to a large value of n computationally or using the analytical results of Table 6.2. For the specific example of G (s) = K/6s (1 + sx) @, one has using Table 6.2 that Im A Go (0, ~) = (Krx/4)6(- r/2m) + tanh (r/2m) @ where m = ~x . Thus using equation (6.21) the solutio n for the limit cycle is (r/2m) - tanh (r/2m) = DhK/x and the ineq uality is satisfied as Re A Go (0, ~) < 0 for all . The DF method for this problem yields the equation m (1 + m2) = 4hKx/D for the normalised frequency, , of the limit cycle. For the particular case of K = 1, x = 1, h/D = 3 , the exact limit cycle frequency solutio n is 1.365 and tha t given by the DF method is 1.352. As expected if the solution is examined as the parameter, x , changes then, since as x reduces the tra nsf er f unctio n is a wo rse lo w pa ss filter, the DF solutio n b eco m es less a ccurate b ut it is less tha n 3% in erro r fo r x > 0.25. When the square wave is asymmetrical, that is not a 1:1 mark to space ratio, the two pulse trains of equation (6.11) will be y1 (t) switching positive at t1 = 0 with Dt1 = Dt , and y2 (t) switching negative at time t2 = Dt and with Dt2 = T - Dt . The required relay inputs of and - at the switching times of 0- (positive) and Dt- (negative) then yield two nonlinear algebraic equations to solve for Dt and . (6.26) (6.25) (6.24)
103
/ Re G (jn~)
/ (1/n) V (0, n~) = /(1/n) Im G (jn~) .
6.4.2 Relay with Dead Zone

To evaluate any odd sym metrical limit cycle in a loop having a relay with dead zo ne, the rela y o utput ca n again be assumed to be the sum of two pulse trains as given in equa tion (6.11) but no w with pulse widths Dt1 = Dt2 = Dt ! 0.5T . The positive switching condition at time t1 , which will be taken as 0, will be x (0-) = d + D and ox (0-) > 0 and the negative one at time Dt x (Dt-) = d + D and ox (Dt-) > 0 The expressions for c (t) and co (t) can be shown to be c (t) = (2h/r)6Im A Go (0, ~) - Im A Go (~Dt, ~) @ and co (t) = (2~h/r)6Re A Go (0, ~) - Re A Go (~Dt, ~) @. (6.30) (6.29) (6.28) (6.27)

INNOVATION, FLAT HIERARCHIES AND OPEN-MINDED PROFESSO RS

CLOSE COLLABORATION WITH FUTURE EMPLOYERS SUCH AS ABB AND ERICSSO N
SASHA SHAHBAZI LEFT IRAN FOR A MASTER S IN PRODUCT AND PROCES S DEVELOPMENT AND LOTS OF INNEB AND Y
HELL TELL YOU ALL ABOUT IT AND ANSWER YOU R QUESTIONS AT MDUSTUDENT. COM
www. mdh.se
104
Thus tha t the switching conditions, for x (t) = - c (t) , are Im A Go (0, ~) - Im A Go (~Dt, ~) = - (r/2h) (d - D) Re A Go (0, ~) - Re A Go (~Dt, ~) < 0 and Im A Go (0, ~) - Im A Go (~Dt, ~) = - (r/2h) (d - D) Re A Go (0, ~) - Re A Go (~Dt, ~) > 0 (6.32) (6.31)
For a specific example one thus has two nonlinear algebraic equations (6.31) and (6.32) to solve for the two unkno wns a nd t. O nce these are k no w n the p ulse wa vef orm is k no w n a nd the limit cycle wa vef o rm ca n b e o b ta ined f ro m th e A locus expression or by other means. Plotting this waveform then allo ws conditio n (3) in sectio n (6.4) to be checked. To investigate the possibility of an asymmetrical limit cycle for the relay with dead zone requires the solution of four nonlinear algebraic equations.
6.5 Limit Cycle Stability

When discussing limit cycle stability using the DF method the perturbatio n equatio n (4.5), which has periodically time varying coefficients, was found and is repeated below. q (D) dx (t) + p (D) n' (x * (t)) dx (t)) = 0 (6.33)
In co ntra st to the DF ca se the exa ct limit cycle solutio n no t a sinusoid al a pp ro xim atio n is given b y the ab o ve a nalysis f or feedback loops with relays. Further if this equation is written in the sta te space form dxo * (t) = A (t) dx * (t) then it can be shown tha t the periodically time varying matrix, A(t), is given by A (t) = A + Bn' (x * (t)) C (6.35) (6.34)
where (A, B, C) is a state space description of G(s). The matrix of equation (6.35) is piecewise constant and equal to A except over the infinitesimal time intervals when the relay switches. The value over the relay switching time may be calculated by ta king the relay as a limiting case of saturation. Fo r exa m ple if the rela y switches f ro m h to + h this is the limit of a n id ea l sa turatio n cha ra cteristic as f " 0 m o ving fro m a level h at to +h at + and the value of the second term in equatio n (6.35) is (6.36)
lim f " 0 BC (h/f) = lim f " 0 BC62h/xo (t1) D @
105
where t i is the tim e a t which switching ta kes pla ce a nd the d uratio n of the switching tim e which tend s to zero as tend s to zero. Thus the product A(t) over the infinitesimal switching time is given by A (t) D = BC 62h/xo (ti) @ (6.37)
The term in the sq ua re bra ck ets is the cha ng e in rela y o utp ut divid ed b y the d erivative of the rela y inp ut a t the switching instant. The value of this result is tha t a necessary and sufficient result, for the stability of a differential equation with a p erio dica lly tim e varying A m atrix which is piecewise co nsta nt, is k no wn[6.9, 6. 10]. The result is that the system is stab le if all the eig envalues of the ma trix Q ha ve ma g nitud e less tha n unity ap art f ro m o ne, corresp o nding to the perio dic so lutio n, which will have unit magnitude. The matrix, Q, is given by
m i=1
(6.38)
where the Ai are the constant values of the A matrix over the time periods Dti and
m i=1
is the lim it cycle p erio d with m values of the A m atrix. Fo r a sym m etrical o dd lim it cycle Q m a y b e co m p uted o ver 0.5 T. Thus for a limit cycle in a feedback system with an ideal relay with levels ! h switching at times 0 and 0.5T, the Q matrix is giv en by Q = exp 6 AT/2 @exp 62hBC/xo (0) @ (6.39)
6.6 Some Interesting Limit Cycle Problems

For most feedback systems with relay elements the limit cycles obtained are of what mig ht be called the fundamental type discussed in section 6.4. That is (1) for a relay with no dead zone the relay will switch twice per period of the limit cycle, namely at times 0 and t in the period, T, where if the limit cycle has odd symmetry Dt = 0.5T and (2) for a relay with dead zone the relay will switch four times per period. In this section, rather tha n consider examples for these fund a m ental ca ses so m e m ore interesting exa m ples are co nsid ered . The results f o r these prob lem s were f o und using the theory described abo ve a nd som e extensions. The solutio ns were obtained using Fortra n prog rams with interactive graphics to allow the display of the limit cycle waveforms.
6.6.1 Example 1 Invalid Continuity Condition

Consider an ideal relay with outputs
2
!1
in a negative feedback loop with the transfer functio n
a nd m o ves o ut to infinity at unit freq uency d ue to the und a m p ed seco nd o rd er term in the d eno m inator. Af ter a p hase cha ng e of -180 a ro und a n infinite semicircle it returns to the origin in the third q ua d ra nt. There is no intersectio n of the neg ative real a xis so the d escribing f unctio n ind ica tes no limit cycle. If the Tsypkin m etho d is used then it ca n b e sho wn
that it gives limit cycle solutio n frequencies of ~ = 1/ (2n + 0.5) for n = 1,2,... Checking the solution waveforms for these freq uencies they are f o und to cross the zero switching level o f the rela y a t tim es other tha n tho se a ssum ed , d ue to the reso na nt na ture of the pla nt t ra nsfer f unctio n. Thus they f ail the co ntinuity co nditio n (2) g iven in sectio n 6.4 a nd a s predicted by the DF no limit cycle exists.
106
Q=
% A Dt
i
T=
% Dt
G (s) = (1 + 2s) / (s + 1) (s + 1), [6.10-6.13]. The Nyq uist lo cus of this tra nsf er f unctio n sta rts in the first q uad ra nt
6.6.2 Example 2 Multipulse Limit Cycles

Here a very interesting p ra ctical exa m ple co ncerning the p ossib ility of m ultip ulse oscilla tio ns in the a ttitude co ntrol system of a co m m unica tio n satellite in a geosta tio na ry o rbit is discussed [6.14 -6.17]. The roll -ya w lo op is co ntro lled b y a pseudo rate co ntroller which co nsists of jet thrusters with f eed ba ck thro ug h a tim e co nsta nt which has a diff erent value when th e thrusters are o n to when they a re off. Thus o n its o wn the controller is a rela y feedba ck loop containing a mode dependent tra nsf er f unctio n. The satellite ha s sola r p a nels f o r p o wer collectio n which m ea ns that the pla nt tra nsf er f unctio n co nsists of the rig id b o d y d yna mics of a d o uble integrator plus lig htly da m p ed seco nd o rd er tra nsf er f unctio ns, corresp o nding to the panels resonant modes, in parallel. A block diagram for the system is shown in Figure 6.2.
Figure 6.2 Block diagram for system of example 2 (copied from reference 6.15)
107
R ef erence 6.15 discusses the d etermina tio n of the m ultip ulse limit cycles using the DF, b ut here the co m m ents will ref er to the use of the Tsyp kin m etho d d iscussed in p a p ers [6.15, 6.16]. There a re two o rigina l asp ects in these p ap ers the first being the extension of the Tsypkin method to evaluate quite complex limit cycles with multiple pulses per period and the seco nd b eing a n extensio n of the m etho d to find limit cycles in m o de dep end ent rela y system s. Here the term m od e dependent is used to refer to the fact that the loop transfer function changes with the relay output level. This topic is considered further in reference 6.17. To determine the possible existence of an odd symmetrical multipulse oscillation of n pulses per half period requires a solutio n to be f o und to 2 n alg eb ra ic eq ua tio ns. R esults were o btained b y this ap p ro a ch in ref erence 6.16 fo r o ne assum ed flexible mode and are given for the pseudo rate controller with equal and, then the usual case, unequal feedback time constants. Results with up to five pulses per half period were calculated and checked with analogue simulatio n results.
6.6.3 Example 3 Limit Cycle with a Sliding Mode

The investig atio n of the step resp o nse of a rela y co ntrol system in sectio n 2.3.2 sho wed tha t a slid ing m o tio n, where the relay switches rapidly, was found. This therefore raises the interesting question of whether a limit cycle can exist with a sliding m o d e a nd if so ho w ca n it b e d etermined theo retically. This p ro blem is discussed in ref erences 6.10, 6.18 a nd 6.19 where a relay with a dead zone of ! 1 , output levels ! 1 and linear transfer function G (s) = 8 (s - 0.5) / (s2 - 0.5s + 1) was studied . The relative deg ree of the tra nsf er f unctio n was tak en a s o ne since it ca n b e sho wn that sliding m otio n und er relay control is not possible for a relative degree greater than unity. The Tsypkin method gave the solution for a limit cycle shown in Figure 6.3, which violates the continuity condition. The interesting point is tha t the waveform reverses im m ediately af ter the switching a t A a nd then soo n reverses a gain to pa ss thro ug h the switching level at C, which is arg ued is an indication of sliding with the first estimate for the time of sliding being from A to C.
Figure 6.3 T sypkin solution for limit cycle (taken from reference 6.19)
108
To see if a sliding solutio n co uld be o btained the rela y o utp ut was a pp ro xim ated b y a straig ht line o ver this p erio d a nd the Tsyp kin m ethod extend ed to ob ta in a solutio n fo r a rela y o utp ut wa vefo rm ha ving this shap e. A new lim it cycle solutio n was then f o und sho wing tha t after the switching the calcula ted rela y inp ut rem ained almo st co nsta nt a s req uired to verif y sliding. By adding a further linear segment to approximate the average relay output waveform in the sliding mode by two linear segments the method gave the limit cycle sho wn in Figure 6.4 with a frequency of 0.386 rads/s, which was in a greem ent with the sim ula tio n result. It ca n b e seen tha t the rela y inp ut is no w alm o st co nsta nt a t the switching level during the sliding period.
Figure 6.4 Limit cycle solution using T sypkin extension (taken from reference 6.19)
The example has also been done in simulink using a fixed step integration algorithm since the default optio n with a varia ble step length stop s when sliding sta rts. The sim ulink diag ra m is sho wn in Fig ure 6.5, where the b lo cks in the up p er left hand corner generate a time vector, z, of equal length to the output vectors x and y.
Figure 6.5 Simulink block diagram for system having limit cycle with sliding
The limit cycle at the input to the relay, x, is shown in Figure 6.6 and in the x-y plane in Fig ure 6.7. The initial conditio n was set slightly off the limit cycle.
109
Figure 6.6. Waveform of limit cycle at the input to the relay.
Figure 6.7 Limit cycle plot in the phase plane
Brain power
By 2020, w ind could prov ide one-tenth of our p lanets electricity needs. A lready today , SKF s innov ativ e know how is crucial to ru nning a large propo rtion of the w orlds w ind turbines. U p to 25 % of the generating costs relate to maintenance. These can be reduced d ramatically thanks to our sy stems for on-line conditio n monitoring and automati c lubricati on. We help make it more economical to create cleaner, cheaper ene rgy out of thin air. By sharing our expe rience, expertise, and creativ ity , industrie s can boost performance bey ond expectat ions. Therefore w e need the best employ ees w ho can meet this challenge !
The Power of Knowledge Engineering
Plug into The Power of K nowledge E ngineering. Visit us at www.s kf.com/kno wledge
110
6.6.4 Example 4 Multiple Limit Cycle Solutions

For a f eed ba ck lo op co ntaining a sing le valued no nlinea rity, which ha s a C(a) lo cus o n the neg ative real a xes a nd a linea r tra nsf er f unctio n which ha s a freq uency resp o nse with m o re tha n o ne intersectio n of this a xis, vario us p ossib le limit cycles m a y result. A stud y of this situatio n using the a pp ro xim ate DF a nalysis, where bo th cubic a nd id ea l rela y no nlinea rities are ta k en, ca n b e f o und in ref erence 6.20. The Tsypkin m etho d ena bles the existence of f und a m e ntal lim i t cycles a nd their stability to be found but no t limit cycles at combinations of the intersection frequencies. To illustrate the situa tion the case of a n id eal rela y in a neg ative f eed ba ck lo op with the linear tra nsf er f unctio n G (s) = 12 (s + 1) 2 /s3 (s2 + 0.3m s + m 2 ) as considered in reference 6.20 is discussed. The Nyquist locus for the transfer functio n with = 4.0 is shown in Figure 6.8. The negative real axis is cut at the two frequencies 1 and 2 and the ratio a = G (j~1) / G (j~2) decreases as decreases. Using the IDF approach of the DF method the limit cycle at frequency 1 is always uns table and that at 2 is stable if > 2 which corresponds to > 4.05. This compares with the Tsypkin solutio n which gave an unstable limit cycle at frequency 1 for > 3.73, but no solution for smaller values of . The limit cycle at 2 was only stable for > 3.73. Simulatio ns gave good agreement with the Tsypkin theory for a single mode limit cycle but it cannot predict the co m bined m o d e. U sing the SSDF a nd t he ISSDF, the increm ental d escribing f unctio n in the presence of two sinusoids, a stable co m bined f req uency lim it cycle was f o und to exist f or < 4.05 in ref erence 6.20. G oo d a g reem ent wa s also fo un d with the amplitude and frequency components in the combined mode.
Figure 6.8 Nyquist plot for example 4
6.6.5 Example 5 Chaotic Motion

Altho ug h no theo retica l solutio n ca n be f o und f o r a cha otic mo tio n there has b een sig nifica nt resea rch d o ne in trying to find co nditio ns fo r its existence in a uto no m o us f eed ba ck loo ps. In these studies b oth the a pp ro xim ate describing f unctio n approach [6.21] and Tsypkin method [6.22] have been used. Useful starting points are provided by the conjectures of Og o rzalek [6.23] a nd fro m the results of ref erence 6.21. Og o rzalek co njectured tha t f o r a neg ative f eed ba ck system with G (s) = - b/ (s n + an - 1 s n - 1 + .......... + a1 s + a0 ) 111
(6.40)
where all the coefficients are real and positive, and with nonlinearity n(x), chaotic motio n was likely to exist if: (i) The Hurwitz sector of the linear system with n(x) rep la ced b y K satisfies K2 < K < K 1 , where K 1 = a0 / - b and K2 is finite. (ii) The nonlinearity n(x) intersects the line through the origin of slope K1 at least once to ensure the existence of an equilibrium point. (iii) The slo p e o f n(x) at all the intersectio n p oints with the lines of slop e K1 a nd K2 sho uld b e either grea ter tha n the upp er b o und K1 o r less tha n the lo wer b o und K2 , thus ensuring the insta bility of the eq uilibria . From their studies using the describing function Genesio and Tesi [6.21] concluded tha t the nega tive feedback system with no nlinearity n(x) and linear dynamics G(s) might be expected to exhibit chaos if (iv) (v) The application of the DF method to the system gives a stable predicted limit cycle x*(t). There exists an equilibrium point E, different from the one tha t generates the predicted limit cycle, which has one eigenvalue in the open right half plane and the others in the left half plane. (vi) (vii) The abscissa xE belongs to the range of the predicted limit cycle, tha t is, x*(t) = xE for some time t. As the f iltering pro p erties of G (s) d eteriora te the m otio n m a y b e exp ected to cha ng e f ro m a lim it cycle to chaotic mo tion.
The investig atio ns in [6.22] co nsid ered, n(x), to b e a rela y a nd used the tra nsf er f unctio n f o rm of eq ua tio n (6.40) f or G (s). It was conjectured that for n > 2 the system was likely to possess a region of chaotic motion if (i) (ii) Equation (6.40) has at least one pair of complex conjugate roots with positive real part The Hurwitz sector lies within the nonlinearity sector
Based o n these assump tio ns 3 rd a nd 4th o rd er rela y systems with chao s were synthesized. For the third o rd er system with an ideal relay with unit levels and = 1, then the upper a nd lo wer bo unds of the Hurwitz sector are K1 = a0 a nd K2 = a 0 - a1 a2 . Satisfactio n of conditio n (iii) above requires 0 < K2 < K1 , which gives a0 > a1 a2 . Thus if the transfer function is chosen with a0 = 4 and a1 = 1.25 chaotic motion ma y exist for 0 < a2 < 3.2 .
112
The equilibria are stable for a2 > 3.2 , become neutrally stable with roots at -3.2 and 1.118j for a2 = 3.2 , and as a2 decrea ses the co m plex ro ots ha ve p ositive rea l pa rts. Using the prog ra m ba sed o n the Tsyp kin m etho d no limit cycle wa s found for a2 > 3.2 but several unstable limit cycle solutions were found as a2 was decreased below 3.2. In particular an almost sinusoidal uns table limit cycle was found together with several spiral type unstable limit cycles. When a2 was red uced to 2.8 o ne of the sp iral limit cycles b eca m e stable a nd is sho wn in Fig ure 6.9. Thi s m a y b e ref erred to as a 3 spiral beca use of the num b er of o scillatio ns. For 2.8 < a2 < 3.2 the va rio us unstab le limit cycles ha d diff erent num b ers of spirals apart from the almost sinusoidal one which bounded all the others. Simulatio ns starting with ini tial conditions inside this bounding limit cycle resulted in chaotic mo tion with the motio n jumping between different spiral type uns table limit cycles as shown in Figure 6.10.
Figure 6.9 Three spiral limit cycle (taken from reference 6.19)
Figure 6.10 Chaotic spiral type motion (taken from reference 19)
113
6.6.6 Example 6 Unstable Limit Cycles and Simulation

An interesting a sp ect of the a b o ve theory f o r find ing limit cycles in rela y systems is that b o th their wa vefo rm a nd stability can be accurately calculated. Stable limit cycles are normally not difficult to find in simula tion studies as they may be rea ched f ro m vario us initial co nditio ns, unstable limit cycles are, ho wever, m uch m o re difficult to lo ca te. Sin ce k no wledg e of the exa ct lim it cycle solutio n a llo ws initially co nditio ns to b e used in a sim ula tio n w hich lie o n a n unsta ble lim it cycle, ref erence 6.23 studied the a bility of a sim ula tio n to tra ck the lim it cycle. It wa s fo und tha t lo ng er tra cking wa s a chieved , as exp ected , b y using a sm aller step size a nd also if the m a g nitud e of the larg est eig envalue of the sta bility m atrix Q wa s not significantly greater than unity.
6.7 Forced oscillations

The o nly m o difica tio n req uired to b e ma d e to the m etho d in ord er to investig ate fo rced o scillatio ns is that the rela y inp ut b eco m es a co m b inatio n o f the f eed ba ck sig nal a nd the f orcing sig nal. Thus, if the rela y inp ut, x (t ) , is the erro r sig nal o f the feedback loop then x (t) = r (t) - c (t) . If a relay output waveform is assumed which switches when the input has a value of a t the switching tim e t then r (Dt ) - c (Dt ) = D a nd xo (t ) m ust ha ve the ap prop riate sig n. An exa m ple is given in ref erence 6.12 f or the theoretical solutio ns to a pplying a sinusoidal inp ut to a f eed ba ck system ha ving a rela y with dea d zo ne. The system g ain is such that it has a limit cycle with no inp ut. A sinusoida l inp ut with a f req uency o f ro ug hly three tim es that of the lim it cycle is a p plied to the system a nd its a m plitud e slo wly increa sed . Assum ing initially a rela y o utp ut with switching s a p pro pria te to the limit cycle the plo t of the rela y inp ut a s the inp ut a m plitud e increa ses co ntains m o re of the inp ut f req uency w hich eventually lead s to a solutio n with false switchings. It is then assum ed tha t the rela y o utp ut will ha ve two rather tha n o ne p ulse p er half cycle a nd a so lutio n co nfirm ed . Thus, em b ed ding the theory in sof tware with graphical o utp ut enables so me results to be obtained for forced respo nses both for systems which may be stable or possess a limit cycle in the auto nomous state.
6.8 Conclusions
This chapter has shown how an analysis based on a method originally formulated b y Tsypkin ca n be used to find the exa ct form of lim it cycles in rela y system s. In a d ditio n to eva luating the lim it cycles the m etho d a lso allo ws their sta bility to be fo und . The evalua tio n of the lim it cycles, ho wever, req uires investig atio n of the solutio n o f no nlinear alg eb raic eq uatio ns whose number increases with the number of pulses per period at the outp ut of the relay. This is not a problem if the ap proa ch is imb edd ed in co m p uter prog ra ms with intera ctive grap hics as d escribed a bo ve f or so m e interesting exa m ples studied using software written in FORTRAN. It is also possible to extend the method to cover systems with more tha n one relay as well as relay feedback systems with harmonic inputs.
6.9 References
6.1 Bohn, EK 1963, The transform Analysis of Linear Systems, Addison Wesley, New York
6.2 Ha m el, B 1950, Etud e Ma them a tiq ue des System es a plusieurs Deg rees d e Lib erte Decrits p a r d es E q uatio ns Linea ires avec un Terme de Commande Disco ntinu. Proc.Journees dEtude des Vibrations, AERA,Paris. 6.3 Tsyp kin, Ya Z 1984, R ela y Co ntrol System s Ca m b ridg e U niversity Press, Ca m b rid g e, E ngla nd (E ng li sh versio n of m uch earlier Russian edition).
114
6.4 Chung, JK-C & Atherton, DP 1966, The Determinatio n of Periodic Modes in Relay Systems Using the Sta te Space Approach, Inter national Journal of Control, Vol 4, pp 105-126. 6.5 Majhi, S & Atherton, DP 2000, On-Line Tuning of Controllers for An Unstable FOPDT Process, lEE Proceedings Control Theory and Applica tions, Vol. 147, No.4, pp. 421-427. 6.6 Atherton, DP 1966, Conditions for Periodicity in Control Systems Containing Several Relays, Paper 28E, 3rd IFAC Congress, London, 6.7 Atherton, DP 1982, Oscillations in Relay Systems, Trans InstMC, Vol 3, pp 171-184. 6.8 Wa d ey, M 1984, E xtensio ns of Tsypk in s Metho d f o r Co m p uter Aid ed Co ntrol System Desig n, DPhil Thesis U niversity of Sussex, England. 6.9 Willems, JL 1970, Stability Theory of Dyna mical Systems. pp 112-113. Nelson, London 6.10 Bala subra m a nia n, R 1981, Sta bility of lim it cycles in f eedb a ck system s co nta ining a rela y IE E Pro c. D Co ntrol Theo ry and Appl. Vol 128, pp24-29. 6.11 Wadey, MD & Atherto n, DP 1982, Limit Cycles in Relay Systems, American Control Conference, Arlingto n, Virginia, USA. 6.12 Atherton, DP & Wadey, MD 1982, Co mputer Aided Analysis and Design of Relay Systems. IFAC Symposium on CAD of MV Tech. Systems, Purdue, USA, pp 355-360. 6.13 Atherton, DP 1993, Analysis and Design of Relay Control Sys tems, CAD for Control Systems, Chapter 15, Marcel Dekker, pp 367-394. 6.14 Rao, UM & Atherto n, DP 1977, Stability Analysis of a Structurally Flexib le Spacecraft. 20th Mid west Symposium on Circuits and Systems, Lubbock, Texas, pp 652-656. 6.15 Rao, UM and Atherton, DP, M ulti-Pulse Oscillatio ns in Relay Systems, Proceedings 7th IFAC World Congress, Helsinki 1978, Vol 3, Pergamon Press, pp 1747-1754.
Trust and responsibility

NNE and Pharmaplan have joined forces to create NNE Pharmaplan, the worlds leading engineering and consultancy company focused entirely on the pharma and biotech industries. Ins Arizaga Esteva (Spain), 25 years old Education: Chemical Engineer You have to be proactiv e and open-minded as a newcomer and make it clear to your colleagues what you are able to cope. The pharmaceutical eld is new to me. But busy as they are, most of my colleagues nd the time to teach me, and they also trust me. Even though it was a bit hard at rst, I can feel over time that I am beginning to be taken seriously and that my contribution is appreciated.
NNE P harmaplan is t he w orlds leading enginee ring and consulta ncy company focused entire ly on the pharma and biotech indu stries. We employ more than 1500 people w orldw ide and offer global reach and lo cal know ledge along w ith our all-en compassing l ist of serv ices. nnep harmaplan.com
115
6.16 Ra o , U M, Atherto n, DP & Ba la subra m a nia n, R 1978, Sim ulatio n a nd Estim a tio n Stud ies o f a Sa tellite Attitud e Co ntrol System. lM ACS Symposium on Simula tion of Control Systems, Vienna, pp 215-221. 6.17 Moeini, A & Atherton, DP 1996, The Determina tion of Limit Cycles in Mode Dependent Relay Systems, IFAC Congress, San Francisco, Vol E, pp 43-48. 6.18 Atherto n, DP, McNa m a ra, 0P, Wa d ey, MD & G o ucem , A 1985, SU NS: the Sussex U niversity No nlinear Co ntrol Systems Software, 3rd IFAC Symposium on Computer Aided Design in Co ntrol Engineering Systems, Copenhagen, pp 173-178. 6.19 Atherton, DP 1990, Predicting the Auto nomous Behaviour of Simple Nonlinear Feedback Systems, IMSE -90 Conference, Arlingto n, Texas. Published in App. Mech. Review, Vol 43, October 1990, pp 251-260. 6.20 Choud hury, SK & Ather ton, DP 1974, Limit Cycles in Hig h-Order Nonlinear Systems. Proceedings lEE, Vol 121, pp 717-724. 6.21 G enesio , R & Tesi, A 1992, Ha rm o nic -Bala nce Metho d s f or the Ana lysis of Cha o tic Dyna m ics in No nlinea r-System s, Automatica, Vol 28, pp531-548. 6.22 Am ra ni, D & Atherto n, DP 1989, Desig ning Auto no m o us R ela y System s With Cha otic Mo tio n, 28th IE E E Co nf erence on Decision and Control, Florida, Vol 1, pp 512-517. 6.23 Wadey, MD & Atherto n, DP 1986, A Simulatio n Study of Unstable Limit Cycles, IFAC/IM ACS Symposium on Control Systems, Vienna, pp 213-218.
6.10 Appendix
The appendix presents formula from which the A loci of various transfer functions can be evaluated. This is done by putting any transfer function into partial fractions in terms of the simple transfer functio ns given in Table 6.1 and then using the results for their A loci given in Tables 6.2 and 6.3.
Table 6.1 Basic Transfer Function Types.
The results fo r tra nsf er f unctio ns with real p oles are def ined in term s of sine a nd co sine series a nd f o r the co m plex po les in terms of so m e p ara m eters. The results are given f or the A lo ci in Ta ble 6.2. The def initio ns of the sine a nd cosine series a nd the results fo r the tra nsf er f unctio ns are g iven in Ta bles 6.2(a ) a nd 6.2(b ). The pa ra m eters fo r eva luating the co m plex transfer functions are given in Table 6.2(c). Similar results for the A loci are given in Tables 6.3, 6.3(a), 6.3(b) and 6.3(c), respectively.
116
Type No. 1 2 3 4 5 1/sT
Transfer Function 1/ (1 + sT) 1/sT (1 + sT) 1/s T

2 2 2
Type No. 6 7 8 9 10 1/s T

3 3
Transfer Function sT/ (1 + sT) 1/ (1 + sT)

2 2 2 3 2 2 2
1/ (s + 2gs~o + ~o) ~o / (s + 2gs~o + ~o)
1/ (1 + sT)
Table 6.2 Results for A loci
117
Table 6.2(a) Results for Sine Series
118
Table 6.2(b) Results for Cosine Series
119
Table 6.2(c) Parameters for Complex Pole Transfer Functions
120
Table 6.3 Results for A loci
121
Table 6.3(a) Results for Sine Series
122
Table 6.3(b) Results for Cosine Series
123
Table 6.3(c) Parameters for Complex Pole Transfer Functions
124
Controller Tuning from Relay Produced Limit Cycles
7 Controller Tuning from Relay Produced Limit Cycles

7.1 Introduction
Control theory is typically taught from the starting point of ha ving a mathematical model of a plant, most often linear, and a controller is then designed to provide a specified performance. Techniques usually regarded as classical control consider simple forms of controller, such as phase lead or PID and the problem then becomes, once the specific form of controller has been chosen, to try and find parameters for the controller so that the system performance meets the sp ecif ica tio ns. Whe n w o rk ing with a ma them atical m o d el of the pla nt m a ny m ethod s of classical co ntrol ca n b e used to find the controller parameters. Dependent on the specificatio ns there may be no solution, a unique solution, or many solutions to the problem. The latter, for example, could be the case if the only specifica tion was on the response to a step input and required an o versho o t less tha n 10% a nd a relatively slo w settling tim e. On the o ther ha nd if it wa s req uired to minimise the integral squared error for the step inp ut then there wo uld be a uniq ue o r, possibly no feasible, solutio n. The most com mon adjustable parameter controller is the PID, with the three terms P- proportional, I integral and D derivative. It has been used in the process industry since the 1940s first in pneumatic, then successively in vacuum tube, transistor, integrated circuit, a nd no w usually in micro pro cesso r f o rm ; a n im p lem entatio n that allo ws sig nifica ntly m o re f unctio nality to be includ ed .
125
When co m missio ning or exa mining co ntrol loo ps engineers or p la nt op erators often f o und that with sug gested o r existing PID p ara m eter setting s the lo o p p erf o rm a nce wa s not satisf a ctory. Thus the o b vio us q uestio n arose, what co uld o ne d o on the job to get better parameter settings. Enter Ziegler and Nichols [7.1] who in their 1948 paper suggested two, in principle, simple approaches:1. Perform an open loop step response on the plant. From this response estimate the parameters of a first order plus dead time (FOPDT) model. Then set the controller parameters using given formulae involving the estimated plant parameters 2. With the plant under closed loop control put the co ntroller into the P mode. Adjust the gain until an oscillation results, measure its frequency and record the P gain. Then set the co ntroller parameters using given formulae involving the oscillatio n frequency and the P gain. They ref erred to their a p pro a ches a s p ara m eter tuning , which seem s log ica l a s wha t wa s d o ne, po ssibly b y a technicia n, was perform a system test a nd from the readings calculate the co ntroller parameters according to the given formula e, q uite a sta nd ard eng ineering ap p ro a ch. To d a y ma ny p a p ers ref er to p ara m eter tuning whe n the values are calculated starting fro m a m a them atical m o del of the p la nt, wh ich d o es not seem the correct terminolo g y, a s o ne never sees the term used for obtaining the parameters of other fixed controllers, such as lead or lag controllers, by this approach. The Z-N methods consist of two parts, which is not often clearly stated, namely a testing identificatio n procedure and then the use of their tabulated formula to set the parameters. The identification procedures in the tests are: 1. An estima ted FOPDT model from the step response. This was chosen because it was judged to be a reasonable model for many simple processes, although other models could be fitted to the response if felt more appropriate. It should be noted that in the early days of control the standard experimental identificatio n tests were step and/or frequency responses. 2. An estima te of the so-called critical point (or gain margin point) of the plant, namely the gain and frequency value, c, when the plant gives a phase change of -180. This, is of course, a small bit of informatio n about the plant, and the major contribution of Z -N was to recognise the value of this informatio n. Their suggested co ntroller pa rameters based o n the measurements from both tests were b ased o n desig ns to give a relatively high overshoot, around 25%, in the closed loop step response. Given the identificatio n result of 1, na m ely a n FOPDT pla nt m od el, then there are b asically a n infinite num b er of a na lytical ap p ro a ches which ca n b e used to calculate PID co ntroller p ara m eters. The p oint is tha t the a pp roa ch used sho uld m eet sp ecified perf orm a nce criteria which a re ap plicatio n d epend ent a nd m a y, of co urse involve co nsid eratio ns rela ted to inp ut step a nd disturb a nce resp o nses, no ise, no nlinea rity a nd o ther f a ctors not usually m entio ned in sim ple a p proa ches. If the sp ecificatio ns ad d ress a desired clo sed lo op step resp o nse b eha vio ur, as m a ny d o , then a n op en loo p f req uency d o main
approach is indirect and in general will require iteration to achieve the desired closed loop performance.
126
Both the Z-N identificatio n m ethods ha ve problems in pra ctice. In the first test noise in the measurement affects the accuracy of getting the FOPDT model a nd it cannot, of co urse, be used for an open loop unstable plant. If used as intend ed, ho wever, for pla nts in which o ne ha s rea so na ble co nfid ence ha ve essentially the assum ed d y na m ic b eha vio ur, a nd no t a ny bla ck b o x m o d el, then the id entified p a ra m eters no rm al p ro vid e a sa tisfa ctory m o d el f or a nalytical d esig n m etho d s [7.2]. The seco nd test is extrem ely d ifficult to im p lem ent in pra ctice as m a ny pro cess pla nts ha ve slo w d yna m ic s, so the rea ctio n to a cha ng e in lo op g ain is slo w. Further the p rinciple is b ased o n linear d yna m ics a nd a ctuators inva riably ha ve d ead zo nes a nd ind eed witho ut saturatio n of so m e f o rm da ng ero us a m plitude o scillatio ns ca n result fro m too large a n increase in the P g ain. It is, of co urse, a test which ca n b e p erf o rm ed id eally in sim ula tio n using a m o d ern d ig ital sim ula tio n la ng ua g e, althoug h it was not easy on an analogue computer simulatio n and can be, as mentioned, a nightmare in practice. R eco g nising these p ra ctical difficulties with the test, Astro m a nd Hag glund [7.3] sug g ested rep la cing the P g ain co ntrol by a n id eal rela y, which b eca m e sim ple to im plem ent with the m o ve to micro p ro cessor co ntrollers in the 80 s. The q uestio n then arises, having replaced the P control by a relay, as to how one uses the resulting limit cycle to obtain informatio n about the plant. The earlier chapters provide the information for doing this.
7.2 Knowledge from the Limit Cycle

This ca n b e o btained b y b asing the a na lysis either o n use of the DF method o r the exa ct m etho d f or limit cycles co vered in the p revio us cha pter. The o riginal p rop o sal [7.3], usually k no wn a s rela y a uto tuning was b a sed o n the DF m etho d a n d like the original Z-N work assumed the plant was unkno wn but typically having a transfer functio n form common for the pro cess ind ustries. R ecently other a pp ro a ches ha ve b een sug gested where b y a ssum ing a specific m o del f orm fo r the plant its parameters are estimated from measurements on the limit cycle.
7.2.1 Relay Autotuning

For a negative feedback loop containing a plant with transfer function, G(s), and relay, use of DF analysis gives N (a) G (j~) = - 1 which can be written N (a) G (j~) = 1 and ArgN (a) + ArgG (j~) = - 180c For an on-off relay of height h and hysteresis then (7.3) (7.2) (7.1)
ar Thus when the nonlinearity is an ideal relay and denoting the limit cycle frequency by o, one has ArgG (j~o) = - 180c and G (j ~ o ) = ar/4h 127
(7.4)
N (a) = 4h e- ji where i = sin- 1 (D/a)
An Introduction to Nonlinearity in Control Systems This shows that the DF analysis yields ~c = ~o and G (j~c) = ar/4h = 1/Kc
(7.5)
for the frequency and mag nitude of the critical point, where Kc is the critical gain. Due to the approximation of the DF method this result is approximate and is normally more accurate the nearer the limit cycle is to being sinusoidal. It sho uld also be noted tha t a is not the amplitud e of the limit cycle at the rela y input but the amplitud e of its fundam ental. Most practical implementatio ns measure the peak to peak limit cycle amplitude and take a to be one half of this. It is po ssible to includ e a d ditio nal co m p utatio n softwa re in the co ntroller to ob tain a b etter estim ate of the f und a m ental a b ut in view of the ina ccura cies in measurem ent it is p ro ba bly not justifia ble in m a ny insta nces.. O n the other ha nd so m etim es easy corrections are possible [7.4] for example if the limit cycle is almost triangular, as it is for an FOPDT plant with sma ll tim e d ela y, then o ne ca n calculate the f und a m ental f ro m the p eak value. Also the theory gives a b etter p ercenta g e accuracy for the frequency tha n the amplitude [7.5]. If time is no object in performing limit cycle experiments then one can:(1) normally get more accurate measurements by averaging over several cycles. (2) vary the hysteresis in the relay to get additional results (3) include kno wn extra linear elements in the loop to get additional results (4) get improved results from the DF method by making the limit cycle nearer to a sinusoid with the inclusio n of a known tuned filter.
128
Typically (2) a nd (3) ca n b e used to g et ad ditio nal p oints [7.6] to that of the critica l point o n the pla nt freq uency resp o ns e a nd iterative tuning of a reso na nt filter [7.6] ca n a chieve (4) b y p rod ucing a p ure sinusoid al lim it cycle a t the pla nt critical frequency, which then yields the exact value of the critical point apart from, of course, measurement errors. The tuned filter wa s intro d uced in [7.6] in a p ap er co ncerned with the d esig n of no nlinea r PID co ntrollers f or no nlinea r pla nts, where to g et b etter inf orm a tio n o n the pla nt freq uency resp o nse a sinuso idal inp ut wa s req uired . A big a d va nta g e of the rela y a utotuning method is tha t the a m plitud e of the limit cycle ca n b e co ntrolled b y the setting of the rela y o utp ut levels, a n d this a voids one of the pra ctical problem s of the second method of Z-N. Further o ne can use this feature to estimate whether the pla nt b eha ves rea so na bly linea rly a nd if no t use the results to d esig n no nlinea r PID terms as d iscussed in sectio n 9.5.
7.2.2 Plant Parameter Identification

It has b een sho wn ho w the limit cycle in a lo op with a k no wn pla nt tra nsfer f unctio n und er rela y co ntro l ca n be ca lcula ted exa ctly using the Tsypkin m ethod o r estim ated using the DF m etho d. In pa rticular fo r the ca se of a n od d sym m etrical limit cycle the first m etho d allo ws exa ct d etermina tio n of the limit cycle a m plitude a nd f req uency at the rela y inp ut in term s of the pla nt p ara m eters. Thus, the m eth o d ca n also b e used f or p la nt tra nsf er f unctio n id entificatio n since m easurem ent of the a mplitud e a nd f req uency of the limit cycle will allo w two unk no wn pla nt pa ra m eters to b e f o und fro m the eq uatio ns, for example the time delay and time consta nt in an FOPDT model when the plant d.c. gain is kno wn. Analysis for an asymmetrical limit cycle and measurement of more parameters in the resulting limit cycle enables additional plant parameters to be calculated, for example four parameters in an SOPDT model. Several papers [7.7, 7.8] describe the identification of plant parameters based on an assumed plant transfer function form, using exact limit cycle analysis and limit cycle measurements. From a practical viewpoint they have the following possible disadvantages. (1) The microprocessor controller software, as well as being able to measure various limit cycle parameters, must be able to solve the appropriate nonlinear algebraic equations to get the plant parameters. Further it must contain a design procedure for calculating the controller parameters for the known plant. (2) There will be errors in the measured limit cycle parameters, due to noise etc, and these may cause problems in solving the nonlinear algebraic equations, too large errors for example can result in no or physically unallo wable solutio ns. It is also p o ssible to a ssum e a pla nt tra nsf er f unctio n a nd d etermine expressio ns f or so m e limit cycle p ara m eters using DF a na lysis rather tha n the exa ct a pp ro a ch. Again o ne ca n then use the eq ua tio ns in reverse a nd calcula te pla nt p ara m eters fro m limit cycle m ea surements. If this is do ne then d ue to the ap pro xim atio n of the DF m etho d, pa rticula rly when dealing with asymmetrical limit cycles, the problem cited in (2) is normally worse.
129
7.3 Tuning the Controller

Ma ny d esig n m etho ds m a y be used to o btain co ntroller p ara m eters to m eet given system sp ecif ica tio ns when the pla n t tra nsf er f unctio n ha s b een estim a ted b y o ne of the m etho d s in sectio n 7.2.2. Bo th cla ssica l a nd m o d ern m etho ds will often turn out to require an iterative approach as they will invariably no t address the required specifications directly but by inference. For example, if the design specificatio n is with respect to the inp ut step response, then if one uses a freq uency d o m ain d esig n o ne will b e relying o n a n im p recise link , sa y b etween p ha se m a rgin a nd o versho ot in the step resp o nse, o r if o ne uses pole pla cem ent the eff ect of a ny clo sed lo op zeros o r neglected po les o n the exp ected respo nse will have to be checked. The a ctua l system b eha vio ur will diff er f ro m the theo retical calculatio ns d ue to the m o d elling erro rs of the id entifica tio n a nd this, assuming the pla nt b eha ves reaso na bly linea rly, ca n b e ta ken into a cco unt b y co nsid ering p ara m eter uncertainty. Muc h is m a d e a b o ut the co ncep ts of ro b ust co ntrol f or such situatio ns b ut no results a d dress the real d esig n pro blem , namely if the plant parameters vary over a certain range will the specifica tions, which normally relate to responses, not sta bility, still b e m et. Fortunately to da y MATLA B is so fa st tha t the ro b ustness of a co ntroller d esig n ca n easily b e check ed simply by doing a set of Monte Carlo simulations over the expected range of variations in a few parameters. In fact a go o d d esig ner m a y b e a ble to id entif y the wo rst ca se , so that the d esig n just m eets the sp ecifica tio n fo r this ca se a nd is better for other parameter values. On the other hand if one uses the limit cycle test to estimate the critical point of the plant transfer function, two issues arise namely how accurate is the critical point estima te and secondly how, kno wing just this bit of information, can the controller parameters be found so that the system meets the desired specifications. The first point has largely been ad d ressed in the previo us sectio ns a nd to so m e extent ho w error s in the critical p oint value m a y aff ect the sub seq uen t desig n ca n be exa mined b y co nsid ering in sim ulatio n ho w the p erfo rma nce diff ers when selecting a f ew slig htly diff eren t values for the critical point. The seco nd p oint is m uch m o re difficult, since e ven if a sim ple step resp o nse sp ecificatio n is given, o ne ca nnot assert fo r given PID p ara m eters that if a pla nt with a given critical point ha s a closed loo p step resp o nse with a given p ercenta g e overshoot then all plants with the same critical point will have the same overshoot. Certainly the closeness of the open loop frequency response of the compensated system to the (-1, j0) point has a significant effect on the form of the closed loop step response but the shape of the response in this region can differ significantly for different plant model transfer functions with the same critical point and the same compensator. It is therefore desirable in using critical point inf orm a tio n to ha ve so m e id ea of the f orm of the pla nt tra nsf er f unctio n. Ha ving ob ta ined the critica l p oint all that ca n b e d o ne in a co ntroller d esig n alg orithm is to select the co ntroller p a ra m eters so tha t, sub ject to realisa bility co nstraints, the compensated frequency response has the critical frequency at a desired point in the Nyq uist plane. This po int ca n b e m o ved aro und , within a finite regio n, b y a djusting the co ntroller pa ra meters. For the ideal PID
co ntroller K61 + sTi + (1/sTd @, Z-N chose to ta ke Ti = 4Td and it is easily shown tha t their choice for Td and K places the critical frequency o n the compensated lo cus at a gain of 0.66 a nd p hase a ngle of -155. The corresponding point for their suggested PI p ara m eters is 0.46 a t a p ha se a ngle o f -191. These choices, a s fo r their f irst m etho d , tend to pro d uce clo se d lo op step responses with a relatively hig h overshoot.
130
Thus the success of PID co ntroller design based o n critical point informatio n is o ne of k nowing where the critical frequency sho uld b e o n the co m p ensated lo cus to, ho p ef ully, m eet the d esig n specifica tio ns. Given so m e id ea of the p la nt tra nsf er functio n it is relatively straig htfo rward to co me up with go o d strategies. Fo r exa m ple, in ref erence 7.9 op tim um pa ra m eters are f o und f o r PID co ntrollers to m inim ise integral of erro r p erf orm a nce ind ices f o r a step inp ut to a clo sed lo op with an FOPDT plant. These are given for minimization of the ISE, ISTE and IST 2E. Typically as the time weig hting increases the o versho o t d ecrea ses, the rise tim e is slo wer a nd the settling tim e increases slig htly. The lo catio n of the critical freq uenc y o n the co m p ensated Nyq uist lo cus ha s also b een calcula ted f o r these d esig ns so that critical f req uenc y shifting d esig ns can be used for other similar plant transfer functions. Minimiza tion of the ITSE typically leads to a design with around 10% overshoot. Many different design approaches are possible for deciding where to place the critical frequency. Also, depending on the application different requirement specificatio ns may have to be met. One m ig ht b e tem pted to sa y tha t k no wing a bit of inf o rma tio n a b o ut a pla nt tra nsf er f unctio n, na m ely its critical p oint, cannot be anything like as good as having an experimentally measured open loop plant step response, in principle a multiplicity of points. The problem is that measured step response results are not ex act with large amounts of noise often present in industrial measured data. To illustrate a point consider the two step responses in Figure 7.1, curve (a) is for a chosen transfer functio n and curve (b) for a transfer function derived from it by a model reduction technique. On this basis, particularly if the response of (a) had resulted from measurements contaminated by noise, one would ha ve clearly co nclud ed tha t the tra nsf er f unctio n of curve (b ) wo uld b e a go o d m o del to use fo r co ntroller d esig n f or th e original transfer function (a). Now go to Figure 7.2 and look at the corresponding frequency responses. It can be seen tha t o ur p revio us co nclusio n is clea rly no t valid a s the tra nsf er f unctio n with the m od el (a) ha s a finite a nd m o d el (b ) a n infinite gain margin. Thus a poor design for say, a constant gain controller, would result from using model (b), one in fa ct which wo uld p rob a bly result in a n unsta ble system with the chosen tra nsf er f unctio n. The two tra nsf er f unctio ns a re 5.435 13s + 10 , respectively. s + 1.739s + 0.5435 (s + 1) (0.5s + 1) (s2 + 3s + 1)
Figure 7.1 Step responses of original and reduced models.
131
and
Figure 7.2 Frequency responses of original and reduced models
Sharp Minds - Bright Ideas!

E m pl oy e e s at F OS S A na l y ti ca l A /S a r e l iv i ng pr o of of t he co m pa ny v a l ue - F i rst - usi ng ne w i nv e nt i ons t o ma k e de di ca t e d sol ut i ons f or our cust o me r s. Wi t h sha r p mi n ds a nd cr oss f un ct i ona l t e a m w or k , we co nst a nt ly st r i ve t o de v e l o p ne w uni q ue pr od uct s W oul d y o u li k e t o j oi n our t ea m? F OS S w or k s di li ge nt l y wi t h i nnov a t i on a n d de v e l op me nt a s ba si s f or i t s gr o wt h. It i s re f le ct e d i n t he f a ct t ha t m or e t ha n 200 of t he 1200 e m pl oy e e s i n F OS S wor k wi t h Re se ar ch & De v e l op me nt i n S ca n di na v i a a nd US A . E n gi ne e r s a t F OS S wor k i n pr od uct i on, de v e l op me nt a nd ma r k e t i ng, wi t hi n a wi de r a nge of di ff e re nt f i el ds, i. e. Che mi st r y, El e ct r oni cs, M e cha ni cs, S of t wa r e , O pt i cs, M i cr obi ol ogy , Che m o me t r i cs. We offer A ch al l en gi n g j ob i n a n in t er na ti o na l a nd i nn o va ti v e co m p an y t ha t i s l ea di ng i n i t s f i el d. Y ou wi ll g e t t h e opportunity to work with the most advanced technology together with highly skilled colleagues. R ea d m or e a bo ut F O S S at w w w .f os s. d k - or g o d ir e c tl y t o ou r s tu d en t si t e w w w. f os s.d k /s ha r p m in ds w h e r e you can learn more about your possibilities of working together with us on projects, your thesis etc.
The Fa mily o w ne d F O SS gr ou p is the w orl d l ea der a s s u pp lier of de di cate d, hi g h-te ch ana lyti cal sol uti o ns w hi c h mea su re a n d co ntr ol t he q ualit y a nd p ro d uc tio n of a gri cu ltur al, f oo d, p harma ce uti cal an d c he mi cal pr o du cts. M ai n a ctiv ities are i niti ate d fro m De n mark, S we de n a n d USA wit h hea d q uarter s d o m ici le d i n Hil ler d, D K. T he pr o du cts are mar kete d gl o bally b y 2 3 sa les co m p ani es a n d a n e xten sive net of distr ib ut ors. I n li ne w ith the co reval ue t o be First , th e co m p any int en d s t o ex pa n d its mark et po siti o n.
Dedicated Analytical Solutions

FOSS Slangerupgade 69 3400 Hillerd Tel. +45 70103370
www.foss.dk
132
7.3.1 An Example
Consider for example a pla nt with the transfer functio n G (s) = K p e - 0.5s (s + 1) 2 with Kp = 1.
The objective is to sho w that go od tuning ca n be o btained fo r a PID co ntroller ba sed o n a n a uto tuning test. Using a rela y of unit height the measured values of the amplitude of the limit cycle, A, and frequency, ~0 , from a simulatio n with no noise introduced were A = 0.287 and ~0 = 1.90 . Using equation (7.5) with a = A gives ~c = 1.90 and Kc = 4.44 which co m pa re with the a ctual va lues of 1.92 ra ds/s a nd 4.69 f o r the k no wn pla nt tra nsf er f unctio n. The norm alized g ain l = K p Kc and assuming K p = 1 has been found by other measurements l = 4.44 . The tuning formula, suggested in ref erence 8, which gives op tim um p ara m eters f or ISTE tuning (minimizatio n o f the integral sq ua red o f tim e m ultiplied by for these tuning parameters in the controller is shown in Fig ure 7.3 (response (1)) together with that for controller parameters of K = 2.42, Ti = 1.99 and Td = 0.63 obtained by minimizatio n of the ISTE criterio n (response (2)) for this kno wn plant transfer functio n. The results are very similar considering the approximations involved and the simplicity of the auto tuning approach.
error) of a FOPDT plant based on l, Kc and ~c , gives K = 2.55, Ti = 2.61 and Td = 0.40 . The unit step set point response
Figure 7.3 Step response comparison for autotuning (1) and IST E minimization (2)
7.4 Autotuning using the Relay in Parallel with the Controller

The sta nd ard a uto tuning a p pro a ch is to repla ce the co ntroller b y a rela y so tha t the f orwa rd p a th o f the f eed b a ck lo o p co ntains a rela y a nd the p la nt. Further if o ne wishes to check certain aspects of the co mp ensa ted system , fo r exa m ple find its g ain ma rgin, then this ca n be do ne b y pla cing the rela y in series with the co ntroller a nd p la nt [7.10]. These ap proa ches involve wha t mig ht b e called sepa rate tuning exp erim ents b ut in so m e cases it m a y b e a pp rop riate to try a nd tune the system whilst it is still operating normally. This can be done by placing the relay in parallel with the controller [7.11]. This procedure was first suggested in reference 7.12, although in the paper the approach was used for plant parameter identificatio n rather tha n critical point tuning.
133
When the relay is placed in parallel with the controller the feedback loop is as shown in Figure 7.4. To determine a limit cycle b y DF a nalysis o ne sim ply op ens the lo op a t the rela y a nd calculates the tra nsf er f unctio n, -Ge , b etween e o, the rela y output, and e i, the relay input, in the figure.
Figure 7.4 Relay in Parallel with the Controller
The basic equations (7.1-7.2) still apply but with G replaced by G e. It is easily shown tha t Ge = G 1 + G Gc (7.6)
which can be written Ge = 1 # Gc G Gc 1 + G Gc = 1 # Tcl Gc (7.7)
where Tcl is the closed loop transfer functio n. Now consider the Nyq uist diagram of GGc , and denote its magnitude at the freq uency o f the lim it cycle, o , b y M a nd tha t of 1 + G G c , b y t M , as sho w n in Fig ure 7.5. The m a g nitud e eq ua tio n in (7.2) can no w be written M = ar which gives t = 1 Gc tM 4h and the correspo nding phase eq uation is z = 180c + Arg (1/Gc) where - z is the closed loop phase angle shown in Figure 7.4. (7.9) 4h ar Gc (7.8)
134
Figure 7.5 Nyquist diagram and triangle for calculations.
Since the transfer functio n in the controller is kno wn |Gc| and Arg Gc are known at the frequency, o, so that equations (7.8) a nd (7.9) ca n b e solved , resp ectively, f or , as a is k no wn f ro m the lim it cycle m easurem ent, a nd z . Using the co sine rule with the a ngle z f o r the tria ngle, since its sid es a re 1, M a nd M , the value of M ca n b e fo und . Arg (G G c ) ca n also b e found from the tria ngle so that both the magnitude a nd p hase of the frequency respo nse of G at the freq uency , o , can be calculated . This will o nly b e the critical f req uenc y fo r G if Arg G c is zero , that is if the test ca n b e d o ne with the co ntroller in the proportional mode. However, if o is near to the critical frequency then settings based on the critical frequency should produce satisfactory results. Alternatively, other theories can be produced for finding controller parameters with various types of plant for other frequencies on the plant frequency response locus.
+LZPNU `V\Y V^U M\[\YL H[ 4(5 +PLZLS

^^^ T HUKPLZLS JVT
135
7.4.1 An Example
As a n example results from sim ulations of the FOPDT plant with tra nsfer f unction G (s) = e K (1 + 2sT d) 2 in parallel with the PID controller having T = 4Td , so that Gc (s) = , are given. 4sTd
- 0.5s
s+1
and a relay with h = 1
The critical f req uency a nd g ain of this p la nt are, respectively, 3.67 rad s/s a nd 3.81 a nd fo r the sim ula tio n the pa ra m eters of the controller were set at the Z-N values of K = 2.284 and Td = 0.214. For the input step response this results in an o vershoo t of 52.5%. The a m plitud e a nd f req uency of the limit cycle at the inp ut to the co ntroller, which is sho wn in Fig ure 7.6, were measured as 0.680 and 4.30rads/s.
Figure 7.6 Plot of limit cycle
For the controller at = 4.30, Gc (j~) = 2.72 and ArgG c (j~) = 33.0c giving = 0.688 and = 147 from eqns. (7.8) and (7.9), respectively. Using the cosine formula for the triangle of Fig ure 7.5 gives M the magnitude of GGc at = 3.40 as 0.617, and applying the formula again gives , the angle below the real axis of M equal to 13.4, so that ArgGG c = -166.6. Using the kno wn values for the mag nitud e a nd arg ument of G c, then gives the mag nitud e a nd arg ument of G as 0.227 and -199.6. Despite the fact tha t the waveform is not a good sinusoid these value s are within 1% of the calcula ted values f ro m the tra nsf er f unctio n. The p rob lem no w b eco m es ho w sho uld this inf orm a tio n b e used to set th e controller parameters? Here use is made of the ISTE critical point design procedure of reference 7.8. For this plant the critical point needs to b e m o ved to 0.58 a t -153, which m ea ns the PID g ives a p ha se lea d o f 27. Thus f or a hig her f req uency it will g ive m o re p hase lead , so that a reaso nable sug gestio n f or the freq uency 3.40 is to increase the shift b y aro und 25% to 35. This gives a p ha se of a pp ro xim ately -165 (199.6 -35) f o r this f req uency o n the co m pensated freq uency resp o nse lo cus. Since this p oint is nearer (-1, 0) tha n the critical freq uency p o int b y 12 a slig ht red uctio n in the m a g nitud e b y a ro und 10 % to th e value of 0.53 is taken. Using the same Ti/Td value as for the critical point shift design of reference 7.8, the required PID pa ra m eters a re K = 1.94, T d = 0.205 a nd T i = 1.152, resp ectively. The inp ut step resp o nse f o r this d esig n is sho wn in Fig ure 7.7 and as expected is compatible with an ISTE design.
136
Figure 7.7 Closed loop step response
7.5 Conclusions
This chapter has shown how informatio n can be obtained abo ut the linear d ynamics of a feedba ck loop from mea surements on limit cycles prod uced b y adding a rela y. In particular, the focus has b een placed on using the results to set the pa rameters of a PID controller, however the concep t can be used for deciding on appropriate parameters for other fixed structure co ntrollers. With m o d ern m icro pro cesso r co ntrollers the a pp roa ch is relatively ea sy a nd p ro vid es a go o d m etho d f or the a uto m atic tuning of co ntro ller p a ra m eters witho ut the need f o r m a them atical m od elling of the pla nt. Care, ho wever, ha s to be taken to ensure that safe plant operation will be possible under relay control.
7.6 References
7.1 Ziegler, JG & Nichols, NB 1942, Optimum Settings for Auto matic Controllers, Trans ASME, Vol 64, pp 759 -768. 7.2 Atherton, DP, 2009 Control Engineering Boo kboon publications at www. bookboon.co m 7.3 str m , KJ & Hg glund , T 1984, Auto m atic Tuning of Sim ple R eg ula to rs with Specificatio ns o n Pha se a nd Am plitud e Margins, Automa tica, Vol 20, pp 645-651. 7.4 Atherton, DP 1997, Improving the Accuracy of Autotuning Parameter Estimates. Proce edings, IEEE Interna tional Conference on Control Applications, Hartford, pp 51-56. 7.5 Po po v, E P & Pal to v, N P 1963, Ap p ro xim ate Metho d s fo r Analyzing No nlinea r Auto m atic System , E ng lish tra nslatio n, Wright Pa tterson Air Force Base translation services. 7.6 Atherton, DP, Beno uarets, M & Nanka-Br uce, O 1993, Design of Nonlinear PID Controllers for Nonlinear Plants, Proceedings IFAC World Congress 93, Sydney, Vol 3, pp 355-358. 7.7 Majhi, S & Atherton, DP 1999, Autotuning and Controller Design for Processes with Small Time Delays. Proc IEE Control Theory and Applica tions, Vol 146, pp 415-425. 7.8 Kaya, I & Ather ton, DP 2001, Parameter Estimatio n from Relay Autotuning with Asymmetric Limit Cycle Data, J. of Process Control, Vol 11, pp 429-439. 7.9 Zhuang, M & Atherto n, DP 1993, Automatic Tuning of Optimum PID Co ntrollers. Proc.IEE Control Theory and Applicatio ns, Vol 140, pp 216-224. 7.10 Astrom, KJ 1996, Tuning and Adaptatio n, Plenary lecture, IFAC Co ngress, San Francisco. 7.11 Atherton, DP 2006, Relay Auto tuning: An Overview and Alternative Approach, Ind. Eng. Chem. Res. Vol 45, pp. 4075-4080. 7.12 Majhi, S & Litz, L 2004, Relay Based Closed-loop Tuning of PID Controllers, Automa tisierungstechnik, Vol.52, pp.202-20.
137
Absolute Stability Results
8 Absolute Stability Results

8.1 Introduction
It has b een seen in previo us chapters that the beha viour of the no nlinear feedba ck loop of Figure 3.1 a nd indeed , its stability, m a y be aff ected b y the inp ut. This ha s theref o re led to two d efinitio ns of sta bility fo r a no nlinear system , the f irst stability of the auto nomous system, that is with r(t) = 0 in Figure 3.1, which is often referred to as stability of equilibrium and to which most early work was directed. The second form of stability is known as bounded -input bounded-o utp ut (BIBO) the sa m e f or a linea r system ; f or which necessary a nd sufficient co nditio ns f or stability ha ve b een k no wn f or m a ny years. On the other hand the general nonlinear problem for the stability of equilibrium remains unsolved. Indeed until the d evelo p m ent of f req uenc y resp o nse b a sed criteria in the 1960s [8.1] the o nly a vaila ble results were tho se fro m the wo rk of Lya p uno v which were no t ea sily usa ble b y a n engineer. The f req uency respo nse b ased criter ia , which will b e presented , only give sufficient conditions for stability and thus often produce very conservative results for a specific nonlinear that the results are very robust because they ensure stability for a large number of nonlinearities which satisfy certain conditio ns. This, of course, means that the system can be proven to remain stable if the conditions are not violated by any perturbations to the nonlinearity. The primary concern of this chap ter is to cover frequency response absolute stability results as they allo w comparisons to be made with both the DF method and limit cycle results for relay systems. In the next section Lyapuno vs metho d is briefly covered as some of the original frequency domain results were obtained using it whilst others were obtained using functional analysis.
sta bility a nd req uires the o utp ut to b e b o und ed f or a b o und ed -inp ut sig na l. These two f orm s of sta bility are, o f co urse,
characteristic as they only use limited information on the no nlinearity. On the other hand a mathematician may argue
8.2 Lyapunovs Method

The direct or seco nd method of Lyap uno v for d etermining the stability of a n a uto nomo us system is a timedomain method using a state space description. The problem is to determine the stability of an equilibrium state (singular point), which without loss of generality may be assumed to be the origin of thestate space, for the system xo = f (x) Before considering Lyapunovs method some preliminary topics must be considered. (8.1)
138
8.2.1 Definitions of Stability

Definition 1 The equilibrium state 0, or the equilibrium solution x(t) = 0, is called stable if for any given positive , there exists a positive , such tha t x (t0) 1 d implies x (t) 1 f for all t 2 t0 , where denotes the Euclidian norm. It should be noted tha t a system exhibiting a limit cycle or solutions for which x (t) increases temporarily will be considered stable by this definitio n ,when and are appropriately chosen, but it rules out solutio ns which grow witho ut bound. For a nonlinear system stability may only exist in some domain x 1 R of the state space. Definition 2 The equilibrium state 0 is called asymptotically stable if it is stable and x (t) " 0 as t " Definition 3 The equilibrium state 0 is called asymptotically stable in the large, or globally asymptotically stable, if it is asymptotica lly stable and the domain R covers the whole sta te space.
Student Discounts
Student Events
Money Saving Advice
Happy Days!
2009
139
8.2.2 Positive Definite Functions

Let V(x) b e a tim e inva ria nt scala r f unctio n of the vector x a nd S b e a clo sed bo und ed regio n in the x spa ce co nta ining the origin. Definition 4 The functio n V(x) is positive definite (semi-definite) in S if, for all x in S (i) V(x) has continuous partial derivatives with respect to the components of the vector x (ii) V(0) = 0 (iii) V(x) > 0 for x 0 (or V(x) 0 ). If the ineq uality sig ns in (iii) a re reversed then the co nditio ns f or a neg ative d efinite (or semi -definite) f unctio n are obtained . Thus, f or exa m ple, in two sp a ce if V(x) = x 12 + x 22 then it is p ositive d efinite a s it is o nly zero a t the origin, b ut the f unctio n V(x) = (x1 + x2) 2 is only positive semi-definite as it is zero everywhere along the line x1 + x2 = 0 . A commo n quadratic form scalar functio n is
T
(8.2)
where the matrix Q, which is nxn for x an n vector, is chosen to be symmetric. In this case it is known that a necessary a nd sufficient co nditio n f or the q ua d ra tic f orm of (8.2) to b e p o sitive d efinite is that the d eterm ina nts of a ll the principle leading minors of Q be positive. For it to be positive semi-definite all the principle minors, not just the leading ones, must be non-negative.
8.2.3 Lyapunovs Theorems

Theorem 1 The null solutio n, or equilibrium state of the origin, of the system of equation (8.1) is (asymptotically) stable if in some neig hbo urho od of the origin there exists a po sitive d efinite f unctio n V(x) such that its derivative Vo (x) with resp ect to the so lutio ns of eq ua tio n (8.1) is neg ative (d efinite) semi -d efinite in tha t reg io n. The co nd itio ns f or a sym p totic sta bility may be relaxed slightly by a theorem due to LaSalle which is given next. Theorem 2 The null solutio n of the syst em of eq ua tio n (8.1) is a sym p to tically sta ble if in so m e neig hb o urho o d R of the o rigin there exists a positive definite function V(x) such that its derivative Vo (x) is negative semidefinitein R and is non-zero in R along any solutio n of equation (8.1) other than the null solution. If the co nditio ns of Theorem 2 are satisfied for all x then the equilibrium state is asymptotically stable in the large. The function V(x), which satisfies the stable conditio n of Theo rem 1, f o r the eq ua tio n (8.1) is k no wn a s a Lya p uno v f unctio n f o r the no nlinea r eq uatio n. It sho uld b e no ted that finding a Lya p uno v f unctio n p ro vid es a suf ficient co nd itio n f o r sta bility of eq ua tio n (8.1). Thus the difficulty o f a pp lyin g Lya p uno vs m etho d is o ne of finding a g o od V(x). Failure to find o ne d oes no t m ea n the system is unsta ble a nd finding one may result in a very conservative condition for stability.
140
Q (x) = x Qx
8.3 Application of Lyapunovs Method

As might be expected it is possible to prove all the known results for the stability of a linear system using Lyapunovs method but good results for nonlinear systems, despite intensive research by mathematicians, are not easy to obtain.
8.3.1 Linear System

For the linear system xo = Ax one can choose V(x) = x Qx . Differentiating and substituting for xo and xo T gives Vo (x) = x T (A T Q + QA) x = x T Px where P = A T Q + QA which is known as the Lyapuno v equatio n, and which can be solved for Q given A and P. 01 A Lyap uno v f unctio n fo r a5/4 1/4 linear system ca n b e f o und b y taking P = -I. Fo r exa m ple if A = e - 2 - 3 o then whit P = -I the solution for Q is Q =1/4 1/4 m and c V (x) = (1/4) (5x12 + 2x1 x2 + x22) which is positive definite. On the o ther ha nd if we ha d assum ed a ny q ua d ra tic f orm f o r V(x) it m a y no t ha ve been p o ssible to p ro ve sta bility. Fo r instance for v (x) = x12 + x22 Then Vo (x) = 21 xo 1 + 2x 2 xo 2 = 2x 1 x2 + 2x2 (- 2x1 - 3x2 ) = - (2x1 x2 + 6x2 2) which is not negative definite a nd no co nclusio n ca n b e dra wn ab o ut sta bility. A p hysical interp reta tio n ca n b e fo und f or this if the eq uatio ns (8.6) a nd (8.7) are d ra wn o n a p hase pla ne f or diff erent co nsta nt values of V to g ether with trajecto ries f o r the system . It will b e seen tha t all the trajecto ries for the system cut the form er co nsta nt V curves o nly o nce in the directio n of decreasing V, which is not the case with the latter. (8.7) (8.6) (8.5) (8.4)
T
(8.3)
141
8.3.2 Nonlinear System

There ha ve b een m a ny eff o rts to try a nd find suita ble V f unctio ns f or no nlinea r system s. One of the m ost successf ul wa s due to Lure who for a feedback system with a single nonlinearity in the feedback path given by xo = Ax + bu u = n (z) z = cT x used a V functio n of the form V (x, z) = x T Qx + b (8.8)
# n (z) dz
(8.9)
He was able to obtain necessary conditions for V(x, z) to be a Lyapunov function with (a) the nonlinearity confined to the sector [0, K] and all the eigenvalues of A in the left half plane and (b) the no nlinearity co nfined to the sector [, K] and all the eigenvalues of A confined to the left half plane or on the imaginary axis. Perhaps most importantly it led to results being found in the frequency domain which are more easily used by the engineer.
8.4 Definitions and Loop Transformations

Before presenting so me freq uency domain results some definitions need to be given a nd also two usef ul loop tra nsformations. The results have been derived from either or both Lyapunovs method and functional analysis. Definition 5 A system is bo und ed inp ut bo und ed o utp ut (BIBO) stab le when a b o und ed inp ut results in a b o unded o utp ut. Typ ically a bounded signal is defined in terms of its L 2-norm, that is
3
x (t) =
0
# x (t ) dt , which is finite for a no n-periodic BIBO sig nal.

2
Definition 6 The A matrix of a transfer function G(s) is Hurwitz if all of its eigenvalues lie in the left ha nd side of the s -plane. This will be denoted A ! " A1, Definition 7 If the A matrix is Hur witz apart from one eigenvalue at s = 0 then A ! " A0, Definition 8 If the clo sed lo op system of Fig ure 3.1 has o nly eigenvalues in the left ha nd side s -pla ne , when the no nlinearity is repla ced by a linear gain k, this will be denoted by Ak ! " A1,
142
An Introduction to Nonlinearity in Control Systems Definition 9 Sector bounded nonlinearity
A no nlinea rity n(x) which is a real, co ntinuo us, single valued , scalar f unctio n o f x, with n(0) = 0 b elo ng s to the cla ss " N , if k1 # n (x) /x # k2 for x ! 0 and is denoted by n (x) ! 6k1, k2 @. Definition 10 Slope bounded nonlinearity A no nlinea rity n(x) which is a real, co ntinuo us, single valued , scala r f unctio n of x, with n(0) = 0 belo ng s to the cla ss " M , if m1 # dn (x) /dx # m2 for x 0 and is denoted by nl (x) ! 6m1, m2 @ Definition 11 Monotonic nonlinearity A nonlinearity belo nging to class - {M} which also has m 1 # 6n (x 1) - n (x 2) @ / 6 x 1 - x 2 @ # m 2 for all finite x1a nd x2x 1 belo ngs the class {M m}
8.4.2 Loop Transformations

Two interesting tra nsf orm a tio ns which are usef ul f or o btaining sta bility results with the ab o ve criteria , a nd ha ve also b een used in the proofs of some of them, are the pole and zero transformations.
143
An Introduction to Nonlinearity in Control Systems 8.4.2. 1 Pole transformation
The clo sed lo op pro p erties f o r the a uto no m o us system co nsisting of a no nlinea rity n(x) a nd pla nt tra nsf er f unctio n G (s) in the fo rwa rd p ath are not cha ng ed if a neg ative f eed ba ck pa th with g ain is pla ced a ro und the pla nt a nd a f eedf orwa rd path of gain is placed around the nonlinearity as shown in Fig ure 8.1.
Figure 8.1 Illustration of pole transformation for loop with no input.
The new autono mous loop has linear dynamics Gt (s) =
and if n(x)! 6 k1, k2 @ the new no nlinearity G (s) 1 + tGof the linear d yna mics changes the sector of (s) nt (x) ! 6 k1 - t, k 2 - t @ . The tra nsformatio n as well as changing the poles
the no nlinearity but not the sector width defined by k2-k1 for n (x) ! 6k1, k2 @ 8.4.2. 2 Zero transformation The added paths of gain in the zero transformation are the reverse of those in the pole transformation, tha t is the k1 feedback path is around the nonlinearity and the feedforward path in parallel with the linear dynamics. The new loo p k2 elements are Gv (s) a nd nv (x) where Gv (s) = G (s) - v a nd nv (x) ! ; E , 1 + vk1 1 + vk2
8.5 Frequency Domain Criteria

Several frequency domain criteria are given in the follo wing sections.
8.5.1 Popov Theorem

A sufficient co ndition for the auto nomous system of Figure 3.1 to be asymptotically stable with Ak ! " A1, for k = 0 a nd n (x) ! 60, k @ is tha t there exists a non-negative q such tha t Re "(1 + j~q) G (j~), + k- 1 $ d 2 0 (8.10)
where is an arbitrary small constant. When A k ! " A0, for k = 0 then the result is tr ue for n (x) ! 6f, k @ where is a small positive constant. Equatio n (8.10) can be written as Re (G (j~) 2 - k- 1 + q~ Im G (j~) (8.11)
If the real and imaginary axes of the Nyquist plot are denoted by x and y then this means tha t for stability the Nyquist plot of G (j~) should lie to the right of the straight line x = - k- 1 + q~y 144
which is a f unctio n of the f req uency , so the line ha s to b e cha ng ed fo r ea ch f req uency o n G (j ~ ) . The result is usua lly interpreted graphically in the frequency domain using the Popov locus Gu (j~) = Re G (j~) + j~) + j~ Im G (j~) The theorem can then be stated as, A sufficient co ndition for the system to b e asymptotically stable is that the Popov lo cus
Gu (j~) lies to the right of a straight line with finite slope passing through the point (-k-1,0), provided lim sG (s) 2 - k- 1 The slope of the line is q-1 .
The graphical interpretatio n of the Popov criterion is shown in Figure 8.2 for two different Gu (j~) loci labelled G1 and G2. Assuming the transfer functions are such that the linear system will be stable for gains, k, in the sector k ! 60, kH @ , which is kno wn as the Hurwitz sector, then it is seen that the Hur witz and Popov sectors are the same for G2 but the latter is smaller for G1.
Figure 8.2 Illustration of the Popov criterion.
8.5.2 The Circle Criterion

There a re two f orm s of a circle criterio n which ca n b e d erived fo r a sector b o und ed no nlinea rity f ro m the Pop o v criterio n. One is the case when q is ta k en eq ua l to zero a nd the criterio n is then usua lly ref erred to as the circle criterio n . Fo r this case it has further been shown to be true for BIBO stability. It is shown in reference [8.1] that if a pole transformation is applied to the autonomo us system, which results in a sector transformation for the nonlinearity and this is taken as n (x) ! " N , then the Popok criterion for a stable G(s) is satisfied if the Nyq uis t plo t of G (j~) do es no t enter the 1 v 1 k ,1 k k with centre a t the p oint - 2 c 1 1 + 1 2 m 2 q~ c 1 1 - 1 2 m a nd which p asses thro ug h the p oints ( - k - 1 , 0) a nd (- k - 1 , 0) . The dep endence of the circle o n is so m etim es no t a pro blem since q ca n b e cho sen la rg e eno ug h to ta k e the circle centre
2
circle
well into the third q ua dra nt f or a ny releva nt values of . If q is ta ken eq ual to zero , ho wever, which is a sp ecial a nd m ore restrictive case of the Popov criterion, then the result is the follo wing theorem.
145
An Introduction to Nonlinearity in Control Systems Theorem 1 The Circle Criterion
The system will be asymptotically stable for n (x) ! " N , and a linear tra nsfer functio n G (s) , here the circle D has its centre o n the negative real axis of the Nyq uist diagram and passes thro ugh the points ( - k1 - 1, 0) and (- k2 - 1, 0) if (i) for 0 1 k1 1 k2 the Nyquist plot of G(j) encircles the circle D, P times counter-clockwise, where P is the number of poles of G(s) with positive real parts (ii) for 0 =1 k1 1 k2 the Nyquist plot of G(j) satisfies ReG (j) 2 - k2- 1 and P = 0. (iii) for k1 1 0 1 k2 the Nyquist plot of G(j) lies in the interior of the circle, D. The criterio n corresp o nds to the Nyq uist criterio n f or linea r system s f o r (i) with the Nyq uist critical point ( -1, 0) b eco m ing the circle D. Figure 8.3 illus trates the situa tion for two stable transfer functio ns G1 and G2 whose Nyquist plo ts close with a semicircle in the right half plane and the circle is for a nonlinearity in the sector [1,3]. The closed loop nonlinear system with G 1 is sta ble a ccording to the criterio n b ut no co nclusio n ca n b e rea ched a bo ut G 2. It sho uld b e no ted tha t as k 1 tends to zero the circle tend s to a vertical straig ht line thro ug h - k 2 - 1 . Ob vio usly, a disa d va nta g e of the criterio n f or autono mous systems, since it is a special case of the Popov criterion, is that it produces more conservative results.

146
Figure 8.3 Illustration of the circle criterion
It is interesting to compare this result with the DF method as for any nonlinearity n (x) ! " N , it was sho wn in sectio n 3.3 that N (a) ! " N , so the DF for the class N of sing le va lued nonlinear ities is the d ia meter of the circle of Fig 8.3. Thus for DF stability G(j) sho uld not cut the diameter of the circle. Aizermann conjectured tha t DF stability was correct b ut the so -ca lled co unter exa m ple o f Fitts g iven in sectio n 5. 5 sho ws this to b e untrue. The f ea ture of the Fitts exa m p le which brea ks DF stability d o wn is that G (j) exists in bo th the seco nd a nd third q ua dra nts (d efined in a co unter clo ck wise direction from the first quadrant within the positive axes). It is interesting to note that if G(j) is confined only to the third a nd f o urth q ua d ra nts then DF stability ca n b e p ro ved b y the m ore general circle criterio n d iscussed at the sta rt of this sectio n. Since this is a typical regio n fo r m a ny Nyq uist lo ci, tha t is the pla nt tra nsf er f unctio n has a p ha se lag between 0 and 180, this result may be regarded as the small phas e shift stability theorem.
8.5.3 The Off-axis Circle Criterion

The off-axis circle criterion applies when the no nlinearity is mono tonic. Its applicability is therefore restricted but it can give less conservative results than the other criteria. Theorem 2 The a uto no m o us system with G (s) such that Am 1 a nd Am 2 ! A1 a nd n (x) ! " M m , with nl (x) ! 6 m 1 , m 2 @ is asym p to tically stab le if the Nyq uist lo cus G (j) fo r ~ ! 6 0, 3 @ d o es not enter the circle D , where D is a circle which p asses thro ug h the po ints (- m1- 1, 0) and (- m2- 1, 0) with centre anywhere in the Nyquis t plane. Remark For the case of m1=0 the circle becomes a straight line throug h (- m2- 1, 0) , so that for stability the Nyquist locus must lie to the right of this line. Figure 8.4 illustrates the situatio n showing the loop stable for G1 but not for G2. However sta bility fo r G 2 ca n b e pro ved b y cho osing a larg er rad ius fo r the circle a nd mo ving its centre f arther a wa y f ro m the x a xis.
147
Figure 8.4 Illustration of the off-axis circle criterion.
Remark Since the dia m eter of the circle ca n b e ma d e very la rg e the limit of the cho rd is the line between the p oints ( - m 1- 1 , 0) a nd (- m2- 1, 0) . Thus for a monotonic no nlinearity instability can only occur if the Nyquist locus exists in both the second a nd third q ua d ra nts. This a g ain sho ws why a G (j) like that in the Fitts co unterexa m ple b reak s d o wn a DF sta bility typ e result based on the slope bounds of the nonlinearity which was conjectured by Kalman.
8.6 Examples
Several examples are presented in this section to allow co mparisons of the various approaches. Example 1 Co nsid er a system with a n o pen loo p tra nsf er f unctio n of G (s) = k (1 - s) /2 (1 + s) . Substituting s = j ~ , with k = 1, gives G (j~) = Gu (j~) = 1 - j~ 2 (1 + j~) 2 1 - ~2 ~ ) (1 +
2
= -
2 (1 j~ ~ ) (1 - + ) 2
2
(1 - ~ 2 ) 2 (1 + ~2 )
j~ (1 + ~ 2)
so that the Popov locus is
j~2 1 + ~2
which on writing Gu (j~) = Xu + jYu and eliminating from the
exp ressio ns fo r X u a nd Yu gives the straig ht line Yu = X u - 0, 5 . The lo cus starts a t (0.5,0) f or = 0 a nd finishes at ( -0.5, -1) for = . Fig ure 8.5 sho ws the Nyq uist a nd Pop o v lo ci, with the latter d ra wn fo r p ositive f req uencies o nly. It is clea r tha t a line ca n b e dra wn thro ug h the o rigin with slo p e a p pro xim a tely 2.0 so tha t the Po po v lo cus lies to its rig ht. This m ea ns a k equal to infinity, however the criterion also requires lim sG (s) 2 - k- 1 , which gives - 2- 1 2 - k- 1 , that is k < 2 . The Hurwitz sector is (-2, 2) so that the Hurwitz and Popov criteria are the same for positive gains.
148
Figure 8.5 Nyquist and Popov Plots
Example 2 Consider a nonlinear feedback system with G (s) = - 1/ (1 - sT1) (1 + sT2) and a nonlinear element in the sector n (x) ! (k 1 , k 2 ) . It is req uired to f ind co nd itio ns f or sta bility of the closed lo o p system . The cha ra cteristic eq ua tio n fo r th e linea r system with a lo op gain of k is sT 1 T 2 + s (T 1 - T 2 ) + k - 1 = 0 so tha t the Hurwitz secto r is (1, ) pro vid ed T 1 >T 2 . Fo r the no nlinea r system the Po po v criterio n ca nno t b e a pp lied directly a s G (s) is no t stable, that is Ak g " A 0, f or k= 0. Ho wever
2
Now G(s) is stable provided T 1>T2 and >1 and since the Popov locus of G(j) is entirely within the lower half pla ne the Popov sector for n (x) is (0,) which gives stability for n(x)! (1,) , provided T1>T2, in agreement with the Hurwitz sector. For n(x)! (1, ) the circle D for the circle criterion is a unit circle with centre at (-0.5, 0). It is easy to show tha t the Nyq uist plot of G(s) will lie within this circle for all T 1, T2 so that stability canno t be proved. Example 3 s + 0, 8 s (s + 0.1) (s2 + 0.3s + 9) the Hurwitz sector is greater than the Popov sector. As ~ " 0 , the real part of G(j) tends to -7.81 so that the locus will lie to the right of a vertical straight line passing throug h (-7.81,0), giving a gain for stability of k<1/7.81=0.128. A Popov line d ra wn to keep the Po p o v lo cus to the rig ht intersects the neg a tive real a xis at a ro und -3.40 so to sa tisf y the Po p o v criterion k<0.29. Similarly a line drawn to keep the Nyquist locus to the right will intersect the negative real axis at -0.48 giving k<2.08 for stability. Finally for the linear system the maximum gain for stability from the Hurwitz sector is 2.61.
149
if a pole transformatio n is used with a gain then Gt (s) = 1/ (s T1 T2 + s (T1 - T2 ) + t - 1 a nd nt (x) ! 6 k1 - t, k 2 - t @ .
Consider a system with G (s) =
with n (x) ! " N , a nd "M , such that n (x) ! "0, k , and
nl (x) ! " 0, k , . The Nyquist and Popo v loci for the tra nsfer f unction are shown in Fig 8.6 from which it ca n be seen that
8.7 Conclusions
In this chap ter so m e a b solute sta bility results ha ve b een presented p rim arily those using freq uency d o m ain criteria . Their major advantages are that they allow co mparisons with the approximate DF method and because of their formulatio n allo w stab ility to b e p ro ven f or cla sses of no nlinea rity, ra ther tha n a sp ecific o ne, so that they a re r o b ust to cha ng es in the nonlinearity. Their disadvantages are that results of their application can prove conservative for a specific nonlinearity and they are restricted in their applicability to a feedback loop with a single nonlinear element.
8.8 References
8.1 Mohler, RR 1991, Nonlinear Systems Vol1 : Dynamics and Control , section 7.2.2, Prentice Hall, New Jersey.
8.9 Bibliography
Atherton, DP 1982, Nonlinear Control Engineering, student editio n Appendix C,Va n Nostrand Reinhold, London. Atherton, DP 1981, Stability of Nonlinear Systems, Research Studies Press, John Wiley, Chiches ter. Aizermann, MA & Gantmacher, FR 1974, Absolute Stability of Regula tor Systems, Holden Day, San Francisco. Cook, PA 1994, Nonlinear Dynamical Systems, Second edition, Prentice Hall, New Jersey. Mohler, RR 1991, Nonlinear Systems Vol1 : Dynamics and Control Prentice Hall, New Jersey. Slotine, J-J & Li, W 1990, Applied Nonlinear Control, Prentice Hall, New Jersey. Taylor, JH & Narendra, KS 1973, Frequency Domain Criteria for Absolute Stability, Academic Press, New York. Vidyasagar, M 1978, Nonlinear Systems Analysis, Prentice Hall, New Jersey.
www.job.oticon.dk
150
Design of Nonlinear Control Systems
9 Design of Nonlinear Control Systems

9.1 Introduction
In a greement with the situa tio n f or bo ok s o n linear co ntrol system s the early cha pters of this b oo k ha ve necessarily b een devo ted to a na lysis method s ra ther tha n d esig n. In p ra ctice the d esig n p ro cess involves co nsid eratio n of the ap p ro a ch to be used to m eet the g iven sp ecif icatio ns, the selectio n of the co m p o nents to be used a nd their m o delling a nd f ina lly the use of a nalytical m etho ds a nd p rob a bly sim ulatio n to sho w tha t the d esig n m eets the sp ecif ica tio ns. This is norm a lly an iterative pro cess req uiring a return to earlier sta g es in the pro cess, m a king cha ng es a nd then co ntinuing . The difficulty if a ny co m p o nents a re no nlinear is, as has b een seen, that the b eha vio ur is d ep end ent o n the m a g nitude a nd typ e of inp ut sig nals to which the system is subjected , no necessa ry a nd suf ficient co nd itio ns f or sta bility a ssessm ent a re k no wn a n d no general system analysis method is available. Thus if the system sp ecifica tio n req uires the inp ut step respo nse to ha ve certain prop erties, then it is not sufficient to k no w tha t this will b e the ca se f or a step inp ut of a sing le m a g nitud e, as with a linea r system , b ut o ne m ust d etermine that the req uirem ents will b e satisfied fo r all step inp uts within the allo wab le ra ng e. Also if the system is sub ject to m ultip le inp uts it is not sufficient to determine the behaviour for the inp uts separately since nonlinear interactions between them may cause unexpected beha viour. The b eha vio ur of high o rder no nlinear systems is d ep end ent on the regio n of the state space in which o peratio n tak es pla ce, as ha s b een d em o nstra ted f or a seco nd ord er system b y use of the p hase pla ne. Once a system has b een d esig ned sim ulatio n pro vid es a n excellent to ol f o r checking the p erf o rma nce within the allo wa ble state space region of operation, however, if it is done without theoretical understanding it can prove inefficient and possibly erroneous if the simula tions fail to cover appropriate scenarios. One mig ht a rg ue that the first ob jective of a desig n sho uld b e to try a nd ad d co mp ensatio n to m ak e the system linea r, so as to allow easy analysis, but this may not be possible or even desirable. The advantage of linearity is that the response is independent of the inp ut amplitude but it ma y only be possible to have linear behaviour between a specific input and output. A very common form of nonlinearity is plant actua tor saturation, for example motor torque saturation in a speed or position control system. It might be possible to avoid this by using a larger motor but if the system satisfies the specifications with a smaller motor then the conseq uence of cha nging to a larger one is a more costly design. An objective of a ny engineering d esig n is to m eet the specifica tio ns a t the lo west co st. The sp ecifica tio ns are p ara mo unt a nd as well as p erf orm a nce a nd reliability issues they m a y co ver other a sp ects such a s o p erating co sts, weig ht o r volum e, etc. a nd enviro nm ental co ncerns. In so m e ca ses when the a ctuator to a pla nt sa turates it ca n ca use p ro ble ms a s the f eedb a ck loo p is then essentially o p erating op en lo op . One extrem e ca se is when the p la nt mo d el is a n unsta ble tra nsf er f unctio n.
151
In ma ny cases where nonlinearity occurs it may be possible to do a control design based on an approximate linear or linearized system model and then check b y simulatio n that the effects of the nonlinearity still allow the req uired p erformance to b e m et. Such a desig n m a y invo lve p ara m eter selectio n f or fixed co ntrollers lik e PID, or o ne of several state sp ace desig n m etho ds. The latter m a y use sta te va ria bles estim a ted us ing o bservers a nd altho ug h unlik e fo r linea r system s there is no rig oro us justifica tio n pro vid ed b y a sep aratio n principle the pro ced ure is effective. Tod a y mo st co ntro llers a re im plem ented using micro p ro cesso rs so that it is just a s easy to im plem ent a no nlinear co ntro l la w as a linea r o ne. When sho uld this be done and why? What are the advantages and disadvantages? In the f ollo wing sectio ns so m e d esig n m etho ds f or no nlinea r system s are m entio ned a nd ho p ef ully these will go so m e way to answering these questio ns. The topics, however, are only covered briefly to give a flavour of what is available a nd a lso b eca use to d o o therwise wo uld req uire the intro d uctio n o f a dd itio nal theoretical co ncep ts. Co m plete texts are a vailable o n m o st of the topics a nd ref erences will b e g iven in the ap p ro p ria te sectio ns. There are also several bo ok s tha t give a g oo d o verview of ma ny of the topics co nsid ered in this cha pter, such a s references 9.1 a nd 9.2. The la tter, ref erence 9.2, beca use of the a uthor s ind ustrial ba c kg ro und is p articularly interesting in off ering a m uch m o re pra ctical p ersp ective than is often the case.
9.2 Linearization
Nonlinear systems are often designed using linear methods based on an approximate linear model. Simulations may show tha t satisfactory performance may be obtained from a design using one linear model but in some situa tions this will not b e p ossible d ue to the larg e cha ng es in the b est linear m o d el with resp ect to the system o p erating p oint. Linea r controllers may then be designed for the different operating point models, or linearizations. The controller parameters are then sched uled in so m e ma nner b ased o n k no wledg e of the o perating point. Ma ny a pp roa ches a re p ossible which ca n b e co nfirm ed to wo rk sa tisfa ctorily in sim ulatio ns a ltho ug h there is so m etim es a la ck of rig o ro us theory to sup p o rt the techniques. The a p pro a ch is pa rticularly co m m o n in the a ero sp a ce ind ustry where a co m m o n a pp roa ch is to em plo y sta te f eedb a ck with the gains scheduled according to the specific operating point. The design is typically done off line in tha t a linear mathematical model is obtained for each operating point and then the required controller gains calculated and stored ready for implementatio n around tha t specific operating point. Various methods are also used to avoid discontinuous cha ng es o f co ntroller gains with cha ng e in the o p erating p oint. Two g o o d introd ucto ry ref erences fo r work in this f ield are [9.3, 9.4]. If a new m a them a tical m o d el is id entified d uring o peratio n then this p r o vid es a sta nd a rd f o rm of ad a ptive control. All co ntrol designs typical involve consideratio ns of system performance under plant cha nges and how severe
these m a y b e exp ected to b e will a ff ect whether a ro b ust d esig n ca n b e a cco m plished b y a sing le co ntro ll er or whether scheduled controllers or an adaptive control is required. Ma ny p ro cesses which need to b e co ntro lled in the chem ical ind ustry of ten ha ve q uite no nlinea r d yna m ics so these are sometimes controlled by PID controllers which allow the three PID terms to be scheduled according to the operating p oint. In recent years there has also b een a sig nifica nt a m o unt o f wo rk o n the to pic of m ultip le m o d els f o r use in b oth modelling and control [9.5]
152
9.3 Frequency Response Shaping

In classical controller desig n methods for linear systems it is a commo n pro ced ure to add kno wn forms of compensator, such as p hase lead o r lag , to achieve a d esired freq uency do m ain p ro perty, such as a minim um allo wab le g ain o r p ha se m argin. In such d esig ns the req uirem ent is then to o b ta in suita ble co ntroller p ara m eters so the system m eets the sp ecificatio ns. This is a gain a sa tisf a cto ry a pp ro a ch fo r a no nlinear system a nd in term s of the DF a p proa ch, a ssum ing a single no nlinear elem ent, b eco mes o ne of sha ping the o p en lo op freq uency resp o nse lo cus so that it l ies at a n ap prop riate dista nce f ro m the C(a) lo cus. The C(a) lo cus ma y be rega rd ed a s a m o ving (-1, 0) p oint o n the Nyq uist p lot so a ga in a nd p ha se m argin may be defined for each value of a. Typically a design will involve shaping the frequency response locus to ensure that the value of the gain and/or phase margin for all a is never less than a given value. Ap a rt fro m the o b vio us wa y of do ing this b y co m p a ring the lo ca tio n of the C(a) a nd G (j ) lo ci o n a Nyq uist plot, or o n a Bode plot, particularly for a single valued nonlinearity, another approach is to make use of methods for determining limit cycles. U sing such a n ap p ro a ch the a m o unt of ga in o r p ha se shif t, the la tter b y a d ding a tim e dela y, which m a y b e added to the loop before a limit cycle results can be calculated.

153
9.4 Nonlinear Compensation

Given the block diagram for a feedback system with a nonlinear element, possibilities may exist, dependent upo n the lo catio n of the no nlinea rity, fo r co m p ensa tio n using a nother no nlinea rity. Fo r exa m p le, if no nlinea rity exists at the inp ut to the pla nt then it m a y b e p ossible to includ e a no nlinea rity in the co ntroller o utp ut to m ak e the lo op beha vio ur m o re linear . Typ ically this mig ht b e used to pro vid e a hig her g ain fo r sm all sig nals to co m pensate fo r a valve d ead zo ne. Fro m p urely blo ck diag ra m co nsid eratio ns it is p ossible to includ e a nother p ath to the system o utp ut in p arallel with the pla nt, but in practice this is not physically possible. However, the closed loop design problem may be made linear by taking such a p a th a nd co nnecting it to the f eed ba ck sig na l a s sho wn in Fig ure 9.1. If in the fig ure the no nlinear elem ent N 2 i s ta ken eq ual to 1-N 1 then it is ea sy to sho w tha t the closed lo op cha ra cteristic eq uatio n is linea r a s is the tra nsfer functio n from the inp ut, R, to the control signal, U, which is U (s) R (s) = Gc 1 + GGc (9.1)
Figure 9.1 A Nonlinear Compensation Method
So m etim es no nlinearity o ccurs in a mino r f eed ba ck p ath, fo r exa m ple, when the d a mping is a no nlinea r f unctio n of speed, unlike visco us friction which is assumed to vary linearly with speed. Obviously, if speed is measured then feeding back a no nlinea r f unctio n of this sig nal ca n be used to m a k e the to tal sp eed f eed b a ck linear. Pra ctically, ho wever, this m ea ns k no wing the f orm of the no nlinea r speed d a mping a nd m a y also req uire a n a dditio na l s ensor fo r speed m easurem ent o r the implementation of a speed estimator. A very commo n use of nonlinear compensation is for the prevention of integral windup, which is provided with most commercial PID controllers. As mentioned previously saturation always eventually occurs in a plant, whether it be the m a xim um torq ue fro m a mo to r, the m a xim um po wer fro m a hea ter, the m a xim um flo w thro ug h a valve a nd at this stag e further increase in the co ntroller o utp ut will not increa se the level of the system o utp ut, a l lo wing of co urse fo r a ny possible dela y in the resp o nse. If the co ntroller co ntains a n integ ra l term then its o utp ut will increa se in m ag nitud e a nd no t start to decrease until its input cha nges sig n. This ca n cause large o vershoots in the system respo nse a nd in the case of a n unstable plant, since the loop is basically open, possibly catastrophic behaviour. Thus, integral wind up protection is achieved by stopping the integration when the co ntroller outp ut signal reaches a certain level. Today, since PID co ntrollers use micro pro cesso rs va rio us software algo rithm s are used to im plement the pro ced ure, b ut in the d a ys of a nalog ue co ntrollers a sim ple ap p roa ch wa s to inco rpo rate a no nlinea r feed b a ck pa th aro und the integrato r. Ano ther a pp roa ch to a voidin g integral wind up [9.6] is sim ply to switch o n the integ ra to r o nly when nea r to the stead y state, which in so m e wa ys
seem s m o re lo gical a s the integ ra tio n no rm ally ha s a d esta bilising ef fect o n the lo op a nd the o nly reaso n it is usually includ ed is to elimina te steady state errors. A slight difficulty is in obtaining a suitable algorithm to implement the switching on
154
An Introduction to Nonlinearity in Control Systems and off conditions without causing transient problems.
9.5 Compensation using DF Models

In the previo us sectio n it was a ssum ed that a ny no nlinearity in the co ntrol system co uld b e represented in a simple b lo ck diag ra m structure. In p ra ctice, suc h m o d elling m a y no t be easily a chieved o r m a y not b e of releva nce fo r the a pp roa c h because of the complexity of the plant model structure with multiple dyna mic and no nlinear elements. In this case a possible approach is to mak e use of a freq uency d epend ent DF model, N (a, ~ ) , for the pla nt, which ma y b e obtained from either a m athem a tical mo d el o f the pla nt or po ssib ly freq uency respo nse id entifica tio n testing . As lo ng a s the freq uenc y response sho ws a do mina nt o utp ut at the inp ut freq uency the mod el should be satisfactory for use in compensator desig n stud ies. Two a p pro a ches a re p ossible, the first sim p ly b eing a DF im p lem enta tio n of the structure of Fig ure 9.1, that is if the no nlinea r pla nt has the mo d el, N 1 (a, ~ ) , then the DF m o del f ed ba ck aro und the co ntroller N 2 (a, ~ ) , sho uld b e suc h that N1 (a, ~) + N2 (a, ~) . G (j~) a linear model. To DF accuracy the open loop is then essentially linear, so that linear compensation ca n be used. An a lterna tive a pp ro a ch is to pla ce a series co m p ensator b ef ore the pla nt which ha s a DF m o d el N 2 (a, ~ ) such that th e series co m b inatio n N 2 (a 2 , ~ ) a nd N 1 (a 1, ~ ) a re a p pro xim a tely eq ual to G (j ~ ) a linea r m o d el. The diff erent sub scripts o n the as ha ve b een used to d enote that the a mplitud es into the DF mo d els will not be the sa m e. Mo re d etails o n the a ctual ap proa ch to a chieving this d esig n ca n b e f o und in ref erence 9 7. The a p proa ch ha s also b een extend ed f or use with rela y a utotuning (see chap ter 7) to imp lem ent no nlinear term s in a PID co ntro ller. W hen the rela y a utotuning a pp roa ch is used for a no nlinear pla nt then the estim ated critical point will cha ng e as the rela y o utp ut a m p litud e is cha ng ed. This not o nly ena bles o ne to estimate the d eg ree of no nlinea rity in the pla nt b ut also enables the d esig n of a no nlinear PID to prod uce roughly a linear input outp ut response [9.8].
9.6 Time Optimal Control

Early work in time optimal control was done for second order systems by making use of the phase plane approach. Ho wever, the topic develo ped rapidly in the 50s a nd a larg e num b er of p ap ers o n the subject were p ublished in the 60 s. Althoug h much of the early wor k was done in the USSR, major developments started in the USA in the 60s prima rily driven by the space program where an important requirement was to develop control strategies using minimal fuel, a pro blem which is clo sely related to time o ptim al co ntrol. The co ncepts used a re b ased o n classical Ha m ilto nia n m echa nics and generally are known as Pontriagins maximum principle or the Bellman-Hamilto n-Jacobi approach. The books by McCa usla nd [9.9], Ja co bs [9.10], a nd Gla d a nd Ljung [9.1] give g oo d intro d uctio ns a nd m o re ad va nced bo ok s a re those b y R ya n [9.11], Atha ns a nd Fa lb [9.12] a nd Lewis [9.13]. The m a them atical theo ry of these a p pro a ches is no t easy so here a sim ple introd uctio n to the to pic is p ro vid ed b y wa y of the p ha se pla ne f or a d o uble integ rator pla nt. The so lutio n to the problem using the classical Hamiltonian approach is given in chapter 6 of reference 9.9.
155
When co nsid ering the initial co nditio n resp o nse of a seco nd o rd er system in sectio n 2.3.2 it was seen that the p ha se pla ne plot fo r a clo sed lo op system co nsisting of a rela y a nd a d o uble integra to r pla nt tra nsf er f unctio n co nsisted of pa rab olas. This corresp o nd s p hysically t o a p ositio n co ntrol system , a ssum ing a co nsta nt inertial lo ad a nd no f rictio n, being driven at constant acceleration, K, with the equatio n of the parabolas being x22 - x202 = ! 2K (x1 - x10) (9.2)
where x1 is po sitio n a nd x2 velo city. If the inp ut to the rela y is - x 1 - m x 2 the switching b etween p ara bo la s co rresp o ndin g to positive and negative acceleration (deceleration) takes place on the line - x1 - mx2 = 0 in the phase plane. It is easily a pp reciated tha t the fa stest po ssible wa y to the o rigin f ro m a n initial co nditio n of ( -h, 0) is to ha ve the m axim um acceleration of +K to a distance h/2 from the origin followed by the maximum deceleration of K to the origin. The final trajectory to the origin is the parabola x22 = - 2Kx1 , or if the motion starts from (h, 0), rather than (-h, 0), the final trajecto ry is x 22 = 2K x1 . Thus to a chieve a minim um tim e resp o nse in the id eal case these two p arts of p ara b olas sho uld form the switching curve no t the line - x1 - mx2 = 0 . A minimum time response will therefore be obtained if the input to the relay is - x2 x2 - 2Kx1 . This has been called a V|V| system [9.14]. It is, however, often implemented by having the switching bo und a ry a s - x2 + 2K sgn (- x 1) error) system. The optimal solution is seen to be dependent on knowing the system model, in this case the parameter K assuming negligible friction. If K is estimated, ho wever, the switching boundary can be adjusted by some adaptive strategy. - x1 = 0 , when it is known as the SERME (sign error root modulus
156
9.7 Sliding Mode Control

The p heno m eno n of sliding was enco untered in E xa m ple 2 of Chap ter 2, where a p hase pla ne plot was co nstructed fo r a seco nd o rd er system co ntaining a rela y with d ead zo ne a nd hysteresis. Wha t hap p ens, a s explained in the exa mple, is tha t the d yna mics of the m otio n cha ng es when the rela y switches a nd it is p ossib le tha t at b oth sid es of the switching line the resulta nt d yna mic m otio n is to ward s the switching line, pro d ucing a m o tio n which is directed with ra pid switching alo ng the switching line. The interesting f eature reco g nised a bo ut this sliding m o tio n is that o nce the sl iding m o tio n is started the trajectory is d efined a nd m a y no t cha ng e even with so m e cha ng es in the system d yna m ics. Sliding m o d e co ntrol is a particular form of variable structure control, where the feedback law is selected from several possibilities by decision rules. Variable structure control systems (VSCS) were first studied in the USSR [9.15]. A second order system will again be considered in the next sectio n before a few comments on the more general situation.
9.7.1 A Second Order System

Consider again the same system as Example 2, Chapter 2, but in this case with the relay no nlinearity replaced by an ideal relay. Again the feedback to the relay is assumed to be - x1 - mx2 and the transfer functio n of the plant after the relay to be K/s2 . The Simulink diagram is shown in Figure 9.2 with m = K = 1 . Assuming an initial condition of (- x10, 0) then since the relay output u = sgn (- x1 - mx2) the initial value of u is positive and the equation of motio n is x22 = 2K (x1 - x10) which meets the switching line x1 + mx2 = 0 at say (x1s, x2s) . The motion after switching is then described by x22 - x2s = - 2K (x1 - x1s) and this will move back to the switching line if its slope is less than - 1/m . That is if - K/x2 < - 1/m , which gives x2 < mK . This conditio n for ensuring sliding is kno wn as the reachability cond itio n. (9.3)
Figure 9.2 Simulink diagram for a system with a sliding mode.
It is normally given in a more general form since if the inp ut to the relay is taken as s so that s = - x1 - mx2 = 0 is the switching line, the control signal, u, is given by
157
u = sgn (s)
(9.4)
reachability co nditio n sso < 0 (9.5)
As the motion in the phase plane is symmetrical the actual reachability condition using equatio n (9.3), or as can easily be shown from equation (9.5), is x2 < mK. (9.6)
When id eal sliding tak es pla ce the resulting m otio n is alo ng the switching line a nd is go verned b y the diff erential eq uatio n dt m disco ntinuo us a nd no t d efined o n the switching line. Ho wever, in ad ditio n whe n sliding tak es pla ce so = 0 a nd theref ore s = - xo 1 - mKu = 0 , giving u = - xo 1 /Km = - x2 /mK which is known as the equivalent control. Figure 9.3 shows the phase plane response for the system of Figure 9.2 from the initial condition (-1, 0). The switching line x 1 + x 2 = 0 is f irst m et with x 2 < 1 , the rea cha b ility co nd itio n of eq ua tio n (9.3) f or m = K = 1 , so sliding tak es pla ce. Fig ure 9.4 sho ws the system o utp ut, x1 , (Sco p e ); - x2 (Scop e3) a nd the a pp ro xim ate eq uivalent co ntrol (Sco pe 2) ob tained by filtering, u , which is shown in Figure 9.5 continuously switching during sliding. The approximate equivalent control can be seen to become equal to - xo 1 = - x2 during sliding as given by equation (9.7) for m = K = 1 . (9.7)
Figure 9.3 Motion in the phase plane for input from (-1,0)
158
then if the switching line is to be reached it requires lim s " 0 so < 0 a nd lim s " 0 so > 0 , which can be expressed as the single
s = 0, which in the ca se of the exa mp le is dx1 + 1 x 1 = 0 . The co ntrol a ctio n req uired to bring a bo ut this m otio n is
Figure 9.4 Signals in the Simulation
Figure 9.5 Control signal u
Let it no w be assumed that the plant model is not an ideal double integrator but has some nonlinear damping. This is illustrated in Fig ure 9.6 where the no nlinear da m ping ha s b een ma d e q uite sig nifica nt, whereas in p ra ctice o ne wo uld not expect such a sig nificant cha nge from the assumed model, so that the cha nge in the respo nse on the phase pla ne is clearly seen in Fig ure 9.7. The switching line is rea ched o n a diff erent tra jectory b ut o nce it is attained sliding tak es p la ce as bef ore.
678'< )25 <285 0$67(56 '(*5((

&KDOPHUV 8QLYHUVLW\ RI 7HFKQRORJ\ FRQGXFWV UHVHDUFK DQG HGXFDWLRQ LQ HQJLQHHU LQJ DQG QDWXU DO VFLHQFHV DU FKLWHFWXU H WHFKQRORJ\ U HODWHG PDWKHPDWLFDO VFLHQFHV DQG QDXWLFDO VFLHQFHV %HKLQG DOO WKDW &KDOPHUV DFFRPSOLVKHV WKH DLP SHUVLVWV IRU FRQW ULEXWLQJ WR D VXVWDLQDEOH IXWXUH ERW K QDW LRQDOO\ DQG JOREDOO\ 9LVLW XV RQ &KDOPHUV VH RU 1H[W 6WRS &KDOPHUV RQ IDFHERRN
159
Figure 9.6 Simulink Diagram for System with Nonlinear Damping.
Figure 9.7 Phase plane response for system of Figure 9.6.
The reachability condition for this example, from equatio n (9.6), is x2 < 1 and it may be that reachability is required for a larger regio n in the p ha se sp a ce. Co nsid er the co ntro l sig nal u to b e a sum m atio n of sta te f eed b a ck a nd the sig nu m functio n, so that u = k1 x1 + k2 x2 + tsgn (s) Then sso = s (- xo 1 - mxo 2) = s (- xo 1 - mK6k1 x1 + k2 x2 + sgn (s) @) Choosing k1 = 0 and k2 = - 1/mK gives sso = s6- mKtsgn (s) @ = - mKt s = - h s where h is a positive design scalar. Selectio n of sso < - h s is known as the h -reachability condition. The control signal u = - (x2 /mK) + tsgn (s) (9.8)
160
which is seen to b e the eq uivalent co ntrol of eq ua tio n (9.7) plus the sig num f unctio n. In this ca se the switching surfa ce is rea cha ble f ro m all reg io ns in the p hase pla ne. Fig ure 9.8 sho ws the Sim ulink dia gra m f o r this system a nd Fig ure 9.9 th e resp o nse f ro m a n initial co nd itio n of (-6,0). Curve (a ) is fo r the co ntrol la w of eq ua tio n (9.4), tha t is the system of Fig ure 9.2 and curve (b) for the control law of equation (9.8), that is the system of Figure 9.8. Also shown for the second case is curve (c) the response from (-6, 2).
Figure 9.8 System with control law of equation (9.8)
Figure 9.9 Phase plane responses for the two different control laws.
Fro m the co nsid eratio ns of a sim ple seco nd order system it has b een seen that when the rea chability co nd itio n, given b y equation 9.5, is satisfied, sliding is obtained down the switching curve, which of course need not be linear. Further the m otio n ca n be reg ard ed as ha ving a n eq uivalent co ntrol a ctio n to pro d uce a slid ing resp o nse which is first o rd er, that is reduced by one, from that of the original system. The form ulatio n presented has used the ideal m athematical sig num f unction, which ca nnot be realised in practice, altho ug h it ca n b e nea rly id eal in so m e im plem entatio ns. Ho wever, the effects of very ra pid switching ma y ca use p ra ctical problem s such as rapid wear of components, so the ideal may be approximated by using hys teresis in the switching or using an s approximation to the sig num functio n such as sgn ( s) . where d is a small positive constant. s+d
161
9.7.2 Further Comments

The p resentatio n of the previo us sectio n ha s ho p ef ully indicated the objectives of sliding m od e co ntrol d esig n, na m ely to design a controller so that the motio n will reach, and subsequently remain on a sliding surface, in contrast to a line for a seco nd o rd er system , in the sta te pla ne. The d yna m ics of the system when co nfined to the surf a ce is d escrib ed a s a n id eal sliding m otio n a nd fo r a system with a single inp ut will be red u ced in order b y o ne f ro m tha t of the original system . The co ntro ller d esig n fo r hig h o rd er system s is no t easy to co ver in a short presenta tio n of the to pic b ut there are m a ny specialised texts available on the subject such as Utkin [9.16],Zinober [9.17], and Edwards and Spurgeon [9.18]. It is sho wn in these texts ho w d esig ns ca n b e a chieved using vario us state sp a ce a p pro a ches f o r MI M O as well a s SISO system s, wha t typ e of pla nt uncertainty will no t aff ect the sliding m otio n, a nd ho w the id eas ca n b e a p plied to no nlinear system s a nd sliding surfa ces, no t just hyperpla nes. Lik e m ost sta te spa ce a pp ro aches all the states ma y ha ve to b e a vailab le for implementation, so additional ins trumenta tion or observers may be required, although some results are available based o nly o n o utp ut feedb a ck . There ha s also b een sig nifica nt co vera g e in recent yea rs o n the d esig n of o bservers usin g sliding mode principles[9.18].

Gr aduat e oppo rtunities Ch eltenh a m | 24,945 + ben efits
One of the UKs int ellig ence services, GCHQs role is two-fold: to gather and analyse intelligence which helps shape Br itains response to global events, and, to p rovide technical advice for the protection of Government communication and infor mation systems. In doing so, our specialists in IT, internet , engineering, languages, information assurance, mathematics and int elligence get well beneath the surface of global affairs. If you thought the world was an interesting p lace, you really ought to exp lore our world of work.
TOP GOVERNMENT EMPLOYER
www .careersinbritishi ntelligen ce.co.uk

Appli cants must be Br iti sh cit i zens. GCHQ values d iversi ty and wel comes app li cants from all sect ion s of the commu nity. We want our workf orce to reflec t the diver sity of our work .
162
9.8 Fuzzy Logic

The co ncept of f uzzy logic was introduced b y Lotfi Zad eh [9.19]. It is perhaps a n unfortunate choice of nam e from a co ntrol d esig n viewp oint b eca use a co ntroller d esig ned using f uzzy lo gic is d eterm inistic ha ving a d efined o utp ut f or a d efined input. It is the procedure of producing the output for a specific input which is the fuzzy aspect. This procedure varies slightly according to the approach used but tha t due to Mamdani [9.20] contains membership functio ns, fuzzifica tio n and defuzzifica tion procedures, and a set of rules. Typically, standard forms for the first three may be available and the d esig ner ha s o nly to p ro vid e a suita ble set of rules. Thus, as stated b y Za d eh, it ena bles a p erso n with no k no wled g e of control theory to design a controller by writing a few logic type rules, which in princip le may be selected based on the observed response of the plant. The first po int to m ak e a bo ut the pro ced ure is that in general the co ntroller desig ned will be no nlinear, so the wisdo m of doing this if the pla nt is essentially linea r is d eb atab le. Seco ndly, it seems to b e a fairly co m m o n pra ctice to write the logic rules f or a co ntroller in the erro r cha nnel in term s of the erro r a nd its d erivative. Thus the co ntroller beco m es a no nlinear controller of the form f (e, eo) . It is not easy to see what adva ntage a co ntroller of this PD form ma y ha ve compared with the less general PD fo rm f1 (e) + f 2 (eo ) , a cla ssica l PD co ntroller with no nlinear P a nd D term s. Ma ny p ap ers d escribing a f uzzy logic co ntroller end up with a no nlinear o ne yet fail to check the respo nse variatio n with the a m plitud e of the inp ut sig na l. A very g o od discussio n o n f uzzy lo gic co ntrollers ca n b e f o und in the results of the DynL AB p ro ject of ref erence 9.21. In pa rticula r it exp la ins that the o utp ut is sim ply a sta tic no nlinear f unctio n o f the inp ut(s) a nd ho w m em b ership f unctio ns affect the shape of the resulting no nlinearity. Finally there seem to be no examples given in the literature of obtaining fuzzy logic controllers for plant transfer functio ns which are kno wn to be difficult to control, for example, plants with rig ht half p la ne zero s, reso na nt m o des o r which a re unsta ble. Fuzzy lo gic id eas ha ve b een used in other a rea s of scienc e and engineering; a text covering fuzzy logic in control is reference 9.22.
9.9 Neural Networks

Neural networ k techniq ues provide a mechanism for obtaining a static or dynamic no nlinear structure using various typ es o f no nlinea r a nd linear elem ents, a nd co nfig uratio ns. Typically p a ra m eters in the structure are cha ng ed ad a ptively (training ) to try a nd m atch given inp ut a nd co rresp o nding o utp ut da ta structures. Prob ably the ea rliest use of this co ncept in co ntrol was a varia nt o n this them e where the o bjective was to p ro d uce a no nlinear f unctio n, thus ad justing the system parameters to p roduce a k no wn outp ut for a given inp ut. This was do ne to achieve required no nlinear f unctions on analog ue co mp uters where typically the individ ual f unctio ns were a no nlinear elem ent co nsisting o f two a djustable linear seg m ents. Using sets of input and the corresponding outp ut data, training techniq ues are used in a neural network to adjust the network coefficients, no nlinearity in the static case, so that they model the system producing the data sets. To test the resulting m o d el it is usual to sub ject it to new sets of inp ut da ta to check ho w clo sely the resulting o utp ut m atches the
corresp o nding k no wn o utp ut da ta f ro m the g iven system . Ob vio usly the success of the a pp ro a ch, a s we ha ve seen f ro m the interesting behaviour of nonlinear systems discussed in chapter 1, will be highly dependent on the structure used in the network rela tive to tha t of the a ctua l system . This will aff ect m a ny fa ctors such a s the a ccura cy of the m o d el, the required training time and indeed whether the behavio ur can be modelled, for example in the case of chaotic motion.
163
Altho ug h this type of black b o x ap pro a ch ca n b e usef ul, the im p orta nce of being able to inco rp orate p hysical k no wledg e into system modelling, particularly the structural connectio ns between linear and nonlinear elements, cannot be over emphasised. The simple example of illustrating the approximatio n of a static nonlinearity by a neural network provides a good illustration of this. Suppose one has two symmetrical odd nonlinear elements (i) ideal dead zone and (ii) a cubic. Further assume there is a choice of neural network structures for doing the modelling co nsisting of either (a) linear segmented nonlinearities or (b) power law characteristics. Both of these functio ns may be described as infinite approximators, as they can approximate any nonlinear functio n exactly by taking an infinite number of them. Thus if the approximators (b) are used for (i) and (a) for (ii), then indeed to get a good approximation a large number will b e required. However, if (a) is used for (i) and (b) for (ii) then only a small finite number will be required to get an exact result. For example, if the power laws are x n with n = 1, 3, 5 etc only the cubic term is required for (ii). Neural networks ha ve been used successf ully in m a ny a reas of science a nd engineering so there a re ma ny bo ok s o n the to pic. Two dealing with their application in control engineering are references 9.23 and 9.24.
9.10 Exact Linearization

In sectio n 9.3 the possibility of making a system linear by the additio n of other nonlinear elements was discussed from a block diagram viewpoint. Some of the situations mentioned are special cases of an approach, based on state space methods, known as exact linearization. The topic is considered in depth in Isidori [9.25] but introductions are given in several textbooks including, Slotine and Li [9.26], Corriou [9.27] and Glad and Ljung [9.1]. This section will follow q uite closely the presentation given in the last reference. The b asic co ncept o f exa ct linea riza tio n is to pro vide a f eed ba ck to the pla nt inp ut, norm ally tak en a s u, which m ak es the response from the new input, r, linear. Consider the state space description xo 1 = x2 xo 2 = n1 (x1) + n2 (x2) + u y = x1 If r is a reference input then a closed loop control can be taken with u = r - n1 (x1) - n2 (x2) - a1 x1 - a2 x2 which gives the closed loop system xo 1 = x2 xo 2 = a1 x1 - a2 x2 + r y = x1 This is a linea r system between the inp ut r a nd o utp ut y a nd the va lues cho sen fo r a 1 a nd a 2 allo w lo ca tio n of the closed loo p p oles. W hen the state spa ce d escriptio n (9.9) has y a s linea r po sitio n the techniq ue is often ref erred to as co mp uted force or when y is a rotary position as computed torque and is often used in robotics. On the other hand if the equatio ns are x1 = x2 + n1 (x1) xo 2 = - x1 - x2 + u y = x1 (9.12) (9.11) (9.10) (9.9)
164
Then it is not p o ssible to directly linea rize n 1 (x 1) b y the choice of u since the two q ua ntities are not in the sa m e eq uatio n. Altho ug h if the block dia gra m is d ra wn it is ea sily seen ho w the system ca n b e linearized if x 1 ca n b e measured o r estimated using a n ob server a nd a n a d ditio na l inp ut given to xo 1 . xo 1 , ho wever, is typically a n internal varia ble which ca nno t b e influenced directly. It is easy to see that a general form of state equation which will allo w exact feedback linearizatio n if all the states are available, known as the controllability cano nical form, with state vector x, is :Jx1N J KO x2 K N O (9.13)
K K . OK . O KK OO K O LxnP L f (x) + b (x) uP with the input u chosen to be u = (1/b (x)) (r - kx - f (x))
(9.14)
Here u consists of three terms, the new input r, the linear state feedback kx and a term to cancel the no nlinearity. The last line in the state equation becomes xo n = r - kx (9.15)
So the eq uatio n is linear between r a nd y a nd the sta te f eedb a ck g ain vecto r k ca n be cho sen to suita bly pla ce the system poles. This lo oks a ll rig ht in principle b ut o ne ob vio us difficulty is if fo r so m e x , b (x ) b eco m es sm all or zero . To co nsid er the situation further a more g eneral form of nonlinear state spa ce rep resentation will be co nsidered a nd so me definitio ns given.
In Paris or Online
Internat ional programs tau ght by professors and professionals from all over the world BBA in Global Business MBA in International Management / International Marketin g DBA in In ternational Business / Internat ional Mana gement MA in Internat ional Edu cation MA in Cross-Cultural Commu nication MA in Foreign Lan guages Inn ovative Practical Flexible Affordable Visit: www. HorizonsUn iversit y.org Write: Adm issions@horizonsun iversity. org Call: 01.42.77.20.66
165
K K x2 O x3 KO . O = KK o
O . O O, y = x1
9.10.1 Relative Degree

Consider a system described by the general differential equation xo = f (x) + ug (x) y = h (x) (9.16)
where f, g a nd h are infinitely differentiable f unctio ns of the nth order state vector x. Obvio usly y does not d epend explicitly on u so that any cha nges in y will occur gradually via x. Differentiating the output equatio n one obtains y = hx (x) xo = hx (x) (f (x) + g (x) u) where hx (x) denotes the partial derivative of h(x) with respect to x, that is uh hx (x) = c ux1 ux2 , ..........n u h m uh, ux Equation (9.11) can be written yo = hx f + uhx g (9.19) (9.18) (9.17)
where the x dep endence of hx , f a nd g ha s b een o mitted f or sim plicity. It ca n b e seen tha t yo no w d ep end s directly o n u for at least some values of x, if and only if hx g ! 0 , and in this case the system will be said to have relative degree 1. If hx g / 0 then one ca n differentia te again to obtain yp = (hx f) xf + u (hx f) x g (9.20)
p The system is said to have relative degree 2 if hx g / 0 and (hx f) x g ! 0 so that y depends directly on u. Otherwise one can continue differentia ting until a derivative of y is directly dependent on u. It is useful to introduce some partial differential operators for the repeated differentiations, namely L f = f1 u + ..........fn u and Lg = g1 u + ..........gn u ux1 uxn ux1 uxn (9.21)
Lf is called the Lie d eriva tive in the directio n of f . If v d eno tes the relative d egree, which is the lo west d erivative of y that depends on v, then from the above results it can be seen tha t v = 1: Lg h ! 0 v = 2: Lg h = 0 Lg h ! 0 v = 3: Lg L f h = 0 Lg L2f h ! 0 etc. One thus has the following definitio n (9.22)
166
An Introduction to Nonlinearity in Control Systems Definition 1 The relative degree v is the least positive integer such that Lg Lvf- 1 h ! 0 .
It can be shown tha t this definition is consistent with that for a linear system where the relative degree is the difference in the degree between the polynomials of the numerator and denominator terms of the transfer functio n.
9.10.2 Input-Output Linearization

From the above for a system of relative degree v, one has y(v) = Lvf h + uLg Lvf- 1 h so that if the feedback, assuming Lg Lvf- 1 h ! 0 , u= 1 Lg Lv f- 1 h (r - Lvf h) (9.24) (9.23)
is introduced then one has y(v) = r (9.25)
a linear rela tio n ship b etween the inp ut r a nd the vt h d erivative of the o utp ut. Here Lvf h a nd L g L vf- 1 h a re f unctio ns of x and in particular if the la tter is zero for any value of the state x the required control signal becomes infinite. If the coordinate change z = { (x) is made with J z1 N J h N K OK O K . 2 O K Lf h O zO K K OK O K f O K zv O K Lv - 1 h O KO O = K O KKzv + 1O K arbitraryO OK K.O K O . K . OK . O KK OO KK OO L zn P LarbitraryP (9.26)
(9.27)
167
An Introduction to Nonlinearity in Control Systems Then the new system description will be of the form J zo 1 N J z2 N KO K O o KzOK z3 O KO K O . K .O K O K zv O Kp (z) + uh (z)O o KO=K O K o zv + 1O K }1 (z, u) O KO K O . K .O K O K zn O K }n - v (z, u) O o KK OO K O z1 LyPL P for some functions p, h and } . The linearizing feedback is u = (r - p) /h which gives the closed loop system J zo1 N J z2 N KO o O KK z3 Kz O O KO K . O K .O K O K zo v O K r O KO =K O K zov + 1O K }1 (z, (r - p) /h) O KO K . O K.O K O K zo n O K}n - v (z, (r - p) /h)O O KK OO K z1 LyP L P
(9.28)
(9.29)
(9.30)
Destination MMU
MMU is proud to be one of the most popular universities in the UK. Some 34,000 students from all parts of the globe select from its curricula of over 1, 000 courses and qualifications. We are based in the dynamic yet conveniently compact city of Manchester, located at the heart of a sophisticated transport network including a major international airport on the outskirts. Parts of the campus are acclaimed for their architectural style and date back over 150 years, in direct contrast to our teaching style which is thoroughly modern, innovative and forward-thinking. MMU offers undergraduate and postgraduate courses in the following subject areas: Art, Design & Performance Computing, Engineering & Technology Business & Management Scienc e, Envir onm ent al Studies & Geogr aphy Law, Education & Psychology Food, Hospitality, Tourism & Leisure Studies Humanities & Social Science 168 Please click the advert
For more details or an application form please contact MMU International. email: internatio nal@mmu.ac.uk telephone: +44 (0)161 247 1022 www.mmu.ac.uk/internatio nal
Thus the whole system has not been linearised and there is possibly some nonlinear dynamics affecting the states zv + 1, .........., zn . This dynamics is not visible in the outp ut and is called the zero dynamics of the system. If the above calculations are performed for a linear system then it is possible to show that the zero dynamics are the zeros of the transfer function from u to y.
9.10.3 Exact State Linearization

It will b e no ted fro m the a bo ve tha t the zero d yna mics disa p pear when v = n. In this case there is a state f eed b a ck which makes the system exactly linear in the state variables z1 = y, z2 = yo , ..........zn = y(n - 1) and the resultant system will be in the controllability companio n form given in equatio n (9.7) in z, that is Jz o1 N J z2 N KO K o z3 K z2 O K O O KO K . O K. O = K O K. O K . O KO K O K zon O KKp (z) + uh (z)O O KO z1 L LyP P and can be linearised with the feedback u = (r - p) /h and the nth line of the above equatio n is zon = r .
(9.31)
9.10.4 Examples
Two examples are now considered to illustrate the method. Example 1 Consider the nonlinear system xo 1 = n (x1) + x2 xo 2 = - x2 + u Obviously the inp ut- outp ut linearisation is dependent on the outp ut y which affects h in equation (9.27). First let the output be taken as y = x1 + x2 . Differentiation of y sho ws that the first d erivative is dependent on u so that the system is of relative degree 1. From equation (9.27) this mea ns y = z1 = x1 + x2 and z2 can be chosen arbitrarily. Choosing z2 = x1 gives zo1 = n (x1) + u = n (z2) + u zo2 = xo 1 = n (x1) + x2 = n (z1) + z1 - z2
169
Choo sing u = r - n (z 2 ) , gives zo 1 = r a nd the no nlinear zero d yna m ics is given b y the ab o ve eq uatio n fo r zo 2 a nd will b e stable if z1 is bounded. Now as a second choice consider the output to be y = x1 then differentiating gives yo = xo 1 = n (x1) + x2 and yp = n' (x1) xo 1 + xo 2 = n' (x1) (n (x1) + x2) - x2 + u where n' (x) = dn (x ) dx . Thus choosing u = r + x2 - x2 n' (x1) - n' (x1) n (x1) gives yp = r . In terms of z
z 1 = x1 z2 = xo 1 = n (x1) + x2 = zo 1 and zo2 = r if the feedback u is taken as u = r - z2 n' (z1) + z2 - n (z1) . More directly from equations (9.24) and (9.27) z1 = y = x1 and z2 = L f h = 1 uh1 + f2 u h2 = (n1 (x) + x2) = x f o u x1 u x2 Lg L f h = g1 L 2f h = f1 x) u 6 n (u1 x 1+ x 2 @ x1) u 6 n ( ux 1+ x 2 @ + f2 + g2 u 6 n (u1x +2x2 @ x)
1
= 0 + 1 = 1 and
u 6 n (ux 1+ 2x2 @ x)
= 6n (x1) + x2 @n' (x1) - x2 giving
u = r + x2 - x2 n' (x1) - n' (x1) n (x1) as before. Consider now tha t it is desired to place both poles for the closed loop response to z1 at -1 then choosing u = r - z1 - 2z2 - z2 n' (z1) + z2 - n (z1) gives 01 0 zo = e - 1 - 2 oz + c1m r which has the required characteristic equation of s2 + 2s + 1 = 0 In particular if n (x1) = - x13 then the feedback in terms of z is u = r - z1 - 2z2 + 3z2 z12 + z2 + z13 a nd Fig ure 9.10 sho ws the sim ulink dia gra m fo r the system . The f eed ba ck in the bo tto m p art of the dia g ra m co rresp o nd s to the last three term s in the ab o ve eq uatio n fo r u, which witho ut f urther terms wo uld m ak e zo 2 = r . There are two positive feed ba ck no nlinear term s in these three to m a ke the system linear between the inp ut r a nd z , a nd the o ther two term s are the linea r o nes to pla ce b oth po les a t -1. The linear respo nses to z1 (Sco p e) a nd z2 (Sco p e1) fo r a step inp ut a s expected corresp o nd to tho se of tra nsf er f unctio ns 1/ ( s + 1) 2 a nd s/ (s + 1) 2 , resp ectively a nd tha t to z1 is sho wn in Fig ure 9.11.
170
Figure 9.10 Simulink diagram for example 1

171
Figure 9.11 Input-output step response for example 1.
This ap pears to b e a ra ther co mp licated b ut interesting wa y o f obtaining a linea r inp ut -o utp ut respo nse b ut what ha pp ens if there a re m o d elling erro rs? To illustrate the pro blem s f or this pa rticular case let the a ctual no nlinea rity b e n (x 1 ) = ax 13 with 10% variations in a, that is with a equal to first 0.9 and then 1.1. Step responses are shown in Fig ure 9.12 for step seco nd the step m a g nitud e. Altho ug h three resp o nses are satisf a ctory the fo urth fo r a step inp ut ma g nitud e of 1.14 with a = 0.9 a ctua lly b eco m es unsta ble. This is no t surp rising in view of the stro ng p o sitive feed b a ck , wherea s the sta bility of the original system is no t a problem.
m ag nitudes of 1 a nd 1.14, resp ectively. The fo ur respo nses a re ma rk ed in the b oxes with the first num b er b eing , a, a nd the
Figure 9.12 Step responses for Example 1 with nonlinearity parameter changes.
172
An Introduction to Nonlinearity in Control Systems Example 2 Consider the third order system xo = x2 - ax1 xo 2 = - bx2 + n (x3) xo 3 = - cx3 + u y = x1 where n (x3) is any nonlinear element. This means x2 - ax1 h = x1 f = f - bx2 + n (x 3)p a nd - cx3 0 g = f 0 p, so that 1
L f h = f1 = x2 - ax1 L2f h = f1 (- a) + f2 = - ax2 + a2 x1 - bx2 + n (x3) L3f h = f1 (a2) + f2 (- (a + b)) + f3 n (x3) = - a x 1 + (a 2 + ab + b2 ) x2 - (a + b) n (x3 ) - cx 3 n' (x3) and Lg L2f h = n' (x3) Thus the required feedback to make zo n = r is u= 1 n' ( x 3)
"r + a3 x1 - (a2 + ab + b2) x2 + (a + b) n (x3) + cx3 n' (x3),
3
or in terms of z apart from n' (x3) u= 1 n' ( x 3)

"r + abz2 + (a + b) z3 + cx3 n' (x3),
Thus if the three poles for the linear system in z are all to be placed at -1 then the required u is given by u= 1 n' ( x 3)
"r + abz2 + (a + b) z3 + cx3 n' (x3) - z1 - 3z2 - 3z3, .
Here if n (x3 ) = x 33 then n' (x 3 ) need s to b e lim ited to a sm all value when x 3 = 0 to limit the value of the sig nal u. It is also clea r that the m etho d ca n ea sily b e used if the no nlinea rity is linear seg m ented a s its slo p e is k no w n. Fig ure 9.13 sho ws the implementation in Simulink for the case of a = 0 and b = c = 1 and the linear segmented no nlinearity n (x3) = x3 = 0.3x3 + 0.7 = 0.3x3 - 0.7 for x3 < 1 for x3 > 1 for x3 < - 1
The Sco pe blo ck s a re co nnected to the co m p o nents of z a nd the o utp ut of the pro d uct b lo ck is the sig nal u. The rela y elements connected in parallel are used to give n' (x3) .
173
Figure 9.13 Simulink implementation for Example 2
9.10.5 Further Comments

To sum marise, the procedure described in the previo us sections for a system with relative degree, v, eq ual to the dimension of the state x allo ws a no nlinea r tra nsf orm atio n to b e fo und fro m x to a new state z a nd a co ntrol sig nal u which will m ak e z linea r with resp ect to the new inp ut r. A f eedb a ck co ntroller ca n then b e d esig ned in term s of the orig inal sta te x, or z, to control the linear dynamics of z with respect to r. If the relative degree, v, is less than n the dynamics of some states of z, k no wn as the zero d yna m ics will rem ain no nlinear a nd need to b e check ed in ca se they a re unsta ble. The req uired all p oints in the st a te spa ce. The m etho d req uires the state x to b e a vaila ble so these values m ust either be m ea sured or estimated using a nonlinear observer. Perha ps the m ajo r a d va nta ge of the m ethod is that it allo ws a sta ble desig n of a no nlinear system b ut req uires a n accurate m a them a tical m o d el. W hen the assum ed m o d el do es no t a ccura tely m a tch the system it m a y b e difficult to assess th e effects it ma y ha ve a nd a s sho wn b y the f irst exa m ple the effects ma y b e undesirable. Also as mentio ned p revio usly it m a y not b e a pp ro p ria te to rem o ve the eff ects of the no nlinearity, in this case b y using co ntrol energ y, a s their existence m a y not be detrimental to the system performance.
control signal u is given in equation (9.24) and since its denominator is a function of the state x it may no t be finite at
174
9.11 General Comments

Severa l m etho d s f o r the desig n of co ntrollers f or no nlinea r system s ha ve b een b riefly co vered in this cha pter. There is a hug e literature o n m a ny o f the to pics so a few ap p ro priate ref erences ha ve b een cited b ut tho usa nd s mo re ca n b e f o un d b y p utting topics a nd k ey wo rds m entio ned in the ab o ve into G oo gle. Co m pa riso n o f the vario us a pp ro a ches is not ea sy excep t p erha p s in the ca se of o p tim al co ntrol where a solutio n is ob ta ined which m eets a certa in criterio n, a nd theref ore by definition is optimal for tha t case. However, even in this case, the criterion may have had to be perturbed from the true requirements in order to obtain a mathematical solution or, alternatively, it might be kno wn that the model used has def iciencies. A la rg e num b er of p ub lished p ap ers a re usually d evoted to a p a rticula r ap pro a ch no t to co m p ariso ns o f m ethod s whilst in p ra ctice the a p proa ch used will b e influ enced b y m a ny fa cto rs no t all of which will b e ea sy to d escribe analytically. Finally, one must again stress the importance of simula tion to support studies of nonlinear control systems no matter what ana lytical theories are used.
9.12 References
9.1 Glad, T & Ljung, L 2000, Control Theory: Multivariable and Nonlinear Methods, Taylor and Francis, London. 9.2 Friedland, B 1996, Ad vanced Control System Desig n, Prentice Hall, New Jersey. 9.3 Rugh, WJ 1991, Analytic Framework for Gain-Scheduling, IEEE Control Systems Magazine, Vol 11, pp 79-84. 9.4 Rugh, WJ & Sha mma, JS 2000, Research on gain scheduling, Auto matica, Vol 36, pp 1401 1425. 9.5 Murray-Smith, R & Johansen, TA, (Editors) 1997, Multiple Model Approaches to Modelling and Control, Taylor and Francis, London. 9.6 Bohn, C & Atherton, DP 1995, An Analysis Package Comparing PID Anti-Wind up Strategies. IEEE Control Systems Magazine, Vol 15, pp 34-40. 9.7 Nanka-Bruce, O and Atherton, DP 1990, Design of Nonlinear Controllers for Nonlinear Plants, IFAC Congress, Tallinn,Vol 6, pp 75-80. 9.8 Atherton, DP, Beno uarets, M & Nanka-Br uce, O 1993, Design of Nonlinear PID Controllers for Nonlinear Plants. Proc. IFAC World Congress, Sydney, Vol 3, pp 355-358. 9.9 McCausland, I 1969, Introductio n to Optimal Co ntrol, Wiley, New York. 9.10 Jacobs, OLR 1974, Introductio n to Control Theory, Chapter 9, Oxford University Press. 9.11 .Ryan, EP 1982, Optimal Relay and Saturating Control System Synthesis, Peter Peregrinus Ltd. Stevenage. 9.12 Athans, M & Falb, PL 1966, Optimal Control. McGraw-Hill, New York. 9.13 Lewis, FL & Syrmos, VL 1995, Optimal Control, John Wiley, New Yor k. 9.14 West, JC 1960, Analytical Techniques for Nonlinear Control Sys tems, Chapter 6, EUP London. 9.15 Emelyano v, SV 1967, Variable Structure Control, Nauka, Moscow. 9.16 Utkin, VI 1992, Sliding Modes in Control Optimization, Springer -Verlag, Berlin. 9.17 Zinober, ASI (Editor) 1990, Deterministic Control of Uncertain Systems, Peter Peregrinus, Stevenage. 9.18 Edwards, C & Spurgeon, SK 1998, Sliding Mode Control, Taylor and Francis, London. 9.19 Zadeh, LA 1965, Fuzzy Sets, Information and Co ntrol, Vol 8, pp 338353. 9.20 Mamdani, E H & Assilian, S 1975, Experiment in Linguistic Synthesis with A Fuzzy Logic Controller, Interna tional Journal of Man-Machine Studies, vol. 7, pp 1-13.
175
9.21 Schmid, Chr. 2004, Dynlab course at http://virtual.cvut.cz/d ynlabmods/syscontrol.pdf. Chapters 14-17. 9.22 Verbruggen, HB & Babuska, R (Editors) 1999, Fuzzy Logic Control; Advances in Applications. World Scientific Publis hing Co. 9.23 Warwick, K, Irwin, GW & Hunt, KJ (Editors) 1992, Neural Networks for Control and Systems, Peter Peregrinus, Stevenage. 9.24 Irwin, G W, Wa rwick , K & Hunt KJ (E dito rs) 1995, Neura l Netwo rk Ap plicatio ns in Co ntrol, Peter Perig rinus, Stevena g e. 9.25. Isidori, A 1995, Nonlinear Control Systems; An Introduction, 3rd E dition, Springer Verlag, New York. 9.26 Slotine, J-J & Li, W 1990, Applied Nonlinear Control, Prentice Hall, New Jersey. 9.27 Corriou, J-P 2004, Process Control; Theory and Applications, Chapter 17, Springer Verlag London.

INNOVATION, FLAT HIERARCHIES AND OPEN-MINDED PROFESSORS

CLOSE COLLABORATION WITH FUTURE EMPLOYERS SUC H AS ABB AND ERICSSON
SASH A SHAHB AZI LE FT IRAN FOR A MASTERS IN P RODUCT AND P ROCESS DEVE LOPMEN T AND LO TS OF INNE BANDY
HEL L TE LL YOU A LL ABO UT IT AN D ANSWER YOUR QUESTIO NS AT MDUSTUDENT.COM
www.mdh.se
176

An Introduction To Non Linearity in Control System

Caricato da

Informazioni sul documento

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

An Introduction To Non Linearity in Control System

Caricato da

Copyright:

Formati disponibili

An Introduction to Nonlinearity in Control Systems