Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Abstract
This paper uses the Popov-Tsypkin multiplier (which has intiinate connections to mixed structured singular value theory) to design robust H , estimators for uncertain. linear discrete-time systems and considers the applicat ion of robust H , estimators to robust fault detection. The key to estiniator-based, robust fault detection is to generate iesiduals which are robust against plant uncertainties and external disturbance inputs, which in turn requires the design of robust estimators. The robust H , estimation pioldem is formulated as a Riccati equation feasibility problem in which a cost function is minimized snbjert to a Riccati equation constraint. The robust H , estimator framework is then applied to the robust fault detection of dynamic systems.
1. Introduction
Fault det.ect,ionof dynamic systems has been an act.iw research area in recent years [l,2, 5, 61. The key step in est.imator-based fault detection methods is to generate residuals which are accent.uated by faults. These residuals are t lien (.ompared wit.11 some threshold values to determine whether faults have occurred. Logically, the existence of uncertainties and disturbance inputs (i.e., plant distiirlmices and measurement noise) obscures the effect of faults and is therefore a source of false alarms. In order to reduce false alarm rates and improve fault detection accuracy, the residuals generated should be robust against uncc3rtainties and disturbance inputs. The residuals used in fault ctetection a.re generated by coniparing the actual measnrrmrnts of the plant, with t.he corresponding estimated cliiantities which are obtained by estimation. The requireiiient*of robust residual generation naturally leads t,o the proldetti of robust estimation which is the key factor in giiaranteeiiig robust fault detection. Diffcwiit estimator-based residual generation apITliis rewarrh \vas supported in part by the National Science Foundation under Grant CMS-9802197 and the National Atwmaiitics and Space -4dministration under Grant NAG32193.
proaches can be used and the widely used methods include the parity space approach [6],the fault detection filter approach [l],and the unknown input observer approach [l]. The motivation behind each of these approaches is to distinguish the effects of faults from those caused by plant uncertainty and external disturbances and thus to achieve robust fault detection. Each of these approaches seeks either to completely decouple the effects of uncertainties and disturbance inputs from those caused by faiilts or to minimize certain norms of the transfer function matrix from disturbance inputs to the residual signals. A necessary condition for the complete decoupling of disturbances from the residual signals is that the number of faults not exceed the number of measurements which limits the application scope of the fault detection techniques. Another approach [5] is to model the uncertainties as complex uncertainties with bounded magnitude and to solve the robust fault detection problem by using the small gain theorem. As discussed below, this approach may lead to conservative results. It is now well known that robust design of fixed quadratic Lyapunov functions is intimately related to the small gain theorem and assumes that the uncertainty is arbitrarily time-varying or complex. This assumption can lead to very conservative designs for systems with timeinvariant, parametric uncertainty. However, mixed structured singular value theory and the associated multiplier theory [3, 7 1 are based on parameter-dependent Lyapunov functions and can lead to much less conservative results. In this paper, the problem of robust H , estimation will be studied by using multiplier theory. In particular, by using the Popov-Tsypkin multiplier [3, 71, the robust H , problem is formulated as a Riccati equation feasibility problem in which a cost function is minimized subject to a Riccati equation constraint. A continuation algorithm that uses quasi-Newton (BFGS) corrections is developed to solve the minimization problem. The algorithm is initialized with an H , estimator obtained by solving a single algebraic Riccati equation corresponding to the nominal system. The initializing multiplier matrices are obtained by solving a linear matrix inequality (LMI) [3].
4408
where
real, complex numbers m x n real matrices, complex matrices 71 x n real diagonal, symmetric matrices 71 x 11 nonnegative definite matrices 71 x 71 positfivedefinite matrices A I , - Ad1 positive definite AI2 - Afl nonnegative definite dimension of A f
~ L ~ P o,,,,(G(eje)) ~ ~ [ ~ . ~ ~ ]
where r,, E RP is the state vector, yp E R p p denotes the , disturbance plant measurements, w E Rd denotes an C signal. atid the uncertainties AA, and AC, satisfy
C O =
and
[ -2 :],
bfi = diag(MAp,l, Mcp,i),M2 = diag(MAp,2,MC,,z). (17) The Robust H , Estimation Problem. Find A, and W such that 1. the augmented system (10) is asymptotically stable over the uncertainty set U A ;and
2. the H , performance satisfies
AAEU~
SUP
IlGzw(z)llcc <
1
-1
(18)
.r,(k
+ Ilk) = A,z,(klk
- 1)
+ Wy,(k)
(7)
where y > 0, and G,,(z) E C q x d is the transfer function matrix from disturbance signal w(-) to the performance variable z(.).
to pstiniate (in some sense to be defined) the state vector i f ]wlieie . A, E R n p X n p and W E R n p x p p are the filter parairieteis to be determined. Define the estimation error as
3. Robust H , Estimation
This section begins by restating a Riccati equation robust stability condition for uncertain linear, discrete-time systems based on the Popov-Tsypkin multiplier. These results are then used to formulate the robust H , problem as a Riccati equation feasibility problem which is solved by a continuation algorithm that uses quasi-Newton (BFGS) corrections.
e(k)
(8)
whirh wing ( l ) , (2), and (7) can be shown to obey the evolution equation
e(k
+ 1)
= A,e(k)
Stability
Conditions
Consider the standard uncertainty feedback configuration given in Figure 1, where G ( z ) E C is square, asymptotically stable and G ( z ) that the uncertainty A satisfies
N
Next, define the error output z E Rq as z ( k ) 2 E,e(k). Then augmenting (1) with ( 9 ) yields
.r(k
[%I.
Assume
+ 1)
(10) (11)
2 ( k ) = E+),
A E U 2 {A E R m x m : MI 5 A 5 M2},
(19)
4409
re23
w 2
> 0 and N 2 0.
Theorem 1 [3] Assume G ( z ) is asymptotically stable. If there ezist real diagonal H > 0 and N 2 0 , P > 0 and E > 0 sirch that D , ( H , N ) + D , ( H , N ) T - B T P B , > 0 and
In order to account for the performance block A p ,the multiplier matrices H and N are redefined as
H
(27) (28)
= ATPA,
where H I E VD('+s)x('+s), H2 E R q x q satisfies H 2 A p = A p H 2 , and N I E V('+s)x('+g). The next theorem forms the theoretical basis for robust H , estimator synthesis.
7 t h ere
Theorem 2 Assume G(z) is asymptotically stable. If there exist real diagonal H > 0 and N 2 0 , P > 0 and E > 0 such that D , ( H , N ) D , ( H , N ) T - B Z P B , > 0 and
C , ( H . N )= [ ( H
(24) the71 the negatzve feedback znterconnectzon of G(z) and A i s nsynipfotically stable for all A E U.
+ N ) ( M z - M1)C D,(H,N) = H + N ,
1,
(23)
= A;fPA,
+ [ B T P A , - C , ( H , N)IT . ( D a ( H ,N ) + D , ( H , N ) T - B:PB,)-l
(29)
[ B T P A , - C a ( H ,N ) ] + I , where
G(Z)
[e] [y]
&?
C O ' (25)
L I
d
Note that matrices A , D , E , Bo, and COare as defined in equations (12), (13) and (16). To consider H , performance, a fictitious complex unrertainty block Ap is inserted into Fig. 2 [8] as shown in ( y. A For ~ ) ease of preFig. 3. It is assumed that O ~ , ~ , ~< sentation. assume that dim(z) = dim(w) = Q, such that A, E P U q Define .
(32) then the uncertain system of Fig. 3 is asymptotically stable addition, for each AA E ,yA.
SUP IIGrw(z)ll, AAEU~
C a ( H ,N ) = [ ( H + N ) ( k 2 - Kf1)C N ] , D,(H, N ) = H N ,
(31)
< -.
Y
(33)
M I = hlorlr-diag{AI1, --yIq}.
A I ,
Proof. Follows from Theorem 1 and the main loop theorem [SI. 0 Theorem 2 poses the robust H , estimation problem block-diag{Mz, -71,). as a Riccati Equation Feasibility Problem (REFP). As discussed in Reference [4], an approach to solving the REFP (26)
441 0
J ( A , , TI')
(34)
Hrre tlw robust H , performance J ( . ) can be chosen as the aititicial cost, function [4]
where 11 . IIp denotes the p norm of a Lebesgue signal and J t h is a threshold value. The residual generated is given by the following equation if estimator (7) is applied to the system described by (38) and (39).
r(k) =
=
(IT) ~ trP. 4
(35)
(42)
+ Rp,zf(k),
Auxiliary minimization problem. The robust H , estimation 111 oblem can be cast as an auxiliary minimization ptoblmi in which (35) is minimized subject to the constraint represented by equation (29). To (*hac'terize the extremals define the Lagrangian
C(F..~,.TT-,H,N,P,Q) = J ( A P , W ) trQ[-P the R.H.S. of (29)]. (36)
Thus the necessary conditions for a solution to (34) are giveri l)r
r ( k ) = 4 k ) + DP,ZW(k) + Rp,Sf(k),
(44)
0c -
- =
oc
36
0,
0.
- - O, all..
dC
ac = 0, dL = 0,
aAe
ON
8.C ac ap = 0, aQ = O .
8H
(37)
(45) where [NO, NI corresponds to a certain time interval. Note that similar norm conditions used for fault detection can be found in Reference [5]. If f ( k ) = 0, equation (45) reduces to
Note that
$$
(47)
and
(48)
(Ap AAp)ap(k) Dp,rw(k) + RP,lf(k), (38) !/,>(k)= CPZP(k) + D,,aw(k) + Rp,2f(k) (39)
=
+ 1)
where rT,.!I!>. and IU are as discussed in previous sections, and f E R"' is the fault vector. The term Rp,lf(k) represents actuator and component faults while Rp,2f(k) denotes tlie sensor faults. The fault distribution matrices R , 1 and R, 2 are assumed to be known. The robust fault detection problem is to generate a robust rpsidual signal r ( k ) that satisfies
11r(k)llp
5. Illustrative Example
The example used is a linearized discrete-time model of a simplied longitudinal flight control system [2] which is is given in the form of (1) and (2) and the elements of the state variable vector z ( k ) [ qy w, 6 , are normal
Jth Jfh
Ilr(k)llp
>
if if
f(k) = 0, f(k) # 0,
(40) (41)
1'
441 1
velocity vu, pitch rate w z , and pitch angle 6,. The system parameter matrices are
6. Conclusion
In this paper, the problem of robust H , estimation for uncertain, linear discrete-time systems and its applications to the robust fault detection of dynamic systems has , problem was then reforbeen addressed. The robust H mulated as a Riccati equation feasibility problem in which an artificial cost is minimized subject to the constraint of an algebraic Riccati equation. The robust H , estimator framework was then used in estimator-based fault detection of dynamic systems. By considering a flight longitudinal system, it was shown that the robust fault detection methology based on the robust H , estimation framework is capable of significantly reducing false alarm rates.
0.8950 -0.1083 -0.3872 0.0015 0.8912 -0.0672 , Cp = 1 3 x 3 , 0 0.7368 0 Dp.l = rliag(0.1, 0.1, 0.01}, D,,z = 0.1 x 1 ~ ~ 3 .
A,,
Bilp =
CA,
66).
Fs,,
References
[l] Chen, J., Patton, R. J., and Zhang, H., Design of Unknown Input Observers and Robust Fault Detection Filters, International Journal of Control, Vol. 63, No. 1, 1996, pp. 85-105.
[2] Chen, J . and Patton, R. J., Optimal Filtering and Robust Fault Diagnosis of Stochastic Systems with Unknown Disturbances, IEE Proceedings-Control Theory and Applications, Vol. 143, 1996, pp. 31-36. [3] Collins, E. G. Jr., Haddad, W. M., Chellaboina, V., and Song, T., Robustness Analysis in the Delta-Domain Using Fixed-Structured Multipliers, Proceedings of the IEEE Conference on Decision and Control, San Diego, CA, 1997, pp. 3286-3291. [4] Collins, E. G. Jr., Haddad, W. M., Watson, L. T., and Sadhukhan, D., Probability-One Homotopy Algorithm for Robust Controller Synthesis with FixedStructure Multipliers, International Journal of Robust and Nonlinear Control, Vol. 7, 1997, pp. 165-185.
The iinrei tain parameters 61 through 6s correspond to 33% parameter fluctuations in the first two rows of matrix
A,.
It is desired to design a predictive filter of the form (7) for which the estimation error is given by (9). For this particular example, the error output is defined as z ( k ) = E,e(k) where Ep 1 3 x 3 = C, which satisfies the residual equation (44). The sampling period used is 0.01 sec. and N - No = 100. It is also assumed that through experimentation using additional sensors, it was found that I I U J / ~ ~ , [ N ~ , N I < 0.7377. In order to illustrate the application of the robust H , estimator to the robust fault detection of the system, a sensor fault is added to the velocity sensor at the time instant t = 6 sec. Specifically, it is assumed that the faulty sensors reading is 0.5 times the actual system velocity. It can be seen from Figure 4 that both the nominal H , estimator and the robust H , estimator can detect this fault since both residuals surpassed the threshold value. But it shoiild be noted that the residual signal generated by the nominal H , estimator tends to give false alarms since it surpassed the threhold values at two time intervals before t = G sec. even though there is no fault in the system in the time interval [O,G sec.). This false alarm resulted from the fact that the nominal estimator cannot distinguish the fault effect from the effects caused by system unrertaint ies.
[5] Frank, P. and Ding, X., Frequency Domain Approa.ch to Optimally Robust Residual Genera.tion and Evaluation for Model-Based Fault Diagnosis, Automatica, Vol. 30, No. 5, 1994, pp. 789-804. [6] Gertler, J. J., Fault Detection and Isolation Using Parity Relations, Control Engineering Practice, Vol. 5, NO. 5, 1997, pp. 653-661. [7] Kapila, V. and Haddad, W. M., A Multivariable Extension of the Tsypkin Criterion Using a Lyapunov Approach , IEEE Transactions on Automatic Control, Vol. 41, 1996, pp. 149-152. [8] Packard, A. and Doyle, J . C., The Complex Structured Singular Value, Automatica, Vol. 29, 1993, pp. 79109.
4412