Robust HOutput Feedback Control of Networked Control Systems with Dynamic Quantizers

Faiz Rasool , Farrukh A. Bhatti , Saeed Ur Rehman , Tariq Shahzad

Electrical and Computer Engineering Department, University of Auckland, New Zealand

COMSATS Institute of Information Technology Sahiwal Campus, Sahiwal, Pakistan

Email: {f ras004, sreh008, f bha004} @aucklanduni.ac.nz, tariq @ciitsahiwal.edu.pk

AbstractIn this paper a robust H dynamic output feedback control is designed for Networked Control Systems (NCSs). A new dynamic quantization technique, based on the difference of signals, is used to quantize data before transmitting it over the network. Dynamic quantizers are assumed both on system input and output. Network induced random delays are modeled using Markov chain and a stochastic stability criteria is proposed using a detailed Lyapunov-Krasovskii functional. On the basis of stability conditions, an innovative control law is proposed in terms of Bilinear Matrix Inequalities (BMIs). At the end, a numerical example is discussed and it is observed that the proposed technique assures stability for systems with fast dynamics.

treated as stochastic processes, modeled by a Markov chain which takes values from a transition probability matrix [16]. By using this method, TDSs and NCSs can be represented as Markovian Jump systems (MJSs) which is another important research eld [14]. MJSs are used to model systems with abrupt changes. In the second approach, packet dropouts and delays can be modeled by a random variable following the Bernoulli random sequence [15]. The use of digital networks and computers in control systems introduces signal quantization because of their nite precision. In the beginning, research was done to minimize the effects of quantization errors on the system performance [1], [2]. However, modern research shows that quantizers are information coders [4]. Here, interest is that how much information is required for system stability? The minimum information required for system stability depends on the quantization density and unstable poles of the plant. It is already proven that logarithmic quantizers assure quadratic system stability with the coarsest quantization density [4]. The quantized feedback control problem can be classied into 1) quantized feedback with static quantizers, and 2) quantized feedback with dynamic quantizers. The rst approach assumes that the quantization value at any instant in time, k , does not depend on previous quantization values [4], [2], [5]. The latter approach considers dynamic quantizers with an internal state which helps to enlarge the region of attraction and reduce limit cycles [6]. In this paper, a new type of dynamic quantizers is investigated.

I. I NTRODUCTION Distributed control systems in which control loops are closed over real time networks are called networked control systems (NCSs). Rapid development in communication systems, along with advancements in control system design, has greatly enhanced modularity, exibility and reduced processing cost in NCSs. This has made them highly popular in industry. However, the incorporation of a network into control loop results in various design constraints; such as time delays and packet dropouts due to limited bandwidth and quantization errors due to the hybrid nature of NCSs. All of these constraints can be compensated for either by designing dedicated communication protocols or by developing efcient control strategies for NCSs. In NCSs, time delays are typically random due to network congestion and the number of nodes and can be modeled together with packet dropouts. To model these constraints, two lines of research are available in the literature 1) Markov chains, and 2) Bernoulli random sequences. In the rst approach, delays are

Logarithmic quantizer are used to quantize difference of current x(k ) and previous (x(k-1) feedback information. After quantization, the information is sent over the network to observer based output feedback controller which calculates control signal for system stability, with given attenuation level , using well known Lyapunov Krasovskii functional. Sufcient conditions for the existence of an output feedback controller is formulated in terms of bilinear linear matrix inequalities (BMIs). Furthermore, an iterative algorithm is used to convert the BMIs into quasi-convex optimization problem which can be solved easily. Numerical examples for 1) plants with slow changing dynamics and 2) plant with fast changing dynamics are provided. It is observed that the technique assures system stability for fast dynamic systems while it is not applicable for the other type. The main contributions of the paper can be summarized as following:

described by the following model: x(k + 1) z (k ) y (k ) = = [A + A(k )]x(k ) + [B1 + B1 (k )]w(k ) + [B2 + B2 (k )]u(k ), x(0) = 0 [C1 + C1 (k )]x(k ) + [D11 + D11 (k )]w(k ) + [D12 + D12 (k )]u(k ) = C2 x(k )

A robust H output feedback controller is designed for NCSs with random delays and dynamic quantizers. Quantizers are assumed both on system input and outputs Design is applied to multiple Control systems and it is observed that it is a suitable technique to assure stochastic stability of fast dynamic systems.
Fig. 1. Networked control systems with sensor-to-controller delay

where x(k ) n , u(k ) m , z (k ) m1 , y (k ) m2 are the state, input, controlled output and measured output, respectively. w(k ) m3 is the disturbance which belongs to L2 [0, ), the space of square summable vector sequence over [0, ). The matrices A, B1 , B2 , C1 , D11 , D12 and C2 are of known matrices with appropriate dimensions. The matrix functions A(k ), B1 (k ), B2 (k ), C1 (k ), D11 (k ) and D12 (k ) represent the time-varying uncertainties in the system which satisfy the following assumption. Assumption 2.1: [ A(k ) B1 (k ) C1 (k ) where Ei = [ Hi = D11 (k ) [ ] = Ei F (k )Hi (II.1)

The rest of the paper is organized as follows. Section II provides system description, modeling of delays, quantizer description and problem formulation. In section III, main results for stability analysis and synthesis of a quantized robust H output feedback controller design are formulated in terms of BMIs. A cone complementarity algorithm is employed to convert these BMIs into quasi-convex optimization problems. Section IV comprises an example in which design is investigated for systems with different dynamics. Finally, the conclusions are drawn in section V.

B2 (k ) D12 (k ) ]

E1 E2 H1

II. S YSTEM D ESCRIPTION AND D EFINITIONS

] H2 H3 (II.2)

Ei and Hi are given matrices which characterize uncertainties. Consider the system framework shown in Figure 1. A class of uncertain discrete-time linear systems under consideration is 2 Furthermore, there exists a positive denite matrix W such that the following inequality holds:

where F T (k )W F (k ) W (II.3) x (k ) =

[ [

x T 1 (k ) A 1 [ 0 0 E1 0 H1 B1 0 B2 0 C1 C2

x T 2 (k )

]T

Let {rk , k } be a discrete homogeneous Markov chain taking values in a nite set S = {1, 2, , s}, with the following transition probability from mode i at k to mode j at time k + 1

= A

] ] 0 ] ] ] 0 C2 ] (II.7)

1 E pij := Prob{rk+1 = j |rk = i} where i, j S . In this paper, the random delays k is modeled by a nite state Markov process as k = (rk ) with 0 (1) < (2) < < (s) . We assume that the controller will always use the most recent data, that is, if there is no new information coming at step k + 1 (data could be lost or there is a longer delay), then x(k k ) will be used for feedback. Thus the delay k can only increase at most by 1 at each step, and we constrain 1 C 2 C 2 B 1 H 1 B

= [ = = [ = [ = = [ [

Consider a quantized dynamic output feedback controller: x (k + 1) = Ac (i) x(k ) + Bc (i)q1 ( y (k (k ))) Cc (i) x(k ) q2 (v (k )) (II.8)

Prob{k+1 > k + 1} = 0 Hence, the structured transition probability matrix [10] is

P = p11 p21 . . . . . . p s1 p12 p22 . . . . . . ps2 0 p23 . . . . . . ps3 j =1 0 0 . . . . . . p s4 ... ... .. . 0 0 . . . p(s1)s pss

v (k ) = u(k ) =

where x (k ) is the controller state; Ac (i), Bc (i) and Cc (i) are the
(II.4)

controller matrices: Quantizer Description: This paper considers a logarithmic static and time-invariant quantizer as follows[5]: 1 1 i i i if 1+ < 1 , > 0, i = 0, 1, 2, q ( ) = 0, if = 0 q ( ) if < 0

. . . ...

with 0 pij 1 and

i+1

pij = 1. A quantizer is used as

an information coder in digital and networked control systems [4]. A good information coder should use least number of bits to represent any information to transmit over the channels with limited capacity. One way of achieving this capability is to quantize the difference of two consecutive signals instead of the current signal. Consider: x 1 (k ) x 2 (k ) Hence
x (k + 1) = 1 ) 1 + +E 1 Fk H x(k) + (B (A 1 Fk H2 )w(k) + (B 2 + E 1 Fk H3 )u(k), E z (k ) = 1 ) 1 + E2 Fk H x(k) + (D11 + (C E2 Fk H2 )w(k) + (D12 + E2 Fk H3 )u(k), y (k) = 2 x C (k) (II.6)

(II.9)

where 0 < < 1 is the quantization density of q (), and is = = x(k ) x(k 1) (II.5) related to by = 1 1+ (II.10)

The associated quantized set U is given by { } U = i , i = 0, 1, 2, {0}. Now dene the quantization error as e(k ) = f (v (k )) v (k ) = q (k )v (k ), 3

(II.11)

(II.12)

then, it has been shown in [5] that q (k ) [, ]. Applying controller (II.8) on system (II.6): ( (k + 1) = ) ( 1 F (k )Hcl (i) (k ) + B 1 + Acl (i) + E ) 1 F (k )H2 w(k ) + Bcl (i)(I + E 2 ( (k (k ))) 1 (k ))C z (k ) = Ccl (i) + E2 F (k )Hcl (i) (k ) + (D11 +E2 Fk H2 )w(k ) where [ (k ) = [ [ 1 E Hcl (i) 1 B Bcl (i) 2 C Ccl (i) = [ = = [ = [ = = [ [ x T (k ) x T (k ) ]T ] (II.13)

The following lemma plays an important role in the derivation of the main results. (k ) = Lemma 2.1: [12] Let x (k ) = x (k + 1) x (k ) and [ T T (k ) T (k (k )) wT (k ) T (k )Hcl (i)F T (k ) ]T T q T (k ) wT (k )H T F T (k ) l , then for any T (k (k ))C 2 2 matrices R nn , M nl and Z ll satisfying [ ] R M 0 (II.16) MT Z the following inequality holds
k 1 i=k (rk )

{ } T (k ) 1 + T + (k ) (II.17) x T (i)Rx (i) Z 1

where 1 = M T [diag{I, 0} diag{I, 0} 0 0 0 0]. Proof is omitted due to space limitations . III. S TABILITY A NALYSIS AND S YNTHESIS OF NCS S

Acl (i) =

B 2 (I + 2 (k ))Cc (i) A 0 1 E 0 1 H 1 B 0 0 Bc (i) 2 C 1 C 0 D12 (I + 2 (k ))Cc H3 (I + 2 (k ))Cc (i) ] ] ] Ac (i) ]

Stability criteria for uncertain discrete-time systems with random communication delays and logarithmic quantizers both on controller inputs and outputs are given in the following theorem. Theorem 3.1: For given controller matrices Ac (i), Bc (i) and Cc (i) where i = 1, , s, quantization densities 1 > 0, 2 > 0 and 2 > 0, if there exist sets of positive-denite matrices P (i), R1 (i), R1 , R2 (i), R2 , W1 (i), W2 (i),W3 (i), Q, Z (i) and matrices

M (i) satisfying the following inequalities [ R1 > R1 (i), R2 > R2 (i) (1 pi(i+1) )R1 (i) + R2 (i) M (i) Z (i) ] 0 (III.1) (III.2)

The problem under our study is formulated as follows. Problem Formulation: Given a prescribed > 0, design a dynamic output feedback controller of the form (II.8) such that 1) the system (II.13) with (II.8) and w(k ) = 0 is stochastically stable, i.e, there exists a constant 0 < < such that { } T E () () < (II.14)
=0

T (i)R1 + (s)R2 ] (i) + T 1 (i)P (i)1 (i) + 2 (i) [ T 2 (i) + 1 (i) + T 1 (i) + (i)Z (i) + (i)(i) < 0

(III.3) Then the closed-loop system is stochastically stable with the prescribed H performance. Proof: is omitted due to space limitations The following theorem provides procedures for designing a

for all (0), r0 . 2) Under the zero-initial condition, the controlled output z (k ) satises { E

} z T (k )z (k )|r0 < 2

k=0

wT (k )w(k )

(II.15)

quantized output feedback controller. Theorem 3.2: For given > 0 and quantization densities 1 > 0, 2 > 0, if there exist positive symmetric matrices X (i) > 0, 4

k=0

for all nonzero w(k ).

 (i) satisfying the following matrices A(i), B (i), C (i), J (i), M inequalities for all i, j S

Time Delay Modes

Y (i) > 0, Y (i) > 0, W1 (i), W1 (i), W2 (i), W2 (i), W3 (i), Q, (i) and Q, N (i), N (i), R1 , R1 , R1 (i), R2 , R2 , R2 (i), S (i, j ), Z

2 1.5 1 0.5 0 0.5

R1 > R1 (i), R2 > R2 (i) [ ] (i) N (i) M 0 (i) Z [ [ ] >0 TT N (i) ] >0

(III.4)
1 0 50 100 150 200 250

Time

(III.5)
Fig. 2. Random Modes

S (i, j )

J T (i) Y (j )

We assume same uncertainty matrices for both system for sim(III.7) plicity and analysis purposes: E1 = .002, E2 = 0.005, H1 = .05, H2 = 0.004, H3 = 0.007. where F (k ) is a random variable follows given assumption.

(1 pi(i+1) )R1 (i) + R2 (i)

(III.8)

N (i)N (i) = Y (i)Y (i) =

I, R1 R1 = I, R2 R2 = I, QQ = I, 1 (i)W1 (i) = I, W 2 (i)W2 (i) = I I, W (III.9)

Delays are modeled using Markov chain which follows transition probability matrix: [ P = 0.4206 0.500 0.5794 0.500 ] (IV.3)

Then the closed-loop system is stochastically stable with the prescribed H performance. Furthermore, a suitable controller is given as follows ) 1 ( i+1 1 (i) p Y ( j ) X Ac (i) = ij ( j =1 ) (i)(AY (i) + B2 (I + 2 )C (i)) Y 1 (i) A(i) X ( ) 1 i+1 1 (i) Bc (i) = (j ) X B(i) j =1 pij Y Cc (i) = C (i)Y
1

This transition probability matrix is obtained after performing series of experiments on Vodafone wireless network. Modes of Markov chain are show in Figure 2 Controller gains are obtained after implementing Theorem 3.2 in LMI ToolBox. Gains for Type 2 systems are: [ Ac (1) = [ Cc (1) = ] , Bc (1) = ] , ] , Bc (2) = ] , [ ] , [ ] ,

(i).

4.4733 3.6675

0.4497 0.2815

0.2169 0.1821

Proof is omitted due to space limitations IV. S IMULATION E XAMPLE In this example we consider two different systems: 1) system with slow dynamics and 2) system with fast dynamics. example of rst type of system: x(k + 1) = z (k ) y (k ) = = x(k ) + w(k ) + 0.5u(k ), x(0) = 0 0.7x(k ) + 0.1w(k ) + 0.15u(k ) 0.5C2 x(k ) (IV.1)

0.0636 0.0694

[ Ac (2) = [ Cc (2) =

4.1733 1.6671

1.4127 0.0815

0.2231 0.1111

0.2636 0.0621

It can be seen that controller gains are double in dimensions due to augmented system given in Equation II.6. This controller assures stability, in stochastic terms, for second type of systems as shown in Figure 3. In Figure 4, it can be seen that design technique does not work for second type of systems. Controller gains for second type of systems are omitted due to space

while second system: x(k + 1) z (k ) y (k ) = 1.09x(k ) + 0.5w(k ) + 1.2u(k ), x(0) = 0 = 0.75x(k ) + 0.5w(k ) + 1.5u(k ) (IV.2) 5 = 2C2 x(k )

limitations.

(i) i) + (i) + T (i) + (i)Z ( 1 1

T (i) 1 i) (

T (i) 2 0 R

T (i) 3 0 0 I

T (i) 4 0 0 0 Q

T (i) 5 0 0 0 0 (i) W 1

T (i) 6 0 0 0 0 0 (i) W 2

< 0

(III.6)

0.012 0.01

0.008 0.006 0.004 0.002 0 0

50

100

150

200

250

Time

Fig. 3.
x 10
5

System state of Type 2 systems

4 3

State x

2 1 0 1 0

10

15

Time

Fig. 4.

System state of Type 1 systems

V. C ONCLUSIONS A new technique, based on quantization of difference of signals, to implement dynamic quantizers for NCSs is developed. An observer based output feedback controller with dynamic quantizers on both systems input and output is developed and implemented. This controller design compensates networked induced delays modeled by Markov chain. Stability criteria is provided using a detailed Lyapunov-Krasovskii functional. On the basis of stability criteria, controller gains are obtained in term of BMIs. A cone complementarity algorithm is used to converts these BMIs into quasi convex LMIs. Control design is supported and elaborated further with a numerical example. R EFERENCES
[1] R. Kalman , Nonlinear aspects of sampled data control systems, In proceedings of the symposium on nonlinear circuit theory, Brooklyn NY, 1956.

[2] D. Delchamps,Stabilizing a linear system with quantized state feedback, IEEE Transactions on Automatic Control, vol. 35, pp. 916-924, 1990. [3] R. Brockett and D. Liberzon, Quantized Feedback Stabilization of Linear Systems, IEEE Transactions on Automatic Control, vol. 45, pp. 12791289, 2000. [4] N. Elia, and S.K. Mitter, Stabilization of linear systems with limited information , IEEE Transactions on Automatic Control, vol. 46, pp. 13841400, 2001. [5] M. Fu and L. Xie, The sector bound approach to quantized feedback control, IEEE Transactions on Automatic Control, vol. 50, no. 11, pp. 16981711, 2005. [6] G. Nair, R. Evans, Exponential Stabilization of multi dimensional Linear Systems, Automatica, vol. 35, pp. 916-924, 2003. [7] S. Tatikonda and S. Mitter, Control under communication constraints ,IEEE Transactions on Automatic Control, vol. 49, pp. 1056-1068, 2004. [8] M. Fu and L. Xie Finite-level Quantized Feedback Control For Linear Systems, 45th IEEE Conference on Decision and Control, pp. 1117 1122, 2006. [9] F. Rasool and S. K. Nguang, Quantized robust H control of discretetime systems with random communication delays, in International Journal of Systems Science, vol. 42, no. 1, pp. 129138 , 2011 [10] L. Xiao, A. Hassibi and J. How, Control with random communication delays via a discrete-time jump system approach, Proceedings of American Control Conference, vol. 3, pp. 21992204, 2000. [11] A. P. C. Goncalves, A. R. Fioravanti and J. C. Geromel, Dynamic Output Feedback H Control of Discrete-Time Markov Jump Linear Systems Through Linear Matrix Inequalities, Proceedings of the 47th IEEE Conference on Decision and Control, Cancun, Mexico, pp. 47874792, 2008. [12] S. Chae, F. Rasool, S K. Nguang and A. Swain, Robust mode delaydependent H control of discrete-time systems with random communication delays, IET Control Theory Applications, vol. 04, no. 6, pp. 936944, 2009. [13] L. Ghaoui, F. Oustry and M. AitRami, A cone complementarity linearization algorithm for static output-feedback and related problems, IEEE Transactions on Automatic Control, vol. 42, no. 8, pp. 11711176, 1997. [14] J. Li, Q. Zhang and Y. Xie, Robust H Control of uncertain networked control systems with dropout compensation and markov jumping parameters, Procedding of the 7th world congress on intellegent control and automation, 2008. [15] F. Yang, Z. Wang, Y. Hung, and M. Gani, Control for networked systems with random communication delays, IEEE Transactions on Automatic Control, vol. 51, no. 3, pp. 511518, 2006. [16] D. Huang and S. K. Nguang, State Feedback Control of Uncertain Networked Control Systems With Random Time Delays, IEEE Transactions on Automatic Control, vol. 53, no. 3, pp.829-834, 2008.

System State