You are on page 1of 5

2016 UKACC 11th International Conference on Control (CONTROL)

Belfast, UK, 31st August - 2nd September, 2016

Improved minimum entropy control for two-input


and two-output networked control systems

Jianhua Zhang Yamin Kuai, Shuqing Zhou, Guolian Hou Mifeng Ren
State Key Laboratory of Alternate School of Control and Computer College of Information Engineering
Electrical Power System with Renewable Engineering Taiyuan University of Technology
Energy Sources North China Electric Power Taiyuan, China
Beijing, China University
e-mail: zjh@ncepu.edu.cn Beijing, China

Abstract—In this paper, the problem of control algorithm presented under the stochastic framework. NCSs were
design for a class of nonlinear two-input and two-output transformed to Markov jump linear systems, the control
(TITO) networked control systems (NCSs) with non-Gaussian algorithms were presented via the iterative linear matrix
random time delays is investigated, where a general non-linear inequality (LMI) approach in [1-2]. In [3], the idea of Q-
auto-regressive moving average with exogenous model is used
to describe the plant. Due to the non-Gaussian random time
learning and concept of adaptive estimator were utilized to
delays involved in the systems, it is insufficient to obtain a develop the stochastic optimal and sub-optimal control
satisfactory optimal control algorithm by only controlling the designs for linear NCSs in the presence of random delays
expected value of the tracking errors. The Renyi entropies of and packet losses. However, most of the developed
the tracking errors and control inputs are adopted to stochastic NCSs mainly investigate linear systems.
characterize the randomness of the closed-loop system. The Moreover, the assumed condition on various induced delays
formulations of the probability density functions (PDFs) of the may be strict or conservative.
tracking errors and control inputs are deduced. By minimizing
the new performance index, a recursive optimal control Since the NCSs are usually nonlinear non-Gaussian
algorithm is obtained. Furthermore, the local stability systems, the stochastic optimal controller [4] and filter [5]
condition of the closed-loop systems is established. Finally, the were designed based on a minimum error entropy criterion.
simulation results are presented to illustrate the effectiveness Following the development in [4-5], a data-driven controller
of the proposed method. is designed for two-inputs and two-output (TITO) NCSs.
Keywords—minimum entropy criterion; networked control
This paper is organized as follows. Section 2 formulates
systems; random time delays; non-Gaussian
the problem of the TITO NCS with non-Gaussian time
delays. Section 3 designs controller for the investigated
I. INTRODUCTION TITO NCS. Section 4 presents the local stability condition
Compared with the traditional point-to-point control of the closed-loop TITO NCS. Section 5 introduces a
systems, networked control system (NCS) has the numerical simulation example to illustrate the efficiency of
advantages in terms of less wiring, low cost, easy to extend the proposed control strategy. The last section concludes
and maintenance, high reliability and so on. Thus, NCS is this paper.
widely applied in many significant fields, such as remote
medical care, intelligent transportation, automotive industry. II. PROBLEM FORMULATION
A major problem in networked control system (NCs) is the
network induced delays which result in the deterioration of Consider the TITO NCS setup in Fig. 1, assuming the
system performance and even destabilize the systems. plant model can be represented by the following non-linear
ARMAX equations:
The delay characteristics on NCSs basically rely on the
type of a network. Delays over random access networks are  y1, k = f1 ( y1, k −1 , y1, k − 2 , , y1, k − n1 , y2, k −1 ,
stochastic in nature. Moreover, a cyclic service network 
 y2, k − 2 , , y2, k − n1 , u1, k ,u1, k −1 , , u1, k − m1 ,
connected to a random access network also leads to random
delay. In recent years, some control algorithms have been  u2, k ,u2, k −1 ,u2, k − 2 , ,u2, k − m1 )

 (1)
 y2, k = f 2 ( y1, k −1 , y1, k − 2 , , y1, k − n2 , y2, k −1 ,
 y2, k − 2 , , y2, k − n2 ,u1, k ,u1, k −1 , , u1, k − m2 ,

This work was supported by China National Science Foundation under Grant  u2, k ,u2, k −1 ,u2, k − 2 , ,u2, k − m2 )
(61374052, 61503271, and 61511130082) and Beijing Natural Science Foundation 
(4142048).

978-1-4673-9891-6/16/$31.00 ©2016 IEEE


u1,k y1 , k where H 2 (⋅) is the quadratic Renyi entropy, E (⋅) is the
mean value. Ri (i = 1, 2, , 6) are constant weights
u2,k y2, k
corresponding to the entropies and mean values respectively.
When the entropy of a random variable is minimized, all
τ 1,k τ 2,k d 2,k d1,k
moments of its PDF are constrained besides the second
moments. Minimizing the entropies of both tracking errors
u2,k _ y 2,k
and control inputs makes the uncertainties of the tracking
_ errors and control inputs as small as possible, it means that
u1,k + + y1,k the NCS may have small uncertainties. In addition, the
r1,k r2,k tracking errors can approach to zero by minimizing the last
two terms on the right side of (5).
Fig. 1. Diagram of a NCS with random time delays The goal of next section is to find optimal sequential
control inputs u1,* k and u *2, k by minimizing the performance
where ui , k and yi , k (i=1,2) are the control inputs to the
index J k in (5).
plant and measured outputs from the plant respectively.
ui , k (i = 1, 2) and y i , k (i=1,2) are the control signal outputs
from the controller node and inputs to the controller node. III. MODEL FREE CONTROL
Without loss of generality, we assumed that f i (⋅) (i=1,2) is In this section, the way to obtain the performance index
continuous, bounded and first-order differentiable functions (5) is introduced, the optimal incremental control algorithm
for all its variables. is then formulated.
Bounded random time delays di , k and τ i , k (i=1,2) exist
A. Calculating performance index
in both links from sensors to controllers and controllers to
Take one of the tracking error e2, k as an example, its
actuators shown in Fig. 1. Let the set points r1, k and r2, k be
bounded as well. Then the tracking errors are quadratic Renyi entropy H 2 (ek ) and mean value E (ek ) are
defined as
ei , k = ri , k − yi , k (i = 1, 2) (2)
b
H 2 (ek ) = − log  γ e2k ( x)dx = − log V (ek ) (6)
where e1 and e2 are defined on [α1 , β1 ] and [α 2 , β 2 ] a

respectively. Moreover, the tracking errors can be described b


E (ek ) =  xγ ek ( x)dx (7)
by a

 e1, k = g1 (η1, k , uk , d1, k , d 2, k ,τ 1, k ,τ 2, k ) where [ a , b ] and γ e are the domain of definition and
 (3) probability density function (PDF) of the tracking error,
e2, k = g 2 (η 2, k , uk , d1, k , d 2, k ,τ 1, k ,τ 2, k ) respectively. The information potential V (ek ) =  γ e2k ( x)dx
b

a
T
where uk = u1, k u2, k  , control inputs through networks increases when the entropy H 2 (ek ) decreases.
u1,k and u2,k are defined on [α 3 , β 3 ] and It can be observed from (6) and (7) that the PDF of
[α 4 , β 4 ] respectively. ηi , k (i = 1, 2) are the known tracking error should be estimated in order to calculate the
performance index (5). Some nonparametric methods can be
information vectors at instant k : employed to estimate the PDF. In this work, the histogram-
based estimation approach is used for estimating the PDF.
ηi , k = [ y1, k −1 , y1, k − 2 , , y1, k − n , y2, k −1 , y2, k − 2 , ,
i Suppose the tracking error sequences within a sliding
y2, k − ni , u1, k −1 , , u1, k − mi , u2, k −1 , , u2, k − mi , window whose window width is N are {ek(1) , ek(2) ,  , ek( N ) }
(4)
y1, k −1 , y1, k − 2 , , y1, k − ni , y 2, k −1 , y 2, k − 2 , , located in [ a , b ] . Equally partition [ a, b] into N intervals,
y 2, k − ni , u1, k −1 , , u1, k − mi , u2, k −1 , , u2, k − mi , ri , k ]T the section points are a = t0 < t1 <  < t N = b , and the width
b−a
Since there are induced non-Gaussian delays in NCSs, of each interval is Δ = ti +1 − ti =, i = 0,1, 2, , N − 1 .
N
the model free controller can be designed by minimizing
following performance index Then the standard histogram PDF estimator with respect to
Δ is given by:
J k ( uk ) = R1 H 2 (e1, k ) + R2 H 2 (e2, k ) + R3 H 2 (u1, k )
(5) γˆe = μ ([ti , ti +1 ]) Δ ,
k
if ek ∈ [ti , ti +1 ] (8)
+ R4 H 2 (u2, k ) + R5 E (e 21, k ) + R6 E (e 2 2, k )
where μ ([ti , ti +1 ]) is the standard empirical measure of Δuk∗ = −ψ k−21ψ k1,ψ k 2 > 0 (18)
[ti , ti +1 ] , i.e.
IV. STABILITY ANALYSI
1 N
N j =1
(
μ ([ti , ti +1 ]) =  Ι ek( j ) ∈ [ti , ti +1 ] ) (9) In this section, a local linearization approach will be
used to study the stability condition of the NCS. Since the
tracking errors are stochastically bounded, the closed loop
where Ι (⋅) is the indicator function.
NCS can be formulated by:
Substituting the histogram PDF (8) into (6), the
 y1, k = F1 ( y1, k −1 , , y1, k − n1 , y 2, k −1 , , y 2, k − n1 , u1, k −1 , , u1, k − m1 ,
nonparametric estimator for Renyi’s entropy can be 
obtained:  u2, k −1 , , u2, k − m1 , u1, k , u2, k , d1, k , d 2, k ,τ 1, k ,τ 2, k , μ1, k )

 μ 2 ([ti , ti +1 ])   y 2, k = F2 ( y1, k −1 , , y1, k − n2 , y 2, k −1 , , y 2, k − n2 , u1, k −1 , , u1, k − m2 ,
H 2 (ek ) = − log  ⋅ ( b − a )  , if ek ∈ [ti , ti +1 ] (10)  u2, k −1 , , u2, k − m2 , u1, k , u2, k , d1, k , d 2, k ,τ 1, k ,τ 2, k , μ 2, k )
 Δ 2  
 
(19)
The mean value of tracking error ek can be calculated by
where μi , k = [ y1, k −1 , y1, k − 2 , , y1, k − n , y2, k −1 , y2, k − 2 , , y2, k − n ,
i i

1 N
u1, k −1 , , u1, k − mi , u2, k −1 , , u2, k − mi ]T (i = 1, 2) .
E (ek ) =
N
e
j =1
( j)
k (11)
Then, system (19) is linearized to be
Similarly, the entropies of the control inputs can be
estimated using (10) as well. Therefore, the performance  n1
∂F1 n1
∂F1 m1
∂F1
index (5) can then be obtained. Do not use abbreviations in  Δy1, k = Δy1, k − i +  Δy 2, k −i +  Δu
the title or heads unless they are unavoidable.  i =1 ∂y1, k − i i =1 ∂y 2, k − i i =0 ∂u1, k − i 1, k − i
 m1
∂F1 ∂F ∂F1
B. Incremental control algorithm  + Δu2, k − i + 1 Δd1, k + Δd 2, k
 i =0 ∂u2, k − i ∂d1, k ∂d 2, k
After obtaining the performance index (5), the optimal 
control input can be solved as follows:  ∂F1 ∂F1 ∂F1
+ Δτ 1, k + Δτ 2, k + Δμ1, k
 ∂τ 1, k ∂τ 2, k ∂μ1, k
uk∗ = arg min J k ( uk ) (12)  n2 n2
Δy ∂F2 ∂F2 m2
∂F2
= Δy1, k − i +  Δy 2, k − i + 
uk
Δu
 2, k ∂y1, k −i i =1 ∂y
 2, k −i 1, k − i 1, k −i
i = 0 ∂u
Denote uk = uk −1 + Δuk , ψ k (uk ) = J k . The Taylor 
i =1

 m2
∂F2 ∂F ∂F2
expansion at uk −1 of ψ k (uk ) will be approximated as  + Δu2, k − i + 2 Δd1, k + Δd 2, k
 i =0 ∂u2, k −i ∂d1, k ∂d 2, k
follows:
 ∂F2 ∂F2 ∂F2
 + Δτ 1, k + Δτ 2, k + Δμ 2, k
1  ∂τ 1, k ∂τ 2, k ∂μ 2, k
ψ k (uk ) = ψ k 0 + ψ k 1 Δuk + ΔukTψ k 2 Δuk (13)
2 (20)
where
where Δy i,k = yi,k − yi,k −1 and Δui,k = ui,k − ui,k −1 .
ψ k 0 = ψ k (uk ) uk = uk −1 (14) Using the unit backward shift operation z −1 to both
sides of (20), we have
∂ψ k (uk )
ψ k1 = uk = uk −1 (15) O( z −1 ,k )Δy k = I ( z −1 ,k )Δuk + Dk Δωk + K k Δμk (21)
∂uk
T T
where Δy k =  Δy1, k Δy 2, k  , Δuk =  Δu1, k Δu2, k  ,
∂ 2ψ k (uk )
ψ k2 = uk = uk −1 (16) Δωk =  Δd1, k Δd 2, k Δτ 1, k Δτ 2, k 
T
∂uk2
T
The optimal control law can be calculated from the Δμ k =  Δμ1, k Δμ 2, k  and
following optimality criterion
 n1
∂F1 − i n1
∂F1 
∂J k 1 −   z  ∂y z −i 
=0 (17) −1  i =1 ∂y1, k − i i =1 2, k − i 
∂Δuk O( z ,k ) = (22)
 n2 ∂F n2
∂F2 
1 −  2
z −i  z 
−i

the optimal law of the NCS in Fig. 1 is then given by  i =1 ∂y1, k − i


 i =1 ∂y
 2, k − i 
 m1 ∂F1 −1 m1
∂F1   1
  z  ∂u z −1   y1 (k ) = (1 + y (k − 1)) 2 (0.8 y1 (k − 1)
 i = 0 ∂u1, k − i 
−1
I ( z ,k ) = m
i =0 2, k − i
(23)  1

 2 ∂F m2
∂F2   +0.2u1 (k − 1) + 0.2u2 (k − 1))
 2
z −1  z −1   (29)
 i = 0 ∂u1, k − i i = 0 ∂u
2, k − i   y (k ) = 1
(0.9 y2 (k − 1)
 2
(1 + y2 (k − 1)) 2
 ∂F1 ∂F1 ∂F1 ∂F1  
 +0.3u1 (k − 1) + 0.1u2 (k − 1))
 ∂d ∂d 2, k ∂τ 1, k ∂τ 2, k 
Dk =  
1, k
(24)
 ∂F2 ∂F2 ∂F2 ∂F2 
  Fig. 3 shows the responses of two outputs of the NCS, it
 ∂d1, k ∂d 2, k ∂τ 1, k ∂τ 2, k 
can be seen that the proposed approach can stabilize both
the outputs around the set points with small oscillation. Fig.
 ∂F1  4 shows the performance index which is decreasing along
 ∂μ 0 
with sampling time. The three-dimensional (3D) mesh PDFs
Kk =  
1, k
(25) of tracking errors is shown in Fig. 5. The PDFs at some
 ∂F2 
 0  typical instants are shown in Fig. 6. Obviously, the shapes
 ∂μ 2, k  of PDFs of tracking errors become narrower and sharper
along with sampling time, which indicates the system
Denote I ( z −1 ,k )Δuk + Dk Δωk + K k Δμk = ξ k , then (21) outputs are driven towards a smaller uncertainty or
can be represented as randomness.
O( z −1 ,k )Δy k = ξ k (26)
VI. CONCLUSION
From the previous discussion, Δuk , Δωk and Δμ k are In this paper, a model free control strategy has been
bounded. Therefore, ξ k is also bounded. In this case, if Δy k proposed for TITO NCSs with non-Gaussian random time
delays based on an improved performance index. Since the
is bounded, then the linearized closed-loop system is stable.
delays are non-Gaussian variables, some traditional
Let
performance indices cannot characterize the randomness in
N the control progress. Thus, the Renyi entropies of the
O( z −1 ,k ) = 1 −  Ai ( k ) z − i , N = max(n1 , n1 , n2 , n2 ) tracking errors and control inputs are considered to establish
i =1 (27) a new performance index. Then, the performance index is
X (k ) = [ x1 (k ) x2 (k )  xN (k ) ]
T
calculated and a recursive optimal control law derived.
Finally, some simulation results are presented to testify the
where x1 (k ) = Δy k − N , x2 (k ) = Δy k − N +1 ,  , xN (k ) = Δy k −1 . effectiveness of the proposed strategy.
Consequently, the following state-space representation of
Δy k can be obtained

 0 1  0  0
         d1,k d2,k

X (k + 1) =  X (k ) +   ξ k 80 80
 0 0  1  0 (28)
    60 60
Delay (m s)

Delay (m s)

 AN ( k ) AN −1 (k )  A1 (k )  1 
40 40
= A(k ) X (k ) + bξ k
20 20
Therefore, the closed-loop stability condition for the
system is A(k ) < 1 . 0 100 200 300 400 500 0 100 200 300 400 500
Samples number Samples number
τ1,k τ2,k

V. SIMULATION RESULTS 80 80

In order to demonstrate the performance of the proposed 60 60


Delay (m s )

Delay (m s )

control strategy for a non-linear TITO NCS with random 40 40


time delays, let us consider the TITO plant described by Eq.
(29). 20 20

In the simulation, the non-Gaussian random time delays 0 100 200 300 400 500 0 100 200 300 400 500
are measured by an experiment for the sake of analysis. The Samples number Samples number

samples and the distribution of the random time delays are


Fig. 2. Random time delays
shown in Fig. 2 respectively.
[1] Y. Shi, B. Yu, “Output feedback stabilization of networked control
2 systems with random delays modeled by Markov chains,” IEEE
system output y 1(k) Trans. Autom. Control, vol. 54, pp. 1668-1674, Jul. 2009.
system output y 1 (k)

1.5
set point r1(k) [2] M. Liu, D.W. Ho, Y. Niu, “Stabilization of Markovian jump linear
system over networks with random communication delay,”
1 Automatica , vol. 45, pp. 416-421, Feb. 2009.
0.5 [3] H. Xu, S. Jagannathan, F.L. Lewis, “Stochastic optimal control of
unknown linear networked control system in the presence of random
0 delays and packet losses,” Automatica, vol. 48, pp. 1017-1030, Jun.
0 10 20 30 40 50 60 70 80 2012.
time(s)
[4] M. Ren, J. Zhang, M. Jiang, M. Yu, J. Xu, “Minimum entropy
2 control for non-Gaussian stochastic networked control systems and
system output y 2(k) its application to a networked DC motor control system,” IEEE
system output y 2 (k)

1.5 Trans. on Control Sys. Tech., vol. 23, pp. 406-411, Jan. 2015.
set point r2(k)

1 [5] J. Zhang, L. Du, M. Ren, G. Hou, “Minimum error entropy filter for
fault detection of networked control systems,” Entropy,vol. 14, pp.
0.5
505-516, Feb. 2012.

0
0 10 20 30 40 50 60 70 80
time(s)

Fig. 3. System responses

1.6
Entropy based control
1.4

1.2

1
J (e)

0.8
k

0.6

0.4

0.2

0
0 10 20 30 40 50 60 70 80
time(s)

Fig. 4. Performance index

1.2
1
0.8
1
γe

0.6
0.4
0.2 80
60
-5 40
0 20
5 0
t
e1

1.2
1
0.8
2
γe

0.6
0.4
0.2 80
60
-5 40
0 20
5 0
t
e2

Fig. 5. PDFs of tracking errors

REFERENCES

You might also like