You are on page 1of 7

 

International Journal of Trend in Scientific 
Research and Development (IJTSRD) 
UGC Approved International Open Access Journal 
 
ISSN No: 2456‐6470 | www.ijtsrd.com |Volume ‐ 1 |Issue ‐ 5 

MIMO Systems with Degraded Actuators and Sensors


Chieh-Chuan Feng
Department of Electrical Engineering
I-Shou University
Kaohsiung City 84001, Taiwan

ABSTRACT

This paper presents a reliable control design technique only within a prespecified subset of the control
for linear, time-invariant, multi-input multi-output components. In addition, the actuators and sensors are
(MIMO) systems with degraded actuators and sensors. subject to either failure or normal operation which
The degradation defined in this paper ranges from neglects the possibility of partial degradation. Fujita
normal operational conditions to complete failure of and Shimemura [2] provide necessary and sufficient
actuators and sensors. We derive linear matrix inequa- conditions by introducing a class of U-matrices for
lity (LMI) conditions ensuring robust stability of the examining system stability against arbitrary feedback-
system using static state feedback. The potential of the loop failures in multivariable control systems with a
proposed technique has been demonstrated by an ex- stable plant and controller. This condition requires an
ample of three coupled inverted pendulums. exponential growth in computation as a function of
plant dimension. Partial failures are not considered.
Keywords—linear matrix inequality; static state For a given stabilizing controller of a plant,
feedback; MIMO; degradation. Vidyasagar and Viswanadham [3] studied the problem
  of designing the second controller such that either
I. Introduction controller acting alone stabilizes the plant and both
Conventional feedback control designs for a MIMO acting together also stabilize the plant. A computing
system may result in instability in the event of the method for the second stabilizing controller which
degradation of actuators and sensors, even though it involves stable coprime factorizations using the plant
may be possible to control the system using only the and the first controller data are established and may
surviving functions of actuator and sensor. It is worth result in the second controller of high order. Cho et al
noting that this condition can occur even if the open [4] used the same methodology as [3] except that the
loop system is stable. It is therefore of interest to second stabilizing controller is of adaptive controller
develop feedback control designs which guarantee which again ends up with high order.
robust stability despite degradation of actuators and In this paper, we design a static state feedback control
sensors. system to tolerate the degradation ranging from
There are relatively few methods for design of reliable normal operational condition to complete failure of
control. Veillette et al [1] develop observer-based actuator and sensor. The degradation is modeled as a
reliable centralized and decentralized control systems multiplicative uncertainty at the plant input or output.
provided the failure of actuators and sensors occur We accomplish this via diagonal weighting and norm
   
 
@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 1 | Issue – 5 Page:90
 
International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470
bounded matrices. To simplify the derivation we will (3) 0 < ur < 1 indicates the rth actuator may fail
first develop and prove LMI conditions which tolerate partially.
sensor or actuator degradation independently. Our
u uk
approach is extended to the simultaneous degradation x  Ax  Bu +

actuators and sensors in the system. x pu u qu u

II. Problem Formulation K

The purpose of this section is to define the framework


on which our approach of reliable control systems is Figure 1 Closed Loop System for Actuator Case
based. The degradation for actuators and sensors will
B. Sensor case:
be considered separately to simplify the derivation.
Sensor case represents the degradation of sensors
A. Actuator case: and is written as follows,
Actuator case represents the degradation of
x  A x  B u
actuators and is shown as follows, (2.2)
xs  x  ps
x  A x  B u qs   sx  
u  uk  pu  
p sl   sl q sl ,  sl ( t )  1.
qu   uuk (2.1)
  where s is the uncertainty weighting of the sensor and
p ur   ur q ur ,  ur ( t )  1.
xs denotes sensor output. The definition for ps, qs, s,
where xRn denotes the state of the system, uk Rm is and s are defined in (2.3) in a similar way as in the
the output of the control gain shown in Figure 1. u  actuator case and can be seen clearly from Figure 2
Rm is the input to the plant. u is the uncertainty where it shows the closed-loop system for sensor
weighting of the control input to the plant. Note that degradation.
pu and qu represent the output and input vectors of
ps = [ ps1 .... psl ... psn ]T, ps l  R
plant input perturbations. pur and qur correspond to the
rth element of vector pu and qu respectively. qs = [ qs1 .... qsl ... qsn ]T, qsl R (2.3)

pu = [ pu1 .... pur ... pum ]T, pur  R s = diag (s1 .... sl ... sn ), sl  R
qu = [ qu1 .... qur ... qum ]T, qur  R s = diag ( s1 .... sl ... sn),  sl  R
u = diag ( u1 .... ur ... um), ur  R The degraded sensor signal xs(t) can be readily
ur(t) is the real valued unknown bounded uncertainty represented as
and it is assumed to be time-varying scalar quantity.
xs(t) = ( I + s(t)s)x(t)
The control signal u(t) can be readily computed as
where the possibility of full or partial failure of sensors
u(t) = ( I + u(t)u)uk(t)
is defined in the similar way as in the actuator case
where the I is identity matrix and u = diag (u1 .... ur ... using variables s and s in place of u and u,
um ), ur  R. Since |ur(t)|  1, i.e. -1  ur(t)  1, respectively.
x u
the degradation is modeled by a priori weighting u as + x  Ax  Bu

follows. ps s qs s
xs
(1) u = 1 indicates the rth actuator may fail fully.
r

(2) ur = 0 indicates there is no possibility that the rth K

actuator fails partially or fully. Figure 2 Closed Loop System for Sensor Case
 

@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 1 | Issue – 5 Page:91


 
International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470
C. Actuator and sensor case Theorem 1. Consider the closed-loop system
Actuator and sensor case is the case where actuator x  ( A  B K ) x  B p u
and sensor are simultaneously degraded shown as qu   u Kx (3.1)
follows,
p ur   ur q ur ,  ur ( t )  1.
x  A x  B u
u  uk  pu If there exist matrices Q, Y, and M satisfying
xs  x  ps
Q = QT > 0
qu   uuk (2.3)
QAT  AQ  BY  YT BT  BMBT Y TuT 
qs   sx  0 (3.2)
 uY M
p ur   ur q ur ,  ur ( t )  1
p sl   sl q sl ,  sl ( t )  1. for a given weighting matrix u, then the following
statements are equivalent,
(1) The closed-loop system (3.1) is asymptotically
x x  Ax  Bu u uk
+ + stable.
ps  s q s
s pu  u q u (2)(A + BK) is Hurwitz for K = YQ-1.
xs u

Proof. See Appendix A for proof and notation.


K

Remark 1. The inequality (3.2) is convex in the matrix


Figure 3 Closed-loop System for
variables Q, Y, and M. Thus, (3.2) can be readily
Actuator and Sensor Case
solved by convex optimization techniques. We write a
The matrices A and B shown in (2.1), (2.2), and (2.3) feasibility problem
are assumed to be real constant matrices with
Find Q, Y, and M
appropriate dimensions. u and s are given a priori
and represent the uncertainty weighting used to model Subject to Q = QT > 0, M > 0, and (3.2).
the degradation for actuators and sensors. We define
Theorem 2. Consider the closed-loop system
u := diag (u (t))r=1,...,m,
r
u (t)  R
r
x  ( A  B K ) x  B K p s
qs   (3.3)
s := diag (sl(t))l=1,...,n, sl(t)  R s x
p sl   ls q sl ,  ls ( t )  1.
where ||u||  1 and ||s||  1.
The control problem is to find, if it exists, static state If there exist matrices P, W, S and  satisfying
feedback that stabilizes the system described in (2.1),
P = PT > 0, S = ST > 0,  > 0,
(2.2), or (2.3).
 PP  W T S 1W  0 (3.4)
III. Controller Synthesis
 PA T  AP  W T B T  BW  BSB T P  Ts 
In this section we present the main results. Lyapunov  0 (3.5)
 sP   1 
stability theory is used to establish stability criteria.
Although there is no trivial method to establish a for a given weighting matrix s, then the following
Lyapunov function, by experience, quadratic statements are equivalent.
Lyapunov functions, i.e. V() = TL, have been
(1) The closed-loop system (3.3) is asymptotically
proven to be efficient and easily implemented [7]. We
stable.
consider static state feedback, i.e. uk(t) = Kx(t) for the
actuator case, u(t) = Kxs(t) for sensor case and uk(t) = (2)(A + BK) is Hurwitz for K = WP-1.
Kxs(t) for combination of actuator and sensor case. Proof. See proof for Theorem 3.
 

@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 1 | Issue – 5 Page:92


 
International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470
Remark 2. The matrix inequalities (3.4) and (3.5) are IV. Numerical Example
inversely coupled through the matrices S and .
Consider the system in Figure 3 consisting of three
Moreover the inequality (3.4) is not an LMI. Thus, to
coupled inverted pendulums of point masses mi, and
solve theorem 2 simultaneously through the use of
length li. The pendulums interact via three springs and
convex optimization is difficult. However, we may
three dampers of stiffness kij and damping bij; i, j =
still solve the problem using of sequential method that
1,2,3, and ij. The distances from attached point of
we solve for (3.5) first and then verify (3.4).
springs and dampers to the platform baseline are ai.
Theorem 3. Consider the following closed-loop The system data are shown in Table 1. Using this
system system, we will demonstrate several examples for
actuator and sensor degradation. The system dynamics
x  ( A  B K ) x  B K p s  B p u
qu   u K ( x  ps ) for pendulums are written in the general form,
(3.6)
qs   sx  3  
  1  ai2  l j ai a j   mi g 1      (4.1) 
p  q ,
r r r
 (t )  1
r  i   bij  i  2 kij i    bij  j  2 kij  j   
mi  j 1, j i   li   li li  l i
i  ui 
l i 
u u u u

p sl   sl q sl ,  sl ( t )  1. u1
u3
m1
u2 m3
m2
k13,b13
l1
If there exist matrices P, W, Su, Sx, Ss, , and  l2 l3
satisfying k12,b12
k23,b23

P  P  0 , Su  S  0 , S x  S  0 , S s  S sT  0 ,
T T T
a2
u x
1 2 3 a3
a1
 > 0,  > 0   

   B T S u1 B  0 (3.6) Figure 3 Three Inverted Pendulums Systems 


Table 1 
 PP  W T ( S x1  S s1    )W  0 (3.7)
m1 m2 m3 b12 b13  b23  k12  k13 k23
 PA T  AP  W T B T  BW   1  1  1  1  .51  1  1  .51 1 
  WT P  Ts 
  BS x B  S u  
T
(3.8) a1=a2=a3=0.5, g=10, l1=1, l2=1.2, l3=1.1.
  T S s      
 1 
 W 0 0
 sP 0  
1
Initial conditions:
 
 
. , 1  011
1  01 . ,  2  0.34, 2  0.23,  3  0.29, 3  133
. .

where     Tu  u , for a given weighting matrix u We will first demonstrate actuator degradation. The

and s, then the following statements are equivalent, weighting is chosen u=diag(1, 0.1, 0.28) which
represents the possibility of degradation for each
(1) The closed-loop system (3.6) is asymptotically actuator. For instance, actuator for m1 may fail during
stable. its operation, actuators for m2 and m3 are subject to
(2) (A + BK) is Hurwitz for K = WP-1. 10% and 28% variation of its nominal operation signal.
We will demonstrate nominal operation of actuators,
Proof. See Appendix B for proof. i.e. u = 0, at time t  0.5. Then, we will show the case
Remark 3. Normally speaking, the matrix inequalities where the actuator for m1 is subject to fail and the
(3.6)-(3.8) should be solved simultaneously. However, actuator for m2 and m3 are partial failure. Detail
it is not possible to solve using convex programming conditions are shown in Table 2.
Table 2
due to inversely coupled terms, Ss, Su, Sx, and .
time  m1  m2  m3 
Thus, we suggest for some given  we solve (3.8) for
P, W, Sx , Ss, and Su and then verify (3.6) and (3.7). 0  t < 0.5  S1  S2  S3 

@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 1 | Issue – 5 Page:93


 
International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470
0.5  t < 2.5  failure  0.9*S2  0.72*S3  chosen s = diag(0.4, 0.5, 0.4, 1, 0.5, 1) which
2.5  t  6  2*S1  1.1*S2  1.28*S3  represents the possibility of

where Si = the nominal operation control signal of


20

actuator associated with each mass. The computed


gain followed by Theorem 1 is 15

 -44.6423 -11.3417 2.7025 -1.3565 1.2500 -0.7005 


 
10

control signals
K  - 362.5824 - 93.3584 - 2.9357 -17.3282 - 4.3807 -10.9399 
 - 25.4243  7.1539 4.5474  16523
.  35.8191  10.0032
5

eigenvalues of (A+BK) = [-45.2643 -3.6514 - (3)

1.13161.6097i -2.13062.1394i]. The simulation 0


(1)

results are shown in Figure 4 and 5 for plant outputs


(2)

−5

and control signals. It is noticed that the control signal


0 1 2 3 4 5 6
time (sec)

for m1 decays relatively fast at the beginning of Figure 5 Control Signals 


operation and then is close to zero as shown in Figure degradation for each sensor. The velocity sensor of m1
4. Thus, even though the actuator #1 fails at time t  is subject ++to 50% variation, while the position
0.5 sec, it has very little effect to the system. Since the sensor has 40% variation. The velocity sensors of m2
actuator signals for m1 varies from twice of its nominal and m3 are subject to the possibility of failure, while
value to zero (which is used to represent the failure of the corresponding position sensors have 40% and 50%
the actuator), one possible way to accommodate such variations. We will demonstrate the extreme case,
a large variation is simply to make the resultant signal which is shown in Table 3. The computed gain
as small as possible such that it has the least effect on followed by Theorem 2 is
the system. The jumps shown in Figure 5 are due to -29.7241 -8.2556 1.0801 -1.0596 1.2811 0.3397
 
K   0.5669 - 0.9255 - 29.5401 - 8.0746 0.9343 0.9577
the switching at time t = 0.5 sec instantaneously. If
 1.2812  0.2544 12151
.  0.9838  29.3428 8.3852 
there are two actuators in the system subject to the
possibilities of failure, e.g. u=diag(1, 1, 0), we are not eigenvalues of (A+BK) = [-6.8259 -2.4785 -5.8912 -
able to find feasible solutions for the Theorem 1. In 3.0449 -3.0952 -6.2916]. The simulation results are
other words, it is not possible to maintain stability of shown in Figure 6 and 7. The sudden jumps in Figure
the system when two actuators failures in the system. 7 show the instantaneous switching at time t = 0.5 sec.
The following results are observed from simulation.
0.4 The system will oscillate if we fail three velocity
0.3
sensors. Since the damping control signals given by
0.2
(2) velocity sensor are zero, the occurrence of oscillation
is not unexpected.
outputs − theta1 theta2 theta3

0.1

(1) Table 3
0

time m1 m2 m3
−0.1

Pos Vel Pos Vel Pos Vel


(3)

−0.2

0 t 1.4* 1.5* 1.4* 2* 1.5* 2*


−0.3

 .5 S1p S1v S2p S2v S3p S3v


−0.4

0.5 0.6* 0.5* 0.6* F 0.5* F


−0.5
0 1 2 3
time (sec)
4 5 6
t 5 S1p S1v S2p S3p
Figure 4 Outputs of Three Pendulums
Pos = position sensor, Vel = velocity sensor, S with all
For the sensor degradation case, the weighting is
 

@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 1 | Issue – 5 Page:94


 
International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470
subscripts = the nominal operational signal associated Proof.
with each sensor (subscript p = position sensor, We consider the quadratic Lyapunov function V(x)=
subscript v = velocity sensor). F = failure. xTLx. The consideration of perturbation will be
incorporated in the derivative of Lyapunov function
0.4
for stabilization of overall system. We have
0.3

. m
V ( x )    r ( q ur q ur  pur pur )  0 (A.1)
T T
0.2

(2) r 1
outputs − theta1 theta2 theta3

0.1

0 (1)
for some r>0. Substituting (2.1) into (A.1) and
rearranging the expression, we obtain
−0.1
(3)

 AT L  LA  LBK  ( LBK)T  
−0.2

x
T
 T T  LB x  (A.2)
p    K u u K   p   0
−0.3

 u   u
−0.4
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5  BLT
 
time (sec)

Figure 6 Outputs of Three Pendulums


where the =diag(1, 2, ..., m). (A.2) is equivalent to
35

30  A T L  LA  LBK  ( LBK ) T  
  LB  (A.3)
  K  u   u K  0
T T
25
 T
  
20
 B L
control signals

15
Expanding by Schur complement and letting Q = L-1,
10
Y = KQ, and M = -1, we may have
5 (3)

QAT  AQ  BY  Y T BT  BMBT Y TuT 


0  0
(1)
(2)  uY M
−5

This completes the proof.


−10
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time (sec)
Appendix B
Figure 7 Control Signals
Proof.
V. Conclusions
We need the following lemma used in the proof.
We ensure the stability of a MIMO system by
Lemma B ([8]) For any matrices M and N with
establishing LMI conditions when the system is
appropriate dimensions, we have
subject to degradation of actuators and sensors. The
theory developed in Section 3 is demonstrated by three M T N  N T M  M T SN  N T S 1 M
inverted coupled pendulums, which show the system
for any positive-definite symmetric matrices S.
can be stabilized.
We consider the quadratic Lyapunov function V(x)=
VI. Acknowledgment xTLx. The system perturbation will be incorporated in
the derivative of Lyapunov function for stabilization
This work was supported by the Center for Autom-
of overall system. We have
ation and Intelligent Research and DARPA/ARO
grant. n m
V ( x)    r (qsr qsr  psr psr )   r (qur qur  pur pur )  0 (B.1)
T T T T

r 1 r 1
Appendix A

@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 1 | Issue – 5 Page:95


 
International Journal of Trend in Scientific Research and Development (IJTSRD) ISSN: 2456-6470
for some r>0 and r>0. Substituting (3.6) into (B.1) [1] R.J. Veillette, J.V. Medanic, and W.R. Perkins,
and rearranging the expression, we obtain “Design of Reliable Control Systems,” IEEE
Transactions on Automatic Control, vol. 37, No.3,
 x T ( A kT L  LA k  K  Tu  u K   Ts  s ) x 
  pp.290-304, March 1992.
  x ( LBK  K  u  u K ) p s  x LBp u 
T T T T (B.2)
  0 [2] M. Fujita and E. Shimenura, “Integrity Against
  p s ( K B L  K  u  u K ) x  p u B Lx 
T T T T T T T

  Arbitrary Feedback-loop Failure in Linear


  p sT (    K T  Tu  u K ) p s  p uT  p u 
Multivariable Control Systems,” Automatica, vol.
where the =diag(1, 2, ..., m) and =diag(1, 2, ..., 24, No.6, pp. 765-772, 1988.
m). Use of Lemma B we have [3] M. Vidyasagar and N. Viswanadham, “Reliable
Stabilization Using A Multi-controller
x T LBpu  puT B T Lx  x T LSu Lx  puT B T Su1 Bpu
Configuration,” Automatica, vol.21, No.5, pp.
x T LBKps  psT K T B T Lx  x T LBS x B T Lx  psT K T S x1 Kps 599-602, 1985.
[4] Y.J. Cho, Z. Bien, and B.K. Kim, “Reliable control
x T K T  T Kps  psT K T  T Kx  x T K T  T Ss  T Kx  psT K T Ss1 Kps
via additive redundant adaptive control,” in Proc.
where     Tu  u . Thus,
1989 Amer. Contr. Conf., Pittsburgh, PA., pp.
(B.2)  1899-1904, 1989.
[5] C.C. Feng and S. Phillips, “Robust Decentralized
  AkT L  LAk  K T (     T S s   ) K   Ts  s  
 xT  x  Control; A Linear Matrix Inequality (LMI)-based
   L( S u  BS x B ) LT
  0
  design,” Submitted to the 2017 CDC.
  p s     K ( S x  S s    ) K  p s  p u     B S u B  pu 
T T 1 1 T T 1

[6] S. Boyd and C.H. Barratt, Linear Controller


Therefore, we require Design - Limit of Performance. Prentice-Hall
International, Inc. 1991.
 AkT L  LAk  K T (     T S s   ) K  (B.4)
  0 [7] S. Boyd, L. El Ghaoui, E. Feron, and V. Bala-
   Ts  s  L( S u  BS x B T ) L 
krishnan, Linear Matrix Inequalities in System and
Control Theory. SIAM books, Philadelphia, 1994.
   K T ( S x1  S s1    ) K  0 (B.5)
[8] K. Zhou and P.P. Khargonekar, “Robust Stab-
ilization of Linear Systems with Norm-Bounded
   B T S u1 B  0 (B.6)
Time-Varying Uncertainty,” Systems and Control
Letters, Vol. 10, pp. 17-20.
Multiplying P = L-1 on both side of (B.4) and applying
Schur complement, we obtain  

 PA T  AP  W T B T  BW  
  WT P  Ts 
  BS x B  S u  
T


 W   T S s      
1
0 0
 (B.7)
 sP 0  
1

 
 

where W=KP. Multiplying P on both side of (B.5) we


have

 PP  W T ( S x1  S s1    )W  0 (B.8)

This completes the proof.


References

@ IJTSRD | Available Online @ www.ijtsrd.com | Volume – 1 | Issue – 5 Page:96