Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

&VSPQFBO$POUSPM$POGFSFODF &$$ 

1PSUP 1PSUVHBM 4FQUFNCFS 

AN LMI APPROACH FOR DESIGNING SLIDING MODE


OBSERVERS FOR FAULT DETECTION AND ISOLATION
C.P. Tan , C. Edwards 

Control Systems Research,


Department of Engineering,
University of Leicester,
University Road,
Leicester LE1 7RH,
United Kingdom.
fax: +44-(0)116-252-2619
e-mail: cpet1@le.ac.uk , ce@sun.engg.le.ac.uk
 

Keywords: Nonlinear Observer, Fault Detection and Diagno- struct the fault signals from knowledge of only the known sys-
sis, Variable Structure Control, Sliding Mode tem inputs (usually the control signal) and measurements of the
system output. In [9] this methodology has been used on the
Ship Propulsion benchmark problem. The examples consid-
Abstract ered in [6, 5], although challenging from certain viewpoints,
This paper considers the use of sliding mode observers for are of dynamical order four and allow for relatively straightfor-
fault detection/disturbance estimation. A new observer design ward tuning of some of the gains in the sliding mode observer.
method is considered which is based on Linear Matrix Inequal- This paper considers a 17th order system which represents a
ity (LMI) optimisation. Sliding mode observers have previ- chemical process [8]. Hand tuning of the gains for an ob-
ously been used for fault detection and disturbance estimation; server for this system is not practical. A further complication is
however the systems for which they have been designed were the presence of nonzero direct feed-through terms in the plant
of low dynamical order. The LMI approach considered here model. In this paper a recent Linear Matrix Inequality (LMI)
allows these ideas to be exploited on more realistic higher or- based approach for the design of the observer gains is consid-
der systems. The efficacy of the approach is demonstrated by ered [11] and is applied to the problem of disturbance estima-
considering a model of a chemical process plant. tion.

1 Introduction 2 Sliding mode observer description


Interest in the use of sliding mode ideas in the context of de- Consider the dynamical system
signing state estimation observers has been maintained over the
last two decades. Sliding mode observers differ from their lin-


                

(1)
ear counterparts in that the state error convergence is obtained         

(2)
via a nonlinear discontinuous injection term which depends on
the output estimation error. This term is designed to guarantee where  ! # % ' %   ! # % ' )   ! # , '

and
%  ! # % ' -

that the state estimation error vector is forced onto and sub- where . Assume that the matrices
. 0 2


and are full rank 

sequently remains on a surface in the error space - often the and the function  4 # 5 6 # % 6 # ) 8 #

is unknown but -

surface associated with the output estimation error being iden- bounded so that : : : :

tically zero. Utkin [12] designed a simple observer, with only


(3)
; < =

        >   

a discontinuous output error dependent term being fed-back


through an appropriate gain. Walcott & Zak [13] designed an
where is a known scalar and is a
< =

> 4 # 5 6 # , 8 # 5

observer which also has the output error being fed-back linearly
known function. The signal will be taken to represent a  

and used a Lyapunov approach to prove stability. The method


fault/disturbance condition within the system in such a way that
in [13] invariably requires a symbolic manipulation package to  

will be considered to represent a fault free condition.


C

solve the design problem which is formulated. Edwards and


Spurgeon [3, 4] proposed a canonical form for sliding mode Assuming that is full rank, it can be shown [3] that there   

observer design subject to certain conditions relating to the in- exists a linear change of coordinates
J J
so that the J
 F8 H I   J

put and output distribution matrices, and also the invariant ze- triple  

in the new coordinate system has the following


   

ros of the system. Their method described in [3] utilises both structure:
linear and discontinuous output error injection and is similar
in structure to that of Walcott & Zak. Recently Edwards et K

The system matrix can be written as


al.[6] considered the use of such an observer as a Fault Detec- LM

= = = N

tion and Isolation scheme, the fundamental idea being to use


 

(4)
N = =

the so-called equivalent output error injection signal to recon-


 

N N OP

N = N

Proceedings of the European Control Conference 2001 481

*4#/
where                  . Proposition 1 - If there exists a positive definite Lyapunov ma-
+ + + +

\
+

When partitioned these matrices have the structure trix , that satisfies P , with the structure P


:
P `

+ + +

(14)
P  P 

and (5)
 

   
+ + + + + +

[ [ [
\ \


P 


        

 

P  P  P 
 b
:

  

+ +

where for some    "  " and       " 


where and , then the error
P       P    

system in equation (13) is quadratically stable.


   

and the pair is completely observable.




# $

%   

    &

Furthermore, the eigenvalues of are the invariant ze- 

 

Proof see [11]. c

ros of . %  (  )

Corollary 1 - An ideal sliding motion takes place on


&

The disturbance distribution matrix has the form d ? O


+

 g

 e ) 

( 

(6) in finite time. Furthermore the sliding dynamics are given by [

the system matrix ( ).


(

    

where (    is nonsingular. Define


 

+
(     
Proof see [11]. c

as the bottom rows of (which therefore includes the




Remarks: If a further linear change of co-ordinates


0 (

matrix ). (

2 h

The output distribution matrix has the form   i


 2

(15)


 2

)   (7) is applied to the triple and its Lyapunov matrix , %


+

 (
+

 )
+

P
+

&

2
the system matrix, disturbance distribution matrix and the out-
where    and is orthogonal. put distribution matrix will be in the form
k k

The input distribution matrices and will be transformed to


+ +
3 4
k



k
 

k
 







be and respectively, and will have no particular structure.


3 4
   

 m


 n
i

(16)
m

Edwards and Spurgeon [3] propose a state observer for of the


+

k
[
2

form 6
where and . In the new co-     

:
     ( 

ordinate system, the Lyapunov matrix will be


m

+ + + +

= > ? @ =

(8)
+ +

7 7

% 8  % 8 3 ; % 8 % 8

& & : & < & : 

A
+
+ +

\
p
2 h 2 h

(9) (17)
P 
 
+

B 7

 2 2 \

 % P %  

% 8  ) % 8 4 ; % 8
& &

& & : &

 P 

where the discontinuous vector A


is defined by The s.p.d. matrix from (10) can now be defined as
+

O \

(18)
2 2

? @ N

if


P   P 

M M

(10)
% 8   ; (  

< F & G G I J K L

 D

otherwise


I J K L

If the error states in this new co-ordinate system are appro- \

? ?
\ \@

priately partitioned they can be written as with


? @ O

where and is symmetric positive 


B

<
B

P    

? ?


definite (s.p.d.) and will be formally defined later. The scalar . Furthermore, the states govern the reduced


   

function satisfies F
 R S  S  U V  R
order sliding motion. As argued by Edwards and Spurgeon p

[4], the fact that is block diagonal Lyapunov matrix for


(11)
k k > k

implies that is stable and hence the sliding


B $ # B

% 8   ;  ; % 8 

  
F & G G : X & : Y Z

< r

motion is stable.


where Y Z
is a positive scalar.
+ +

= > =

The gain matrices


=
+
and are to be determined, however 3 Synthesis procedure
is assumed to have the structure
+
In this section a method utilising Linear Matrix Inequalities [1] [
+

=
+

>

will be described to synthesise the gain matrices and and


[
2 \

(12)
<

the Lyapunov matrix which represent the design freedom in


2 \




the problem.
+

[
[  [

where and
+ +

= >

       

? O
   

As described in [11] P and will be chosen so that the matrix


and is given in (7). If the state estimation error ,
2

+ +

7 5

<

inequality
then it is straightforward to show from equations (1) and (8) 6
+

\
+ + + + + +

=
+

> t =
+

\
+

that (19)
>

P P ` P s P P P

: < <

+ + +


? ? =

(13)


% 8  % 8 ( _ % 8   ;

is satisfied, where the a-priori user-specified weighting matri-


& & : < &

 A

+ + + +

= > t

where 

<
) . ces and s are assumed to be symmetric
     

Proceedings of the European Control Conference 2001 482




positive definite, and has the structure in (14). The ratio- to calculate the optimal observer gain
nale for the matrix inequality (19) will be given later and is  
  = > ?
 
< = > ? 
  

described in detail in [11] . Substituting for , the inequality < = > ? 

(19) can be written as: The associated optimal cost is given by # $ % & '
.
< 
3

 
  
  
     
  

(20) It is argued by Tan and Edwards [11] that if is any ma- <
3
< = > ?

trix satisfying (27) then . Hence the requirement of < 

minimising follows from the desire to approach the


 
  

where . Using standard matrix manipulations, in- # $ % & '

< = > ? 

equality (20) is identical to true minimal cost given by . Of course a particular # $ % & '

  
<

sub-optimal cost is enforced by the requirement that


 

       
          

has the structure of (14).


  
      

(21) Remarks: In inequality (27), is the performance weighting




matrix for the observer, and is the co-variance matrix of the


Using inequality (21), necessary and sufficient condition for system’s sensor noise. As in classical LQG theory the choice
 

of and can be used to trade off performance and noise




(20) to hold is that satisfies





 


      
 
amplification.
(22)
Depending on the system, there may exist some freedom for
since choosing designing the sliding motion matrix . The designer may B
 


   

(23) add additional LMIs to the inequalities (25) and (26) to force
the eigenvalues of to lie in certain regions using pole- B
 

eliminates the third term in (21). The problem considered here 




placement techniques. However, simply satisfying inequality
is one of minimising subject to satisfying in-
 

# $ % & '

 
(25) alone will guarantee that is stable. For details see the B
 

equality (22). The observer gain can then be directly calcu- work of Tan and Edwards [11].
lated as  
   

  

(24)
3.1 Synthesis procedure implementation
which follows from equation (23) and the definition of . The 

matrix inequality in (22) is equivalent to Writing from (14) as


(

 

   D
  
       3

(29)
 
 

 
 
 )

 
  )

(25)  D
D D

 , - E / 0 E /  , - 0

where and
 G H  G H G G

  D D

* , - / 0 /

by using the Schur complement. If is symmetric 

with
D

, it follows
  I


 D 
 J


 D 
, - E /

 G H
0 E

G  M H

positive definite, then (again using the Schur complement) the there is a one-to-one correspondence between the variables
Linear Matrix Inequality (LMI) 

and 

since 

5


 D 

5


D D



5 N 5


D


(
  

(30)
 2

  

 

2 *

(26)  
 


N
 
 D 

(31)
  
     

* 3  


D D D  D

(32)
is equivalent to . Thus minimising sub-
   

 D
 

# $ % & '


*

ject to (22) is equivalent to minimising # $ % & ' subject to the It follows that the constrained optimisation problem: minimise *


LMIs (25) and (26). # $ subject to, (25) and (26) is a convex optimisation
% & '

   *

problem with regard to and . Standard LMI    D  D D

The motivation for the choice of the inequality posed in (19),


5 5




software, such as [7] can be used to synthesise numerically


and the minimisation of subject to (22) and (26),
 

* 

and . Once has been determined can be determined


# $ % & '

   
 

is that if and in (3) then as the ob-


N

   

from (31). can be determined from (24), from (12) and


$ 4 # 5 6 7 8 9

server tends to a sub-optimal Linear Quadratic Gaussian (LQG) 

from (18).
D

formulation. This will be demonstrated using an argument sim- <


 

ilar to that on page of [1]. Defining , then pre and


 

: : ;

<
The remainder of the paper explores the use of the sliding
post multiplying inequality (22) by , the following inequality mode observer described in previous sections for fault detec-
can be obtained: tion/disturbance estimation.
 < <   <  <  
     

(27)
 

<

4 Estimating the faults/disturbances


and . The objective is thus to minimise


 


  

< 

subject to (27). Once the parameters



and consequently 

 

5
 D 

5
D D

have been synthesised, the coordinate transforma-


# $ % & '

 D

5 N 5

The standard LQG optimal observer design method as de- < = > ?
tion in (15) is completely determined and the state estimation
scribed in [10] uses the stabilising solution to the Alge- error system associated with (16) can be written as
braic Ricatti Equation P




Q  

'


# B
 

'


# B
 D 

' R # (33)
 < = > ? < = > ?   < = > ?  < = > ?    Q   T 
          S

(28) (34)
P

D   D D D D U

' R # B ' # B ' R # #

Proceedings of the European Control Conference 2001 483


where and are appropriate partitions of the linear output  


In the linearisation of [8], the model was scaled. The follow-
error injection matrix after the coordinate transformation.  ing tables indicate the scalings that were used. The scalings
Notice this is precisely the canonical form for disturbance re- were multiplied with the corresponding quantities during the
construction proposed in [6]. linearisation.
Once a sliding motion has been attained and ,   
Quantity Scaling
and the discontinuous output error inject term can be replaced 
Purge down stream pressure (psia) /

by the so-called ‘equivalent output error injection’ required


   Cooling-effluent-temperature ( ) &  0 

to maintain a sliding motion [6, 3]. From (33) - (34), during


the sliding motion Table 1: Scalings used for the disturbances
            (35)
Quantity Scaling Range
(36)
Flash inlet temperature ( )
            ! #    

&   1 2 0 3

Since    is stable, and hence from equation (36),   


Production rate (BTU/hour)  1 2 3 

    # 
. The disturbance can be estimated from Product purity (mol percentage) 4   N/A
 
Hydrogen to aromatic ratio 0  1 2

(37) Flash vapour outlet pressure (psia)




  # #    #   
 1 2 2

 

since is full rank. In reality the equivalent output error in-


# 

Table 2: Scalings used for the outputs. The range indicates the
jection can be approximated to any degree of accuracy by
  

allowable perturbation within specifications.


replacing (10) with



Quantity Scaling
 !         
(38) Benzene column splitter: reflux ratio  1 2 5 0

Pre-flash cooler duty (BTU/hour)


    

2 5 0 6 1 7

where is a small positive constant which governs the accuracy



Gas feed flow (lb mol.hour) / 6 1 0

to which the equivalent control is approximated. Note the right Toluene feed flow (lb mol/hour) 0 0 1 6 7 0

hand side of (38) can be computed on line and hence estimates Purge outlet valve opening / 1 4 6 8

of can be computed online. For further details see [6, 4].


  

Table 3: Scalings used for the actuators


5 An application: a chemical plant A proportional-integral-derivative (PID) controller was de-
signed to provide good disturbance rejection and tracking of
In this section, the observer described and designed in the pre-
production rate changes. Using the PID structure
vious sections will be used for fault detection/disturbance es-
timation in a linear model of a chemical process plant for the 9 9
4
9 ;
:
9 =

production of Benzene. A detailed nonlinear simulation of the




 :   

Hydroealkylation (HDA) process, developed by [2], was lin-


: 4  < :

earised and the original 270 state system model-reduced to a where , Table 4 shows the details of the controller pa-
<  1 4

17th order system by Hermann et al. [8]. rameters. (In closing the loop the -th output was connected to >

The model of the chemical plant can be represented by the -th input.) >

The method of fault/disturbance estimation proposed in [6]







   
  

    


     


   (39) cannot be used directly because of the presence of the non-zero


   
  

    


     


   (40)  term in (40).




where are the states, are the outputs,


 "
   !  ! $  ! $

are the control inputs and  !




are the disturbances to the 5.1 Augmenting the system


system. Specifically
%
In this particular example can be verified to have full rank 


Purge down stream pressure (psia) (i.e. rank ) and thus there exists an orthogonal matrix
 A 
? @


Cooling-effluent-temperature ( ) &  '


!

such that %

Flash inlet temperature ( )


, -
@ 

-  
'
& 

Production rate (BTU/hour)


-

where and is non-singular. Let


 

 !

)*
Product purity (mol percentage)  

Hydrogen to aromatic ratio


*

*+

Flash vapour outlet pressure (psia)


=

 

@ 

  '

Benzene colum splitter: reflux ratio


-
,

Pre-flash cooler duty (BTU/hr) Using equation (40) it follows


-

 )*
Gas feed flow (lb mol/hr)
Toluene feed flow (lb mol/hr)
*

.

(41)
               

*+

Purge outlet valve opening


@    D

                     

Proceedings of the European Control Conference 2001 484



Loop 1 Loop 2 Loop 3 Loop 4 Loop 5 J
   L M




J
 N 


-0.3783 15 0.22 5 150  V W F X Y Z V Y

- - -
 [ V X \ V \ ] F Y ^ X W \ _

-0.5 50 0.15 4 80
 [ W X V V [ ] V X _  X F \ ^ [ ] Z [ X [ ^ ^ W _

Q Q Q

- - - - -

0 0 0.05 0 0
 V X F [ ] F X [ F _  Z X Z \ \

- -

 W W X F ^ [ [ ] W \ X V F Z _  X Y ] ^ Z X W [ [ _

Q Q Q

- -

 [ V X ^ Z  ^ ^ X F ^ Z ] F X F V _

Q Q Q

Table 4: PID controller details  W ^ X V Y Y \ ] V Z X

Q
V F Y _  V V X [ F \ [

 W F X \ [ ^ F  Z B Z V \ \

 V F X Y Y Y V ] [ X F \ Y _  F X [ [ \ \

where and represent an appropriate partition of the


 

 
 F F X W Y F ] Y X W W F
Q

Y _  ^ X Z F Y ] F X W Y Z [ _

scaled output distribution matrix . 

 ^ X W [ Y \
Q

] F X
-

Z _
Q

Q Q

The idea is to filter the vector according to


-

 F X [ W Y

 ^ V X \ F

                        !   

 "

   #

$
  

(42)
Table 5: Parameters of the observer
where is a diagonal matrix whose elements are the
  & ' ( ) (

inverse time constants of the first order filter in each channel. 10

8
10

Choosing and as the new outputs, it follows that (39) and  

6 6

(42) can be combined to form the augmented system 4 4

2 2

* * *


 -

0 0
        

−2 −2


     +      +      +


−4 −4

. / 0 1

−6 −6

−8 −8

−10 −10
0 5 10 15 20 25 0 5 10 15 20 25
* 4


* 7


Hours Hours

"
   

$
  

(43)
Figure 1: The purge-down-stream pressure disturbance (left) and its
+ +
  !   #

. / 0 1 . / 0 1

5 8
reconstruction (right). Unit is ` a .

and the new modified output is 60 60

* * *

-
40 40
    !

(44)
9

       

- < -

+      + +
"

20 20

. / 0 1 . / 0 1

0 0
=
>

−20 −20

Notice is full rank since is full rank. The sliding


 @
  #

−40 −40

mode observer strategy described earlier may now be employed −60 −60
0 5 10 15 20 25 0 5 10 15 20 25

for the system in (43) - (44) since it can be shown by numerical Hours Hours

calculation that from [8] has no invariant zeros.   B

@
B




Figure 2: The cooling-effluent-temperature disturbance (left) and its


reconstruction (right). Unit is b c d e
.
6 Simulation results
16 16

The augmented system developed in 5 has 19 states, 2 input C


14 14

faults/disturbances and 5 outputs. A sliding mode observer was 12 12

designed using the synthesis procedure in 3. Due to space C


10 10

8 8

constraints the parameters will not be shown. In this example, - - < < I
6 6

the design matrices were set to be and . D




- <
F

G H


4 4

The filter matrix was chosen to be . Table 5 shows the   F

2 2

eigenvalues of the linear part of the observer and J


   L M




0 0

0 5 10 15 20 25 0 5 10 15 20 25

the eigenvalues of the sliding motion of the observer J


 N 

Hours Hours

that were obtained from the synthesis procedure in 3.


Figure 3: The production rate demand (left) and the production rate
C

Note that in this case it would have been impossible to do the response (right). Unit is lb mol/hour.
design by hand using the pole placement design freedom avail-
able in [6]. A simulation was done with both disturbances (in
Table 1) and a change in the production rate demand applied si- Figures 1 and 2 show the faults/distrubances as well as their $

multaneously to the system. The observer was used to then es- reconstructions using equation (37). It can be seen that the ob-
timate the fault/disturbance using equation (37). In this sim- $

- - S T
server faithfully reconstructs the fault despite the presence $

ulation, and from (38) were chosen to be and O P Q


F R F

of sensor noise. Figures 3, 4, 5 and 6 show the response of the


respectively. The scaled sensors were subjected to white noise system to the disturbances and production rate demand change.
of standard deviation 0.08. It can be seen that the production rate tracks its demand, and

Proceedings of the European Control Conference 2001 485


2
low dynamical order. The LMI approach considered here al-
1.5
lows these ideas to be exploited on more realistic higher order
1 systems. The efficacy of the approach was demonstrated by
0.5
considering a model of a chemical process plant. In this pa-
per however, there is no direct consideration of the robustness
0
of this fault detection methodology against system uncertainty.
−0.5 Research is currently being carried out in this area.
−1

−1.5 References
−2
0 5 10
Hours
15 20 25 [1] S.P. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan.
Linear Matrix Inequalities in Systems and Control The-
Figure 4: The response of the flash inlet temperature    . ory. SIAM: Philadelphia, 1994.

x 10
−3 [2] Y. Cao, D. Rossiter, D. Edwards, J. Knechtel, and
4
D. Owens. Modelling issues for control structure selec-
2 tion in a chemical process. Computers and Chemical En-
gineering, 22(ss):411–418, 1998.
0

[3] C. Edwards and S.K. Spurgeon. On the development of


−2
discontinuous observers. International Journal of Con-
−4
trol, 59:1211–1229, 1994.

−6
[4] C. Edwards and S.K. Spurgeon. Sliding Mode Control:
Theory and Applications. Taylor & Francis, 1998.
−8
0 5 10 15 20 25
Hours
[5] C. Edwards and S.K. Spurgeon. A sliding mode observer
based FDI scheme for the ship benchmark problem. In
Figure 5: The response of the hydrogen to aromatic ratio. Proceedings of European Control Conference, Karlsruhe,
0.04
1999.
0.03
[6] C. Edwards, S.K. Spurgeon, and R.J. Patton. Sliding
0.02 mode observers for fault detection and isolation. Auto-
0.01 matica, 36(4):541–553, 2000.
0
[7] P. Gahinet, A. Nemirovski, A.J. Laub, and M. Chilali.
−0.01
LMI Control Toolbox, Users Guide. The MathWorks,
−0.02
Inc., 1995.
−0.03

−0.04
[8] G. Hermann, S.K. Spurgeon, and C. Edwards. Model-
based control of the HDA-plant, a non-linear, large
−0.05
0 5 10
Hours
15 20 25
scale chemical process, using sliding mode and ap-
proaches. UKACC Control 2000, 2000.
Figure 6: The response of the flash vapour outlet pressure     .
[9] R. Izadi-Zamanabadi and M. Blanke. A ship propulsion
system as a benchmark for fault-tolerant control. In Pro-
all other parameters stay within their specified ranges (Table 2) ceedings of the IFAC Symposium - Safeprocess ’97, pages
despite the presence of the disturbances. This shows that the 1074 – 1081, 1997.
PID controller provides good disturbance rejection (the distur-
[10] J.M. Maciejowski. Multivariable Feedback Design.
bances have major effects on the open loop system performance
Addison-Wesley, 1989.
[8]) and more importantly, the faults/disturbances could be re-
constructed faithfully even though the outputs in the closed [11] C.P. Tan and C. Edwards. An LMI approach for designing
loop appear unaffected by the disturbances. sliding mode observers. IEEE Conference on Decision
and Control, CDC 2000, Sydney, pages 2587–2592, 2000.
7 Conclusion [12] V.I. Utkin. Sliding Modes in Control Optimization.
This paper describes the use of sliding mode observers for Springer-Verlag, Berlin, 1992.
fault detection/disturbance estimation. A new observer design [13] B.L. Walcott and S.H. Żak. State observation of nonlin-
method was considered which is based on Linear Matrix In- ear uncertain dynamical systems. IEEE Transactions on
equality optimisation. Sliding mode observers have previously Automatic Control, 32:166–170, 1987.
been used for fault detection and disturbance estimation. How-
ever the systems for which they have been designed were of

Proceedings of the European Control Conference 2001 486

You might also like