Proceedings of the 2005 IEEE TuA02-5

International Symposium on Intelligent Control

Limassol, Cyprus, June 27-29, 2005

Model-based fault diagnosis using sliding mode

observers to Takagi-Sugeno fuzzy model

B. Castillo-Toledo and J. Anzurez-Marin


Abstract— In this paper, we present some results obtained
from the application of a class of sliding mode observers to the
model-based fault diagnosis problem in non-linear dynamic
systems. A Takagi-Sugeno fuzzy model is used to describe the
system and then Sliding mode observers are designed to
estimate the system state vector, from this the diagnostic
signal-residual is generated by the comparison of measured
and estimated output. The technique was proven satisfactorily
in simulation and real-time in a two-tank hydraulic system.
Keywords: Fault diagnosis, Fuzzy system, Takagi-Sugeno
fuzzy models, Sliding Mode Observers.

I. INTRODUCTION
Fig. 1. Model-based fault diagnosis structure

T he availability, cost efficiency, reliability, operating

safety and environmental protection are very important
characteristics in modern control systems. For critical
The majority of model-based fault diagnosis methods are
based on linear system models. For non-linear dynamic
safety systems, the consequences of faults can be extremely systems, the fault diagnosis problem traditionally has been
serious in terms of human mortality, environmental impact analyzed in two steps: First, the model is linearized around
and economic losses. Therefore, there is a growing operating point, and second, some techniques are applied to
necessity for supervision and fault diagnosis to increase the generate signal-residual, e.g. kalman filter, observers,
reliability of such systems. parity relations, parameter estimation [11].
There are three main approaches to deal with fault Therefore, in this work, we use the model-based
diagnosis [2], [4], [5], [7]: approach in order to deal with the fault diagnosis problem
x Signal analysis-based. and the technique used in [3]. The strategy is based in
x Knowledge-based. obtaining a Takagi-Sugeno fuzzy model and then we
x Model-based. propose to use sliding mode observers to estimate the
The model-based approach to fault diagnosis in dynamic system state vector; from this the diagnostic signal-residual
processes has been receiving considerable attention since is generated by the comparison of measured and estimated
the beginning of 1970, both in the research context and in output.
the domain of application on real processes [1], [2]. Figure The paper is organized as follows: In section 2 some
1 shows the general and conceptual structure of a model- basic concepts on fault diagnosis theory are shown. In
based fault diagnosis system comprising two main stages: section 3 we describe the problem and outline a solution. In
Residual generation: Its purpose is to generate a signal- section 4 we describe an application example. In section 5
residual, indicating a fault. we show the simulation and real-time results and finally we
Decision Making: The residuals are examined for the present some conclusions.
likelihood of fault and a decision rule is then applied to
determine if some fault appears. II. MODEL-BASED FAULT DIAGNOSIS BASIC
CONCEPTS.
Work supported by Universidad Michoacana de San Nicolas de Changes in the behavior of the system, due to external
Hidalgo with grant UMICH-132.
B. Castillo-Toledo and J. Anzurez-Marin are with the Department of
inputs, outside of the limits of tolerance specified according
Automatic Control, CINVESTAV-IPN. Unidad Guadalajara, A.P. 31-438, to engineering approaches will be considered as "faults";
Guadalajara, Jalisco 44550, Mexico therefore, the fault diagnosis concept will refer as the
(email:[toledo], [janzurez]@gdl.cinvestav.mx)
problem of detecting and localizing the fault, i.e., not only

0-7803-8936-0/05/$20.00 ©2005 IEEE 652

recognize the presence of a fault, but also which part of the observers is not an easy task. Therefore, in this work we
system causes it [8]. use of the so-called Takagi-Sugeno (T-S) fuzzy model,
A traditional approach to fault diagnosis in the wider where the non-linear trajectories are linearized around
application context is based on “hardware or some operating points. The set of linear models is used to
physical/parallel redundancy” methods, which use multiple realize a fuzzy interpolation and to achieve an
range sensors, actuators, computers and software to approximation of the non-linear system behavior [9].
measure and/or control a particular variable. It is possible
to use different measured values and their combinations
instead of duplicating each component individually to
smooth the conflict between the reliability and the cost due
to the additional components. This is the concept defined as
"functional or analytical redundancy", because takes
advantage of the redundant analytical relationships between
several measured variables in the monitored process [11].
In analytical redundancy schemes, the resulting
difference (signal) generated from the comparison of
measured and estimated output is called "residuals or
symptoms, r". The residual should be zero-valued when the Fig. 2. Residual generation scheme with sliding mode observers.
system is in normal condition and naturally should diverge
from zero when a fault (f(t))occurs in the system; therefore, The fuzzy model proposed by Takagi and Sugeno is
this property of the residual is used to determine whether or described by fuzzy IF-THEN rules which represent local
not faults appears in the system [6], [7], i.e. the residual linear input-output relations of a nonlinear system. The
must satisfy the following condition: main feature of a Takagi-Sugeno fuzzy model is to express
the local dynamic of each fuzzy implication (rule) by a
­°| 0 Ÿ f t 0 (normal) (1) linear subsystem [9]. To be precise, the ith rules of the T-S
r t ®
°̄  0 Ÿ f t z 0 (faulty) fuzzy models are given by the following form:

The major advantage of the model-based approach is that IF z1 t is M i1 and ! and z p t is M iP (2)
no additional components (hardware) are needed in order to THEN
perform the fault detection and identification (FDI) x
½
x t Ai x t  Bi u t  Ei d t °
algorithm; because the analytical redundancy uses a ¾ i 1....r.
mathematical model of system is that this approach is y t Ci x t °¿
known as "model-based approach". where; x t n is the state vector; y t q is the output
Figure 1 show that an essential problem in the model-
vector, u t m is the known input vector and d t p is
based FDI is the generation of the residual, since if it is not
designed correctly it could lose important information on an unknown input vector, usually containing external
the faults; therefore, different methods exist for the disturbance or faults in the system. Ai, Bi, Ci and Ei are time
residuals generation [11]. In this paper, we develop the invariant matrices of appropriate dimensions. The vector
output observer method and in particular, sliding mode z(t) is termed the premise variable, whereas Mij is a fuzzy
observers. This class of observer is very useful and was set and r is the number of IF-THEN rules.
developed for many reasons, among which; Given the input vector u t , the global state and output
i) The possibility of working with reduced observation of the system are inferred as follows:
error dynamics.
ii) A finite time convergence for all the observables < r

states. x t ¦Z
i 1
i z t ª¬ Ai x t  Bi u t  Ei d t º¼
(3)
iii) Robustness under parameter variations is possible. r
The residual generation scheme may be summarized as y t ¦Z z t Ci x t
i
in the figure 2. i 1

where; Zi z t is the grade of membership of the z(t) in

III. OBSERVERS DESIGN
Mij and notice that Zi z t should satisfy the following
In the model-based approach is necessary to build the
constraints [9]:
system mathematical model to be observed. In the general
case of non-linear dynamic systems, the design of

653
 r < r r

¦Z
i 1
i z t 1 (4) e t ¦¦Z i z t Zj z t ª Ai  Li C j
¬ (9)
i 1 j 1

0 d Zi z t d 1, i 1,..., r x t [ t  E i d t  M i t º¼
A. Sliding mode observers to Takagi-Sugeno fuzzy This way;
r r
model. <
e t ¦¦ Z i z t Z j z t ª¬ Aij e t  Ei d t  Mi t º¼ (10)
In a non-linear dynamic system described by its i 1 j 1

corresponding T-S fuzzy model (3), for the fuzzy observer where; Aij Ai  Li C j .
design, it is assumed that the fuzzy system model is locally
observable, i.e., all (Ci,, Ai) (i=1,…r) pairs are observable
Let the following Lyapunov function;
and it is required to satisfy: x t  [ t o 0 as t o f .
where; [ t   n denotes the state vector estimated by the v t eT t Pe t (11)
fuzzy observer.
We propose the design of a fuzzy observer structure with Deriving:
sliding mode or sliding mode observer (SMO), which uses < < <T

a number of local linear time-invariant observers; each v t eT t P e t  e t Pe t (12)

local observer is associated with each fuzzy rule given as:
Using (10)
IF z1 t is Mi1 and……and zp t is MiP (5) < ­ r r ½
v t eT t P ® ¦ ¦ Z i z Z j z ª¬ Aij e t  Ei d t  M i t º¼ ¾ 
THEN ¯i 1 j 1 ¿
< T
½ ­ r r ½
[ t Ai[ t  Bu t  Li y t  yˆ t Mi t °
i
¾ i 1!r. ® ¦ ¦ Z i z Z j z ¬ª Aij e t  Ei d t  M i t ¼º ¾ Pe t
yˆ t Ci[ t °¿ ¯ i 1 j 1 ¿
(13)
where M i t   n is the discontinuous vector of sliding mode
and z1 ,! , z p do not depend on the state variables < r r

^e T
P Aij  Aij T P e t 
v t ¦¦Z i 1 j 1
i z Zj z t
estimated by a fuzzy observer.
We used the idea of PDC (“Parallel Distributed 2 e T t P E i d t  2 e T t PM i t `
Compensation”) analyzed in [9], the overall state (14)
estimation is a combination of individual local observer Then;
outputs. < 2
v t d J e t 
r r
x
[ t ¦Z i z t ^ A[
i t  Bu
i t  Li ª¬ y t  yˆ t º¼  Mi t ` ¦Z i z ^2 eT t P Eid t  2 e T t PM i t `
i 1 (6) i 1

r (15)
yˆ t ¦Z i z t Ci[ t <
i 1 In order to have v t d 0 , we propose that the
where the weights Zi z t are the same as the weights
discontinuous term has the following form:
used in T-S fuzzy model (3): eT t P
To analyze the convergence of the SMO, the state Mi t ki T Ÿ Mi t ki sign eT t P (16)
estimation error ( e t x t  [ t ) dynamics given by the e t P

following equation is examined. where; ki ! 0 , ki = constant and P ! 0 so that the following

Lyapunov inequality is satisfied;
x x x
e t x t [ t (7)
PAij  Aij T P  0 (17)
Using (3) and (6); From this, lim e(t ) 0 if taking (15) and (16) is fulfilled
r r t of
<
e t ¦ ¦Z z t Zj z t ª¬ Ai x t  [ t  that;
i 1 j 1
i
(8)
r ­° eT t P ½° (18)
Ei d t  Li C j x t  [ t  M i t º¼ ¦ Zi z ®2 eT t P Ei d t  2eT t Pki ¾d0
eT t P ¿°
i 1|
¯°
i.e.

654
 r < Ke 1 (24)
2 eT t P ¦ Zi z
i 1
^ `
Ei d t  ki  0 (19) w1 t
T
v1 t  w1 t
T
< Ke 1 (25)
w2 t v2 t  w2 t
Then we found the following condition; T T
where f e is input flow, w1 and w2 are the opening of the
ki ! Ei d t (20) valve 1 and 2 respectively, h1 and h2 are the heights of the
liquid in each tank, K e is the static gain, T is the constant
where ki is chosen sufficiently big to satisfy (20), i.e. we
of time, v1 and v2 are inputs of control to system, At is a
tune the value of ki in order to have Ei d t ! ki when a
transversal area of each tank and f s1 .and f s 2 they model
fault appears.
the faults in the sensors 1 and 2 of level of the dynamic
Furthermore, it is known that an observer with
system.
converging state estimation can be referred as a stable
In the non-linear model of the system (22)-(25), we
observer, it can be proved that the stability of the (10) can
consider faults in the sensor 1 and 2. Then in this paper we
be verified by mean of the theorem of asymptotically
reported the simulation results when faults in both sensors
stability of fuzzy observers given by Tanaka and Wang in
happen and the application in real-time when a fault
[9]. For witch the inequalities (17) can be efficiently solved
appears in sensor 1. So far we are not considering the faults
via numerical approach within LMI framework [13] and at
in actuators (Eq (24)-(25)).
the same time to obtain the gains Li of (6) like in [3].
In equation (22) and (23), we can observe that the two-
Therefore, if a positive definite matrix P exists, the sliding tank system is sensitive to certain faults ( f s1 , f s 2 ) which
mode fuzzy observer is defined.
Once the state or output is estimated, the signal-residual causes that the signal-residuals is active or not, usually
can be generated by the comparison of measured and represented by the binary values “1” and “0”. This case is
estimated output. show in table 1 termed fault signature or fault code [11];
TABLE I
r t y t  yˆ t (21) FAULT SIGNATURE
Faults fS1 fault in sensor 1
fS1 fS2 fw1 fw2 fS2 fault in sensor 2
IV. APPLICATION EXAMPLE. r1 1 1 0 0 fw1 fault in actuator 1
The case of application is a non-linear dynamic system, r3 0 1 0 0 fw2 fault in actuator 2
which is formed by two interconnected tanks as is illustrate
A. Takagi-Sugeno fuzzy models
in figure 3. The system output is composed by the levels of
the two tanks and the inputs are the voltages to the valves. For T-S fuzzy model, the rule number is normally
determined by the modeling accuracy required. A large rule
number generally leads to higher accuracy; however, model
complexity should also be considered. For the two-tank
hydraulic system analyzed in this paper only 2 rules are
used, i.e. 2 models working in operation points that
correspond at two reference levels: h1=25 centimeter and
h2=35 centimeter. The rules are summarized in table 2.

TABLE II
RULES FOR TAKAGI-SUGENO MODELS, LEVELS IN METERS
Rule h1 h2
1 25 25
Fig. 3. Two-tank hydraulic system schemes. 2 35 25
h1=level in tank 1 in centimeter; h2=level in tank 2 in centimeters.
The dynamic model of the system is defined as follows
[10]; For the fuzzy sets associated to the premise variables
z(t), we considered the membership functions that are
< f e w1 t (1  f s1 )h1 t
(22) illustrated in the figure 4, .where z(t) is directly formed by
h1 t 
At At the state variables h1 and h2, i.e. the level of liquid in the
tanks.
< w1 t (1  f s1 )h1 t w2 t (1  f s 2 ) h2 t Then, the T-S fuzzy model for specific operation points
h2 t  (23)
At At in the two-tank system can be defined by (3).

655
 ª w01 h01 º 1) Case 1: System without faults.
«  0 0 »
« 2 At h01 At Figure 5 shows the residuals 1 and 3 for a fault-free
where »,
« 1 » system during 18 minutes. Where we can observe that the
« 0  0 0 »
Ai « T » signal-residual remain without change, around zero.
« h01 h02 »
« w01 
w02
 »
« 2 At h01 At 2 At h02 At »
« »
« 1 »
«¬ 0 0 0 
T ¼»
ª1 º
ª 0 0 º
«K » «0»
ª1 0 0 0 º and
« e 0 » , Ci «0 0 1 0 » Ei « »
Bi «T » ¬ ¼ «1 »
« 0 0 » « »
«
«
»
Ke » ¬0¼
«¬ 0 T »¼

Therefore, with the following operating points (X01, X02)

we obtain the fuzzy models or subsystems.
x01 ª¬h01 0.25, w01 0.20795e3, h02 0.25, w02 0.20795e3 º¼
Fig. 5. Residual due to systems faults-free
x02 ª¬h01 0.35, w01 0.17574e3 , h02 0.25, w02 0.20795e3 º¼
2) Case 2: System with faults.
Figure 6(a) shows the behavior of residuals during 40
minutes. An abrupt fault was simulated in sensor 1 (h1) at
20 minutes, i.e. Ei d t ! ki . As we indicate in table 1 only
the residual 1 (r1) should be activated, in this simulation the
residual 3 is also activated but its magnitude is smaller than
residual 1, therefore we may assume that the fault appears
only in the sensor 1. Figure 6(b) shows residual 1 and 3
active when we simulated an abrupt fault in sensor 2 (h2) as
we indicate in table 1.

Fig. 4. Membership functions of the fuzzy model.

The SMO (6) is designed to estimate the system output

and generate the diagnostic signal-residual (21) that
indicates whether faults appear or not. As two rules are
involved in the system model, two local observers are
required in the fuzzy observer. In this case we chose a
ki 10 to satisfy (20) and it is used in (16).

V. RESULTS
The method was proven in the two-tank system both in
simulation and real-time, the initial condition for the system
was x(0)=[0.25 0 0.25 0]T and has been carried out around
Fig. 6(a). Residual due to fault in sensor 1 and 6(b). Residual due to fault
the an operating point. In this work we only have been in sensor 2
considered abrupt faults in the liquid level sensors, i.e. the
total lost of the element in the system. 3) Case 3: Parametric variations.
Figure 7 illustrates the result of the system due to
A. Simulation parametric variations on the plant. In a test of 40 minutes
Comparing of measured and estimated output and taking we modify the parameters by means of a matrix change, Ai.
considerable attention in the signal-residual obtained, we We can observe that the residuals have small changes,
analyzed three cases mainly: whose behavior is similar to the shown in the figure 5, i.e.
the SMO from this point of view is robust under parametric

656
variations. fuzzy model. In the example of section 4 we can observe
that robustness is achieved under parametric variations of
the system, i.e. the residual only depend of external inputs
or fault in the system. Therefore, the sliding mode
technique provides a way to find a possible solution to this
problem. Furthermore, the application in real-time shows
the validity of this approach and the convergence proof of
the aggregate observer is included.

Fig. 7. Residual to faults-free system with parametric variations

B. Real- time
In real-time we applied the algorithm of fault detection
proposed, to the two-tank dynamic system. In this case, we
disconnected (fault) the sensor 1 between t1= 20min and
t2=21.5min.
Fig. 9. Zoom of residual due to fault in sensor 1 in real-time.
Figure 8 shows the result of the test during 2400sec, at
1200sec the fault (disconnection) was presented in sensor
REFERENCES
1, this sensor still broke 1.5 minutes and then we can
[1] J. Chen & R.J. Patton, Robust Model-Based fault Diagnosis for
observe that the behavior is very similar to the result of Dynamic Systems, Kluwer Academic Publishers, 1999.
simulation shown in the figure 6(a), i.e. the residual 1 is [2] R. Isermann, Process Fault Detection Based on Modeling and
active and the residual 3 change very little. Thus the fault Estimation Methods- A survey, Automatica, Vol. 20, No.4, pp 387-
appeared in sensor 1, as it were indicated above in table 1. 404, 1984.
[3] E. A. Garcia & S. S. Flores, Fault Detection and Isolation based on
Takagi-Sugeno modeling, Proc. IEEE ISIC, Houston, Texas, pp
673-678, Octubre 2003.
[4] R.J. Patton, P.M. Frank & R.N. Clark, Fault Diagnosis in Dynamic
Systems: Theory and Applications¸ Prentice Hall, 1989.
[5] Ch. Edwards & S. K. Spugeon, Sliding Mode Control:Theory and
Applications, Taylor & Francis Ltd, 1998.
[6] E.Y. Chow & A.S. Willsky, Analytical Redundancy and the Design
of Robust Failure Detection Systems, IEEE Trans on Automatic
Control, Vol. 29, No.7, pp 603-614, July 1984.
[7] P.M. Frank, Fault Diagnosis in Dynamic Systems Using Analytical
and Knowledge-Based Redundancy-A survey and some new
results, Automatica, vol 26, No. 3, pp 459-474, 1990.
[8] E.A.García, Detectando fallas mediante redundancia analítica,
Ingenierías-FIME-UANL, vol IV, No. 11, 2001.
[9] K.Tanaka & H.O. Wang, Fuzzy Control Systems Design and
Analysis: a Linear Matrix Inequality Approach,John Wiley & Sons,
INC, 2001.
[10] M Maldonado, Teoría del Regulador Lineal, Tesis de maestría
Fig. 8. Residuals due to fault in sensor 1 in real-time
Cinvestav unidad Guadalajara, Junio 1999.
[11] J.J. Gertler, Fault Detection and Diagnosis in Engineering Systems,
Figure 9 is a zoom of figure 8 during 500sec around the Marcel Dekker, 1998.
fault; we can observe that in fact the residual 3 change very [12] Ch. Edwards, S. K. Spurgeon and R.J. Patton, Sliding mode observer
for fault detection and isolation, Automatica, vol. 36, pp 541-553,
little. 2000.
[13] S. Boyd, L.E. Ghaoui, E. Feron and V Balakrishnan, Linear Matrix
VI. CONCLUSION Inequalities in System and Control Theory, Society for Industrial
Applied Mathematics (SIAM), Philadelphia,1994.
We present in this paper an analysis for signal-residual
generation in the fault detection process by means of
sliding mode fuzzy observers design for Takagi-Sugeno

657