You are on page 1of 6

Copyright © 1996 IFAC .

13th Triennial World Congress, San FrancIsco, USA


7f-066

SLIDING MODE OBSERVERS FOR ROBUST SENSOR


MONITORING
F.J.J. Hermans· M.B. Zarrop·
* Control Systems Centre, UMIST, PO Box 88, Manchester M60 lQD, U.K.
e-mail: hermans@csc.umist.ac. uk

Abstract. The aim of this paper is to introduce the sliding mode observers in fault
detection and to investigate their potential and advantages. This is carried out within
a general framework which focusses on the design requirements relevant to fault de-
tection. To illustrate the potential, the techniques are used for sensor monitoring of
a non-linear plant under feedback control.
Keywords. Sliding Mode, Observer, Robust Estimation, Fault Diagnosis, Sensor Fail-
ures, Uncertainty.

1. INTRODUCTION zero. If the deviations of the model and real plant are
within a certain acceptable range the residuals should be
The early detection and isolation of system malfunctions virtually zero so that no fault is indicated. In the sliding
is an indispensable requirement for any system. There mode context one says that the observer is sliding i.e. it
are two major approaches to fault detection,using ei- is situated on its surface. If a fault occurs i.e. the mis-
ther non-model-based or model-based algorithms. Model match moves outside the acceptable range, the observer
based methods are reviewed in (Isermann 1984, Gertler is disturbed from its surface and sliding ceases. This slid-
1988, Frank 1990). ing effect can be used for fault detection and diagnosis.
Model based fault detection depends on the availability Despite this clear intuitive link between change detec-
of a mathematical model which is a good representation tion and sliding mode observers, relative few researchers
of the plant or system under consideration. However, have investigated the area. For example, (Sreedhar and
when using a model to check for plant malfunctions, Fernandez 1993) discusses the use of sliding mode ob-
there is always a possibility that modelling errors will servers if full state measurements are available.
give rise to a significant fault signal even when the sys- The paper is constructed as follows. Section 2 reviews
tem is operating properly. Techniques must therefore be the design of sliding mode observers. The section fo-
developed to produce residuals which are robust with re- cusses on the parts of the design which are considered
spect to such mismatches. One way to deal with this is important for its application in fault detection. In sec-
by a statistical analysis of the residuals. Another way tion 3, the effects of faults on the sliding mode observer
of achieving robustness is to make the residual gener- are investigated. Most time is spent looking at sensor
ator, which in most cases is observer based, invariant faults because this is the problem which is dealt with
to a certain set of mismatches. Methods like eigenvector for the real plant. Finally in section 4, the algorithm
placement (Patton and Kangethe 1989) and unknown in- is tested on a real plant using logged data. For con-
put observers (Clark 1989) achieve this objective up to a fidentiality reasons it is not possible to reveal specific
certain degree. In these algorithms the system structure information about the system under investigation.
is used to cancel possible mismatches but no attempt
is made to estimate them.
Another way to produce robust residuals is the use of
a Sliding Mode Observer, adapting ideas from the field 2. SLIDING MODE OBSERVER
of sliding mode control (Utkin 1992, Utkin 1977) and
(Edwards and Spurgeon 1994b, Edwards and Spurgeon 2.1 Canonical Form
1994a). The main idea is to generate unbiased estimates
of the system states despite the modelling errors. This is An uncertain dynamical system can be described in its
achieved by specifying extra conditions on the observer. most general form as follows:
These extra conditions involve the Sliding Surfaces. The
sliding surfaces are situated where the output errors are x(t) = Ax(t) + Bu(t) + E~(x, u, t) (1)
yet) = Cx(t) + FtfJ(u, t)

6530
with A E IRnxn , B E IRnxm, C E IRPxn, E E IRnxq the number of outputs (r ::; p). The requirement for this
and F E IRPxr. The vector function E{(x, u, t) rep- uncertainty vector is that it is bounded i.e.:
resents system uncertainties like system non-linearities
and model mismatches. The vector function F4>(u, t) IltfJ(u, i)11 ::; (T Vu E IRm, t "?, 0
models the possible uncertainties in the feed-forward
path of the system. 2.3 Design of Sliding Mode Observer
To be able to design a sliding mode observer, the math-
ematical model is transformed into a Canonical Form. 2.3.1. Matched System
The existence of this transformation is a necessary con-
In the case of matched uncertainties (q ::; p S n) and
dition for designing a sliding mode observer. The trans-
no feed-forward uncertainties (F = 0), the observer pro-
formation is given in (Edwards and Spurgeon 1994a).
duces error free state estimates when it is operating in
The canonical form representation of an autonomous sliding mode. A sliding mode observer is determined by
state space system looks like (input omitted for sim- its observer gain K and sliding gain K8/ (Edwards and
plification) : Spurgeon 1994a). For the design of these gain matrices
the system is considered in its canonical form, which is
XI(t) = AUXI(t) + AI2Z(t) + EI{(XI, z, t) (2)
given in (2), with El = 0 and F = O. The observer has
z(t) = A2lxdt) + A22Z(t) + E 2{(XI,z,t) the following structure:
y(t) = Iz(t) + F4>(u, i)
i l = AllXI + A l2Z + KI(Z - y) (4)
with All E IR(n-p)x(n-p) invertable, Al2 E IR(n-p)x p , i = A 2lXI + Anz + K2(Z - y) + K.w
m
A21 E IRPx(n- p ), An E pxp . El E m(n-p)x q , E2 E
m pxq , and I E IRpxI'. The vector Xl E m n - p are the with 11 a discontinuous function dependent on the ob-
free moving stat.es and the vector z E mp are the output server error. The observer gain K = [~~] is designed
related states.
to estimate the unobservable states. The sliding mode
gain K8/ on the other hand is such that it must compen-
2.2 Structure of Uncertainties sate for the system uncertainties. To be able to achieve
Depending on the structure of E, the system uncertain- this, the sliding surfaces of the observer must be at-
ties can be classified into two groups. The first kind of tractive. This means that after being disturbed from its
disturbance is one which can be matched. In this case surface, the observer is able to pull the state estimates
the uncertainty vector {(x, U , t) has q independent com- back. The condition of attraction is :
ponents with q ::; p and is bounded i.e. :

11{(x,u,t)ll::;p VxEIRn,uEmm,t"?,O (3)


where s = Z - y are the sliding surfaces.
This structure of the uncertainty interaction matrix E First the observer gain is considered. The only informa-
will make it possible to compensate perfectly for the un- tion available is y. This means that the dynamics of the
certainty at any point in time and for any {(x, u, t) pro- n - p non-output related states are not changeable by
file satisfying its bounds (Edwards and Spurgeon 1994a). feeding back the output errors. This is the reason why,
This results in state and output estimates which will be by generating the canonical form, a stable All is re-
error free. In canonical form (see (2)), El has all zero quired. For the p output states it is clear that feeding
entries. back the output errors enables a stable An to be de-
signed according to any requirement. The observer gain
The second kind of system uncertainty is referred to as
matrix then has the following structure:
non-fully-matchable. In this case q "?, p. The uncertainty
vector {(x, u, t) again has q independent components. If
Al2 ]
q = n then there areas many independent disturbances K = [A A8tab/e
22 - 22
as states. The El disturbance interaction matrix of (2)
no longer has only zero entries. Under certain situations The gain of the non-output states decouples these states
the condition on {(x, u, t) is the same as in (3). For this from the output related ones.
set of uncertainties, it will not be possible to get error For the sliding mode gain the following is considered.
free state estimates as well as outputs. Take the output equation in (2). Transform the z vector
The uncertainties Fc,)( u, t) in the feed-forward path are with the transformation matrix T to z = Tz. The reason
always matchable. The number of independent compo- for, and the exact form of, this transformation will be
nents in this vector is always smaller than or equal to discussed in section 3. The sliding mode gain K8/ of

6531
the transformed system must be designed such that the where e z = z - z is the error on the output state. The
surfaces s are attractive. This is obtained by calculating disturbance term is then:
the unique solution P2 of the Lyapunov equation for the
stable A2~able and symmetric positive definite Q2. The
sliding mode gain for the output related equation in (2)
Normally, { and tjJ are two independent uncertainties
is then:
such that for designing a sliding mode observer a bound
on en
instead of the two separately must be considered.
The design of the observer gains is similar to that for
the matched system where the uncertainty interaction
The term IIE;P;II is introduced to compensate for the matrix is the unity matrix.
scale of Q2.
The discontinuous function v is given by:
psign [T(z-y)] ifz-Y:f0 3. FAULTS AND SLIDING MODE OBSERVER
v ={ 0 otherwise
3.1 Introduction
2.3.2. Non-Fully-Matchable System The aim of fault detection is to generate residuals which
As explained in section 2.2, it is not possible to trans- are sensitive to certain sets of faults. These residuals are
form the effect of all the inaccuracies in the output normally functions of the output errors. Here, in the case
states. This means t.hat El does not only have zero en- of sliding mode observers, it means that the residuals
tries. The feed-forward uncertainties are again not con- are functions of the sliding surfaces. The first thing to
sidered (F = 0). The effect of the uncertainties on the do is to investigate how a fault interacts with the sliding
Xl states is:
surfaces and how the sliding performance of the observer
is affected.
(5) The faults which are of interest are actuator and sensor
This means that the disturbance on the output related faults. Actuator faults are uncertainties which interact
with the process equations. These faults are represented
states becomes:
by Mfl(t) in (7). The sensor faults only affect the out-
puts and are not influenced by the system dynamics.
These faults are modelled by f2(t) in (7).
To calculate the sliding mode gains, the unit matrix
must be used as interaction matrix and a bound on en x(t) = Ax(t) + Bu(t) + E{(x, u, t) + Mfl (t) (7)
instead of { must be used. In the situation where the dis- y(t) = Cx(t) + FtjJ(u, t) + f2(t)
turbance does not change very quickly the disturbance
en can be approximated as follows: with M E IRnxr , fl and f2 are respectively rand p
dimensional vectors.
en ~ A2IAIIIEI{ + E 2{ (6)
To generate residuals for fault detection and diagnosis,
This expression should now be used instead of the sim- a robust sliding mode observer is used. An indication
ple E2e to calculate the sliding mode gain K,l, where that the real system and its mathematical model corre-
E2 = A21A 11lEl + E2 is the new uncertainty interac- spond, within the uncertainty bounds, is obtained by the
tion matrix. Using this E2 instead of E2 in the design fact that the observer is sliding. When the mismatch is
explained for the matched situation, it is possible to outside the robustness boundaries, a fault has occurred
generate an observer with sliding mode characteristics. and the observer does not slide anymore. This means
The output related state estimate errors are zero but that the sliding surfaces can be used to form residuals
the non-output state estimates are not error free. This =
i.e. r(t) s(t).
error is given in (5).
3.2 Actuator Faults
2.3.3. Matched systnn with Feed-forward uncertainties
To investigate actuator faults, the situation of matched
For this situat.ion, the interaction matrix El has zero uncertainties is considered because non-fully matchable
entries and F is not. zero. Taking feed-forward uncer-
ones only affect the non-error-freeness of the state es-
tainties into consideration, the error related to the esti- timates and not the sliding capabilities. The system is
mation of t.he out.put. relat.ed states z is affected by the considered to be transformed into its canonical form.
uncertainties tjJ( u, t) as follows: The dynamics of the state errors and the sliding condi-
ez = A21 ex, + A~~ableez - E2~ - K2Fq. - F~ + K.w tions for a system with actuator faults are:

6532
6533
6534
6535

You might also like