Auromnrico, Vol. 33, No. 2, pp.

217-225, 1997
0 1997 Elsevier Science Ltd
PII: sooos-10!&%)00155-0 Prtnted m Great Bmam. All rights reserved
OCQ5.1098/97 S17.W + 0.00

Brief Paper

On the Use of Adaptive Updating Rules for Actuator

and Sensor Fault Diagnosis*
HONG WANG,-F ZHEN J. HUANGS and STEVE DALEY§

Key Words-Fault detection and diagnosis; adaptive updating rules; linear systems; observers;
actuators, sensors.

Abstract-A novel approach is presented for the fault In observer-based FDD approaches. observers are
detection and diagnosis (FDD) of faults in actuators and constructed based upon a knowledge of the parameter
sensors via the use of adaptive updating rules. The system matrices in the state-space representation of the healthy
considered is linear time-invariant and is subjected to an system. This enables fault detection to be performed by
unknown input that represents either model uncertainty or checking the observation error in relation to a pre-specified
unmeasurable disturbances. First, fault detection and threshold (Frank, 1994). As for diagnosis, either a single
diagnosis for linear actuators and sensors is considered, observer (Massoumnia, 1986; White and Speyer, 1987) or a
where a fixed observer is used to detect the fault whilst an number of observers (Ge and Fang, 1988) are connected in
adaptive diagnositic observer is constructed to diagnose the parallel to the system, whereupon residual signals are
fault. Using the augmented error technique from mode1 generated, with each being sensitive to one particular fault
reference adaptive control, an observation error model is but insensitive to others. Since the parameter matrices of the
formulated and used to establish an adaptive diagnostic system are assumed to be exactly known, difficulties will most
algorithm that produces an estimate of the gains of actuator certainly arise when applying these methods to practical
and the sensor. An extension to the fault detection and systems. As a result, researchers have sought to enhance the
diagnosis to cover nonlinear actuators is also made, where a robustness of observer-based FDD such that the generated
similar augmented error model to that used for linear residual signals are sensitive to the fault but insensitive to
actuators and sensors is obtained. As a result, a convergent model uncertainties and external disturbances. This has lead
adaptive diagnostic algorithm for estimating the parameters to the use of unknown input observers (Viswanadham and
in the nonlinear actuators is developed. Two simulated Srichander, 1987; Phatak and Viswanadham, 1988; Wuen-
numerical examples are included to demonstrate the use of nenberg and Frank, 1988: Chen and Zhang, 1991; Hou and
the proposed approaches. 0 1997 Elsevier Science Ltd. All Muller, 1994), where the model uncertainties and slowly
rights reserved. varying parameters are treated as unknown inputs to the
system. However, these methods are still limited to systems
1. Introduction whose dominant parts are linear and known (Ding and
Owing to the increasing demand for high reliability for many Frank, 1993).
industrial processes, fault detection and diagnosis (FDD) Since in general the residual signals are functions of faults,
algorithms and their application to a wide range of industrial model uncertainties and unknown parameters of the system,
and commercial processes have been the subjects of intensive adaptive observers have been proposed (Sidar, 1983; Li and
investigation over the past two decades. Many methods have Zhang, 1993: Frank ef al., 1991: Ding and Frank, 1993) to
been developed, and examples include the robust- further enhance the robustness. In these approaches,
redundancy-relation approaches (Chow et al.. 1986; Bas- adaptive observers are used to track the slow variations of
Seville, 1988; Frank, 1994) the observer-based approaches unknown system parameters, and the estimated parameters
(Patton et al., 1989; Patton and Chen. 1991; Viswanadham are then used to evaluate the residual signals in order to
and Srichander, 1987; Ge and Fang, 1988; Park and Rizzoni, reduce the number of false alarms. As a result, the sensitivity
1994), identification-based methods (Isermann, 1984; Freyer- of residual signals to both model uncertainties and slowly
muth, 1991) and the neural-network-based methods (Zhang varying parameters is decreased. In fact, adaptive observer-
and Roberts, 1992: Patton et al., 1994; Wang et al., 1994). based FDD combines the observer-based FDD methods with
These methods have proven to be capable of successfully with parameter-estimation FDD approaches, and is capable
detecting and diagnosing certain types of system faults, which of simultaneously estimating unknown parameters, slowly
are generally assumed to be unexpected changes of some varying faults and the state of the system (Sidar. 1983; Li and
parameters in the system (Daley and Wang, 1993). Zhang, 1993: Ding and Frank, 1993). Since fault detection
and diagnosis for abrupt faults is performed through the
* Received 3 April 1995: revised 3 November 1995; revised evaluation of adaptive residual signals, the implementation of
15 April 1996; received in final form 20 June 1996. This paper the overall structures are somewhat complicated. This is true
was not presented at any IFAC meeting. This paper was especially when a number of adaptive observers are
recommended for publication in revised form by Associate connected to the system in order to diagnose different fault
Editor M. Basseville under the direction of Editor Torsten types. An added limitation is that the conditions for
Soderstrom. Corresponding author Dr. Hong Wang. Tel. convergence and stability are very complex. and a priori
+44 161 200 3859; Fax +44 161 200 3858; E-mail knowledge of the norms of the state and inputs is required
hong.wang.@umist.ac.uk. (Sidar, 1983).
t Department of Paper Science. UMIST, P.O. Box 88, As a result, simplified approaches for fault diagnosis are
Manchester M60 lQD, U.K. still in need of further development. This forms the main
$ Radio Technology Department, The South East Univer- purpose of this paper, where a simple adaptive diagnostic
sity, Nanjing, P.R. China. algorithm, together with the use of an observer-based FDD
5 Mechanical Engineering Centre, European Gas Turbines framework, is developed for the diagnosis of sensor and
Ltd, GEC Alsthom, Leicester LE8 3LH, U.K. actuator faults. which can be either abrupt or incipient.
217
218 Brief Papers

2. Detection observer matrices L.,, gu and fu in (4) and (5) are fixed. Therefore the
Assume that the system to be considered can be expressed observer (4). (5) is also referred to as a detection observer
by for the system (l), (2).
f(t) = Ax(t) + Bf(r)u(r) + v(t), (1)
3. An adaptive diagnostic observer and its error equation
YW = idtww + W(rMt)lv (2) To diagnose the fault after the alarm (9) has been
where x(t) E W” is the unmeasurable state vector, u(t) E W” generated, the adaptive diagnostic observer
is the measurable input vector and y(r) E w’ is the
i,,,(r) = Ax,(r) + Bf(r)u(r) + L&(r), (10)
measurable output vector; v(t) is a unmeasurable term
representing model uncertainties and the input noise of the E(t) = d(f)-l[Ym(o - YO)l, (11)
system. It also represents the output noises through the
dynamic equivalent principles. A, B, C and D are known Y,(f) = ~(r)[wl(r) + mrMr)l (12)
parameter matrices of dimensions n X n, n X m, r X n and is constructed, where x,(t) E tRmis the observer state vector,
r X m. respectively; f(t) and g(f) are m X m and r X r L E IV” is a pre-specified observer gain matrix that makes
time-varying matrices representing the gains of actuator and A + LC stable, andf(r) and g(t) are the estimates off(t) and
sensor connected to u(r) and y(r), respectively. Both sensor g(t) respectively. The initial values of f(t) and d(r) are set to
and actuator faults are considered. It is assumed that the pair fH and g, respectively. It is assumed that after t >fr, a fault
(A, C) is observable; (ii) g(r) is always a non-singular matrix;
occurs such that
(iii) v(t) is uniformly bounded (i.e. there is a known positive
number 6, such that [iv(r)I[ c Sc), and (iv) when no fault f(f) =f = constant +fH, (13)
occurs, we have
g(r) = g = constant f g,, (14)
dr) = gH. f(f) = fH. (3)
and that g is nonsingular. Moreover, the faults may occur
where g, and fH are known matrices and the subscript H either simultaneously or at different times. Denote
stands for healthy sensors and actuators.
The purpose of FDD is therefore to generate an alarm e,(t) =x,(r) - x(r). (15)
signal when a fault occurs and to produce an accurate
estimate of the matrices f and g that define the faulty Then it can be obtained from (1) and (10) that
behaviour of the actuator and sensor. For this purpose, the P,(t) = Ae,(r) + B&r) -f(t)]u(r) + LE(I) - v(t), (16)
following observer is constructed:
Y,(r) -Y(r) = g(tw,(r) + d(o~m4)
i(t) = Af(r) + &Hut?) + Ld[F(t) -Y(t)], (4) - gCx(r) - gDfu(t),
9tr) =gHtCftr) + DfHu(r)lt (5)
= g(r)[Q&) + Df(r)lu(t) +g(tk-‘Y(t), (17)
where Z(t) E KY’ is the state vector of the observer and f(1) =f(r) -J g(r) =8(r) -g. (18)
La E IX”“’ is the gain matrix such that A + Ldg,C is a stable
matrix. Define From (11) and (17). e(r) can be further expressed as
e(f) = f(f) -x(f), Wa) e(r) = Cc,(r) + Df(t)u(r) + [g-’ -g-‘(t)]y(r). (19)
edr) =9(f) -y(r). (6b) Substituting (19) into (16), it can be further obtained that
f(l) =fH -f(f)? g(t) = g, - g(t). (k)
C,(r) = (A + LC)e,(r) + (B + LD)y(r)u(r)
Then it can be shown that the equality
+ m -’ - g-‘(r)]y(t) - v(t). (20)
% =f(d -Y@) = gHcf(t) + gHDfHu(l)
As a result, the combination of (19) and l’20) gives the
- gWx(t) - gww(~M~) relationship between the estimation errors (j(t), g(t)) and
=gHfCetr) +D?(r)u(r)l +&?(tkm’(t)y(r) (7) the observation error c(t). The purpose of fault diagnosis is
to use available c(r), u(t) and y(r) to construct a diagnostic
holds. As a result, the observation error equation algorithm for d(f) and j(t) such that
p(r) = (A + L,g,C)e(t) + (B + LdgHD)f(rju(r)

+ Llmg-‘Y(r) - v(t) (84

Compared with the detection observer in Section 2, the
can be obtained. If no fault occurs, (3) holds and 7 = g = 0. adaptive diagnostic observer uses c(r) instead of the output
As a result, the second and third terms in (7) and (Sa) are error. Otherwise, the matrix A + LC in (20) will be of the
zero. This means that the observation error satisfies form A + Lg(t)C, which may not be regarded as a stable
t!(r) = (A + L,g,C)e(r) - v(r), WI
matrix. Since (19) and (20) are in a similar form to the error
equation between the reference model and the adjustable
edr) = gHCe(t). @cl part of the system found in model reference adaptive control
(Landau, 1979; Narendra and Annaswamy, 1989: AstrSm and
Since 11v(t) 5 a,,, it can be further obtained that Wittenmark, 1989; Sun and Ioannou, 1992) in the following
section an adaptive updating rule, which has been widely
iie,,(r)iI ‘zz; IigHccjw~ -A - &gHC)mlII 60 = h. (8d) used in adaptiye control, will be constructed and applied to
tune &f(r) and f(r).
which is calculable because all the matrices involved in (8d)
are known. Therefore the fault detection can be readily 4. Single-input and single-output systems
carried out as follows: Denote
A,,=A+LC, B,,=B+f.D. h=g-‘. (224
Ilec,(r)ll = IIf -y(r)11 1:: ii C~~U,fso~~~~ed, (9)
h(C) =g-‘(I). h(r) =h -Ii(r). (22b)

where A is the threshold defined by (Sd). Then (20) can be expressed as

This is the well known observer based fault-detection
method (Patton et al. 1989: Frank, 1994), where the gain i,(t) = A,,e,(r) + E,$(t)u(r) + LL(t)y(r) - v(r). (23)
 Brief Papers 219

When the system (1). (2) is single-input and single-output, can be applied. In (35), r = rl’>O is a pre-specified gain
three transfer functions matrix that defines the tuning rate.
The tuning rule (35) is referred to as the adaptive
W,(s) = C(s1 -A<,)-‘&, + D, (24) diagnostic algorithm for the actuator and sensor fault
W,(s) = C(sl - A,,)-‘L + 1, (25) diagnosis of the system (l), (2). Letting r = diag (y,, y2) and
&(s) = -C(sl -A,))-‘. (26) using the adaptive diagnostic algorithm (35) and the
definition of the vectors e(t) and C(t) in (29) and (30), it can
can be readily defined, which lead to the following be further shown that
input-output expression of the error equations (19) and (20)
in Laplace-transform form: d(t) &l(t)w(S)U(~) (36)
T = -y’ 1 + ((r)TIgt) ’
44 = wVwuw + K~WYW + ww). (27)
Since under fault conditions g(r) and f(t) take constant d&r) ~l(~)~Z~~Pw~)Y(~)
(IIE,(t)ll ‘h). (37)
values different from g, and f,,, it can be shown, using the dr = -y2 1 + cgt)Tg(t)
augmented error technique (Narendra and Annaswamy,
1989). that It is well known in adaptive control that a persistently
exciting signal u(t) can improve the accuracy of the
estimation. This signal may therefore be required in practice
for an improved diagnosis. However, the application of such
signals can cause problems for practical systems in which case
an exponentially decaying persistent signal should be used to
improve the accuracy of adaptive diagnostic algorithm (36).
(37). The overall fault detection and diagnosis procedure is
(30) shown in Fig. 1.
Although the term 8(t) in (28) tends to zero exponentially,
where 6(t) decays exponentially (Narendra and Annaswamy, the speed of this decay depends on the zeros of the
1989) and can be neglected. Since the second and third terms denohinators of transfer functions W,(s) and W,(s), which
of (28) are all measurable, we can define are part of the eigenvalues of matrix A + LC. Therefore, to
ensure the validity of neglecting 6(r) and to provide a good
El(t) = a4 - wdsm -fow,(s)lw convergence of the adaptive diagnostic algorithm (35), the
+ Ms)m - mw)1Yw (31) fixed gain matrix L should be selected such that the
eigenvalues of the matrix A + LC are not only placed in the
as the augmented error signal for the system. As a result, it left-hand side of the complex plane, but are also far from the
can be seen that imaginary axis.
PI(I) = eT(Mr) + VII. (32)
vu = Ws)W). (33) 5. Mulriple-input and multiple-output systems
When n or r>l. the original system (l), (2) is
l/V,~ll 5 ‘E7-,“; llw,ciw)ll 6, = A,. (34) multiple-input and multiple-output. In this case, the transfer
functions W,(s) and W,(s) are r X m and r X r matrices with
Since W3(t) is known after L is selected, A, is a known WY(S) and WY(S) being their ijth elements respectively.
number, which can be used to construct a dead zone for an However, owing to the non-commutative property of matrix
adaptive adjusting algorithm. Equation (32) is in the standard multiplications, a compact matrix form such as (28) cannot
format of the error equations widely used in adaptive control be obtained directly. In this section, we shall use the
(Narendra and Annaswamy, 1989, equation (8.24) of Chapter component form to derive the required error equation (32)
8). The adaptive updating laws used in adaptive control can for MIMO systems.
therefore be directly applied to tune e(t). This is summarised Since L can be selected such that A + LC is stable, it can
by the following theorem. be seen that all th_e WY(S) a_nd w$(s) are stable transfer
functions. Denote Jj(r) -and h,(r) as the ijth elements of
Theorem 1. With the error model (32) and the fact that v(s) the matrices f(r) and h(t) respectively. Then the MIMO
(i = 1,2, 3) are stable, and that s(t) decays exponentially and version of the error equation (27) becomes
can be neglected, the tuning rules
e,(r) = 2 5 w:'(s)ff~(t)u,(t) + i i wF(s)hi,(r)Yj(r)
-r E1(05W
,=,,=I ,=I ,=I
(II~l(~)ll ‘Al).
e(t) = 1 + CT(W)
(35)
+ vlI.k (k = 1. 2.. , r), (38)
0 (Il&,(t)ll SA,)
where ek(t) is the kth component of the vector I, v,,,,(t) is
realizes a bounded e(t) and b(t) e L2. the kth component of vector v,,(l). u,(t) and y,(r) are the jth
components of the vectors u(t) and y(l), respectively. Using
The proof of this theorem is omitted here, since the same the same technique as in Section 4, it can be seen that
formulations presented by Narendra and Annaswamy (1989)
w:‘(~fJ~,(t)u,(t) =~,~t~~:‘(s)u,(~)
+ [w:‘(slf;,o) -f#,wdY~N~,u) + h(t), (39)

W?(SYi,WY,(~) = ~,,(a4’(S)Y,(~)

+ [wS’(s)h^q(t) - h^t,(r)w~‘(s)lYj(t) + sk(r), C40)

where s,(t), again, decays exponentially (Narendra and

Annaswamy, 1989). Let

5ktr) = ,z ,z {[w?‘(s)~~(t) -ff~(~)wff(s)luj(t)~

r I
+ c c Ilw:‘(sh,w - fi&h:‘(s)lYjw~
,=-L ,=I
Fig. 1. Structure of the proposed method. + 26,(f). (41)
220 Brief Papers

Then it can be obtained from (38) that Because the signals r&(t), uj(r) and yj(r) are ah measurable,
and W,(s) and W,(s) are known after L is selected, the
%tr) = E 2 fij(r)w:i(s)uj(r) + i i htj(r)w:'(s)Yj(f) adaptive diagnostic algorithm (53) is realizable in practice.
i-l j-1 i-1 j-1 Using h(r) in (53b), g(r) can be readily obtained via the
+ [k(r) + v0.k. (42) inverse
Define
g(r) =/;-*(I). (54)
e = Vll, jl2, . . , L. 121, . , fh, ) fd, . ,
La, . . , I-, .. Since the adaptive diagnostic algorithm (53) is obtained via
the use of gradient descent rule and the perfomance function
RI,, 5 L?. . , R,,, . 9 L, / is quadratic with respect to 0, a global minimum of the
l-i,& . ., R,,lT E R”“+“, (43) performance function J can be guaranteed if X;=r {k(r)tl(r)
(k(r) = [wt’(s)u,(r), w:‘ts)u2(r) . Y w:‘(s)k,(t) is uniformly positive-definite (i.e. there is a positive number
w:2(s)u,(r), . . ) w:2(s)u,(r), , (I such that zi=, t(r)lT(‘(t) > al). However, this is a difficult
precondition for many systems. Further studies are
w:“(s)u,(I), . . I w:m(s)u,(r) necessary in order to tune f(r) and g(r) to ensure
w:‘(Sly*(t), 4’(S)Y*(f), . 1 wt’(s)Yr(t), .. , convergence and realize (21). However, in the case when the
w:‘(s)y,(r), . ) wf’(s)y,(r)]* E F”“‘“, original system is multiple-input and single-output, f(r) and
k = 1, 2, . . . , r. hence h(r)) becomes a scalar. This means that the vector u(r)
(44 consists of only one element. As a result, the output error
Then (42) can be expressed as equation (46) becomes
&k@) = @ck + tk + vO,k. (45) u,(r) = eTb(r) + v,, (55)
As a result, a measurable augmented error u*(t) can be which is in the same form as (32) of Section 4 without 8(t).
defined as In this case, the compact form such as the adaptive diagnostic
ok(r) = &k(r) - &k(r), (k = 1, 2,. , r). (44 algorithm (35) can still be used, which can guarantee the
convergence of the diagnostic algorithm when u(r) is
Using (45) it can be seen that persistently exciting.
uk(r) = eTlk(r) + vo,, (k = 1, 2,. . . , r). Wb) Another issue is the selection of threshold A, dead zone
A,, and adaptive learning rates yl, y2, Aij and pLlj. As for the
Similarly to Section 4, the adaptive diagnostic algorithm for threshold, (8d) is normally used for fault detection purpose.
matrices f(r) and h(r) should be such that This means that the threshold is selected as the maximum
lim a(r) = 0. (47) allowable magnitude of the output error e,,(r) when no fault
,-== occurs. Since 6, is assumed known, this threshold is
However, since all the equations in (46b) (k = 1,2,. . . , r) completely calculable from (8d). The dead zone A1 is in fact
are subjected to a single parameter estimation error vector 0 clearly defined by (34) since both the transfer function
and the existence of term vo, a compact form such as (35) for (matrix for MIMO cases) W,(s) and the upper bound on the
tuning f(r) and h(r) cannot be obtained. As a result, an input noise So are known. Therefore the most flexible parts
adaptive tuning algorithm for f(r) and h(r) should be selected are the adaptive learning rates, which determine the speed of
to minimize the magnitude of &k(r). This can be achieved by the fault diagnosis for the actuators and sensors. In general,
choosing the following performance index function: high learning rates will lead to fast convergence, whilst at the
same time resulting in adaptive diagnostic algorithms too
sensitive to noises. As a result, care should be taken in
selecting these parameters, and a good compromise is
needed. An optimal selection of these learning rates can be
Using the well-known gradient descent rule in the found in Wang (1987).
continuous-time domain and the dead zones in Section 4, the
adaptive diagnostic algorithm for f(r) and h(r) can be
formulated to give 6. Extension to nonlinear actuators
In the above sections, we have assumed that the actuators
and sensors are all linear. However, there are many practical
systems where the characteristics of actuators and sensors are
(49) nonlinear. For example, the saturation shown in Fig. 2 is
exhibited by many of the hydraulic valves and electric
actuators that are used widely in a number of industries.
Therefore it is necessary to extend the fault detection and
(50) diagnosis algorithms to cope with nonlinear actuators and
sensors in dynamic systems. This forms the purpose of this
where Arj and pJ are positive numbers representing the section, where the extension to nonlinear actuators will be
learning rates. From (46b) and (48) we have particularly addressed.
& =zl uk(r)wf’(sb,(r)s (51)

L = i uk(l)w:‘(s)yj(t). (52)
&jtr) k=l

The adaptive diagnostic algorithm (49)-(52) constitutes an

estimate of the matrices f(r) and h(r), which can be further
expressed as

-j&j C 2 uk(r)w’i’(s)yj(r) (Iluk(f)II > Ad,

k=l k=l

(53b) Fig. 2. Saturation in actuators and sensors.

 Brief Papers 221

System representofion. To simplify the presentation, the Theorem 2. Suppose that

system considered is expressed by (i) the gain matrix L for the adaptive observer (61) can be
selected such that the following conditions are satisfied:
i(f) = Ax(f) + Bf(O(f), u(f)) + v(f), (564
Y(f) = Cx(r), (A + IX)~P + P(A + LC) = -Q, (67)
(56b)
where f(6’, u(l)) (f: W’+m~R”‘) is an M x 1 known vector C = BTP, (68)
mapping representing the characteristics of the nonlinear
actuator and 0 E W is a parameter vector, which is assumed where P and Q are two positive-definite matrices;
to change unexpectedly when faults occur in the actuator. (ii) f(e. u(r)) is continuous and first-order differentiable;
Moreover, it is still assumed that the pair (A, C) is (iii) there is a known positive function K,(u(f))>O and a
observable and that, when no fault occurs, the equality positive known number K, such that for any e,, Bz E IF
the inequalities
e(r) = r3” (57)
me.u(N
holds, where OH is a known vector and the subscript H stands
for the healthy actuator. The purpose of fault detection and
Ipu4.4) -f(h u(r))- ae 1e=.¶,(6 - e21~~
diagnosis is therefore to generate an alarm signal when a 5 K&(Q). (69)
fault occurs and produce an accurate estimate of the vector 0 llell SK, (70)
that defines the faulty behaviour of the actuator.
hold, where @(0. u(t))/ae is the Jacobian matrix of
Fadf Similarly to the detection observer in Section
derecfion. f(t? u(r)) with respect to 0.
2, the following nonlinear observer is constructed:
Then the adaptive diagnostic algorithm

V(&L
u(r))
1T
I(f) = A.?(f) + Bf(e,. u(f)) + L&qf) -y(r)], (58a)
j(r) = Cf(f). (58b)
--(T&r) -
a&r) dt) (c(r), B(r) EDd,
where E(t) E lR8”is again the state vector of the observer and
y^(t) E 08”’ is the output vector of the observer. Since it has d&f)
-= T
been assumed that the pair (A, C) is observable, Ld E RnXm dt E(f) (E(f), B(r) E 0,)
I
can be selected such that A + L,C is a stable matrix. Define
(71)
e(f) = a(f) -x(f), Wa)
eO(r) = g(t) - y(f) = Cc(f). (59b) guarar$ees that, within a finite time period, the variables
(E(r), e(r)) converge to the set D, exponentially at a rate
Then the observation error becomes greater than ems’, where
e(t) = (A + &C)e(f) + B[f(&. u(f)) -f(e(f), u(r))] - v(f).
(60) (knin(Q)
S=min ~A,,,(P)‘u I , (72)

By taking Ile,,(f)II as a detection signal. the fault detection

can be readily ‘carried out in a srmilar way as has been DR= ((8(r). 6(r)) 1y lle(r)ll’+~ 11~0112
discussed in Section 2.
>K:+~~,,JQ) IIQ-w
Fault diagnosis. To diagnosis the fault, the observer
K,,(r) = Ax,(r) + W(h), u(r)) + L[Cx,(r) -y(r)] (61) x [ IIB ll’K:(Nr)) + S:,] + 61 , (73)
1
is co!structed, where x,(t) E R” is the observer state vector
and e(t) is an estimate of O(r). The value of h(l) is set to 0” D,=j(&(r).B(r)~~,,E(r),,2+~,,8(r)l12
until a fault is detected. It is assumed that after r 2 I, (or
Ile,,(t)(l > A in (9)) a fault occurs such that
-K:+$[MQ) IIQ-‘PII’
e(t) = e= constant f eH. (62)
Denote x [IIBII K:(u(r)) + G,] + aKi]), (74)

e,(r) = x,(r) -x(r), (63) and A,,,(Z) and A,,,(Z) are for the minimum and maximum
E(r) = cX,(f) -y(f) = Cc,(f). (64) eigenvalues of the positive-definite symmetric matrix Z.
Then it can again be obtained that
Proof. See the Appendix.
&n(r) = (A + LCk,(r) + B[f(&r), 41))
-00, u(t))1 - ~0) 0 > rd. (65)
7. Simulation examples
As a result. the purpose of fault diagnosis is to find a The system to be considered in the simulation is second
diagnostic algorithm for e(t) such that order and is given by

lim E(f) = 0, lim B(t) = e (66)

I--r= ,--rl 1(r) = i:F
L
Ii:,45]x(r) + [:]f(r)u(r)

Again, (64) and (65) are in a similar form to the error + 0.01 sin (23.6r)
equation between the reference model and the adjustable I -0.02 I ’
part of the system found in model reference adaptive control
(Narendra and Annaswamy, 1989). As a result, the following y(f) = g(r)[-3.5 5.51x(f) + g(r)f(r)u(r). (76)
theorem can be obtained, which produces a convergent
adaptive diagnostic algorithm for estimating the fault vector It can be seen that in this case S,,= 0.0224 and the system is
8. observable. It is assumed that the healthy actuator and
222 Brief Papers

sensor have the gains fH = 1 and g, = 1. For this system, the

observer gains are selected as
1
L,=L= -
[I
2 ’

which makes (A + LC) stable, with two real eigenvalues,

-1.35 and -6.39. As a result, the following transfer functions
can be obtained:
s= + 5.75s + 5.15
W,(s) =
s= + 7.75s + 8.65 ’
s* + 0.25s + 0.15 0.2 ’ I
W(s) = 0 5 10 15 20 25 30
s2 + 7.75s + 8.65 ’
3.5s + 1.5 5.5s + 5 f (1).
sz + 7.75s + 8.65 sz + 7.75s + 8.65 1 (80) Fig. 4. Fault diagnosis for

Using these transfer functions, it can be shown that the As for nonlinear actuators, we reconsider the system with
threshold and the size of dead zones are
A = A, = 0.0194. (81) i(f) = it4’ Ii:y45]*(1) + [ y] tan-r [e(r)u(t)]
[
To illustrate the use of the algorithm, faults are created as
+ 0.01 sin (23.6r)
follows:

(82)
[ -0.02 1 ’
(86)

y(t) = [-3.5 5.5jX(t). (87)

(83)
1.1,
With r = diag (30.2.5) and the sampling period being 0.01, I
the simulation results are shown in Figs 3-6, where the
system is subjected to a step input at t = 0. It can be seen that
both f(r) and g(r) can track the behaviour of the actuator 1.05 I I(f) -
and sensor faults in a desired manner.
Although many actuator and sensor faults lead to an
abrupt change in gain, in practice the faults can also cause
the gain to drift. Such faults are extremely difficult to detect 1 L
immediately from a simple visual inspection of the output
signals. To simulate this phenomena, it is assumed that the
faulty actuator and sensor for the system (75). (76) are given
0.95
by 0 5 10 15 20 25 30
fH = 1 (I 5 7 s),
Fig. 5. Fault diagnosis for g(r).
f(I) = 1 - O.O5(t - 7) (t>7s). (84)
1 0.375 + O.O5(t ~ 20) (t > 20 s).
I (2 5 7 s).
g(r) = 1 ~ O.Ol(r - 7) (t > 7 s), (85)
0.5 -
1 0.87 + O.Ol(r - 20) (I 2 20 s).

Because the adaptive diagnostic algorithm (35) can only 0 I Li

follow constant or slowly time-varying faults, the driftings
(84) and (85) arc made very slow in order to maintain our
-0.5
assumptions (13) and (14). However, to increase the tracking
ability of the adaptive diagnostic algorithm (35). the learning
rate for the tuning of g(t) is modified to y2 = 10. Figures -1 i
7-10 show the responses for both detection and diagnosis for
the system (75). (76) subjected to (84) and (85). The drifting
gain faults in both the actuator and sensors can again be 0 5 10 15 20 25 30
diagnosed, and @(t),g(r)) follows (f(t). g(r)) with the
Fig. 6. The augmented error e,(t).
desired accuracy.

0.51
O.*~ 0.4 -
I

0.3 - Y eo(Ol

0.2 -

o’:bi.l0 5 10 15 20 25 30
O-l/
01
0 5
J
/
,
10
4

Fig. 7. Detection
15 20

signal.
25 30

Fig. 3. Detection signal.

 Brief Papers 223

0.15 -

II e&l

0.05
t

g(r)
*‘I- i
0.2 ’ I
0:
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Fig. 8. Fault diagnosis for f(t). Fig. 11. Detection signal.

1 n

where
f(6)(t),
o’95y
v0.9

0.85’

the
0

nonlinear
5
Fig. 9. Fault

u(f)) = tan-’ [e(f)u(t)].

actuator
10

is
15
diagnosis

characterized
20
for g(t).

by
It can be seen that the gain
25
I
30
0.8 -
0.6 -

0.4
0

In the simulation,
5

the healthy
fault is created as follows:
10

Fig. 12. Fault

15
It

20

diagnosis.

actuator
25 30

has OH = 1. and the

matrices in (77) makes the triple (A + LC, B, C) strictly
positive real, with 1 (I >7s).

P=
3
-3.5
-3.5
1 [
5.5 ’ Q =
26.2
~43.425
-43.4250
75.25
1 (88) e(f)= 1.4
1 0.7
(r~7~),
(t > 20 s).
(91)

Moreover, with the choice of the nonlinear actuator in (86), Based on the principles outlined in Section 6. Kz = 2 and
it is assumed that 0.5 5 110115 2. Using the well-known v = 1. The simulation results are shown in Figs 11 and 12,
Cauchy theorem. it can be shown that where the threshold is again equal to 0.0194. It can be seen
that e(t) tracks the faulty actuator behaviour in a desired
manner.

where 0, and @*-are two arbitrary numbers in the closed 8. Discussion and conclusions
interval [0.5.2], 6 is related to 0, and Q2 and is also in In this paper we have developed an adaptive algorithm for
10.5.21. As a result, the following inequality can be fault detection and diagnosis of actuators and sensors in
formulated: deterministic dynamic systems. Two observers-a fixed
detection observer and an adaptive diagnostic observer-
/tan-’ [e,u(t)] - tan’ [e2u(t)] - a tanmijeu(r)’ ~BZc (e, - e2)1 were constructed separately to detect and diagnose the fault.
/ A new form of observation error equation for the adaptive
11(r) U(l) diagnostic observer was formulated and used to construct an
= 1 + [&(I)] - 1 + [eilc(r)]* I81 - eZ’ adaptive diagnostic algorithm, which can be obtained directly
via the use of the augmented error signal technique from
2 Iww + iw 4 lU(f)l
-=* = K,(u(t)). model reference adaptive control. Single-input and single-
51 + min (ie,i. iezl)%(t)2Cl +0.25u(r)2-lu(r)l output systems were first considered, and it was shown that a
(90) compact form of adaptive diagnostic algorithm with dead
zones can be obtained. Extension to arbitrary multi-input
and multi-output systems was then considered, and a
0.06 ( gradient descent rule in the continuous-time domain was
1
used to construct an adaptive diagnostic algorithm. A further
0.04 extension to nonlinear actuators was performed via the use
of the same fault detection and diagnosis for linear actuators
and sensors. Two simulated examples have been included to
0.02
demonstrate the applicability of the proposed methods and
encouraging results have been obtained.
0 A bounded single term v(t) has been used to represent
l-+7 ( model uncertainties and input and output noises through the
-0.02 dvnamic equivalent principle. However, the mode1 uncer-
t u tainties in matrices (A, B. C. D) will lead to the following
-0.04 1 equations for the system:
0 5 10 I5 20 25 30 f(f) = (A + bl))x(t) + (B + M)f(r)u(r) + v(t).
Fig. 10. The augmented error c,(t). y(t) = g(r)[(C + M’)x(r) + (D + ~ID)f(t)u(t)].
224 Brief Papers
In this case, similar error equations to (19) and (20) can still Patton, R. J., J. Chen and T. M. Siew (1994). Fault diaenosis
be obtained, but the unmeasurable term will include (AAx, in nonlinear systems via neural networks. In Proc.
ABfu, gACx, gADfu) as part of its component. As a result, International Conf. on Control ‘94, Vol. 2, pp, _-
1346-l 351.
.- _--_.
the assumption on the boundness of the unmeasurable term Phatak, M. S. and-N. Viswanadham (1988): Actuator fault
no longer holds. Two alternative approaches may therefore detection and isolation in linear systems. Inc. J. Svst. , Sci..
be used to treat this case. The first is to use normalized 19,2593-2603.
signals in combination with dead zones (Narendra and Sidar, M. (1983). Implementation of failure detection system
Annaswamy, 1989) in the design of an adaptive diagnostic with adaptive observers. In Proc. American Control Conf.
algorithm. The other approach is to choose the observer gain Boston, MA, pp. 1205-1211.
matrix .L.such that error equation is made insensitive to these Sun, J. and P. Ioannou (1992). Robust adaptive LQ control
model uncertainties. This can be achieved via the use of scheme. IEEE Trans. Autom. Control, AC-37,100-106.
parametric design in robust fault diagnosis observers (Wang Viswanadham, N. and R. Srichander (1987). Fault detection
et al., 1993). using unknown-input observers. Control Theory Adu.
References Technol., 3.91-101.
Astrom, K., J. and B. Wittenmark (1989). Adaptive Control. Wang, H. (1987). A new method for improving the dynamic
Addison-Wesley, Reading, MA. properties of MRAS. In Proc. 10th IFAC World Congress
Basseville, M. (1988). Detecting changes in signals and on Automatic Control, Munich, Vol. 10, pp. 2128-2132.
systems: a survey. Automatica, 24,309-326. Wang, H., M. Brown and C. J. Harris (1994). Fault detection
Chen, J. and H. Y. Zhang (1991). Ronust detection of faulty for a class of non-linear systems via associative memory
actuators via unknown input observers. Inc. J. Syst. networks. Proc. JMechE, 208,101-108.
Science, Z&1829-1839. Wang, H., H. Kropholler and S. Daley (1993). Robust
Chow, E. Y., X. C. Lou, G. C. Verghese and A. S. Willsky observed based FDI and its application to the monitoring
(1986). Redundancy relations and robust failure detection. of a distillation column. Trans. Inst. Meas. Control, 15,
In Basseville, M. and A. Benveniste (Eds). Detection of 221-227.
Abrupt Changes in Signals and Dynamic Systems pp. White, J. E. and J. L. Speyer (1987). Detection filter design:
275-294. Lecture Notes in Control and Information spectral theory and algorithm. IEEE Trans. Autom.
Science, Vol. 77. Springer-Verlag, Berlin. Control, AC-32,593~603.
Daley, S. and H. Wang (1993). Fault detection using reduced Wuennenberg, J. and P. M. Frank (1988). Model based
order models. Proc. JMechE, 207, 153-163. residual generation for dynamic systems with unknown
Ding, X. and P. M. Frank (1993). An adaptive observer- inputs. In Proc. 12th IMACS World Congress on Scientific
based fault detection scheme for nonlinear dynamic Computation, Paris, Vol. 2, pp. 435-437.
systems. In Proc. 12th IFAC World Congress, Sydney, Vol. Zhang, J. and P. D. Roberts (1992). On-line process fault
8, pp. 307-311. diagnosis using neural network techniques. Trans. Inst.
Frank, P. M. (1994). Enhancement of robustness in Meas. Control, 14, 179-188.
observer-based fault detection. Inc. J. Control, 59,
955-981. Appendix-Proof of Theorem 2
Frank, P. M., X. Ding and J. Wochnik (1991). Model-based Since in the nonlinear adaptive observer (61). e(t) = OH
fault detection in diesel-hydraulic driven industrial trucks. until a fault is detected (see (62)) the adaptive diagnostic
In Proc. American Control Conf, Boston. MA, pp. algorithm (71) is only utilised after t > tf, where, based upon
1528-1533. the assumption (62) 8(c) = 0. Therefore, when t > tr, the
Freyermuth, B. and R. Isermann (1991). Model incipient following Lyapunov function can be chosen for the system
fault diagnosis of industrial robots via parameter (64), (65) and (71):
estimation and feature classification. Proc. 1991 European V(e,(t), e(c)) = e’,Pe, + Ilh(t) - 611’. 64.1)
Control Conf., Grenoble, pp. 115-121.
Ge, W. and C. Z. Fang (1988). Detection of faulty It can be shown that
components via robust observation. Inc. J. Control, 47,
581-599. ti(e,(t).f(t)) = -e,(t)TQe,(t) + k,(t)TPWf(@t)v u(t))
Hou. M. and P. C. Muller (1994). Fault detection and ^
isolation observers. Inc. J. Control, C&,827-846. -f(e, U(C))] + 2[&c) - elTd$+ 2e,(t)rPv(c). (A.2)
Isermann, R. (1984). Process fault diagnosis with parameter
estimation methods. In Proc. IFAC Digital Computer Using assumptions (ii) and (iii) of the theorem, it can be seen
Application to Process Control, Vienna, pp. 51-60. that
Landau, Y. D. (1979). Adaptive Control-The Model
Reference Approach. Marcel Dekker, New York.
Li, Y. and H. Zhang (1993). Robust detection of faulty f(&c). U(C))- f(e, U(C)) = af(@$c)) 1 [B(t) - e] + A,
e=&,
actuators via adaptive unknown input observer. In Proc.
International Conf on Fault Diagnosis, Toulouse, France, (A.3)
pp. 296-300. where the term A satisfies 1IA][5 K,(u(t)). Using (71). it can
Massoumnia, M. A., G. V. Verghese and A. S. Willsky be obtained that
(1989). Failure detection and identification. IEEE Trans.
Autom. Control AC-34,316-321. ti(e m(I) Be(c)) = -e m(t)‘.Qe m(t)
Narendra. K. S. and A. M. Annaswamy (1989). Stable + 2e (?)TPB
Adaociue Svstems. Prentice-Hall. Encrlewood Cliffs. NJ. m
Ono, ‘T., T: Kimamaru, A. Maed; S. Sagara and K.
Kumamaru (1987). Influence matrix approach to fault + 2[e(t) - elT ‘T + 2e,(~)~Pv(c) = -e,(r)TQe,(t)
diagnosis of parameters in dynamic systems. IEEE Trans.
Indust. Electron, E-34,285-291.
Park, J. and G. Rizzoni (1994). An eigenstructure assignment + ~~,,,(c)~P[BA + v(t)] - 2a[&(t) ~ .9]“&t). (A.4)
algorithm for the design of fault detection filters: IEEE Using assumption (iii) of the theorem, it can be shown that
Trans. Autom. Control. AC-39. 1521-1524.
Patton, R. J. and J. Chen (1991). A robust parity space v(e,(t), 8(t)) = -[e,,,(c) + Q-‘P(BA + v)lTQ[e,(t)
approach to fault diagnosis based on optimal eigenstruc-
ture assignment. In Proc. International Conf on Control + Q-‘P(BA + v)] + (BA + v)~PQ-‘P(BA + v)
‘91, Vol. 2. pp. 1056-1061. - 2a]B(c) - e]‘r[B(t) - e] - 2u[e(t) - elTe
Patton, R. J., P. Frank and R. Clark (1989). Fault Diagnoses
5 -6V(e,(t), B(r))
in Dymamic Systems: Theory and Application. Prentice-
Hall. Englewood Cliffs, NJ. + &n,,(Q) IIQ-‘PlIZ[ll~l12 K:(4)) + S:,l + 4 (A.5)
 Brief Papers 225
On the other hand, it can be seen that the following equality of the time period when the pair (E(I), 8(t)) are inside DR.
holds when (I, e(t)) ED,: Then it can be shown that

* > Li”(P) IImll* + f Il&~)l12- G

-II@11 - ,,q,*

+.tQ) IIQ-‘f’II* [IIW K:@(r)) + S;] + UK:} (A4

Using the inequality* and the definition of DR, it can be
I
’ IdVk&), &))I
5 f 8 l~(~,WW)) - Gl
^
W(e,(r), e(r)) - G? < +m
seen that V(e,(r), O(r)) <O (e(r), e(r)) E DR). As a result, = +og (A.8)
~V(4tf), Wf)) - EZ
the dynamic system represented by (64), (65) and 171) is
stable. Since the inequality (AS) holds for all (E(r), f?(r)) E when (E(r), d(r)) E DR, where
DR. the pair (e,(r), e(r)) are therefore uniformly bounded
and converge to DR exponentially at a rate greater than e-“. + Gl + OK: (A.9)
G, = LdQ) IlQ-L~l12[ll~llz~~~~(~))
Denote by m,{r I (&(I), O(r)) E DR} the Lebesgue measure 0