Robust Fault Detection in Open Loop vs. Closed Loop

Henrik Niemanny Jakob Stoustrupz

Version: Robust_FDI4.tex { Printed 15h 47m, February 9, 1998

Abstract
The robustness aspects of fault detection and isolation (FDI) for uncertain systems will be
considered in this paper. The standard setup for FDI will be considered here. The FDI design
problem will be analyzed both in the case where the control input signal is considered as a known
external input signal (open loop) and when the input signal is generated by a feedback controller.

Keywords: Robust FDI, Feedback control, uncertain systems.

 This work is supported by the Danish Technical Research Council under grant no. 95-00765.
y
Department of Automation, Technical University of Denmark, Building 326, DK-2800 Lyngby, Denmark. E-
mail: hhn@iau.dtu.dk.
z
Department of Control Engineering, Aalborg University, Fredrik Bajersvej 7C, DK-9220 Aalborg East, Denmark.
E-mail: jakob@control.auc.dk. W3 :
http://www.control.auc.dk/  jakob/.

1
1 Introduction
In the three papers [NS97], [SG96], [SGN97], the combined setup for both feedback and fault detection
lter design problem has been considered. The design setup is shown in Figure 1, where a standard
formulation is applied [ZDG95].

zc   wd
G(s)

yc

fs - +?
k y - K (s)
uc -
+k fa

Figure 1: Control system with actuator fault, fa , and sensor fault, fs.

A complete analysis of the combined feedback controller/fault detection lter has been given in
[SGN97] for both nominal systems as well as for uncertain systems. The results of this analysis
is that there is a separation between the design of the feedback controller and the fault detection
lter in the nominal case which does not exist in the uncertain case. The reason for this missing
separation in the uncertain case is that there is a trade-o between performance in the feedback loop
and performance for the fault detection lter.
By using the setup shown in Figure 1, we are looking at both the feedback controller and the fault
detection lter at the same time. The other approach in fault detection is to consider only the system
without taken care of how the control signal is calulated. This is shown in Figure 2.

 v
G(s)  f


y - u

- 
F (s)
- f^

Figure 2: General structure for fault detection in open loop. v is disturbance input and f is fault
input signal. f^ is the estimate of f .

This setup has been considered in several papers, see e.g. [Fra90], [PFC89], [Fra96] and the references

2
therein.
The main issue in this paper is to give an analysis of the FDI design problem both in the case
when the relation between u and y is known and when it is not known. Nominal systems as well as
uncertain systems will be considered.
The rest of this paper is organized as follows. The system setup is given in Section 2. The analysis
results are given in section 3 followed by an analysis of a special case in Section 4. A conclusion is
given in Section 5.

2 Problem Formulation
Consider the setup given in Figure 3, which is an extension of the setup shown in Figure 2.

-
(s)


 v
G(s)  f


y - u

- 
F (s)
- f^

Figure 3: General setup for robust fault detection in open loop.

The system G in Figure 3 has the following form:

0 1
! !B w CC
z
= Gzw Gzv Gzf Gzu BB v CA (1)
y Gyw Gyv Gyf Gyu @f
u
where z and y are the external output signal and the measurement output signal respectively. The
inputs are external input w from the uncertain block
, disturbance input v, fault input signal f
and the control input signal u, respectively. It will further be assumed that the perturbation block

is scaled such that k
k  1; 8! and the scaling function is included in G. There is no assumption
about the structure of
. Further, it is also assumed that all other relevant weight matrices are
included in G.

3
The design problem is to design a FDI lter F (s) such that the estimation error e dened by
e = f ? f^ (2)
= f ? (F (s)y ? F (s)Gyu u)
is minimized in some sense.
In the next section, the estimation error will be analyzed in the general case given by (1). A special
case of (1) will be analyzed in Section 4. case.

3 General Case
The main analysis results will be given here in four dierent cases.

3.1 Open Loop FDI Analysis for "Nominal" Systems

In this case, the system in (1) takes the following form:
0 1
 B v C
y= Gyv Gyf Gyu @fA (3)
u

The estimation error is then given by:

eopen = (I ? F Gyf )f ? F Gyv v (4)

It turns directly out from (4) that it is not possible to make estimation of f if the disturbance v is
in the same frequency range as the fault signal f is and has the same direction at the system. There
is a trade-o between fault detection and disturbance attenuation.

3.2 Closed Loop FDI Analysis for "Nominal" Systems

The control input signal u is given by
u = K (s)y (5)
where K (s) is a stabilizing feedback controller. The closed-loop system from (3) with (5) is then
given by:
 ?1 ?1
 v!
y = (I ? Gyu K ) Gyv (I ? Gyu K ) Gyf f
  v! (6)
= SGyv SGyf f
where S is an output sensitivity function. The fault estimation error e for the nominal closed-loop
system is given by:
eclosed = f ? f^ (7)
= (I ? F SGyf )f ? F SGyv v
The only dierence between the estimation error for the open loop in (4) and for the closed loop
in (7), is that the sensitivity function S is included in (7). So, if the lter in (7) is selected as
Fclosed = Fopen S ?1 , we get exactly the same equation for the estimation error as in the open loop
case. The open loop and the closed loop cases are equivalent in the nominal case.
4
3.3 Open Loop FDI Analysis for Uncertain Systems
When the loop from z to w is closed by the uncertain block
, w =
z , the system in (1) takes the
following form:
0 1
 B v C
y = Gyw
S
Gyv + Gyv Gyw
S
Gyf + Gyf Gyw
S
Gyu + Gyu @fA
0 1 u
(8)
  v
= G yv G yf G yu B@ f CA
u
where S
= (I ? Gzw
)?1 .
The estimation error is then given by:
eopen = (I ? F Gyw
S
Gyf ? F Gyf )f
?(F Gyw
S
Gyv + F Gyv )v (9)
?F Gyw
S
Gyuu
With the perturbation block present in the system, three additional terms appear in the equation
for the estimation error, (9), compared to the nominal case.

3.4 Closed Loop FDI Analysis for Uncertain Systems

As in Section 3.2, the control input signal u is based on a feedback controller K (s). Closing the
loop in (8) with a feedback controller K (s), we get the following system (with G yv , G yf , and G yu as
dened in (8)): !
 1 1
 v
y = (I ? G yu K ) G yv (I ? G yuK ) G yf
? ?
f (10)
The fault estimation error in the closed loop for uncertain systems is then given by:
eclosed = (I ? F (I ? G yu K )?1 G yf )f ? F (I ? G yuK )?1 G yv v (11)
The above equation for the fault detection error signal is quite complicated in the general case.
Therefore, a simplied case will be considered in next section, which make it possible to give a clear
analysis.

4 A Special Case
A special case of the more general case considered in Section 3 will be considered here.
Let us consider a system given by:
y = G(s)(I +
)(f + v + u) (12)
where
is a multiplicative perturbation at the plant input. Using the same formulation as in Section
3, the system in (12) is given by:
0 1
! !B w CC
z 0 I I I B v
y = G G G G B
@f CA (13)
u

5
Note that both the fault signal f , the disturbance signal v and the control input signal u enter the
system at the same place. It there is not a separation in frequency between the fault signal and
the disturbance input signal, it will not be possible to separate the fault signal from the disturbance
signal.
In the nominal case, the two estimation error signals for open loop and for closed loop are by:
eopen = (I ? F G)f ? F Gv (14)
and
eclosed = (I ? F SG)f ? F SGv (15)
respectively, where S = (I ? GK )?1 is the sensitivity function.
As in the general case, there is no principal dierence between fault detection in open loop and closed
loop when we only consider nominal systems. To get exactly the same estimation error in the closed
loop as in the open loop, we just need to use the lter given by Fclosed = Fopen S ?1 . Note that when
S is small at low frequencies (as it is in general), the gain of Fclosed will increase equivalently at low
frequencies. This is the drawback of fault detection in closed loop compared with open loop.
Now, let us consider the uncertain case. The two estimations errors are given by:
eopen = (I ? F G(I +
))f ? F
u ? F G(I ?
)v (16)
for the open loop and
eclosed = (I ? F (I ? G(I +
)K )?1 G(I +
))f ? F G(I ? (I +
)KG)?1 (I +
)v
= (I ? F G(I ? KG ?
KG)?1(I +
))f ? F G(I ? KG ?
KG)?1 (I +
)v
= (I ? F G(SI?1 ?
KG)?1(I +
))f ? F G(SI?1 ?
KG)?1 (I +
)v (17)
= (I ? F GSI (I ?
KGSI )?1 (I +
))f ? F GSI (I ?
KGSI )?1 (I +
)v
= (I ? F SG(I ?
TI )?1 (I +
))f ? F SG(I ?
TI )?1 (I +
)v
for the closed loop, respectively, where TI = GK (I ?GK )?1 is the complementary sensitivity function.
The picture gets much more complex when we are looking at uncertain systems. In the open loop,
an additional term from the control input signal appears in the estimation error due to the model
uncertainty
. This term has been removed in (17) when a feedback controller has been applied.
Further, note that the robustness of the feedback loop has a major in uence on the estimation error
in (17). This is in accordance with the results in [NS97], [SGN97]. Here we can see the connection
between the performance in the feedback loop and the performance in the fault detection lter. To
see this connection, assume that the performance in the feedback loop is maximized be the design
of K (s). This mean that I ?
TI is close to the robust stability bound, i.e. j
TI j ' 1. When the
inverse of I ?
TI appear in (17), we will get a very poor estimation of the fault signal f .
To give a more complete analysis of the FDI problem for uncertain systems, let us separate the
estimation error into:
e = enom + eunc (18)
where enom is related to the nominal case and eunc is related to the model uncertainty. Further, as
mentioned before, let us assume that the fault signal f and the disturbance signal v is separated in
frequencies.
The estimation error for the open loop is then given by:
eopen = enom + eunc (19)
= (I ? F G)f ? F G
f ? F G(I +
)v ? F G
u

6
and for the closed loop
eclosed = enom + eunc
= (I ? F SG)f ? F SG
f ? F SG
TI (I ?
TI )?1 (I +
)f (20)
?F SG(I +
)v ? F SG
TI (I ?
TI )?1 (I +
)v
Without loss of generality, we will only consider the fault input signal in the frequency range where
f is separated from v. Only the fault signal will be considered in the following.
The estimation error for the open loop is then given by:
eopen = enom + eunc (21)
= (I ? F G)f ? F G
f ? F G
u
where the uncertain part is given by:
eunc;open = ?F G
f ? F G
u (22)
and for the closed loop
eclosed = enom + eunc (23)
= (I ? F SG)f ? F SG
f ? F SG
TI (I ?
TI )?1 (I +
)f
where the uncertain part is given by:
eunc;closed = ?F SG
f ? F SG
TI (I ?
TI )?1 (I +
)f (24)

From the uncertain part in (22) for the open loop, it can be seen that the estimation error, eopen , will
increase if the model uncertainty
is too large in the frequency range where we want to estimate the
fault signal f . Further, the control input signal u will also reduce the estimation precision in the open
loop uncertain case due to the term F G
v. The control input signal's in uence on the estimation
error can be removed by using the closed-loop formulation. As a consequence of closing the loop
by a feedback controller K (s), the eect from the control input signal has been removed from (24),
but another term appear in the equation for the estimation error. The new term in (24) includes
(I ?
TI )?1 which is related to the robust stability of the closed-loop system. If the performance of
the closed loop system is optimized, j
TI j will be close to 1, due to the robust stability condition
for the system in (12), see [SP96]. The consequence is that (I ?
TI )?1 can be very large and the
estimation error will increase. Consider e.g. the case where the feedback controller is designed with
50% robustness margin, i.e. j
TI j' 0:5. Then j
TI (I ?
TI )?1 j' 1. The estimation error from
(23) is then given by:
eclosed ' (I ? 2F SG)f ? 2F SG
f (25)
or
eclosed ' enom ? F SG(I + 2
)f (26)
As a consequence of (26), we can see that even if the uncertainty is small, a quite large estimation error
is obtained due to the term F SGf . If the nominal estimation error enom is small, i.e. j enom j 1,
we will have that j F SG j' 1 in the frequency range where we want to make fault detection. Using
this approximation, (26) is then given by:
eclosed ' ?(I + 2
)f (27)
in a certain frequency range.
(27) shows that we get more than 100% estimation error in the case where the feedback controller is
designed with 50% robustness margin.
7
These results are in accordance with the results from [NS97]. In [NS97] it is shown that it is not
possible to separate control and fault detection in the uncertain case when a compact setup is applied.
The results here give an indication of how the estimation error will increase when a separated design
of the controller and the fault detection lter is applied.
At rst glance it seems a little strange that we get a worse estimation error when we applied the
information of how the control input signal is derived. The reason is that we in practice decrease the
information available for the fault detection. In the open loop case we use the control input signal
directly, where as the control input signal is only used indirectly in the closed loop case. As a matter
of fact, when we use the open loop formulation, the uncertainty will not be fed back, because we use
the real signal. Hence, we have more information available in the open loop case which makes the
dierence. The analysis of the closed loop case above shows what can happen when this information
can not be used.

5 Conclusion
Fault detection in open loop vs. closed loop has been considered in this paper for both nominal
systems as well as for unceertain systems. In the nominal case, there is in principle no dierence
between open loop and closed loop fault detection. This is not the case for uncertain systems. In
this case, there is a trade o between good fault detection and good performance of the closed loop
system.

References
[Fra90] P.M. Frank. Fault diagnosis in dynamic systems using analytic and knowledge-based re-
dundancy - A survey and some new results. Automatica, 26:459{474, 1990.
[Fra96] P.M. Frank. Analytical and qualitative model-based fault diagnosis - A survey and some
new results. European Journal of Control, 2:6{28, 1996.
[NS97] H.H. Niemann and J. Stoustrup. Integration of control and fault detection: Nominal and
robust design. In SAFEPROCESS'97, England, 1997. Submitted for publication.
[PFC89] R. Patton, P. Frank, and R. Clark. Fault diagnosis in dynamic systems - Theory and
application. Prentice Hall, 1989.
[SG96] J. Stoustrup and M.J. Grimble. Integrated control and fault diagnosis design: A polyno-
mial approach. In Modelling and Signal Processing for Fault Diagnosis, Leicester, U.K.,
September 1996. IEE.
[SGN97] J. Stoustrup, M.J. Grimble, and H.H. Niemann. Design of integrated systems for the control
and detection of actuator/sensor faults, 1997. Submitted for publication.
[SP96] S. Skogestad and I. Postlethwaite. Multivariable feedback control - Analysis and design.
John Wiley & Sons, 1996.
[ZDG95] K. Zhou, J.C. Doyle, and K. Glover. Robust and optimal control. Prentice Hall, 1995.