Вы находитесь на странице: 1из 6

On the Integration of Fault Detection and

Isolation in Model Based Fault Diagnosis




V. Puig
1
, 1. Quevedo
1
, T. Escobet
1
, B. Pulido
2




1
Automatic Control Department (ESAII) - Campus de Terrassa
Universidad Politecnica de Catalunya (UPC) - Spain
e-mail: vicenc.puig upc.es
2
Departamento de InIormatica - Universidad de Valladolid - Spain
Abstract. Model based diagnosis can be divided in two subtasks:
Iault detection and Iault isolation. Fault detection tests try to check
the consistency between the modelled and observed behaviours
while Iault isolation tries to isolate the component that is in Iault.
Since these two tasks can be executed in parallel there should be
some interIace between them. Typically a binary vector containing
the result oI the Iault detection tests is used as the input to be
analysed by the Iault isolation module. However, in the process oI
binary codiIying these tests some inIormation is lost. In this paper,
we try to pinpoint the drawbacks oI this classical approach and to
suggest some possible improvements.


1 INTRODUCTION

Model-based diagnosis has been approached over the last two
decades Irom two diIIerent scientiIic communities: ArtiIicial
Intelligence, also known as the DX approach |11||22|, and
Automatic Control, also known as FDI approach |2||10| |17|.

The DX approach relies upon a well-Iounded and logically based
theory Ior diagnosis oI static systems. From a logical point oI view,
Iault detection is perIormed through a consistency-check,
organized around the conIlict concept. In this approach, Iault
localization or isolation is automatically derived Irom the conIlict
detection stage, which usually relies upon some kind oI
dependency-recording.

On the other hand, the FDI approach consider Iault diagnosis as
two separate tasks: Iault detection and Iault isolation based on
generating and evaluating a set oI analytical redundancy relations
obtained oII-line Irom elementary component models oI the
physical systems.

Elements oI comparison between both approaches have been
included in |7|, which provides a common Iramework Ior research
Irom both communities. In such trend, some recent works have
tried to merge the best oI each approach |18| (BRIDGE project). In
this merging process the Iault diagnosis task has been again
separated in the task oI Iault detection and isolation. Fault detection
traditionally has been more deeply investigated in the FDI
community using a broad set oI techniques: parity equations,
observers, parameter estimation,... and looking at the perturbing
eIIects oI noise, perturbations and model uncertainty (robust Iault
detection) |4|. On the other hand, Iault isolation has been more
deeply investigated in the DX community thanks to the logical
diagnosis theory developed by Reiter |22|. However, when this
theory has been applied to dynamical systems some problems has
appeared that prevent to use it directly. In Iact, the theoretical
Iormalization Ior diagnosis oI dynamic systems it is nowadays an
open issue |8|. These problems have motivated Iurther research on
analysing oII-line the set oI dependencies which could become
conIlicts as in the FDI approach |20|, based on the common
Iramework provided by |7|.

However, most oI the research in both communities has been
Iocused either in the detection or in the isolation tasks but very Iew
papers deal with the interIace between these two modules. Only
recently some researchers has signalled the importance oI this
interIace but only Irom one oI the two approaches |5|.

The aim oI this paper is to emphasize that iI Iinally a two step Iault
(detection and isolation) diagnosis approach is adopted, the
interIace between the modules should be careIully designed in
order to not losing inIormation that derive in decreasing the quality
oI Iinal Iault diagnosis result, specially when we are diagnosing
dynamic systems.

This paper is organized as Iollows: Section 2 analyses the
limitations oI the basic existent interIace between detection and
diagnosis decision based on binary codiIying detection tests and
the Iault signature matrix is outlined; Section 3 proposes a
motivational example that illustrates isolation problems when
dealing with the case oI diIIerent time oI appearance oI symptoms;
Section 4 proposes some improvements oI existent FDI and DX
Iault isolation schemes to deal with problems presented in Section
3. Finally, Section 5 presents main conclusions obtained.


2 LIMITATIONS OF FDI AND DX FAULT
ISOLATION

Model based Iault detection tests are based on the evaluation oI a
set oI relations r
i
(k) (also known as analytical redundancy
relations (ARRs) in the FDI community or potential conflicts in
the DX community) derived Irom the elementary models oI system
components and the available measurements coming Irom sensors
available:


)) k ( ), k ( ( f ) k (
el mod measured
y y r = (1)

according to Ploix |18|.

Using either an approach coming Irom FDI (as Ior example,
Staroswiecki`s structural analysis |23|) or an approach coming
Irom DX (as Ior example, Pulido`s GDE based algorithm |19|) the
whole set oI ARRs Ior a given system and a given set oI sensors
can be generated.

Each detection test (r
i
) should be evaluated on-line in order to
decide iI it is or not violated at a given time instant (typically a
binary codiIication is used: 0 indicated not violation and 1
violation):

<
=
) fault ( ) k ( r if 1
) fault no ( ) k ( r if 0
) k ( s
i i
i i
i

(2)

where
i
is the threshold associated to the detection test r
i
(k). This
test is constitutes the detection phase oI the diagnosis process.

Finally, the evaluation oI each detection test will be the actual (or
observed) fault signature oI the system:
[ ] ) k ( s , ), k ( s ) k (
n 1
= s that it is set oI symptoms that will be
supplied to the Iault isolation module in order to isolate the Iault.

2.1 FDI fault isolation

Given a set oI symptoms [ ] ) k ( s , ), k ( s ) k (
n 1
= s , and a set oI
considered Iaults: f
1
,f
2
,...f
m
, the theoretical fault signature matrix,
, can be deIined binary codiIying the eIIect or not oI a Iault in
every symptom. This matrix has as many rows as symptoms and as
many columns as Iaults are considered. The element
if
oI this
matrix is equal to 1 meaning that the f
th
Iault appears in the
expression oI the i
th
symptom generator, otherwise is equal to 0.
Then, Iault isolation will consist in looking Ior the theoretical fault
signature in the fault signature matrix that matches with the
observed signature. For example, in case oI having the Iollowing
Iault signature matrix

f
1
f
2
f
3

s
1
1 1 1
s
2
1 0 1
s
3
1 1 0

Table 1. Fault signature matrix


then, the Iollowing logical tests will allow to isolate the diIIerent
Iaults without considering that noise or perturbations may cause
detection errors:

3 2 1 3
3 2 1 2
3 2 1 1
s s s f
s s s f
s s s f
=
=
=
(3)

ThereIore, FDI approach to Iault isolation Iollow a column view
approach.

2.2 DX fault isolation

Fault detection can be considered as equivalent using ARRs in FDI
and potential conIlicts in DX |7|, however Iault isolation can be
tackled in two diIIerent ways within the DX community:
consistency-based, CBD, or abduction-based |11|. While
consistency-based diagnosis tries to reject those behavioural modes
which are not consistent with current observations, abduction-
based diagnosis tries to explain current observation with a
consistent behavioural mode assignment. In such sense, abduction-
based diagnosis is closer to the FDI approach than consistency-
based. Moreover, Ior real complex dynamic systems there is no
direct translation Irom the static consistency test to the dynamic
one |3|.

Based on the proposed Iramework by Cordier |7|, Iault signature
matrix presented in Table 1 is interpreted in DX CBD approach to
Iault isolation considering separately each line corresponding to a
violated ARR, isolating R-conIlicts (i.e., a set oI components that
are to be considered abnormal in order to be consistent with the
observed Iault signature) beIore searching Ior a common
explanation, i.e, Iollows a row (or column) view oI the table.

2.3 Common limitations

When applying DX and FDI approaches to dynamic systems, since
they may exhibit symptoms with diIIerent dynamics (Iast, slow,
...), the use oI binary codiIication oI the result oI the Iault detection
test associated to a given ARR (or potential conIlict), some
inIormation has been lost. In particular, among others:

a) the sign oI the symptom is lost
b) the sensitivity oI the symptom respect to each Iault is lost
c) the order oI symptom apparition oI a given ARR with respect
to the others is not recorded, since it is a static test,
d) the persistency oI the symptom indicator oI each ARR aIter a
Iault is independent oI the others,
e) the time to wait the appearance oI a set oI symptoms times to
isolate the Iaults when it exists diIIerent appearance times is
not considered,
I) and instability oI the Iault detection test indicator (chattering)
since the presence oI noise and the binary test used,

In authors opinion, the diagnosis provided by each approach could
be improved iI that inIormation was considered. However, in the
remainder oI this work we will Iocus just on the temporal issues
described c) and e), using the approaches coming Irom both
communities, agreeing with one oI the tasks proposed in the
BRIDGE Iramework.


3 MOTIVATIONAL EXAMPLE

3.1 System description

In order to motivate the temporal limitations presented in the
previous section, a part oI a water transport network is presented
here (Figure 1). This network corresponds to a common strategy
used in real water networks. It consists in continuously suppling
water to consume areas (here, c
1
and c
2
) Irom reservoirs (v
1
and v
2
,
in this case) and simultaneously, during nightly hours (lower
electrical Iares), it raises the water to the maximum level oI the
head reservoirs to stock the water Ior the daily hours (bigger
electrical Iares). The water transport between reservoirs is
modelled as an open Ilow channel and the transport time delays (
1
and
2
)

may be identiIied Irom the real data recorded Irom the
network.


Pump

Tank 1
Tank 2
delay 1 delay 2
c1
c2
y1
y2


Figure 1. A part oI a water transport network.

The discretised and linearised non-linear dynamic equations are:

v
1
(tt) v
1
(t) - k
1
c
1
(t) k
2
u
pump
(t-
2
)
- k
3
u
out1
(t) (4)

u
out1
(t) k
1
v k
4
v
1
(t) (5)

u
pump
(t) k [a(h-v
2
)
2
b(h-v
2
)c]
k
5
k
6
v
2
(t) (6)

v
2
(tt) v
2
(t) - k
7
c
2
(t) k
8
u
out1
(t-
1
)
- k
9
u
pump
(t) (7)

where: t is the sampling time.

II the measured variables are: v
1
, v
2
, c
1
and c
2
, the linearised model
equations to predict one step ahead
1
v and
2
v are:

1
v (tt) (1-k
3
k
4
).v
1
(t) - k
1
.c
1
(t)
k
2
k
6
.v
2
(t -
2
) k
2
k
5
(8)

2
v (tt) (1-k
9
k
6
).v
2
(t) - k
7
.c
2
(t)
k
8
k
4
.v
1
(t-
1
) k
9
k
5
(9)

Then, the analytical redundancy relations will be:

r
1
(tt)v
1
(tt)-
1
v (tt)
v
1
(tt) [(1-k
3
k
4
).v
1
(t) - k
1
.c
1
(t)
k
2
k
6
.v
2
(t -
2
) k
2
k
5
] (10)

r
2
(tt)v
2
(tt)-
2
v (tt)
v
2
(tt) [(1-k
9
k
6
).v
2
(t) - k
7
.c
2
(t)
k
8
k
4
.v
1
(t-
1
) k
9
k
5
] (11)


Then, using these relations and deIinition (2), the Iault signature
matrix considering only as possible single Iaults on level and
consume sensors will be:





f
v1
f
v2
f
c1
f
c2

s
1
1 1 1 0
s
2
1 1 0 1

Table 2. Fault signature matrix

where: f
v1
, f
v2
, f
c1
and f
c2
means Iaults in the sensor that measures
v
1
, v
2
, c
1
and c
2
respectively.

Looking at this matrix, Iaults in consume sensors (f
c1
and f
c2
) are
isolable since their signature is diIIerent Irom the signature oI f
v1

and f
v2
. However, Iaults in these last sensors (f
v1
and f
v2
) are not
isolable since they have the same signature.


3.2 Fault isolation

The isolation approaches presented in Section 2 uses a set oI binary
detection tests to compose the observed Iault signature. Since, the
isolation decision at time k only depends on the detection tests at
time k, this scheme is purely static, independently oI the approach
Iollowed. This can be at the origin oI Ialse isolation decisions,
especially when some detection tests have a transient behaviour
(especially in dynamic slow/delayed systems as in this example) in
response to the Iaults.

When considering the FDI or DX isolation approaches looking at
the problem oI the diIIerent time oI symptom appearance, their
behavior is very diIIerent during the transient time. As described in
Section 2 since FDI isolation algorithms Iollow the column view, iI
the diIIerent '1s corresponding to a given Iault signature do not
appear at the same time and the isolation decision is taken
continuously an erratic diagnosis will be provided.

Figure 2 shows the time evolution oI symptoms when a Iault in
sensor v
1
appears. The symptom s
1
is activated at time t
1
, while the
second symptom s
2
is not activated at same time because oI the
transport delay
1
. This symptom is activated at time t
2
. Table 3
illustrates the FDI and DX CBD diagnostic corresponding at the
time t
1
and t
2
in which each symptom appear. Using the Iault
signature presented in Table 2, FDI approach, that Iollow the
column view, provides the Iollowing diagnosis: between t
1
tt
2
the
observed Iault signature will correspond to that oI Iault in consume
sensor c
1
, and at t~t
2
, the observed Iault signature correspond to
that oI the Iault in sensor v
1
or v
2
. As a conclusion, while all
symptoms associated to a given Iault are not activated because oI
the diIIerent appearance times, an FDI isolation approach will
provide erratic diagnosis during the transient (see Table 3).

On the other hand, iI a DX CBD isolation approach is used since it
Iollows the row view, those components associated with activated
symptoms will be considered as possible Iaulty candidates. Using
the same example than in the FDI case, a Iault in sensor v
1
, the
activation oI symptom s
1
will provide as Iaulty components sensors
v
1
, v
2
and c
1
, considering only single Iaults to be comparable with
FDI. Then, the activation oI symptom s
2
will allow to reduce the
set oI Iaulty components to the sensor v
1
and v
2
. DX approach then
allows to reIine incrementally the Iirst diagnosis that already
include the Iinal Iaulty components without providing erratic
diagnosis in the transient (see Table 3).


Figure 2. Sequence oI Iault symptoms

Time t
1
t
2
Symptoms s
1
1 s
2
0 s
1
1 s
2
1
FDI f
c1 f
y1
f
y2
DX CBD f
c1
f
y1
f
y2
f
y1
f
y2


Table 3. Diagnosis with FDI and DX CBD approaches (in black
the provided diagnostic )

4 APPROACHES TO DEAL WITH
ISOLATION WITH DIFFERENT TIME OF
SYMPTOM APPEARANCE


AIter a review oI existent approaches, a set oI improvements to
deal with isolation with diIIerent time oI symptom appearance will
be proposed. The Iinal goal is to include these improvements in the
existent approaches (either FDI or DX) without increasing
complexity.

4.1 Introduction

Delayed/slow systems are an important class oI dynamic systems,
which generally are composed by many subsystems with important
delay times in their behaviour or with large transient responses.
Some examples oI delayed/slow systems are energy and water
distribution networks, chemical processes, etc. In this class oI
systems, when a Iault occurs, their symptoms never appears
simultaneously. II the decision procedure is static, as it is shown in
the previous example, the propagation oI the Iault produces a
diagnostic that change in time. This problem has been noticed by
Koscielny |12||13|, Console |6| and by Quevedo |21| among
others.

In both DX and FDI communities, there are recent works where the
inclusion oI temporal inIormation associated with the model and
the symptom appearance in isolation algorithm improve the
diagnosis results in case oI dynamic systems.

Most AI methods take advantage oI implicit temporal knowledge
in the structure oI the model, which could be stated as a set oI
QSIM/QDEs |9| |16|, temporal causal graphs |14| or causal
inIluences |24|. These pieces oI knowledge would help the
isolation process in the transient period caused by the Iault (it is
well known that Ieedback eIIects make extremely diIIicult to do
Iault isolation in steady state because they tend to compensate Iault
eIIects |15|). Temporal inIormation in the models (integral or
derivative operators within the constraints) can be used to predict
later deviations or to estimate delayed consequences. At the same
time, they can be used to Iavouring diagnosis candidates according
to the dynamics oI their symptoms |15| |24| or to reject them iI
their eIIects are not present at a given time |1|. In general these
techniques use rules that giving a symptom at time t, compute the
diagnosis taking into account the observations in times earlier than
t, and not exploited the ability oI discriminating with the
inIormation acquired at subsequent snapshots. They arrive to
diagnostics aIter a sequence oI transient diagnostics. Console |6|
proposes temporal decision trees that take into account temporal
inIormation on the symptoms and temporal constraints on the
recovery actions to be perIormed, increasing the Iault
discrimination. In this work each node is labelled with a pair
svmptom, time delav~, observed symptoms give multiple decision
trees, each edge is labelled with symptoms, and at the end the
diagnostic is obtained. The waiting time Ior diagnostic is intrinsic
in the proposed algorithm.

On the other hand, in the FDI community the works oI Koscielny
|12||13| propose a new Iault signature matrix replacing binary 1`s
by an interval oI time |t
min
,t
max
|,where t
min
and t
max
are the minimum
and maximum time period between Iault occurrence and symptom
appearance. The conclusion oI this approach is that taking into
account the time oI symptom appearance is possible to increase the
Iault isolation, and suggest the possibility to limit the number oI
time periods needed Ior a diagnostic.

4.2 Isolation with different time of symptom
appearance

One possible approach to deal with the problem oI diIIerent time oI
symptom appearance, independently oI the isolation approach,
consists in not allowing an isolation decision until a preIixed
waiting time (T
w
) has elapsed, Irom the Iirst symptom appearance
assuming permanent symptoms. This T
w
must be calculated Irom
the larger transient time response (T
lt
) Irom non-Iaulty situation to
any Iaulty situation. Moreover, depending on the nature oI the
Iaults (abrupt or incipient), there is usually a detection time that
should be added. OI course, Ior some Iaults this approach can be
very conservative since each Iault has its own time. In next
sections, additional proposals will be introduced in order to make it
less conservative.

In the example described in Section 3, the T
w
must be calculated
Irom the largest transient time response Irom non-Iaulty situation
to any Iaulty situation, being in this case the transient time
response Irom s
1
s
2
0 to s
1
s
2
1 is T
lt
max(
1
,
2
). Moreover,
taking into account the uncertainty oI the delays
1
and
2
and the
detection time, a security time should be added to T
w
in order to
make a diagnostic decision. In this example, then T
w
~ T
lt

max(
1
,
2
) (see Table 4).

In the case oI some DX approaches, it is also possible to exonerate
components not involved in not violated detection tests aIter the
waiting time T
w
. This will allow to obtain the same isolation results
than FDI approach.

Time t
1
t
2
T
w
Symptoms s
1
1 s
2
0 s
1
1 s
2
1 s
1
1 s
2
1
FDI I
c1 I
y1
I
y2
f
y1
f
y2
DX CBD I
c1
I
y1
I
y2
I
y1
I
y2
f
y1
f
y2


Table 4. Diagnosis with FDI and DX CBD approaches (in black
the provided diagnostic)
t
t
s
1
s
2
0
0
1
1
t
1
t
2
4.3 Isolation considering the order of symptom
appearance

A second improvement to the previous approach is to take into
account the order oI symptom appearance.

The Iault signature matrix gives a static relation between Iaults and
symptoms. Taking into account the order oI symptom appearance
could increase the capacity to discriminate one Iault Irom the rest,
because it will allow to discriminate those Iaults that have the same
static Iault signature but diIIerent dynamic Iault signature, i.e,
diIIerent order oI activation oI their associated symptoms.

ThereIore, a new Iault signature matrix is introduced named
dvnamic fault signature matrix (DFSM) (
d
) |21|. This matrix can
be composed using an analysis oI the propagation oI each Iault.
Each element oI this matrix has two digits: the Iirst digit (index)
codiIies the result oI detection test (0 means non-Iired test while 1
means Iired test) and the second digit (sub index) codiIies the
logical sequential order oI appearance oI the symptom Ior each
Iault. This sub index describes in a simple way the propagation oI
the Iault eIIect on the system and can be inIerred Irom the
analytical redundancy relations or by analysis oI the Iault eIIects in
the model.

Considering again the Iault signature matrix presented in Table 2,
the Iaults f
v1
and f
v2
have identical static Iault signature and
consequently are not isolable. However, taking into account the
order oI appearance (f
v1
produces an abnormal value oI s
1
and then
oI s
2
, f
v2
produces an abnormal value oI s
2
and then oI s
1
while in
the case oI f
c1
and f
c2
only one residual is activated), then they can
be isolated since their associated symptoms appear in diIIerent
order. The dynamic Iault signature matrix is shown in Table 5.

f
v1
f
v2
f
c1
f
c2

r
1
1
1
1
2
1
1
0
r
2
1
2
1
1
0 1
1


Table 5. Dynamic Iault signature matrix

Now, regarding the columns oI the dynamic signature matrix in
Table 5, all the theoretical Iault signatures are diIIerent, and it is
possible to isolate Iaults on f
v1
and f
v2
. Then, we can say that this
two Iaults are sequentially isolable but not static isolable. In Table
6, the same diagnosis problem than the one presented in Table 4
when considering the order oI symptom appearance.


Time t
1
t
2
T
w
Symptoms s
1
1 s
2
0 s
1
1 s
2
1 s
1
1 s
2
1
FDI

I
c1
I
y1
f
y1
DX CBD I
c1
I
y1
I
y1
f
y1


Table 6. Diagnosis with FDI and DX CBD approaches (in black
the provided diagnostic)


4.4 Isolation considering the time of symptom
appearance

A third improvement that can be introduced is to consider not only
the order oI appearance but also their corresponding interval oI
time oI appearance [ ] [ ]
max i min i i
d , d d = where d
imin
and d
imax
are
the minimum and maximum time oI appearance oI all symptoms oI
the Iault f
i
Irom the Iirst symptom appears.

Regarding Table 5, iI the system was aIIected by a new Iault f
5

with the same dynamic Iault signature than f
v1
, it will not be
possible to isolate the two Iaults even considering the order oI
appearance. But iI the time needed Ior each Iault to aIIect all the
symptoms is known, these two Iaults will be isolable iI they have
disjoint interval oI time oI appearance oI all symptoms. Moreover,
knowing this time will allow to shorten the waiting time beIore a
isolation decision is taken.

We propose to include in the DFSM, the interval oI time oI
appearance oI all symptoms Ior each Iault to make a diagnostic
decisions, according Table 5, Ior the proposed example including
the Iictitious Iault f
5
.

f
v1
f
v2
f
c1
f
c2
f
5
s
1
1
1
1
2
1
1
0 1
1

s
2
1
2
1
1
0 1
1
1
2

[d] |d
1
| |d
2
| |d
3
| |d
4
| |d
5
|

Table 7. ModiIied dynamic Iault signature matrix (assuming the
case d
1max
d
5min
)

In order to check this modiIied dynamic Iault signature matrix,
three diIIerent single-Iault scenarios corresponding to Iaults f
v1
, f
5

and f
c1
are considered and analysed in Table 8, Table 9 and Table
10, respectively. These tables present how this matrix allows to
isolate two Iaults that are not isolable using the order oI symptoms
since it is known the time interval in which symptoms should
appear (Table 8 and 9) and also to reduce the diagnostic time.

Time t
1
d
1min
t
2
d
1max

Symptoms s
1
1 s
2
0 s
1
1 s
2
1
FDI I
c1
f
y1

DX CBD I
c1
I
y1
I
5
f
y1


Table 8. Diagnosis with FDI and DX CBD approaches (in black
the provided diagnostic)

Time t
1
d
5min
t
2
d
5max
Symptoms s
1
1 s
2
0 s
1
1 s
2
1
FDI I
c1
f
5
DX CBD I
c1
I
y1
I
5
f
5


Table 9. Diagnosis with FDI and DX CBD approaches (in black
the provided diagnostic)

Time t
1
max(|d
1
|,|d
3
|,|d
5
|)

Symptoms s
1
1 s
2
0 s
1
1 s
2
0
FDI I
c1
f
c1
DX CBD I
c1
I
y1
I
5
f
c1


Table 10. Diagnosis with FDI and DX CBD approaches (in black
the provided diagnostic)

5 CONCLUSIONS

This paper presents a set oI problems that appear when Iault
detection and isolation are handled by separated modules as usually
the case in FDI methods. Because there is current trend in DX
community to separate also this two tasks when integrating
methods oI both communities using Ior example detection tests
based on FDI techniques and isolation methods based on DX
techniques, the problems presented in this paper should be oI
interest and also could be considered by DX community.

The paper identiIies the problems that are produced by the use oI a
binary codiIication oI the detection tests and the use oI static
reasoning scheme when applied to dynamic systems especially
those who present important transient in the detection tests aIter the
appearance oI Iault. The solutions proposed in the paper consider
some improvements to the basic FDI and DX CBD schemes
considering temporal aspects in the appearance oI symptoms. An
motivational example is used to illustrate how these improvements
work.

ACKNOWLEDGMENTS

This paper is supported by CICYT (DPI2002-03500 and DPI2002-
01809), by Research Commission oI the 'Generalitat de
Catalunya (group SAC reI.2001/SGR/00236) and by DAMADICS
FP5 European Research Training Network (reI. ECC-TRN1-1999-
00392).

REFERENCES

|1| C.J. Alonso, J.B. Pulido, J.J. Rodriguez, C. Llamas. 'Diagnosis
oI dynamic systems: a Iramework to integrate Iault modes and
learned domain knowledge. In Current trends in qualitative
reasoning and applications. 2000 Edition. Pages 33-47. Edicion
Digital tres. June, 2000. Spain.
|2| Blanke, M., Kinnaert, M., Lunze, J., Staroswiecki, M.
'Diagnosis and Iault-tolerant control. Springer-Verlag, 2003.
|3| Chantler, M.J., Daus, S., Vikatos, T. and Coghill, G.M. 'The
use oI quantitative dynamic models and dependency recording
engines". In Proc. oI the Seventh International Workshop on
Principles oI Diagnosis (DX-96). Pp. 59-68. Val Morin,
Quebec, Canada. 1996.
|4| Chen J., Patton, R.J. 'Robust Model-Based Fault Diagnosis Ior
Dynamic Systems. Kluwer Academic Publishers, 1999.
|5| Combastel, C., Gentil, S., Rognon, J.P. 'Toward a better
integration oI residual generation and diagnostic decision.
IFAC SaIeProcess 2003, Washington DC. USA.
|6| Console, L., Picardi, C., Thesider, D. 'Temporal Decision
Trees or the lazy ECU vindicated. IJCAI-INT Joint ConI. on
ArtiIicial Intelligence, 545-550, Seattle, 2001.
|7| Cordier, M., P. Dague, M. Dumas, F. Levy, J. Montmain, M.
Staroswiecki, and L. Trave-Massuyes, 'A comparative analysis
oI AI and control theory approaches to model-based
diagnosis. 14
th
European ConIerence on ArtiIicial Intelligence
(ECAI 2000), Berlin, Germany.
|8| de Kleer, J. 'Fundamentals oI Model-based Diagnosis. 14th
International Workshop on Principles oI Diagnosis, DX 2003.
|9| Dvorak, D. and B. Kuipers. 'Model-based monitoring oI
dynamic systems. Readings in Model Based Diagnosis. pp
249-254. 1992.
|10| Gertler J.J. 'Fault Detection and Diagnosis in Engineering
Systems. M. Dekker. 1998.
|11| Hamscher, W. and L. Console and J. de Kleer(Eds.). Readings
in Model based Diagnosis. Morgan KauImann Pub., San
Mateo, 1992.
|12| Koscielny, J.M. 'Fault isolation in industrial processes by the
dynamic table oI states method. Automatica 31(5), pp. 747-
753, 1995.
|13| Koscielny, J.M., Zakroczymski, K. 'Fault isolation method
based on time sequences oI sympton appearance. In
Proceedings oI IFAC SaIeProcess 2000, Budapest. Hungary.
|14| Mosterman, P.J. "Hybrid dynamic systems: a hybrid bond
graph modeling paradigm and its applications in diagnosis",
PhD Thesis. Vanderbilt University, Tennessee, USA, 1997.
|15| P. Mosterman and G. Biswas. 'Diagnosis oI continuous
valued systems in transient operating regions. IEEE T.on
Systems, Man, and Cybernetics, vol. 29, n6, pp. 554-565,
1999.
|16| Ng, H.T. 'Model-based, multiple Iault diagnosis oI time-
varying, continuous physical devices. 6
th
ConIerence on
ArtiIicial Intelligence Applications, Santa Barbara, USA, pp. 9-
15, 1990.
|17| Patton, R.J., P.M. Frank, R.N. Clark. 'Issues oI Iault diagnosis
Ior dynamics systems. Springer Verlag, 2000.
|18| Ploix, S., TouaI, S., Flaus, J.M. 'A logical Iramework Ior
isolation in Iault diagnosis. IFAC SaIeProcess 2003,
Washington DC. USA.
|19| Pulido, B., C. Alonso, F. Acebes. 'Lessons learned Irom
diagnosing dynamic systems using possible conIlicts and
quantitative models. In Lecture Notes in ArtiIicial
Intelligence. International ConIerence on Industrial and
Engineering Applications oI ArtiIicial Intelligence and Expert
Systems, IEA/AIE 01. Budapest, Hungary. June, 2001
(corrected version oI the article in DX, 2001).
|20| Pulido, B., C. Alonso. 'Possible conIlicts, ARRs, and
conIlicts. In Proc. oI the 13th International Workshop on
Principles oI Diagnosis, DX02. Semmering, Austria. Pages
122-128. May, 2002.
|21| Quevedo, J., Puig, V., Nejjari, F., Escobet, T. , Aguilar, J.,
Pulido, B. 'An improvement oI model based Iault isolation
based on a dynamic signature matrix Ior delayed systems.
Workshop on Advanced Control and Diagnosis, Duisburg,
2003.
|22| Reiter, R. 'A theory oI diagnosis Irom Iirst principles.
ArtiIicial Intelligence, 32:57-95, 1987.
|23| Staroswiecki, M., P. Declerk, 'Analytical redundancy in non
linear interconnected systems by means oI structural analysis.
IFAC Advanced InIormation Processing in Automatic Control
(AIPAC-89), Nancy, France, 1989.
|24|Trave-Massuyes, L. R. Milne. 'TIGER: Gas Turbine condition
monitoring using qualitative model based diagnosis. IEEE
Expert Intelligent Systems and Applications, 1997.

Вам также может понравиться