A
Comparison
of
Model-based Reasoning and Learning approaches to Power
Transmission Fault Diagnosis'
Ramesh
K.
Rayudu Sandhya Samarasinghe Don Kulasiri
Dept. Natural Resources Eng.
Ctr. for Computing
&
Biometrics
Dept. Natural Resources Eng.
Lincoln Univetsity,
Canterbury,
New Zealand.
rayudur@iincoln.ac. nz
sandhya @lincoln.ac.nz
kulasitd @lineoh. ac. nz
Abstract
An
application
of
model-based reasoning and model-
based learning to an operative diagnostic domain such
as electrical power transmission networks is presented.
Most
of
the research in model-based diagnosis
is
based on maintenance diagnosis. Operative diagnosis,
on
the other hand,
is
done while the system is still in
operation even afer the fault.
We plan to develop an
efficient algorithm
for
operative diagnosis which can
handle large domain
of
faults and multiple faults in
real time.
In
our search toward a better algorithm, we
develop and compare
two
different reasoning methods:
diagnosis based
on
model based reasoning, and
diagnosis based
on
heuristic rules leamt ji-om model
based reasoning. This paper presents the results
of
the
comparison.
Introduction
Operative diagnosis (OD) of a physical system is
the process
of
detecting faults while it is in operation
whereareas in maintenance diagnosis (MD)
[l],
the
fault diagnosis is done offline.
Operative diagnosis is
needed for systems which cannot be stopped for
maintenance (as it is too expensive), and the diagnosis
involves
the
consideration of symptoms
and
state
which can change with time.
In electrical power
transmission networks, operative diagnosis is confined
to alarm readings in real time while the effects of the
faults are still propagating through the network.
OD
is
heuristic in nature and often provides a challenging
task for experts involved. Experts find that the pattern
recognition of
alarms
triggered by a fault in the system
is relatively easier task compared to the identification
of the physical origins of the fault from a list of alarms.
This difficulty could be due to several components
malfunctioning at the same time within the network.
Power transmission networks carry power from
supply utilities to the consumer and any fault in the
network directly affects the consumers. The hazards of
performing fault diagnosis in this domain incorrectly
and too slowly result in notable accidents such
as
1977
New York City blackout where the power restoration
took several minutes causing inconvenience to
consumers. Earlier research
[2]
has shown that
decision-support systems can aid system controllers
during emergency situations. This paper presents the
proceedings of our attempt to develop an efficient fault
diagnostician for power transmission network in New
Zealand. Towards this development, model based
techniques
are
applied to the domain and the results are
presented in this paper.
Model
Based Diagnosis
Model based diagnosis (MBD) is a form of
diagnostic reasoning which incorporates operational
principles of the devices in the form of models where
the problem can be simulated under ideal conditions
and the output (predicted) is compared with the
observed output
[l].
This comparison would reveal
useful information for problem solving such as the
status of the equipment in the power system network
'
The
funding from Trans
Power NZ
Ltd.,
New
Zealand,
for
the
project
is
greatly appreciated.
218
0-8186-7174-2195 $04.00
0
1995
IEEE
and the authenticity of information reaching the control
centres. Figure
1
shows operation of the fault
diagnosis system using the model based approach.
I
Figure
1.
Model based reasoning.
As model based diagnosis or MBD is ireasoning
based on system model’s behavioural comparison, it
does not use any heuristic information about system
failures
[I].
Due to this reason model based diagnosis
can be more advantageous over heuristic reasoning.
Some of these advantages are listed below
[
101:
MBD
covers a wide range of fault sciniiios than
heuristic reasoning because MBD is basad on the
system’s behavioural analysis.
MBD can detect deviations from the expected
behaviour.
MBD can detect malfunctioning equipment in the
early stages.
MBD can predict the effects of the fimlts and
unnecessary alarms as it simulates the faults.
MBD can handle multiple faults efficiently because
the cascading effects of the faults can be simulated
and analysed.
Adopt rule-based system’s goal driven reasoning
POI.
Fault Diagnosis In Power Transmission
Networks
A prototype model based system for fault analysis
and diagnosis (MoBFAD) of electrical power
transmission was developed and was implemented in
Prolog
on
L”uX/486. The motivation for its
development was to produce and test
a
generic
diagnostic system for electrical power networks.
MoBFAD has a generic reasoning mechanism and can
be setup to work for a particular network by encoding
the model of that power network in the system.
MoBFAD is based on Reiter [3] and de Kleer,
et
al.’s [4] definition of model based diagnosis
represented as a 3-tuple (system description,
components and observations) and Igor Mozetic’s
[
13
hierarchical model based diagnosis. The advantage of
Reiter’s approach
is
that it can infer diagnoses in terms
of
components from a set of observations. The
important feature of these diagnoses is that they are
minimal2. Hierarchical structure is employed because
the power system can easily be divided into different
levels. In addition, hierarchical model based diagnosis
is efficient to handle complex systems with large
number of components or states of components and this
helps a great deal when multiple faults are considered
for diagnosis.
New Zealand’s power transmission network
comprises 13,005 route
Kms
of both AC and DC
power cables and transmits around 30,000 Gw of
power per annum. For a network of this size, time and
accuracy plays a very important role on its fault
diagnosis. Solving the diagnostic problem in
hierarchical levels has proven [5] to be faster than
conventional model based reasoning. Diagnosis
in
MoBFAD is done in 3 levels:
1.
Observation level.
2.
Synopsis level.
3.
Abstract level.
Observation level is the level where the components
are observed for their behaviour. Any change in the
network is reflected on this level; therefore, it
is
always
active. The components are checked almost
immediately for their operation and the component
operation information
is
then passed onto synopsis
level (level
2)
where it is used
to
derive more abstract
diagnoses. Abstract level (level 3) is a generalised
heuristic level where the domain experts’ knowledge is
encoded and is used in the diagnosis at “network”
level. Abstract level also has heuristic knowledge
relating to problems (for example, cable faults) which
do not need detailed diagnosis. Figure
2
shows the
organisation of MoBFAD’s hierarchical levels.
‘igure
2.
Hierarchical modelling
in
MoBFAD.
’
A
minimal diagnosis is such
that
by
changing
a
status
of
any
abnormal component
to
normal would make the diagnoses
inconsistent with the observations.
2
19
Goal driven reasoning in MoBFAD
Model
Based
Learning
MoBFAD starts with “abstract” level by trying to
solve the problem with its shallow reasoning
knowledge. Observation and Synopsis levels
are
triggered by changes in the network and hence they
operate in real time. If the shallow reasoning process
of “abstract” level cannot find a solution, then the
outputs (components’ behaviour) of “observation” and
“synopsis” levels are considered for the diagnosis. The
behavioural component of the model is expressed
as
rules describing its qualitative properties. The model
does not deal with quantitative representation of the
components; instead, it symbolically describes the
components’ characteristic features. Examples of
qualitative modelling can be found in
[5]
and [6] where
the technique is used to model the electrical activity
of
heart and electrical power components, respectively.
An example of alarms input to MoBFAD would
illustrate how model based reasoning works in practice.
When the alarm list is input to MoBFAD, the list
is
analysed by “observation” level to find any
malfunctioning of the components. Then heuristic
rules are applied from the “abstract” level to solve the
problem using shallow reasoning. If the problem could
not be solved using shallow reasoning, then rules in
“synopsis” level use the output from “observation”
level to deduce some preliminary diagnosis.
This
preliminary diagnostic output
is
then used by “abstract”
level to arrive at a solution.
Depending upon the problem, MoBFAD takes
0.13
to
2.5
secs (with multiple fault alarms) to arrive at
a diagnosis. It is inferred that problems which can be
solved using shallow reasoning are solved faster than
model based reasoning3. This brings our focus on
using machine learning algorithms to learn from
simulated output
of
models and create generalised
heuristic rules. The algorithm would search through
the logical aspects of the reasoning process and record
the common search paths by generalising them.
The
work done by several others has also prompted our
utilisation of machine learning techniques. Mozetic
[
514 discusses the compression
of
rules using Induction
and Fattah
&
0’
Rorke[7I5 discuss the learning of
association rules from models using Explanation based
learning. Due
to
our
domain’s
generalised nature, we
chose to investigate some algorithms based on
explanation based learning
[
81.
Tests
were conducted with same
alarm list
on both reasoning
techniques. Model based reasoning
took
0.24
secs
to
find the faulty
device where
as
shallow reasoning took
0.19
secs to solve the same
problem.
This research is applied
to
operative diagnosis.
This
research is applied to maintenance diagnosis.
Model based learning can be done in two ways:
1.
learning knowledge by analysing the goal driven
reasoning of model based reasoning [7].
2.
learning meta-knowledge
from
existing knowledge
using a qualitative model of the domain
[8].
The first method was used to develop the learning
mechanism. Two machine learning algorithms which
are generally based on Explanation Based
Generalisation (EBG), Peter Clark’s “Lazy Partial
Evaluation” (LPE) [8] and Mitchell
et
al.’s
“Explanation Based Generalisation” (EBG) [SI, were
chosen for this purpose.
Lazy
Partial Evaluation
Lazy Partial Evaluation (LPE) is a learning
algorithm which is a hybrid
of
explanation based
generalisation (EBG) and partial evaluation (PE)
algorithms. Learning of LPE is same as that of EBG
but,
it
also
includes PE’s ability to generalise and store
the failed proofs. LPE replaces the original theory by a
more generalised equivalent thereby making the run-
time of the solution faster.
The main advantages of
LPE over EBG are
[XI:
l+
LPE eliminates EBG’s repeated computation as it
saves the total work done in exploring proofs other
than the main proof.
l+
LPE’s “alternative proof saving” eliminates EBG’s
“masking effect6” which enhances the quality of the
solution.
Explanation Based Generalisation
Explanation based generalisation is an articulation
of the common aspects of various explanation based
learning systems [9]. It is based on generalisation of a
proof for a positive example by synthesising an
operational definition of the proof in terms
of
its sub-
goals. The new definitions thus created are then added
to the domain for future reasoning.
Model based learning in MoBFAD
Learning in MoBFAD is primarily based on Fattah
and
O’Rorkes’
[7]
proposal
of
integrating
EBG
and
model based reasoning. However, their model is based
The Masking Effect of EBG is its inability
to
find altemative
solutions
as
this
may sacrifice the learning
of
most efficient solution;
EBG learns only one solution it finds at the fist instance.
220
on maintenance diagnosis and need several additions.
Our
system adds the “state of the system” and learns
the association rules with respect to MoBFADl’s model
hierarchy. In general, the learned rules will be of the
following format:
top-level-rule(Y)
if
state-of-sy stem(SD),
state-of-component(Comps ),
state-of -observations(Obs).
The variable
Y
is a function of logical variables
appearing in the rule.
The resulting systems of this integration of
machine learning algorithms and model based
reasoning are called MOBFAD&, (EBG+MBR) and
MOBFAD,, (LPE+MBR). Both algorithms learn rules
in advance.
Comparative Results And Analysis
To evaluate the algorithms, seven training
(4
positive,
3
negative) examples are chosen.
The
results
are shown in Table
1.
MOBFAD,, generates fewer
hypotheses and goal expansions than MoBFAIDhg.
CPU
time
Figure
3
shows the graph of CPU time utilised by
each algorithm to solve* all seven examples. All the
examples for this analysis are taken from only one
range of problems (a range with no pilot wire faults and
observations are incomplete). During this an,alysis, the
state of the power system is also considered constant.
Under these circumstances, MOBFAD,, takes more
time to solve the proofs, but for the rest of the
examples it used less time to solve.
This
is because
MoBFAD1, reformulates the original definitions into
several definitions using some initial examples and
uses them on the following examples. On the other
hand, MoBFADhg does not use much time to learn
in
the initial stages, but it incrementally learns with the
The performance comparison is
based
on
Peter
Clark‘s
criteria
@I.
*
Solving in here relates to algorithms’ attempt to satisfy the goal
and learn concept definitions (rules) simultaneously.
number of examples. Eventhough MoBFAD1, is
slower in the initial stages, it is more efficient than the
other algorithms.
MoBFADhg is also efficient than
MoBFAD but sometimes it uses more time than
MoBFAD as in example
5
(Figure
3).
-------____
-----__
----..--____
----__
Figure
3.
Performance graph of CPU-time.
Number
of
rules learnt
From the CPU time graph, it can be estimated that
MoBFADI, learns more rules than MoBFADh,. For a
given example, MoBFADl, learns more concept
definitions (see figure
4)
than MOBFAD&,. This is
because MoBFAD1, learns
“all”
possible “good”
explanations of the example and it also learns some
“not relevant” explanations along with explanations
which lead to “failures”. MoBFADhg learns only the
explanations which satisfy the example. Hence its
upgradation of rules is linear. MoBFADl,, on the
other hand, replaces the original definitions with new
ones and this changes number of rules accordingly.
1234567
Learning Examples
4gure
4.
Graph
of
number
of
rules learnt.
Discussion
The above tests are done by considering constant
“state”
of
the
power
system
network. The need
for
this
assumption is to test the efficiency of the algorithms
221
under conditions used in other papers. Figure
5
shows
the performance of each algorithm under dynamic
conditions of power system state. The examples were
chosen from a wider range of problems. According to
the graph, reasoning with model based learning
algorithms is faster than model based reasoning. On
the other hand, it is also evident that the number of
inferences used to find a solution increases with the
increase
in the number
of
rules. MOBFAD,, is a
preferred algorithm for fault diagnosis under constant
network state but due to its longer learning time, it is
not recommended for networks under dynamic
conditions. In addition, MOBFAD,, may not be able to
deal with larger domains such
as
power transmission
networks efficiently
[SI.
MOBFAD&, is consistent
under both (constant and dynamic) conditions of the
network but clearly demonstrates its inability to learn
an efficient solution
to
the problem (see example 3
&
5
in Figure 5). At this stage, it can be concluded that
MOBFAD, is preferred over MoBFADl, and
MoBFAD because MOBFAD&, takes less time on
average to arrive at a solution.
3
E
1500
%
=
1000
0
z
”..
500
0
k
t
1
t-----t-c
234567
Examples
Figure
5.
Performance graph of all algorithms
under operative diagnostic conditions.
Conclusion
We have presented three algorithms applied
to
operative diagnosis based on model based reasoning
and machine learning. MoBFAD, a pure model based
system,
is
based on Igor Mozetic’s Hierarchical Model
Based Diagnosis and incorporates
the
domain
models
in three levels. MoBFADl, and
MOB FAD&^
are
hybrid systems combining model based reasoning and
machine learning. The machine learning algorithms,
LPE and EBG, are used to learn top level rules
from
the model based diagnosis. All three algorithms are
compared and their performance analysed.
The
hybrid
algorithms work faster than model based diagnosis
unless the top level rules grow unacceptably large.
Under dynamic system conditions however,
MOBFAD, seems to provide the required solution
faster.
Acknowledgments
Thanks to Peter Clark for letting
us
use his Prolog
version of LPE and EBG algorithms.
Thanks to Control Centre Staff at SICC
&
NICC,
Trans Power NZ Ltd. for their feedback and support.
Thanks to Dr. Eduardo Morales
of
Institute of
Electric Power, Mexico and Dr. Rambabu Adapa
of
EPRI, USA for their help and support.
References
1.
Mozetic I.,
Hierarchical Model-Based Diagnosis,
Intl. Jnl. Man-Machine Studies, 35 (3), 1991,329-362.
2. Wollenberg B. F.,
Feasibility study for an Energy
Management System Intelligent Alarm Processor,
IEEE Trans. On Power Systems, May 1986, pp 241-
247.
3. Reiter R.,
A Theory
of
Diagnosis from First
Principles,
Artificial Intelligence 32, 1987, pp 57-95.
4. Johan de Kleer
&
Williams B.
C.,
Diagnosing
Multiple Faults,
Artificial Intelligence 32, 1987, pp 97-
130.
5.
Mozetic I., Diagnostic efficiency
of
deep and surface
knowledge in KARDIO, AI in Medicine 2 (2), 1990,
6.
Hau-Wagenbauer M., and Nejdl W.,
Integrating
Model-based and heuristic features in a
real-time
Expert System,
IEEE-EXPERT, August 1993, pp 12-
18.
7.
El Fattah
Y.
&
O’Rorke P.,
Explanation-Based
Leaming for Diagnosis,
Machine Learning, 13,1993,
8.
Clark
P.
&
Holte R.,
Lazy
Partial Evaluation: An
integration
of
explanation based generalisation and
partial evaluation,
Proc. 6th European Conf. On
Machine Learning (ECML-93), 1993, pp 360-365.
9. Mitchell
T.
M., Keller R. M., and Kedar-Cabelli
S.
T.,
Explanation based generalisation: A unifying view,
Machine Learning, 1,1987, pp 47-80.
10.
Wang
X.,
Dillon T.,
“A second generation expert
system for fault diagnosis”,
Electrical Power
&
Energy
Systems, Vol
14,
No
213, ApriVJune 1992, pp 212-216.
pp 67-83.
pp 35-70.
222