ADVANCED CONDITION MONITORING AND FAULT

DIAGNOSIS OF TURBO-ALTERNATORS USING AN

EMBEDDED EXPERT SYSTEM
J.R. McDonald, R.W.Stewart, B.D. Gemmell, E.J.M. Georgin - Centre For Electrical Power
Engineering, University of Strathclyde, Glasgow, Scotland.

B.J. Weir - Technology Development Centre, ScottishPower, East Kilbride, Scotland.

1 INTRODUCTION
Research in the area of artificial intelligence has recently focussed on a new problem solving approach
that involves the use of knowledge based expert systems. Expert systems are capable of applying
rules developed by human experts to heuristically solve problems in a specific domain. While such
a system will never surpass the capability of an expert, it does provide the opportunity to emulate
the reasoning process of an expert when one is not available. Expert system development provides a
consistent mechanism for preserving, transferring and documenting human expertise. Not only can
such an approach be used for problem solving, but it also provides a mechanism for giving the reasoning
behind conclusions reached or the questions asked. Expert systems can also handle the uncertainties
associated with the use of heuristic rules or real-world data.
It should be emphasised that expert system development is not a universal remedy for solving every
type of problem. Such an approach is viable for problems that require only cognitive skills and not
common sense. Also, the process of gathering human expertise and converting it into a knowledge
base is an expensive, time-consuming process. However, expert system development is justified in
cases where payback is high, where human expertise is scarce or being lost, when expertise is required
in many locations at once, or when intelligence is required in a hostile environment.

MONITORING AND DIAGNOSTIC EMBEDDED EXPERT

SYSTEM

This paper discusses a Monitoring and Diagnostic Embedded Expert System (MODEES) which forms
part of a more general Condition Monitoring and FAult Diagnosis System (COMFAD). This is a
rule based system with the ability to categorise power plant performance and recognise impending
failure or dangerous trending, and the need for preventative maintenance or remedial action. COMFAD/MODEES is targeted at operator personnel who must decide within a short period if the power
plant is fit for continued or prolonged service, and if not make a thoroughly reasoned "guess" at what
faults have occurred and how they can be remedied. On-line monitoring via real-time processing is
also necessary to provide sufficient information even with erroneous or missing measurements (McDonald1991). COMFAD/MODEES effectively accepts the processed sensor data input in real-time,
and provides diagnostic information output to the operator.

2.1

ON-LINE MONITORING

The use of sensor measurements for monitoring the electromechanical condition of large turbo-alternator
units is very important in maintaining the availability of the most economic$ units and in helping to
minimise the costs of planned and unplanned outages. The differential costs between high merit and
alternative source plant can exceed 1100,000 per day (Brooke1988).
Traditionally, data analysis and interpretation is carried out by expert engineering staff who advise
power station operators on which course of action should be taken to remedy any fault which may be

diagnosed. The effectiveness of this approach depends not only on the ability to analyse the sensor
data in detail but also on the skill and experience of the engineering personnel in interpreting the
results using any relevant results obtained previously.
AS a further development engineers have identified the potential for knowledge based technology to
interpret sensor d a t a in a similar fashion to the human analyst, and provide a statement about the
health of the turbo-alternator unit. The approach is to automatically apply specialist knowledge to
plant data and to aid the turbine staff to optimise plant performance. In line with these objectives
the knowledge based system will be required to give early warning of departure from accepted normal
operating modes and warning of the failure of components before such conditions actually occur. As a
first stage towards the development of a Diagnostic Support Environment (DSE), a system t o validate
turbine sensor output is required. This ensures that turbine diagnosis is performed using only data
that reflects the state of the turbine plant.
Both fault diagnosis and data validation knowledge based systems reason using qualitative descrip
tions of sensor data similar to those used by the human expert. An important part of this work is
the development of techniques for abstracting numerical data reported by the turbine sensors into
qualitative data descriptions about the characteristics of the reported data. The resulting system is
described as a knowledge based data abstraction and validation system which interfaces t o an On-Line
Vibration Monitoring System (OLVMS) and whose outputs are statements about the validity of the
sensor data together with qualitative statements which will be used as input to a knowledge based
diagnostic system.
An appropriate starting point for the diagnostic rules is the various bearings on the turbo-alternator
unit. These represent important plant items, the state of which provide symptoms leading to fault
diagnosis. The DSE classifies the turbo-alternator problems in terms of severity, origin, frequency and
interrelation. This enables the production of an effective monitoring system.

2.2

THE PROCESS OF KNOWLEDGE ELICITATION

The success of an expert system is largely dependent on the quality of knowledge obtained from the
domain expert. For this reason, the difficult process of obtaining and representing human knowledge is
critical. Knowledge elicitation requires an intermediary capable of extracting and structuring human
decision making knowledge and then transforming it into an acceptable representation system. There
are many different knowledge elicitation techniques (Diaper1989), however the methodologies best
applied are: Live Interviews; Teach Back; Protocol Analysis; Case Studies. The most efficient means
of obtaining domain expertise is to establish frequent, sharply focussed knowledge elicitation sessions
with the domain expert. Knowledge refinement is an important part of the knowledge elicitation
process.
The methodology adopted in the representation of the domain knowledge is the one advocated by
Keravnou and Johnson (Keravnoul986). Their methodology provides a set of tools and techniques
for eliciting modela of competence and mapping these models and the domain specific knowledge into
knowledge representation schemes.

2.3

SIGNAL PROCESSING SUBSYSTEM

A vital element in COMFAD/MODEES is the presentation of data to the expert system. Hence two
interfaces are required, the first to extract relevant information from raw sensor data, and the second to
perform signal to symbol transformation to allow meaningful presentation of power plant information
to the expert system. The first interface requires a signal processing subsystem. This consists of a
suite of algorithms for both time and frequency domain analysis of the sensor data. Plant failures can
occur over a long or short period of time. Hence the time domain algorithms include both short and
long time scale analysis techniques in order that trending can be performed over a wide range of time

912

scales, and failures perhaps detected before they occur. A novel feature about COMFAD/MODEES
is that the expert system can feed back to the signal processing subsystem. There will be occasions
when for example, high resolution data is required at particular harmonics of vibration frequencies
in order to accurately diagnose fault conditions. COMFAD has the agility to request and provide
this information. A DSP56001 is used to both receive (digital) sensor data, and implement the signal
processing algorithms. These algorithms are stored locally on the COMFAD hardware.

CROSS-SENSOR CORROBORATION

The harsh environment of a large turbo-alternator sometimes results in the malfunction of system
sensors. To provide sufficient confidence in the monitoring system, it is important to determine that the
data reported is a true reflection of the turbine state. A knowledge based strategy for sensor validation
is based on a technique employed by turbine experts called Cross-Sensor Corroboration (CSC). Baed
on experience of the plant and knowledge of its operation the analyst builds and maintains an intuitive
model of sensor relationships. The analyst uses this model to perform CSC by assessing whether the
relationships between pairs of sensors are consistent with what he expects. This information can be
elicited and encoded in the CSC knowledge base. It is clear that three classes of knowledge are used
during this activity. They are:
1. Knowledge about the physical structure of the turbo-alternator in terms of its components
and their connectivity. The turbine structure can be modelled as a semantic net where nodes
represent components and arcs form the connectivity between components (Weir1990).
2. Knowledge about which sensor sets can be causally related as a result of the transmission of
the physical process through the turbine structure. CSC attempts to identify descriptions of
significant relationships between pairs of sensed turbine parameters and then interpret them to
determine the validity of data transmitted by these sensors.
3. Knowledge on how to draw conclusions about the validity of data from a particular sensor on
the basis of pairwise comparisons between that sensor and a set of other sensors. There can of
course exist multiple causal interactions between turbine parameters transmitted by the sensors.
The resulting web of inter-sensor relationships provides the basis for performing CSC analysis,
which aggregates all relevant evidence provided.

DEVELOPMENT OF COMFAD/MODEES
Turbo Alternator
Set

SCADA

SignalProcessing

System

Subsystem

.-

4
I

V
Deep Plant

Models

MODEES
2.

(Monitoring and Diagnostic

Embedded Expert Systun)

Shallow Plant

Models

CorroborPcion)

(Diagnostic Suppon
Environment)

Operator
Interface

Figure 1 : Power plant expert system for condition monitoring and fault diagnosis (COMFAD).

913

The development of COMFAD/MODEES included a learning element allowing it to interact with

a domain expert to generate new rules t o be added to the knowledge base. The expert system is
structured to use several rule sets classified to cover : detection of sensot failure; identification of
plant model level based on data availability and required response time frame; selection of appropriate
signal processing algorithms for the analysis of techniques for particular modes of turbealternator
operation.
A general block diagram for COMFAD is shown in Figure 1. The system represents an advance in
the utilisation of measurement data to provide more automated plant monitoring. At present, most
condition monitoring systems concentrate on the data acquisition which is then presented to a human
operator. COMFAD/MODEES provides the potential for closing the loop in the plant monitoring and
control cycle as it embodies the appropriate domain expertise (illustrated by the dotted line in Figure
1).

CONCLUSIONS

This paper has shown that an expert system can intelligently process sensor data, emulating the
diagnostic strategies adopted by a domain expert. An expert system for turbo-alternator condition
monitoring and fault diagnosis can be extremely helpful either as an aid to the expert analyst or as
a source of knowledge when an expert is not available. While not intended to completely replace the
human expert, an intelligent knowledge based system for diagnosing rotating machines can be used
to solve easy problems, freeing the expert for more difficult and interesting problems.

REFERENCES

Brooke R.N.T., Clark D.M., Millar, A.G.. On-Line Vibration Monitoring Of Turbine-Generators In
The S.S.E.B. IMechE Seminar : Minimising Steam Turbine Outage Times, 18th October 1988.
Diaper D. Knowledge Elicitation : Principles, Techniques and Applications. Ellis Horwood Limited
(1989).
Keravnou E.T., Johnson L. Competent Ezpert Systems : A Case Study In Fault Diagnosis. Kogan
Page Limited (1986).
McDonald J.R., Stewart R.W., Gemmell B.D., Weir B.J. An On-Line Vibration Monitoring System :
An Embedded Ezpert System For Condition Monitoring And Fault Diagnosis On Turbo-Alternators.
Proceedings of COMADEM 1991, Southampton Institute, 2-4July 1991.
McDonald J.R., Burt G.M., Young D.J.. Alarm Pnxessing And Fault Diagnosis Using Knowledge
Based Systems For Transmission And Distribution Network Control. Proceedings of IEEE PES 1991
Transmission and Distribution Conference, Dallas, September 1991.
Taylor T., Lubkeman D.. Applications of Knowledge Based Pmgmmming To Power Engineering
Problems. IEEE Transactions on Power Systems, Vol. 4, No. 1, February 1989.
Weir B.J., Glen J.A.. Knowledge Based Sensor Data Validation. Paper submitted to the Institute of
Measurement and Control, 18 October 1990.