Sunteți pe pagina 1din 7

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 15, NO.

2, APRIL 2000 585

Model-Aided DiagnosisA New Method


for Online Condition Assessment of High
Voltage Circuit Breakers
Michael Stanek and Klaus Frhlich, Senior Member, IEEE

AbstractIn recent years, online condition assessment of high personnel in identifying a problem by analyzing the monitored
voltage equipment is increasingly asked for by power utilities. data in much the same way a human expert would. This can
However, most systems on the market today are merely monitoring help to reduce outage times of equipment or even prevent major
solutions with hardly any diagnostic capabilities. This report
reviews the basic principles for condition diagnosis from the field failures by pre-warning of developing problems before they be-
of artificial intelligence and introduces a novel strategy, named come serious.
Model-Aided Diagnosis (MAiD). It was initially developed for This report presents a novel method for condition diagnosis of
condition assessment of high voltage circuit breakers but is also high voltage circuit breakers. It was developed with the prospect
suitable for other technical devices. Theoretical considerations of implementation on a computer with low computational power
as well as practical results show that MAiD is a useful strategy
for condition assessment. In addition, with its low hardware in mind. Therefore, it can easily be incorporated into a dedicated
requirements it allows online condition diagnosis of high voltage local monitoring device or into an intelligent substation control
circuit breakers locally by a low-cost monitoring device. system.
Index TermsFault diagnosis, circuit breakers, modeling.
II. REVIEW OF PRINCIPLES FOR CONDITION DIAGNOSIS
I. INTRODUCTION
As defined in [4], diagnosis is the process of determining
URING the last decade, power utilities have displayed in- the nature of a fault or malfunction, based on a set of symp-
D creasing interest in ways to extend the life-cycle time of
their high voltage equipment. The most discussed approach is
toms. Input data (the symptoms) are interpreted and the under-
lying cause of these symptoms is the output. In practical ap-
condition monitoring and diagnosis in order to find developing plication on a high voltage circuit breaker it is to identify the
faults of the equipment at an early stage, or at least to identify faulted component(s) and the kind of fault in the breaker from
the nature of a suddenly occurred problem from measured data. the data provided by its sensors.Note that in this context the
The findings of the second international enquiry on the reli- word sensor is used in a broad sense to include such elements
ability of high voltage circuit breakers [1] by CIGRE working as voltage or current transformers and limit switches.
group 13.06 indicate that almost half the major and minor faults Three basic strategies for assessing the condition of a system
occurring are located in the operating mechanism. Hence, the under observation are known, namely
greatest emphasis on condition monitoring and diagnosis should rule-based diagnosis,
be placed on this component. This recommendation was also model-based diagnosis,
followed in the present work. case-based diagnosis.
During recent years, several devices for condition assessment
Since they are the basis for the decision to develop a new
of high voltage circuit breakers have been introduced (see e.g.
approach they will be introduced briefly with their respective
[2]) by circuit breaker manufacturers and also by independent
strengths and shortcomings. For details the interested reader is
companies. However, close examination of these devices reveals
referred to literature on expert systems [4] and artificial intelli-
that these are plain monitoring solutions i.e. they record sensor
gence [3].
data and perform boundary checks and similar operations on
them. In most cases, the operator is left to drawing his own con-
clusions on the cause and nature of the fault, based upon the A. Rule-Based Diagnosis
numbers presented by the monitoring device. Rule-based diagnosis is based on a set of heuristic relation-
In the field of artificial intelligence (AI), much effort has been ships between symptoms and causes which form the knowledge
devoted to condition diagnosis [3], which may also be applied to base for the diagnostic process. These relationships are called
high voltage circuit breakers. Its purpose is to assist substation rules. In condition assessment, a diagnosis is found when all
rules pertaining to it are fulfilled.
Manuscript received January 26, 1999. This work was supported by ABB Strengths: The knowledge base can be entered fairly
High Voltage Technologies Ltd., Switzerland, and ABB Power T&D Company straightforward by a human expert to reflect his or her ex-
Inc., Greensburg PA. perience. It is very efficient in terms of required computer
The authors are with the High Voltage Laboratory, Swiss Federal Institute of
Technology Zurich, 8092 Zurich, Switzerland. performance and computation time. Small changes to the
Publisher Item Identifier S 0885-8977(00)03476-2. knowledge base can be implemented fairly easily.
08858977/00$10.00 2000 IEEE
586 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 15, NO. 2, APRIL 2000

Shortcomings: Rules are mostly heuristic in nature cases) is still substantial, so far no universally applicable auto-
(shallow knowledge) and lack a deeper understanding of the mated methods have been proposed.
principles governing the operation of the system. A human
expert needs to design the rules, which is a tedious process and D. Application to High Voltage Circuit BreakersTodays
carries the possibility of introducing errors into the knowledge Situation
base. It is almost impossible to handle missing or unexpected Following the evaluation of the three basic diagnosis strate-
data as well as unforeseen problems. The method is not flexible gies listed above it was determined that the model-based ap-
in creating a knowledge base for a new or enhanced system proach is the best suitable one for condition diagnosis of high
(different type of circuit breaker) from an existing one. In many voltage circuit breakers. However, a market study revealed that
cases, diagnosing faulty sensors requires a large additional the presently available software packages for model-based di-
number of rules, increasing the complexity of the knowledge agnosis are not suitable for this purpose, for several reasons:
base. System requirements: The model-based systems under con-
sideration require a UNIX workstation to run, one for each cir-
B. Model-Based Diagnosis cuit breaker.
Model-based diagnosis uses a computer model which con- Price: In addition to the necessary computer platform the
tains the structure as well as the parameters of the device, im- model-based diagnosis software must be purchased, one license
plicitly comprising the physical principles which govern its op- for each installation, which is often more than a thousand US$
eration and possible malfunctions. The diagnostic engine runs for a single run-time version.
the model base and automatically finds all diagnoses that can Data Transmission: For a centralized workstation-based
explain or at least are consistent with the observed anomaly. monitoring system as described above, all the sensor data
Strengths: The model contains deep knowledge i.e. it is need to be transmitted to the substation computer for evalu-
based on the laws of physics governing the operation of the de- ation. However, substations are not generally furnished with
vice. Simulation permits access to hidden quantities (not ac- high-speed data links between switchgear and the station house,
cessible for direct measurement). The model base can be modi- the standard is very often still conventional wiring. Hence, it
fied for use on a different device fairly easily. No previous field would be more useful to perform the monitoring and diagnosis
experience is necessary, making this a suitable technique for task locally at the breaker, by a specialized device, and only
novel devices. Previously unforeseen or multiple faults can be report problems to the substation personnel. The previous
detected provided they can be explained through the model. A arguments make it clear, though, that this is not a feasible
faulty sensor can be detected as easily as a fault in the system solution with the systems available on the market today.
itself. Therefore, the remaining options were either to use a strategy
Shortcomings: The initial effort for creating and adjusting different from the model-based one or to find a new approach to
the model is substantial. Computing the diagnosis requires using it. The latter way was chosen and the result is described
high computer power. Structural changes in the device (e.g. a in the next section.
bridge fault between two adjacent wires) are almost impossible
to identify. III. MODEL-AIDED DIAGNOSIS AS A NEW METHOD FOR
CONDITION ASSESSMENT
C. Case-Based Diagnosis The approach for diagnosis introduced in this report is a
Case-based diagnosis uses an explicit database of past prob- novel combination of the case-based and model-based diag-
lems with their symptoms in order to find a solution for a new nosis strategies, in order to retain certain important advantages
problem situation. As opposed to the rule-based and model- of both. This is achieved by evaluating the model off-line line
based approaches, the knowledge base for case-based diagnosis to simulate the anticipated fault modes of the circuit breaker. A
does not need to be assembled explicitly by a human expert, search module, which can be rather simple, uses the resulting
rather it is generated automatically by analyzing actual cases fault cases to find the most likely diagnosis. This approach is
from the past and generalizing them for future use. A diagnosis called Model-Aided Diagnosis (MAiD); it is described in detail
is found by retrieving the best matching case from the knowl- below.
edge base, modifying it if necessary, and applying it to the new
problem situation. The result can be stored as a new case for fu- A. Basic Concept
ture use. Model-aided diagnosis is based on a functional computer
Strengths: The knowledge base is generated automatically, model of the system under observation. This model can
reducing the overhead and error probability of implementation. accurately simulate the systems behavior under normal and
The system learns with every incident. The knowledge base fault conditions. When it is fed with the internal and external
is always consistent and correct. A faulty sensor can be detected operating conditions it will yield the same data that would be
as easily as a fault in the system itself. measured by the sensors under these conditions.
Shortcomings: Since a history of fault cases is required to fill From the sensor data, meaningful quantities for assessing
the knowledge base, this approach is hardly suitable for novel the condition of the system are extracted. These are called
systems. Cases contain no deep knowledge. The initial effort features. The set of all features at a given time is called system
for designing the system (storing, retrieving, and matching of status. In addition, the confidence of the diagnosis can be
STANEK AND FRHLICH: MODEL-AIDED DIAGNOSIS 587

MAiD can be used for novel types of equipment (HVCBs)


even before any operational experience has been gained since
all conceivable faults can be simulated beforehand. Thus, a pow-
erful system for condition monitoring and diagnosis can be sup-
plied with the device from its first installation.
Since the actual diagnosis (Part 2 of the algorithm) requires
no more than finding the best matching case in a pre-defined
database, it can be executed on a low-performance computer
or monitoring device. Or, for even lower computing demands
on the peripheral hardware, only the feature extraction can be
performed locally and the actual diagnosis is performed by a
central diagnosis computer which holds the databases of several
pieces of equipment.
Adaptation to similar types of equipment can be performed
fairly easily by modifying the model operating parameters to
suit the new application. After that, all that needs to be done is
generating a new database and installing it onto the diagnostic
computer. Since the database is generated automatically, the in-
formation contained therein is always consistent and complete.
If a fault not originally considered is encountered later it can
easily be added to the database if desired and if the diagnosis
software permits this option. In order to maintain consistency
with other installations, though, the recommended procedure is
to incorporate this fault into the fault list, generate a new data-
base, and install it on the diagnosis system.
Shortcomings: MAiD works best for analysis of transient
events; for continuous supervision, it is necessary to define
events for conducting the diagnosis, either periodically or from
Fig. 1. Algorithm for Model-Aided Diagnosis. certain changes in system operation. This is due to the fact that
simulation of the operating condition needs to be conducted be-
increased by compensating for dependencies of the features on forehand, and the simulation can only run for a finite duration.
environmental influences such as temperature prior to further However, careful pre- and post-processing of the sensor and
processing. simulation data as well as selection of the features for diagnosis
The basic strategy for MAiD is two-fold. First the computer can to a large extent mitigate this limitation.
model is used to simulate all possible faults; the results are If the occurrence of more than one fault at the same time shall
stored in a database. This database is used on-line for finding be taken into account the database can grow very large quickly,
the most likely diagnosis when a deviation from normal opera- increasing the storage requirements on the diagnostic system.
tion has been detected. Fig. 1 gives the algorithm of MAiD in In calculating the necessary memory size it must be considered
form of a structogram. For details the interested reader is re- that the number of simulated cases N is
ferred to [5], where each step is covered in depth.
N = E 1 S; (1)
B. Discussion
E ... number of elements in the list from Part I Step 1,
MAiD is a powerful combination of the model-based and S ... number of operating states e.g. closing or opening the
case-based approaches to condition diagnosis. Originally devel- breaker, CB on-line or not, etc.
oped for diagnosing high voltage circuit breakers, it can be used
for condition assessment of most kinds of technical equipment, IV. APPLICATION TO HIGH VOLTAGE CIRCUIT BREAKERS
as long as a model of the device exists, with moderate expendi-
ture. As mentioned before, the initial application for model-aided
Part I (Preparation) of MAiD is essentially the model-based diagnosis is on high voltage circuit breakers. A live tank breaker
part. The computer model is used to conduct experiments by with spring-hydraulic mechanism was selected for a first test
simulating various operating conditions. The results are used to implementation, for two reasons: First, it is a well-established
build a knowledge base for the case-based second part (Diag- product with thousands of installations worldwide. Second, one
nosis), where the best matching case is looked up to produce a single-phase exemplar of this breaker type, was available to the
diagnosis for the current condition. author for functional analysis and measurements.
Strengths: Much like true model-based diagnosis, MAiD
permits access to quantities that are not or cannot be measured A. Computer Model
in the system, helping to really pinpoint the location and kind The circuit breaker employed is an SF6 -insulated self-blast
of the problem. high voltage ac circuit breaker with two interrupters on one
588 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 15, NO. 2, APRIL 2000

Fig. 2. Simulink model of high voltage circuit breaker. PT = Potential Transformer, CT = Current Transformer, R = resistance. Other abbreviations on port names
refer to the origin or destination block of the signal, e.g. TC = Trip Coil.

column, in T shape, per phase. It is operated by an hydraulic TABLE I


DIAGNOSTIC FEATURES FOR TEST IMPLEMENTATION
mechanical drive where the mechanical energy for the opera-
tions is provided by a plate spring assembly. Energization of a
control coil actuates a change-over valve which in turn applies
or removes the system pressure to/from the main piston head,
causing it to move from OPEN to CLOSED position or vice
versa. The oil volume lost from the high pressure storage is re-
placed by a hydraulic pump, which is controlled by the position
of the spring column via a limit switch.
These functions were implemented in Simulink (a tool for
dynamic system simulation for MATLAB) to form the circuit
breaker model. All components were modeled individually and
finally assembled into the complete system. Sinceaccording
to the CIGRE study [1]the drive mechanism is the most fault-
prone part of an HVCB, the highest effort went into modeling
this component with all its relevant elements.
The complete model is shown in Fig. 2. No detailed descrip-
tion can be given here, rather its purpose is to give the reader an
impression of the complexity involved.
This model can simulate the function of the circuit breaker
and each component, which is the important part for diagnosis,
rather than the physical structure. It is called and evaluated by
means of several MATLAB functions.

B. Test Implementation
most cases this results in a fault; the diagnosis attempts to
For simulating various faults and operating conditions in the find out which input parameter was changed in what manner.
circuit breaker, a number of parameters can be changed in the The diagnostic features were selected such that even
computer model. When changed from their default values, in without special training they would be meaningful to operating
STANEK AND FRHLICH: MODEL-AIDED DIAGNOSIS 589

Fig. 3. Diagnosis report from test implementation, with reduced hydraulic flow cross-section (cf. Fig. 5, curves b). The default and estimated values in the upper
section of the window refer to the parameter which caused the fault. The lower section contains a complete list of features (scrollable) with their normal and
measured values and the corresponding values of the best matching case from the database.

Fig. 4. Excerpt from circuit breaker model, containing those components and
quantities which directly influence the piston position.

Fig. 5. Simulated (dotted lines) and measured (solid lines) travel curves for
circuit breaker opening; (a) no fault, (b) reduced hydraulic flow cross-section.
personnel. The specific list of features used in this test im-
plementation is presented in Table I. They can be derived
easily from the measured quantities. In addition, most of the C. Example
required sensors are standard equipment of new circuit breaker
monitoring devices or even the circuit breakers them selves. The process of condition assessment by model-aided diag-
A simple user interface was also created in MATLAB, which nosis shall be illustrated by an example. Consider the model
presents the most likely diagnosis together with a confidence segment depicted in Fig. 4, which generates the position of the
value contains list of features (Fig. 3). This list contains the main piston at simulation time. Parameters directly influencing
default, measured and estimated values of each feature, thus its operation are, among others, the initial position and the hy-
making the diagnostic process more transparent to the users. draulic damping.
590 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 15, NO. 2, APRIL 2000

During the preparation part of MAiD, simulations were con- upper half of the window gives the circuit breaker operation
ducted for both closing and opening the circuit breaker, with (open) and the connection status (disconnected from the net-
three different values for the damping parameter, which has a work). The remaining fields are the diagnosis results. Here it
default value of 1. Simulation of a normal opening operation has been found that the hydraulic damping of the main piston
from initial position = 1 (fully closed) resulted in the travel curve was increased from its default value of 1 to approximately 5,
shown by the dotted line (a) in Fig. 5. From this curve, the fea- with a confidence or 35.4%. With the scrollable features list in
tures contained in the Normal column in the user interface the lower part of the window, the user is enabled to follow the
(Fig. 3) were extracted. Another simulation with the hydraulic reasoning that led to this diagnostic decision.
damping parameter set to 5 produced the dotted line (b) with the
features contained in the Matched column. Both sets of fea-
tures were stored in the database. V. CONCLUSIONS AND OUTLOOK
When the real circuit breaker was operated in normal condi-
tion, a travel curve according to the solid line (a) in Fig. 5 was Model-aided diagnosisthe novel strategy for condition di-
recorded. Its features are rather close to those from the no-fault agnosis introduced in this reportis a useful method for con-
simulation. Then the hydraulic flow cross-section for opening dition assessment of technical devices. Its feasibility has been
was reduced by an arbitrary amount. The travel curve from the demonstrated on a real high voltage circuit breaker.
next opening operation is shown by the solid line (b); its features The benefits for the user of model-aided diagnosis include,
are shown in the Measured column. Calculating the distance Determining the location and cause of a problem in a
from all stored simulated cases, the best match was found to be technical system, possibly even before it becomes serious.
the one discussed above (hydraulic damping = 5). Hence, the Thus the time for identifying the problem and with it the
according diagnosis was reported to the user. outage time can be reduced. This also helps to reduce the
costs for operation and maintenance.
D. Practical Results Access to quantities which are not accessible for direct
measurement, inside the system. This provides more de
The parameters of the computer model were manually ad- tailed insight into the overall condition of the equipment.
justed to match the particular circuit breaker under observation. Since model-aided diagnosis does not require high com-
With that, the diagnostic database was created and used on the puting power in each installation it can easily be integrated
high voltage circuit breaker described above. into new or existing monitoring or EMS hardware. In this
The following quantities were measured on the circuit manner, the capabilities of a condition monitoring system
breaker: close and trip coil energization (before the auxiliary can be enhanced significantly at hardly any increase in
contacts), close and trip coil voltage (after the auxiliary con- cost per unit.
tacts), close and trip coil current, linkage travel, motor voltage, Further work on the concept of model-aided diagnosis should
motor current. These quantities were recorded by means of a include: self-adaptation of the model parameters to the specific
PC-based monitoring system and evaluated off-line at a later device to be diagnosed; finding the best methods for pre- and
time. No spring travel sensor was available, hence the spring post-processing of the sensor data; incorporation of trend anal-
position after each operation was estimated and manually ysis and extrapolation with diagnosis; interfacing a model-aided
entered by the operator. It is believed that monitoring this diagnosis system with intelligent EMS.
quantity directly would have further increased the diagnostic
confidence.
Several abnormal operating conditions were applied to the ACKNOWLEDGMENT
circuit breaker, namely,
The authors would especially like to thank Mr. A.C. Carvalho
1) the hydraulic flow cross-section was reduced, causing a
for his support and valuable discussions.
low opening speed (Fig. 5);
2) a differently aligned pump motor switch was used for con-
trolling the motor, resulting in low spring charge;
3) the control voltage for opening or closing the breaker was REFERENCES
changed from 110 to 80 V or 130 V, respectively. [1] A.L.J. Janssen et al., A Summary of the Final Results and Conclusions
At first, one change was made at a time; later, up to three of the Second International Enquiry on the Reliability of High Voltage
changes were applied simultaneously. In every case, the ab- Circuit Breakers, in CIGRE 1994 session, 1994, paper no. 13-202.
[2] J. Reason, Coming soon! Circuit breakers with on-line condition moni-
normal condition (or one of them in case of multiple faults) was toring, McGraw-Hills Electrical World, vol. 209, no. 2, February 1995.
correctly identified from the measured data. Obviously, the con- [3] G.F. Luger and W.A. Stubblefield, Artificial Intelligence: Structures and
fidence values were higher for the single fault cases than for the Strategies for Complex Problem Solving, 3rd ed. Harlow: Addison
Wesley Longman, 1998.
multiple faults. In addition, the confidence was further increased [4] A.A. Hopgood, Knowledge-Based Systems for Engineers and Scien-
by quantizing the simulation results to match the resolution of tists. Boca Raton: CRC Press, 1993.
the measured data. [5] M. Stanek, M. Morari, and K. Frhlich, Model-Aided Diagnosis: An In-
expensive Combination of Model-Based and Case-Based Condition As-
Fig. 3 shows the diagnosis for the fault case of reduced hy- sessment, IEEE Transactions on Systems, Man and Cybernetics, 1999,
draulic flow cross-section (Fig. 5). The first row of fields in the submitted for publication.
STANEK AND FRHLICH: MODEL-AIDED DIAGNOSIS 591

Michael Stanek was born in Vienna, Austria, in Klaus Frhlich was born in 1945 in Salzburg, Aus-
1965. He studied electrical engineering at the Vienna tria. He received a Master of Electrical Engineering
University of Technology, Austria, where he ob- and a Ph.D. degree in technical science from the
tained his Dipl.-Ing. degree in 1993. Presently he is Vienna University of Technology, Austria. After 11
studying for his Ph.D. at the High Voltage Laboratory years in Switchgear and High Voltage Technology
of the Swiss Federal Institute of Technology Zurich. with BBC (later ABB) in Switzerland he became a
His research activities are focused on the intelligent full professor at the Vienna University of Technology
high voltage circuit breaker, in particular controlled in 1990. Since 1997 he has been a full professor
switching and on-line condition monitoring and of High Voltage Technology at the Swiss Federal
diagnosis. In 1998, Michael Stanek received the Institute of Technology Zurich, Switzerland. Klaus
Innovation Award of the Swiss Electrotechnical Frhlich is a senior member of IEEE, a member of
Association for his work on a controlled switching device for high voltage CIGRE Study Committee 13, and the convenor of CIGRE Working Group
circuit breakers. 13.07 (Controlled Switching).

S-ar putea să vă placă și