ALARM ANALYSIS
• When a fault occurs on the network, automatic
protection and isolation occurs at a number of
substations, opening switches to ensure that power is
no longer fed to the faulty line, and that the
disturbance on the remainder of the grid is effectively
minimised.
• Details on which of the switches have operated are
passed on to the control centre which continually
polls round the different substations to see what
switchgear information each has.
• No direct information is obtained on what type of
fault has occurred and where, but rather information is
obtained in the control centre which indicates which
switches and automatic protection devices have
operated in response to the fault.
• The switching messages, along with further analogue
information, can be regarded as 'symptoms', whose cause can
be inferred from the overall message pattern.
• Unfortunately interpretation of the messages depends on such
variables as network topology, the current network state, and
which intertrips are in service.
• All of these go together to make the analysis of alarms and
subsequent diagnosis of faults a non-trivial exercise.
• It is the job of a control engineer, based in the control centre, to
infer the cause of a fault, from the string of messages received.
• Normally for a one off fault, a control engineer can make a
relatively swift and accurate diagnosis; however there are
times, such as stormy weather conditions, when alarm activity
is very high and this stretches and sometimes even exceeds
the human ability to cope with the sheer volume of information.
• These can severely reduce the speed of diagnosis and hence
the overall network efficiency.
• In these circumstances there is an apparent need for an
automatic, computer based system which can be used to
assist the control engineer.
• On simple faults, with which the control engineer can cope, the
computer-based system can act as a supporting, confirming
device or as a suggestive mechanism to speed up the
diagnosis.
• Where difficulties occur, however, it may easily become a front
line tool, firing out suggestions to the engineer on request.
• A real-time alarm handling and fault analysis (AHFA) expert
system is used, this being seen as more than simply an alarm
processor.
• AHFA is able to perform its diagnoses directly in real-time
without recourse either to situation specific heuristics or to
model-based 'generate and test'.
• The AHFA expert system was first developed in 1990 and was
scaled up to demonstrate complete area control two years
later.
• An adaptive form of AHFA based on a learning
classifier system (LCS), and this uses a fault
simulator to direct the entire fault diagnosis.
• Unlike a conventional expert system, rules are not
programmed into the LCS, although that is quite
possible.
• On the contrary, the LCS is trained by example. The
diagnostics are then adapted through genetic
algorithms (GAs), which directly operate on the rule
strings.
• From an input alarm stream, the LCS produces a
fault diagnosis along with a list of switching
information which supports the solution.
• This chapter is intended to give an overview of the
problem area of alarm analysis for power system
transmission, along with a couple of approaches to help
tackle the problems using computer-based technology.
• Firstly, time stamps given to pieces of alarm data are not
directly indicative of the actual time of occurrence, in that
one switch could have operated before another one in real-
time; however, due to sub-station polling, they can be
reported to the control centre in a distorted chronological
order.
• Secondly, when serious faults occur, switching messages
can arrive at the rate of 600 messages/minute.
• Numerous switchgear will operate due to routine
switching, because of spurious transients or due to
switchgear malfunction, whether or not a fault occurred.
Problem area and requirements
• At the present time the problem area is essentially one of assisting
the control engineer with alarm analysis and acceptance.
• An important sub-area of this is dealing with alarms associated with
switchgear operation, this including circuit breakers and isolators.
• At present when switchgear open/close telemetry arrives at the
control centre, the control engineer must accept these as individual
alarms, one for each open or close operation.
• The immediate need is for an on-line aid to group together such
alarms where they are all associated with the same event.
• The engineer must accept each alarm, and in doing so the system is
then aware that that alarm has been noted by the control engineer.
• AHFA, which involves fault diagnosis as well as alarm processing,
was employed live for the first time in a control room environment,
in 1993. At that time it was realised that the basic principles
underlying AHFA could be generalised to also allow for the
recognition of different types of routine switching.
• The display needs to be concise and to convey
sufficient information to allow the control engineer to
judge the situation for himself, based partly on
information outside the realistic scope of AHFA, e.g.
telephone conversations with staff at remote sites.
The immediate aim was for an engineer aid rather than
a replacement.
• A more worrying problem was that even a well
designed and proficient fault diagnosis tool might
occasionally misinterpret or ignore particularly unusual
fault situations.
• Inherent in the design is therefore a requirement that
when in trouble, perhaps due to bad data, the
diagnosis system clearly indicates that it has given up,
possibly with a brief explanation as to why.
• Because of this it is necessary for the system to make
as much use of available information as possible,
including such information as time sequences, as well
as heuristics.
• A further requirement of any system of this type is
to be easily extendible to include new types of
fault situations, as and when they arise, through
either knowledge acquisition or on-line experience.
• Overall the trend has been to incorporate fault
diagnosis and alarm processing as part of the
data and display infrastructure rather than as a
stand-alone tool.
• This naturally necessitates a general
infrastructure for machine readable data,
information and knowledge.
Fault diagnosis difficulties
• Occasionally a fault may develop on the grid, typically a power
line affected by a lightning strike or a short circuit to earth.
• In this situation, automatic protection arrangements will, if
operating correctly, open the circuit breakers nearest to the
fault so as to isolate it from the system and interrupt the fault
current.
• The most desirable property of only those circuit breakers
which are nearest to the fault opening is known as
discrimination.
• In the control centre, the engineer will see alarms relevant to
individual circuit breakers when a fault occurs. For example,
opening a circuit breaker at the Exeter 400KV substation would
result in the following message:
• 14/05/96 13.43.02 EXET4 X405 OPEN
• If there are many number of breakers are operated then the
fault diagnosis problem requires that characterisation of the
switching is based primarily on the sequence of events
occurring immediately following the onset of fault conditions.
• The situation is complicated due to the polling, on top of which
some data arrives late and is not time-stamped at source.
AHFA
• AHFA was initially designed as a stand alone tool
for fault diagnosis, the main idea being to
significantly reduce the amount of information
displayed to the control engineer.
• In simple terms the mission brief was for AHFA to
display a single line of text — the fault diagnosis,
with the intention that this would replace several
lines of low-level alarm information.
• In practice, after accepting switching information,
the one line summary is then displayed in an
interactive window, from where reports in greater
detail are available.
• A number of common drawbacks appear with standard expert
system fault diagnosis tools.
• The first is the timelag involved in waiting for all the messages
to arrive before a hypothesis is formed.
• Secondly, multiple faults can cause the need to rework the
hypothesis in that the system needs to detect and indicate that
a multiple fault scenario is in place.
• A final drawback is the timing problem caused by inadequate
knowledge of the exact sequence of messages.
• When AHFA was initially being put together, one of the first
jobs was to discuss their role with senior control staff, in the
form of review meetings and interviews as well as obtaining
example diagnoses.
• One firm observation was the engineer's approach of an
incremental diagnosis.
• Given one piece of switching information, such as a particular
circuit breaker opening, the control engineer became ready — it
could be routine or it could be a fault.
• On receipt of a second such message, from elsewhere, the
possibility of a fault would become more likely and so on.
• The main object to be focused on was deemed to be an Islet
which contains the faulty component and which is bounded by
circuit breakers or open disconnectors.
• The components under consideration are typically
transformers, bus-bars and power lines such that a fault on any
one of these will cause the circuit breakers to trip.
• In a perfect world, the overall pattern of switching provides a
unique fault signature.
• Failure of one circuit breaker would lead to circuit breakers
tripping in an associated, linked Islet, thereby providing back-up
protection but complicating the fault diagnosis.
• In this way AHFA can actually look ahead and suggest a
particular fault, before all of the supporting evidence has
appeared.
• Each Islet is roughly speaking equivalent to a
protection zone, in that it is a group of neighbouring
plant items, in an area of the grid network, bounded
by the most immediate circuit breakers.
• When all of these circuit breakers are open, the Islet
is electrically separated from the rest of the network.
• When a 'circuit breaker open' alarm or a 'protection
operated' alarm is received, this acts as a trigger
event to establish a monitoring process on the
associated areas of the network where there might
be a fault.
• For example, in the case of a circuit breaker open
alarm then a pair of Islet monitors are formed, one
for each side of the circuit breaker.
• If further circuit breakers open around a given Islet,
so the confidence of a fault being in that Islet is
increased.
• Each Islet monitor thus checks the expected
behaviour of associated circuit breakers and
switches, raising or lowering its confidence
dependent on the sequence which actually occurs.
• The beauty of this approach is that when multiple
faults, in different Islets, occur, so each Islet monitor
can carry out its own diagnosis in a completely
independent way.
• In total AHFA is made up of two main parts, firstly a
model of the transmission network and secondly a
diagnosis operative. This has all been implemented
using Smalltalk-80, an object oriented programming
language.
• The network model indicates the present state of
play as far as network topology is concerned, with
isolation, circuit breakers and their protection
systems all included.
• Once a diagnosis has been started, due possibly to a circuit
breaker opening, each Islet monitor is responsible for its own
diagnosis, although communication between monitors is quite
possible to resolve shared events.
• Received switching messages are only routed to the relevant
Islet monitors and on checking, if an Islet monitor decides that
it has something of interest for the operator, then this can be
reported.
• Intermediate stages of diagnosis are not usually reported to
the operator; however once an Islet monitor has signalled that
it does have something of interest, it can incrementally report
on this such that the operator is kept updated.
• This minimum reporting threshold is quite useful as it avoids
reporting on such events as routine switching operations.
AHFA diagnoses: an example
• This example serves to demonstrate how an AHFA diagnosis is
carried out and one of a number of ways in which the results
can be displayed.
• Suppose three different but concurrent faults have occurred
on the network, as follows:
(i) Persistent (failed reclose) fault on Power Line EXET4/AXMI4,
Exeter to Axminister.
(ii) Permanent fault on SuperGrid Transformer AXMI4-SGT1.
(iii) Transient (successful reclose) Fault on Power Line
CHIC4/EXET4, Chickerell to Exeter.
• It can take from one to two minutes for switching associated
with a persistent or permanent fault to complete and hence
for all the relevant data to arrive at the control centre.
• On the other hand data from a transient fault, i.e. where a
successful reclosure has occurred, usually takes about 15
seconds.
• The actual total time taken for AHFA to diagnose a fault,
given no time delays and a block of relevant data, is 2
seconds.
• Given that the diagnosis is carried out incrementally as
pieces of information are received, the actual time to
indicate a fault correctly, from the time of receipt of the
last relevant data item, is negligible.
• Stage 1
• The first message to arrive is:
'CHIC4 EXET4 CCT MAIN PROT OPERATED'
• This message means that the CCT (circuit), which is
the main power line from Chickerell to Exeter, is faulted,
although it does not show the type of fault, i.e.
persistent or transient.
• The first diagnosis from AHFA reflects this uncertainty:
— Nothing: PROT (but. SWI) -> CHIC4/EXET4
which indicates that a protection message has been
recorded implicating CHIC4/EXET4 but AHFA awaits
expected switching alarms and hence nothing further
can be said at this stage about the type of fault.
• Stage 2: After a while, when more messages
have arrived, the original diagnosis is modified
to:
— prob. Fault: PROT & prob. SWI -> CHIC4/EXET4
• At this time AHFA has also become aware that
there are possibly two concurrent faults, and this
is reported as:
— prob. Fault: prob. SWI -> AXMI4-MC1? EXET4/
AXMI4? AXMI4-SGT2?
• AHFA effectively has an indication that a fault is
likely, but more data is needed in order to nail
down the fault type and/or location.
Stage 3:
• The original diagnosis is now largely complete, sufficient
messages having been received. At this point AHFA can
classify the fault type, and thus shows:
— Transient: PROT & SWI -> CHIC4/EXET4
• which was the third of the faults which have actually occurred.
Also, what was considered to be a probable fault at Stage 2,
has now received further confirmation, thereby moving it up to
a possible persistent fault and reducing the number of
potential causes to two:
— poss. persistent: poss. SWI -> AXMI4-MC1? EXET4/AXMI4?
• A third fault has, at this point, also been identified:
— prob. Permanent: prob. SWI (no. PROT) -> AXMI4-SGT1?
• which is a probable permanent fault, although an expected
message indicating 'Protection operated' has not yet been
received.
• There is however only one component within the Islet which
could have caused this particular fault, namely AXMI4-SGT1.
Stage 4:
• Due to the further messages that have arrived there are now
two definite and one probable diagnoses. These are indicated
as:
— Transient: PROT & SWI -> CHIC4/EXET4
— prob Persistent: prob. SWI -^ EXET4/AXMI4
— Permanent: SWI (no. PROT) -> AXMI4-SGT1?
• the transient fault on CHIC4/EXET4 is no longer indicated by a
dark region, the fault having cleared.
• Because both relevant disconnectors have opened, there is
complete confidence that SGI has faulted.
• Conversely, the situation is less clear for the Islet on the left of
AXMI4-X120. The fault type is certainly Persistent and it is
between the circuit breakers.
• There are, though, two components which could have caused
this fault type, namely the power line EXET4/AXMI4 and the
mesh-corner AXMI4-MC1.
• At this point both of the options are possible, although the
power line is favourite as one of its disconnectors has opened.
Stage 5:
• This is the final stage in the diagnosis. At this point all Islet
monitors have either observed all the expected messages or
they have timed-out, which is typically 2 minutes.
• The diagnoses are now fixed, as is indicated by * ; however
question marks are retained where no confirmatory protection
messages were received.
• *Transient: PROT & SWI -> CHIC4/EXET4
* Persistent: SWI -> EXET4/AXMI4
* Permanent: SWI (no. PROT) -> AXMI4-SGT1?
• In this particular example, AHFA was required to deal with 3
close concurrent faults.
• In practice it has been found that AHFA can actually deal with
at least 20 concurrent faults, should that prove necessary.
Adaptive alarm processing
• An alternative to AHFA is described, making use of an adaptive
on-line diagnostic method which automatically diagnoses
faults from switchgear indications.
• The system uses a fault simulator to provide corrective
information for a learning classifier system (LCS), as shown in
below fig.
• The diagnostics can either be initially programmed in, as is the
case with AHFA, or can be learned on-line, through experience.
• Whichever initialisation method is selected, subsequent
adaptation is obtained through a genetic algorithm procedure.
• A number of adaptive alarm processors have, in the past, been
based on a neural network approach
• Several adaptive systems which retain a meaningful structure
are hybrids between newer learning schemes and more
classical artificial intelligence
• An LCS is itself also a hybrid in that a genetic
algorithm is combined with a classical expert
system.
• The method uses switchgear messages, just as
AHFA does, to excite a fault diagnosis system;
however it also employs a genetic algorithm
approach to dynamically update the association
between switching information and fault hypotheses.
• Here emphasis is placed on the learning and
adaptive features of the system, as provided by the
genetic algorithm.
Adaptive system for fault diagnosis on the electrical power transmission
network
