Fault Signatures, Pattern Recognition, and Classifiers
This page examines fault signatures, pattern recognition, and pattern classification as part of the white paper
A Guide to Fault Detection and Diagnosis.
Pattern recognition is a general approach that directly uses the observed symptoms and compares them to a set of known symptoms for each possible fault, looking for the best match. For instance, an observed combination of several particular high temperatures and high energy consumption may be known to be the symptoms of a known failure mode. These signatures might be derived from first principles models or observed empirically.
We can represent the “pattern”, or “fault signature”, as a vector (1 dimensional array) of symptoms for each defined fault. For qualitative models, this fault signature vector consists of values 0 or 1 (or -1 in a signed case). For the unsigned case, a 1 would typically correspond to an observed symptom of a fault, while a zero would indicate normal behavior. A fault signature of all zeros would represent normal operation. The collection of fault signatures can be organized as a matrix, with each column corresponding to a fault and each row corresponding to a particular symptom.
For quantitative models, the values in the fault signature might be any real number. These “features” might be direct measurements, model residuals, or other transformations of measurement values.
In either the qualitative or quantitative case, a “distance” can be calculated between the observed signature and each signature associated with known failure mode. If there is no exact match, the nearest signature is used if it is close enough (based on a tuning parameter chosen for the distance). For binary-valued fault signatures, a “Hamming distance” would usually be used. This is simply a count of the number of symptoms that are different.
When an algorithm chooses exactly one signature as the best match, it is called a “classifier”, solving the “classification problem”.
In some cases, such as the typical training scenarios for neural networks built as classifiers, the signatures corresponding to known failure modes will not be explicit. Instead, the symptom information is the input to the neural net, and the resulting output is ideally 1 for a particular diagnosed failure and 0 otherwise.
One weakness of many pattern recognition approaches is that they implicitly make a single-fault assumption. If there are two independent faults, usually no single fault signature will match the observed symptoms. This could be addressed by also storing signatures for pairs or triples of failures, but this does not scale well.
Other forms of pattern recognition can be used in event-oriented diagnostics, discussed in later sections.
Case-Based Reasoning (CBR) includes pattern recognition to recognize previous similar “cases” (which could be fault signatures). However, it generally includes this in a larger cycle of learning new cases and generalizing them.
(Return to A Guide to Fault Detection and Diagnosis)
Copyright 2010 - 2013, Greg Stanley