Model Completeness Assumptions - Novel Faults, Missing Links
This page examines the effect of incomplete models, such as faults or cause/effect links not included in the models, as part of the white paper
A Guide to Fault Detection and Diagnosis.
It is difficult to model or even imagine every possible fault scenario. A fault that arises that was not considered during application development is called a “novel fault”. Similarly, there may be other missing parts of the model, such as a missing cause/effect link.
Quite often, novel faults will not interfere with fault detection based on models of normal operations. Those models don’t include faults anyway. If any fault is serious enough, it will probably affect some aspect of the “normal operation” model, so that a fault will be detected even if it isn’t modeled.
The problem with novel faults comes in during fault isolation. Once the system detects a problem, it will try to assign root causes. If the system picks any of the defined faults, it will be wrong. If the system is based on a single fault assumption (such as those based on Bayes rule or most pattern matching techniques), the mismatch between predicted symptoms and observed symptoms will usually be large. For techniques that can diagnose multiple faults, the mismatch might not appear as severe, because there are often faults affecting just a few symptoms that can always be blamed. Multiple failures will probably be reported, and might even exactly support the observed symptoms. But the diagnosis will still be wrong.
The ideal diagnostic system should recognize when a novel fault or missing link is present, since none of the faults can explain the observed symptoms. Pattern matching techniques can be sensitive to novel faults, but the problem can be recognized: generally, if the observed pattern vector isn’t close to the signature of some known fault, the system can provide a warning.
Conventional (backpropagation) neural nets used as pattern matching classifiers are especially susceptible to the novel fault problem, and may not be able to provide any warning. However, related techniques such as RBFN (Radial Basis Function Networks), come with their own built-in error analysis, and can recognize when the known faults can’t explain the observed symptoms.
Copyright 2010 - 2013, Greg Stanley
(Return to A Guide to Fault Detection and Diagnosis)