Optimum fault isolation by statistical inference

Abstract

Fault Isolation is always based on a statistical model of fault occurrence and measurement error. Usually the statistical assumptions are impilcit and unstated. Making them explicit and systematically exploring their consequences proves to be an extremely powerful method of developing optimum fault isolation techniques. A criterion is developed here for optimum isolation of catastrophic faults. The technique evolved serves, in fact, to determine the actual probability of each fault based on the observed values of the measured quantities. These probabilities provide ambiguity statistics for the degree of isolation attained. A method is also developed for optimally employing the measurements to detect and isolate noncatastrophic faults when catastrophic faults are absent. Information theoretic criteria are derived for the optimum selection and sequencing of candidate tests. These can be applied at test time in a refined version of branched "go-no go" logic, and also at test-development time to screen a large set of candidate tests down to a small efficient subset.