Measurement Selection and Diagnosability of Complex Physical Systems

Abstract

Functional redundancy techniques in diagnosis rely on models of a system to infer deviating system variables without directly measuring them. To avoid an overwhelming or excessive number of observations and unnecessary cost, it is important to establish a minimal number of required observations to achieve a certain degree of diagnosability. Complete diagnosability allows for unique isolation of faults. In general, fault effects propagate to affect multiple system variables, but different faults impose different temporal signatures on the measurements. However, faults may cascade, or propagating effects often trigger compensating mechanisms which may mask the original fault signature. These aggregate behaviors make it hard to reliably determine the initial fault. For dynamic, continuous systems, it is important to quickly isolate faults by using more observations, thus reducing the chance for unique signatures to be masked by other effects. The trade-off between lesser observations and quick fault isolation is investigated in this work, and a measurement selection algorithm is presented to derive a minimal set of measurements to achieve complete diagnosability.