Improved flagging for pattern classifying diagnostic systems

Fault detection and isolation (diagnosis) is based on residual generation and residual analysis. The model-based approach flags the residuals through thresholding, to isolate the effect of faults from noise, and performs diagnosis by mapping the residuals to a residual space with prespecified fault signatures. The main problem with this approach is that thresholds are not always able to differentiate between the effect of faults and noise, so this approach suffers from false alarms, undetected faults, and misdiagnosis. As an alternative to prespecified fault signatures and to cope with their variability, the use of pattern classification techniques has been proposed. However, since tile fault signatures established by these classifiers are formed irrespective of diagnosability, this approach is also prone to misdiagnosis. In this paper the authors demonstrate the application of a flagging unit that enhances the quality of fault signatures. This unit, which relies on a training set to tune its parameters, is shown to improve detection, reduce the number of false alarms and enhance diagnostics