A Comprehensive Framework for Model Based Diagnosis

This paper reviews our work on monitoring, prediction, and fault isolation methods for complex dynamic systems affected by abrupt faults. The key to this work has been our ability to model the transient behavior in response to these faults in a qualitative framework, where the predicted transient effects of hypothesized faults are captured in the form of signatures that specify future behavior for the fault with higher order time-derivatives. The dynamic effects of faults are analyzed by a progressive monitoring scheme that avoids direct measurement of second and higher order derivatives from real signal values, which are unreliable when signals are noisy. However, generating qualitative characteristics from real time varying signals is still a challenging problem. More recently, we have been investigating statistical techniques for reliable change detection and labeling in noisy signals. We discuss some of these techniques, and study the tradeoff between speed of detection and sensitivity of analysis. The integrated framework for monitoring, prediction, and fault isolation is being tested on real data generated from an automobile engine test bed