A Bayesian Approach to Efficient Diagnosis of Incipient Faults

Safe, reliable, and efficient operation of complex dynamical systems requires the ability to detect, isolate, and identify degradation in system components. Degradations are typically modeled as incipient faults, which are slow drifts in system parameters over time. This paper presents an efficient approach for the detection, isolation, and identification of incipient faults under uncertainty using a Dynamic Bayesian Network (DBN) approach. Initially a DBN is used as an observer to track nominal system behavior. Once a fault is detected, incipient fault hypotheses are generated using a variation of our qualitative TRANSCEND approach for abrupt fault isolation. A modified DBN that includes the active fault hypotheses is then used to isolate the true fault and estimate the rate of change in its parameter value.