Model extraction for fault isolation

This paper presents a simulation-based approach for fault isolation in complex dynamic systems. A machine learning technique is used to extract, from simulated data, models representing regularities in system behavior. A heuristic based on the degree of coverage of the model on the data is then applied to isolate faults. To test tolerance to incomplete models, our simulation model only requires i/o functions of relevant system processes that can be observed. We view our approach as an incremental filtering process which is useful for diagnosis of large scale systems. To illustrate the approach, we describe experiments in two examples including a well known three-tank system. Preliminary results show that, on the average, different types of faults at different locations such as a leaked tank and a blocked pipe can be isolated effectively more than 99% at a time. Results are promising but more in-depth study is required.