Detecting and Learning Unknown Fault States in Hybrid Diagnosis

Discrete diagnosis systems have traditionally included the ability to model unknown faults. However, this ability has not been available in hybrid diagnosis approaches. This is because one weakness of the standard approaches to tracking the hybrid state—Kalman filters and particle filters—is that it is hard to tell when they are performing poorly. In this paper we show that by using a particle filter with look-ahead we can measure in a principled way when the filter is performing badly and hence the system may be in an unmodelled state. We show that this approach can be used to detect unknown faults in a small realistic domain. In addition, we show that for a common class of faults we can learn a model of the unknown state so as to predict it better in the future. This is one step on the way to being able to enhance an existing hybrid model by learning unmodelled states and adding them to the model.