Probabilistic fault diagnosis based on incomplete training data with application to an automotive engine

In this paper a Bayesian approach is used for fault diagnosis of complex systems. The posterior probabilities that different faults are present, given observations from the system, are computed. Generally in previous methods for Bayesian fault diagnosis, a model of the system under diagnosis is estimated. In the present paper a step further is taken, and it is shown that it is not necessary to estimate any model. Instead the diagnosis is based on training data directly. The Bayesian method allows for handling of incomplete training data, i.e. where only a small amount of, or even no training data is available from some fault cases. This is important in the diagnosis application, since there are faults which are uncommon or even unknown, and from which it is difficult to obtain training data. The Bayesian method also makes it possible to combine information from training data with prior knowledge of the kind used in many previous diagnosis algorithms. To illustrate the Bayesian diagnosis method it is applied to on-line diagnosis of an automotive diesel engine, using discrete observations and data from real driving situations. In this application study, there are faults that cannot be diagnosed with only prior knowledge, and the importance of combining it with training data is shown. Further it is shown that in the case of discrete observations, the method results in simple computations because of an efficient representation of the probabilities.