Two fault classification methods for large systems when available data are limited

In this paper, we consider the problem of fault diagnosis for a system with many possible fault types. Two approaches are presented that are useful for initial diagnosis of system-wide faults, assuming that no data are available before commissioning the system but the possibility of the occurrence of each symptom is known for each fault. The first method uses a fault tree approach to reduce the solution space before applying the geometric classification method, the assumption being that no unwanted symptoms are possible. This method is nonparametric and thus does not require any data to estimate the underlying distribution of faults and symptoms. The second method is based on the Bayes classification approach to utilize the subjective information and the limited data that may be available. The two methods are generic and applicable to a variety of industrial processes.