Unsupervised Clustering for Fault Diagnosis in Nuclear Power Plant Components

The development of empirical classification models for fault diagnosis usually requires a process of training based on a set of examples. In practice, data collected during plant operation contain signals measured in faulty conditions, but they are 'unlabeled', i.e., the indication of the type of fault is usually not available. Then, the objective of the present work is to develop a methodology for the identification of transients of similar characteristics, under the conjecture that faults of the same type lead to similar behavior in the measured signals. The proposed methodology is based on the combined use of Haar wavelet transform, fuzzy similarity, spectral clustering and the Fuzzy C-Means algorithm. A procedure for interpreting the fault cause originating the similar transients is proposed, based on the identification of prototypical behaviors. Its performance is tested with respect to an artificial case study and then applied on transients originated by different faults in the pressurizer of a nuclear power reactor.