Detection of incipient fault using fuzzy agglomerative clustering algorithm

This paper depicts an adaptive diagnostic system based on a fuzzy pattern recognition approach. The proposed system is designed to operate on-line and to deal with the following characteristics: on-line adaptation of classes, detection of slow or abrupt changes and stabilization in a new state, on-line creation of new classes. To meet these requirements, classes are constructed sequentially with a fuzzy agglomerative clustering procedure. Such a clustering procedure requires only one pass through the data, the fuzzy prototypes are created or adapted as new observations are gathered. A prototype is labelled as it is created by using a k-nearest neighbours rule with a distance reject option. This labelling rule is well adapted for stationary states and abrupt changes. However, this rule does not operate in case of incipient faults. To deal with this limitation, we define the concept of temporary prototype. To decide if this prototype is representative of a stationary state or a transient one we introduce a progressive hypotheses test based on the activation rate of the prototype. The results of a robustness study are presented. Finally, the diagnosis system operation is demonstrated on a simulated example