Abstract: To solve the problems of insufficient fault-samples and diagnosis-knowledge, and according to the merit of Support Vector Machines (SVM) that can be trained with small-sample, a SVM based unsupervised clustering model is presented. By modifying the decision-function of One-Class Support Vector Machine (1-SVM), which has the ability to find outliers from a dataset without any class of information but rarely is applied to pattern-recognition for its algorithm limits, a Decision-Improved 1-SVM (1-DISVM) is formed. Based on it, multi-pattern training and classing method is designed, then an unsupervised clustering model is constructed. The simulation and diagnostic experiment results of a helicopter’s gearbox show that this clustering model can not only recognize the unknown fault patterns adaptively and precisely, but also learn the nature of the input-patterns from small samples and diagnose the faults successfully.

Key words: fault diagnosis, clustering, support vector machine, unsupervised learning