Intelligent fault diagnosis using an unsupervised sparse feature learning method

Abstract

Feature learning is an integral part of intelligent fault diagnosis. Sparse feature learning methods have been shown to be effective in learning discriminative features. To learn features with optimal sparsity distribution, an unsupervised sparse feature learning method called variant sparse filtering is developed. Variant sparse filtering uses a sparsity parameter to determine the optimal sparse feature distribution. A three-stage fault diagnosis method based on variant sparse filtering is then developed to identify rotating machinery faults. The method is validated using a rolling bearing dataset and a planetary gearbox dataset and is compared with other diagnosis methods. The results show that the developed diagnosis method can identify single faults and compound faults with high accuracy.