Unsupervised feature learning with reconstruction sparse filtering for intelligent fault diagnosis of rotating machinery

Sparse filtering (SF), as a recently emerging unsupervised feature learning method, has drawn much attention in intelligent fault diagnosis of rotating machinery. Generally, SF implements feature extraction by the trained basis vectors. However, similar features may be extracted by SF due to the lack of effective restrictions on the basis vectors during training, which leads to an adverse effect on the diagnostic performance. To address this drawback, reconstruction sparse filtering (RSF) is proposed based on SF, which explicitly constrains the basis vectors via a soft-reconstruction penalty (SRP). In particular, SRP enables RSF to learn a group of independent basis vectors so as to extract dissimilar and diverse features. These features contain comprehensive and rich fault information that can precisely describe the rotating machinery health conditions, so RSF can perform significantly better. Based on RSF, an intelligent diagnosis method is developed, and it is evaluated through experiments on a gear and two bearing datasets. The results testify that RSF is able to extract dissimilar features from the vibration signals, and it has better feature learning ability than SF and nine other popular unsupervised feature learning methods. Moreover, the superiority of the developed diagnosis method is verified by comparing with several state-of-the-art intelligent diagnosis methods on two famous bearing datasets. © 2021