TSN: A novel intelligent fault diagnosis method for bearing with small samples under variable working conditions

Traditional deep learning methods rely on big data heavily, which makes bearing fault diagnosis with small samples under variable working conditions a tricky problem. The extremely tough data status that only few samples are available renders methods of tradition deep learning unworkable. In this paper, we propose a novel method called transferable Siamese network (TSN) to solve this problem. TSN can fully utilize the small samples, explore the similarities and differences between samples, achieving the maximum utilization of existing samples. In TSN, Siamese structure is constructed to solve the problem of insufficient samples in the way of data matching, and to solve the problem of variable working condition by transferring initial weights. The feature extraction network is the main executor of TSN, and a deep network for feature extraction including one-dimensional convolutional network, residual structure, and improved attention mechanism is constructed. Verification results from two related datasets demonstrate that the proposed method is effective and feasible, and its diagnostic accuracy is superior to some existing methods of generative adversarial networks. The proposed method provides a promising solution for bearing fault diagnosis under tough data circumstances. The application of this method is beneficial to ensure the reliability and safety of industrial equipment.

Automated summary

Traditional deep learning methods face challenges in bearing fault diagnosis with small samples and variable working conditions. In this paper, a novel method called TSN is proposed, which utilizes Siamese structure and a feature extraction network to address the issues of insufficient samples and variable working conditions. Experimental results demonstrate the effectiveness of the proposed method compared to existing methods.