CSWGAN-GP: A new method for bearing fault diagnosis under imbalanced condition

Most intelligent bearing fault diagnosis methods are conducted with balanced datasets, which is not in line with the reality of industry. Suffering from this problem, intelligent methods are prone to misclassify minority data as majority class. So, it is difficult to develop a robust method to conduct bearing fault diagnosis with imbalanced datasets. Therefore, this work develops a new diagnosis method namely cosine similarity-based self-attention Wasserstein generative adversarial network with gradient penalty (CSWGAN-GP). First, original sampled data are preprocessed to facilitate subsequent analysis. Then, transformed samples are input to the CSWGAN-GP to generate new samples. Penalty terms based on cosine similarity are utilized to constrain the optimization objective. The utilization of self-attention mechanism increases the ability to obtain interested fault features. Finally, bearing fault diagnosis is performed on the new dataset which is rebalanced with the generated samples. The proposed method is assessed through a wide range of metrics and compared with other state-of-the-art methods. From the experiment results, it can be concluded that the CSWGAN-GP presents encouraging performance on the bearing fault diagnosis under imbalanced condition.

Automated summary

This paper proposes a cosine similarity-based self-attention Wasserstein generative adversarial network method to solve the bearing fault diagnosis problem in imbalanced datasets, where most intelligent methods tend to misclassify minority data. The proposed method preprocesses the original data, applies CSWGAN-GP to generate new samples, and performs bearing fault diagnosis on the rebalanced dataset.