An Imbalanced Fault Diagnosis Method for Rolling Bearing Based on Semi-Supervised Conditional Generative Adversarial Network With Spectral Normalization

In actual industrial applications, rolling bearings are under normal working conditions most of the time, and the fault data that can be collected are insufficient, so they are prone to data imbalance. Due to the high cost of labeling all fault data, most fault data are unlabeled. In this study, the Semi-supervised Conditional Generative Adversarial Network with Spectral Normalization (SN-SSCGAN) is proposed to solve these problems. Its core idea is to generate new samples with similar distribution by using partially labeled minority fault samples to balance the dataset. First, this method first applies wavelet transform to preprocess a vibration signal and obtain a time-frequency matrix. Second, the partially labeled time-frequency fault data are taken as the input of SN-SSCGAN, Nash equilibrium is achieved through adversarial training, and then data with similar distribution are generated. Lastly, the generated fault data are added to the dataset for balancing, and a convolutional neural network is used for fault diagnosis. The effectiveness of the proposed method is verified with comparative experiments in the CWRU bearing dataset. Results show that this method can generate high-quality samples and determine satisfactory results in bearing fault diagnosis when only a small number of labeled samples and the remaining unlabeled samples are used.

Automated summary

The study proposes a method using a Semi-supervised Conditional Generative Adversarial Network to address data imbalance in rolling bearings. By generating new samples with similar distribution, the dataset is effectively balanced, and satisfactory results are achieved in bearing fault diagnosis.