Data Augmentation Fault Diagnosis Method Based on Residual Mixed Self-Attention for Rolling Bearings Under Imbalanced Samples
Jiuyuan Huo, Chenbo Qi, Chaojie Li, N.C. Wang
Abstract
The imbalanced data in the collected samples affects the generalization performance and the accuracy of the fault diagnosis model due to the low frequency and short duration of industrial bearing failures in the actual production. In this study, an industrial bearings fault diagnosis technique under class-imbalance based on residual mixed self-attention Wasserstein conditional generative adversarial network and one-dimensional convolutional neural network (RMA-WCGAN-1DCNN) is proposed. To begin, the RMA mechanism is proposed to extract the time-domain and frequency-domain features of one-dimensional time-series vibration signal. Second, the RMA-WCGAN fits the data distribution without data preprocessing to generate specified class of high-quality samples to balance the dataset. Finally, the RMA-1DCNN is trained for fault diagnosis. By comparing nine sampling methods and eleven classification models under various conditions, it is demonstrated that the generated samples of the RMA-WCGAN-CNN method proposed in this paper have validity and reliability, significantly improve the classification performance of the fault diagnosis model in the case of imbalanced datasets. When the RMA-WCGAN and RMA-1DCNN are combined, results for high-precision fault diagnosis can still be obtained despite high class-imbalanced rates.