关键词: 辅助分类生成对抗网络, 类小波变换, 轴承故障诊断, 数据生成

Abstract: Using data generation method to generate high-quality data which made time-domain and frequency-domain features consistent with the real signals of bearing faults, and constructing balanced dataset, were of great significance for the establishment of an efficient diagnostic model of bearing faults in the case of data imbalance. In order to address the limitations of the existing data generation methods, which focused on a single feature in time or frequency domains, WLT-ACGAN was proposed herein. Firstly, a WLT network was constructed with a multi-layer neural network based on the principle of wavelet transform. The wavelet transform and inverse transform were simulated, and the mapping relationship between time-domain signal and frequency-domain signal was established. Secondly, the WLT network was embedded into ACGAN model as the primary component of model generator. Finally, two discriminators were constructed with different functions, enabling the improved ACGAN to learn time-domain and frequency-domain feature information of authentic bearing vibration signals concurrently. Experimental results show that the bearing vibration signals generated by WLT-ACGAN model exhibit consistent time-domain and frequency-domain features with those of the actual bearing vibration signals. Furthermore, the fault diagnostic model constructed with the balanced dataset augmented by the generated signals exhibits a high degree of accuracy when the data are imbalanced.

Key words: auxiliary classifier generative adversarial network（ACGAN）, wavelet-like transform（WLT）, bearing fault diagnosis, data generation