Imbalanced Bearing Fault Diagnosis Based on RFH-GAN and PSA-DRSN

Bearings in actual working environments typically operate in healthy conditions, resulting in an imbalance in the data collected data. The majority of the collected data are related to bearings in healthy conditions, with insufficient data related to faults. This imbalance leads to accuracy and stability issues in deep learning models used for diagnosis purposes. To address this issue, we propose employing a residual factorized hierarchical search-based generative adversarial network (RFH-GAN) and a residual shrinkage network with pyramidal squeezed attention (PSA-DRSN) for unbalanced fault diagnosis. The process involves transforming vibration signals collected from bearings into time-frequency (TF) domain images through the utilization of the continuous wavelet transform (CWT). The enhanced RFH-GAN generates synthetic fault samples with authentic characteristics, while the PSA-DRSN performs fault diagnosis. The experimental findings substantiate that our method improves the quality of the generated samples, mitigates the data imbalance issues that are inherent in conventional diagnosis methods, and attains heightened precision and efficacy in fault diagnosis tasks.