A bearing fault diagnosis method based on MCDCGAN and DPTAN under data imbalance
TL;DR
A diagnostic framework that combines generative-adversarial data augmentation with dual-branch time–frequency representation learning to improve feature quality and fault classification and achieves better performance than baseline models and maintains robustness under data imbalance and noisy conditions is proposed.
Abstract
Bearings are critical components of rotating machinery, yet reliable fault diagnosis remains challenging under complex conditions due to signal non-stationarity and data scarcity. To address these issues, this paper proposes a diagnostic framework that combines generative-adversarial data augmentation with dual-branch time–frequency representation learning to improve feature quality and fault classification. Firstly, the bearing vibration signals are transformed into two-dimensional time-frequency representations by utilizing the continuous wavelet transform. Subsequently, a multi-attention conditional deep convolutional generative adversarial network (MCDCGAN) is employed for conditional augmentation under class imbalance, integrating attention mechanisms and stabilization strategies to generate more reliable samples for minority fault classes. Finally, a dual-branch parallel time-frequency attention network (DPTAN) is designed to jointly learn temporal and spectral feature representations and then fuse them for fault classification. Experimental results on the CWRU and HIT datasets demonstrate that the proposed method achieves better performance than baseline models and maintains robustness under data imbalance and noisy conditions.