Cross-condition bearing fault diagnosis under data imbalance using adversarial domain adaptation signal alignment digital twin

Abstract Bearing fault diagnosis plays a key role in industrial equipment maintenance. However, real-world scenarios often involve cross-condition variability and severe data imbalance, which hinder the generalization ability of existing methods. While digital twin (DT) technology can generate supplementary fault data, most existing approaches overlook uncertainty estimation and lack effective alignment mechanisms across varying conditions. To address these issues, we propose a DT method based on adversarial domain adaptation signal alignment. A signal response twin model synthesizes basic fault signals, and a conditional generative adversarial network enhances data diversity and fidelity. To align distributions and time series across conditions, we introduce dynamic weighted maximum mean discrepancy and dynamic time warping. Additionally, a variational autoencoder is introduced to regularize the latent representation during signal generation, while Bayesian inference is used in the diagnosis stage to estimate predictive uncertainty, thereby improving diagnostic reliability. Experiments on two benchmark datasets show that our method achieves over 98% diagnostic accuracy under cross-condition and imbalanced settings, outperforming existing approaches. The results demonstrate the method’s effectiveness in synthetic signal generation, uncertainty modeling, and adaptive alignment.