Cross-Domain Fault Diagnosis of Rotating Machinery Based on Class-Imbalanced Joint Domain Adaptation Network

Abstract

The widespread prevalence of data distribution differences between different operating conditions of rotating machinery seriously affects the performance of fault diagnosis methods. Domain adaptation technology can improve the generalization performance of models by aligning different distributions in the latent space. However, when there is class imbalance in the target domain, it will cause serious interference to the domain adaptation effect. To solve this issue, a class-imbalanced joint domain adaptation network (CIJDAN) is proposed to tackle the cross-domain diagnosis challenge under class imbalance conditions. First, the predictive information is fully leveraged to extract more discriminative features during domain adversarial training, while an augmented balancing strategy is adopted to alleviate the negative effects caused by the class imbalance. Additionally, by combining the prior knowledge of the source domain label and the distribution estimation and confidence measure of the target domain prediction, a boundary amplifier based on dual-weight subdomain distribution alignment is designed to help the classifier establish clearer and more reliable decision boundaries. As a result, the CIJDAN method combines domain adversarial training and distance measurement to eliminate domain bias and extract reliable and robust domain-invariant features. The CIJDAN method is validated in two cases, demonstrating its feasibility and reliability. Experimental results indicate that CIJDAN achieves high precision and robust diagnostic performance in various class-imbalanced cross-domain scenarios, highlighting its potential for practical engineering applications.

Keywords: Rotating machinery, Fault Diagnosis, Class Imbalance, Domain adaptation, subdomain distribution alignment.