Abstract Fault diagnosis techniques are essential for preventing equipment failures, reducing maintenance costs, and enhancing operational efficiency by promptly identifying anomalies. The widespread deployment of industrial sensors has significantly increased the availability of machinery data, facilitating extensive research in data-driven fault diagnosis. However, real-world datasets frequently exhibit label scarcity and severe class imbalance, where fault instances are substantially fewer than normal samples. To address these challenges, this study proposes a robust unsupervised domain adaptation framework that synthesizes fault signals by interpolating real healthy samples with domain-specific knowledge. Although this synthetic augmentation effectively expands training data, the resulting distribution often deviates from actual fault scenarios, limiting model generalizability. To alleviate this domain discrepancy, our framework incorporates Conditional Domain-Adversarial Networks (CDAN) for domain-invariant feature extraction, complemented by structured pseudo-labeling to assign reliable predictions to unlabeled target samples. Subsequently, a Locality Preserving Projection (LPP) module constructs a shared latent space to achieve both domain alignment and enhanced class discrimination. Experimental evaluations conducted on a synthetic dataset derived from the CWRU bearing benchmark demonstrate that the proposed method achieves accuracies of 91.10% under imbalanced conditions and 84.65% in balanced scenarios, surpassing current state-of-the-art methods by 12.87% and 3.57%, respectively. Ablation studies further underscore the significant contribution of structured pseudo-labeling to the overall performance, confirming the proposed approach’s efficacy and robustness in real-world unsupervised industrial fault diagnosis tasks.
Liu et al. (Wed,) studied this question.