A Multi-Scale Cost-Sensitive Contrastive Domain Adaptation Network for Imbalanced Fault Diagnosis

Rotating machinery constitutes the core component of industrial systems, with its operational health directly determining production safety and efficiency. However, traditional methods for rotating machinery fault diagnosis are often inadequate for complex industrial environments due to challenges like cross-condition data discrepancies, label scarcity, sample imbalance, and significant signal noise. To address these limitations, this paper proposes a novel multi-scale cost-sensitive contrastive domain adaptation network. The proposed framework integrates a multi-scale feature extraction module to capture robust time-frequency features, enhancing noise resistance. Furthermore, it combines multi-scale contrastive domain adaptation with cost-sensitive learning to simultaneously mitigate both cross-domain distribution shifts and class imbalance issues. Extensive experiments on the XJTU Rolling Bearing Fault Simulation Dataset demonstrate that this synergistic approach effectively reduces noise interference and the effects of class imbalance. The resulting solution achieves superior diagnostic accuracy and enhanced generalization, offering a practical framework for intelligent predictive maintenance systems.