ScholarWorks@Y-Scholar Hub

초록

Modern industrial systems require reliable fault diagnosis under various operating conditions, with diagnosis models often encountering feature scarcity, class imbalances, and changes in distribution between domains. This paper introduces WALKMAN, a unified framework designed to improve diagnostic performance under these constraints. The proposed approach incorporates wavelet-based feature extraction to address sample scarcity, a class-aware large-margin softmax that adjusts decision boundaries based on class distribution, and conditional weighted adversarial training to reduce discrepancies in both covariate and label distributions. Collectively, these components enable the learning of discriminative and generalizable representations even in severely imbalanced and shifted environments. Experiments on two bearing fault datasets from Jiangnan University and Machinery Failure Prevention Technology find that the proposed method outperforms existing unsupervised domain adaptation techniques, particularly in scenarios with long-tailed distributions. The proposed framework not only improves the accuracy of fault classification but also enhances robustness against minority class suppression and negative transfer. Due to its improved accuracy, stability, and adaptability, the WALKMAN framework represents a practical solution for condition monitoring in real-world industrial environments.