Statistically-aligned feature augmentation for robust unsupervised domain adaptation in industrial fault diagnosis

Abstract

Effective Unsupervised Domain Adaptation (UDA) remains challenging due to significant distributional discrepancies between labeled source and unlabeled target domains. In industrial fault diagnosis tasks specifically, these discrepancies severely impair model generalization when applied across heterogeneous operating conditions. To mitigate this problem, we propose a novel framework termed Statistically-Aligned Feature Augmentation for Domain Adaptation (SAFA-DA), which addresses domain disparities through statistically-driven feature augmentation strategies. SAFA-DA progressively constructs an intermediate representation by dynamically selecting samples with high prediction confidence from both source and target domains. Utilizing statistical insights derived from this intermediate domain, the framework employs mean-based feature alignment and covariance-based feature augmentation to iteratively align the source domain distribution toward the target domain. Furthermore, SAFA-DA incorporates an adaptive control mechanism based on the Maximum Mean Discrepancy metric, effectively moderating augmentation intensity to ensure stable convergence and prevent overfitting. Extensive experiments on the Case Western Reserve University and the Paderborn University datasets demonstrate that SAFA-DA significantly outperforms existing state-of-the-art methods, achieving an average accuracy improvement of 18.11%. Importantly, SAFA-DA exhibits notable robustness under realistic industrial conditions, consistently maintaining high accuracy despite severe noise interference and class imbalance, underscoring its practical utility in industrial informatics applications.