Reciprocal temporal-frequency distillation with uncertainty awareness for cross-domain mechanical fault diagnosis

Abstract Domain adaptation (DA) techniques have proven highly effective for cross-domain classification in mechanical fault diagnosis, where operating conditions exhibit substantial variation. DA efficacy fundamentally depends on two critical factors: transferability and discriminability. However, existing methods typically neglect the distinct contributions of temporal and frequency features, applying a uniform alignment strategy that inadvertently damages the discriminability of domain-sensitive frequency features while failing to enhance the discriminability of transferable temporal features. To address this limitation, this paper proposes a Reciprocal Temporal Frequency Distillation with Uncertainty Awareness (RTFD-UA) method. First, a period segmentation frequency feature extractor is designed to enhance the discriminability of fault-related frequency bands, while a temporal convolutional network extracts highly transferable fault features from the temporal domain. Second, a reciprocal temporal frequency distillation mechanism facilitates bidirectional knowledge exchange between temporal and frequency features across source and target domains. Through dynamic role alternation between teacher and student networks, this mechanism enhances temporal feature discriminability in the source domain while improving frequency feature transferability in the target domain. Third, adversarial training is implemented within the temporal frequency fused space to minimize domain discrepancy across individual temporal and frequency features and their inherent interrelations. Finally, an uncertainty awareness mechanism with selective gradient enhancement is incorporated to suppress prediction uncertainty of unlabeled target samples, thereby stabilizing the optimization process. Extensive cross-domain fault diagnosis experiments demonstrate that RTFD-UA outperforms existing DA approaches in classification accuracy and adaptation capability.