Domain adaptation (DA) is crucial for cross-domain mechanical fault diagnosis under varying operating conditions. The success of DA relies on two key factors: transferability and discriminability. However, existing methods often neglect the distinct contributions of temporal- and frequency-domain features in vibration signals to these factors. This paper presents a Reciprocal Temporal-Frequency Distillation (RTFD) method that leverages both feature types to enhance cross-domain performance. First, a period-segmentation frequency extractor is designed for discriminative fault-related features while a temporal convolutional network extracts transferable temporal features. Subsequently, a reciprocal distillation mechanism enables knowledge exchange between temporal and frequency features across domains through dynamic teacher-student role alternation. Finally, adversarial training on fused features is conducted to minimize domain gaps in both individual features and their interactions. Experimental results demonstrate that RTFD achieves superior classification accuracy and adaptation capability compared to existing DA methods.

Domain adaptation; Fault diagnosis; Temporal-frequency distillation; Transferability and discriminability.