A Hilbert–Schmidt Independence Criterion based self-supervised method for motor non-destructive fault detection and diagnosis

Scarcity of samples and labelled data and frequent variations in operating conditions have long constrained the performance and stability of the non-destructive fault detection and diagnosis model. Under this background, how to overcome these challenges without affecting the structure and performance of the monitored motor has become a key research direction in non-destructive motor fault detection and diagnosis. In this work, a self-supervised method is developed, which exploits both vibration data from vibration sensors and current data from Hall sensors to conduct non-destructive monitoring and diagnosis. First, an ensemble feature encoder is constructed for multimodal data with comprehensive multi-scale feature extraction and effective modelling of temporal dependencies. Subsequently, an alignment loss based on the Hilbert–Schmidt Independence Criterion is designed. This loss achieves alignment between heterogeneous data by maximising the distributional dependence of features across modalities at the batch level, and when integrated with the cross-correlation loss, it enables multi-modal feature consistency alignment and discriminative enhancement during the pre-training stage. Finally, the effectiveness of the proposed framework is validated through three different device cases. The results show that the model achieves over 95.13% with 1 sample per class in intra-domain tasks, fully demonstrating its robustness, generalisation capability and broad applicability.