Abstract
Convolutional Neural Networks, with their excellent capabilities for automatic feature discrimination and learning, have been widely applied in the field of mechanical fault diagnosis. However, in real-world operating environments, acquiring large amounts of fault data as training samples is often challenging, which limits the applicability of traditional methods. To address this issue, this study proposes a frequency-adaptive fault diagnosis method for high-speed motors under small-sample scenarios. Specifically, this paper designs an innovative data augmentation technique that effectively expands the diversity and coverage of the training dataset and is seamlessly integrated into the fault diagnosis model. Furthermore, to enhance the richness of feature representations and strengthen information exchange between different feature channels, this paper proposes a frequency-adaptive convolutional layer (SCNET), which significantly optimizes the performance of Bidirectional Gated Recurrent Units (BiGRU) in fault feature extraction. Based on these technological improvements, we have constructed an efficient intelligent fault diagnosis model named RS-SCBiGRU. Experimental validation shows that, compared to various advanced fault diagnosis methods, the RS-SCBiGRU model achieves a significant improvement in accuracy and demonstrates stronger noise resistance capabilities.