Cross-Domain Self-Supervised Contrastive Learning with Multi-Scale Feature Fusion for Bearing Fault Diagnosis Under Limited Labels

Deep learning is extensively used in fault diagnosis. Actual implementations frequently encounter the issue of limited labeled data. This paper proposes a cross-domain self-supervised contrastive learning method with multi-scale feature fusion (CD-MSSCL) for bearing fault diagnosis under limited samples. The method first employs a specialized time-domain data augmentation strategy to capture the complexity of industrial vibration signals and enhance model generalization. A designed encoder backbone network (MSF-SEA) combines multi-scale features with a SENet attention mechanism using pyramid fusion. The network performs self-supervised pre-training on unlabeled samples to effectively capture multi-frequency fault features critical for bearing fault diagnosis. Limited labeled samples then fine-tune the model to transfer pre-trained features to specific tasks. Experimental results show CD-MSSCL outperforms traditional deep learning and current contrastive learning methods in accuracy and domain adaptation under limited labels. The approach significantly reduces data collection and labeling costs through effective unsupervised knowledge extraction and transfer.