A cross-attention based contrastive learning method for bearing fault diagnosis under limited labels

Deep learning-based methods have demonstrated significant success in the field of bearing fault diagnosis. However, these methods hinge on access to substantial volumes of labeled data for training, but in practical applications, labeled fault data are often difficult to obtain, making it difficult for traditional deep learning methods to be useful. To address this challenge, this paper proposes a self-supervised method for bearing fault diagnosis based on cross-attention mechanism and contrastive learning. The method employs abundant unlabeled data for pre-training, maximizes the agreement between positive sample pairs through temporal and similarity contrast and learns the fault feature representation of the vibration signals from them for application to downstream tasks, and then achieves good performance in the fine-tuning stage using only a minimal amount of labeled data for fine-tuning. In addition, a sample dynamic weight adjustment strategy is introduced to enhance the representation quality and learning efficacy of contrastive learning. Through a series of experiments on the CWRU public dataset, the results show that the proposed method achieves high diagnostic accuracy and outperforms existing methods in the presence of limited labeled data availability.