Cross-sensor contrastive learning-based pre-training for machinery fault diagnosis under sample-limited conditions

Recently, data-driven approaches have been extensively used in fault diagnosis. However, most existing methods are based on single-sensor fault data, which is hard to suit for complex industrial systems. Extracting complementary fault features from multi-sensor monitoring data is imperative, especially under limited labeled samples. Inspired by the success of self-supervised learning in handling unlabeled data, we propose a cross-sensor contrastive learning-based pre-training method for machinery fault diagnosis under sample-limited conditions. In the initial pre-training phase, we introduce an innovative cross-sensor contrastive framework to capture complementary features among different sensors for enhancing the acquisition of discriminative fault features. Then, in the fine-tuning phase, a novel cross-sensor interactive attention is designed for effective feature fusion to provide a more robust feature representation. The proposed method is validated on three benchmark datasets, demonstrating superior diagnostic performance under limited labeled samples and well-adapted to different working conditions.