Multi-Modal Self-Supervised Learning for Cross-Domain One-Shot Bearing Fault Diagnosis

Deep learning methods have achieved impressive results in bearing fault diagnosis, but they typically require large amounts of labeled source domain data. This can be time-consuming and expensive to collect and annotate. Moreover, fault data are difficult to collect in some practical applications, and limited data fail to train complex deep fault diagnosis models. To address these challenges, a novel multi-modal self-supervised learning cross-domain one-shot bearing fault diagnosis method is proposed in this paper. This method uses unlabeled vibro and acoustic data collected under the same working conditions and faulty state to pre-train a model. Benefiting from the complementary nature of vibration and acoustic data, the model can learn inheritable and robust fault features from unlabeled data. The pre-trained model is then applied to one-shot fault diagnosis in different working conditions. Experimental results demonstrate the effectiveness of the proposed method, achieving competitive performance compared to state-of-the-art techniques.