Few‐shot multiscene fault diagnosis of rolling bearing under compound variable working conditions

Abstract

Abstract As one of the most widely used rotating machinery components, whether the bearing can operate stably is related to the reliability of the equipment and the safety of the staff. Therefore, efficient and accurate intelligent fault diagnosis (IFD) technology is necessary for modern industrial equipment. Bearing fault diagnosis based on deep learning methods has made great progress in recent years. However, most methods rely heavily on massive data and the domain shift phenomenon caused by the high‐level compound variable working conditions would greatly affect the performance of the model. To solve the data sparsity and domain shift problem simultaneously, an effective feature disentanglement and restitution (FDR) few‐shot method is proposed for IFD under multiple scenes. First, the vibration signals are preprocessed and input into the metric‐based neural network. The model is trained based on the meta‐learning method to extract task‐level general features to alleviate the data sparsity problem. Then, the FDR method extracted task‐related features from the information discarded by the convolution kernel at different scales and fused them with the output features of the embedding module to reconstruct task‐specific features and alleviate the phenomenon of domain shift. Finally, the relational module automatically extracts the nonlinear relations between features and classifies them. A number of high‐level compound variable working condition tasks were constructed on two experimental platforms, and the fault diagnosis tests with small samples and multiple scenes were carried out. The results show that our proposed method has superior accuracy and transferability under compound variable working conditions.