A Multi-task Learning Method for Few-Shot Fault Diagnosis Based on Metric Learning

As a critical component of engineering machinery, the performance and reliability of rolling bearings directly affect the operational status of the machinery. In recent years, intelligent fault diagnosis has achieved significant advancements and gained increasing attention in the field of bearing fault diagnosis due to its powerful feature learning capability. Existing intelligent diagnosis approaches commonly assume an abundant amount of fault samples. However, in practice, the lack of fault samples has turned out to be a thorny problem. Therefore, to address the problem of poor fault diagnosis performance caused by inadequate data or the emergence of new fault types, this paper proposes a multi-task architecture based on metric learning. Multi-task learning includes similarity metric and fault classification. Firstly, a novel Siamese network framework is designed, which can reduce the demand of data volume, and allow for effective learning of fault characteristics in few-shot scenarios. Secondly, by employing parameter sharing, a single-branch network with a classifier is constructed to perform the final classification. Experimental results from different datasets validate the effectiveness of the proposed method in few-shot scenarios and unknown class samples.