Multi-subspace self-attention siamese networks for fault diagnosis with limited data.

It has always been an important issue to diagnose mechanical equipment faults with limited training data. Specifically for the problem of bearing fault diagnosis, a multi-subspace self-attention siamese network (MSSASN) is designed for fault diagnosis with limited training data. In MSSASN, multi-subspace self-attention block is developed to assign higher weights to the fault-related information during learning. Particularly, input features are divided into multiple sub-paths in the channel dimension, and the spatial attention features are calculated separately on sub-path and then merged. In this way, the cross-channel information can be effectively learned, while multi-scale feature learning is carried out. Finally, contrastive learning is carried out on the fault features of different samples using siamese networks to deal with the problem of limited training samples. The proposed method is verified by the vibration dataset collected from the three-phase asynchronous motor experiment platform in Zhejiang University of Technology. The results show that the proposed method can identify rolling bearing faults more accurately with limited training data. [ABSTRACT FROM AUTHOR]