An individual generalization framework based on independent samples towards a more reasonable fault diagnosis benchmark

Domain offset is an inevitable phenomenon in industrial signals for fault diagnosis. This article discusses a neglected problem of the traditional evaluation benchmark for data-driven fault diagnosis, i.e., the presence of identical individual information in both the training and testing sets. An individual generalization framework is explored using independent test individuals towards a more reasonable fault diagnosis benchmark. Furthermore, an improved lightweight transformer is applied to enhance the dynamic global feature extraction and irrelevant information filtering. Comprehensive experiments are performed on the Paderborn University bearing dataset and a machine-level motor dataset collected from real production lines. The results show that the traditional benchmark cannot effectively evaluate the screening ability for fault-irrelevant features and the generalization ability for new individuals. The proposed lightweight transformer achieves the highest generalization performance with great application potential.