A federated cross-machine diagnostic framework for machine-level motors with extreme label shortage

Abstract

A start-up company is usually only able to collect normal samples, resulting in extreme label shortage and inability to establish effective intelligent diagnostic models. Especially for machine-level motors with more expensive experimental labeling, cooperation from multiple partners is often required. The phenomenon of signal global domain shift caused by inherent assembly differences and dynamic testing environments becomes more pronounced in scenarios from multiple sources. Furthermore, due to data security and privacy concerns, it is not advisable to directly share local data from multiple companies. This article proposes a federated cross-machine diagnostic framework (FedCMD) to address these challenges, which not only considers the security of various companies but also evaluates the generalization ability of machine-level motors on new individuals. A multiscale pruning transformer (MSPT) is developed with three strategies to provide an effective solution. Comprehensive experiments and visual analyses are performed on real production line signals, which demonstrates the superiority of the proposed model with highest generalization performance of 97.42% mean cross machine accuracy (CMA).