Filter-Match-Interact Transfer Framework for Machineries Open-Set Fault Diagnosis

Abstract

The available labeled samples are scarce when high-end machineries work in different operating-load conditions. There are often new faults present when conducting transfer fault diagnosis, leading to performance degradation. How to accurately identify them under dynamic load-variable conditions is a more challenging issue. Therefore, the filter-match-interact transfer framework (FMI-TF) is proposed, which consists of three interactive networks. Open samples filtering: learn the known–unknown samples classification hyperplane by designing the progressive filtering discriminator, achieving target samples distraction and progressive outliers filtering. Weighted auxiliary matching: align domain distributions and tighten known–unknown samples boundary through the entropy-modified weighted matching mechanism, the auxiliary distracting classifier, and the high-confidence negative probabilities of unknown samples. Interactive refinement rectification: mine and cultivate information interaction and coupling within two networks by improving the differentiated interactive updating module, and achieving positive network transfer. The FMI-TF has been validated on different mechanical testbeds.