Adversarial subdomain adaptation method based on multi-scale features for bearing fault diagnosis

Due to the variable working environment of bearings, collected data often follow different probability distributions. It is hard directly use trained models identify bearing fault with operating conditions. In addition, it a high cost label samples for every work condition. To solve these problems, multi-scale adversarial subdomain adaptation diagnosis method proposed, which based on Continuous Wavelet Transform (CWT) and our constructed Multi-scale Adversarial SubDomain Adaptation Network (MASDAN). Firstly, extract features non-stationary signals frequency bands, CWT used convert continuous vibration into two-dimensional time-frequency images. Secondly, enhance correlation across ConvNeXt adds module obtain scales. Finally, reduce distribution discrepancy between source domain target avoid feature confusion in domains, adaptive alignment network introducing information constructed. Two modules are included: Domain Alignment Classification Module (DACNM) Multi-kernel Local Maximum Mean Discrepancy (MK-LMMD) (DANM) discrimination. Thus, MASDAN consists module, DACNM, DANM, can realize extraction adaptively align at domains. The experimental results Qingdao University Technology (QUT) dataset Case Western Reserve (CWRU) demonstrate proposed effectively diagnose faults under