A Fault Diagnosis Method of Rolling Bearing Under Variable Working Conditions Based on Multi-Source Domain Adaptation

In the prognostics and health management (PHM) of rolling bearing, domain adaptation serves as a reliable fault diagnosis method issues under variable working conditions. As opposed to single source domain, gathering signals from multiple source domains allows for the extraction of copious feature information. Data bias occurs in the multi-source domain situation not only between target and source domains but also among several source domains. It can interfere with each other to combine all source domain data during learning. Therefore, a multi-source domain adaptation fault diagnosis method is proposed for complex working conditions. First, variational mode decomposition (VMD) is applied to obtain intrinsic mode functions (IMFs), and feature extraction is performed on the IMFs and original samples to pick up the fault feature set. Second, it maps each group of source and target domains into a group-particular subspace and reduces their distribution difference. Finally, multiple fault diagnosis sub-models are built based on different subspace data, and they are optimized by improving the accuracy while reducing the bias of the predicted results among sub-models. Experiment results prove the effect of model on multi-source domain tasks.