A Global-Local Dynamic Adversarial Network for Intelligent Fault Diagnosis of Spindle Bearing

Abstract

Transfer learning-based intelligent fault diagnosis methods for smart spindle bearings have been constantly developed in the recent years. The existing methods generally either assume different domains belong to the same label spaces or the number of fault categories in source and target domains are equal. Nevertheless, this assumption is unrealistic since the unknown fault class will unexpectedly occur when changing working condition. To solve this problem, a global-local dynamic adversarial network is proposed for unexpectedly fault detection of smart spindle bearing, in which global and local data alignment are introduced and the relative proportion of two distributions are dynamically calculated to extract domain-invariant features. In addition, new fault classifier is designed to separate unknown fault from known fault. Experimental on a smart spindle bearing dataset demonstrate the proposed method is promising for new fault detection.