A Novel Clustering Based on Consensus Knowledge for Cross-Domain Fault Diagnoses

Abstract

In recent years, cross-domain fault diagnosis problems based on knowledge transfer have attracted considerable attention from researchers, some of whom have adopted domain adaptation algorithms for model transfer under various working conditions. Such algorithms typically assume the samples of the target and source domains share the same fault mode sets, and the source and target domains to be interrelated. However, such prior knowledge is difficult to obtain in actual industrial fault diagnoses. To address this issue, we propose a universal domain adaptation network model based on domain consensus clustering (DCUAN). This model adopts cluster matching and domain consensus scores to mine effective knowledge from the class and sample levels, thus eliminating the dependence on prior knowledge of the label set. Furthermore, we introduced a contrastive domain discrepancy to jointly optimize clustering. Experimental results on two diagnostic datasets demonstrate the proposed DCUAN method to be robust and effective, it yields higher performance than existing state-of-the-art cross-domain fault diagnosis methods.