Self-supervised learning-based dual-classifier domain adaptation model for rolling bearings cross-domain fault diagnosis

Abstract

The dual-classifier domain adaptation methods concentrate on the output consistency between two classifiers, while ignoring the classification determinacy of each classifier and reducing the fault identification capability of the model. To address this challenge, a self-supervised learning-based dual-classifier domain adaptation model (SLDDA) is presented for cross-domain fault diagnosis of bearings. Firstly, a dual-classifier classification determinacy metric is formulated to alleviate the output ambiguity between classifiers, which simultaneously considers the joint determinacy between two classifiers along with the individual determinacy of each classifier. Secondly, a self-supervised learning approach based on the clustering of target data is proposed to extract abundant target features from the target data. Furthermore, two validation experiments are conducted on bearing datasets of Paderborn University (PU) and H0205. Comparing with the Maximum Classifier Discrepancy method, the proposed SLDDA improves the diagnostic accuracy by an average of 14 % of the PU dataset, which efficiently fulfills the cross-domain fault diagnosis of rolling bearings under variable operating conditions.