A novel multi-consistency adversarial transfer diagnosis method for gearboxes refined with pseudo-labels

Abstract Intelligent fault diagnosis methods based on deep learning and transfers learning have been widely applied in the transfer fault diagnosis of rotating machinery. Semi-supervised learning, which leverages a large amount of unlabeled data, faces challenges when existing algorithms primarily focus on domain shift caused by data distribution during domain adaptation, potentially resulting in poor performance when cross-domain label distribution differences exist. Optimizing model parameters using high-threshold pseudo-labels can alleviate label imbalance, but incorrect pseudo-labels may lead to model over-reliance and misalignment. To address this issue, this paper proposes a novel multi-consistency adversarial transfer diagnosis method for gearboxes, refined with pseudo-labeling. Firstly, feature alignment between source and target domain data is achieved both globally and locally using information entropy and partial domain adversarial methods. Secondly, multiple prediction consistency is applied to select high-confidence pseudo-labels from target instances under various data transformations. Finally, the model is trained with a small amount of labeled data from the target domain to achieve cross-condition transfer fault diagnosis of gearboxes. Experiments conducted on gearbox datasets under various operating conditions validate the effectiveness and superiority of the proposed method.