An Ensemble Domain Adaptation Network With High-Quality Pseudo Labels for Rolling Bearing Fault Diagnosis

The unsupervised domain adaptation (UDA) technique is aimed at minimizing the distribution discrepancy of interdomain (DDID) and, thus, bears immense potential for label scarcity in the target domain. Currently, most UDA methods frequently utilize pseudo labels to measure the distribution of target domain features. However, the level of pseudo labels in the reported studies is always modest, thus restricting the performance of UDA. To address the aforementioned deficiency, we develop an ensemble domain adaptation network (EDAN). It leverages ensemble learning (EL) to generate high-accuracy pseudo labels and couples domain adaptation (DA) as well as EL to guarantee pseudo labels’ robustness. Specifically, an EL network (ELN) with high generalization to the target domain is constructed based on multiple multiscale convolutional neural networks (CNNs) and a self-reinforce soft voting mechanism. Moreover, we dynamically couple ELN and the weighted balance distribution adaptation (WBDA) to upgrade EDAN’s classification stability, as well as the applicability of each CNN to the target domain. Twelve cross-condition fault diagnosis tasks and seven cross-device fault diagnosis tasks are designed based on four open-access rolling bearing datasets to verify the effectiveness of the proposed method. The research results indicate that EDAN outperforms five related approaches.