BiST-DA: A Bi-Branch Self-Training Domain Adaptation Network for Bearing Misalignment Diagnosis

Bearing misalignment diagnosis under varying operating conditions is challenging due to severe distribution discrepancies between training and testing data, and traditional data-driven methods perform poorly on unlabeled target domains. This paper proposes a bi-branch self-training domain adaptation network (BiST-DA) for cross-condition bearing misalignment diagnosis, consisting of a bi-branch backbone network, an adversarial adaptation branch, and a self-training branch. The backbone fuses Swin Transformer V2 (global contextual features) and ResNeXt (local discriminative features) for complementary feature extraction; the adversarial branch reduces source-target domain distribution gaps; the self-training branch with pseudo-label filtering refines target-domain predictions iteratively. Experiments show BiST-DA outperforms state-of-the-art domain adaptation methods, verifying its effectiveness and robustness under varying operating conditions.