Graph structure learning guided multi-source-free domain adaptation for mechanical fault diagnosis

Recently, multi-source-free domain adaptation (MSFDA) methods have been proposed to address the privacy concerns in cross-domain fault diagnosis. However, existing methods often overlook the structural relationships among samples and the varying transferability of source models during target domain adaptation. To address these issues, this paper proposes a graph structure learning guided MSFDA method (GSFDA), which constructs a KNNGraph and employs a graph neural network to capture and propagate latent structure relationships. In addition, an unsupervised fusion strategy based on dynamic weights is designed to optimally combine multi-source domain models, and an adaptive self-training mechanism is developed enables effective fine-tune using unlabeled target data without accessing source samples. Experiments on eight transfer tasks across rolling bearing and planetary gearbox datasets demonstrate that GSFDA achieves an average diagnostic accuracy of 98.84 %, outperforming multi-source domain adaptation methods that require source data access, and effectively improves the generalization ability and robustness of cross-domain fault diagnosis.