SFUGDA: Source-free unsupervised multiscale graph domain adaptation network with privacy-preserving for cross-domain fault diagnosis of offshore wind turbines

Unsupervised Domain Adaptation (UDA) has achieved tremendous success in the task of solving new unlabeled target domains by leveraging its knowledge learned from labeled source datasets and has been widely utilized in wind turbine fault diagnosis. However, in the actual industrial scenarios, data privacy, commercial confidentiality, and transmission efficiency constraints make the source domain data inaccessible. In addition, internal structured information modeling and multi-structured fusion of wind turbine data under non-stationary operating conditions is not sufficiently taken into account. To solve the aforementioned problems, a privacy-preserving source-free unsupervised multiscale graph domain adaptation network (SFUGDA) is proposed. Specifically, the proposed SFUGDA consists of two main stages, the pre-training stage of the source diagnostic model and the adaptation stage of the target domain. In the pre-training stage, a novel multi-scale multi-structured network with hybrid attention mechanism is implemented, which can effectively fuse deep and shallow features to extract multi-scale node information. Meanwhile, the node information is further fused with the topology information to capture robust, structured, and discriminative feature. In the domain adaptation stage, we consider novel loss functions and constrain the target domain through neighborhood clustering, regularization, and predictive diversity for self-training to achieve high-precision fusion and clustering, obtaining a diagnostic model for the final target domain. To verify its effectiveness and superiority, we evaluate SFUGDA in a variety of experiments including comparison and ablation experiments about gears and bearings of the wind turbine under variable operating conditions, especially time-varying operating conditions. Experimental results indicate that SFUGDA yields state-of-the-art results among multiple advanced comparison methods.