Unsupervised graph transfer network with hybrid attention mechanism for fault diagnosis under variable operating conditions

In recent years, unsupervised domain adaptation (UDA) has gained widespread application in addressing intelligent fault diagnosis under variable operating conditions. However, how to effectively model data structure information and integrate it into UDA has hindered the application of intelligent fault diagnosis in the industry. To solve this issue, a multi-scale and multi-structure information-embedded unsupervised graph transfer network for fault diagnosis is proposed. Specifically, a novel node feature extractor is first designed for feature embedding. To better fuse multi-scale information and obtain more effective features, a multi-scale convolutional layer and a hybrid attention module are utilized. Secondly, an adaptive similarity graph-constructing method based on the inner-product kernel is adopted to convert the node features into graph data. Next, the graph neural network (GNN) is introduced to obtain graph-structured information. Finally, a joint domain adaptation module is designed to cope with the covariance drift problem in cross-domain fault diagnosis. The proposed method exhibited state-of-the-art performance in the experiments of three case studies.