Dual-Contrastive Multiview Graph Attention Network for Industrial Fault Diagnosis Under Domain and Label Shift

Recently, domain generalization (DG) methods have been actively researched for the complex industrial fault diagnosis, which aims to learn generalized representations from historical working conditions to build a diagnosis model that can perform well on unseen working conditions. However, these methods ignore the interactions between monitoring variables, which may fail to learn the feature representation with topological structure in non-Euclidean space. In addition, these methods assume the same label distribution across historical and unseen working conditions, which is generally challenging in practice, as the probability of faults varies across different working conditions. This label shift problem can negatively impact the generalization performance. To address these issues, a novel dual-contrastive multiview graph attention network (DMGAT) is proposed in this article. Specifically, a multiview graph attention network (GAT) is designed to explore the intrinsic topological structure of the data, which learns an optimal graph structure that best serves DG by integrating both graph learning and graph convolution in a unified network architecture. In addition, a novel dual-weighted contrastive learning strategy is developed. The intradomain contrastive learning facilitates the extraction of expressive node features, while interdomain contrastive learning simultaneously considers the alignment and separation of semantic probability distributions to extract shared feature representations for multiple source domains under both domain and label shifts. Furthermore, a label sampling probability is used to weight the interdomain contrastive loss and the source domain classification loss, to encourage the model to learn from minor classes in fault diagnosis. Experiments on two cases demonstrate the superiority of the proposed method.