Self-supervised cross-domain open-set fault diagnosis method based on graph learning under time-varying speed conditions

When conducting cross-condition transfer diagnosis of rotating machinery in industrial scenarios, novel fault types that are unseen during training may emerge, leading to the open-set fault diagnosis problem. Meanwhile, the operating speed of equipment frequently varies over time, causing domain shifts in monitoring signals and resulting in blurred feature boundaries among different fault categories. However, existing open-set diagnosis studies predominantly focus on constant speed conditions, neglecting the feature entanglement and transfer challenges introduced by time-varying speeds, thereby limiting their practical applicability. To address this issue, this paper proposes a graph learning-based self-supervised open-set threshold adaptive method for cross-domain open-set fault diagnosis of rotating machinery under time-varying speeds. Specifically, graph representation learning is introduced to mine key latent fault information in time-varying signals. A shared-class prototype guided self-supervised contrastive learning approach is employed to enhance sample-level feature discriminability. Using pseudo-label sequences combined with a confidence threshold adaptive mechanism, decision boundaries between shared and unknown classes are dynamically adjusted. Additionally, a shared-class subdomain alignment strategy is integrated to enhance the model’s cross-domain generalization ability. Cross-domain open-set diagnosis tasks are designed on two time-varying speed fault datasets for comparative experiments. Experimental results validate that the proposed method achieves superior fault recognition accuracy and generalization performance, providing a reliable solution for the health state identification of rotating machinery in industrial scenarios.