Fault diagnosis of rotating machinery across operating conditions based on capsule graph attention transfer network

Under complex operating conditions, the fault features of rotating machinery in vibration-based non-destructive testing (NDT) often exhibit significant distribution drift, while the target domain usually lacks labelled samples. These challenges are common in practical NDT scenarios, including weak fault signals, variable operating states, and environmental noise, which limit the effectiveness of traditional diagnosis methods. To address these issues, this paper proposes the Capsule-Attentive Graph Network (CAGNet), an unsupervised deep transfer model specifically designed for vibration-based NDT applications. The model aims to improve fault recognition accuracy and transfer robustness under unlabelled target-domain conditions. The model is synergistically optimised at three levels: feature extraction, cross-domain distribution alignment, and target-domain training regulation. A structurally collaborative multimodal feature extractor is constructed by integrating the local spatial modelling capability of the capsule network, the sample relationship representation capability of the graph convolutional network, and the channel selection capability of multi-head self-attention, achieving deep fusion of local and structural information. Rényi divergence is introduced as a metric for domain alignment to enhance generalisation in boundary regions. An exponentially weighted pseudo-label regulation mechanism dynamically adjusts the contribution of target-domain pseudo-labels during training, effectively mitigating initial pseudo-label errors and improving convergence and stability. Extensive experiments on WT planetary gearbox, Huazhong University of Science and Technology (HUST), and Southeast University (SEU) bearing datasets, comprising 26 cross-working condition tasks, demonstrate that CAGNet outperforms existing methods, achieving average accuracies of 99.50%, 96.34%, and 99.69%, respectively. Feature visualisation reveals that CAGNet effectively compresses the source-target domain feature gap, resulting in clearer inter-class separation and inter-domain alignment. Overall, this study provides a robust and generalisable framework for vibration-based NDT under varying operating conditions, offering enhanced defect identification capability without requiring labelled target-domain data and demonstrating strong potential for practical industrial applications.