A Conditional Adaption Alignment Dynamic Graph Neural Network model for unsupervised fault diagnosis of rotating machinery

In rotating machinery, the diverse operating conditions and cross-machine discrepancies lead to significant distribution shifts in monitoring data, with label scarcity persisting under new or specific conditions. These challenges motivate the application of domain adaptation techniques in cross-domain fault diagnosis. In recent years, unsupervised fault diagnosis methods based on graph neural networks still face several limitations. The geometric structure of data distributions is often neglected by existing conditional distribution adaptation methods. To address this problem, a Conditional Adaptive Alignment Dynamic Graph Neural Network (CA-DGNN) cross-domain unsupervised model is proposed. First, a Dynamic Graph Neural Network integrated with Gaussian edge features is constructed, where inter-sample correlations are learned through dynamic topological structures and embedded into graph-level fault feature representations. Subsequently, the relationship between graph-level fault features and labels is explicitly established through the Conditional Maximum Mean Discrepancy (CMMD), which is formulated within the Reproducing Kernel Hilbert Space (RKHS) using the conditional covariance operator. The CMMD is then used to measure the domain discrepancy of feature-conditional distributions, and a conditional adaptive loss is designed to realize the domain alignment. The intra-class knowledge transfer is enhanced compared to traditional marginal alignment. Additionally, mutual information between fault features and predicted labels is utilized to extract discriminative information and improve the reliability of pseudo-labels. The proposed method is evaluated through experiments on two varying operational condition cases and one cross-machine case, with the results demonstrating that the model is more effective than other models in cross-domain fault diagnosis tasks. The codes of CA-DGNN model are released at: https://github.com/Pear-so/CA-DGNN.