An open domain fault diagnosis method for rotating machinery based on dynamic convolutional graph network and bi-classifier adversarial learning

Abstract To address the challenges of distributional discrepancies and partially inconsistent class labels between source domain and target domain data in real-world engineering applications, this study proposes an open-domain adaptation method based on a Dynamic Convolutional Graph Network and bi-classifier adversarial learning (DCGNDAT). First, a dynamic convolution module is introduced to replace conventional convolution layers, enhancing the model’s ability to extract fault-related features from vibration signals. Second, a graph network is employed to encode structural information, thereby improving the model’s representation of complex data. Finally, a bi-classifier adversarial training mechanism is designed, incorporating entropy maximization and minimization strategies for the source and target domains, along with a binary cross-entropy scheme for target domain outputs, to accurately delineate the boundaries between known and unknown classes. Fault diagnosis experiments conducted on bearing and self-priming centrifugal pump datasets demonstrate that the proposed method achieves significantly higher recognition rates for unknown fault categories compared to existing algorithms, validating its superior diagnostic performance and robust open-domain adaptability.