Machine Cross-Domain Fault Diagnosis Using A Dual-Dynamic Domain Adaptation Network

Accurate fault diagnosis in unlabeled mechanical scenarios is a critical and urgent practical requirement. Unsupervised domain adaptation has emerged as a dominant approach to address this challenge. However, most existing methods rely on static learning mechanisms and cannot adapt to variations in target tasks, which limits their diagnostic performance. To overcome this limitation, a dual-dynamic domain adaptation network (DDAN) is proposed. The dual-dynamic framework fuses a dynamic multi-mode and a dynamic multi-source domain. Specifically, a dynamic multi-mode domain adaptation network is built to dynamically adjust training modes during optimization, enabling the model to gradually improve domain-adaptive diagnostic capabilities by creating task-oriented and precise diagnostic decision boundaries. In addition, a dynamic multi-source domain adaptation loss is designed to adaptively align knowledge from multiple source domains, creating an effective cooperative diagnostic mechanism that can be dynamically adjusted across source domains. The effectiveness of the proposed DDAN is validated on an aero-engine fault dataset, where an average diagnostic accuracy of 98.58% is achieved across multiple unsupervised cross-domain diagnostic tasks, indicating its strong diagnostic capability.