A Cross-Source Domain Contrastive Learning-Guided Invariant Adversarial Network for Mechanical Fault Diagnosis Under Unseen Conditions

Although domain adaptation methods are commonly used in mechanical fault diagnosis to mitigate the domain shift problem, they typically rely on target domain data being available during training. To address this issue, this paper proposes a cross-source domain contrastive learning-guided invariant adversarial network (CSDCL-IAN). The model first employs a residual attention network to construct a feature extractor, aiming to enhance discriminative fault information. Subsequently, a cross-source domain contrastive learning mechanism is designed, which extracts common features across multi-source domains by making intra-class features closer and inter-class features more distinct. Finally, unseen-condition data are input into the trained CSDCL-IAN to realize cross-domain fault diagnosis. In transfer diagnosis experiments on a planetary transmission system test rig, CSDCL-IAN yields an average diagnostic accuracy of 98.15% on across six transfer tasks, which significantly verifies its superior domain generalization ability and cross-domain diagnostic performance.