Toward Instance Relationship Learning: A Progressive Domain Fusion Network for Transfer Fault Diagnosis

Intelligent fault diagnosis (IFD) is crucial for effective health monitoring and maintenance of mechanical systems. Unsupervised domain adaptation has been widely applied to IFD by learning domain-invariant fault features to address distribution shifts. However, existing methods primarily focus on reducing inter-domain differences at a macro level, overlooking the inherent diversity and similarity of instances. This oversight can lead to negative transfer and unclear predictions of ambiguous fault samples near the decision boundary. To overcome these issues and enhance diagnostic performance, we propose a novel progressive domain fusion network (PDFN) with synergizing instance diversity and similarity from an instance relationship learning view. Specifically, a progressive domain fusion mechanism is developed to construct fusionable subdomains by exploiting instance diversity. These subdomains are progressively fused with the source domain for adaptive training, effectively narrowing the inter-domain distribution gap. Additionally, we design an instance similarity-relationship insight to construct a similarity matrix, capturing the affinity between source and target instances. The adverse effects of ambiguous samples could be alleviated by an improved multi-instance contrastive loss based on the similarity matrix. Ultimately, extensive experiments were conducted under varying working conditions on both a widely-used open-source dataset and a self-collected dataset from a practical diagnosis platform, with each experiment repeated five times to ensure the reliability of the results. The final results reveal that PDFN achieved accuracies of 96.27% and 94.72% on the two datasets, respectively, demonstrating superior performance compared to state-of-the-art IFD models.