Entropy-Embedded Partial Domain Adaption Network for Digital Twin-Enhanced Rolling Bearing Fault Diagnosis

Accurate and reliable bearing fault diagnosis is critical for ensuring the operational reliability of industrial machinery, yet it remains a challenge due to scarce labeled fault data and prevalent unknown health conditions. To enhance diagnostic reliability under such uncertainties, this article proposes a digital twin (DT) augmented framework that integrates entropy-embedded partial domain adaptation. The approach constructs a high-fidelity virtual bearing model to reliably simulate fault dynamics and generate comprehensive fault signatures. By systematically applying entropy-based criteria to evaluate the uncertainty and complexity of features, the framework extracts and purifies the most informative and reliable characteristics between the physical and virtual domains. These purified features are aligned through a weighted block-diagonal structure, effectively mitigating domain shift caused by unknown health states and improving the reliability of knowledge transfer. This process ensures robust cross-domain diagnostics while minimizing negative interference from outlier conditions. Experimental validation across multiple bearing datasets confirms that the proposed framework reliably transfers diagnostic knowledge from virtual to physical systems, maintaining high fault identification accuracy even with completely unlabeled measurements. By leveraging entropy-based domain adaptation within a DT environment, this work achieves reliable fault diagnosis. Notably, the proposed approach maintains high efficacy despite the lack of annotated data and the presence of unexpected faults.