Physics-constrained twin network: A cross-equipment domain adaptation diagnosis framework based on digital twin information transfer

作者：Xing Zhikai, Yongbao Liu, Qiang Wang, Mo Li, He Yuxuan · 发表于：Results in Engineering · 年份：2025 · DOI：10.1016/j.rineng.2025.107649 · 被引用次数：4 · 研究领域：Fault Detection and Control Systems、Machine Fault Diagnosis Techniques、Digital Transformation in Industry

To address the challenges of insufficient model generalizability for cross-equipment fault diagnosis in real industrial scenarios, including significant domain discrepancy between Digital Twin (DT) and physical equipment, and difficulties in information transfer under noisy conditions, a Physics-Constrained Twin Network (PCTN) based domain adaptation framework is proposed. This approach constructs bidirectional feature mapping pathways between the virtual domain of the digital twin and the target domain of physical equipment, with embedded physical prior information to enable robust joint representation learning despite noise interference. Initially, a Noise-aware Feature Disentanglement Module (NFDM) is introduced to isolate equipment-independent background noise. This is achieved by integrating real-world operational data indicative of equipment health states with simulated data from the DT. Subsequently, a physics-constrained domain adaptation module is designed, incorporating dynamic feature alignment and physics-based constraints in a joint optimization strategy. This effectively mitigates the distributional shift between the DT domain and the target equipment domain. Finally, a Gated Sparse Regularization (GSR) mechanism is developed. The gating coefficients adaptively strengthen the discriminability of critical cross-domain fault features as training progresses. Experimental evaluations on both publicly available datasets and a self-developed diagnostic platform demons...