DOI: 10.11648/j.ajset.20261103.21 ISSN: 2578-8353

A Physics-Data Dual-Channel Collaborative Cross-Domain Transfer Learning Fault Diagnosis Method for Main Circulation Pump in Converter Valve Cooling Systems

Gao Lei, Wang Wenchao, Wu Heng, Zhang Yu, He Biao
Show PDF Cite
To address the three major challenges in fault diagnosis of main circulating pumps in converter station valve cooling systems—small-sample scenarios, cross-domain distribution shift, and lack of physical interpretability—a physics-data dual-channel collaborative cross-domain transfer learning framework is proposed. The physics channel designs a physics-constrained conditional generative adversarial network (PC-cGAN), which explicitly embeds the pump's Euler equation into the loss function to constrain the hydraulic characteristics of generated samples. The data channel constructs a multi-source signal attention fusion network (MSAF-Net), achieving cross-modal fusion of vibration, temperature, and pressure features through a physically guided mechanism. Theoretically, this paper establishes a quantitative relationship between the Wasserstein distance and the HΔH divergence, deriving an explicit upper bound for domain adaptation error. Validation on measured data from 69 main circulating pumps across five UHV converter stations shows that under the stringent condition where only 5% of target domain samples are labeled, the cross-domain diagnostic accuracy reaches 89.7%, an improvement of 13.4 percentage points over the optimal baseline. The Pearson correlation coefficient between the model's decision logic and the fault physical mechanism is 0.82 ( p < 0.001 ), indicating that the diagnostic results possess clear physical interpretability.

More from our Archive