Imbalance-Robust Feature Transferred Method for Single-Phase PWM Rectifier Fault Diagnosis

Open-circuit faults in high-speed train pulse rectifiers critically compromise traction system reliability. Practical maintenance scenarios face dual challenges: 1) severe data scarcity induces imbalanced class distributions; and 2) conventional diagnostic methods exhibit degraded accuracy under cross-domain conditions. To address these limitations, a feature transfer diagnostic framework is developed. First, domain-invariant feature representation is constructed through transfer learning to bridge source-target domain discrepancies. Second, a dynamic weight adaptation mechanism automatically prioritizes minority-class features during model training, effectively mitigating the impacts of class imbalance. The classification module synergistically combines transferred features with original labels through synergistic optimization, achieving cross-domain robustness. Finally, validation on an edge computing platform demonstrates the framework's computational feasibility and performance in the simulated scenarios based on the actual system.