Open-Circuit Fault Diagnosis for Power Converters: Meta-Learning-Guided Physics-Informed Neural Network With Prior-Knowledge Integration

Open-circuit faults in power converters are prevalent yet difficult to diagnose under nonstationary operating conditions, where conventional approaches often fail to balance diagnostic accuracy and efficiency. To address this issue, a meta-learning-guided physics-informed neural network (PINN) framework with prior-knowledge integration is proposed. Specifically, a bidirectional Mamba-based temporal backbone is adopted to construct a hidden state mapper (HSM), through which long-sequence features are extracted from three-phase current signals. Moreover, a physics-guided regulator (PGR) is devised so that latent nonlinear partial differential equations (PDEs) can be learned from data. The learned physical constraints are embedded into the loss function together with two engineering priors, namely, constant phase differences among the three-phase currents and among their derivatives. Furthermore, an independent first-order meta-learning strategy is incorporated so that inner- and outer-loop optimization can be performed over a task family that is composed of voltage-load combinations, which reduces the iterations required to reach the target accuracy and enhances training stability. Experiments are conducted on a Vienna rectifier hardware platform that covers various voltage levels, load conditions, and open-circuit fault types. The results demonstrate that the proposed framework achieves a diagnostic accuracy of 99.96% and provides greater robustness and superior training efficiency than representative existing techniques, which highlights its practical value for open-circuit fault diagnosis in power converters.