Gradient-aligned physics-informed neural network for performance analysis of permanent magnet eddy current device under complex operating conditions

The permanent magnet drive (PMD) is a non-contact mechanical power transmission device that demonstrates broad application prospects in industrial fields due to its exceptional torque control performance and structural reliability. However, with the increasing demand for real-time condition monitoring and the advancement of high-power-density PMD design, achieving efficient and accurate performance analysis has become a critical challenge. Artificial intelligence-driven digital twin approaches offer a novel perspective to address this issue, among which Physics-Informed Neural Networks (PINNs) exhibit significant potential by integrating data with physical knowledge. Nevertheless, under complex operating conditions, the skin effect of electromagnetic fields and the scale variation of solution domains in PMD severely limit the prediction accuracy and physical consistency of PINNs. To address the aforementioned challenges, this paper proposes a Gradient-Aligned Physics-Informed Neural Network (GA-PINN) method for performance analysis of PMD under complex conditions. GA-PINN incorporates electromagnetic governing equations and boundary conditions into the loss function to guide the model training process. By identifying and aligning the gradient information of the loss terms, it dynamically balances the model’s constraint capacity across multi-scale solution domains. Additionally, a dynamic sampling strategy is introduced, which reconstructs the distribution probability of collocation points by identifying the gradient characteristics of the solution function, thereby addressing the mismatch in the global distribution of collocation points caused by localized high residuals. Through experimental validation across multiple operating conditions and structural parameters, GA-PINN achieves efficient and accurate performance analysis of PMD in complex scenarios, providing a reliable theoretical basis for its performance monitoring and optimization design. Dataset link:https://github.com/wongsihan/GA-PINN.