考虑温度影响的车用驱动电机磁场分步泛化实时求解方法

Accurate and fast quantification of the effect of temperature on electromagnetic performance is of great significance for reliable operation, condition monitoring, and optimized design of motors. However, the working conditions of automotive permanent magnet synchronous drive motors are complex, and the working temperature is dynamic and variable. The motor's excitation characteristics change with temperature, and its magnetic circuit exhibits highly nonlinear behavior. It is difficult to obtain a fast, accurate solution for the magnetic field that accounts for temperature using traditional finite-element or thermal-network methods. This paper proposes a fast solution method for the magnetic field with stepwise generalization across different temperatures and motor operating points. First, an electromagnetic simulation model of the motor accounting for temperature is established, and the magnetic field matrix is extracted from the simulation results. Secondly, a step-by-step generalization-solving method is proposed based on a migration-learning neural network and an interpolation algorithm. The magnetic field matrix is reduced using the POD algorithm to construct a low-dimensional feature space, and a neural network is trained to learn magnetic field features at different temperatures. The neural network and Kriging interpolation algorithm construct the mapping relationship of operating the point-temperature-magnetic field step by step. Finally, a prototype experimental platform is built to verify the proposed method. The results show that the proposed method calculates the magnetic field over a complete electric cycle in 0.95 s, which is 1.12% of the time required by the finite element method. The relative error of the reconstructed average magnetic density in the generalized range does not exceed 2.34%. The proposed step-by-step generalization strategy provides a new way to solve the complex temperature-magnetic field interactions in multi-field coupling of electric machines. The following conclusions can be drawn. (1) The proposed method has good generalization performance and meets the requirements of fast and accurate magnetic field reconstruction. (2) The proposed method effectively scales down the training cost of the neural network. Compared with a traditional data-driven neural network trained on 50% of the training set, the parameter freezing method further reduces network training time. The total training time of the network in the proposed method is 26.74% that of the neural network in the control group. (3) An experimental prototype with search coils is prepared, and the proposed method is verified.