Optimal Design Of Permanent Magnet Synchronous Motor Based On Deep Learning Modeling Method

Vibration reduction and noise reduction is an important research direction in the optimal design of permanent magnet synchronous motor,and the stress analysis of motor is of great significance in the research of motor noise reduction.At present,the stress analysis of the motor is mainly simulated by the finite element calculation.In this paper,the finite element calculation is carried out by establishing the electromagnetic mechanical coupling numerical model.Every time the model is modified,it needs to be recalculated.In the process of parameter optimization,the number of variables is increased by using the parameter scanning function of finite element software,and the number of scanning is increased into geometric number,and a model is calculated It takes several hours or several weeks,and sometimes the system crashes,making it difficult to complete the calculation task.Therefore,We begin to study the method of surrogate model,and tries to use the data generated by finite element software to train the deep neural network,use its deep learning ability to learn the physical laws,and use its generalization ability to make the trained neural network model applied to large-scale finite element calculation.Therefore,it can realize the motor noise reduction with higher efficiency and optimize the design of permanent magnet synchronous motor..The research group proposed an innovative and practical noise reduction method.The negative magnetostrictive material was filled in the stator of the motor,and the position and radius of the hole were adjusted to find the best position to reduce the most noise.In this paper,firstly,the electromagnetic mechanical coupling numerical model of the motor is established,The stator core model of a 4-pole permanent magnet synchronous motor is simulated by finite element method.Five groups of different stator parameters scanning simulation experiments are carried out.At the same time,the numerical data of five groups of simulation experiments and the stress deformation diagram data of the motor are derived,which provides training and test data set for the following deep learning algorithm.Secondly,the deep learning algorithm is carried out to process the numerical data,and the BP neural network prediction model based on multiple regression is established.It can fit different punching positions with the minimum stress value of the motor corresponding to the position,and can output the minimum stress corresponding to the position within 1s under the determined drilling position.The prediction accuracy of the neural network model is up to 98%.It provides a great help to find the best drilling position faster and more accurately,and greatly speeds up the analysis process of finite element calculation results.Finally,the deep learning algorithm is carried out to process the motor stress deformation image data,and three different generative adversarial network models are trained,which are pix2 pix HD,Star GAN and Cycle GAN.These three models can learn the mapping relationship between the data sets of the motor stress deformation diagram,and generate high-resolution and high-quality images Compared with the results of finite element calculation,stargan has the best generation effect,which has85% pixel accuracy,80% Per-class accuracy and 78% class IOU.In terms of time efficiency,the three network models are far better than the finite element calculation,which can replace the image mapping process of finite element simulation calculation,for deep learning in multiple physical fields Modeling and optimization of PMSM design direction has great practical significance.