基于子域自适应迁移学习的PWM供电下永磁同步电机损耗快速预测方法OA

Previous research has conventionally utilized the finite element analysis(FEA)to calculate losses of interior permanent magnet synchronous machines(IPMSM)under pulse width modulation(PWM)supply,which takes a lot of time to calculate the losses over the entire operating ranges.Semi-analytical methods and small-signal methods are proposed in some research to fasten the PWM voltage excitation(PVE)loss calculation.However,these approaches does not always have universality due to the complex structure of IPMSMs and software operation.To address this issue,this paper proposes a fast prediction method for PWM losses based on subdomain adaptive transfer learning deep neural network(SA-TLDNN). Firstly,the losses under sinusoidal current exciton(SCE)and PVE are calculated to validate the similarity of the distribution under the two excitation modes.Besides,the SCE losses over entire operating range and PMW losses at several operating points are collected for training the model.The loss distribution is found to be complex and therefore,the subdomain adaptive method is utilized to divide the working points into several subdomains according to the control strategy.The loss distribution within each subdomain follows a regular pattern,which simplifies the model effectively.Aiming at the condition of small sample number of PMW losses,the gradient sequential sampling(GSS)method is utilized to fully capture the characteristics of PMW loss data.Based on the above methods,a SA-TLDNN model is established to maintain high prediction accuracy in the case of small samples.Finally,the model is trained using the previously prepared SCEFEA and PVEFEA data. The root mean squared error(RMSE)is chosen to be the evaluation criteria.To validate the influence of sampling methods,the train progress is conducted under uniform sampling and GSS,respectively.The RMSE of the stator hysteresis loss is significantly decreased over 50%when GSS is applied.Subsequently,the operating range is divided into two subdomains to simplify the model.The RMSE of the stator hysteresis loss is reduced by 45.10%when the subdomain adaptation is conducted.The model is also trained with different sample numbers to validate the performance under few-shot condition.It is found that SA-TLDNN has the highest accuracy among other deep neural network models.The model-predicted results are compared with the FEA and experimental results.The relative error of PWM losses between predicted and simulated results are within 10%,which verifies that the proposed fast PWM loss calculation method has sufficient accuracy and can significantly reduce the calculation time. The following conclusive are drawn from this research.(1)Transfer learning approach is well suited to the entire-range PWM loss calculation issue.(2)The subdomain adaption method can effectively reduce the model complexity and improve the prediction accuracy.(3)The source domain data feathers are effectively captured in few-shot condition.The loss prediction method based on SA-TLDNN adopted in this paper is able to realize the fast prediction of PWM losses under entire operating ranges,which can effectively facilitate the accurate and fast analysis of losses in the machine design process.