Research On Aging State Evaluation Method Of IGBT Module Based On Machine Learning

IGBT module is a core device for power transmission and transformation,and has a wide range of applications in areas such as smart grids,rail transit,and distributed generation of renewable energy.The internal structure of IGBT module will gradually age and fail when the module is subjected to a long period of thermal stress and mechanical stress.Therefore,the aging state evaluation of IGBT module is of great significance for the residual life prediction and reliability analysis of the converter and for ensuring the stable operation of the system.Based on the reasonable selection of aging characteristic parameters,this paper proposes an aging state evaluation method for IGBT modules based on extreme learning machine,and also conducts an in-depth study on the optimization of extreme learning machine.Firstly,the structure and working principle of IGBT module are introduced.The mechanism characteristics of chip failure and package failure are analyzed in detail.The common aging failure characteristic parameters of IGBT module are summarized.After comprehensive comparison,the saturation voltage is selected as the characteristic parameter of the aging state evaluation of IGBT module in this paper,and the influence of collector current and junction temperature is considered.Secondly,the correlation between saturation voltage and junction temperature and collector current is analyzed theoretically.The saturation voltage,junction temperature and collector current dataset of IGBT module at different aging degrees are obtained by power cycle accelerated aging test and single pulse measurement test.The experimental data analysis proves that the combination of the three electrical and thermal characteristics can characterize the aging state of IGBT module.Based on this,an aging state evaluation model of IGBT module based on extreme learning machine(ELM)is proposed.The model is trained and tested by the obtained test dataset,and the feasibility of the model is verified.Thirdly,for the shortage of ELM that its input weight and hidden layer biases random generation,the dandelion algorithm is used to optimize the extreme learning machine to obtain the DA-ELM algorithm.The validity of the DA-ELM algorithm is verified by using five regression prediction benchmark datasets.Then the aging state evaluation model of IGBT module based on DA-ELM is proposed,and the model is tested by aging test dataset.The experimental results show that the prediction effect of the model is better than the ELM evaluation model and the genetic algorithm optimized extreme learning machine evaluation model.Finally,aiming at the shortcomings of whale optimization algorithm,a hybrid improved whale optimization algorithm(HIWOA)is proposed.Six typical test functions are selected to test the performance of HIWOA.The results show that the improved whale algorithm has significant improvements in convergence accuracy and convergence speed.Then using HIWOA to optimize the input weights and biases of ELM,the aging state evaluation model of IGBT module based on HIWOA-ELM is established.The experimental results show that the prediction effect of HIWOA-ELM model is better than that of ELM model,WOA-ELM model,DA-ELM model and so on,which verified the applicability of the proposed algorithm in the aging state evaluation of IGBT module.