Research On IGBT Performance Degradation Prediction Method With Multiple Characteristic Parameters Based On Machine Learning

With the improvement of modern science and technology,the performance of semiconductor devices has been continuously improved.Therefore,semiconductor devices are widely used in industries,energy,communications and other fields.Due to the long-term effects of environmental stress,the performance of semiconductor devices is prone to degenerate and eventually failure,which may cause the electronic equipment to fail to work normally,or even serious accidents.Therefore,the performance degradation prediction at the level of semiconductor devices is very important to ensure the reliability of electronic equipment.Insulated gate bipolar transistor(IGBT)has become a typical representative of semiconductor devices due to its low driving power and large current density.Taking IGBT as the research object and machine learning methods such as kernel principal component analysis(KPCA)and least squares support vector machine(LSSVM)as the core algorithms,this paper researches on characterization of IGBT performance degradation,performance degradation prediction method and parameter optimization for performance degradation prediction model.Firstly,for the characterization of IGBT performance degradation,this paper proposes that a single characteristic parameter cannot effectively characterize the state of IGBT performance degradation.Therefore,based on KPCA,weighted Markov distance,negative function and wavelet denoising,this paper proposes a method to extract IGBT health indicator based on multiple characteristic parameters.By reducing the redundant information among the multiple characteristic parameters,distinguishing the importance between the main components and reducing the impact of noise factors,the IGBT health indicator obtained by this method can effectively characterize the state of IGBT performance degradation.Secondly,for the parameter optimization for IGBT performance degradation prediction model,this paper uses genetic algorithm as the optimization algorithm.By analyzing the characteristics of the existing genetic algorithms,this paper proposes an explosion operator to achieve a local explosion with self-regulating range when the evolution of population stagnates.Combining the explosion operator and the catastrophe operator,and adopting the elite retention strategy and the adaptive cross-rate strategy,this paper proposes an improved genetic algorithm.This improved genetic algorithm improves search efficiency of parameter optimization for IGBT performance degradation prediction model.Finally,for the IGBT performance degradation prediction,by comparing the commonly used machine learning algorithms,LSSVM is selected as the core algorithm.Using the advantages of several algorithms,this paper combines LSSVM,particle filter(PF)and improved genetic algorithm to obtain a prediction method to achieve IGBT performance degradation prediction and uncertainty expression based on health indicator.The short-term prediction error,long-term prediction error and the standard deviation of multiple prediction error of this prediction method are 1.2%,5.2%,and 1.6%,which proves that this method has good prediction accuracy and stability.