Research On Fault Prediction Of IGBT Based On Deep Learning

Insulated Gate Bipolar Transistor(IGBT)has many advantages,such as high current density,high input impedance,and is widely used in various aspects of the power industry.However,IGBT may gradually aging and even fail when it works in high temperature,high pressure and other harsh environments for a long time.Therefore,the study of IGBT fault prediction is of great significance to reduce the occurrence of accidents and improve the stability of power system.Most researchers use mathematical statistics or machine learning to predict IGBT faults,but fail to make full use of time series information of degraded data,which results in unsatisfactory prediction.Therefore,this thesis studies the problem of IGBT fault prediction,and proposes a method that makes use of deep learning hybrid network model for IGBT fault prediction,which is applied to the software system of fault prediction.The main contents of this thesis are as follows:(1)Firstly,the failure mechanism of IGBT is deeply analyzed and studied.The peak voltage of collector-emitter turn-off is selected as the characteristic parameter of fault prediction,and the characteristic parameter is validated by the aging data of NASA PCoE research center.(2)The time series fault prediction of IGBT is studied,and a deep learning recurrent neural network method is proposed for IGBT fault prediction.This thesis explores and designs a network of LSTM(Long Short-term Memory)and GRU(Gated Recurrent Unit)time sequence prediction.To solve the problem of insufficient output information of historical time in network,an optimized network model based on full connection layer Ds(Denses)is proposed,and LSTM-Ds and GRU-Ds networks are designed for IGBT fault prediction.(3)IGBT fault prediction experiments are carried out for the established LSTM,GRU,LSTM-Ds and GRU-Ds networks.Aiming at the problem of strong data fluctuation and large harmonic periodic disturbance in LSTM-Ds and GRU-Ds network prediction,this thesis proposes a network improvement scheme.Finally,the hybrid network SES-GRU-LSTM-Ds after the Second Exponential Smoothing Method(SES)is used to predict,and the optimal network model with RMSE(Root Mean Square Error)of 0.0756 and MAE(Mean Absolute Error)of 0.0593 is obtained.(4)A visual software system of IGBT fault prediction for converter valve equipment is designed,and the optimal hybrid network model is integrated into the software system for application,which has certain practicability.In summary,this thesis takes IGBT fault prediction as the research direction,and predicts IGBT fault by using the hybrid network model of deep learning algorithm,which has reference significance for engineering application.