As an important high-power switching device in power electronic equipment,IGBT(Insulated Gate Bipolar Transistor)module has been applied to many fields,and has a promising future.However,IGBT module will gradually aging when working long hours in high strength environment,even failure.So,real-time fault prediction of IGBT module not only can reduce the daily maintenance costs,more to avoid major accidents due to device failure.To solve this problem,this thesis studies the IGBT fault prediction technology based on machine learning algorithm.The main works are as follows:Firstly,study on IGBT working characteristics,the reason of failure and degradation parameters.The transient peak voltage when device switch-off is selected as the observation variable.To get the peak voltage experiment data,the raw aging data is processed,which provided by the NASA PCoE Research CenterSecondly,focus on the regression algorithm and neural network algorithm in machine learning.The algorithm model is built by Google's TensorFlow open source platform used to train the peak voltage data.The results show that LSTM(Long Short-Term Memory)recurrent neural network with RMSProp and batch normalization optimized has high prediction accuracy and high training speed,which can predict the IGBT degradation data.Finally,in order to improve maintenance efficiency,make an IGBT real-time fault prediction software system using B/S architecture.The prediction of historical data is accelerated by stochastic gradient descent method;Real-time fault prediction is realized by rolling method,which avoids error accumulation problem due to the long time prediction.The system implements the IGBT fault prediction based on the recurrent neural network algorithm,which has practical significance. |