基于神经网络的IGBT结温预测  

Junction temperature prediction of IGBT based on neural network

作　　者：李国元[1] 严伟 周斌 肖庆中 LI Guoyuan;YAN Wei;ZHOU Bin;XIAO Qingzhong(School of Electronic and Information Engineering,South China University of Technology, Guangzhou 510640, China;Science and Technology on Reliability Physics and Application of Electronic Component Laboratory, The Fifth Electronics Research Institute of the Ministry of Industry and Information Technology, Guangzhou 510610, China)

机构地区：[1]华南理工大学电子与信息学院,广东广州510640 [2]工业和信息化部电子第五研究所电子元器件可靠性物理及其应用技术重点实验室,广东广州510610

出　　处：《华中科技大学学报（自然科学版）》2019年第7期68-72,共5页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：装备发展部技术共性资助项目(JAD1628130);广东省自然科学基金资助项目(2016A030310361);电子元器件可靠性物理及其应用技术重点实验室基金资助项目(614280601041705)

摘　　要：针对大电流下绝缘栅型双极晶体管(IGBT)饱和压降和集电极电流与结温之间的非线性关系带来的结温预测难题,搭建了大电流下IGBT饱和压降测试系统,获取了结温和集电极电流与饱和压降之间的非线性关系曲线,分析了关系曲线变化规律对应的物理机制.采用Matlab软件建立了误差反向传播(BP)神经网络模型和径向基函数(RBF)神经网络模型进行结温预测.与多项式数学模型预测结果对比表明:两种神经网络模型的预测相对误差和预测误差90%置信区间比多项式数学模型更小,结温预测精度更高;并且BP神经网络模型的预测精度高于RBF神经网络模型,结温预测模型选择时应优先考虑BP神经网络模型.In order to solve the problem of insulated gate bipolar transistor(IGBT)junction temperature prediction resulted from the nonlinear relationship among junction temperature,collector current and saturation drop voltage under large collector current,a test system was designed to measure the saturation drop voltage under different collector current and temperature.The curve among junction temperature,collector current and saturation drop voltage was obtained,and the curve was analyzed.Back propagation(BP)neural network model and radical basis function(RBF)neural network model coded in Matlab software were established.It is found that the relative prediction errors and confidence interval with 90%confidence level of two neural network models are smaller than that of polynomial fitting model,indicating that the neural network models provide more accurate prediction results than polynomial fitting model,and BP neural network is more accurate than RBF neural network,so the BP neural network is the preferential model.

关 键 词：绝缘栅型双极晶体管(IGBT) 结温预测 饱和压降 神经网络 置信区间 

分 类 号：TN386[电子电信—物理电子学]