基于机器学习算法的IGBT模块故障预测技术研究

发布时间：2018-10-16 18:25

【摘要】：作为电力电子设备中重要的大功率开关器件,IGBT(Insulated Gate Bipolar Transistor,绝缘栅双极晶体管)已经被应用到众多领域,发展前景十分广阔。然而长时间工作在高强度环境中的IGBT模块会逐渐老化,甚至失效。因此能够对IGBT模块进行实时故障预测,不但可以减少日常维护成本,更能避免因器件失效而造成重大事故。基于此问题,本文结合机器学习相关算法对IGBT故障预测技术进行研究。本文主要工作如下:首先,对IGBT模块的工作特性、失效原因以及退化参数进行研究,选取关断状态的瞬态尖峰电压作为观测参数,对NASAPCoE研究中心提供的老化数据进行处理,得到预测实验所需要的尖峰电压退化数据。其次,深入研究机器学习领域中的回归算法和神经网络算法,利用Google开源平台TensorFlow搭建算法模型,对关断尖峰电压退化数据进行训练和预测。结果表明:使用RMSProp和批标准化优化的长短期记忆循环神经网络具有较高的预测准确度和训练速度,可以对IGBT退化数据进行预测。最后,为了提高设备维护效率,设计一套采用B/S架构的IGBT实时故障预测软件系统。历史数据的预测使用随机梯度下降法对训练过程进行加速;实时数据的预测采用滚动预测方式实现,有效避免因长时间预测带来的误差积累问题。该系统实现了基于循环神经网络算法的IGBT故障预测功能,具有实际意义。
[Abstract]:As an important high-power switch device, IGBT (Insulated Gate Bipolar Transistor, insulated gate bipolar transistor (IGBT), it has been applied in many fields and has a broad prospect. However, IGBT modules that work long hours in high-intensity environments may gradually age, or even fail. Therefore, the real-time fault prediction of IGBT module can not only reduce the cost of daily maintenance, but also avoid serious accidents caused by device failure. Based on this problem, this paper studies the IGBT fault prediction technology combined with machine learning algorithm. The main work of this paper is as follows: firstly, the working characteristics, failure reasons and degradation parameters of IGBT module are studied, and the transient peak voltage of turn-off state is selected as the observation parameter to process the aging data provided by the NASAPCoE research center. The peak voltage degradation data needed to predict the experiment are obtained. Secondly, the regression algorithm and neural network algorithm in the field of machine learning are deeply studied, and the algorithm model is built by using Google open source platform TensorFlow to train and predict the voltage degradation data of off peak. The results show that the long and short term memory cycle neural networks optimized by RMSProp and batch standardization have high prediction accuracy and training speed, and can be used to predict IGBT degradation data. Finally, in order to improve the efficiency of equipment maintenance, a IGBT real-time fault prediction software system based on B / S architecture is designed. The prediction of historical data is accelerated by stochastic gradient descent method, and the prediction of real-time data is realized by means of rolling prediction, which effectively avoids the problem of error accumulation caused by long-term prediction. The system realizes the function of IGBT fault prediction based on cyclic neural network algorithm, which has practical significance.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN322.8

【参考文献】