【作者】 王振宇； 吴坚； 程天笑； 姜志远； 赵一鸣； 王威； 孙凤举； 李兴文； 邱爱慈；

【Author】 WANG Zhenyu;WU Jian;CHENG Tianxiao;JIANG Zhiyuan;ZHAO Yiming;WANG Wei;SUN Fengju;LI Xingwen;QIU Aici;School of Electrical Engineering,Xi’an Jiaotong University;

【机构】 西安交通大学电气工程学院；

【摘要】 快脉冲直线变压器驱动源(FLTD)采用"化整为零"的拓扑结构,其模块化设计在提升能量转换效率的同时,也因系统包含数万至十余万量级的开关与电容器单元,导致故障诊断面临海量元器件状态监测的挑战。为解决FLTD装置支路自放电故障定位难题,本研究基于12级串联FLTD实验平台,提出数据驱动的故障诊断方法。首先,通过集成于水介质传输线内电极的D-dot电压传感器,在充电升压过程中成功捕获了有限样本条件下的故障特征信号;其次,构建了包含磁芯磁滞效应的12级FLTD等效电路模型,采用拉丁超立方采样策略进行参数空间探索,并基于实测数据特征注入量化噪声进行数据增强,有效扩展了不同自放电故障模式的样本空间;最终,建立一个一维残差时间卷积网络分类模型,在实测故障数据集中实现了100%的故障模式识别准确率,且各分类结果的Softmax置信度均超过0.8,与人工判读结果完全一致。该方法突破了传统故障诊断对器件级传感器布置需求的依赖,为包含海量元器件的脉冲功率装置提供了智能运维解决方案,对保障大型FLTD系统的长期运行可靠性具有重要工程应用价值。更多还原

【Abstract】 The fast pulse linear transformer drive(FLTD) adopts a "modularization" topology,and its modular design improves the energy conversion efficiency,but also leads to the challenge of monitoring the status of a large number of components for fault diagnosis because the system contains tens to hundreds of thousands of switches and capacitors.In order to solve the problem of brick prefire fault localization of FLTD facilities,this study proposes a data-driven fault diagnosis method based on a 12-stage FLTD facility.First,the fault characteristic signals under limited sample conditions are successfully captured during the charging process by the D-dot voltage sensor integrated in the inner electrode of the water transmission line.Secondly,a 12-stage FLTD equivalent circuit model containing the hysteresis effect of the magnetic core is constructed,and the parameter space is explored by using the Latin hypercubic sampling strategy,and the quantization noise is injected to perform the data enhancement based on the measured data characteristics,effectively expanding the sample space for various prefire fault modes.Finally,a one-dimensional convolutional neural network classification model is established,and 100% fault mode identification accuracy is achieved in the measured fault data set,which is completely consistent with the manual interpretation results.The softmax confidence of each classification result is more than 0.8.This method breaks through the dependence of traditional fault diagnosis on the demand for device-level sensor arrangement,provides an intelligent operation and maintenance solution for pulsed power devices containing a large number of components,and has an important engineering application value for guaranteeing the long-term operational reliability of large-scale FLTD systems.更多还原