基于多通道卷积神经网络的甚低频/低频雷电辐射电场波形分类方法  

Classification Method for VLF/LF Lightning Radiated Electric Field Waveforms Based on Convolutional Neural Networks

作　　者：肖力郎 陈维江 王宇[3,4] 贺恒鑫 傅中 向念文 XIAO Lilang;CHEN Weijiang;WANG Yu;HE Hengxin;FU Zhong;XIANG Nianwen(State Key Laboratory of Advanced Electromagnetic Engineering and Technology,Huazhong University of Science and Technology,Wuhan 430074,China;State Grid Corporation of China,Beijing 100192,China;Wuhan NARI Co.,Ltd.of State Grid Electric Power Research Institute,Wuhan 430074,China;Hubei Province Key Laboratory of Lightning Risk Prevention for Power Grids,Wuhan 430074,China;Electric Power Research Institute,State Grid Anhui Electric Power Company,Hefei 230022,China;School of Electrical and Automation Engineering,Hefei University of Technology,Hefei 230009,China)

机构地区：[1]华中科技大学电气与电子工程学院强电磁工程与新技术国家重点实验室,武汉430074 [2]国家电网有限公司,北京100192 [3]国网电力科学研究院武汉南瑞有限责任公司,武汉430074 [4]电网雷击风险预防湖北省重点实验室,武汉430074 [5]国网安徽省电力有限公司电力科学研究院,合肥230022 [6]合肥工业大学电气与自动化工程学院,合肥230009

出　　处：《高电压技术》2024年第11期5184-5191,共8页High Voltage Engineering

基　　金：国家电网有限公司科技项目(5500-202120583A-0-5-SF)。

摘　　要：雷电过程中产生多类雷电辐射电场波形,基于特征值的传统分类方法易误分类。为准确分类雷电辐射电场波形,该文提出了一种基于多通道卷积神经网络的甚低频/低频雷电辐射电场信号分类方法,该方法采用深度网络直接处理电场波形以减少先验知识依赖,设计多通道并行卷积核以有效提取雷电波形多尺度特征,引入shortcut连接以加速模型收敛。基于合肥地区实测雷电数据,该文建立了回击、初始预击穿、窄双极性脉冲以及云闪4分类波形数据集,模型训练结果表明该模型识别准确率达到99.4%,与经典机器学习方法以及主流深度神经网络模型分类性能相比,所提模型在分类准确率及训练收敛速度上均达更优水平。基于该模型,该文采用知识蒸馏方法获得了适用于低算力计算平台的部署模型,部署模型在合肥某雷暴活动中单次分类耗时仅59ms,算力需求降低66%,分类准确率为99.0%,实现了模型在低算力计算平台上的可靠应用。The lightning process generates multiple types of lightning electric field waveforms.Traditional classification methods based on waveform parameters are prone to make misclassification.To address this issue,we proposed a method of VLF/LF lightning electric field signal classification based on a multi-channel convolutional neural network.This method uses a deep network to directly process the field waveforms,reducing dependency on prior knowledge.The net-work was constructed with multiple convolutional kernels to effectively extract the multi-scale waveform features.The shortcut connections were introduced to accelerate model convergence.Based on the data collected in Hefei,a training dataset of four typical waveforms,namely,return stroke,preliminary breakdown,narrow bipolar event,and intracloud,was established.The training results show that the model achieves an accuracy of 99.4%.Compared with classic machine learning methods and deep learning models,the proposed model performs better in classification accuracy and training convergence speed.By using the knowledge distillation method,a model suitable for low-computing-power platforms can be obtained.The distilled model takes only 59 ms for single classification,with a 66%reduction in computing power re-quirements and a classification accuracy of 99.0%,demonstrating reliable application of the proposed model on low-computing-power platforms.

关 键 词：卷积神经网络 VLF 雷电辐射电场 波形分类 模型部署 

分 类 号：P427.3[天文地球—大气科学及气象学] TM866[电气工程—高电压与绝缘技术]