摘要
辐射源定位结果的有效性判定能够排除噪声定位结果干扰,保留真实有效的辐射源定位点,进而获取一个清晰连续的闪电成像图.基于电磁时间反转(electromagnetic time reversal,EMTR)的雷电甚高频辐射源定位方法具有较高的定位精度,但其定位结果有效性判定方法依靠主观设定的阈值,无法准确区分弱辐射源和噪声定位结果;其次,该方法定位速度较慢,时效性较差.为了改善这些问题,本文提出了一种基于神经网络辅助决策的定位方法,构建了一个双通道二维卷积神经网络分类模型.首先对滑动窗口的时域信号进行离散傅里叶变换,将其频点幅值及相位信息输入模型进行分类预测,判断其是否为辐射源;而后仅保留辐射源滑窗数据进行定位计算,减少了滑窗运算量;最后通过密度聚类算法对定位结果进行筛选并得到最终定位结果.模型在实测的人工引雷数据上的分类精度达到了99.73%.使用梯度可视化热力图对模型所学习到的特征以及分类依据进行物理涵义分析,增强了模型的可解释性以及合理性.相较于现有的EMTR方法,本文提出的方法不仅定位速度提高了21倍,同时模型具有较好的迁移泛化能力,对于未曾学习过的人工触发闪电以及自然闪电数据均具有较好的识别能力,在这些数据上的辐射源定位数量增加了55.71%,在排除噪声干扰的同时,获得了更为精细的通道结构图,并保留了更多的雷电发展分支结构.
Determining the validity of radiation location results is to eliminate the noise point interference,retain the real radiation location points,and obtain a clear and continuous lightning map.The VHF radiation source location method based on electromagnetic time reversal(electromagnetic time reversal,EMTR)has high location accuracy under low signal-to-noise ratio.However,the current EMTR location validity judgment method mainly relies on subjective threshold,which cannot accurately distinguish the weak radiation source and noise location results.Secondly,the global searching is slow and timeliness poor.To solve the above issues,this paper proposes a localization method with neural network assisted decision and constructs a double-channel two-dimensional convolutional neural network classification model.Firstly,discrete Fourier transform was performed on the time domain signals in turn to obtain the frequency domain signal,and its frequency and phase information were input into the model for classification and prediction to determine whether there was a radiation source.Then,the sliding window that was judged to contain the radiation source was used for locating,which reduced the amount of calculation.Finally,the density clustering algorithm was used to screen the location results and obtain the final mapping result.The classification accuracy of the model on the artificial lightning data reaches 99.73%.The gradient visualization heat map is used to analyze the physical meanings of the features learned by the model,which enhances its interpretability and rationality.Compared with the existing EMTR method,the proposed method not only improves the localization speed by 21 times but also has better generalization ability for recognizing unlearned artificially triggered and natural lightning,and the number of radiation source localization on these data increases by 55.71%.A more detailed channel structure is obtained while noise interference is eliminated,and more branching structures of lightning development are kept.
作者
杜双江
李云
邱实
罗小军
石立华
DU ShuangJiang;LI Yun;QIU Shi;LUO XiaoJun;SHI LiHua(Army Engineering University of PLA,the National Key Laboratory on Electromagnetic Environmental Effects and Electro-Optical Engineering,Nanjing 210007,China)
出处
《地球物理学报》
北大核心
2025年第9期3367-3385,共19页
Chinese Journal of Geophysics
基金
国家自然科学基金(51977219,42105077)资助.
关键词
雷电定位
电磁时间反转
卷积神经网络
辐射源判别
可解释分析
Lightning mapping
Electromagnetic time reversal
Convolutional neural network
Radiation identification
Interpretable analysis
