摘要
闪电哨声波是一种重要的电磁波动,了解其传播特征及传播过程有助于揭开圈层电磁耦合机理.从卫星观测资料识别闪电哨声波通常需要将原始电磁波形进行滤波处理再转化为时频图像,最后采用目视方法识别图像中的色散状形态,整个过程消耗大量人机时间和内存资源,不能满足张衡一号(ZH-1)卫星观测的海量电磁场数据处理的需求.针对该问题,鉴于闪电哨声波原始波形数据能够通过播放器产生降调的声音,本文打破以视觉分析为主的闪电哨声波研究惯例,首次采用语音智能技术研究其自动识别算法.首先,以张衡一号卫星感应磁力仪(SCM)的VLF波段的波形数据为研究对象,截取时间窗口为0.16 s的波形数据作为音频片段;然后对该片段进行去趋势处理;基于梅尔频率倒谱系数(MFCCs)能够刻画人耳的听觉机理,提取闪电哨声波的MFCCs特征;其次,构建长短期记忆(LSTM)神经网络并输入波形数据的MFCCs特征训练分类模型;最后利用MFCCs特征和训练得到的LSTM分类模型实现闪电哨声波自动识别.通过对10200数据集(5100段包含闪电哨声波,5100段无闪电哨声波)上开展实验发现:该方法的准确率为96.7%,召回率为84.2%,调和平均得分(F1-score)为90.0%,AUC(Area under Curve)评分为90.1%,而且消耗的时间成本是2.28 s,消耗内存资源是82.89 MB;当前最优的基于时频图的闪电哨声波识别算法在本数据集上的准确率为97.3%,内存消耗为233 MB,在CPU上处理0.16 s的片段数据所消耗的时间是6.71 s,内存消耗和时间消耗比较严重.相比而言,基于智能语音的闪电哨声波识别算法准确率略低0.6%,但能够节约66%的时间成本以及65%的内存资源.这表明该算法不仅仅适合从卫星观测的海量数据中快速准确识别出闪电哨声波,且更适合应用于星载识别.
The lightning-induced whistler is a type of important electromagnetic wave in space,by studying its characteristics and propagation we can understand and uncover the electromagnetic coupling mechanism of the lithosphere-atmosphere-ionosphere.To recognize lightning whistler from satellite observation data,it is usually necessary to filter the original electromagnetic waveform data and transform them into the time-frequency domain,and finally,visually identify the dispersive shape in the wave spectral image.Such a process consumes a lot of human-computer time and high memory resources of the computer,cannot meet the needs of mass electromagnetic data processing of ZH-1 satellite observation.To solve this problem,according to that the original waveform data of lightning whistler wave can produce a falling tone sound through the audio player,this paper firstly adopts the automatic recognition algorithm by using intelligent voice technology for the first time,instead of the conventional visual analysis method.Firstly,using the very low frequency(VLF)band waveform data of the search coil magnetometer(SCM)onboard the ZH-1 satellite,a segment of waveform data with a time window of 0.16 seconds is extracted as an audio clip which is then detrended.Then,based on Mel frequency cepstrum coefficients(MFCCs)which can describe the auditory mechanism of human ear,the MFCCs features of lightning whistler are extracted.Additionally,the long-and short-term memory(LSTM)neural network is constructed,and the MFCCs feature is input to train the classification model of waveform data;Finally,MFCCs features and LSTM classification model are used to realize automatic recognition of lightning whistler.Experiments were carried out on 10200 data sets(5100 segments including lightning whistler and 5100 segments without lightning whistler).The results show that our algorithm have the precision of 96.7%,the recall of 84.2%,the F1 score of 90.0%,the AUC(Area Under Curve)score of 90.1%,the costing time of 2.28 s,memory consumption of 82.89 MB respectively.Compared to the current best lightning whistler recognition algorithm which has a memory consumption of 233 MB,the accuracy of 97.3%,and a CPU processing time of 6.71 s,which means that the algorithm in this study has a slightly lower accuracy of 0.6%,but can save 66%of the costing time cost and 65%of the memory space.In comparison,the proposed algorithm is more suitable for fast and accurate recognition of lightning whistler from massive satellite observations.
作者
袁静
王子杰
泽仁志玛
王志国
丰继林
申旭辉
吴鹏
王桥
杨德贺
王统领
周乐
YUAN Jing;WANG ZiJie;ZEREN ZhiMa;WANG ZhiGuo;FENG JiLin;SHEN XuHui;WU Peng;WANG Qiao;YANG DeiHe;WANG TongLing;ZHOU Le(Institute of Disaster Prevention,Sanhe Hebei 065201,China;National Institute of Natural Hazards,Ministry of Emergency Management of China,Beijing 100085,China;Tsinghua University,Beijing 100084,China;Institute of Physical Education,Huzhou University,Huzhou Zhejiang 313000,China)
出处
《地球物理学报》
SCIE
EI
CAS
CSCD
北大核心
2022年第3期882-897,共16页
Chinese Journal of Geophysics
基金
中国地震局教师科研基金(20150109)
国家自然基金(41874174)
国家重点研发项目(2018YFC1503501)
防灾科技学院教研教改项目(JY2021A06)
基础性科研院所稳定支持项目(A132001W07)
国际合作项目the APSCO Earthquake Research Project PhaseⅡand ISSI-BJ project联合资助。
