基于STFT和CNN的地震信号分类识别研究  

Research on seismic signal classification and recognition based on STFT and CNN

作　　者：黎炳君 黄汉明[1] 王婷婷[2] 王鹏飞 王梦琪 施佳朋 薛思敏 LI BingJun;HUANG HanMing;WANG TingTing;WANG PengFei;WANG MengQi;SHI JiaPeng;XUE SiMin(College of Computer Science and Information Engineering,Guangxi Normal University,Guilin 541004,China;Institute of Geophysics,China Earthquake Administration,Beijing 100081,China)

机构地区：[1]广西师范大学计算机科学与信息工程学院,桂林541004 [2]中国地震局地球物理研究所,北京100081

出　　处：《地球物理学进展》2021年第4期1404-1411,共8页Progress in Geophysics

基　　金：国家自然科学基金(41264001);专项资金(075440);广西重点研发计划(桂科AB18126045)联合资助。

摘　　要：地震信号分类——即信号震源类型的分类,尤其是天然地震和人工爆破的分类对于地震目录的清洗和强震实时预警等具有重要意义.本文首先对原始波形信号进行必要的预处理,然后对预处理后的信号进行分帧加窗,对窗内信号采用短时傅里叶变换,将原始波形信号从时域转换成时频域信号,生成时频谱图,每条波形生成一个时频图像;经过反复多次试验比较,将原尺寸时频谱图统一缩放为32×32像素大小的灰度图像,该灰度图像作为卷积神经网络(CNN)的输入,此时的分类效果和计算效率最优.采用以上方法对2010年至2016年发生在河北三河和北京通州等地区,震级为1.5~2.8级之间的54个天然地震事件的1674条波形和63个人工爆破事件的1509条波形,采用五折交叉验证法进行分类,得到的平均准确率为97.39%,与传统的支持向量机(SVM)方法和多层感知器(MLP)方法对比,分类准确率有大幅提升,与使用梅尔倒谱系数(MFCC)构建的CNN分类器方法对比,本文方法信噪比更低,分类准确率提高约1.5%.实验结果表明,采用短时傅里叶变换提取地震信号时频域特征,生成时频谱图,使用卷积神经网络对地震信号进行分类具有良好的分类效果.The seismic signal classification-signal source type classification,in particular,the classification of natural earthquakes and artificial blasting is of great significance for the cleaning of earthquake catalogs and real-time early warning of strong earthquakes.In this paper,first of all,the original waveform signal is preprocessed,then the preprocessed signals are divided into frames and windowed,the Short-time Fourier Transform(STFT)is applied to the signal in the window,the original waveform signal is converted from the time domain to time-frequency domain,spectrum diagram at generation time,each waveform generates a time-frequency image,after repeated experiments and comparisons,the original size of the spectrogram is uniformly scaled to a grayscale image of 32×32 pixels.This grayscale image is used as the input of the Convolutional Neural Network(CNN),and the classification effect and computational efficiency are optimal at this time.We used the above method to analyze the waveform data of 1674 natural earthquakes and 1509 artificial blasts that occurred in Sanhe,Hebei and Tongzhou,Beijing from 2010 to 2016,with a magnitude of 1.5~2.8,use the 5-fold cross validation method for classification,the average accuracy obtained is 97.39%,compared with the traditional SVM method and MLP method,the classification accuracy has been greatly improved,Compared with the CNN classifier method constructed using Mel-Frequency Cepstral Coefficients(MFCC),the proposed method has a lower SNR and a classification accuracy improvement of about 1.5%.The experimental results show that the use of short-time Fourier transform to extract the time-frequency domain features of seismic signals,generate time-spectrograms,and use convolutional neural networks to classify seismic signals has a good classification effect.

关 键 词：地震信号分类 卷积神经网络 短时傅里叶变换 天然地震 人工爆破 

分 类 号：P315[天文地球—地震学]