基于卷积神经网络的天然地震与人工爆破识别研究

Research on identification of natural earthquake and artificial blasting based on convolutional neural network

地震台网记录的人工爆破事件波形特征与天然地震有相似之处,如果不能及时的予以识别和筛选,会混淆依此记录所建立的地震目录,影响日后地震学的研究工作.因此连续地震信号中的天然地震和人工爆破的识别分离有助于破坏性构造地震的监测预警.同时,随着地震仪的大规模部署以及建设行业活动的持续增强,记录的爆破事件增多,增大了识别的难度和工作强度.为了解决人工识别天然地震与爆破存在的问题,本研究基于卷积神经网络,设计了一个具有13层深度的网络模型(CNN-Epq13),选用福建数字地震台网历史地震记录和人工爆破记录,利用TensorFlow深度学习框架,采用TFrecord文件形式作为训练集,训练得出事件类别识别模型.此模型在训练中识别准确率为98.438%,损失为0.0646.用新的数据进行测试,模型能准确识别事件类别,区分天然地震和人工爆破,可以进一步在日常工作中进行应用.

The waveform characteristics of artificial blasting events recorded by seismic networks are similar to those of natural earthquakes.If we cannot identify and screen them in time,the seismic catalogue established based on these records will be confused and the future seismological research will be affected.Therefore,the identification and separation of natural earthquake and artificial explosion in continuous seismic signals is helpful to the monitoring and early warning of destructive tectonic earthquakes.At the same time,with the large-scale deployment of seismometers and the continuous enhancement of construction industry activities,the recorded blasting events increase,which increases the difficulty and intensity of identification.With the advent of seismic big data and artificial intelligence,scholars at home and abroad have started a new round of recognition algorithm research.Some use BP genetic network for seismic and blasting identification.Some first use wavelet packet for 4-layer decomposition to obtain relevant features,and then combine Support Vector Machine(SVM)for classification.Some first use HHT transformation to obtain signal features and construct feature data sets,and then combine with SVM for classification.In some cases,the attenuation rate of seismic S wave and its spectral morphological characteristics are extracted first,and then the feature data are classified and recognized by single and combined classification using Least Square Support Vector Machine(LS-SVM).Some researchers think that BP neural network is easy to fall into local optimal,and support vector machine method lacks effective means to select appropriate kernel function,so they propose to use INTEGRATED learning method:BP-Adaboost method for classification,and achieve good results.Some researchers believe that source events are stationary signals within a certain time range.First,Mayer Frequency Cepstral Coefficient(MFCC)algorithm is used to extract features,and then Convolutional Neural Network(CNN)is used for classification and identification.All the above researches firstly extract the features of signals by some technical means,then establish feature sets and use some classification algorithms in machine learning to identify them.In the process of feature extraction,it is inevitable to lose some possible feature parameters.In order to solve the problem of artificial identification of natural earthquake and blasting,At the same time,in order to avoid the feature loss problem caused by some feature extraction algorithm.In this study,a network model(CNN-Epq13)with a depth of 13 layers was designed by using the excellent classification performance of the convolutional neural network and directly sending waveform signals into the convolutional neural network without using the method of extracting data features in advance.The historical seismic records and artificial blasting records of Fujian digital seismic network were selected,the TensorFlow deep learning framework was used,the TFrecord file form was used as the training set,and the event category identification model was trained.The recognition accuracy of this model in training is about 98.438%,and the loss is 0.0646.By testing with new data,the model can accurately identify event categories and distinguish between natural earthquakes and Artificial blasting,which can be further applied in daily work.