基于卷积神经网络的地震波形智能识别研究

随着世界上多个国家和地区的地震预警系统投入运行，误报和漏报等问题逐渐突显，特别是将标定以及强干扰波形误识别为大震事件，快速、精确地区分地震与其他波形是一个难题。针对于此，该研究提出了基于卷积神经网络地震波形智能识别方法。首先收集并处理了2012—2017年中国境内福建以及周边邻省共683个地震和478个爆破事件，并对这些样本筛选、截取和基线校正等预处理，共得到了27 500条三通道波形。在此基础上，构建了3 s波形输入的卷积神经网络模型(SW-CNN)。结果表明：模型对地震、噪声、爆破和异常波形的识别率分别为97.9、99、99.2和99.3%。相比于人工手动分类识别，该模型更省时和更稳定，为地震预警目前所面临的问题提供了一个新的解决方法。

Intelligent recognition of seismic waveform based on convolutional neural network

With the operation of earthquake early warning in many countries and regions in the world, the problems of false alarm and missing alarm are gradually highlighted, In particular, calibration and strong interference signals are misidentified as large earthquake events, it is an problem to distinguish seismic waveform from other waveform quickly and accurately. For this, in this study, an intelligent recognition method of seismic waveform based on convolutional neural network is proposed. Firstly, the seismic data of Fujian, Taiwan and neighboring provinces from 2012 to 2017 were collected and processed, a total of 683 earthquakes and 478 blasting events; there samples were preprocessed by screening, interception and baseline correction, 27500 three-channel waveforms were obtained. On this basis, a convolutional neural network model(SW-CNN)was constructed by inserting three-second data records. The results show that the recognition rates of earthquake, noise, blasting, abnormal waveforms are 97.9%, 99%, 99.2%, and 99.3%, respectively. Compared with manual classification recognition, the model is more time-saving and stable for waveform classification, it provides a new solution for the problems faced by earthquake early warning at present.