基于生成对抗网络与随机森林组合模型的地震与地脉动区分研究

Discrimination of earthquake and microtremor based on generative adversarial network and random forest combination model

地震事件识别是地震监测业务的基础性工作,特别是随着大规模地震台站建设、海量地震数据汇聚以及地震预警的发展,从连续波形记录中自动区分地震与地脉动噪声显得更加重要。以准确识别地震事件为目标,提出了一种基于生成对抗网络(generative adversarial network,GAN)与随机森林(random forest,RF)的地震事件识别组合模型,该模型先利用生成对抗网络提取波形信号特征、再利用随机森林基于提取的波形信号特征将地震事件识别转化为地震与地脉动的分类问题。地震与地脉动各5378条数据的测试集研究结果表明,该模型对地震事件与地脉动的分类准确率均可以达到99%以上,地震事件识别率比较传统的长短时窗方法(short term averaging/long term averaging,STA/LTA)提高了23.56个百分点,表明该模型可以从地脉动中准确识别地震事件,并在地震监测与地震预警中具有应用前景。

Seismic event identification is the basic work of seismic monitoring service.Especially,with construction of large-scale seismic stations,aggregation of massive seismic data and development of earthquake early warning,it is more important to automatically distinguish earthquake and microtremor noise from continuous waveform records.Here,aiming at accurately identifying seismic events,a combined model of seismic event identification based on generative adversarial network and random forest was proposed.With the model,firstly,the generative adversarial network was used to extract waveform signal features,and then based on extracted waveform signal features,the random forest was used to convert seismic event identification into classification problems of earthquake and microtremor.Testing sets study results of 5378 data of earthquake and microtremor,respectively showed that the proposed model’s classification accuracy for earthquake events and microtremors can reach more than 99%;compared with the traditional short term averaging/long term averaging(STA/LTA)methods,the model’s recognition rate for earthquake events increases by 23.56%,so the proposed model can accurately discriminate earthquake events from microtremors,and have application prospects in earthquake monitoring and earthquake early warning.