Research On Source Type Identification Methods Based On Deep Learning

With the large-scale and high-intensity activities of human beings on the earth,the quality of seismic observation data has been seriously affected.The seismic signals recorded by the seismic monitoring networks in various places contain signals of many unnatural seismic events,such as blasting and collapse.If these non-natural seismic signals can’t be removed in time,it will have a great impact on seismology research.However,if a large number of seismic detection signals are manually identified and classified,a lot of time and energy will be wasted.Therefore,the realization of high-efficiency,high-precision,strong generalization performance of source type recognition and classification algorithm is particularity important in the era of seismic big data.Therefore,it is the main content of this study to classify and identify the source types(natural earthquake,blasting and collapse)according to the seismic signals collected by the seismic monitoring station.Firstly,based on the traditional seismic feature extraction algorithm,a new feature extraction algorithm is proposed.At the same time,considering that the traditional method only uses the single component data in the three-component data,which may cause the loss of information,so the three-component data is spliced,and the BP neural network is used to identify the source type.Secondly,this paper proposes a ResNet depth enhancement model to solve the problem of low significance of source type recognition through feature values.This method combines the advantages of residual module and SE module,and designs a method to explore the image feature correlation between channels,adaptively recalibrate the characteristic response between channels,and use the waveform graph and the spectrogram as sample data for model training.Finally,an ensemble learning method for source type recognition is implemented in this paper,and designs an integration strategy to integrate the sub-model with the eigenvalue,waveform and spectrogram as training samples to give the final source event type.Compared with the recognition algorithm of a single sub-model,the final performance has a certain improvement.And using the monitoring data of different regions for verification,it is found that the method has good generalization performance.