Automatic Seismic Phase Identification Method Based On Deep Learning

Earthquake disasters occur frequently in China,and accurate earthquake early warning is one of the effective means to mitigate earthquake disasters.As the core link of earthquake early warning,seismic phase identification accuracy directly affects the accuracy of early warning.However,with the increasingly complex seismic data acquisition environment,the signal to noise ratio of seismic signals gradually decreases,and even effective seismic signals are submerged in noise,which affects the accuracy of seismic phase identification.Therefore,accurately identifying the seismic phase of low signal to noise ratio seismic signals has become an important topic in the current field of seismology research.Traditional seismic phase identification methods often cannot cover all features in seismic waveforms,and are not suitable for accurate seismic phase identification in complex environments;Due to the strong mapping ability of deep learning methods for nonlinear signals,it is possible to learn all the characteristics of the signal,which is conducive to improving the accuracy of seismic phase identification.Therefore,this paper has conducted research on seismic phase identification of low signal-to-noise ratio seismic signals in complex environments based on deep learning methods.The main research work includes:(1)Aiming at the problems of low accuracy and high missed detection rate of seismic phase recognition of low signal-to-noise ratio seismic signals,a new seismic phase recognition model UBAN(U-net-Bidirectional Gated Recurrent Unit-Attention Network)is designed based on U-net neural network U-net,combined with Bi-GRU and Attention mechanism.In this model,a bidirectional gated recurrent unit Bi-GRU and an attention mechanism Attention are added between the U-net coding layer and the decoding layer.Bi-GRU is suitable for processing long time series signals,and the past and future seismic phase characteristics can be integrated through the positive and negative gated recurrent units to realize the mining of time series characteristics in seismic signals and reduce the missed detection rate.Attention is used to pay attention to the time series characteristics of seismic phases,ignoring the advantages of useless features such as noise and peaks,improving the network ’s perception of arrival time features and improving recognition accuracy.The UBAN seismic phase recognition model is trained and tested using the STEAD data set of Stanford University,and compared with the U-net series model.The root mean square error is reduced by 1.3 ms～3.3 ms,the accuracy is increased by 2.26 %～4.43 %,and the missed detection rate is reduced by 1.9 %～5.8 %,indicating that the UBAN model significantly improves the seismic phase recognition effect of low signal-to-noise ratio seismic signals.(2)The UBAN seismic phase recognition model has an ideal seismic phase recognition effect on low signal-to-noise ratio seismic signals,but the seismic phase recognition effect is slightly worse when the noise is impulse noise.In this paper,a new seismic phase recognition model R-UBAN(Residual-U-net-Bidirectional Gated Recurrent Unit-Attention Network)is designed based on the UBAN seismic phase recognition model,combined with step convolution and residual module.Using the difference between impulse noise and polarization direction information of seismic phase,the R-UBAN recognition model replaces the maximum pooling layer with step convolution.The step convolution learns the feature weights in the convolution kernel through back propagation,which can comprehensively consider all the information of the feature map and realize the recognition of polarization direction information of seismic phase.However,the introduction of step convolution will deepen the number of deep network layers,resulting in gradient disappearance and network degradation.Therefore,the residual module is further introduced to avoid the problem of gradient disappearance caused by the large number of deep network layers and the difficulty of identical mapping of redundant layers.The R-UBAN phase recognition model is trained and tested by using the Stanford University STEAD data set.The experimental results show that compared with the UBAN model,the R-UBAN model shows good performance in the phase recognition of low signal-to-noise ratio seismic signals with impulse noise.The root mean square error of P-wave and S-wave phase recognition is reduced by 1.8 ms and 1.0 ms,the correct rate is increased by 1.29 % and 4.57 %,and the missed detection rate is reduced by 0.64 % and 1.98 %.This shows that the R-UBAN phase recognition model can identify the polarization direction information that distinguishes impulse noise from phase arrival.Improve the effect of seismic phase identification with impulse noise.(3)In order to further test the effectiveness of the R-UBAN seismic phase recognition model,the actual data monitored by China ’s seismic stations with lower signal-to-noise ratio were selected for model testing.The recognition accuracy of P-wave and S-wave was 89.11 % and 71.29 %,the missed detection rate was 4.62 %and 0.75 %,and the root mean square error was 78.8 ms and 234.4 ms.Compared with the U-net series model,the root mean square error decreased by 0.1 ms～9.2 ms,the correct rate increased by 0.6 %～5.04 %,and the missed detection rate decreased by 0.33 %～6.95 %.This shows that the R-UBAN seismic phase recognition model can still identify the polarization direction information that distinguishes impulse noise from seismic phase arrival when the seismic signal with lower signal-to-noise ratio is identified,and the ideal seismic phase recognition effect is obtained.The generalization ability of the R-UBAN seismic phase recognition model is further tested by using the data of Hebei,Shanxi and Beijing with different noises and epicentral distances.The root mean square errors of P-wave and S-wave arrival time are 79.6 ms～86.2 ms and 225.8 ms～237 ms.The recognition accuracy is81.79 %～87.44 % and 65.42 %～76.23 %,and the missed detection rate is0.35 %～8.09 % and 0.53 %～4.12 %.This shows that the R-UBAN seismic phase recognition model still has a good seismic phase recognition effect on seismic signals with various noises and too high or too low epicentral distance,which proves that the R-UBAN seismic phase recognition model has generalization.