A Study Of Seismic Event Detection And Wave First Arrival Picking Based On Deep Learning

In recent years,with the increasing investment of the state in protection against and mitigation of earthquake disasters,the seismic monitoring network has been gradu-ally improved.Seismic stations have recorded a large number of waveform data,these waveform data provide data support for seismic positioning,seismic analysis and other research.When processing continuous waveform data,the first step is to detect the waveform fragments which contain seismic events from the continuous waveform data,and then pick up the first arrival time of wave in the waveform fragments which contain seismic events.Initially,relevant processing is manually done by well-trained practi-tioners,which requires a lot of manpower and time costs.Therefore,automatic seismic event detection and wave first arrival picking methods have become hot topics in aca-demic research.Seismic stations receive waveform data to collect continuous waveform records.When processing waveform records,seismic events are first detected,and then the first arrival of wave is picked.Seismic event detection refers to searching the waveform segments containing seismic events from the continuous waveform data.Picking first arrival of wave refers to the identification of time when the P wave(Primary Wave)and S wave(Secondary Wave)first arrive at the observing station in the seismic events.At present,two popular automatic processing methods,STA/LTA(Short-and Long-Time Average Ratio)and AR-AIC(Autoregressive Akaike Information Criterion)are mainly based on signal processing,designing waveform attributes,setting thresholds and re-quiring a lot of effort on tuning parameters to achieve the optimal effect.STA/LTA and AR-AIC not only require the designer to have a good command of seismic knowledge,but also require a lot of effort of tuning parameters.Although the processing is fast,its effect is greatly affected by noise.With the development of arti-ficial neural network,some scholars began to study the automatic processing methods based on artificial neural network.Although these methods save a lot of effort of man-ual parameter tuning,they still need to design the input to the artificial neural network.The manual designed inputs hardly make full use of the information contained in the waveform data.In this paper,the background of seismic event detection and waveform first arrival picking is firstly introduced,the related algorithms of seismic event detection and wave first arrival picking as well as the research progress of deep learning are analyzed and introduced.For the task of seismic event detection,most of the current convolutional neural network based methods only train one detecting network for all stations,utilize a large number of training data,and achieve good performance,but the effect will be greatly degraded when the amount of data is reduced.For the task of picking the first arrival of wave,the traditional STA/LTA and AR-AIC can not fully utilize the sequen-tiality of waveform data and are greatly affected by noise.Some deep learning based methods have achieved good picking effect by taking waveform data as input.How-ever,due to the long sequential characteristic of seismic data,these methods can only be used for mirco-earthquakes.Facing with the above challenges and the characteristics of seismic data,the main contributions of this paper are as follows:(1)Aiming at the problem of seismic event detection with few samples,consid-ering the difference caused by the geographic location of seismic stations,a multi-task learning-based seismic event detection model is proposed in this paper.Firstly,based on the distance between seismic stations,the seismic stations are clustered and the de-tection of seismic events by cluster of stations is regarded as a single learning task.Secondly,a hard-sharing based deep multi-task learning method is applied for multiple tasks,and a separate seismic event detection network is established for each learning task by combining multi-station data.Experiments show that this method improves the seismic event detecting effect of convolutional neural network based method in the case of few samples.(2)For task of wave first arrival picking,this paper proposes a deep learning based picking model by making full use of the sequentiality of seismic data and utilizing se-quence labeling.Firstly,this model designs a network,which takes the sequence data as input,reduces the dimension of the sequential data,and obtains the rough wave first arrival by using sequence labelling.Then,the rough wave first arrival is used to slice waveform,and the traditional picking method is applied on sliced waveform in order to pick wave first arrival precisely.Experiments indicate that the proposed model is ro-bust to noise and achieves high picking success rate and small picking deviation.Also,experimental results vertify the transferability of proposed network.