Seismic Time Window Selection Based On Fully Convolutional Neural Network

Using seismic full waveform data to invert the structure of the earth is one of the most advanced inversion methods in the field of seismic imaging,to overcome the influence of noise and nonlinearity on the inversion convergence rate,it is necessary to select time windows of seismic data that meet certain criteria for inversion.When dealing with massive data,the traditional time window picking method can no longer meet the efficiency requirements of inversion,and it is necessary to develop a more intelligent and efficient waveform picking method.In the past years,artificial intelligence technology has been widely used in various fields.Here we shall use artificial intelligence to help to select the time window of seismic waves.Based on the neural network's learning ability,we design a fully convolutional neural network(FCN)to automatically select complex seismic signals.We select 37,960 waveform data from the actual exploration seismic data and resample each waveform data to 4992 sampling points.First,we use the traditional method(PYFLEX software)based on the long/short-term average ratio(STA/LTA)to pick the time window of the waveform.The PYFLEX software is designed for natural seismic events and depends on the travel-time table of the seismic event,it is difficult to accurately pick the window when dealing with exploration data.We manually refine the labeled sample by PYFLEX and construct the data set for network training.We build 36 convolutional layers,there are 7 down-sampling layers and 7 up-sampling layers.After 1000 iterations of training,our network can achieve a high recognition rate.Compared with the traditional method,the FCN window picking method is about10,000 times faster than PYFLEX.When processing millions of exploration seismic data samples,the FCN method has incomparable advantages over traditional methods.This will provide an important boost of time-window selection for the full waveform inversion method.