Application Of Convolutional Neural Network In Velocity Picking Of Seismic Exploration

Seismic exploration data is very precious and very scarce,but China's increasingly better economy has put higher demands on oil and natural gas production.At present,most of the reserves of oil fields in China are low-permeability oil and gas,so it is very difficult to mine,which puts higher requirements on the technology of seismic exploration.In conventional seismic processing,superposition is one of the three core technologies.It can not only suppress random interference,but also improve the signal-to-noise ratio of signals in seismic exploration.The speed picking can provide the best stacking speed for the stack,and the speed is the key parameter in the stack,so the accuracy of the speed picking in the field of seismic exploration is very important.However,the traditional speed picking method has different defects,some of which have relatively high computational complexity,and some have defects in the method itself and need to be used together with other methods.Convolutional neural networks in deep learning are very widely used techniques in recent years.By constructing deep convolutional network models to extract features in data,we can predict the results we want well.This paper intends to use the "end" to "end" convolutional neural network to represent the problem of superimposed velocity oicking as the regression task of the image,take the local pictures of the CMP gathers corrected by NMO as input and the speed estimate as output.In addition,the CTC loss function is introduced.Through a large amount of data and experiments,when the U-NET is selected on the network,the effect of the cross-entropy loss function and the CTC loss function is compared.The improved U-NET has an effective accuracy improve,and to a certain extent,increase the adaptability of the data,This method has good application prospects in seismic exploration velocity picking.