Research On The Method Of Seismic Data Denoising Based On Symmetric Dilated Convolutional Neural Network

In seismic exploration,due to the interference of various kinds of noises,useful seismic events of effective waves are often difficult to distinguish in a seismic record.This would affect the subsequent data processing and interpretation.Therefore,data denoising is an important part of seismic data processing.Conventional denoising methods often have some limitations,such as low efficiency of denoising,or damage of effective information.Therefore,the design of a method that can effectively denoise and protect the effective signals has always been a research topic of denoising.In recent years,the Convolutional Neural Network(CNN)de-noising has opened up a new way in addition to the conventional de-noising methods,and has made remarkable progress.In this paper,Dn CNN(Denoising Convolutional Neural Network)is improved.In its convolution kernel,the form of dilated convolution is used.The convolution kernel parameters and cavity factors in the network structure are set to the form of mirror symmetry.A new Symmetric Dilated Convolutional Neural Network(SDCNN)is designed.Our SDCNN retains the advantages of the batch normalization of Dn CNN,so that it can improve the accuracy of network training steadily;and the introduction of the form of dilated convolution can increase the receptive field to the input seismic data,which is conducive to making full use of the context information to distinguish the signal and noise in the seismic data.Through the forward modeling on Marmousi model,we synthesize a set of experimental data to train our SDCNN.Then,a comparative numerical experiment,on the denoising effect of the trained network,is performed using the low-pass filtering method,wavelet transform method,Dn CNN method and our network model to denoise the same batch of data.The results show that after denoising with our SDCNN,the peak signal-to-noise ratio of the data is 9-10 d B,5-6 d B and 1-2 d B higher than that of the low-pass filter,wavelet transform and Dn CNN,respectively.In terms of maintaining the effective signal waveform,our SDCNN has a very prominent advantage,which is more conducive to identifying different types of wave signals than the other three methods.In addition,the parameters of SDCNN are greatly reduced compared with the conventional convolution neural network,which improves the training efficiency and reduces the time cost.These together show that our SDCNN has advantages over the other methods.The practical application of SDCNN is also tested using real seismic data.The test demonstrates that SDCNN has made excellent achievements in denoising various types of noise with strong randomness.The results show that SDCNN not only has better de-noising ability than f-x deconvolution filtering,low-pass filter,wavelet transform and Dn CNN,but also has better protection for effective signals.The seismic events of the de-noised seismic record by SDCNN is clearer than other methods.It shows that our SDCNN has excellent performance in de-noising of real seismic data.