Denoising Of Seismic Data Based On Deep Learning

As a new method in seismic exploration,simultaneous seismic source technology has the advantages of high acquisition efficiency,low exploration cost,etc.,but blending noise will inevitably be introduced into the seismic data,which will affect the subsequent processing and interpretation of seismic data.The deblending technology can denoise the blending noise and deblend the simultaneous source data into normal single source data.However,the traditional deblending method is difficult to select parameters,high in calculation cost,and highly dependent on relevant practitioners.Especially with the development of geophysical exploration technology,the amount of seismic data collected has grown explosively,and the manpower and material costs required for seismic data processing have gradually increased,It is urgent to explore a new intelligent method to reduce costs.The deblending method based on deep learning has low manual dependence,fast calculation speed after training,and can intelligently batch process massive seismic data,which is a hotspot of current research.Therefore,this paper has carried out research on deblending based on depth learning,improved the structures of DnCNN and Vision Transformer models respectively according to the characteristics of seismic data noise denoising task,and applied them to blending noise denoising,and achieved good results.The main work is as follows:(1)The classical deep learning noise reduction model DnCNN in visual field is applied to deblending.According to the additive characteristics of blending noise and Gaussian random noise,seismic data containing blending noise will participate in DnCNN training in image form,and the trained model will be used for deblending of real seismic data.The experimental results show that DnCNN can accurately extract the blending noise in noisy seismic data samples.After processing,the signal-to-noise ratio of seismic data is significantly improved,and the separated single source data samples are relatively clear,meeting the needs of subsequent seismic data processing and interpretation.(2)An improved DnCNN is proposed to solve the problems of poor stability,frequent loss curve oscillation and overfitting in the training process of DnCNN.The improved DnCNN uses the hybrid dilated convolution module to replace the ordinary convolution in the hidden layer of the original model to expand the receptive field,eliminate the gridding effect,and reduce the overfitting problem with Dropout.The experimental results show that the improved DnCNN can still converge rapidly and stably in the case of large learning rate,and has stronger ability to extract blending noise than DnCNN,and can properly protect the effective signal,so that the weak part of the effective signal in the denoise result is also clearly visible.(3)According to the characteristics of seismic data denoise task,the Vision Transformer,which was originally used for classification task,was improved and applied to deblending.Vision Transformer combines the Self-Attention to further improve the model’s ability to extract blending noise features.The experimental results show that,on the premise of having a large number of training data,compared with the CNN model composed of convolution,residual and other structures,Vision Transformer is more accurate and comprehensive for extracting blending noise,and has stronger noise reduction performance.Using Vision Transformer to process seismic data samples containing blending noise significantly reduces the residual effective signal in the noise,which is conducive to simultaneous-source data interpretation and processing.