Research On Seismic Data Denoising Method Based On Convolutional Neural Network

In seismic exploration,due to the complexity of the field operation environment,the accurate interpretation of seismic data images requires a high signal-to-noise ratio.Therefore,denoising the collected seismic data is a key link in seismic data processing.In recent years,the development of deep learning in the field of artificial intelligence is in the ascendant and has made rapid development at home and abroad.Because its convolution neural network has a significant effect on the feature learning,recognition,classification and prediction of image data,therefore,it has been gradually applied to seismic data denoising.Aiming at the problem of removing a large amount of random noise and retaining effective texture information.In this paper,the standard convolution are used to construct the cyclic generation adversarial network model.In the generator of the network model,the nonlocal convolution network module is used as the residual connection to maintain the long-term dependence of the global data.In the discriminator of the network model,the shallow convolution neural network is used to judge the output of the generator,and the adaptive moment is used to estimate and optimize the network parameters.After experimental analysis,The network algorithm has good denoising effect on seismic data.To restore the fine lines of the generated data image,In the generator,this paper uses standard convolution and dilate convolution to form sparse blocks to increase the depth and perception domain of the network and obtain more image texture features.In the discriminator network,patchgan is used instead of full connection as the output mode.Compared with the experimental results before improvement,it is found that the fine texture is restored.Matlab software is used to preprocess the simulated seismic data and actual seismic data by cutting,rotating,equalizing and normalizing,and then the network model is compiled by the deep learning framework tensorflow to train the data.In the process of training many times,the parameters such as network depth,convolution number,convolution kernel size,learning rate and optimizer parameters are adjusted to optimize the training of weights.The experimental results of the test data show that the deep learning model constructed in this paper can remove a large amount of random noise in the seismic data and obtain the seismic data with high signal-to-noise ratio,which lays a foundation for the interpretation work.