With the development of the oil and gas industry,large amount of seismic data and low signal-to-noise ratio of seismic data have become a challenge for seismic data processing.Therefore,how to process large amounts of seismic data quickly and effectively without parameter input by human is a new problem of seismic data processing.Machine learning method is based on statistical theory,thus,supervised learning can turns into an optimization problem of empirical risk function or structural risk function.Through a brief summary of autoencoder and convolutional neural network,it is believed that Convilutional Auto Encoder(CAE)deep neural network can extract data features better.In this study,we proposed an improved CAE architecture consisting of multiple convolutional modules.Seismic data processing method based on CAE deep neural network take the spatial characteristics of seismic data into consideration,and treat the problem of seismic noise attenuation and seismic data reconstruction as a extraction problem of effective signals in different environments.By designing the underground velocity structure of different geological models,we get the seismic data for training form the numerical simulation of seismic waves.During the training process of CAE deep neural network,L1 loss function and Adam optimization algorithm is used to enhance the robustness and improve calculation efficiency.The CAE method can applied to random noise attenuation,linear noise attenuation,surface wave attenuation and seismic data reconstruction effectively which is proved by numerical experiments.The seismic data processing method based on the CAE deep neural network is a data-driven method.This method without the mathematical constraints in traditional method,and does not require set the manual input of threshold,therefore has higher processing efficiency. |