Seismic Multiple Wave Adaptive Subtraction Method Based On Optimization Algorithm Driven Deep Neural Network FISTA-Ne

Multiple suppression is one of the difficult problems in seismic data processing for petroleum exploration.Adaptive subtraction is a crucial step in the prediction and subtraction method.The adaptive subtration method based on traditional linear regression adopts the Fast Iterative Shrinkage Thresholding Algorithm(FISTA)to solve the matching filter,which has high computational efficiency,but the algorithm requires human trial to select hyperparameters.The incorrect selection of hyperparameters can result in errors in solving the matched filter,leading to residual multiple or damaging primaries.Compared with traditional linear regression method,the existing data-driven deep neural network method can suppress multiple better,but the method lacks interpretability and relies on a large amount of labeled data during network training.Therefore,this thesis proposed an adaptive subtraction mehod driven by optimization algorithms based on deep neural network due to the issue of low accurary of traditional adaptive subtration,poor interpretability of network structure in data-driven neural network methods,and reliance on large amounts of labeled data for supervised training.The main content of this thesis is summarized as follows:(1)This thesis proposes the adaptive subtraction method based on supervised FISTANet in order to solve the problem of poor interpretability of data-driven deep neural network method and low accuracy of traditional linear regression method.The proposed method expands the iterative steps of the sparse optimization algorithm FISTA in the traditional linear regression method into a network layer,and uses the U-Net that is classical datadriven neural network instead of primarys sparse promotion,and the regularization parameters in FISTA are adaptive estimation in the network training process.Therefore,the proposed FISTA-Net has certain interpretability and can be interpreted as the iterative steps of FISTA.The results from the synthetic data processing show that the adaptive subtraction method based on supervised FISTA-Net achieves a better multiple suppression effect than that of the improved U-Net method and the traditional linear regression method,achieving a higher signal-to-noise ratio.The results from field data processing show that the proposed method can suppress multiple more effectively than the improved U-Net method.(2)This thesis proposes the adaptive subtraction method based on unsupervised FISTA-Net in order to solve the problem that the construction of labeled training samples in the field data is cumbersome and the network generalization ability under supervised training is low.Unsupervised FISTA-Net uses L1 norm minimization constraints on the primaries,and utilizes unlabeled training samples for unsupervised training due to the primaries conforms to the non-Gaussian distribution.The results from synthetic data and field data processing show that the unsupervised FISTA-Net method suppresses multiples and protects primaries better than traditional linear regression method.(3)The thesis prosposes the field data processing method based on transfer learning in order to improve effciency.Due to the adaptive subtraction of synthetic data and field data belongs to similar tasks,and synthetic data and field data have certain similarities in statistical distribution.Therefore,fine-tuning the model based on training with synthetic data to achieve adaptive subtraction of field data.The processing results of field data show that the proposed method is able to accelerate the convergence speed of the model while maintaining the accuracy of multiple separation.