Seismic Noise Suppression And Data Reconstruction Via Unsupervised Feature Learning

Seismic noise suppression and data reconstruction are an important part of seismic data processing,which can provide high-quality input data for seismic imaging and inversion.This dissertation discusses in detail the unsupervised feature learning theory based on neural network and low-rank matrix approximation,and their applications in seismic noise suppression and data reconstruction.There are many forms of noise in seismic data,and each type of noise has its own unique formation mechanism.According to the coherence between noise and seismic signal,seismic noise can be divided into two types:coherent noise and incoherent noise.This dissertation studies the method of incoherent noise suppression.Based on different unsupervised feature learning theories,this dissertation proposes three new methods for seismic incoherent noise suppression.Specifically:1)Construct a three-layer autoencoder network based on unsupervised learning,add sparse constraints on the hidden layer,optimize the network to learn the hidden layer features that can represent the effective signal,and finally use the learned hidden layer features to represent the noisy seismic data to achieve noise suppression;2)Construct a deep convolutional neural network based on unsupervised learning,adjust the network structure to enable it to reduce the dimensionality of the data.Due to the participation of nonlinear activation functions and convolution operations,the low-dimensional complex hidden features obtained after network optimization also contain the local spatial correlation features of the data.Using this hidden layer feature to express noisy seismic data is beneficial to improve the fidelity of the effective signal while suppressing noise;3)Based on the low-rank matrix approximation theory,first perform Hankel matrixization of the seismic data in the frequency domain,and use the kernel norm of the signal matrix and the L₁ norm of the noise matrix to construct a low-rank constraint convex optimization problem,and then use the orthogonal subspace learning method to optimize the objective function to learn the low-rank features of the noisy Hankel matrix,which can finally reconstruct the denoised seismic data.The proposed robust low-rank matrix approximation method can not only effectively suppress Gaussian random noise,but also has strong robustness in the face of abnormal incoherent noise with excessive amplitude.Due to the limitation of field acquisition conditions,the original seismic data may be irregular or under-sampled in spatial distribution.The seismic data reconstruction technology can reconstruct it into regular and complete seismic data in spatial distribution.Based on different unsupervised feature learning theories,this dissertation proposes two new methods of seismic data reconstruction to recover missing traces.Specifically:1)First,build a multilayer neural network with increasing dimensions,and only use the observed data with missing traces to participate in the training of the network,so that the network can learn the low-dimensional hidden layer features of the observed data in an unsupervised learning way and use them to reconstruct the complete data.The proposed method can effectively recover the missing traces of 3D seismic data;2)Based on the low-rank matrix approximation theory,the restoration of missing traces for incomplete seismic data can be regarded as a matrix rank-reduction problem.Based on the data reconstruction framework of the multi-channel singular spectrum analysis method,this dissertation uses the robust matrix rank-reduction method to replace the traditional truncated singular value decomposition or random singular value decomposition to obtain more reliable low-rank features.The proposed method can reconstruct the five-dimensional seismic data,and has better reconstruction performance than the multi-channel singular spectrum analysis method under the interference of abnormal incoherent noise with excessive amplitude.