Random Noise Suppression For Seismic Exploration Based On Novel Convolutional Neural Network

Seismic data exploration is the key to underground structure detection and oil and gas exploration.However,the complexity and unpredictability of random noise in seismic data bring great difficulties to imaging and interpretation of subsequent seismic data.How to effectively extract signal from noisy data to improve the SNR and resolution of seismic data is an important research direction in seismic signal processing.According to the easy saturation of convolutional network performance and the limitation of feature extraction of convolutional network.In this paper,two random noise suppression methods for seismic surveys based on novel convolutional neural networks are proposed.The main studies are as follows:1.Batch renormalized denoising networkFor deep convolution neural network convergence difficulties,network performance tends to saturation and batch normalization(BN)technology is limited by hardware.Moreover,the BN performance grades seriously when the batch size is too small.Based on these problems,the paper proposes batch renormalization network.The network combines two parallel branch networks to increase the network width for obtaining more characteristics.Specifically,the upper network is composed of residual learning(RL)and batch renormalization(BRN)and the lower network including BRN,RL and expansion convolution.The batch renormalization technique solves the batch problem of BN.The network uses overall RL and expansion convolution to promote network training and extract more features for denoising.BRDNet is applied to synthetic seismic recording and actual seismic recording,and compared with f-k filtering,wavelet denoising,time-frequency peak denoising and other denoising methods,the results show that BRDNet can suppress random noise more effectively.2.Multiscale attention mechanism networksAiming at deep convolutional networks ignore the influence of shallow networks on deep networks and only focus on single scale local feature information.Thus,the paper proposes a multi-scale feature extraction attention mechanism network which combines pyramid compression(PSA)module and attention mechanism module.The PSA module can effectively integrate local and global attention and learn more multi-scale information.Attention mechanisms are used to reinforce the most useful characteristic information.Moreover,the extended convolution is used to expand the sensing domain to extract characteristics and reduce the computational cost.The network combines residual learning and batch normalization to prevent over-fitting and improve network performance.The MFE-ADNet is applied to synthetic seismic records and actual seismic recordings,and compared with the traditional denoising algorithm and attention mechanism network,MFE-ADNet has efficient denoising ability and clearer texture of the recovered seismic signal.