Research On Noise Suppression Algorithm For Seismic Exploration Based On Convolutional Neural Network

Seismic exploration is one of the most important tools in the exploration of hydrocarbon and mineral resources.In recent years,along with the depletion of easily explored and extractable resources,the need to exploit unconventional resources in complex strata has placed higher demands on the quality of exploration records.The seismic records are usually mixed with a large amount of noise due to the acquisition conditions,which adversely affects the acquisition and identification of effective reflection information,resulting in incoherent and difficult to identify homogeneous axes and low Signal-to-Noise Ratio（SNR）of seismic data,causing great difficulties in the subsequent processing of seismic data.The existing traditional methods for processing seismic data extracted from such complex geological structures are not satisfactory.In addition,seismic random noise usually exhibits complex characteristics such as non-smooth,non-Gaussian and coincidental signal mixing,which is relatively difficult to suppress.To effectively improve the processing of exploration data,convolutional neural networks have been introduced to the field of seismic exploration for efficient and intelligent denoising of seismic exploration data,and have achieved better processing results than traditional methods.However,the existing convolutional neural network architecture is relatively simple and lacks a high-fidelity dataset adapted to exploration data processing,resulting in a deficiency of the denoising network in terms of noise suppression capability and data processing accuracy.Therefore,two new intelligent denoising convolutional neural networks based on Feed-Forward Denoising Convolutional Neural Networks（DnCNN）are designed and applied to the problem of random noise suppression in real seismic surveys in desert survey areas.To address the complex exploration noise suppression problem,this topic analyzes the problems of DnCNN networks in seismic exploration data denoising and proposes a Diverse Branch Block feature enhancement Convolutional Neural Networks（DBBCNN）.The network uses the DBB module to enrich the feature space and the null convolution to expand the perceptual field,and uses long path operations to fuse global and local features to improve the network’s ability to represent detailed features.On this basis,in order to further improve the network performance,this topic continues to improve the DBBCNN network by proposing a novel Double Multi-scale Feature Fusion Denoising Network（DMFF-Net）,which improves the network performance by increasing the network width In order to improve the performance of the network by increasing the network width,the features extracted from the two layers of different subnetworks are spliced and fused by channel to extract more comprehensive features.At the same time,in order to continuously improve the network’s abatement capability for complex random noise,a high-fidelity training dataset is constructed by combining forward modelling with actual exploration noise to meet the demand for highly realistic training sample data for high-precision noise suppression networks.In order to verify the effectiveness of the proposed algorithm,the trained network model is processed and analysed against simulated data and actual exploration data measured in the Tarim region,and the results are compared with those of classical traditional algorithms and deep learning algorithms.Both the simulated and actual data processing results show that the two methods proposed in this work can effectively suppress the complex random noise in the desert seismic data,and have good recovery capability for the deep weak reflection information in the desert survey area.In summary,the proposed noise suppression network can achieve intelligent suppression of complex exploration noise and reconstruction of weak reflection information,which can provide reference for the design of subsequent noise suppression networks.