Random Noise Reduction In Seismic Exploration Based On Hybrid Multi-resolution Network

Seismic exploration is an important means to explore oil and gas mineral resources.With the increasing demand for energy,easy-to-find and easy-to-capture resources have been gradually depleted,and seismic exploration is deepening into more complex topographic areas,which brings challenges to the existing seismic exploration technology.Distributed Acoustic Sensing(DAS)has been introduced to seismic exploration as an emerging acquisition technology in recent years,which has the advantages of low cost,anti-electromagnetic interference,and high acquisition density.However,there is often a large amount of noise in the seismic data acquired by seismic exploration.The noise causes discontinuities in the reflection events,making it difficult to identify the reflection events in the seismic records,and the signal-to-noise ratio is low.It seriously affects the subsequent data interpretation and processing.Therefore,an effective noise suppression method for seismic records is particularly important.The noise in seismic records has non-smooth and non-Gaussian characteristics,and it is often mixed with the effective signal at the same frequency.Traditional denoising methods rely too much on a priori knowledge and noise characteristics and cannot achieve satisfactory denoising effects.In recent years,with the development of machine learning methods,Convolutional Neural Networks(CNN)have been applied to seismic record noise suppression.This method aims to establish a nonlinear mapping of noisy records and pure records,which has a better denoising effect and stronger generalization than traditional methods.However,the early networks have disadvantages such as over-simplified network structure and relatively single scale,which can reduce the feature extraction ability and generalization of the network and thus affect the denoising effect.In addition,the convolutional neural network relies heavily on high-precision data,and it often requires matching noisy data with effective signals to build the training set in the network training,but it is impossible to obtain completely pure effective signals in the field seismic exploration,which seriously affects the application of the convolutional neural network in the field seismic data noise suppression.Based on the above two problems,a hybrid multi-resolution denoising network and a training set of high-quality seismic data are designed in this work.First,to address the problem that the scale and resolution of traditional CNN networks are too single,this paper designs a hybrid multiresolution denoising network: Dilate Pyramid Attention Network(DPA-Net),which is a multiresolution and multi-scale denoising network with codec characteristics.The pyramid module outputs multi-resolution eigenfaces for subsequent modules;The Serration Dilate Convolution increases the perceptual field;The multi-scale module extracts rich multi-scale features;And the attention module is used to enhance the reflective axis edge information.At the same time,this paper constructs a high-quality training set,constructs a high-quality pure signal by forward modeling,and then superimposes field seismic noise to meet the demand of the denoising network for a high-quality and high-fidelity training set.In order to verify the effectiveness of the denoising network in this paper,DPA-Net is tested with simulated and field seismic records,and compared with traditional denoising methods and CNN-based denoising methods.The comparison shows that DPA-Net outperforms the comparison methods in terms of noise suppression and signal retention,and meets the requirements of seismic data noise suppression.In summary,the research results of this paper have some positive effects on complex noise suppression in DAS seismic data,and can provide some reference roles for the subsequent research on denoising networks,with some research and application prospects.