Application Of Attention-guided Convolutional Neural Network In DAS-VSP Data Denoising

Oil and gas resources have always been one of the most important energy resources on the earth,and due to their non-renewable nature,the increasing demand also means that the difficulty of exploration and acquisition by countries around the world is gradually increasing.China is also a major country in the use and exploitation of oil and gas resources,and the in-depth exploration of oil and gas resources is currently the focus of seismic exploration.In recent years,vertical seismic profiling(VSP)and distributed optical fiber sensing(DAS)have become hot technologies in the field of seismic exploration because the DAS-VSP data collected using these technologies have higher spatial and temporal resolution,enabling more precise detection of stratigraphic structure and the location of oil and gas reservoirs.However,the actual DAS-VSP data collected often contains various types of complex noise,which will annihilate the effective signals,which seriously affects the analysis of seismic exploration signals,and poses great difficulties for subsequent inversion imaging and the exploration of oil and gas locations.However,traditional seismic exploration denoising methods are difficult to achieve good results in processing DAS-VSP data due to their limitations.Therefore,more effective denoising methods urgently need to be proposed.This paper proposes an attention directed convolutional neural network that can be applied to DAS-VSP data processing by learning convolutional neural networks and improving the network using attention mechanisms and deep learning methods such as extended convolution.The new network utilizes attention mechanisms to assign different weights to information of different importance,thereby extracting more important information,enabling the network to learn the characteristics of signals and noise more effectively.Expanding convolution expands the receptive field of convolution operations by injecting holes into the convolution core,improving the efficiency of learning.In order to train a new neural network,this paper constructs a training set of DAS-VSP data,constructs an effective signal training set through simulation synthesis,and constructs a noise training set by intercepting noise from actual records.Through training,the network can effectively suppress noise and restore effective signals,improving the signal to noise ratio of DAS-VSP data.In the experimental part of this paper,the advantages of this method have been verified by a variety of comparative experiments.In the processing of synthetic data,the overall and local aspects have also been adopted,and the denoising results have been evaluated by using such indicators as signal-to-noise ratio,average absolute error,kurtosis and skewness,which further shows the advantages of this method in denoising.Through the processing of the actual records,it is verified that the method in this paper can effectively process the actual DAS-VSP records,and has good generalization.Finally,the interpretability of neural network is also studied,and the training of neural network and the role of attention mechanism are explained from the general trend by visual method.