| Along with the sustained and rapid development of our economy and the constant growth of the demand for energy,the contradiction between supply and demand of domestic energy is becoming increasingly prominent.From 2011 to2021,China’s dependence on foreign energy continues to grow,with crude oil alone increasing from 54.4 percent in 2011 to 73 percent in 2021.In order to maintain national energy security and ensure healthy economic development,the“14th Five-Year Plan” puts forward the requirements of increasing domestic oil and gas exploration.As one of the important means of oil and gas exploration,seismic exploration has played an extremely important role in the exploration and development of various oil and gas fields.With the decrease of conventional easy-to-explorable oil and gas reserves in middle and shallow layers,oil and gas exploration targets are transitioning from the middle and shallow layers to deep and ultra-deep layers,and the resource exploration types are gradually advancing from the conventional fields to unconventional fields.“Deep Earth” exploration is an important development direction of oil and gas exploration in China,and it is also an important guarantee for opening up new space for resource exploration and maintaining national energy security and stability.Faced with “deeper”exploration target layers and “more complex” exploration resource types,the problems of “deep,thin,hidden,low,and difficult” faced by seismic exploration technology have become increasingly prominent,making conventional seismic exploration technology gradually failed to meet the needs of “deep earth”exploration,and it is an urgent need to further improve the level of seismic exploration technology.In recent years,the advent of distributed acoustic sensing(DAS)technology has opened up a new round of research on the transformation of high-precision seismic exploration technology.In 2021,DAS was listed as a “Key scientific issue of transformative technology” in the national key research and development Program.DAS has significantly improved the spatial sampling rate of the borehole observation system,and has the potential to support high-resolution deep formation imaging.In addition,DAS also has significant advantages such as low deployment cost,short well occupation time,strong tolerance to harsh environments,and one-shot whole-well reception,which largely makes up for the shortcomings of conventional geophones.Based on the above advantages,DAS has become a research hotspot in the international industry,and is considered to be a feasible solution to replace conventional geophones for “finer” vertical seismic profile(VSP)data acquisition,and it is expected to alleviate the problems of “deep,thin,hidden and difficult” faced by seismic exploration to a certain extent.However,due to the limitations of DAS technology,the signal-to-noise ratio(SNR)of DAS VSP data is generally lower than the VSP data collected by conventional geophones.In addition,DAS VSP data contain a variety of new types of noise,such as random noise,fading noise,coupled noise,low-frequency noise,horizontal noise and checkerboard noise,and the noise intensity(level)usually shows the characteristics of “temporal and spatial inhomogeneity”,which brings new challenges to seismic exploration tasks and puts forward higher data processing requirements.In view of the processing limitations of DAS VSP data,this paper innovatively combines deep learning to carry out a series of study on DAS VSP data denoising,aiming at realizing the “three high” DAS VSP data processing of“high SNR”,“high resolution” and “high fidelity”.The study focuses on the bottleneck problems faced by DAS VSP data processing at the present stage,such as “complex noise types”,“unstable noise intensity” and “lack of labeled datasets”.Moreover,the paper proposes some specific solutions including“multi-type noise intelligent suppression technology based on convolutional neural network”,“human-computer interactive noise suppression technology based on denoising level adjustment” and “label-free noise suppression technology based on self-supervised learning”,and is committed to promoting the“high efficiency”,“high performance” and “convenient” DAS VSP data processing.First of all,in response to the “complex noise types” characteristic of DAS VSP data,a multi-type noise intelligent suppression technique based on convolutional neural network is proposed in this paper.Conventional denoising algorithms can usually only suppress one type of noise.In the face of DAS VSP data affected by multiple types of noise,it is necessary to select multiple algorithms suitable for different kinds of noise for step-by-step suppression,which is likely to cause multiple damages to the effective signals.In addition,traditional methods require a lot of manual parameter adjustment based on experience,and the processing efficiency is low.To solve the above problems,this paper utilizes the excellent feature extraction ability of convolutional neural network to simultaneously excavate the potential features of various DAS noises and effective signals,and implicitly removes various noises through the superposition of multi-layer network according to the feature differences between signals and noises.Compared with traditional methods,the proposed method has better universality of noise,and can simultaneously suppress multiple noises in DAS VSP data through a single model,effectively reducing the energy loss of effective signals.At the same time,the proposed method can realize adaptive noise suppression by a well-trained denoising model,avoid manual parameter adjustment,and significantly improve the efficiency of data processing.Secondly,in response to the “unstable noise intensity” characteristics of DAS VSP data,a human-computer interactive noise suppression technology based on denoising level adjustment is proposed in this paper.Since the noise level of DAS VSP data collected at different times or in different regions varies greatly,traditional deep learing-based denoising algorithms are usually obtained by training on data with a specific noise level,so they are only suitable for processing seismic data with a specific noise level.In the face of DAS VSP data whose noise level varies with time or space,multiple denoising models need to be trained for processing data with different noise levels,which greatly increases the computational cost and time cost.To solve this problem,this paper starts with the visualization of the “filters” of models with different denoising levels,and explores the denoising mechanism of models with different denoising levels by analyzing the commonality and differences of the “filters”.The study found that corresponding “filters” in the models with different denoising levels are very similar in terms of visual patterns,only their weight statistics such as mean and variance are different.Based on this observation,this paper introduces an adaptive filter modification layer in the traditional network framework to adjust the “filters” weight statistics.Then,by controlling the denoising level adjustment factor λ in the adaptive filter modification layer,the denoising level of the model can be flexibly adjusted according to the needs,so as to solve the problem that the existing deep learning-based denoising model can only process data with a single noise level or multiple discrete noise levels within a certain range.Finally,in response to the “lack of labeled datasets” problem of DAS VSP data,based on the Noise2 Noise theory,a self-supervised learning-based denoising framework named Sample2 Sample is proposed in this paper.Traditional supervised deep learning-based denoising algorithms often require massive labeled datasets to drive model parameter updating.However,due to the current study on DAS data denoising is still in its infancy and data labeling is a time-consuming and expensive process,the lack of labeled datasets hinders the denoising performance of such algorithms.In this context,self-supervised learning-based denoising algorithms that can discard labels have attracted extensive attention from geophysicists.Self-supervised learning can effectively avoid the network’s dependence on labels,and can use only noisy data to learn noise suppression.However,traditional self-supervised learning-based methods usually require training data to meet strict prerequisites,such as training pairs need to contain the same signal components,but such data is difficult to obtain in practical applications.To solve this problem,this paper improves the self-supervised learning idea and proposes the Sample2 Sample network framework.Starting from a single noisy DAS VSP data,it introduces a training pair generation strategy to generate paired sub-samples that approximately meet the Noise2 Noise training requirements.In addition,in order to compensate for the denoising error caused by the subtle differences between the signals in the generated sub-samples,this study further introduces a regularization loss to reduce the loss of signal details,so as to achieve high-fidelity signal recovery without the guidance of labels.In conclusion,this study aims to improve the quality of DAS VSP data by means of signal processing,solve the severe challenges faced by DAS VSP data processing at the present stage,and further promote the level of DAS VSP data processing,so as to provide high-quality data support for the application of DAS VSP technology in “deep earth” exploration. |
PDF Full Text Request