| Energy is an important material foundation and driving force for the progress of human civilisation,and is a matter of national security and livelihood.The "14th Five-Year Plan" for modern energy system clearly puts forward to increase domestic oil and gas exploration and development,enhance energy science and technology innovation capabilities,and take intelligence and efficiency as one of its basic principles.With the distribution of shallow resources in key domestic basins largely proven,the deployment of resource exploration will shift to the development of deep oil fields,deep shale gas and tight gas,which asks for higher capacity in oil and gas exploration and development technology.Seismic exploration is one of the most effective methods in oil and gas exploration and development.Due to the limitation of cost and difficulty of deployment,it is difficult for traditional detectors to obtain large-scale and high-precision seismic exploration data that meets the requirements of "three highs and one standard" in the new era.In recent years,distributed optical-fiber acoustic sensing(DAS)technology has been applied to seismic exploration internationally,and is gradually replacing traditional detectors due to its low cost and high accuracy of obtaining data.It has many superior advantages of full well coverage,high sampling density and strong tolerance compared to conventional geophones in harsh environments.However,due to the outdated DAS instrumentation and inadequate fibre optic deployment technology,the collected DAS data are usually contaminated by a variety of complex noises with different characteristics,making the recovery of weak signals extremely difficult and detrimental to the subsequent stratigraphic inversion,imaging and interpretation work.Therefore,the noise suppression method for DAS seismic survey data is of great significance to promote the intelligent digital development of domestic DAS technology and to enhance the modernisation of the energy industry chain.To date,a number of classical noise suppression methods and their developments have been verified to be successfully applied to seismic exploration noise suppression,but only one or a few types of noise can be specifically removed,and the denoising process is strongly dependent on threshold functions or manual parameter settings.When faced with a large number of complex noises with different characteristics in DAS data,their efficiency and effectiveness are even less satisfactory.In recent years,coinciding with the rise of deep learning,Convolutional Neural Network(CNN)algorithms have also been widely used in various fields.Although the existing network methods have certain suppression effect and good adaptiveness for DAS noise,suppression effect is at the cost of serious signal loss or lower efficiency.Therefore,SSPNet is suitable for DAS seismic survey data is proposed in this thesis to achieve intelligent and efficient processing of DAS data.A series of basis vectors are generated from input to form a subspace domain,where a set of feature basis vectors are learned with the help of self-attention and guide the decouple projection of noise components from the signal components.Unlike using the attention module for region or feature selection,SSPNet uses self-attention mechanism in subspace to direct projection.Self-attention mechanism,which is both focused and holistic,performs better at capturing the internal relevance of data.Projection decomposition of vectors maintains fine structure information of the input,making it possible that the reconstructed result keep most original useful information.Finally,synthetic and field data experimental results illustrate that after the processing of SSPNet,the effective signals,whether in shallow layers disturbed by strong noise or in a deep layer where the signal energy is weak,both gets well recovered.Its superior performance to some traditional and the existing network methods is fully verified and analyzed. |
PDF Full Text Request