Research On Denoising And Evaluation Methods For High-dimensional Seismic Data Based On Deep Learning

Gas and oil have become strategic assets that every major country must control.Seismic exploration technology is being practically applied in the exploration of underground precious resources,with geoscientists using seismic data to obtain the required geological structures for interpretation and research.The interpretation of seismic data depends on clear,intuitive,and complete seismic images.However,in reality,seismic images are affected by many noise factors such as data incompleteness,which can interfere with the interpretation work of geological researchers.Therefore,this thesis proposes a high-dimensional seismic data denoising algorithm and an unsupervised seismic data quality control method.The innovative research is as follows:(1)This thesis proposes an unsupervised tensor network-based method for noise suppression of high-dimensional seismic data.Deep learning methods have demonstrated excellent performance in seismic data denoising;however,in most cases,clean seismic data is not available in practical production.Currently,deep learning noise suppression methods based on a matrix perspective are powerless when faced with complex highdimensional seismic data.To address this challenge,this thesis proposes a framework transformation domain-based p-order tensor neural network(FPTNN)model that uses data-driven high-dimensional prior knowledge to implicitly learn the typical behavior of clear multidimensional seismic images.(2)A no-reference seismic image quality assessment method is proposed in this thesis.Currently,image quality assessment methods have not been applied in the seismic data field,and most neural network evaluation methods rely on good reference images or labeled data.Faced with a wide variety of seismic noise types,the current methods’ training methods for specific datasets are somewhat inadequate and require improvement.Therefore,to address the above issues,this thesis proposes an unsupervised deep learning-based seismic image quality assessment method that can intelligently evaluate seismic image quality in a data-driven manner without relying on any prior information,measure data amplitude preservation,and thereby improve the accuracy of reservoir prediction for lithological oil and gas reservoirs.