Deep Learning-Based Seismic Pre-Stack Inversion Method And Application Research

Compared with the post-stack seismic inversion technique,the pre-stack seismic inversion technique can provide richer elastic and physical parameters,which play an important role in fluid prediction and lithology identification.The field pre-stack seismic data generally has problems such as random noise interference and low-resolution,which affect the accuracy of inversion results.To address these problems,the seismic data need to be processed with random noise suppression and resolution improvement before inversion.With the increasing size of seismic data,conventional seismic data processing and inversion methods exhibit problems such as high computational effort,low accuracy of results,and many restrictions.Deep learning techniques have overcome these limitations to a certain extent by virtue of their ability to learn the intrinsic laws of sample data and the representation capabilities of feature levels.However,limited by the characteristics of seismic data and the defects of mainstream deep learning frameworks,deep learning-based seismic processing and inversion methods still have some limitations in practical applications.To solve the problems of existing traditional methods and deep learning methods,this paper proposes unique deep learning solution strategies for the problems of random noise suppression,high-resolution processing,and pre-stack three-parameter inversion.First,a study on the noise suppression method of seismic data by Bernoulli sampling self-supervised SEU-Net is carried out.The method completes data enhancement by directly Bernoulli sampling a single noise-bearing seismic data profile,and introduces the squeeze-and-excitation module in the constructed network to adaptively calibrate the feature responses of different channels,so as to obtain the characteristics of noise more accurately and achieve blind denoising of seismic data.Second,by combining the data-driven capability of deep learning and the time-frequency localization capability of time-frequency analysis methods,a high-resolution processing method for seismic data by using time-frequency domain complex-valued U-Net is proposed.The method constructs feature extraction and mapping for the time-frequency spectrum of low-resolution and high-resolution seismic data extracted using S-transform to achieve more accurate high-resolution processing from the time-frequency domain.Finally,to solve the low efficiency and low accuracy problem caused by the conventional deep learning-based pre-stack three-parameter inversion method ignoring the correlation between multiple tasks,research on the pertinent multi-gate mixture-of-experts(PMMOE)U-Net-based pre-stack three-parameter inversion method is proposed.The PMMOE network is divided into an expert network to process the input data,a gate network to give weights,and a tower network to output the predicted data.The expert network contains shared expert module and special expert module to process different input data.The data test results demonstrate that this expert module classification strategy can help the network to achieve effective information sharing while excluding invalid information interference,and thus effectively improves the inverse prediction accuracy of P-wave velocity,S-wave velocity,and density.