Seismic Data High-resolution Processing Method Based On Deep Learning And Wavelet Analysis

The exploration and development of oil and gas resources in China has gradually shifted to deep,thin and complex tectonics,which puts forward higher requirements for seismic exploration related technologies.High resolution and high-precision seismic data processing and interpretation methods have become one of the research hotspots.The existing high-resolution processing technologies for seismic data broaden the frequency band and improve the main frequency to some extent,but its ability to characterize thin reservoirs is limited.Seismic inversion technology calculates physical parameters that can reflect the lithological characteristics of the formation based on conventional seismic attributes.The seismic inversion using logging constraints can introduce high-frequency information from logging,and the inversion results have high vertical resolution,but there is a problem of multiple solutions.This thesis studies a high-resolution processing method for seismic data and a seismic impedance inversion method based on deep learning.The main work of thesis is as follows:1.Implement a high-resolution seismic data processing method based on synchronous squeezing wavelet transform(SSWT).As an improvement of wavelet transform(WT),SSWT has better time-frequency focusing and can more accurately locate the energy distribution of seismic signals in the time-frequency spectrum.According to the principle of SSWT and harmonic decomposition,high-frequency and low-frequency information is compensated based on the effective frequency band of the original seismic data,expanding the frequency band and increasing the main frequency.The vertical resolution of seismic data can be increased to within 20 m,providing a good data foundation for subsequent work,and is one of the important factors in improving the resolution of seismic impedance inversion.2.Propose high-resolution inversion of seismic impedance based on deep learning methods.High resolution comes from broadband seismic data input and the model constraints of well logging impedance.Utilizing the original seismic traces along the wellbore,SSWT broadband seismic traces,seismic meme inversion impedance,and logging impedance curves to form a multi-scale feature dataset,enriching the feature set and facilitating model learning.This thesis propose a high-resolution impedance prediction models(1D Dense Net)and a super-resolution impedance prediction model(1D SR Dense Net).The models are based on one-dimensional convolutional neural networks(1D-CNN)and introduce dense connection modules of Dense Net to achieve shallow and deep multi-scale feature fusion.Further introducing Pixel Shuffle to upsample input features,so that the impedance prediction results contain more well logging detailed information,and vertical resolution is further improved.3.Conduct high-resolution three-dimensional impedance prediction experiments in a gas field in the AM area.Firstly,the SSWT based frequency extension method is used to perform high-resolution processing on the initial seismic data.Extract seismic data from the target layer and select logging data,construct a multi-scale feature dataset of well seismic data.Train and tune the impedance prediction network model.The impedance prediction results show that the vertical resolution has been significantly improved,effectively integrating the horizontal trend of seismic data with the well logging layer details.The predicted results are highly consistent with the well logging interpretation results,and the thickness prediction of the reservoir and interlayer is accurate,which can effectively solve the identification problem of thin reservoir and interlayer in the work area.