Intelligent Recognition Of 3D Seismic Faults And Caves Based On PSF Modeling

Fault and karst cave recognitions are crucial to seismic structure interpretation and reservoir description.Due to the increasing size of the acquired 3D seismic data volume,it is difficult for manual and traditional recognition methods to meet the production demand for efficient and precise interpretation.With the rise of artificial intelligence technology and the improvement of computer performance,intelligent seismic structure interpretation based on deep learning has gradually been attracting attention and has made significant achievements in recent years,which has greatly improved the efficiency and accuracy of seismic data interpretation.Recognition of faults and caves by deep learning algorithms is generally considered a supervised learning task,which requires numerous diverse training samples and corresponding accurate labels.Because field data with faults and caves labeled by experienced interpreters are difficult to acquire,most of the current studies use synthetic seismic data with accurate labels as training samples.However,differences between synthetic seismic training images and actual seismic images always exist,which seriously affects the generalization ability of the neural network trained with synthetic seismic data sets in practical applications.Given this,this thesis conducts research on methods to synthesize more realistic seismic training sample sets quickly and efficiently for two specific tasks,namely,seismic fault and cave intelligent recognitions.The main contents of this thesis include: 1.To deal with the shortcomings of the 1D wavelet convolution method in seismic imaging,investigate the traditional seismic imaging principle and resolution theory,and analyze the analytical calculation method of the point spread function(PSF)to develop a fast and efficient method for synthesizing a realistic seismic training sample;2.With the guidance of geological knowledge,study the automatic construction method of 3D seismic fault and cave models based on mathematical modeling;3.Study the method of building 3D seismic fault and cave training data sets based on the PSF convolution;4.Study the intelligent identification method of 3D seismic faults and caves based on the U-net image segmentation network,and verify the effectiveness,reliability,and practicality of the method presented in this thesis by the processes of theoretical and field seismic data,as well as pointing out the advantages of this method.Through this thesis,the following conclusions are obtained: 1.Since the deep neural networks mainly use kinematic features such as waveforms on seismic images when recognizing faults and caves,and the size of a seismic training sample is a small range relative to a field seismic image,the PSF can be quickly calculated using the analytical Green’s function;2.With the convolution of the PSF computed by the analytical Green’s function and the reflectivity model,the seismic training samples containing the migration and acquisition aperture effects can be simulated quickly and efficiently,thus reducing the differences between the training samples and the real seismic images.In this thesis,the PSF approach takes approximately 15 min to generate a seismic image with 5 faults of size 128 × 128 × 128 voxels,whereas it takes approximately 20 h to generate a seismic image of the same size using reverse time migration;3.The U-net trained with the PSF convolution samples has a strong fault recognition capability,which can recognize faults with minor displacement on the seismic images,and the recognition results provide high continuity and accuracy;4.The U-net trained by the PSF convolution samples can not only accurately detect the strong bead-like seismic reflection generated by the caves on the seismic images,but also precisely delineate the size and shape of the corresponding actual caves based on the bead-like reflection features.In addition,such a well-trained neural network has a high resolution for cave recognition,which can distinguish the combined bead-like abnormal seismic responses generated by the adjacent caves and characterize the corresponding cave bodies.The major innovation of this thesis is that the presented scheme in this thesis is the first successful attempt to generate seismic training sets by the PSF convolution,which minimizes the gap between the synthetic seismic training samples and the actual seismic images to some extent and improves the recognition accuracy and generalization ability of neural networks for faults and caves in field seismic images.