Intelligence Recognition Of Fault Networks

Various types of deep and large fault systems are widely developed in the Tarim Basin and their own characteristics control the static and dynamic spatial distribution of hydrocarbon reservoirs.The complexity and specificity of faults make them often combine transporting and sealing properties.Faults with different scales,periods,and segments modify the reservoir performance of ultra-deep strata.Therefore,fault structures are an important way to find superior reservoirs with ideal storage and seepage properties,high oil and gas production rates,and long stable production periods.However,fault interpretation tasks become more and more difficult as the complexity of seismic exploration increases,especially for ultra-deep seismic data.Recently,numerous researchers have utilized automatic interpretation techniques based on deep learning to improve the efficiency and accuracy of fault prediction.With the application of artificial intelligence in geological structure analysis,deep learning methods raise the demand for diversity in labeled structural learning sets.To improve the generalizability and flexibility of the training sets,a three-dimensional structural modeling framework is established in this thesis.Firstly,the three-dimensional fold pattern is approached by the Fourier series and Gaussian equation.Secondly,to supplement the deficiency of the stochastic simulation algorithm in simulating listric faults,an ellipsoidal surface method with random perturbation is established.Thirdly,the near-field displacement of oblique-slip faults is modeled under the assumption of rotational consistency.Finally,the fault drag is defined by the magnitude and direction of near-field displacement and the drag radius.By randomly combining parameters in some predefined ranges,the proposed modeling framework can automatically construct numerous structural models with rich geological information.To validate the applicability of the proposed modeling framework,the generated models are used as learning sets to train a U-shaped fully convolutional neural network.Experiments using synthetic and field seismic data for fault interpretation show that the trained network based on the proposed modeling framework can provide better fault interpretation results compared to conventional algorithms.These results show that the proposed geological models have better generalizability and can effectively improve the applicability of machine learning.Although deep learning methods have powerful data information processing capabilities,the applicability of deep neural networks may still be limited by the range of learned information.Therefore,we develop a new technique called structural data augmentation to enhance the diversity of the datasets.Concretely,we utilize different geological structure theories to incorporate virtual folds and faults in the field seismic data to improve the diversity and generalization ability of the training datasets.To cope with the multi-stage and multi-scale complex structures developed in ultra-deep strata,the proposed augmentation workflow increases data diversity by generating various virtual structures containing multi-scale folds,listric faults,oblique-slip displacement fields,and multi-directional fault drags.Tests on the field seismic data show that our method not only outperforms conventional seismic attributes but also has advantages over other machine learning methods.To benefit from the revolutionary advances in computer vision,we propose a new network architecture called FGSA UNet Transformer(FGSA Unetr)for 3D seismic data fault characterization.Compared to the Swin Transformer-based neural network architecture,the proposed approach can save close to 50% of the video memory footprint in the fault characterization tasks.In this network,a module called Fast Global Self-Attention(FGSA)is designed in this thesis,whose computation is implemented using the Fast Fourier Transform.Computation using the Fast Fourier Transform gives the Self-Attention module higher computational efficiency and less memory usage,with the flexibility of global attention and a cheaper computational cost.The results of fault identification from field seismic data show that the method proposed in this thesis can be effectively applied to seismic data interpretation,and the integration of global information can effectively improve the applicability of machine learning.