Fault Recognition And Reconstruction Using Intelligent Methods In 3D Seismic Data

Fault is the fracted zone stretched or extruded from continuous layer.Fault interpretation is the key to seismic interpretation,and it directly affects the efficiency and effect of oil and gas exploration.With the larger-scale and the shorter-period of the exploration and the increasing demand for fine interpretation,the intelligent solution of fault recognition and construction has attracted the attention of the academic community in the industry.The fault itself is complex in shape and difficult to find.fault interpretation based on two-dimensional seismic data will lose a lot of spatial features.Convolutional neural network has the characteristics of high-dimensional complex information mining,automatic target feature extraction,and end-to-end integration process.It is a state of art research for 3D fault recognition and reconstruction based on convolutional neural network.In order to improve the effect of fault recognition and reconstruction,this paper studies intelligent methods on fault recognition and reconstruction in three-dimensional data,emphasizing more on volume instead of point or section.The specific work and innovation of the article are summarized as follows:1)This paper proposes a 3D-Unet++ convolutional neural network.The neural network uses three-dimensional synthetic seismic data as data samples,and the model has good generalization ability.For different data types such as synthetic data and area seismic data,the results show good fault recognition capabilities.The model is designed with batch normalization to suppress the overfitting problem,with Focal loss as loss function to solve the indistinguishability between faults and background details,which greatly improves the accuracy of fault recognition.2)This paper proposes a neural network based on video streaming.The neural network views the seismic section as a video frame and the seismic data as a video stream composed of seismic sections.Guiding with a small amount of manual interpretation of seismic section,learn the correlation between seismic profiles.The data sample is a set of continuous multi-frame sections.The training of the neural network uses a progressive iterative method to segment the output of the t-frame instance as the previous frame input of the t + 1 frame.The samples at the inference stage are not limited to fixed fault sections,but are set with interpretation ranges of interest to experts.This semi-supervised method combining expert knowledge and machine learning experience is convenient for experts to implement fault interpretation quickly.The significance of fault instance segmentation is that it not only splits faults and non-faults,but also fault types of different combinations in different periods.This provides the main data basis for fault reconstruction and facilitates rapid reconstruction of faults.In this paper,the validity and practicability of the method are verified respectively by synthetic data and area seismic data.