Intelligent Segmentation Algorithm Of 3D Fault Image

A fault is a structure which caused by the tectonic activity underground.When the received stress is greater than the tectonic strength the broken structure caused the faults.Faults are good storage spaces for fossil energy such as oil and gas.Therefore,interpretation and segmentation of faults are of great significance to the study of the distribution of underground oil and gas.The traditional fault segmentation work is interpreted manually.Manual interpretation consumes a lot of labor and time cost,meanwhile it introduces subjective errors by different interpreters,and it's hard to interpret faults with complex structures.With the increasing amount of seismic data in the current situation,the manual fault segmentation method has been unable to meet the demand.In recent years,with the development of computer image processing technology,seismic researchers have introduced image processing technology to the field of fault segmentation.With the development of deep learning in the fields of computer vision and the great success in the fields of image recognition,target detection and image segmentation,it has also been applied to the field of fault segmentation.In this thesis,two different models are used to improve the fault interpretation result from both 2D and 3D.For the problem that the CNN-based fault segmentation network has limited ability to extract global features,this thesis proposes a hybrid model based on CNN-ViT(Vision Transformer),this model can benefit from both CNN and Transformer,which can make full use of CNN's inductive bias and Transformer's feature extraction ability to get global information from seismic data.This method improve the accuracy and continuity of fault interpretation,and achieve good results in actual work area data.For the problem that the single task model generalization in fault segmentation is insufficient,and the segmentation result varies greatly in different areas,this thesis proposes a multi-task model to improve it.By introducing the two related tasks(dips detection and strikes detection),using the shared representation between the main task and related tasks,to improve the generalization of the model.It makes the model perform better in seismic data with different distributions and features.Meanwhile,this thesis proposes a method to improve the fault prediction process.The methods proposed in this thesis solve the problems of limited ability of network to extract features from seismic data and the poor model generalization in fault interpretation.The fault interpretation result shows improvements on both theoretical seismic data and real seismic data,which fully proves the effectiveness and practicability of the methods.