Application And Research On Fault Detection Based On Deep Learning Semantic Segmentation

As the global population grows and modernization continues,the demand for energy is increasing.Although research into new and renewable energy sources has never stopped,oil and natural gas are still essential and vital resources that are needed to sustain society in all walks of life,so continued research into oil and gas exploration remains vital.Faulting is formed by the rupture and dislocation of underground layers under tectonic stress.Fault is both the boundary of oil and gas field and the channel for oil and gas transportation and accumulation.Fault interpretation is a very important task in oil and gas exploration and development.The distribution and morphology of faults play a key role in the identification and description of oil and gas reservoirs,and have a very important impact on the development of the earth's crust and the extraction and distribution of natural gas and oil.In recent years,with the continuous improvement of seismic exploration technology,the volume of seismic data has become larger and larger,and the use of conventional methods for fault interpretation is not only very trivial and time-consuming,but also more difficult and unrepeatable to verify,and the complex interpretation process also requires high professionalism of the interpreters.With the explosion of artificial intelligence applications in various industries in recent years,various methods of artificial intelligence have been applied to various industries.Deep learning is a hot topic in the research of artificial intelligence methods,which provides an idea to use deeper neural network methods to take human experience for computers to learn so as to solve certain problems instead of human judgment.Image segmentation refers to dividing an image into several disjoint regions based on its grayscale,color,spatial texture,geometry and other features,which simply means separating the target from the background in a pair of images.In the field of fault interpretation in seismic exploration,seismic amplitude images contain a large amount of useful information about subsurface geological structures,which can be used to achieve good fault interpretation.Therefore,the advantages of deep learning methods can be fully utilized to allow computers to learn the patterns of seismic fault interpretation and realize the recognition of faults in seismic data from computer vision,which can greatly reduce human intervention and errors,and significantly reduce the time required for fault interpretation.In view of this,this paper addresses some problems faced in fault interpretation in seismic data,combines the most popular deep learning methods,proposes a fault identification method based on semantic segmentation of deep learning,and investigates the problems of difficult to obtain effective samples and sample imbalance in deep learning performed on seismic data,and conducts a study on optimal post-processing of fault identification results.The main specific works are as follows:(1)This paper analyzes the problem of seismic fault identification in deep learning semantic segmentation,and designs an end-to-end deep learning semantic segmentation network suitable for seismic amplitude image,realizes the recognition of seismic fault by computer vision,and tests the network performance on synthetic seismic data in many aspects.(2)This paper studies the sample selection of seismic fault interpretation.In view of the unstable effect of synthetic data training model in predicting the real seismic data,it proposes to select part of the 2D slice data from the 3D seismic data to be interpreted to artificially interpret the fault as the training sample,and obtain enough learning samples through data augmentation.(3)The use of a weighted cross-entropy loss function to maintain a deep learning neural network for fitting learning in an accurate direction is investigated for the case of severe imbalance between the number of faults and non-faults in seismic amplitude images.(4)This paper studies the optimized post-processing of the fault identification results of deep learning semantic segmentation method.The noise in the identification results is eliminated by using the isolated small connected region removal method.The fault predicted by the model is refined by using the skeletonization method,and the redundant branches generated in the skeletonization process are removed by using the pruning operation.The eight-neighborhood endpoint detection method is used to find the identified fault line break point and connect the discontinuous faults.(5)Combined with deep learning semantic segmentation fault identification network and fault optimization post-processing,this paper proposes a fault identification process based on deep learning semantic segmentation,and carries out the process on the actual seismic data.It also makes qualitative and quantitative analysis on the fault effect identified by different data and different methods,and deeply demonstrates the stability and reliability of this method.The main innovation points of this paper are as follows:(1)An end-to-end deep learning semantic segmentation lightweight network structure for seismic fault recognition is designed by simplifying and improving based on VGG16 network.Applying the testing process on synthetic and real seismic data,it not only reduces the training time and prediction time,but also can achieve good fault identification with fewer fault samples and outperforms other deep learning methods.(2)To address the problems in deep learning semantic segmentation results,a process is proposed to optimize the post-processing of fault identification results.The process includes noise removal by isolated small connected region removal,distance transform-based skeleton extraction algorithm to obtain refined faults and pruning algorithm to remove redundant branches,and a method based on eight-neighborhood endpoint detection to find and connect the disconnected points belonging to the same fault line.The fault identification results of deep learning semantic segmentation are effectively improved in synthetic and in real seismic data.