Research On Fault Recognition And Surface Reconstruction Using Deep Learning

In the field of oil and gas exploration,faults,as key geological features,not only serve as channels for oil and gas transport,but also provide reliable information for determining optimal drilling locations.Therefore,accurate identification of faults is of great significance to geological research and actual production.Traditionally,faults are interpreted manually,which is costly,time-consuming and labour-intensive,and the subjective judgement of the interpreter may also bring errors.As the scale of seismic exploration expands,it is difficult for manual interpretation methods to meet the demand of high precision and high efficiency,so many kinds of intelligent and automated fault interpretation methods come into being.Especially the improvement of computer computing speed and the gradual improvement of statistical learning algorithm theory,the use of machine learning and deep learning algorithms to intelligently identify fault features has attracted the attention of many scholars.In this paper,based on the deep learning model,an integrated algorithm of fault grouping and surface reconstruction is proposed,which firstly segments the fault instances from a two-dimensional perspective,and then reconstructs the fault instances of different profiles in combination.Firstly,based on Cycle GAN,a random straight line image is used as input to generate a synthetic fault image matching it.The generated synthetic fault image not only increases the number of training samples for the instance segmentation network,but also can increase the diversity of samples,which provides a new idea to solve the problem of insufficient number of samples in geological data.Secondly,a fault labeling method based on semi-supervised K-means clustering algorithm is proposed,which adopts spatial density similarity metric instead of Euclidean distance,and then guides the clustering of unlabeled data in the synthetic tomographic image by a small amount of labeled data obtained from the linear image to obtain the instance labels prepared for the training of the instance segmentation network.Finally,a lightweight instance segmentation model is used for training,and the trained model is applied to fault prediction in real seismic areas.On the basis of two-dimensional fault instance segmentation,the fault profiles are regarded as time series,and the similarity between adjacent profiles is utilized,and theories such as Kalman filtering are applied to the tracking of the fault instances to realize fast fault combination and reconstruction.