Deep Learning For Recognition Of Sedimentary Microfacies With Logging Data

The study on sedimentary microfacies has important theoretical and practical significance.It is an indispensable part of the process of oil and gas exploration.The rich stratum sedimentology information contained in logging data is an important indicator to identify sedimentary microfacies.The recognition of sedimentary microfacies via the manual logging is highly subjective and time-consuming.Traditional machine learning methods such as support vector machines and neural networks have been used in the automatic identification of sedimentary microfacies.Noted that the logging curve changes the morphological characteristics through depths.But none of these methods could track these depth-related changes.Therefore,curve partitioning,feature extraction and classification need to be carried out separately.It is difficult to obtain satisfactory results for geologists.To solve this problem,this thesis converted the problem to identify different logging sedimentary microfacies into an ‘image semantic segmentation' task.We also researched the automatic recognition of sedimentary microfacies on logging curves using deep learning,for the purpose of building models that are suitable for logging sedimentary microfacies identification based on different characteristics of these logging curves.Plus,we studied logging data processing methods to improve the accuracy of model identification.Two sections can be divided accordingly in this paper.(1)For the multi-scale feature of logging curves,adjustments have been made based on the U-net network.(1)Remove the pooling layer to reduce the loss of spatial feature information;(2)Introduce multi-scale convolution blocks to achieve multi-scale feature mining(3)Add a one-dimensional convolution layer to achieve single-direction segmentation,thereby a well-logging sedimentary microfacies recognition model named Improved U-net with multi-scale feature constraints is built.A data-enhancement method,which is suitable for the identification of the sedimentary microfacies from the logging curve,is therefore proposed to enhance the performance of the model.To further improve the accuracy of this identification in the process of feature extraction,the slope value,the azimuth value,the variance value and the lithological characteristic value of the curve are calculated by the attribute extension method to reduce the data complexity.The prediction accuracy of the model has been improved to 86.49%,which is much higher than those measured by the traditional SVM and ANN methods.Therefore,the model can better identify the sedimentary microfacies,such as distributary channels,channel margins and interdistributary bays at different scales.(2)For the time sequence feature of the logging curve,the SegCaps network is selected as a basic structure to reduce information loss so that we can better retain the directionality of the logging curve.In addition,the knowledge guidance module that use human experience and knowledge to guide the network and learn more useful information has been added.The function of this module is to improve the accuracy of logging sedimentary microfacies identification.Accordingly,a logging sedimentary microfacies identification model named KD-SegCaps with temporal feature constraints has been built with a prediction accuracy of 86.73%.Compared with the Improved U-net,the KDSegCaps focuses on the longitudinal prosodic features of logging data,and can better identify sedimentary microfacies such as sand flat,sand mud flat,and mud flat.Test results show that this method can automatically obtain fine sedimentary microfacies characteristics from the 2-D image formed by the logging curve,improve the accuracy of identifying sedimentary microfacies by logging data,and solve the problem that morphological change characteristics are often overlooked.Therefore,this method successfully makes it possible to complete curve partitioning,feature extraction and classification at the same time by making the recognition more intelligent.