| Fully mining information on reservoir lithology plays a crucial role in geological exploration and development.China is currently in an important stage of unconventional oil and gas resource development,and the requirements for reservoir lithology identification are becoming more refined.Logging data are one of the important data sources for mining reservoir lithology information.There are many types of Logging data and the ways to process them are becoming more diverse.From the initial manual identification to the current intelligent identification of Logging data,the deep mining technology for lithology information is constantly improving.This article focuses on the problem of feature extraction from Logging data,and conducts in-depth research on the relationship between deep learning network models and reservoir lithology features.A deep learning network model suitable for reservoir lithology recognition,feature attention fusion network model,is proposed.The main work is carried out from the following three aspects:Firstly,the logging curve is preprocessed for the problem of field logging curve.In this paper,the original logging curve is mainly processed by depth correction and smooth filtering.The preprocessed logging curve is more accurate and smooth,and the error caused by the original data problem is eliminated.The sensitivity analysis experiment of logging curve is carried out,and the sensitivity verification analysis of the selected logging curve is carried out.It is found that different Logging data have different sensitivities to different rocks,and the reservoir lithology can be comprehensively identified by multiple Logging data.Secondly,the effects of three deep learning network models,VGG16,Goog Le Net and U-Net,in reservoir lithology identification were studied and compared.By comparing the accuracy,recall,and confusion matrix generated by model training of the three network models,the accuracy range of U-Net was 71% to 78%,with an average accuracy rate of74.86%,a recall rate range of 73% to 76%,and an average recall rate of 74.81%,The analysis shows that the U-Net model has higher accuracy and recall rate and smaller fluctuation range when identifying reservoir lithology.Finally,a reservoir lithology identification model based on U-Net model-FAF-Unet model is proposed.The U-Net model is selected as the basic network,and the defects of the U-Net model are mainly improved : Adding residual blocks in the downsampling process to reduce the ’ degradation ’ phenomenon.In order to improve the extraction of key information by the network model,the high-level channel information of feature extraction is combined with low-level features to increase the weight of some low-level feature channels,and a feature attention fusion block is proposed.FAF-Unet is a network model that combines residual blocks and channel attention mechanisms.It can fully mine the information between rock sedimentary sequences and feature channels,and can effectively improve the performance of reservoir lithology classification. |
