融合SMGC-ECAs-Resnet的致密砂岩岩相识别方法研究

Study on lithofacies identification method of tight sandstonewith fusion SMGC-ECAs-Resnet

针对目前深度学习模型通常只能提取单一尺度岩相特征,无法获得多尺度信息且没有充分适应测井曲线自身形态特点影响岩相识别的问题,基于深度学习以Resnet50为基础网络,设计开发多尺度特征提取模块SMGC(strip-pooling and multi-scale group convolution),并加入改进的ECAs(efficient channel attention strengthen)注意力模块增强对测井曲线条形纹理信息关注度,提出一种SMGC-ECAs-Resnet致密砂岩测井曲线岩相识别方法。以松辽盆地三肇凹陷扶余油层为例,对测井曲线数据预处理构建图像数据集,利用SMGC-ECAs-Resnet网络模型对其进行识别得到分类结果,设置对比试验和鲁棒性实验证明模型有效性。结果表明:所提出的SMGC-ECAs-Resnet网络岩相识别准确率达到91.9%,为复杂致密砂岩岩相的测井识别提供了新的方法。

The task of lithofacies identification and division is important in the exploration and evaluation of tight reservoirs.Currently,the deep learning model usually only extracts single-scale lithofacies features,which fails to obtain multi-scale information and does not fully adapt to the influence of the morphological characteristics of logging curves on lithofacies identification.Based on deep learning and Resnet50-based network,a multi-scale feature extraction module SMGC(Strip-pooling and Multi-scale Group Convolution)is designed and developed.An improved ECAs(Efficient Channel Attention Strengthen)attention module is added to propose a lithofacies identification method of SMGC-ECAs-Resnet tight sandstone logging curve.Fuyu oil layer in Sanzhao sag of Songliao Basin is taken as an example.First,the image data set is built by preprocessing the logging curve data.Then,the SMGC-ECAs-Resnet network model is employed to identify and obtain the classification results.Finally,the validity of the model is verified by comparative and robustness experiments.Our experiments show the proposed SMGC-ECAs-Resnet network reaches the optimal 91.9%in lithofacies recognition accuracy demonstrating its effectiveness in logging identification of complex tight sandstone lithofacies.