基于深度学习的煤系页岩孔隙结构定量表征

Quantitative characterization of pore structure in coal measure shales based on deep learning

储层孔隙结构是页岩气勘探开发的重要影响因素。为准确表征煤系页岩储层中纳米孔隙结构,以青海木里地区煤系页岩为研究对象,利用扫描电镜采集页岩孔隙结构图像,建立页岩孔隙图像数据集,并基于深度学习技术设计出针对页岩孔隙图像分割的HAFCN模型。将孔隙识别效果与3种经典语义分割模型(FCN模型,U-Net++模型,OCRNet模型)做对比,结果表明:HAFCN模型分割效果明显占优,其平均交并比(mIoU)达到0.8576,像素准确率达到0.9702,实现了快速分析页岩孔隙扫描电镜图像的目的,并获得了孔隙结构各项参数。将识别后的孔隙参数与原始孔隙参数值(Ground-truth)对比,发现两者孔隙结构参数相近,证实了模型的可靠性;所测煤系页岩样品的孔径以小孔及中孔为主;小孔、中孔及大孔孔径段的平均形状因子分别为1.65、2.38、4.10,其平均长宽比分别为2.97、2.76、3.01,说明随着页岩孔隙的增大,孔隙形态越偏离理想球形,形状越不规则。

Reservoir pore structure is an important factor affecting shale gas exploration and development.In order to accurately characterize the nanopore structure in coal-measure shale reservoirs,this paper takes coal-measure shale in the Muli area of Qinghai province as the research object,and uses shale pore structure images collected from scanning electron microscopy(SEM)to establish a shale pore image data set.A semantic image segmentation model named HAFCN(Hypercolumns Attention Fully Convolutional Networks)was proposed for shale pore segmentation based on deep learning technology.Compared with other three classical semantic segmentation models(FCN,U-Net++,OCRNet models)for pore images recognition,the HAFCN`model had better pore recognition results than other models,with an average intersection-over-union ratio(mIoU)of 0.8576 and a pixel accuracy of 0.97,so that the purpose of rapid analysis of shale pore SEM images was achieved,and various parameters of pore structure was obtained.Compared with the identified pore parameters with the original pore parameter values(Ground-truth),it is found that the pore structure parameters of the two are similar,which confirms the reliability of the model.The average shape factors of small,medium and large pore diameter sections are 1.65,2.38,and 4.10,respectively,and their average aspect ratios are 2.97,2.76,and 3.01,respectively,indicating that with the diameter increase of shale pores,the pore shape is more irregular.