【作者】 张帆；

【导师】 何小海； 崔俊芳；

【作者基本信息】 四川大学 ， 电子与信息（专业学位）， 2022， 博士

【摘要】 油气储层岩心的微观三维结构直接影响其宏观物理性质(如储集空间、运移特性等),对油气勘探开发具有重要意义。数字岩心技术可以对岩心内部微观结构进行建模,并利用数值模拟等方法,定量分析岩心的各种特性。准确获取岩心微观三维结构,是模拟和分析岩心渗流特性和宏观物理性质的前提条件。近年来,利用岩心二维图像重建微观三维结构逐渐成为研究热点。传统的三维重建算法(如模拟退火算法、多点地质统计算法等)存在难以解决的问题,如重建效率低,重建尺度小,重建结构与真实结构形态相似性不足等。因此,进一步探索新的重建算法、提升样本重建尺度、提高样本重建的精度与效率,并且完成对非均质等特殊形态岩心的三维重建,是重建算法走向实际应用的关键。随着深度学习技术和理论的不断突破,其在图像特征自动提取及快速前向推断方面展现出较为明显的优势。其中,深度生成模型由于可以有效学习并表征高维图像数据的形态特征及数据分布,建立不同数据空间的映射关系,特别适用于解决从二维图像到三维结构的逆向反解问题。本文基于深度生成模型相关理论,围绕重建样本尺度扩展、三维重建范式、重建准确性与通用性以及非均质岩心重建等难点问题,提出基于生成对抗网络和循环神经网络的岩心三维重建模型,相关工作及创新点总结如下:(1)基于混合深度生成模型的岩心图像三维重建算法。该算法将变分自编码器与生成对抗网络相结合,解决了单纯使用生成对抗网络时,网络训练困难、生成样本尺度较小、模型易出现模式坍塌等问题。同时,模型完成对二维参考图像先验信息的特征学习及参数化表达,构建二维图像到三维结构的映射关系,将重建样本的尺度从64×64×64提升至128×128×128,有较好的训练稳定性。通过三维视觉观测、统计特征参数比较以及两相渗流仿真实验证明了模型的有效性。(2)基于循环神经网络的岩心图像逐层三维重建算法。该算法将循环神经网络应用于岩心微观图像的三维重建,将端到端的三维重建问题转化为基于二维参考图像的逐层预测重建问题,大幅优化了模型的解空间,提高了样本的重建尺度。模型充分利用自编码网络与循环神经网络的优点:一方面,利用自编码网络学习图像的形态特征,完成岩心图像空间与特征隐向量空间的相互转化;另一方面,循环神经网络利用隐空间的特征向量,学习图像间的层间关系,从而联合完成基于二维参考图像的三维结构逐层循环重建。该方法最少仅需一个三维样本即可完成对模型的训练,且训练过程稳定。实验表明,用该算法重建尺度为256×256×256的岩心三维结构时,其准确性、多样性及泛化性都有较好的表现。(3)基于特征图优化的深度循环生成重建模型。为了解决更大尺度(如512×512×512)均质岩心图像的三维重建问题,提出一种基于特征图优化的深度循环生成重建模型。该模型充分考虑均质岩心图像层间变化的稳定性,在跳跃连接层引入了一种轻量的卷积注意力模块,加强编码器特征图对解码器的引导。同时,提出一种基于VGG16的截面损失函数,增强生成图像深度方向的约束,提升生成模块对均质岩心图像的层间学习能力,优化模型在深度方向的重建性能。通过对尺度为512×512×512的均质岩心图像重建实验,验证了模型的有效性。(4)面向非均质岩心的逐层三维重建算法。目前,非均质岩心图像的三维重建研究还比较初步,其理论基础还不够完善。本文提出一种生成对抗网络与循环神经网络融合的逐层重建算法。该算法融合了生成对抗模型与逐层重建模型的优点,通过构建生成对抗损失,使模型更好地感知非均质岩心图像的整体形态特征,提高模型生成能力。同时,鉴于非均质图像层间变化的非平稳性与随机性,弱化编码器对于解码器的确定性引导,同时在解码器内部增加自我跳连模块。这一方面使解码器可以学习到上一帧图像的主体形态特征,同时增加了生成图像局部孔隙形态的多样性与随机性。通过在非均质岩心上开展相关的重建实验,表明该模型可以较好地学习到非均质岩心图像的局部及全局孔隙形态特征,验证了算法的有效性。更多还原

【Abstract】 The micro-structure of core can affect its macro physical properties directly(e.g.,reservoir space and transfering property).It is of great significance to oil and gas exploration and development.The digital core technique can model the mircro-sturcture of core and analyse the physical properties quantitatively by using numerical sitimulation method.To access the three-dimenstional(3D)micro-strucuture of core accurately is a prerequisite to model and analyse the transferring and macro physical properties.In recent years,reconstructing 3D microstructure from2D core images has gradually become a research hotspot.The traditional reconstruction methods(e.g.,Simulated annealing or Multi-point geostatistic)have some problems that are difficult to solve,such as low reconstruction efficiency,small reconstruction scale,insufficient morphological similarity between the reconstructed structure and the real 3D structure,etc.Therefore,the key to the practical application of the reconstruction algorithm is to further explore new reconstruction algorithms,improve the scale of synthetic realizations,improve the accuracy and efficiency of reconstruction,and complete the 3D reconstruction of heterogeneous core.With the continuous breakthrough of deep learning technology and theory,it shows obvious advantages in automatic image feature extraction and fast forward inference.Among them,the deep generative model can effectively learn and characterize the morphological characteristics and data distribution of high-dimensional image data,and establish the mapping relationship between different data spaces.It is especially suitable for solving the problem of reverse reconstruction from two-dimensional images to three-dimensional structures.Based on such theory,this thesis focuses on some difficult problems,such as the scale expansion,the paradigm of 3D reconstruction,reconstruction accuracy and efficiency and the heterogeneous media reconstruction.The generative adversarial network(GAN)and recurrent neural network(RNN)based models are proposed.The related work is summarized as follows:(1)A hybrid deep generative model for 3D reconstruction of core image.This model combines the advantages of GAN and variational autoencoder together to cope with the problems of difficulty in training,small scale of generated samples and model collages problems when applying the GAN alone.Meanwhile,the prposed model learns the spatial features of piror reference 2D image and establish the mapping between 2D image and 3D structure.The proposed model is not sensitive to the hyper-parameters and has a good performance on training stability.Thus,the size of reconstructed samples is increased from 64×64×64 to 128×128×128.Evaluated by 3D visual inspection,statistic comparison and two-phase permeability simulation experiment,the effectiveness of proposed model is verified.(2)A layer-by-layer 3D reconstruction model based on RNN.The RNN is applied to solve the 3D reconstruction problem of digital core and transforms the end-to-end 3D reconstruction problem into a layer-by-layer prediction problem based on the 2D reference image.The algorithm greately optimizes the solution space of the model and improves the scales of synthetic realizations.The algorithm integrates autoencoder and recurrent neural network together to makes use of the advantages of the thoses models effectively.On one hand,the model uses autoencoder to learn the morphological features of the image,and completes the transformation between the image space and the feature vector space.On the other hand,the recurrent neural network makes use of the feature vectors from the hidden space to learn the spatial relationship between layers.Thus,the 2D-to-3D reconstruction is completed due to the cyclically generation.For the training of the proposed model,only 1 training sample is need at least and the stability is enhanced as well.Experiments show that the algorithm has good accuracy,diversity and generalization when reconstructing the3D structure of core with the size of 256×256×256.(3)The deep recurrent generative model based on the optimization of feature maps.In order to reconstruct a 3D realization with larger scale(e.g.,512×512×512),a deep recurrent generative model based on feature map optimization is proposed.Considering the stable variation of homogeneous porous media between adjacent layers,a lightweight convolution attention module,namely CBAM,is introduced to the skip-connection module to strengthen the guidance of encoder’s feature map to decoder.At the same time,a section loss based on VGG-16 model is proposed to improve the learning ability of generative model for homogeneous core images and enhance the reconstruction performance of the realization along depth direction.By conducting the reconstruction experiments on core image with the scale of 512×512×512,the effectiveness of the proposed model is proved.(4)A layer-by-layer deep generative model for heterogeneous core.At present,the research on 3D reconstruction of heterogeneous core image is still preliminary,and its theoretical basis is not perfect.In this chapter,a layer-by-layer reconstruction model based on GAN and RNN is proposed to make use of the advantages from the both models.By constructing the adversarial loss,the proposed model can better perceive the overall morphological characteristics of heterogeneous 3D structure and improve its learning ability between layers.Meanwhile,in view of the non-stationarity and randomness between layers of heterogeneous images,the proposed model weakens the guidance of feature map from encoder to the decoder and introduces the self-connection within the decoder.With this improvement,the decoder can learn the main morphological features of the previous frame,and increase the diversity and randomness of the local pore of the generated image.The reconstruction experiments carried on heterogeneous core show that the model can better learn the local and global pore morphological characteristics of heterogeneous cores,which proves the effectiveness of the algorithm.更多还原