摘要
针对现有基于机器学习的裂隙网络骨架方法仅通过裂隙自身的拓扑属性和物理属性来识别关键裂隙,没有考虑邻居裂隙的重要程度,且无法控制所提取骨架裂隙的规模及连通性等问题,提出一种基于多聚合器图神经网络的裂隙网络骨架提取方法。首先,多聚合器图神经网络利用多个聚合器汇聚邻居裂隙信息,生成归纳嵌入表示,以此评估每个裂隙的重要程度;然后,利用可控骨架规模及连通性的贪婪算法完成裂隙骨架提取。实验结果表明,基于多聚合器图神经网络的裂隙网络骨架识别算法性能优于现有方法,在骨架规模为35%时,统计量K_(s)和散度值K_(L)分别为0.09±0.03和0.02±0.01;骨架规模超过10%时,识别出的裂隙骨架网络特性与原始裂隙的网络特性非常相似。骨架规模从10%降到5%时,骨架突破曲线同原始网络的差异急剧变化。通过控制裂隙的骨架规模,可以实现骨架提取的准确性与效率间的平衡。
To address the limitations of existing machine learning-based methods for fracture network backbone extraction,which identify key fractures only through topological and physical properties,ignore the importance of neighboring fractures,and cannot control the scale and connectivity of extracted skeleton fractures,a fracture network skeleton extraction method based on multi-aggregator graph neural network(MA-GNN)was proposed.MA-GNN first employs multi-aggregators to integrate features from neighboring fractures and generates inductive embeddings to assess the importance of each fracture.Then,a greedy algorithm was proposed to perform skeleton extraction while controlling its scale and connectivity.The experimental results show that the performance of the fracture network skeleton recognition algorithm based on multi-aggregator graph neural network is superior to existing methods.When the skeleton scale is 35%,the statistical K_(S) and divergence value K_(L) are 0.09±0.03 and 0.02±0.01,respectively.When the backbone scale exceeds 10%,the network characteristics of the identified fracture skeleton are very similar to those of the original fracture network.When the skeleton scale decreases from 10%to 5%,the difference between the skeleton breakthrough curve and the original network changes sharply.By controlling the size of the fracture skeleton,a balance can be achieved between the accuracy and efficiency of skeleton extraction.
作者
邱家川
杜海涛
郑天骥
王志晓
Qiu Jiachuan;Du Haitao;Zheng Tianji;Wang Zhixiao(Henan Zhenglong Coal Industry Co.,Ltd.,Yongcheng 476600,China;School of Computer Science and Technology,China University of Mining Technology,Xuzhou 221116,China)
出处
《能源与环保》
2025年第11期15-20,共6页
CHINA ENERGY AND ENVIRONMENTAL PROTECTION
基金
国家自然科学基金面上项目(61876168)。
关键词
离散裂隙网络
骨架裂隙
图神经网络
多聚合器
贪婪算法
discrete fracture networks
skeleton fractures
graph neural networks
multi-aggregators
greedy algorithm
