基于图神经网络的路基智能决策算法

Intelligent Railway Subgrade Decision-Making Algorithm Based on Graph Neural Networks

随着智能化技术的发展,路基决策的智能化水平提升成为可能。本研究通过图神经网络(GNN)结合专家知识构建图本体,探索其在路基设计智能决策中的应用,以提高决策的准确性和效率。本文在铁路路基设计领域中的地基加固方案设计数据集上训练了GNN模型,并进行了预测效果验证。然后,使用GNNExplainer对模型决策过程中的影响因素权重进行分析,揭示决策逻辑。研究结论为:(1)GNN在地基加固方案预测中达到约75%的准确率,且常用方案的准确率可达85%;(2)采用GNNExplainer算法对神经网络权重进行分析,实现了对决策逻辑的深层次的解释,为理解具体设计逻辑提供了全新的视角;(3)采用热力图可视化与实际经验结合进行相关性分析,证实了设计方案的多样性并揭示了算法的决策逻辑。本文的研究成果提升了陆路交通领域路基设计的智能决策水平,为相关领域的决策系统提供了有力的技术支持,为后续的智能决策算法奠定了基础。

With the advancement of intelligent technologies,enhancing the level of intelligent decision-making in roadbed construction has become feasible.This study aims to utilize Graph Neural Networks(GNN)combined with expert knowledge to construct a graph ontology,exploring its application in intelligent decision-making for roadbed design to improve decision accuracy and efficiency.In this study,a GNN model was trained on a dataset involving subgrade reinforcement scheme design in the railway roadbed design domain and its prediction effectiveness was validated.Subsequently,the GNNExplainer was used to analyze the weight of influencing factors in the model's decision-making process,revealing the logic behind the decisions.Research results are as follows:(1)GNN achieved an accuracy rate of approximately 75%in predicting ground reinforcement schemes,with common schemes reaching an accuracy of 85%;(2)The use of the GNNExplainer algorithm for analyzing neural network weights enabled a deep interpretation of the decision-making logic,offering a new perspective for understanding the specific design logic;(3)The combination of heatmap visualizations with practical experience for correlation analysis confirmed the diversity of design schemes and revealed the decisionmaking logic of the algorithm.The results of this study have enhanced the level of intelligent decision-making in the field of land transportation roadbed design,offering robust technical support for decision-making systems in related areas and laying a foundation for subsequent intelligent decision-making algorithms.