基于机器学习的高速铁路级配碎石压实质量主控特征与预测研究  

Main control features and prediction of compaction quality of graded crushed stones for high-speed railway subgrade based on machine learning

作　　者：肖宪普 李新志 谢康 郝哲睿 邓志兴 李泰灃 XIAO Xianpu;LI Xinzhi;XIE Kang;HAO Zherui;DENG Zhixing;LI Taifeng(School of Civil Engineering,Shijiazhuang University of Railways,Shijiazhuang 050043,China;School of Civil Engineering,Central South University,Changsha 410075,China;Railway Engineering Research Institute,China Academy of Railway Sciences Co.,Ltd.,Beijing 100081,China)

机构地区：[1]石家庄铁道大学土木工程学院,石家庄050043 [2]中南大学土木工程学院,长沙410075 [3]中国铁道科学研究院集团有限公司铁道建筑研究所,北京100081

出　　处：《振动与冲击》2024年第21期319-328,共10页Journal of Vibration and Shock

基　　金：河北省自然科学基金(E2021210124);河北省教育厅自然科学类项目(ZD2019034);高速铁路过渡段可调高路基结构研究(2021YJ071);中国国家铁路集团有限公司科技研究开发计划(P2021G053)。

摘　　要：完善高速铁路路基级配碎石压实标准对实现振动压实质量高精度智能预测具有重要意义。首先,开展振动压实试验,基于多参数协同测试方法,探究级配碎石最大干密度ρdmax确定方法;其次,在大量试验数据基础上建立级配碎石特征与ρdmax之间的关系,并采用灰色关联度分析算法明晰影响ρdmax的主控特征;最后,将级配碎石主控特征作为输入特征建立预测ρdmax的机器学习(machine learning,ML)模型,并基于ML模型预测性能三层次评价方法确定最优ML模型。结果表明:力学参数动刚度K rb曲线“拐点”对应的压实时间T lp为级配碎石最佳振动时间,进一步通过T lp确定级配碎石ρdmax;明晰影响级配碎石ρdmax的主控特征为最大粒径d max,级配参数b、m,扁平细长颗粒Q e以及洛杉矶磨耗LAA;综合三层次优选结果,各ML模型综合评价指标CEI由小到大分别为:ANN(artificial neural network)模型(1.8797)、SVR(support vector regression)模型(2.9646)、RF(random rorest)模型(4.5040)、Ridge(ridge regression)模型(6.2394)和DT(decision tree)模型(7.1319),ANN模型预测性能最优。研究成果可为高速铁路路基压实质量控制提供新标准,并对路基智能施工提供理论指导。Improving compaction standards for graded crushed stones of high-speed railway subgrade is of great significance for realizing high-precision intelligent prediction of vibration compaction quality.Here,firstly,vibration compaction tests were conducted to explore the method to determine the maximum dry densityρdmax of graded crushed stones based on a multi-parameter collaborative testing method.Secondly,based on a large amount of test data,the relation between features of graded crushed stones andρdmax was established,and the grey correlation analysis algorithm was used to clarify the main control features affectingρdmax.Finally,the main control features of graded crushed stones were taken as input features to establish a machine learning(ML)model for predictingρdmax,and the optimal ML model was determined based on the 3-level evaluation method of ML model prediction performance.The results showed that the compaction time T lp corresponding to the inflection point of mechanical parameter dynamic stiffness K rb curve is the optimal vibration time for graded crushed stones,and further T lp is used to determineρdmax of graded crushed stones;main control features affectingρdmax are the maximum particle size d max,gradation parameters b and m,flat and elongated particles Q e,and Los Angeles abrasion LAA;synthesizing 3-level optimization results,comprehensive evaluation index CEI values of various ML models from small to large are artificial neural network(ANN)model(1.8797),support vector regression(SVR)model(2.9646),random forest(RF)model(4.5040),ridge regression(Ridge)model(6.2394)and decision tree(DT)model(7.1319),ANN model has the optimal predictive performance;the study results can provide new standards for compaction quality control of high-speed railway subgrade,and theoretical guidance for intelligent construction of subgrade.

关 键 词：高速铁路 级配碎石 最大干密度 机器学习 三层次评估模型 

分 类 号：U213.1[交通运输工程—道路与铁道工程]