基于PSO-ML-AdaBoost模型的级配碎石最优压实参数智能预测研究  

Research on intelligent prediction of optimal compaction parameters for graded gravel based on PSO-ML-AdaBoost model

作　　者：陈晓斌[1] 郝哲睿 谢康 闫宏业 李泰灃 尧俊凯 邓志兴 CHEN Xiaobin;HAO Zherui;XIE Kang;YAN Hongye;LI Taifeng;YAO Junkai;DENG Zhixing(School of Civil Engineering,Central South University,Changsha 410075,China;Railway Engineering Research Institute,China Academy of Railway Sciences Corporation Limited,Beijing 100081,China)

机构地区：[1]中南大学土木工程学院,湖南长沙410075 [2]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：《铁道科学与工程学报》2024年第12期5042-5056,共15页Journal of Railway Science and Engineering

基　　金：国家自然科学基金资助项目(51978674);中国铁道科学研究院科技研究开发计划(2023QT002)。

摘　　要：为实现高铁路基级配碎石填料最优压实参数快速准确的确定,对填料的最优压实参数及其智能预测展开研究。首先,基于共振作用下振动压实参数确定方法,综合压实物理和力学指标得到级配碎石填料最优压实状态下的最优频率f_(op)和最优含水率w_(op);其次,通过填料性能试验建立级配碎石填料特征与f_(op)和w_(op)的关系,并采用灰色关联度分析算法明确影响f_(op)和w_(op)的主控特征;最后,将主控特征作为输入特征建立预测f_(op)和w_(op)的3种典型机器学习(Machine Learning,ML)模型,并融合Ada Boost算法解决基础ML算法的不足,建立PSO-ML-Ada Boost模型。结合三层次预测模型评价体系确定最优预测模型,并基于消融分析进一步验证最优预测模型的可靠性。结果表明:取w_(op)为临界含水率,f_(op)为填料的固有频率,可获得级配碎石填料压实状态最优的试样;揭示影响f_(op)和w_(op)的主控特征为最大粒径d_(max)、级配参数b和m,粗骨料细长比Ei、洛杉矶磨耗L_(aa)、吸水率W_(ac)和W_(af);综合三层次评价结果,得到PSO-BPNN-Ada Boost模型的综合评价指标Cei(f_(op)/w_(op))值为12.2645/1.8382,低于其他ML融合算法,为最优预测模型;结合消融分析结果发现,PSO-BPNN-Ada Boost模型的输入参数对于f_(op)和w_(op)预测结果的影响程度与灰色关联度分析算法所得结果一致,进一步说明最优预测模型预测结果的可靠性。研究成果可为路基填料最优压实参数的确定提供新思路,并对高铁路基的压实质量智能评估提供理论指导。To achieve the rapid and accurate determination of optimal compaction parameters for graded gravel in high-speed railway subgrades,the research was conducted on optimal compaction parameters and the intelligent prediction of fillers.First,based on the method of determining vibration compaction parameters under resonance,the optimal frequency(f_(op))and optimal water content(w_(op))for graded gravel under the optimal compaction state were obtained by combining the compaction physical and mechanics indicators.Second,the relationship between the features of graded gravel and f_(op)/w_(op)was established through filler performance tests.The Grey Relational Analysis(GRA)algorithm was applied to determine the key features influencing f_(op)and w_(op).Lastly,the dominant features were used as input parameters to establish three typical Machine Learning(ML)models for predicting f_(op)and w_(op).The PSO-ML-AdaBoost model was established using the AdaBoost algorithm to address limitations in basic ML algorithms.The optimal prediction model was determined based on a three-level assessment system for prediction model,and the reliability of the optimal prediction model was further verified through ablation analysis.The results indicated that setting w_(op)as the critical water content and f_(op)as the inherent frequency of the filler can optimize the compacted state of graded gravel.The dominant features influencing f_(op)and w_(op)included the maximum particle size(d_(max)),grading parameters(b,m),coarse aggregate aspect ratio(Ei),Los Angeles abrasion(L_(aa)),and water absorption rates(W_(ac),W_(af)).Based on the comprehensive evaluation results of the three-level assessment system,the comprehensive evaluation index(Cei)values for predicting f_(op)/w_(op)using the PSO-BPNN-AdaBoost model were 12.2645/1.8382,which were lower than those of other ML integration algorithms,suggesting it as the optimal predictive model.Combining the results of the ablation analysis,it was indicated that the input parameters of the PSO-BPNN-AdaBoo

关 键 词：高铁级配碎石 振动压实 主控特征 机器学习 消融分析 

分 类 号：U213.1[交通运输工程—道路与铁道工程]