摘要
为寻求一种预测速度快、准确率高的城市轨道交通地下线振动源强预测模型,基于55个非减振轨道测试断面数据,经过数据清洗、分析和标签化后,建立了涵盖典型车型和主要线路参数取值范围的8 000多条实测数据库。分析地铁环境振动的影响因素,利用斯皮尔曼相关系数得到各类影响因素与振动源强的关系强度。分别建立基于卷积神经网络(CNN)、随机森林(RF)、支持向量机(SVM)等5个机器学习模型,对比分析了不同模型对振动源强的预测效果。使用麻雀搜索算法(SSA)和遗传算法(GA)优化BP神经网络模型的结构、超参数、权重及阈值,对比SSA-GA-BP、SSA-BP、GA-BP神经网络对振动源强的预测精度。最终使用4个差异明显且未经模型学习的新断面验证SSA-GA-BP模型的泛化能力。结果表明:5种机器学习模型中BP神经网络的非线性回归拟合能力最强,验证集MAE损失为1.55 dB,决定系数为0.948;SSA-GA-BP模型对振动源强的预测精度高于SSA-BP和GA-BP,验证集MAE、MAPE和决定系数分别为1.289 dB、1.856%和0.967,有80.11%数据的平均绝对误差在2 dB以内;SSA-GA-BP模型对4个经典的新断面数据预测效果良好,4个断面汇总数据的MAE、MSE和MAPE误差值分别为1.21 dB、2.18 dB和1.67%,决定系数为0.977,有70%数据的预测误差在2 dB以内,证明了SSA-GA-BP模型有较强的泛化能力。SSA-GA-BP振源预测模型具有较好的预测精度和快速预测能力,研究可为轨道交通地下线路设计阶段的减振降噪设计提供参考。
To develop a vibration source strength prediction model for urban underground rail transit lines that offers fast prediction speed and high accuracy.Based on data from 55 non-damping track test cross-sections,a database of over 8000 measured entries was established after data cleaning,analysis,and labeling,covering typical vehicle types and main line parameter ranges.The influencing factors of environmental vibration in subways were analyzed using Spearman correlation coefficients to determine the strength of the relationship between various factors and vibration source strength.Five machine learning models—including Convolutional Neural Networks(CNN),Random Forest(RF),and Support Vector Machine(SVM)—were separately established to compare and analyze their prediction performance.The Sparrow Search Algorithm(SSA)and Genetic Algorithm(GA)were used to optimize the structure,hyper-parameters,weights,and thresholds of the Backpropagation(BP)neural network model.The prediction accuracies of SSA-GA-BP,SSA-BP,and GA-BP neural networks were compared.Finally,four new cross-sections with significant differences and not used in model training were employed to verify the generalization ability of the SSA-GA-BP model.The results show that among the five machine learning models,the BP neural network has the strongest nonlinear regression fitting ability,with a validation set MAE loss of 1.55 dB and a coefficient of determination of 0.948.The SSA-GA-BP model's prediction accuracy on vibration source strength is higher than that of SSA-BP and GA-BP,with validation set MAE,MAPE,and coefficient of determination being 1.289 dB,1.856%,and 0.967,respectively.The 80.11%of the data have a mean absolute error within 2 dB.The SSA-GA-BP model showed good prediction performance on four classical new cross-section data,with aggregated MAE,MSE,and MAPE error values of 1.21 dB,2.18 dB,and 1.67%,respectively,and a coefficient of determination of 0.977.The 70%of the data have prediction errors within 2 dB,demonstrating the model's strong generalization ability.The SSA-GA-BP vibration source prediction model exhibits good prediction accuracy and rapid prediction capability.This research can provide a reference for vibration reduction and noise control design during the design phase of underground rail transit lines.
作者
刘庆杰
刘博亮
冯青松
徐璐
罗信伟
刘文武
LIU Qingjie;LIU Boliang;FENG Qingsong;XU Lu;LUO Xinwei;LIU Wenwu(State Key Laboratory of Performance Monitoring and Protecting of Rail Transit Infrastructure,East China Jiaotong University,Nanchang 330013,China;Guangzhou Metro Design&Research Institute Co.,Ltd.,Guangzhou 510010,China)
出处
《铁道科学与工程学报》
北大核心
2025年第5期2355-2366,共12页
Journal of Railway Science and Engineering
基金
国家自然科学基金资助项目(51868024)。
关键词
城市轨道交通地下线
振动源强
预测
BP神经网络
麻雀搜索算法
遗传算法
urban rail transit underground line
vibration source strength
prediction
BP neural network
sparrow search algorithm
genetic algorithm
