GA-XGBoost模型对路基压实质量的预测

Prediction of dynamic deformation modulus of subgrade based on GA-XGBoost model

摘要:

为提升智能压实(intelligent compaction,IC)质量的实时检测与评价精度,提出一种基于GA-XGBoost模型的连续压实质量预测方法,以提高动态变形模量(E_vd)的预测精度。模型以动态变形模量为目标,建立机器学习模型,主要采用决策树算法,构建XGBoost模型对压实质量进行预测分析。通过引入遗传算法(genetic algorithm,GA)对模型超参数寻优,以提高模型的预测精度和可靠性。首先,通过现场工程试验,测量压路机碾压时振动加速度,分析加速度信号,计算信号统计量并采用快速傅里叶变换（FFT）得出谐波频率,初步建立各项特征因子与E_vd之间的系统联系；其次,筛选各个时频域特征,进行相关性分析,选用相关性较高的特征来建立预测模型；最后,验证了GA-XGBoost预测模型可以较好的预测E_vd。研究结果表明:遗传算法(GA)可以高效地确定XGBoost算法的超参数,且较单一的XGBoost模型表现出更优的收敛速度；通过优化特征因子,改变输入参数,提高了GA-XGBoost模型的预测精度,优化后均方误差为3.9%,相关系数为0.748；同时对比了传统CMV拟合E_vd的方法,该机器学习模型可以大幅度提高预测精度。

Abstract:

In order to improve the real-time detection and evaluation accuracy of intelligent compaction (IC) quality, a continuous compaction quality prediction method based on GA-XGBoost model was proposed to improve the prediction accuracy of dynamic deformation modulus (E_vd). The model takes the dynamic deformation modulus as the goal, establishes a machine learning model, mainly uses the decision tree algorithm, and constructs the XGBoost model to predict and analyze the compaction quality. In order to improve the prediction accuracy and reliability of the model, genetic algorithm (GA) is introduced to optimize the hyperparameters of the model. Firstly, through the field engineering test, the vibration acceleration of the roller is measured, the acceleration signal is analyzed, the signal statistics are calculated and the harmonic frequency is obtained by fast Fourier transform (FFT), and the system connection between the characteristic factors and E_vd is preliminarily established. Secondly, the characteristics of each time-frequency domain are screened, the correlation analysis is carried out, and the characteristics with high correlation are selected to establish the prediction model. Finally, it is verified that the GA-XGBoost prediction model can better predict E_vd.The results show that the genetic algorithm (GA) can efficiently determine the hyperparameters of the XGBoost algorithm, and it shows better convergence speed than the single XGBoost model. By optimizing the feature factors and changing the input parameters, the prediction accuracy of the GA-XGBoost model is improved. The optimized mean square error is 3.9% and the correlation coefficient is 0.748. At the same time, compared with the traditional CMV fitting E_vd method, the machine learning model can greatly improve the prediction accuracy.