基于机器学习的混凝土侵彻深度预测模型

The test data about concrete penetration are often limited in quantity and unevenly distributed, which leads to the poor accuracy of machine learning-based models for predicting the penetration depth. This study aims to mitigate the unevenness of data distribution and increase the amount of available data to obtain an optimal machine learning model under the limitation of limited penetration test data. Based on collecting a large amount of penetration test data, the penetration data are extended by data augmentation methods such as linear interpolation and adding Gaussian noise. The genetic algorithm and greedy algorithm are used to optimize the hyperparameters of four common machine learning models: multilayer perceptron, radial basis neural network, support vector regression and extreme gradient boosting tree. The prediction of concrete penetration depth based on machine learning is realized. Sensitivity analysis method is used to analyze the influence of input factors on the penetration depth. The results show that the problem of insufficient data can be effectively alleviated by using linear interpolation and adding Gaussian noise. The accuracies of multilayer perceptron, radial basis neural network, support vector regression and extreme gradient boosting tree are improved by 2. 49%, 0. 99%, 0. 74%, and 0. 72%, respectively, after using data augmentation. The diameter, impact velocity and mass of projectile have the dominant influence on penetration depth. In addition, the average error of the optimal concrete penetration depth prediction machine learning model is 8. 28%, and its global accuracy is better than the commonly used empirical formulas for predicting the penetration depth.