Prediction of surface chloride concentration of marine concrete using ensemble machine learning
Prediction of surface chloride concentration of marine concrete using ensemble machine learning
复制标题
DOI：
10.1016/j.cemconres.2020.106164
复制
发表时间：
2020-10-01
影响因子：
11.4
通讯作者：
Ma, Hongyan
中科院分区：
文献类型：
--
作者：
Cai, Rong;Han, Taihao;Ma, Hongyan
针对海洋环境下混凝土结构耐久性设计和使用寿命预测中的重要参数--混凝土表面氯离子浓度（C-s），提出了一种基于集成机器学习（ML）的预测模型。为此，一个数据库包含642个数据记录的现场暴露数据的C-s（沿着与相关的混合比例参数，环境条件和暴露时间）的基础上建立了广泛的文献调查，其中包括飞溅，潮汐，淹没区在世界各地。该数据库用于训练五个独立的ML模型，即线性回归（LR），高斯过程回归（GPR），支持向量机（SVM），多层感知器人工神经网络（MLP-ANN）和随机森林（RF）模型，以及基于集成加权投票的ML模型，并随后用于比较它们的预测性能。结果表明，通过将RF，MLP-ANN和SVM的预测进行元分析组合，与本研究中测试的所有独立ML模型相比，集成ML模型产生了更高的预测准确性。基于所选的ML测试数据集，分析了8种常用的C-s定量预测模型的预测性能。结果表明，采用更多样化的数据集和考虑更多的因素，在传统的模型可以提高其预测性能。建立在大型数据库上的集成ML模型，可以很容易地考虑数据库中的12个影响因素（这是常规模型难以做到的），并且与常规模型相比，具有上级的预测性能，同时具有更好的时间效率。
This paper develops and employs an ensemble machine learning (ML) model for prediction of surface chloride concentration (C-s) of concrete, which is an essential parameter for durability design and service life prediction of concrete structures in marine environment. For this purpose, a database containing 642 data-records of field exposure data of C-s (along with the associated mixture proportion parameters, environmental conditions and exposure time) is established based on extensive literature surveying, which covers splash, tidal, and submerged zones in various areas in the world. The database is used to train five standalone ML models, that is, linear regression (LR), Gaussian process regression (GPR), support vector machine (SVM), multilayer perceptron artificial neural network (MLP-ANN) and random forests (RF) models, as well as an ensemble weighted voting-based ML model, and subsequently used to compare their prediction performances. It is shown that, by metaheuristically combining predictions of RF, MLP-ANN, and SVM, the ensemble ML model produces higher accuracy of prediction compared to all standalone ML models tested in this study. The prediction performances of eight conventional quantitative models for C-s prediction are also analyzed based on the testing dataset selected for ML. The results show that adoption of more diverse datasets and consideration of more factors in conventional models can improve their prediction performance. The ensemble ML model established on a large database, can easily consider the twelve influencing factors (which is difficult for conventional models) in the database, and has superior prediction performance, yet better time-efficiency, compared to conventional models.