摘要
我国现行规范中的碳钢大气腐蚀预测模型准确度较低,亟需提出更准确的预测模型。针对我国典型环境地区各类钢材,收集10个大气暴露试验站腐蚀数据,构建大气腐蚀数据集,提出基于集成图神经网络(BGNN)的钢材大气腐蚀预测方法,即关联特征相似的数据点,构成结构化数据集。通过随机抽样,在结构化数据集基础上生成结构化采样子数据集和余特征数据集。训练预先构建的基本图神经网络(GNN)回归器,并将若干回归器集成得到BGNN模型。采用不同方法预测我国钢材大气腐蚀数据集并进行对比,结果表明,规范公式预测误差最大,SVM和LSTM预测效果较好,BGNN的预测误差最低,预测效果最好。相比单一GNN,BGNN预测误差更小,对超参数的选取相对不敏感。BGNN模型能够充分考虑不同地区、不同钢材和不同气候环境条件间的相似性,增强对关联数据的分析能力,使对新预测目标归因分析更明确、预测准确率更高。
The accuracy of the atmospheric corrosion prediction model for carbon steel in our current standards is low,and it is urgent to propose a more accurate prediction model.Aiming at various types of steel in typical environmental regions of China,the corrosion data obtained from 10 atmospheric exposure test stations are collected.The atmospheric corrosion dataset is constructed,and an atmospheric corrosion prediction method for steel structures based on the batched graph neural network(BGNN)is proposed.The method associates data points with similar features to form a structured dataset,generates a structured sampling dataset and a residual feature dataset based on the structured dataset through random sampling.Trains a pre⁃constructed basic graph neural network(GNN)regressor,and integrates a number of regressors to obtain a BGNN model.Different methods are used to predict the atmospheric corrosion dataset of steel in China and compared.The results show that,the formula in current standards has the largest prediction error.SVM and LSTM have better prediction effect.BGNN has the lowest prediction error and the best prediction effect.Compared with single GNN,BGNN has smaller errors and is relatively insensitive to the selection of hyperparameters.The BGNN model is capable of sufficiently considering the similarity between different regions,different steels and different climatic conditions,and enhancing the ability to analyze the correlated data,which results in a clearer attribution analysis of the new prediction targets and a higher prediction accuracy.
作者
吴柯娴
何雨宸
刘博扬
何晓宇
金伟良
WU Kexian;HE Yuchen;LIU Boyang;HE Xiaoyu;JIN Weiliang(Institute of Structural Engineering,Zhejiang University,Hangzhou,Zhejiang 310058,China;Zhejiang Institute of Communications Co.,Ltd.,Hangzhou,Zhejiang 310006,China;Yangtze Delta Industrial Innovation Center of Quantum Science and Technology,Suzhou,Jiangsu 215000,China)
出处
《施工技术（中英文）》
CAS
2024年第17期115-120,共6页
Construction Technology
基金
国家自然科学基金面上项目(52178176)
浙江省交通运输厅科技计划(2020003,2023007)。
关键词
钢材
图神经网络
大气腐蚀
深度学习
数据集
steels
graph neural network
atmospheric corrosion
deep learning
dataset
