摘要
为实现建筑结构安全的快速评估,提出基于神经网络的建筑结构安全评估方法。基于《民用建筑可靠性鉴定标准》的调查与检测要求并考虑数据易获取性,选择45个涵盖承载力、耐久性、历史记录和环境情况等变量作为输入参数,以《民用建筑可靠性鉴定标准》中的安全等级作为输出参数,采用深度置信网络学习输入参数与输出参数间的非线性映射关系。对输入参数的选择、样本缺值问题、小样本问题和神经网络评估的可靠性进行探讨和验证。结果表明:在无法准确判断输入参数与输出参数相关性的前提下,采用全部输入参数的评估模型具有更高的鲁棒性;迷失森林算法相较其他常用的缺值插补算法有更好的插补性能;采用变分自编码器扩充训练样本集能有效提高神经网络的泛化能力和分类精度;对深度置信网络引入加权交叉熵损失函数加以改进可增加训练时对不安全类别的敏感性,牺牲少量不安全类别的查准率可以大幅提高其查全率;基于神经网络的结构安全评估模型能较好地预测结构的安全等级,具有快速且大批量运算的优势,是实现大范围建筑群结构安全监测的有效手段。
To achieve fast civil structure safety assessment, safety assessment model of building structure based on neural network was proposed. Based on the requirements of inspection and investigation of Standard for Appraisal of Reliability of Civil Buildings, 45 input parameters included bearing capacity, durability, historical condition, and environmental situation were selected as the input parameter. The safety level in the Standard for Appraisal of Reliability of Civil Buildings was selected as the output parameter. Deep belief network was adopted to study the nonlinear map relationship between the input and output parameters. The selection of input parameter, the problem of missing data, small sample problem and the reliability of neural network were discussed and validated. The results show that the evaluation model using all input parameters has higher robustness under the premise that the correlation between input parameters and output parameters cannot be accurately judged. Miss forest algorithm has better interpolation performance than other commonly used missing interpolation algorithms. Using variational autoencoder to expand training dataset can effectively improve the generalization ability and classification accuracy of neural network. Improved deep belief network by introducing weighted cross entropy loss function can increase the sensitivity of training to unsafe categories, and the recall rate can be greatly improved by sacrificing a small amount of precision rate of unsafe categories. The safety assessment model of building structure based on neural network is capable to predict the safety level of the structure. With the fast and mass operation ability, it is an effective mean to realize structural safety monitoring of large-scale buildings.
作者
王俊杰
焦柯
彭子祥
WANG Jun-jie;JIAO Ke;PENG Zi-xiang(Guangdong Architectural Design&Research Institute Co.,Ltd.,Guangzhou 510010,Guangdong,China)
出处
《建筑科学与工程学报》
CAS
北大核心
2022年第4期174-182,共9页
Journal of Architecture and Civil Engineering
基金
住房和城乡建设部科学技术计划项目(2019-K-157)。
关键词
结构安全评估
神经网络
小样本问题
缺失数据插补
查准率
查全率
structural safety assessment
neural network
small sample problem
missing value imputation
precision rate
recall rate
