基于不同机器学习模型的框架结构损伤识别对比研究  

Comparative study of frame structure damage identification based on different machine learning models

作　　者：滕辉 康帅[1] 尹俊红[1] 王永锋[1] TENG Hui;KANG Shuai;YIN Junhong;WANG Yongfeng(School of Civil Engineering and Architecture,Henan University,Kaifeng 475004,China)

机构地区：[1]河南大学土木建筑学院,河南开封475004

出　　处：《四川建筑科学研究》2023年第3期44-53,共10页Sichuan Building Science

基　　金：河南省高等学校重点科研项目(21A560005)。

摘　　要：针对框架结构损伤特征建立了不同的机器学习模型,对不同模型下的损伤识别结果进行了对比分析。以5层钢框架结构为例,以模态曲率差作为数据集,分别建立了支持向量机(support vector machine,简称SVM)模型、Bagging模型、BP神经网络模型、卷积神经网络(convolutional neural network,简称CNN)模型4种机器学习模型,并通过贝叶斯法和经验法对模型进行了超参数优化,对框架结构的单损伤和多损伤工况进行了损伤位置和损伤程度的识别,在此基础上将数据集进行了噪声处理,分析了4种模型的抗噪能力。结果表明:在无噪声影响下,4种模型对框架结构的损伤识别都有较高的精度,其中Bagging模型在损伤程度识别中的精度相对其他3种模型较差;在数据集加入噪声后,CNN模型相较于其他3种模型损伤识别正确率下降幅度较小,说明CNN模型相较于其他3种模型具有较好的抗噪性。Different machine learning models were established for frame structure damage characteristics,and the damage identification results under different models were compared and analyzed.Taking the five-story steel frame structure as an example and the modal curvature difference as the data set,four machine learning models,namely,support vector machine(SVM)model,Bagging model,BP neural network model and convolutional neural network(CNN)model were established.The models were optimized by the Bayesian method and the empirical method,and the damage location and damage degree of the frame structure under single damage and multiple damage conditions were identified.On this basis,the data set was processed with noise,and the anti-noise ability of the four models was analyzed.The results show that the four models have high accuracy for damage identification of frame structures without noise,and the accuracy of the Bagging model in damage degree identification is slightly worse than that of other models.After adding noise to the data set,the CNN model has a smaller decline in the accuracy of damage identification compared with the other three models,indicating that the CNN model has better noise immunity than the other three models.

关 键 词：结构损伤 机器学习 模态曲率 支持向量机 BAGGING BP神经网络 卷积神经网络 

分 类 号：TU317[建筑科学—结构工程]