基于多分类器协同训练的结构损伤识别
Structural Damage Identification Based on Multi-Classifier Cooperative Training
秦世强 1杨睿 1苏晟1
作者信息
- 1. 武汉理工大学土木工程与建筑学院 武汉 430070
- 折叠
摘要
文中提出一种基于多分类器协同训练(multi-classifiers co-training,MCCT)的结构损伤识别框架.该框架结合多层感知机(multilayer perceptron,MLP)和支持向量机(support vector ma-chine,SVM)进行协同训练,从无标签样本中挑选置信度高的样本标注伪标签,扩大样本训练集,并采用加速度响应的功率谱密度(PSD)作为输入特征,用于识别结构损伤.结果表明:协同训练方法能够从无标签样本中选取置信度较高的样本,为损伤识别提供更充足有标签样本.相较于MLP和SVM,该方法在多种工况下,损伤识别准确率分别提升约4.7％和6.3％.
Abstract
A structural damage identification framework based on multi-classifiers co-training(MCCT)was proposed.The framework combined multilayer perceptron,MLP)and support vector machine,SVM)for collaborative training,and selected samples with high confidence from unlabeled samples to label pseudo-labels,thus expanding the sample training set.The power spectral density(PSD)of ac-celeration response was used as the input feature to identify structural damage.The results show that the collaborative training method can select samples with high confidence from unlabeled samples and provide more labeled samples for damage identification.Compared with MLP and SVM,the damage i-dentification accuracy of this method is improved by about 4.7％and 6.3％respectively under various working conditions.
关键词
结构损伤识别/协同训练/半监督学习/伪标签/神经网络Key words
damage identification/co-training/semi-supervised learning/pseudo label/neural network引用本文复制引用
出版年
2024
NETL
NSTL
万方数据