This paper presents a damage identification method based on wavelet analysis and deep learning,which provides improved damage identification accuracy for steel-concrete composite girders.Six composite girder specimens consisting of steel I girders and concrete slabs and revealing different levels of damages were prepared,the surfaces of these specimens were impacted by the freely-falling steel balls to make damages,and the strain signals of the specimens with different levels of damages were collected using fiber bragg grating strain sensors.The collected strain signals were denoised by five wavelet basis functions,including haar,sym2,sym4,db2 and db4.Additionally,six deep learning models(ResNet-18,ResNet-50,ResNet-101,InceptionV3,InceptionResNetV2,and MobileNetV2)were built to train and predict the strain signals before and after the denoising,aiming to select the model with the highest prediction accuracy,and finally to achieve the categorization and localization of damages in the specimens.It is concluded that the denoising effect of haar is superior to the other four functions,and the prediction accuracy of ResNet-50 is higher than the other five models,showing the average prediction accuracies of 96.73％and 97.91％before and after denoising,respectively,and the wavelet denoising enables the prediction accuracy of the ResNet-50 to be improved by 1.18％.The ResNet-50 model reveals a prediction accuracy of 96.82％even afar damages.This method serves as an alternative for maintenance and damage prediction of the steel-concrete composite girders.
维普
万方数据
