CNN-based Damage Identification Method For Arch Bridge Using Spatial-Spectral Information

In the field of structural health monitoring,damage detection has been commonly carried out based on the structural model and the dynamic signatures related to the model.However,the structures in the real-world are often affected by numerous uncertain factors,and the extracted features are also subjected to various error,which makes the pattern recognition for damage detection still challenging.Recently,researches on machine learning(ML)based damage identification have attracted extensive attention.Typically,ML algorithms have been applied to extract features from the dynamic responses of structures and learn the correlation between the changes of modal properties and the locations and indices of structures damage,so as to achieve the purpose of damage identification.However,most researchers used structures' original time-history dynamic response under random white noise or the modal properties obtained through feature extraction as input of the machine learning algorithms.However,white noises cannot simulate environmental excitations of real structures and the time-history dynamic response of structures cannot show the changes dynamic modal properties clearly.Besides,the extraction of modal properties requires complex computation and might result in the loss of important information.In this study,a Convolutional Neural Network(CNN)based damage identification method with Spatial-Spectral Information is presented for automated operation using raw measurement data without complex procedure for feature extraction.CNN is a kind of deep neural networks which typically consists of convolution,pooling,and fully connected layers.A numerical simulation study has been carried out for multiple damage detection in hangers of a tied-arch bridge using ambient wind vibration data.To accelerate the generation of numerous samples for the training and testing,VFIFE method which was developed by author's research group,is used in this study.The work and research results conducted in this study is summarized as follows:(1)An arch bridge model was constructed based on the vector form intrinsic finite element(VFIFE)method,and the accuracy of the model was verified by comparing the modal properties from the vector finite element formulation with the arch bridge model built by Midas Civil?(2)The time-history fluctuating wind speed at each point on the bridge considering the spatial coherence was computed and the buffeting wind force at each point on the arch bridge model were simulated.(3)The basic principle and training procedure of CNN are introduced.A self-designed convolutional neural network is built with TensorFlow,and some normalization methods are applied to prevent overfitting and improve CNN's generalization ability.(4)In order to conduct the damage identification on multiple hanger cables,the Fourier amplitude spectra(FAS)of acceleration responses on the bridge deck are used as an input to the CNN.Numerical results show that the current CNN using FAS data performs better under various damage states than the CNN using time-history data and the traditional neural network using FAS.Robustness of the present CNN has been proven under various level of observational noises and wind speeds.