Artificial Intelligent Damage Identification Methods With Simulated Long-term Monitoring Data

Structural health monitoring(SHM)systems become indispensable parts for practical engineering.They acquire data of structural responses in order to monitor the state of structures,ensure the safety and provide scientific basis for maintenance and reinforcement.However,the capabilities of SHM systems are far away for their initial objective.Though a great quantity of data has been accumulated,it is still difficult to extract structural damage information through monitoring data and evaluate the current state of bridges.This is due to the reason that the relation between response data and structural states is non-linear.As a result,how to extract damage-sensitive features from monitoring data for damage identification has been a challenge.In this paper,the monitoring data of different types of bridge under a large number of damage scenarios are obtained by numerical simulations,providing the data basis for the following damage identification studies.Subsequently,novel damage identification methods based on machine learning algorithms are proposed.Results show that the novel methods in this paper improve the sensitivity and resolution for damage identification.The contents and novelty of this dissertation are summarized as follows:(1)A simulated monitoring database of a planar beam and a continuous rigid frame bridge under a large number of damage scenarios is established.The database not only provides a foundation for the following studies in this paper,but also provides data support for method validation of other researchers in this field.(2)Since the robustness of present structural damage identification methods is poor,which limits the application of the methods for damage identification in large-scale bridges,we propose a novel double-window principal component analysis(DWPCA)method.This method introduces both space and time windows in the traditional principal component analysis(PCA)method.The time window is used to discriminate damage and non-damaged states,while the presence of the space window aims to exclude the damage-insensitive data from those sensors located far from the damage.Results show that the damage detectability of the proposed method is improved compared with previous methods.(3)Two damage indices based on DWPCA are also proposed in this paper.Specifically,the length of the eigenvector variation(LEV)and directional angles of the eigenvector variation(DEVs)are proposed as the two damage indices instead of the components of eigenvector variation(CEVs)by the moving principal component analysis(MPCA)in the previous studies.Results show that the proposed features exhibit better performance in damage identification,localization,and quantification,compared with the CEV by MPCA usually used in previous studies.Noise immunity investigation further indicates that the proposed features by DWPCA has a good anti-noise property and have great potential in damage identification for practical engineering.(4)The eigenvectors of structural responses derived from MPCA are proposed as the input features for machine learning to identify damage.The results demonstrate that,as compared to strains and frequencies,their eigenvectors as input features for machine learning algorithms render better performance in damage identification.It is demonstrated that this method is capable of detecting damages that are located far away from sensors or have small damage severity.