Recurrent Neural Network-Based Correlation Modeling Method For Regional Responses Of Bridges

With the rapid development of sensor technology,the structural health monitoring systems have been widely used in civil structures.However,how to mine the information about structural state from the monitoring data remains an important question.Traditional damage identification methods use modal parameters as damage sensitive features,which is insensitive to local damage.An alternative way is to model the local structure to extract the local damage sensitive features.But there are few effective methods for modelling nonlinear structures like long span bridges.Also,the damage identification methods are easily influenced by environmental and operational variation.Therefore,this thesis focuses on developing a deep learning-based method to model nonlinear local structure that can also take temperature into account.Main contents are included as follows:A recurrent neural network-based correlation modeling method is proposed.Firstly,the mathematical proof of describing the linear structure dynamic system using time series model is introduced.Next,several situations where input-output model can be transformed to correlation model among outputs are analyzed.Then,considering the geometry nonlinear behavior of responses of bridges due to large deformation,instead of traditional linear time series modeling methods,recurrent neural network is used to construct a correlation model among these responses.Also temperature information is incorporated in the model as an input to take temperature’s influence over the correlation into consideration.An error level-based damage identification method is proposed.Since model’s predication error is related to the structure’s state,the standard deviation of the error is calculated as the damage sensitive feature.To eliminate the disturbance of the noise,monitoring data is divided into blocks,and each block is used to calculate a standard deviation to construct an error level.And by comparing to the error level of healthy state,whether the structure is damaged can be determined.The proposed method is validated on the monitoring data of a rigid hinge of a long span cable-stayed bridge.Firstly,a correlation model is built using the proposed recurrent neural network-based method and trained using the monitoring data when the structure is in healthy state.Then,the model is used to predict the responses in unknown state and calculate the corresponding error level.And by comparing these error levels to the error level of the healthy state,the change of state of the rigid hinge due to the change of boundary condition is successfully detected,which prove the feasibility of the proposed method.