Model Updating Method Based On Artificial Neural Network In Substructure Test

Substructure test,also known as hybrid test,is a structural seismic test method which combines physical loading test and finite element numerical simulation.It has an important application prospect for revealing the seismic performance of large and complex structures.Generally,the key components which enter into the nonlinearity first are regarded as the test substructure for loading test,and the remaining part is taken as the numerical substructure with assumed numerical model in simulation.However,for the key components that may enter into the nonlinear state but can not be tested,there may be large model errors in the assumed numerical model: the first kind of errors come from the model defects caused by the over simplification of the numerical model;the second kind of errors are from the uncertainty of parameters of the numerical model.In order to reduce the model errors,the concept of model updating is applied to hybrid test.The constitutive model of the test substructure is identified online by the measured data of the test substructure,and the corresponding part in the numerical substructure is updated synchronously.The neural network can identify the system which can not be described by the definite numerical model,so as to avoid the risk of structural response distortion caused by the defects of numerical model.Thus,a model updating method based on neural network in substructure test is proposed.The main contents and conclusions are as follows:(1)The framework of the model updating method in hybrid test based on neural network is proposed.Firstly,the robustness control of the online neural network algorithm is studied from two aspects of stability and adaptability.The two-stage test method framework of "offline training + online fine tuning" is adopted.The neural network model is firstly trained offline by the existing test data of the component,and the pre-trained neural network model is set as the initial model of the corresponding components in the numerical substructure.Then the online neural network algorithm is applied to fine tune the inaccurate neural network model to ensure the stability of the online identification system.The combination of "random offline samples + dynamic window samples with forgetting factor" are adopted as the online training samples of each loading step.Some samples are randomly selected from the offline samples to ensure the good performance of the neural network in the old tasks;the other samples are the latest batch samples of the test substructure in the current loading step.The more new the samples are,the greater the training weight is given to improve the adaptability.The framework and implementation process of the proposed method are summarized.Taking a 2-Dof nonlinear system as an example,the numerical verification of the proposed method under similar loading histories is carried out in MATLAB.Four sets of offline samples with different nonlinear degrees are used to pre train the neural network model.The results show that: under the similar loading histories,compared with the traditional hybrid test and the hybrid test based on neural network offline calibration,the proposed method can effectively reduce the model errors between the preset numerical model or the pre-trained neural network model and the real model to varying degrees with higher accuracy,robustness and applicability.(2)A hybrid test system based on neural network model updating(MATLAB-Openfresco-MTS hybrid test system)is built.The communication connection between Openfresco,MTS loading control system and MATLAB(finite element analysis and model updating calculation of neural network)is described.Taking a six story frame structure with buckling restrained braces as an example,the effectiveness of the hybrid test system is verified by numerical simulation,and the numerical verification of various model updating methods in hybrid test under different loading histories is carried out.The test substructure is simulated and loaded by the built-in experimental element in Openfresco,and the Bouc-Wen-Baber-Noori model is adopted to explore the recognition ability of different model updating methods for strong nonlinear model with large model error.The results show that the proposed hybrid test system is feasible and effective.In the case of large model error,compared with the traditional hybrid test,the hybrid test based on neural network offline correction,the hybrid test based on traditional neural network model updating,and the hybrid test based on UKF model updating,the proposed method can greatly reduce the model error under different loading histories,and can learn the hysteretic behaviors that do not exist the initial model and has good recognition effects for strong nonlinear model.And it has high prediction accuracy of restoring force in the entire time history and at the peak of response,showing good adaptability,accuracy and robustness.In addition,the average time of each step in numerical simulation is in the range of 0.12s-0.15 s,and the computational efficiency of the proposed method meets the requirements of slow pseudo dynamic substructure test.(3)Taking a two-story two span frame structure with bending dampers as an example,the test substructure of bending damper is designed,and the hybrid test based on neural network model updating is carried out.The experimental results show that: compared with the traditional hybrid test,the hybrid test based on neural network offline calibration and the hybrid test based on UKF model updating,the proposed method can effectively reduce the model errors between the preset numerical model or the pre-trained ANN model and the real model.It can learn the nonlinear behaviors that do not exist in the original model and realize the recurrence of the hysteresis curve.It can predict the restoring force well in the entire time history and at the peak of response,which further improves the accuracy of hybrid test.In addition,the average time of each step in hybrid test verification is 0.13 s,and the proposed method can be well applied to slow pseudo dynamic substructure test.