Application Of High-performance Integration Method In Hydird Test

In today’s rapidly developing seismic theory,people’s research on structural seismic resistance has gradually transitioned from the theoretical stage to the experimental stage,and hybrid testing is the most commonly used structural seismic test in recent years.Among the three core components of hybrid test,numerical integration method is the most important,but the numerical integration method currently used in hybrid test will lose stability when solving nonlinear problems,so how to apply high-performance integration method in hybrid test is an urgent problem.This article establishes a hybrid experimental platform based on Matlab-OpenSees to address the above issues,and combines the GMDH neural network prediction algorithm with the average acceleration method to perform hybrid simulation on a single-layer,single-span planar frame.The main research content of this article is as follows:(1)The internal functions of m language and Tcl language were used to complete the data exchange between Matlab and OpenSees.The effectiveness of communication between Matlab and OpenSees was verified through the transfer of simple content;The feasibility of the hybrid experimental system based on Matlab-OpenSees was demonstrated through eigenvalue analysis,transient analysis,and hybrid simulation of a planar single-layer single span frame by introducing Matlab-OpenSees communication interface into the OpenSees-Open Fresco-MTS hybrid experimental system.(2)Introduced the basic idea of the Data Processing Combination Method(GMDH)and completed the programming of the GMDH neural network prediction algorithm;The reliability of GMDH prediction was demonstrated through displacement prediction of single degree of freedom structures,and the influence of training ratio on GMDH prediction performance was analyzed;Three methods for applying GMDH to the average acceleration method were proposed,and the applicability of GMDH neural network prediction algorithm in the average acceleration method was discussed by solving a single degree of freedom system with undamped free vibration.By solving the single degree of freedom structure,it can be concluded that the solution process using GMDH prediction is usually time-consuming,but the main time-consuming process is the prediction process of GMDH;The displacement predicted by the GMDH neural network is very close to the displacement obtained by iterating to the convergence criterion.Therefore,compared to the solution process without prediction,the number of iterations can be reduced by about 27% from the point where the GMDH predicted displacement value is iterated until the accuracy requirements are met.(3)Solve the planar single story single span frame structure under both free and forced vibration conditions.Without considering the deformation of the upper beam of the frame,the frame structure can be simplified as a column and horizontal spring structure,thus simplifying the calculation process.The restoring force model of the spring is set as a double curved tangent model;Using a hybrid experimental architecture based on Matlab-OpenSees and an average acceleration method based on GMDH,a simplified frame model was mixed and simulated.The spring was set as the experimental unit,which can be replaced by the experimental truss unit;An actuator is horizontally arranged for the experimental unit to simulate boundary conditions.The displacement calculated at the Matlab end is sent to the experimental unit through the actuator,and the restoring force model at the MTS end is set to hyperbolic tangent.The calculated restoring force is sent to Matlab for calculating the next step of displacement.The simulation results show that after a certain incremental step,GMDH has the best prediction effect by training and predicting at each step,with the least number of iterations,and has a very obvious advantage in solving free vibration structures;However,when the structure is subjected to earthquake action,applying GMDH neural network to the solving process will not achieve very significant results,and the number of iterations can only be reduced by about 5%;Compared to traditional calculation methods that do not perform predictions,using GMDH neural networks for prediction will take longer.