| As the complexity and parameter uncertainty of engineering system becomes more and more prominent,it is found that the dynamic control equation based on the principle of mechanics is often difficult to meet the demand of precision.With the continuous development of artificial intelligence technology,system modeling using accurate simulation and experimental data has been widely concerned and become a research hotspot,but there are still some limitations.This paper carries out research on datadriven system modeling and response prediction,mainly including the following aspects:(1)The research significance of data-driven modeling method is introduced,and the research status of parametric modeling method and non-parametric modeling method is summarized.The necessity of developing and perfecting the theory and method of neural network modeling is pointed out.(2)Propose a data-driven modeling method based on system excitation and response data.The main process are as follows: using dynamic state equation,the system response data under different external excitation is numerically simulated to train the neural network;A loss function containing the known relationship between training data is constructed to improve the model accuracy.The trained neural network is a data model expressing the relationship of system states,which is incorporated into the ordinary differential equation solver to predict the system responses under different excitation.The modeling method is applied to the spring mass system with cubic and gap nonlinearity respectively,and the calculation results show that the accurate system states data model can be established based on the low frequency response data,and the hysteretic and jumping phenomena of the system in the main common amplitude frequency response can be predicted.The study also shows that the more training data,the more complete the data coverage state,the better the accuracy of the data model,and the smaller the error of the prediction response.(3)In view of the problem that the increase of system freedom degree will lead to the slow training speed and decreased accuracy of neural network,a substructure datadriven modeling method is proposed to transform the modeling problem of high freedom degree system into multiple low freedom degree system modeling problem.On the basis of the substructure modeling method,the fusion data-driven modeling method is also proposed,which combines the known part of the equation of state with neural network training to reduce the amount of neural network training and improve the efficiency of operation.Six degrees of freedom by cubic nonlinear spring mass system of numerical example show that the substructure data-driven modeling method and the fusion data driven modeling methods are greatly improving the calculation precision of the data model,and the required training for a long time but the substructure method,and the structure data fusion modeling method needs known part,so the two methods have their applicable engineering background.(4)For aeroelasticity,a typical nonlinear dynamics engineering problem,the research on binary wing data modeling and flutter velocity prediction based on response data is carried out.The responses of the binary wing system under aerodynamic force and aerodynamic torque were simulated numerically,and the data were used to train the neural network.The system responses under different flow velocities and the critical flutter velocities were predicted by using the obtained data model.The global dynamic characteristics of the system are studied by using cell mapping method and the initial attraction domains of various motion types are predicted. |
