Application Of Machine Learning In Nonlinear Dynamic Response Analysis Of Structures

Earthquake is a sudden natural disaster,which often causes serious damage to the building structure.Building structure seismic design and seismic measures is an effective method for disaster prevention and mitigation,the seismic time history on structural dynamic time history analysis is the conventional method in seismic design,putting forward along with the new structure form and the application of new materials,the structure becomes more complex,the dynamic response analysis of structures is also put forward higher request,tend to produce nonlinear dynamic response under seismic action,The traditional dynamic analysis theory can not meet the design requirement gradually.Constantly breakthroughs and artificial intelligence theory,machine learning is gradually applied to all aspects of the engineering field,solved many traditional algorithm can not solve the problem,if we can machine learning method combined with structural dynamic response analysis,will bring efficiency of structure dynamic response analysis,provide a lot of new structure forms a new analysis method,It is of great engineering significance to further promote the development of structural dynamic analysis theory.Based on this,this paper mainly carries out the following research:(1)GRU neural network has the ability to learn the nonlinear characteristics of sequences.This paper proposes to use GRU network with exponential sliding window function to predict the nonlinear dynamic response of structures.For structures with nonlinear components,first by trigonometric series method,producing artificial seismic waves under different conditions,and then using the Simulink simulation method to calculate the dynamic response of structure under seismic action on seismic wave and the response data and index window to get the neural network input and the corresponding output,the network is trained as training set,The nonlinear dynamic response of the corresponding structure is predicted by using the windowed GRU network model completed by training.(2)Numerical substructure method to nonlinear component segregation analysis alone,but for the nonlinear analysis of the components is still the traditional method,this paper puts forward the numerical substructure method based on BP neural network,the main structure USES the linear elastic analysis,the nonlinear component in isolation,through BP neural network for nonlinear component is analyzed.The node displacement transmitted by the main structure is used as the network input to predict the substructural force of the nonlinear component and then transmitted back to the main structure for further calculation.The node displacements and substructural forces of nonlinear components are obtained by simulation or experiment,which are used as the training set of BP neural network to predict the substructural forces of nonlinear components.(3)For the windowed GRU neural network model,the single-degree-of-freedom(SDOF)and multi-degree-of-freedom(MDOF)container structures with hell-in are verified.Artificial seismic wave is used as input to predict the model,and the displacement time-history curve is used for qualitative evaluation.The determination coefficient and peak response error are proposed to quantitatively evaluate the model prediction results.For the numerical substructure method based on BP neural network,the framework of four-layer vessel with clutch is verified,and the force of substructure is compared,and the time history curve error is analyzed.The results show that the proposed method can accurately predict the nonlinear dynamic response of structures.