Research On Building Structural Earthquake Response Algorithm Based On Machine Learning

Due to the strong uncertainty and complexity of earthquakes,urban buildings are highly likely to suffer damage from seismic ground motions stronger than the fortification intensities.Emergency rescue,seismic intensity assessment,seismic damage assessment and urban seismic resilience assessment urgently need to obtain the safety and damage status of building structures in the shortest possible time,so it is crucial to estimate the seismic response of urban building structures quickly after an earthquake.The strong nonlinearity and non-stationarity of ground shaking and the complexity of urban building composition lead to the problems of large dispersion and unstable estimation results of linear and nonlinear regression methods for estimating the seismic response of building structures.In this paper,machine learning methods are applied to estimate the structural seismic response,and the paper mainly accomplishes the following works:(1)Apply the support vector machine method to estimate the maximum seismic response of building structures.Using the actual structural seismic observation data from the Engineering Strong Earthquake Data Center(CESMD)in the U.S.and the Architectural Research Institute(BRI)in Japan,a database of the seismic response of building structures is formed,and a total of 41 characteristic parameters in two categories,namely ground shaking parameters(seismic information,spectral information,and timeholding information)and structural information,are selected as inputs to construct a support vector machine model for estimating the maximum interstory displacement angle of structures with physical information(Drift Support Vector Machines(DSVM)model with physical information.The model validation results show that the DSVM model can estimate the maximum interstory displacement angle of the structure.Compared with the traditional logistic regression method,the maximum interstory displacement angle estimation results are less discrete and have a smaller error range.Compared with the traditional support vector machine regression model with a single type of parameter input,the DSVM model has the input of the structure type and the model estimation results are more stable.In addition,the input parameters were selected based on feature importance scores and empirical methods,and the Reduced Drift Support Vector Machines(RDSVM model)was developed for estimating the displacement angle between layers of the reduced-dimensional structure with a small amount of information.The test results after feature selection show that the RDSVM model can reliably estimate the maximum interstory displacement angle of the structure under the condition of less information.(2)Application of recurrent neural network to estimate the seismic time response of building structures.Based on the Bouc-Wen hysteresis model,a single-degree-of-freedom system seismic response time data set containing 100 sets of data samples is established.The long and short time memory network in recurrent neural network is applied to establish the long and short time neural network model(Response-LSTM model)for structural response estimation by using the structural response time course dataset with structural substrate ground shaking as the input and structural displacement as the output.The results of the trained model on the test dataset show that the model is able to estimate the structural displacements quickly and accurately with the given inputs.(3)A recurrent neural network model with embedded physical mechanisms is developed to estimate the time-dependent response of the building structure.Applying the structural kinematic equations as soft constraints to enable the model to learn mathematical formulas with physical meaning in addition to the data,a long and short term memory network model(Loss-LSTM model)with applied loss function is established;using the characteristics of recurrent neural network,the model is trained by fitting a set of data with the building structure parameters as the model trainable parameters,and the model is finally able to obtain the structural parameters and LossLSTM,compared with the Reponse-LSTM model Phy-RNN can obtain better estimation results with less observations.