Research On Prediction Method Of Structural Pathological Problems Based On LSTM

During the daily service period of large-scale equipment and building structures,due to various uncertainties,natural factors and human factors,the structure will be damaged and some unpredictable consequences will occur.Therefore,it is very important to monitor,diagnose and predict the conditions of the structures during the service period.This paper studied the prediction method of structural ill-conditioned problems based on LSTM.The basic theoretical knowledge of LSTM network was discussed.From the perspective of time series prediction,the time series prediction method based on LSTM was studied.Firstly,the Z-score standardization method was used to normalize the simulation time series data and then use it as the input of the prediction model.Then,the number of iterations and the number of hidden layer neurons were changed several times to determine the iteration number and the number of hidden layer neurons in the LSTM prediction model.Finally,in order to illustrate the superiority of the LSTM model,the prediction performance of the model was compared with the RBF neural network and the wavelet neural network.The experimental results showed that the selection of model parameters had different effects on the prediction results.When the number of iterations and the number of neurons in the hidden layer were 100 and 200respectively,the prediction accuracy of the model was better;Moreover,compared with the other two traditional prediction methods,the LSTM network had better prediction performance and higher precision,which provided a basis for the prediction application of engineering data.From the perspective of the application of practical problems,the prediction method of structural ill-conditioned problems based on LSTM was studied.The multi-degree-of-freedom time-varying dynamic system and the ASCE model were used to verify the engineering reliability of the prediction method.Since the instantaneous frequency can represent the damage information of the structure,the engineering simulation signal and the actual engineering data were used for HHT.The instantaneous frequency was used as an input to the prediction model.The experimental results showed that the LSTM prediction model can effectively predict the instantaneous frequency after the change of the stiffness of the multi-degree-of-freedom dynamic system.For the different damage conditions of the structure,the prediction accuracy of the LSTM method was higher than that of the wavelet neural network,which indicated that the method had strong fitting ability to nonlinear time series,and further demonstrates the effectiveness and reliability of the prediction of ill-conditioned vibration signals in practical engineering.The small sample prediction of engineering actual data based on the integration method of SVR and LSTM was studied.The actual engineering signals and bridge vibration signals in Chapter 3 were used as experimental data,and the integrated methods were compared with the predicted performance of a single SVR or LSTM model.The experimental results showed that the integrated method had higher prediction accuracy and better nonlinear prediction ability for the smaller engineering data.In the ASCE actual engineering data prediction,the integrated method had a 12%improvement over the R~2 of the single LSTM model.