Relationship Modeling Between Multi-source Parameter And Deflection Of Cable-stayed Bridge By LSTM

Deflection is an important index of bridge measurement,which represents the vertical displacement of bridge under the action of load or dead weight,reflects the overall characteristics of bridge structure,and plays a key role in the evaluation of bridge health condition.+Based on the health monitoring system of Luoxi Bridge,this paper mainly analyzes the change rule of main beam temperature,cable temperature,humidity,cable modal frequency of the bridge and the correlation with the deflection,aiming to reveal the influence of various monitoring parameters on the bridge deflection,and make a fitting prediction of the short-term trend and long-term trend of the bridge deflection.The main contents and achievements of the study are as follows:(1)The variation rules of main beam temperature,cable temperature,humidity,cable modal frequency and deflection monitoring data are analyzed.It was found that all monitoring data showed a periodicity of 1 day(2)The data of a whole month were analyzed by taking 10 min as the analysis time interval,and the correlation between the deflection and the main beam temperature,the cable temperature,the humidity and the cable frequency were studied respectively.It is found that the deflection is positively correlated with the main beam temperature and the cable temperature,but negatively correlated with the humidity and the cable frequency.There is a high correlation between the deflection and humidity,and the long-term deflection of the bridge increases with the decrease of humidity.The higher the humidity,the greater the range of deflection.According to the curve quality,CV value and correlation scatter plot,it is found that the correlation between 3-5 order frequency and deflection is the best,and the comprehensive effect of the 5th order frequency is the best.(3)The preprocessing of data sets can improve the accuracy of the model.The time delay will affect the fitting effect of the correlation model,and can be calculated and eliminated through the cross-correlation coefficient.Then,the data set is normalized so that all data are squeezed in a unified interval [0,1],thus improving the training speed and fitting effect.At the same time,an improved damage function is introduced to reduce the contribution of low amplitude noise to the loss and amplify the effect of real structural response with larger amplitude.(4)Based on LSTM neural network,a multi-source parameter-deflection global and local correlation model is established.The global model based on mid-span cross-section has good accuracy.Then,the multi-source parameters were selected for dimensionality reduction by correlation coefficient,and finally the temperature of main beam,humidity and fifth-order frequency of cable were selected as the input of local parameter model.It is found that humidity contributes a lot to the variation of bridge deflection.Then the local parameter combination was extended to other sections of the bridge to investigate its generalization ability.According to the evaluation indexes such as goodness of fit and residual error,the LSTM neural network model after dimension reduction also has good fitting prediction ability.