桥梁挠度的DWT‒LSTM分离预测模型及工程应用OA

Objective Accurate prediction of deflection variation holds significant importance for bridge operation and maintenance.The complex and non-linear dynamic characteristics of bridge deflection consistently challenge traditional prediction models,as hysteresis in the deflection response and interference from irregular waveforms in historical monitoring data reduce prediction accuracy.This study proposes a deflection separation-prediction model for bridges by integrating wavelet optimization and long short-term memory networks to capture multi-scale features of deflec-tion signals and account for external influences. Methods Firstly,an Internet of Things monitoring system was employed to investigate the deflection behavior of in-service bridges.Focusing on the Xiongshang High-speed Railway Bridge over the Daguang Expressway,sensors installed on the structure were utilized to record variations in deflection,dynamic load,and temperature.Secondly,given the decoupling of different deflection components across multiple time scales,a wavelet-based optimization approach was applied to decompose historical monitoring data into trend deflection generated by prestress loss and noise deflection induced by external influences such as temperature and dynamic loads.Thirdly,based on the decomposed deflection components and the associated external factors,two LSTM-based time series prediction models were developed,including a multi-factor model for noise de-flection and a single-factor model for trend deflection.Vehicle load,temperature,and noise deflection served as the inputs for the noise model,while trend deflection was used as the sole input for the trend model.Separate predictions were conducted,and the final cumulative bridge deflec-tion was obtained by summing both predicted components based on the principle of time series superposition.Traditional models were adopted for comparison across short-term,medium-term,and long-term periods to evaluate prediction accuracy.Prediction performance was assessed us-ing three metrics:correlation coefficient(R),root mean square error(ERMS),and mean absolute error(EMA).Fourthly,a comparative analysis was performed between the proposed model and the single LSTM prediction model to demonstrate the necessity of the combined model for forecast-ing bridge deflection.Fifthly,in order to verify the necessity of incorporating external factors,the proposed model was compared to a time series model including a single external factor and another excluding external influences,emphasizing differences in prediction accuracy.Sixthly,the maximal information coefficient was introduced to identify the dominant factors affecting noise deflection by analyzing its correlation with tem-perature and dynamic load. Results and Discussions 1)Comparison of the prediction results for short-term,medium-term,and long-term periods with the BP neural network and LSSVM models showed that the prediction accuracy for all three models remained similar in the short and medium periods.However,in the long-term deflection prediction,the DWT‒LSTM-based bridge deflection separation model achieved the highest accuracy and demonstrated stronger generalization ability,with correlation coefficients of 0.86 and 0.77,ERMS of 2.18 and 2.20 mm,and mean absolute errors(EMA)of 2.05 and 1.91 mm.In contrast,the LSSVM model produced ERMS values of 2.82 and 3.52 mm,with EMA values of 2.45 and 3.13 mm.The BP neural network produced ERMS values of 3.06 and 3.53 mm,with EMA values of 2.89 and 3.24 mm.Compared to the LSSVM model,the DWT‒LSTM deflection separation model reduced ERMS by 22.70%and 37.50%and reduced EMA by 39.26%and 38.98%.Compared to the BP neural network,the DWT‒LSTM deflection separation model reduced ERMS by 28.76%and 37.68%and reduced EMA by 29.07%and 41.05%.2)Compared to the DWT‒LSTM deflection separation model,the prediction accuracy decreased when the single LSTM model was used.The ERMS reached 3.74 mm,and the EMA reached 3.45 mm.These relatively large deviations indicated that this model had limited suitability for bridge deflection predic-tion.3)Compared to the time series models that considered only temperature,only vehicle load,or excluded external factors,the model that ex-cluded external factors exhibited the lowest prediction accuracy,with ERMS of 3.91 mm and EMA of 3.38 mm.Among the models that considered a single external factor,the time series model that considered load showed higher prediction accuracy,with ERMS of 2.81 mm and EMA of 2.65 mm,outperforming the temperature-only model,which had ERMS of 2.97 mm and EMA of 2.83 mm.In contrast,the DWT‒LSTM deflection separation model achieved the highest accuracy,with ERMS of only 2.18 mm and EMA of only 2.05 mm.4)Analysis of the dominant factors that influenced noise deflection using the Maximal Information Coefficient(MIC)showed correlation coefficients of 0.35 for temperature and 0.51 for load,indi-cating that vehicle load has a greater impact on noise deflection than temperature. Conclusions This study presents a DWT‒LSTM-based bridge deflection separation prediction model that is suitable for predicting long-term de-flection variation patterns.Compared to traditional prediction models,the proposed model shows higher accuracy,reduced errors,and improved capability in addressing time-lag effects,providing a new approach and method for long-term bridge deflection prediction.