Bridge Deflection Separation Prediction Model Based on DWT-LSTM and Its Engineering Application
ObjectiveAccurate prediction of deflection variation holds significant importance for bridge operation and maintenance. Complex and nonlinear dynamic characteristics of bridge deflection often challenge traditional prediction models, as hysteresis in deflection response and interference from irregul...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Editorial Department of Journal of Sichuan University (Engineering Science Edition)
2025-01-01
|
| Series: | 工程科学与技术 |
| Subjects: | |
| Online Access: | http://jsuese.scu.edu.cn/thesisDetails#10.12454/j.jsuese.202401045 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | ObjectiveAccurate prediction of deflection variation holds significant importance for bridge operation and maintenance. Complex and nonlinear dynamic characteristics of bridge deflection often challenge traditional prediction models, as hysteresis in deflection response and interference from irregular waveforms in historical monitoring data reduce prediction accuracy. This paper proposes a deflection separation-prediction model for bridges by integrating wavelet optimization and long short-term memory networks, aiming to capture multi-scale features of deflection signals and account for external influences.MethodsFirstly, an Internet of Things monitoring system was employed to investigate deflection behavior in in-service bridges. Focusing on the Xiongshang High-Speed Railway Bridge over the Daguang Expressway, sensors installed on the bridge were used to monitor variations in deflection, dynamic load, and temperature. Secondly, considering the decoupling of different deflection components across time scales, wavelet-based optimization was applied to decompose historical monitoring data into trend deflection caused by prestress loss and noise deflection induced by external factors such as temperature and dynamic loads. Thirdly, based on the decomposed deflection components and external factors, two LSTM-based time series prediction models were developed: a multi-factor model for noise deflection and a single-factor model for trend deflection. Vehicle load, temperature, and noise deflection served as inputs for the noise model, while trend deflection was used as the input for the trend model. Separate predictions were conducted, and the final cumulative bridge deflection was obtained by summing both predicted components according to the principle of time series superposition. In order to evaluate prediction accuracy, traditional models were used for comparison across short-term, medium-term, and long-term periods. Prediction performance was assessed using three metrics: correlation coefficient (R), root mean square error (RMSE), and mean absolute error (MAE). Fourthly, in order to demonstrate the necessity of the combined model for predicting bridge deflection, a comparative analysis was conducted between the proposed model and the single LSTM prediction model. Fifthly,in order to validate the necessity of incorporating external factors, the proposed model was compared with a time series model including only a single external factor and another excluding external influences, focusing on differences in prediction accuracy. Sixthly, maximal information coefficient was introduced to identify dominant factors affecting noise deflection by analyzing its correlation with temperature and dynamic load.Results and Discussions1) Comparison of prediction results for short-term, medium-term, and long-term periods with BP neural network and LSSVM models showed that prediction accuracy for all three models was 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 strongest generalization ability, with correlation coefficients of 0.86 and 0.77, root mean square errors (RMSE) of 2.18 mm and 2.20 mm, and mean absolute errors (MAE) of 2.05 mm and 1.91 mm. In contrast, the LSSVM model had RMSE values of 2.82 mm and 3.52 mm, with MAE values of 2.45 mm and 3.13 mm. The BP neural network had RMSE values of 3.06 mm and 3.53 mm, with MAE values of 2.89 mm and 3.24 mm. Compared with the LSSVM model, the DWT-LSTM deflection separation model reduced RMSE by 22.70% and 37.50%, and MAE by 39.26% and 38.98%. Compared with the BP neural network, the DWT-LSTM deflection separation model reduced RMSE by 28.76% and 37.68%, and MAE by 29.07% and 41.05%. 2) Compared with the DWT-LSTM deflection separation model, prediction accuracy decreased when using a single LSTM model. Root mean square error reached 3.74 mm, and mean absolute error reached 3.45 mm. Relatively large deviations indicated limited suitability for bridge deflection prediction. 3) Compared with time series models that consider only temperature, only vehicle load, or exclude external factors, the model excluding external factors exhibited the lowest prediction accuracy, with a root mean square error (RMSE) of 3.91 mm and a mean absolute error (MAE) of 3.38 mm. Among the models considering a single external factor, the time series model considering load had a higher prediction accuracy, with RMSE of 2.81 mm and MAE of 2.65 mm, outperforming the temperature-only model, which had RMSE of 2.97 mm and MAE of 2.83 mm. In contrast, the DWT-LSTM deflection separation model achieved the highest accuracy, with root mean square error of only 2.18 mm and mean absolute error of only 2.05 mm. (4) Analysis of the dominant factors influencing noise deflection using the Maximal Information Coefficient (MIC) showed correlation coefficients of 0.35 for temperature and 0.51 for load, indicating that vehicle load has a greater impact on noise deflection than temperature.ConclusionsThis paper presents a DWT-LSTM-based bridge deflection separation prediction model, suitable for predicting long-term deflection variation patterns. Compared to traditional prediction models, the proposed model demonstrates higher accuracy, smaller errors, and superior capability in handling time-lag effects, providing a new approach and method for long-term bridge deflection prediction. |
|---|---|
| ISSN: | 2096-3246 |