Correlation model of deflection, vehicle load, and temperature for in‐service bridge using deep learning and structural health monitoring

• Published in: Structural control and health monitoring Vol. 29; no. 12
• Main Authors: Deng, Yang, Ju, Hanwen, Zhai, Wenqiang, Li, Aiqun, Ding, Youliang
• Format: Journal Article
• Language: English
• Published: Pavia John Wiley & Sons, Inc 01.12.2022
• Subjects: Bridge loads; correlation model; Decomposition; Deep learning; Deflection; Extreme values; GRU neural network; Influence coefficient; Influence lines; Load distribution; Modelling; Neural networks; Predictions; Structural health monitoring; suspension bridge; Suspension bridges; vehicle load; Windows (intervals)
• ISSN: 1545-2255; 1545-2263
• DOI: 10.1002/stc.3113

Summary Deflection is an important issue in bridge structural health monitoring. An accurate deflection–vehicle load–temperature correlation model is critical to abnormal data identification, deflection prediction under extreme conditions, and bridge structural assessment. However, because of the discrete distribution in time domain of vehicle load and the extreme complexity of the deflection–vehicle load–temperature correlation, the correlation modeling method needs further studies. A novel deflection–vehicle load–temperature correlation modeling method is developed in this study. Based on the concept of deflection influence line (DIL), the raw vehicle load monitoring data are transformed into time‐continuous vehicle influence coefficient (VIC). By using gated recurrent unit (GRU) neural network, a correlation model with inputs of VIC and environmental temperature data and output of deflection data is established. Taking a suspension bridge in China as an example, the prediction accuracy of short‐, medium‐, and long‐term correlation models is tested. Moreover, based on the correlation model, a decomposition method of temperature‐ and vehicle‐induced deflection components is proposed. The results show that the predicted deflection of the short‐term correlation model is basically consistent with the real‐time monitoring data, while the medium‐ and long‐term correlation models have accurate prediction ability for the deflection extreme values in a certain time window. The temperature‐ and vehicle‐induced deflection components separated by using the correlation model are in good agreement with the wavelet decomposition (WD) results, with clear physical meaning and independent of empirical judgment.