Forde, Mike

Lu, Yong

Xu, Donghui

2024-04-29T14:17:53Z

2024-04-29T14:17:53Z

2024-04-29

Structural Health Monitoring (SHM) systems are being installed on long-span bridges worldwide to facilitate condition assessment and damage detection. The vast volume of data generated by these SHM systems presents significant challenges for analysis. This thesis introduces innovative research into big data analysis, drawing on SHM data from two long-span bridges in Scotland: the Queensferry Crossing (QC) and the Forth Road Bridge (FRB). Initially, data integrity was addressed, as raw data from both bridges exhibited errors. The errors were detected by visual analysis and statistical analysis, e.g. z-score. The data available for this project related to the first year of collection from the QC - before the final snagging of the system was undertaken by the contractor. Ensuring data integrity is crucial for subsequent analysis. A novel Random Forest model based on relavant sensor data was proposed to impute substantial segments of missing data. The z-score filter and Exponentially Weighted Moving Average (EWMA) were utilised to identify and rectify outliers. The proposed methods were found to effectively reconstruct SHM data, interpolating missing segments and identifying and substituting outliers. Subsequently, to investigate the vehicle-induced structural response, a new traffic load simulation model was developed. This simulation utilised a car-following model combined with a lane-changing model to replicate the trajectory of vehicles on the Queensferry Crossing. This was based on vehicle attribute distributions captured by the dynamic weigh-in-motion system positioned at the bridge entrance. The deflection influence line at the south midspan was then derived from the bridge’s finite element model. Ultimately, the deflection at the south midspan was determined using both the simulated vehicle data and the influence line. The findings revealed a strong correlation between the actual monitored and simulated deflections in terms of both magnitude and variations. Once the simulation model was validated, a recurrent neural networks training, validation, and testing strategy was proposed to predict deflections based on the vehicle information at the bridge entrance using the simulated data. Parameter analysis was performed to investigate the impact of hyper-parameters, e.g. neuron numbers and number of layers, on the prediction performance. It was discovered that multi-layer Long-short Term Memory (LSTM) networks outperform single-layer models with a large number of neurons. The performance of LSTM was evaluated by the mean square root between the observed values and predicted values. The results showed accurate generalisation of deflection data based on traffic information. The study also delved into the exploration of thermal-induced deflections. The initial step involved identifying the QC’s critical sections using its finite element model. The analysis revealed that significant bending moments were observed at the north and south midspans, as well as the deck over the southern and northern auxiliary piers. Considering the placement of SHM sensors, we chose south midspan as the research target. Then, a temperature field analysis was performed to investigate temperature differences at different locations on the bridge. Next, Generalised Pareto Distrbution (GPD) was employed to estimate the maximum vertical temperature difference and tower temperature difference in 120 years based on SHM data. Further, finite element analysis was performed to compare south midspan deflections when thermal loads were applied to different parts of the QC. It was observed that thermal loads on the stay cables have the most substantial impact on deck deformation. Then, correlations between temperatures and deflections were explored by using wavelet transform. The study culminated with the application of LSTM to predict deflections based on temperature data sourced from various bridge locations. This research explores the feasibility of machine learnring applications to bridge SHM data analysis, mapping complex relationships between environmental & operational conditions and bridge structural responses. The outcome of this research provides researchers and bridge engineers with new perspectives on how to utilise SHM data for bridge maintenance.

https://hdl.handle.net/1842/41715

http://dx.doi.org/10.7488/era/4438

en

The University of Edinburgh

Xu, D., Chong, H., Main, I., Mineter, M., De Bold, R., Forde, M.C., Gair, C., Madden, P., Angus, E. and Ho, C., 2019. Using statistical models and machine learning techniques to process big data from the forth road bridge. In International Conference on Smart Infrastructure and Construction 2019 (ICSIC) Driving data-informed decision-making (pp. 411-419). ICE Publishing.

Xu, D., Xu, X., Forde, M.C. and Caballero, A., 2023. Concrete and steel bridge Structural Health Monitoring-Insight into choices for machine learning applications. In Construction and Building Materials, 402, p.132596.

Xu, X., Xu, D., Caballero, A., Ren, Y., Huang, Q., Chang, W. and Forde, M.C., 2023, August. Vehicle-induced deflection prediction using long short-term memory networks. In Structures (Vol. 54, pp. 596-606). Elsevier

Structural Health Monitoring systems

SHM systems

long-span bridges

Queensferry Crossing

Forth Road Bridge

damage detection

condition assessment

Random Forest model

Exponentially Weighted Moving Average

vehicle-induced structural response

thermal-induced deflections

Generalised Pareto Distrbution

Big data analysis on long-span bridge structural health monitoring systems

Thesis or Dissertation

Doctoral

PhD Doctor of Philosophy