Integrated IoT, computer vision and machine learning technologies for smarter bridge health monitoring and prediction: Project Overview

Bridge displacement is important to understand to determine how the structure responds to different traffic loads, especially heavy loads. However, conventional contact-type displacement sensors, such as the linear variable differential transducer (LVDT), require a stationary reference point that is often difficult to find in the field. Additionally, their short measurement range of less than 1 meter limits their application to large-span bridges. As an alternative, a computer vision-based method was tested. Lab tests demonstrated good agreement with LVDT sensor data. Due to the limited range of LVDT, direct on-site validation was challenging, and only indirect checking of the data was possible. The observations were in line with the displacement influence line theory, and the top ten detected displacements correspond to heavy traffic patterns recorded by the traffic camera, indicating the reliability of the collected data. During the test, the physical dimensions of the bridge were needed to translate the number of pixels in the video to real distances. As the side of the bridge was inaccessible for physical measurement, design drawings were used instead. The field test also identified that environmental factors such as wind-induced camera motion and lighting conditions could affect the accuracy of results. The research team proposed possible solutions for future research to address these challenges.
Vibration data can be used to detect variations in the bridge’s natural frequency over time, serving as an alarm for potential structural damage. A prototype IoT unit was built to measure bridge vibration using an accelerometer. The data was successfully transferred to the AWS S3 cloud storage over an IoT link, and a dashboard was created for accessing the collected accelerometer data. A comparison with an industry-grade wired sensor showed that the chosen IoT sensor lacked sufficient resolution, leading to the purchase of a better model. Although the field test opportunity was missed, lab tests demonstrated comparable results between the new IoT sensor and a smartphone accelerometer. The prototype has overall demonstrated the cost-effectiveness of IoT for collecting bridge vibration data when IoT data coverage is available. Further considerations include addressing power supply issues in remote areas and determining optimal attachment locations for the unit on the bridge.
The ML part of the project set out to develop proof-of-concept models to predict the bridge’s vibration and displacement responses given the observed traffic loads from videos. Both tasks were proven to be difficult, although the vibration prediction achieved comparatively better results. The available literature indicates the inherent difficulty of mathematically inferring bridge displacement from vibration data. Extensive testing was conducted using various ML models, yet the challenge persisted, further confirming the complexity of the task. The ‘phantom displacements’ in the data also contributed to the difficulty. Nevertheless, the model estimated influence curve seemed reasonable, which suggests that it captured some of the underlining mechanisms. Interestingly, combining the vibration data with traffic videos did not increase accuracy.