Technologies for smarter bridge health monitoring and prediction

Taking advantage of IoT, computer vision and machine learning, the completed Smart bridge health monitoring and maintenance prediction project investigated technologies to supplement physical bridge health assessments in remote areas where it can be costly, time-consuming and unsafe to physically inspect. The project’s final report is available for download below.

Over a long service life, bridges are prone to performance degradation owing to material deterioration, natural hazards and load conditions. Regular, efficient and reliable bridge health monitoring is essential for the long-term protection of these valuable road assets through timely maintenance responses.

This monitoring presents significant challenges in areas such as regional and remote Western Australia, which is where new technology has the potential to make bridge assessments safer, easier and more affordable.

With iMOVE on this project were Main Roads Western Australia, University of Western Australia (in partnership with Curtin University), and the Planning and Transport Research Centre (PATREC).

Background and objectives

Western Australia features almost 3,000 bridges, of which Main Roads WA owns 42%. The vast majority of these are located at considerable distances from Perth, and thus only inspected every several years.

This project aimed to investigate the feasibility of using an integrated technology package to support smart bridge health monitoring and maintenance, such as measuring bridge displacement responses. It aimed to provide ongoing monitoring and reduce the need for manual inspections.

Project objectives were:

1. Develop loT-based solutions for collecting, visualising and transmitting vibration response data;
2.  Develop computer vision-based solutions for measuring bridge displacement;
3.  Develop an integrated proof-of-concept model, based on machine learning, for predicting bridge displacement under heavy traffic loads; and
4.  Make recommendations on new techniques for efficient bridge monitoring, a wider roll-out of the technology and suggested next steps.

Methodology and technical details

The project’s technologies included loT-based vibration sensors, vision-based techniques for measuring bridge displacement and machine learning techniques for relating the displacement measurement data with traffic load.

The vertical displacement (deflection) under traffic load is usually selected as a critical parameter for evaluating bridge performance. Thanks to rapid advances in computer vision, non-contact vision sensing has emerged as a promising alternative to conventional contact sensors for structural dynamic response measurement and health monitoring.

After issues with the Libelium Waspmote and RAK wireless WisTrio RAK5010 single-board computers (SBC), the Arduino MKR NB1500 was selected as the foundation of the IoT sensor. It transferred data to AWS cloud storage via the Telstra Narrow-Band IoT network.

A Sony PXW-FS5 camera with a Sony ALC-SH135 28-135 mm lens and in-built GPS module was selected for visual monitoring. A portable 606 Wh power station allows 24 hours of continuous recording.

Traditional camera calibration requires installing a calibration panel on the target structure, which can be difficult to do in the field. Instead, the prototype system selects a Region of Interest on the bridge, containing multiple high-quality feature points, from the first video frame.

Lab and in-situ validation

This project included laboratory and field tests to validate the accuracy of the IoT sensor and the vision-based displacement tracking.

During laboratory tests, the ground truth data for displacement was measured using a laser displacement sensor and compared to the displacement measured by the vision-based method.

A series of in-situ bridge tests were also carried out on the Stirling Bridge, consisting of seven spans, which carries the Stirling Highway over the Swan River.

Machine learning

When comparing the linear and machine learning models, both performed similarly despite the fact they are very different in nature. Given the simplicity and interpretability of the linear model, it seems like a sensible choice, considering their similar performances.

In some cases, the linear model outperformed the machine learning model and vice versa. This led to the belief that further refinement and combination of both techniques, for instance using stacking or blending, might yield better results.

Report findings

Laboratory tests verified the target-free vision-based method can accurately and effectively identify structural displacement responses by tracking natural features on the structural surface.

The project also demonstrated the accuracy of cost-effective IoT-based accelerometer sensors for data collection, transfer over IoT link and storage at AWS S3.
Further in-situ validation results revealed that the vision-based displacement subjected to traffic load aligns well with the traffic pattern. However, accuracy can be affected by environmental factors.

To mitigate this, the project recommends using a relatively heavy and solid camera tripod, utilising a case to cover the video camera to avoid wind effects and developing signal processing techniques to eliminate camera motion-induced displacement identification errors.

To mitigate inaccuracies resulting from poor light conditions, it is recommended to adjust the filming angle, shield the camera and avoid filming during the times of intense direct sunlight.

Conclusion

The project provides a solid foundation for the development of integrated IoT and computer vision for smarter bridge health monitoring and prediction. Further research and expansion of this project can apply the developed methodologies to a wider population of bridges in Western Australia and beyond.

The research produced a proof-of-concept to demonstrate the efficacy and feasibility of an integrated package of technologies for first-level bridge health screening and as an early warning system, reducing the need for traditional physical inspections and instrumentation.

The benefits of the project include contributing to reducing maintenance, operation costs and risk, while achieving a safer transport infrastructure network and increasing productivity.

Recommendations for further research

There are several areas where further research can be considered to fully realise the potential of the developed techniques, :

1. The IoT-based sensor data transmission to AWS needs further investigation to reduce transmission time;
2. Bridge performance indicators and safety thresholds should be defined, to estimate the performance of bridges and facilitate decision-making;
3. A graphical user interface would allow bridge inspectors and engineers to interpret data and make informed decisions about bridge maintenance and repair strategies; and
4. Monitoring of more bridges under different conditions is necessary to develop a comprehensive database of experience.

Expected project impacts

On completion of the project, the Main Roads (WA) chair of the project steering committee reported that the project had successfully demonstrated that the advanced algorithms and technologies can help Road Authorities to inexpensively monitor bridges and their performances.

As an integrated package, IoT, computer vision and machine learning technologies offer to supplement physical bridge health assessment particularly in remote regional contexts which can be costly, time consuming and unsafe to inspect.

Conducting regular, efficient, and reliable bridge health monitoring is essential for the long-term protection of valuable road assets through timely maintenance responses.

Download the report

Download your copy of the final report, Integrated IoT, computer vision and machine learning technologies for smarter bridge health monitoring and prediction, by clicking the button below.