AI Technology

King’s and University of Surrey | Deep Learning Improves Structural Health Monitoring of Civil Infrastructure

This research demonstrated that deep learning can contribute to the traditional discipline with much better performance than existing methods.

Communities and countries depend heavily on major infrastructure such as dams and bridges, but these things do not last forever. Just as a doctor monitors a patient’s health condition with an annual check-ups, someone or something has to regularly check on major infrastructure — what’s called structural health monitoring (SHM). The process of manually inspecting structural conditions and damage however is time-consuming, error-prone, and can be dangerous for the human inspectors.

A team of researchers from King’s College London and the University of Surrey have introduced a new deep convolutional neural network, SHMnet, for identifying damage to bolted steel frames based on the vibration response at just one point of the structure, which is a particularly challenging structural condition determination. The novel research approach is based on AlexNet, with a few adjustments, and is published in the paper SHMnet: Condition Assessment of Bolted Connection with Beyond Human-level Performance.

Paper co-author Ying Wang is an Associate Professor in the Department of Civil and Environmental Engineering at the University of Surrey. Wang told Synced that due to the large investment, long service life and environmental and other challenges and uncertainties associated with civil infrastructure systems, it is necessary to hold them to very high safety standards.

Although the civil engineering profession may be conservative and skeptical of new technologies, the timely and efficient identification of potentially problematic structural conditions can surely help infrastructure asset managers better maintain these infrastructure systems. Explained Wang “This research demonstrated that deep learning can contribute to the traditional discipline with much better performance than existing methods. With sufficient training data, SHMnet could be used to identify structural conditions with very high accuracy.”

image.png
image.png
image.png

Bolted connections have been used in metallic structures for centuries, and the recent development of off-site modular construction techniques has made bolted connections even more practical and popular.

Once metallic structures enter the service stage, almost all their bolted connections will lie outside their ideal condition, i.e. semi-rigid, which can result in differences between simulated performance results and actual structural performance. Bolted connections also naturally deteriorate over time due to fatigue, corrosion, and other factors. Even one damaged bolt can result in significantly reduced structural performance, and moreover accelerate the degradation of other nearby bolts.

Efficient and accurate condition assessment of bolted connections in metallic structures is therefore critical for the effective maintenance and safe operation of these structures.

The research team set up vibration and impact hammer tests to examine how accurately SHMnet could identify subtle condition changes of bolted connections on a steel frame under 10 different damage scenarios (one bolt loosened, etc.). They found that SHMnet significantly improved the identification performance across all scenarios compared with traditional modal-based methods, and scored up to 100 percent identification accuracy in some tests.

The-location-of-sensors-and-impact-hammer-for-the-modal-testing.png
Sensor and impact hammer placement for modal testing
image.png

Wang identified a couple of priorities for further research, starting with data. “Civil infrastructure systems are usually unique, as materials such as concrete are somewhat random in nature. Therefore, unlike with mechanical systems, monitoring for civil infrastructure system needs to be specifically designed, and real training data for different damaged cases may be limited. How to train a reliable model with limited data is the first priority in this field. Another area for future research is transfer learning, which can contribute to the first effort. With high fidelity numerical simulations, we can simulate more damage scenarios than real cases. By using numerical data as training data together with transfer learning, we may be able to significantly improve the research outcome.”

The paper SHMnet: Condition Assessment of Bolted Connection with Beyond Human-level Performance is on researchgate. The Python Code for SHMnet is open for downloading on GitHub.


Journalist : Fangyu Cai | Editor: Michael Sarazen

Advertisements

0 comments on “King’s and University of Surrey | Deep Learning Improves Structural Health Monitoring of Civil Infrastructure

Leave a Reply

Your email address will not be published.

%d bloggers like this: