Corrosion is natural, and it’s everywhere. While there are many techniques that can be applied to stave off corrosion, nothing lasts forever when exposed to the elements of nature, so consistent and planned evaluation is essential to keep the things we build in good shape. A research team at the University of Illinois (U. of I.) led by Shengyi Wang, a Ph.D. candidate in the Department of Civil and Environmental Engineering, recently used the National Center for Supercomputing Applications (NCSA) Delta supercomputer, an ACCESS-allocated resource, to help streamline the process of evaluating infrastructure for corrosion.
You probably have a bridge in your town, and you certainly have pipelines and water systems. Your city is but one of tens of thousands of cities across the country – every one of them with infrastructure that needs to be maintained. Maintenance is costly, especially when you consider everything that has to be looked after.
“Corrosion poses significant challenges to various infrastructure assets,” Wang said, “including bridges, pipelines, military equipment and water systems. It can lead to safety hazards, substantial economic losses and environmental risks. Notably, the United States allocates 40% of its National Maintenance Budget to corrosion-related repairs.”
Wang’s research involves training an AI using images with and without labels. The idea is that you give some guidance to the AI, in this case, pictures of corrosion that are labeled by human experts, and then allow the AI to learn by example how to detect corrosion in unlabeled pictures. This method is called CNN-based semi-supervised learning (SSL). Wang further explains how he’s using this method in his research.
“The methodology involves collecting high-resolution digital microscopy images of corroded steel panels, which were subjected to accelerated weathering per ASTM D1654 and ISO 12944 guidelines. These images are then annotated, processed, segmented, and augmented for training. A mean teacher-based SSL method using DeepLabv3+ with a ResNet-34 backbone is employed, allowing the model to learn from both labeled and unlabeled data. Additionally, a patch-merging smoothing module is introduced to integrate high-resolution image patches seamlessly and reduce edge artifacts. The model is tested using multiple performance metrics, including precision, recall, F1-score, and IoU, and compared with state-of-the-art models.”
Wang’s team recently published their work in the journal Structural Health Monitoring, and the team continues to refine their work. As they move forward, they’ll incorporate more real-world corrosion images to enhance model generalization. Wang also intends to make the AI more adaptable to different scenarios.
“I will also implement domain adaptation techniques to improve the model’s ability to perform well across different environments and corrosion types,” he said. “Additionally, I will explore transformer-based models for improved feature extraction and segmentation accuracy.”
He also plans to work with industry partners to test out the model in the field. “I will collaborate with industry partners such as the USACE and infrastructure maintenance teams to deploy and test the AI system in real-world corrosion monitoring applications.”
A lot of Wang’s work couldn’t be completed in such a short time without ACCESS-allocated resources. High-Performance computing helps research teams like Wang’s get their results quickly, speeding up innovation across all domains that employ research computing.
The ACCESS program has significantly accelerated my research on developing deep learning-based corrosion detection. By utilizing high-performance computing resources, I am able to efficiently develop and train complex deep learning models on high-resolution images, substantially reducing computational time. ACCESS enables research that would otherwise be hindered by hardware limitations, fostering innovation in infrastructure monitoring and maintenance.
–Shengyi Wang, PhD candidate, Department of Civil and Environmental Engineering, U. of I.
“The NCSA Delta GPU system helped a lot in accelerating my research,” said Wang. “Training deep learning models on high-resolution images is computationally expensive, and using High-Performance Computing (HPC) resources significantly reduced training time from several days to under 20 hours per experiment. The large-scale memory and powerful GPUs allowed me to process large image datasets efficiently without considering hardware limitations too much, and enabled me to focus on developing the novel method to solve the problem.”
You can read more about this story here: Safer Infrastructure Through AI
Resource Provider Institution(s): National Center for Supercomputing (NCSA)
Resources Used: Delta
Affiliations: University of Illinois Urbana-Champaign
Funding Agency: NSF
Grant or Allocation Number(s): CIV230015
The science story featured here was enabled by the U.S. National Science Foundation’s ACCESS program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296.