Using convolutional neural networks for structural safety

Brown University researchers have developed a new technology for evaluating the structural integrity of metal structures, such as pipelines, to prevent catastrophic failure.

Sudden failures of load-bearing structures can cost the oil and gas and other industries billions of dollars a year and have dire environmental and human consequences. In two studies published in the International Journal of Solids and Structures and Engineering with Computers, respectively, Vikas Srivastava, a Howard M. Reisman Assistant Professor of Engineering at Brown, and Sijun Niu, a recent Ph.D. graduate, demonstrated that applying machine learning (ML) techniques to pipeline quality control can vastly improve the ability to detect and characterize non-visible flaws in metal structures. 

Flaws refer to a variety of small problems that can occur in metal structures. It can include tiny cracks, corrosion, and more. “If pipes carrying hazardous materials have cracks in them, then they can grow to cause leaks or ruptures, and those are environmentally very catastrophic,” Professor Srivastava said on the issue of structural failure. “People’s lives are at stake because if it is natural gas and you have a leak and an explosion, it can take several blocks out.”

Currently, the industry tests for structural flaws using ultrasounds. The ultrasound wave will have small perturbations due to flaws. These waves are then analyzed by humans. But since the changes in the waves can be extremely subtle, they are prone to interpretation error. 

“The current uncertainty in the detection and quantification of hidden flaws is quite large,” Srivastava said. “Forget about measuring the (flaw) length, (people) will miss detecting some of the big cracks.”

Srivastava and his colleagues wanted to see if the use of convolutional neural networks (CNN), a type of ML technology, could be applied to the traditional practice of testing for flaws using ultrasound waves, a non-destructive technique. In order for a CNN to work, it would need to be trained using a variety of sample flaws. The experimental or field data is not available. So Srivastava and his team replicated thousands of experiments on the computer to create a data set that could then be used to train the CNN. 

A notable distinction in Srivastava’s CNN is that their CNN evaluates metal structures through a one-dimensional signal. This makes his data collection/inspection and CNN run far faster than other CNNs running through an image-based 2D signal. In a field where this technology could be used to potentially check hundreds of miles of pipelines, this speed for fast, non-destructive technique scanning is essential. 

Once the data set had been made and the CNN was trained, Srivastava and his team could go on to experiment with real-life data. He and Niu set up lab samples with embedded cracks of different sizes, in different locations, and with different orientations. They used ultrasound equipment used in the industry and applied simulation-trained-CNN to see if it could spot and quantify the cracks. The results of the experiment were incredibly promising, CNN could find the crack length, location, and orientation with an accuracy of more than 90 percent. 

The entire process of conducting both modeling and experiments took more than two years to complete.

Still, Srivastava and his colleagues decided to take the simulation and CNN-based technology a step further. 

“There are a lot of different kinds of flaws, but the two flaws that are most critical in metallic, steel, and iron-based structures are cracks and corrosion,” Srivastava said. “We wanted to show that (the CNN) can also classify not only a single crack but corrosion and crack with corrosion. We wanted to extend this work to what we call ‘multiple interacting anomalies.’”

The second paper followed a similar structure to the first. Srivastava and his team trained this CNN on a data set of 2,000 ultrasound signals. Once it was trained, they used lab samples to test the CNN’s effectiveness in catching no flaws, single cracks, single wall corrosion, two cracks, and combined cracks with corrosion. Here, the CNN demonstrated close to 100 percent accuracy.

Both papers demonstrate that CNNs could be a viable, and far more effective way of testing for  and characterizing flaws in load-bearing structures than traditional methods. But Srivastava urges that more work by the industry practitioners is needed before any CNNs can be applied in the field. 

“We have focused on the fundamental methodology that needs to be further applied to specific problems of interest,” he said. “Neural networks have their limitations. It is only going to predict things that you train it for. A clear understanding of all the variables and well defined application boundaries is essential. If real life has more variations going on that are outside the boundary of what the neural network is trained for, it will not be able to correctly predict it.”

Regardless, it is a very promising start. This initial research is proof of concept that CNNs can be applied to the non-destructive technique field. And the technology Srivastava and his team have developed can be applied to broader problems. Srivastava and his students are now working on applying ultrasounds and CNNs to improve the detection of breast cancer. 

“Ultrasounds using CNNs may be able to help detect the stiffness, the size, and the density of a tumor, just by looking at the signal. As a signal travels through soft breast tissue, it will reflect from any anomaly such as a tumor, and a neural network-based evaluation of the small variations in the reflected signal can help detect and quantify a cancer tumor at a much earlier stage than the current methods,” Srivastava said. 

The research was funded by the U.S. Department of Transportation. The views and opinions expressed in this article are those of the authors and do not necessarily reflect the official policy or position of any agency of the U.S. government.