INL researching machine learning for infrastructure utilization 

Researchers at the Idaho National Laboratory (INL) are using machine learning to teach computers how to recognize utility infrastructure from photographs, including satellite imagery.

The project came about to answer a need, said Shiloh Elliott, a modeling and simulation scientist at INL.

“There’s a large amount of critical infrastructure that’s constantly changing and privately held,” Elliott said. “This is an entry way into automating and creating a machine learning pipeline with high fidelity. The point is to be able to use open-source data and identify pieces of critical infrastructure with a high degree of accuracy.”

What does utility infrastructure look like?

The first part of the project is teaching the system what utility infrastructure looks like. “If you train a machine learning model to identify something, it can go out and do that on new data, data it’s never seen before,” Elliott said. “It’s not perfect, but it does reduce a lot of the manpower.”

The machine learning software, Densenet 161, runs on a high-performance computer using a graphics processing unit, as opposed to the computers typically used in business, which use a central processing unit that performs more than one task at a time. “We ran it once on a CPU, and it ran very slowly, but it did run,” Elliott said.

Once trained, the software can process 1,000 images in about a minute, Elliott said. The data comes from open-source datasets such as those from the U.S. Department of Agriculture (USDA), she said. That data is updated about every three years. “The infrastructure doesn’t change that much in three years unless you have a reconstruction,” she said.

How do you know it’s right?

The other aspect is making sure the system is correct. Elliott described one example where researchers developed a system to tell photos of huskies from wolves, and were delighted at their accuracy — until they discovered all the photos of wolves had snow in the background, and the system was actually learning to recognize snow.

Normally, developers and users would not be sure why a machine learning model is making the decisions it’s making — like how it knows whether a photo is of a wolf or a husky. Consequently, INL is working with the University of Washington to help reverse-engineer the decision the model is making, Elliott said. “That gives us the other degree of trust, or comfort, that the model is actually identifying what it’s supposed to be identifying,” she said. “The idea is we catch those model errors in the development process.”

The project is also partnering with the Georgia Institute of Technology.

What could it be used for?

INL is covering nine facility types of the 16 defined by the Department of Homeland Security, including electricity, water, health care and petroleum and natural gas, Elliott said.

For now, this is just a research project, but in time, it could be turned into a product that utility companies could use. “It’s definitely within the realm of possibility,” Elliott said. “The model we’ve developed could be deployed, whether by Idaho Power or another utility.” INL has not yet started talking to utilities directly about the project, she said.

For example, a company like California’s Pacific Gas & Electric could use the software to help monitor tree growth near transmission lines, to help prevent some of the wildfires the state has suffered in the past few years. Based on the imagery, the system could alert a human to check out a particular sector, Elliott said.

And collecting the data isn’t putting the infrastructure at risk. “We’re not discovering anything you can’t discover from Google Earth. You drive to work, you see a power line,” Elliott said. “There’s nothing proprietary you would be overly concerned about. The USDA isn’t going to put anything in their data that would concern their users. We’re just doing it in an automated fashion to take an accounting of critical infrastructure.”