New Technology Uses Machine Learning to Find Small Leaks in Oil Pipelines

R&D 100 Staff
May 3, 2018, 3:22 p.m.

While large oil leaks are obvious environmental problems that cost millions of dollars to fix, smaller leaks, which are often overlooked, also pose a problem. Small leaks—defined as leaks that make up less than 1 percent of the flow volume that is traveling through the pipeline per day—can still make a significant impact, said Maria Araujo, the R&D manager at Southwest Research Institute (SwRI)

“Some of the largest pipelines carry about 800,000 barrels a day, 1 percent of that is 8,000 barrels a day, which is a lot,” she said in an interview with R&D Magazine. “The technology that exists today, the primary one being computational pipeline monitoring, are unable to detect leaks that are less than 1 percent because it is below their detection capability.”

Araujo and her team at SwRI have created a real-time autonomous system called the Smart Leak Detection (SLED) system to combat this problem. The system uses optical sensor data and machine learning to detect small liquid pipeline leaks and differentiate between different hydrocarbons that may be leaking out.

SLED, which won an R&D 100 Award in 2017, can be used in conjunction with existing leak detection technologies to detect small leaks of crude oil, diesel, gasoline and mineral oil, said Araujo.

“This technology would be complimentary, existing systems are still effective in detecting the larger leaks,” Araujo said. “[Creating SLED] was challenging because it can actually differentiate between different types of hydrocarbons.”

SLED detects leaks using machine learning techniques and infrared and visible light sensors, which can detect the chemical fingerprint of small liquid leaks from the reflectance response of typical petroleum products.

The system could eventually be deployed at stationary platforms of pumping stations, which are located approximately every 50 miles along pipelines and are considered high-risk areas for leaks because valves and other specialized equipment can fail.

The technology can also be deployed on drones, helicopters, and manned aircrafts to fly over long stretches of pipelines between stations.

Araujo explained that while the color of crude oil makes it easier to detect, translucent hydrocarbons like gasoline are difficult to visually detect because they are indistinguishable from water once they hit the ground.

Araujo and her team built SLED based on their decade-plus focus on developing new technology for autonomous vehicles.

“When I learned about this problem, I wondered if we could apply some of that expertise we have in computer vision and machine learning and apply it to this problem,” she said.

SLED development team
However, the researchers had trouble in the early iterations of SLED with different environmental conditions. They eventually developed sensors that can image a variety of different hazardous liquids in different operational environments like grass, pavement and gravel, under various weather conditions, including lightning and high temperatures.

“When we started the algorithm development we had some initial successes, but then we noticed that you had a cloudy day or shaded conditions it wouldn’t do as well,” Araujo said. “We switched the algorithm architecture to go into deep learning and that improved things a lot.”

Araujo said the next step for the researchers would be to refine the technology and eventually commercialize it.

“Right now we are in discussions with two or three organizations discussing commercializing this technology,” Araujo said. “We need [pipeline operators] involved in this, we need to field test this technology. Working with the operators is a key item.”

Araujo explained that SLED currently ranks at a six or seven on the Technology Readiness Level scale, a rubric that ranks emerging technologies on a one to nine scale based on how ready they are to be commercialized.

She also said there is a similar project ongoing with the U.S. Department of Energy to create a detection system for methane leaks.

Between 2007 and 2012, leaks in the U.S. hazardous liquid pipelines network exceeded 100,000 barrels a year, a 3.5 percent increase from the previous five-year period.