Using Machine Learning for Methane Leak Detection and Quantification

What is Methane Gas?

Methane is a colorless, odorless gas that occurs commonly in nature. Its chemical formula is CH4, a member of the paraffin series of hydrocarbons, and a potent member of known greenhouse gases. Now that you have a good understanding of what it is, let’s discuss the challenges of detecting and quantifying methane.

Where can we find Methane gas?

Methane leaks can occur near valves, compressors, and pipelines where natural gas is gathered, transported, or distributed. It is a common “fugitive emitter,” which typically means gas leaks from a pressurized environment ranging from underground systems to heavy industrial equipment. Methane can also be found normally in nature. It’s produced by the anaerobic decomposition of vegetable matter and is also produced by animals such as cows and even termites.

While methane is hard to see, it is important to keep track of where it is in the world. Long-term undetected methane leaks can lead to serious environmental and safety hazards due it being a flammable greenhouse gas.

How do you Detect Methane?

Older technology used “sniffer” techniques that required instruments to be placed directly in the path of methane to detect and determine concentration levels. This process was laborious and made it difficult to accurately find the source of fugitive emissions. Operating a sniffer is hazardous to humans because it requires close interaction in methane rich environments.

Currently, methane leaks are detected by routine manual inspections which survey methane-energy plants with optical gas imaging (OGI) capable cameras. An OGI is similar to a video camera that operates in the midwave infrared (MwIR) band (methane band), allowing operators to scan a site from a distance and visualize the otherwise invisible methane gas. It is important to note that methane is several times more absorbent than Carbon Dioxide (CO2) and Nitrous Oxide (NO2), other prominent chemicals in this band that make the midwave spectrum optimal for detecting methane. Methane is invisible to the human eye, but in the midwave band methane energy is absorbed, making it appear as a smokey gray/black plume in the image. OGI cameras are expensive, and methane is still difficult to see in the midwave spectrum, thus special training for camera operation is required for methane plant inspectors.

One of the problems with inspection using OGIs comes from motion. Detecting methane from a single static image is difficult. Even with the methane absorbing energy, it can appear translucent on an OGI image and blend in with its surroundings. One of the best ways to find a leak is by observing the methane’s motion in the image(s) relative to everything else that is static in the image. This problem is exacerbated by the fact that inspectors are typically using these cameras by hand, or more recently, on an unmanned aerial system (UAS) platform. This makes spotting a methane leak a difficult task, without stopping and staring at any given component on a site. To help in improving the efficiency and reliability of methane inspections, SwRI partnered with the Department of Energy’s National Energy Technology Laboratory to apply machine learning and deep learning tools to this space.

Why use Machine Learning for this problem?

Firstly, machine learning, or ML, is automatic and does not require a human to supervise it. This means that a camera can be deployed in areas that may have methane leaks, not require a worker to be placed in a potentially hazardous environment and can survey for leaks 24/7. Automated detection also has the benefit of their algorithms never getting bored, tired, or distracted while accomplishing a task, which is a significant improvement over a human inspector. ML also leverages computer vision techniques to look at many frames of the OGI video at a time, making it sufficiently easier to find methane in images. Computer vision also paves the way for more complex detection methods, such as from an aerial drone or even from space by applying preprocessing to get rid of the noise in the image. Machine learning can also be deployed on UAS in real-time allowing for quick actionable information when methane is detected. 

Practicalities of applying Machine Learning to a Problem

The application of ML to any domain requires data, and most of all understanding that building a robust algorithm takes time—and understanding that it will fail before it works. Quality data has many components, including the presence of enough diversity to represent the problem in a robust manner covering all probable operating conditions, and having a lot of examples embedded in the data to learn the desired task. This can be a challenge in the Oil and Gas Industry as leaking natural gas can be expensive, so it’s important to be able to design good models that can excel with minimal training data.

To begin applying machine learning to the problem, the dataset is separated into three parts: training, validation and testing. This data separation allows for tuning of network hyperparameters and preprocessors in training, while using the validation set to evaluate how well the model has learned the problem, and the testing set is used when a model has been selected from the training cycles to determine how well the model performs on data it has not been trained or tuned to.

In order to apply machine learning to the methane leak detection problem, the team had to release controlled methane leaks while observing the leaks with an OGI-capable camera. After collecting the data, the SwRI team had to research many different preprocessing techniques, model architectures, and physical properties to best address this problem. Eventually, the team started to train good models, and the Smart LEak Detection (SLED) system was born. SLED detects both liquid hydrocarbon and gaseous leaks.