Temporal Deep Learning Image Processing Model for Natural Gas Leak Detection Using OGI Camera

Abstract

Abstract Natural gas extraction systems often encounter manufacturing defects or develop defects over time, leading to gas leaks. These leaks pose challenges, causing revenue losses and environmental pollution. Detecting gas leaks in the vast array of extraction, transfer, and storage equipment within these systems can be arduous, allowing leaks to persist unnoticed. Additionally, natural gas leaks are not visible to naked eyes, further complicating their detection. We developed a novel deep learning image processing model that utilizes videos captured by a specialized Optical Gas Imaging (OGI) camera to detect natural gas leaks. The temporal deep learning algorithm is designed to identify patterns associated with gas leaks and improve its performance through supervised learning. Our model incorporates algorithms to detect background environments, motion, equipment, and classify gas leaks. Our model employs leak identification algorithms to determine the presence of gas leaks. These algorithms calculate the probability of detected motion indicating a gas leak based on long-term and short-term background subtraction, detected motion, motion duration, equipment location, and telemetry data. To minimize false positives, we have developed image segmentation and object detection models to identify known objects, such as equipment, people, and cars, within the video footage. To train our model we collect more than 10,000 short videos from real fields and include simulated data with known rate controlled gas release in different situations. Data consist of wide range of weather situations including different temperature, wind speed, humidity in sunny, rainy, and snowy fields. We validated our model by conducting experiments involving actual footage from the field. The model achieved a 98% true positive rate, and a 100% true negative rate, correctly refraining from sending an alarm for all non-releases. Additionally, we developed a postprocessing algorithm capable of estimating the gas leak rate based on the volume of gas leaks observed in the video footage and their distance from the camera. Our experimental results demonstrate that the detected leak rates exhibit an accuracy exceeding 78%. By employing this deep learning image processing model, natural gas extraction systems can significantly enhance their ability to detect gas leaks promptly, reducing revenue losses and mitigating environmental impact.