Enhancing Optical Gas Imaging (OGI) Leak Detection by Transforming Temporal Dynamics into Spatial Features Using Channel Stacking

Le résumé fourni par la source

Abstract This research confronts the persistent challenge of reliably detecting subtle and diffuse invisible hydrocarbon gas plumes using Optical Gas Imaging (OGI) in the oil and gas industry. While OGI is indispensable for visualizing such emissions, accurately identifying leaks from single, static infrared frames is often compromised by the inherent limitations of single-channel data and the complexity of industrial backgrounds. Conventional image processing techniques frequently fail to isolate gas plumes effectively under these conditions. In our previous work we show full video processing can capture essential plume motion, however its significant computational overhead typically prohibits real-time deployment and misinterprets non-gas related movements as a gas leak within the scene. This paper introduces "Channel Stacking," a novel and efficient methodology designed to significantly improve OGI-based leak detection. By transforming critical temporal gas motion dynamics into robust spatial features, this approach enhances detection accuracy and reduces false positives while using much less computational resources. The Channel Stacking technique involves capturing a sequence of consecutive single-channel infrared frames from standard OGI video. These frames are then systematically stacked to synthesize a single multi-channel, RGB-like image, wherein each channel explicitly represents a distinct temporal point in the sequence. This transformation encodes the plume's motion over a defined time window into a rich, spatially-coherent representation. A deep learning model, subsequently trained on these Channel Stacked images, learns to discern the unique spatio-temporal signatures characteristic of gas plumes from background noise and other environmental variations with enhanced robustness. This method advances beyond previous temporal deep learning approaches by creating a more direct and computationally efficient spatial encoding of motion dynamics. Notably, it utilizes approximately less than half of the computational resources required by more complex temporal deep learning models, significantly improving its scalability and practical viability for applications. The efficacy of the Channel Stacking model was rigorously evaluated using a comprehensive and diverse OGI dataset. This dataset contains more than 33000 images acquired from multiple operational sites, encompassing a wide spectrum of environmental conditions, including variations in weather (e.g., clear, rain, snow), ambient temperature, and wind speed. A critical aspect of the evaluation focused on the model's capability to reliably distinguish genuine gas emissions from other motions, such as thermal signatures from flares or hot equipment components, which is crucial for minimizing false alarms in real-world industrial scenarios. The Channel Stacking approach demonstrated a substantial improvement in operational reliability, reducing false positives by nearly 50% when compared to the temporal video processing model. Concurrently, a detection accuracy of more than 91% was achieved under these conditions. This work presents a significant advancement in OGI-based leak detection by introducing Channel Stacking, a method that uniquely translates the temporal behavior of gas plumes—often the most distinctive indicator of a leak—into spatially-encoded features amenable to efficient deep learning analysis. Unlike traditional single-frame or temporal deep learning models, this technique provides a richer, motion-aware input. This enables deep learning models to achieve superior differentiation between actual gas plumes and confounding environmental visual noise, leading to more robust and accurate leak identification. The proposed method offers a practical and computationally efficient pathway to enhance OGI system performance, contributing directly to improved environmental stewardship, operational efficiency, and overall safety within the demanding landscape of the oil and gas sector.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.