_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper OTC 34756, “Temporal Deep-Learning Image-Processing Model for Natural Gas Leak Detection Using OGI Camera,” by Mehdi Korjani, David Conley, and Mark Smith, Clean Connect. The paper has not been peer reviewed. _ The authors have developed a novel deep-learning (DL) image-processing model that uses videos captured by a specialized optical gas imaging (OGI) camera to detect natural gas leaks. The temporal DL algorithm is designed to identify patterns associated with gas leaks and improve its performance through supervised learning. Modeling This study was designed to exploit the technical strengths of OGI technology in tandem with temporal DL techniques. The central objective of the study was to engineer a robust model capable of accurate and efficient detection and quantification of methane leaks from video data. This resulted in the inception of the temporal DL image-processing model (TDLP-NG). The operational phase commenced with the strategic deployment of OGI cameras at select locations throughout natural gas facilities. Equipped with specialized infrared sensors sensitive to the spectral signature of methane gas, these cameras served as the eyes of the study. Their arrangement considered a multitude of variables—distance, angle, and environmental factors—to calibrate the instrumentation for optimal gas-leak visualization. TDLP-NG. The cornerstone of the methodology is the discussed DL architecture crafted to detect and analyze methane emissions within OGI video data. The model harnesses the combined power of convolutional neural networks (CNNs) and long short-term memory (LSTM) units to deliver spatial and temporal analysis capabilities. The model’s CNN component was designed to extract high-level feature representations from individual frames. The temporal aspect was addressed by integrating LSTM units following CNN feature extraction. LSTMs are an ideal choice for capturing the dynamic evolution of methane patterns over time. The LSTM layers analyzed the sequence of CNN-derived features to predict the presence and characteristics of emissions. A custom loss function was designed to simultaneously optimize the spatial precision of detected plumes and the temporal correlation across frames. The model was trained on a curated data set using backpropagation through time. Validation and Testing. The validity of the TDLP-NG’s detection was comprehensively assessed through a validation phase involving a set of reserved video data. A comparison between model predictions and expert annotations served to fine-tune the model parameters and solidify its detection efficacy. A battery of tests also was conducted on entirely unseen data to challenge the TDLP-NG’s generalization capabilities. Rate Estimation. To estimate gas-leak rate using an optical flow-based model, various factors were considered, including the properties of the plume captured by OGI cameras, the distance of the camera from the leak, and the camera’s lens angle. Each pixel of the image represents larger areas for more-distant leaks vs. close ones. A workflow for estimating gas-leak rates with an optical flow-based model is provided in the complete paper. By integrating optical flow with machine learning and accounting for camera specifications, the model provides accurate and robust leak-rate estimations. Gas properties such as temperature, pressure, and methane composition also are integrated to further refine the estimation of the leak rate.
Read full abstract
Jun 1, 2025
Yan Yin + 4
Just Published
Open Access