Detecting Fugitive Gas Emissions from IIoT Telemetry

Abstract

Abstract The oil and gas industry faces increasing pressure for continuous, measurement-informed methane emissions monitoring, driven by voluntary frameworks such as OGMP 2.0 and regulations such as the U.S. EPA's OOOOb rules (U.S. EPA 2024b). Optical Gas Imaging (OGI) is the established detection method, but ubiquitous continuous deployment faces severe scalability barriers due to high capital costs, heavy edge-compute requirements, and environmental vulnerabilities (e.g., performance degradation during high winds, rain, snow, or low thermal contrast). The objective of this paper is to demonstrate a scalable, software-first daily screening methodology. By utilizing existing Industrial Internet of Things (IIoT) telemetry to execute dynamic, process-level anomaly detection, this framework targets fugitive emissions at per-well resolution without the capital expenditure or weather constraints of fixed optical hardware. In this study, per-well supervisory control and data acquisition (SCADA) data (flow meters, pressures, temperatures) are ingested into time-series foundation models that compute a daily well-level forecast residual coming from observed telemetry against each well's expected normal behavior, predicted over a 72-hour horizon. Statistical process control and machine learning anomaly detection algorithms monitor these residuals for persistent deviations exceeding cumulative model uncertainty thresholds. Temporal aggregation improves detection sensitivity from instantaneous meter noise levels to sustained losses accumulated across the scoring horizon. To validate the model, out-of-sample testing was conducted. Operator-recorded leak events, logged independently of the telemetry, were paired with the IIoT anomaly detection algorithms to establish a field ground-truth correlation between recorded leaks and process data deviations. Field validation demonstrated that the IIoT anomaly detection model independently identified fugitive emission events without OGI cameras. The approach detected recorded leak events purely through deviations in the wells' own process telemetry, attaining an area under the receiver operating characteristic curve (AUC) of 0.894 across 90 wells. On similar events, OGI would be constrained by cost, visual coverage, and weather, none of which affects this method. The detection envelope analysis reveals that while instantaneous meter uncertainty limits single-reading sensitivity to approximately 3–5% of facility throughput, temporal aggregation over a 72-hour window removes this limitation. On 57 wells with no recorded leak, using this 72-hour window produced 1.221 false-alarm incidents per well-month at 80% leak recall. This paper demonstrates that IIoT data operators already collect can screen wells daily for fugitive methane leaks, without dedicated optical camera or atmospheric sensors. This methodology introduces a time-series forecasting approach with no installed hardware, shifting continuous leak detection and repair (LDAR) from a hardware-intensive paradigm to a software-based screening layer.