A Machine-Aided Framework for Real-Time Gas Leak Detection: Case Study of the JK-52 Gas Processing Plant

International audience Natural gas is considered a fossil energy source from under the earth's surface. Its main component is methane, but it can contain non-hydrocarbon gases and natural gas liquids as well. Leaks are considered very dangerous since they can build up into an explosive concentration. Gas leaks can be hazardous to health as well as the environment. Leak detection is a method in which the existence of a leak within a system is determined. Many fluid leak detection mechanisms rely on observation of volume changes and physical evidence of a leak, which may take hours, days and sometimes weeks or months to be seen. This is a concern in gas plants where the proximity of the leakage may constitute environmental pollution as well as health hazards for personnel in the vicinity. Economic losses have also resulted from delays in mitigating a gas leak problem due to late detection.This study applies a machine learning technique to develop an algorithm that can detect gas leaks in real-time, where the only possible delay is the lag-time between the inlet gauges at the upstream valve and the outlet gauge at the downstream valve. The dataset for the study was collected, and a pre-processing phase was performed, which included cleaning the data, attempting a linear regression model and other regressions (Random Forest). This study also proposed a model and evaluated its performance. In this case study of the JK-52 gas processing plant, the difference in pressure gauge readings was calibrated against the volume of the gas in the inlet section to quantify the leak volume. Because gaseous fluids do not present a physical indication of volume, a pressure-based method was used for the detection, where a drop in gauge pressure due to depressurisation indicates leakage in the absence of a recorded gas supply or collection. To build the model, the dataset was divided into samples to train and test the model. Python coding language, using Jupyter and PyCharm Integrated Development Environments (IDEs), was used for the programming. The machine learning algorithm analyses the incoming streaming pressure versus time datasets from the gauges during the residual and ramp-up flow phases to set the acceptable pressure difference cut-off. A minimum difference in gauge reading may be normal within an acceptable error margin. The change in the consistency of reading within this acceptable window defines the tolerance. The system is set up to blare an alarm when there is leakage, usually based on a cut-off or tolerance, to be detected by the machine-aided process. Even if no immediate event triggers the alarm, a leak can still be suspected and later confirmed through further analysis. Over time, the model becomes predictive, improving detection accuracy as it learns from experience.