Detection of Leaky Pneumatic Actuators in Real-Time for Fugitive Methane Emissions Management

Abstract

Abstract Pneumatic actuators are pervasive in oil and gas infrastructure, from wellhead sites to transmission lines. Domestic field studies (industry and government stakeholders) identify them as the leading culprit for unintended fugitive methane leaks. We developed a real-time computational approach that utilizes streaming observations of an actuator’s inlet pressure, rate, and displacement to identify whether it is leaky. The approach can be used in integrated modeling workflows to account for the implications of controls on emissions. Our approach is informed by the fluid-structure interaction (FSI) between compressible gas flow and diaphragmatic elastodynamic. Canonical models are established with a dimensionless parameter space, given cataloged device specifications. High-fidelity three-dimensional FSI simulation is applied to generate a large synthetic dataset spanning operations of the pneumatic actuator under normal and dysfunctional settings. The dataset provides a statistically meaningful sampling of diaphragm rupture events during transient and steady periods with the addition of noise to account for measurement errors and background vibration. Dimensional analysis is applied to design predictor features, and several supervised-learning methods are applied within a hyper-parameterized dimensionless space. Training, validation, and testing are performed, and confusion matrices and prediction accuracy are computed to assess the predictive capacity of the methods. The accuracy of the methods ranged between 80% and 97% for binary leak/no-leak predictions. Predictions for multicategory leaks (tear geometry and size) show improved accuracy with tear size. Combined pressure transient and elastodynamic predictors improve performance significantly compared to the use of acoustic data only. Generalization within the spectrum of relatively narrow designs is promising. Previously reported leak detection methods exploited the dynamics of rapid negative pressure waves as a key discriminator. However, given the rapid timescales and sensitivity to sensor location, predictive accuracy is limited. This work augments acoustics with mechanics to obtain a stark improvement in predictive accuracy. Furthermore, the predictors are evaluated exclusively using process flow observations and device specifications, rendering our methods amenable to use in integrated surface facility models and SCADA systems.