Methane Leak Quantification on Edge Devices Using Deep Learning

Abstract

Oil and gas domain deals with a varied set of problems ranging from methane leaks to prognostics and health management. This study demonstrates a solution to identify whether there is a methane leak or not and classify on what level the leak occurs based on different flow rates (5.3 g/h – 2051.6 g/h) at 5 distances (4.6 m - 15.6 m), using various spatial and temporal preprocessing techniques and deep learning models. For this study, we are using the GasVid methane leak dataset which consists of videos taken from infrared cameras on 2 different separators with a frame rate of 15 frames per second. Firstly, we applied a series of preprocessing steps, including contrast enhancement, Gaussian blur, three different background subtraction methods, namely moving average background subtraction method, KNN Gaussian mixture model, and Gunnar Farneback optical flow analysis. To save computational resources, temporal down-sampling was applied on the video frames. Thereafter, experiments were conducted using the 3D CNN model by modifying hyperparameters. It was found that on comparing the Adam and Lion (EvoLved Sign Momentum) optimizers, the Lion Optimizer increased accuracy by more than 34% and achieved state-of-the-art accuracy of 41.36% on 4.6 m videos. Further, the moving average background subtraction method's performance surpassed other background subtraction techniques. In addition, applying spatial preprocessing and down-sampling raw videos were compressed from 2.3GB to ~200MB which is a reduction factor of 11.