Detection and localization of fugitive emissions in industrial plants using surveillance cameras

Abstract

Industrial plants commonly generate gas emissions that are not caused intentionally. These emissions are known as fugitive emissions. Early detection of fugitive emissions helps to find points of failure in the different processes and avoid sources of pollution, helping to reduce danger to the environment and to respect legislation. Despite the importance of the problem, there are no published solutions in the specialized literature about the location and automated detection of fugitive emissions in industrial plants. Therefore, this article proposes an effective approach based on convolutional neural networks for semantic segmentation. The proposed solution takes advantage of existing surveillance cameras to apply state-of-the-art image segmentation methods, in particular, the semantic segmentation network DeeplabV3 + . This work explores aspects such as the ability to differentiate gases like water vapor and clouds from fugitive emissions, the possibility of reusing models in different industrial plants, the differences between multi-class and binary classification, the importance of proportions in the number of images in each class, the use of weights to balance classes, the comparison of a standard size test versus a real use case test, and the feasibility of an area-based alarm system to warn of emissions. This paper describes a methodology to configure the proposed solution for a specific industrial facility.