Aplicação de redes neurais convolucionais para detectar vazamentos de gás natural em imagens de cabeça de poço petrolíferos onshore

Abstract

Large natural gas leaks lead to accidents that often are lethal to humans and highly destructive to property, as well as release methane into the atmosphere, negatively impacting our environment. Therefore, the detection of natural gas leaks is an important action of the oil industry to prevent accidents. The literature provides different techniques for detecting natural gas leaks. However, except for previous studies proposed by the author of this thesis on the subject, there is still a gap in the literature on the detection of clouds of natural gas leaks through digital images, without the need for sensors or specific cameras calibrated for the spectrum of methane molecules. In previous studies, image processing techniques associated with a novelty filter classifier were used to investigate the presence or absence of a visible cloud of hydrocarbon vapors, that is, a plume of natural gas in closed circuit television frames installed on terrestrial oil wellheads. In this thesis a new method for detecting visible clouds of hydrocarbon vapors is presented, which consists of convolutional neural networks, which are applied to classify images (closed circuit television frames) as belonging to classes with or without leakage of natural gas in onshore wellheads, improving the results obtained previously. To carry out the leak detection task, the study proposes and presents the results of the three convolutional neural networks architectures, in addition to investigating the learning transfer technique with the pre-trained architectures DenseNet-201, GoogLeNet, MobileNetV2 and ResNet-18, and also the feature fusion technique between these pre-trained models. The study also provides a visual explanation providing the region of the input image that was relevant for the classifier to decide that a leak occurred, through the Gradient-weighted Class Activation Mapping Algorithm. The experimental results showed that the best performing model presented an accuracy of 100.00% and a false negative rate of 0.00%, surpassing the performance of previous methods of detecting leakage of natural gas by closed circuit television images.