Gas Leakage Recognition Using Manifold Convolutional Neural Networks and Infrared Thermal Images

Gas leakage detection in different industrial sectors is enormously important for safe operation. It is vital to quickly and automatically detect and identify the type of gas in order to prevent environmental damage and protect human lives. Existing approaches mainly rely on electronic noses which have several limitations and should be kept within the leakage region. Lately, novel approaches for gas detection have been proposed based on thermal infrared sensors which can capture heat patterns at a distance far from the gas leakage. Motivated by the success of artificial intelligence such as deep learning in several industrial applications. Combining deep learning with infrared thermal images could effectively improve gas leakage detection accuracy. In this study, a deep learning-based pipeline is proposed based on thermal infrared imaging to detect gas leakage and differentiate between gas categories. Multiple convolutional neural networks (CNN) models are used in the proposed pipeline for feature extraction leading to spatial deep features. These features are then analyzed via the fast Walsh Hadamard transform (FWHT). Next, these features are integrated using principal component analysis and then fed to several machine learning classifiers which are used for gas detection. The detection accuracy attained via the proposed pipeline is 98.0% which suggests that the proposed integration method has improved the performance of gas detection.