Toward Automated Air Leak Localization: A Machine Learning-Enhanced Ultrasonic and LiDAR-SLAM Framework for Industrial Environments

Pressurized air leaks are the source of a significant amount of energy loss in industrial installations. Automating the detection process of these air leaks can help prevent this loss of energy, as detecting and localizing them is the first necessary step in efficient maintenance. Typically ultrasonic microphone arrays are used to detect the leaks, since the escaping air causes turbulences around the leak opening. When the inspection is done manually, the operator of the ultrasonic sensor can evaluate whether or not an ultrasonic sound is originating from an air leak. However, during the automated process of pressurized air leak detection using ultrasonic microphone arrays, many interfering sources arise which lead to the detection of false-positive leak locations. While some of these locations can be easily filtered out using known information of the environment, such as the locations of the compressed air network, many others are more cumbersome to filter out. In order to efficiently suppress these spurious leak signals, we propose the usage of machine learning algorithms to perform classification of the signal sources. To this end, we trained a series of classifiers using real-world experiment data, and show that, by using the best performing classifiers in our leak detection algorithm, we are able to drastically reduce the amount of false-positive leaks, without any meaningful impact on the true-positive leak detection performance. We demonstrate the efficacy of the proposed approach using real-world experimental data gathered in an operating industrial setting.