Flow rate quantification of small methane leaks using laser spectroscopy and deep learning

To ensure safety in the transport and loading processes involving hazardous gasses, the detection of gas leaks in technical installations and the subsequent quantification of the released mass flows is of great public interest. One common approach to obtain this information is the usage of gas cameras, but they typically provide only inaccurate concentration data and are not sensitive enough to detect small gas leaks. In this work, we use a new gas camera approach that is based on laser absorption spectroscopy combined with mid-infrared imaging to solve these challenges. We gain videos of artificial gas leaks that containing more accurate concentration length information in ppm*m while visualizing small flow rates between 5 and 60 ml/min. For the ease of experimentation, we utilize methane while the camera can be adapted to other possibly hazardous gasses. On the dataset obtained, we train a deep-learning model aiming to continuously predict the flowrate. Additionally, we evaluate it on an eight-class classification task to compare accuracy metrics. We demonstrate an average difference of 1.6 ml/min from the mean predicted values to the true flow rates as well as an accuracy of (54 ± 2) % in the classification task.