Research on Deep learning based Ultrasonic Image Learning to Develop a Gas Leak Detection Model

If a gas leak occurs in an industrial area, identifying the location of the gas leak and predicting the scale of the accident are challenging owing to the invisible nature of the gas. In this study, we developed a deep learning-based gas leak detection model that can obtain not only the gas leak status, but also the gas leak location and flow rate information, by using technology to visualize the ultrasonic waves generated during gas leaks. Research methods are broadly categorized into data collection and model learning methods. First, data was collected using an ultrasonic camera to capture ultrasonic images at different measurement distances (1 and 3 m) and gas leak flow rates (0-8 L/min). YOLO (You Only Look Once) was used for image learning, and the model was trained after setting the class according to the gas-leak flow range. The clarity of the collected ultrasonic images decreased as the measurement distance increased. In addition, there was little difference between the images for each leakage flow rate, posing challenges in distinguishing them with the naked eye. However, the model learning results showed high accuracy, with a precision of 0.960, recall of 0.967, and mAP (IoU (Intersection over Union) 50%) of 0.987. Applying this model as a gas safety management technology at industrial sites, enables the accurate determination of gas leak status, gas leak location, and gas leakage flow. This information is expected to guide appropriate accident responses for workers.