Development of a deep learning-based hazardous gas source localization and distribution mapping system using gas sensor grids and CFD simulation

Abstract

Gas leakage poses a significant threat to public safety, and accurate gas source localization is crucial for effective hazard management. Traditionally, gas leakage monitoring systems use bulky and expensive equipment to accurately identify the characteristics of target gas leakages. However, conventional gas sensing systems encounter technical constraints associated with spatial coverage, precise leakage characterization, and source localization. To confront these limitations, recent research in the field of mobile robotic olfaction and sound source localization have focused on developing gas source localization techniques. Nonetheless, issues with reliability, versatility, and maintenance costs persist as significant barriers to their effective integration. This study introduces a novel deep convolutional neural network (CNN) model-based gas source localization and concentration prediction solution using fixed gas sensor grids and computational fluid dynamics (CFD) simulation. The proposed system employs various photoionization detector (PID) grid configurations in a $400\times 500\times 210 mm^3$ chamber apparatus. The sensors are integrated to an Arduino, which simultaneously operates the individual sensors by powering their ultraviolet (UV) lamps and retrieves sensor data signals. To optimize the number of sensors and their configuration to form the grid, the performance metrics for leakage point classification was evaluated for the single-point leakage scenario. The sensor configuration with the highest accuracy involved installing four PIDs at each corner of the chamber. (Leakage Accuracy: 95.6%, Leakage Location Classification Accuracy: 96.5%, Leakage Location MSE: 0.14) Additionally, this study reports experimental results identifying two-point leakages and single-point leakage conditions with a simple barrier within the chamber. The multi-task deep learning-based gas source localization and monitoring system proposed in this study demonstrates the potential to enhance the safety of industrial sites. By addressing the limitations of current methods and advancing the integration of CFD simulations and deep learning, this study contributes to the development of more effective gas source localization techniques. The proposed framework holds promise for improving the safety and reliability of gas leak detection systems in various real-world applications.