Hydrogen leakage diagnosis and visualization of density and pressure using Background Oriented Schlieren with Dual-Channel MGR-Net

Abstract

Accurate detection of hydrogen leakage is critical for safety in high-pressure hydrogen systems used in fuel cell vehicles, where even minor leaks can lead to severe deflagration risks. Conventional detection methods, relying on fixed-location sensors and manual inspection, are labor-intensive, slow, and unsuitable for large-scale, real-time monitoring. Background-Oriented Schlieren (BOS) imaging offers a non-invasive visualization approach, but its diagnostic accuracy is limited by conventional displacement extraction algorithms such as cross-correlation and optical flow. This study introduces a novel dual-channel Multi-Granularity Residual Network (MGR-Net) with transfer learning to overcome these limitations. The architecture employs a dual-channel input structure and specialized MGR Block and MGR Module for enhanced feature fusion, significantly improving diagnostic precision. Experimental results show a Dice coefficient of 81.88 %, exceeding ResUNet++ (77.34 %) by 4.54 percentage points, while reducing model size by 78 % and maintaining real-time processing at 42 frames per second, demonstrating both superior accuracy and efficiency for BOS-based hydrogen leak detection.