Deep learning‑enhanced ultrasonic beamforming for gas leak detection and localization in noisy industrial environments.

Reliable gas leak detection is vital for both industrial safety and environmental protection. Ultrasonic beamforming offers a non-contact solution by capturing broadband emissions from high-pressure gas flow, yet conventional delay-and-sum (DAS) beamforming suffers owing to interference from ambient sources and reverberant conditions. This study introduces a deep-learning-assisted enhancement of beamforming maps using a convolutional neural network. Rather than proposing a new beamforming algorithm, the framework operates as a lightweight post-processing stage on the beamforming maps produced by commercial ultrasonic cameras, without requiring access to raw multichannel signals. The network is trained on hybrid datasets combining industrial ambient recordings with simulated acoustic fields, suppresses background artifacts while highlighting true leak signatures. Experimental validation across three increasingly complex environments achieved 84% successful leak localization in the most challenging scenario, with a mean angular error below 2° and suppression of non-leak components exceeding 40 dB. Complementary simulations further demonstrate superior precision over DAS, MVDR, and DAMAS beamformers under low signal-to-noise conditions. With only 1.47 million parameters and 0.20 GFLOPs per inference, the model supports real-time deployment on edge hardware, offering a practical solution for reliable ultrasonic gas leak monitoring.