Trace gas sensor based on photoacoustic spectroscopy and deep learning nested U-shaped network (U-Net++)

This study presents what we believe to be a novel trace gas sensor integrating a rollar-type resonator photoacoustic spectroscopy (PAS) system with a nested U-shaped deep learning network (U-Net++) to achieve high sensitivity, noise suppression, and long-term stability. The rollar-type acoustic resonator, optimized via finite element simulations, amplifies photoacoustic signals by 21% compared to conventional cylindrical designs, with a resonance frequency of 1040 Hz and a quality factor of 80. The U-Net++ architecture, featuring nested skip pathways and deep supervision, dynamically processes spectral data to suppress noise, achieving a three-orders-of-magnitude reduction in standard deviation (12.91 ppm to 0.0138 ppm) for methane (CH₄) measurements. Allan variance analysis demonstrates enhanced stability, with a minimum detectable concentration (MDC) improving from 1.41 ppm to 0.173 ppb, corresponding to an enhancement factor of over 8100. Experimental validation across varying CH₄ concentrations (5-30 ppm) confirms significant signal-to-noise ratio (SNR) improvements (e.g., 5.5 to 293.3 at 5 ppm) and reduced variability, highlighting the sensor's precision and robustness. This work advances PAS technology by synergizing an optimized resonator design with deep learning, enabling high-precision gas detection for environmental monitoring, industrial safety, and medical diagnostics.