Generative adversarial networks-enhanced quasi-square wave modulated photoacoustic spectroscopy: A highly sensitive NH3 detection method under strong background noise

We present a photoacoustic spectroscopy reconstruction generative adversarial network (PASR-GAN), which boosts the NH₃ detection performance of the photoacoustic spectroscopy sensor through collaborative optimization of the light modulation mode, especially under strong background noise. Instead of a sinusoidal wave, a quasi-square wave is used as the modulation waveform due to its higher signal excitation efficiency, achieving a 37 % signal enhancement. PASR-GAN suppresses noise and reconstructs corresponding clean signals by establishing a nonlinear mapping between noisy and clean signals, overcoming the limitations of traditional algorithms that rely on prior assumptions and are difficult to eliminate complex noises. For inherent noise and sudden noise, PASR-GAN exhibits 7.5 times and 172 times noise reduction, respectively. A detection limit of 32.44 ppb and a 0.99999 linear coefficient of determination within the 0<b>-</b>1000 ppm range demonstrate the concentration prediction capability of PASR-GAN. PASR-GAN provides a robust, data-driven approach for signal reconstruction under complex noise environments.