Photoacoustic carbon monoxide gas sensor using deep neural networks for nonlinear noise suppression

In photoacoustic spectroscopy (PAS) based gas detection, measurement accuracy and sensitivity are frequently constrained by noise interference, particularly under low signal-to-noise ratio (SNR) conditions. Conventional filtering methods are generally unable to effectively suppress noise while preserving signal integrity and retaining characteristic features, thereby restricting further improvements in detection performance. To address this limitation, a novel signal-denoising framework, termed U-LSTM, is proposed by integrating the symmetric encoder-decoder architecture of a U-shaped neural network (U-Net) with Long Short-Term Memory (LSTM) modules. This hybrid architecture enables the simultaneous extraction of local signal features and the modeling of global temporal dependencies, thereby attaining superior suppression of complex, nonlinear noise. The results demonstrate that the proposed U-LSTM model substantially outperforms conventional filtering methods. For example, when processing the second-harmonic (2 f ) signal of carbon monoxide (CO), the SNR increased from 15.4 for the original signal to 390.4 after denoising, demonstrating the excellent noise-suppression and signal-reconstruction capabilities. The proposed method provides a robust and effective denoising solution for PAS-based trace-gas detection and significantly improves detection accuracy and system reliability in noisy environments.