Enhancing photoacoustic trace gas detection via a CNN-transformer denoising framework

We present a novel approach for gas concentration measurement using a differential resonant photoacoustic cell combined with a deep learning-based signal denoising model. This method addresses the persistent challenge of noise interference in 2 f signals at low gas concentrations, where conventional processing methods struggle to maintain signal fidelity. To resolve this, we propose a deep learning model that integrates 1D Convolutional Neural Networks (1D CNNs) for local feature extraction and Transformer networks for capturing global dependencies. The model was trained using synthetic signals with added noise to simulate real-world conditions, ensuring robustness and adaptability. Applied to experimental 2 f signals, the model demonstrated excellent noise suppression capabilities, enhancing the signal-to-noise ratio (SNR) of 500 ppb acetylene signals by a factor of approximately 70. Furthermore, the determination coefficient (R²) improved, reflecting better accuracy and linearity in signal reconstruction. These results underscore the model's potential for improving detection sensitivity and reliability in trace gas measurements, marking a significant advancement in spectroscopic signal processing for gas detection.