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

Area of Science:

• Spectroscopy
• Artificial Intelligence
• Chemical Sensing

Background:

• Photoacoustic spectroscopy (PAS) is crucial for gas concentration measurement.
• Low gas concentrations present challenges due to signal noise interference in 2f signals.
• Conventional signal processing methods often fail to maintain signal fidelity in noisy conditions.

Purpose of the Study:

• To develop a novel deep learning-based signal denoising model for enhanced gas concentration measurement.
• To address the limitations of conventional methods in handling noise at low gas concentrations.
• To improve the signal-to-noise ratio (SNR) and accuracy of 2f signals in PAS.

Main Methods:

• Utilized a differential resonant photoacoustic cell.
• Developed a deep learning model combining 1D Convolutional Neural Networks (1D CNNs) and Transformer networks.
• Trained the model using synthetic signals with simulated noise for robustness.
• Applied the model to experimental 2f signals from acetylene measurements.

Main Results:

• Demonstrated excellent noise suppression capabilities.
• Achieved an approximate 70-fold enhancement in the signal-to-noise ratio (SNR) for 500 ppb acetylene signals.
• Showcased improved determination coefficient (R²), indicating better accuracy and linearity.
• Successfully reconstructed signals with enhanced fidelity.

Conclusions:

• The deep learning model significantly improves detection sensitivity and reliability in trace gas measurements.
• This approach represents a substantial advancement in spectroscopic signal processing for gas detection.
• The method offers a robust solution for overcoming noise limitations in photoacoustic spectroscopy.