End-to-end methane gas detection algorithm based on transformer and multi-layer perceptron.

Area of Science:

• Spectroscopy
• Chemical Sensing
• Machine Learning

Background:

• Tunable diode laser absorption spectroscopy (TDLAS) is crucial for gas sensing.
• Processing noisy spectral data from TDLAS presents challenges for accurate gas concentration determination.
• Existing digital filters may not be optimal for complex spectral analysis.

Purpose of the Study:

• To develop an end-to-end algorithm for methane (CH4) gas detection using TDLAS.
• To enhance the accuracy and efficiency of spectral data processing in gas sensors.
• To introduce a novel deep learning approach for denoising and concentration prediction.

Main Methods:

• An end-to-end algorithm combining a Transformer-based U-shaped Neural Network (TUNN) for filtering and a multi-layer perceptron (MLP) for concentration prediction was designed.
• The algorithm processes noisy transmission spectra directly to derive CH4 concentrations from denoised spectra.
• No intermediate spectral processing steps were employed, ensuring an integrated approach.

Main Results:

• The TUNN filtering algorithm demonstrated superior performance compared to traditional digital filters.
• The concentration prediction network achieved a high determination coefficient (R²) of 99.7%.
• Significant accuracy was maintained even at low methane concentrations, with R² reaching up to 89%.

Conclusions:

• The proposed end-to-end algorithm offers a more efficient, convenient, and accurate method for spectral data processing in TDLAS-based gas sensors.
• The integration of TUNN and MLP provides a robust solution for methane detection.
• This approach advances the capabilities of TDLAS technology for environmental and industrial monitoring.