Adaptively Optimized Gas Analysis Model with Deep Learning for Near-Infrared Methane Sensors

Area of Science:

• Spectroscopy
• Machine Learning
• Environmental Sensing

Background:

• Traditional direct absorption spectroscopy (DAS) faces limitations in accuracy and stability due to noise.
• Accurate methane concentration retrieval is crucial for environmental monitoring and industrial safety.

Purpose of the Study:

• To develop an advanced gas analysis model for improved methane concentration retrieval from noisy spectral data.
• To enhance the accuracy, stability, and sensitivity of methane sensors using deep learning.

Main Methods:

• Developed an adaptively optimized gas analysis model (AOGAM) comprising a neural sequence filter (NSF) and a neural concentration retriever (NCR).
• Utilized deep learning algorithms, including an encoder-decoder structure with attention for NSF and PCA with a fully connected layer for NCR.
• Trained the model on both computationally generated and experimental datasets for methane transmission spectra in the near-infrared region.

Main Results:

• The NSF significantly improved signal-to-noise ratio by 7.3 dB compared to traditional and state-of-the-art filters.
• The AOGAM achieved higher accuracy in methane concentration determination than the traditional DAS method.
• The optimized sensor demonstrated precision, stability, and a minimum detectable column density of 1.40 ppm·m (1σ).

Conclusions:

• Deep learning combined with absorption spectroscopy offers a more effective, accurate, and stable solution for gas monitoring systems.
• The AOGAM provides a robust approach for extracting methane absorption information from noisy spectra.
• This technology holds significant promise for real-time methane sensing applications.