Neural network optimization algorithms for high-precision TDLAS gas spectroscopic detection.

Area of Science:

• Spectroscopy
• Chemical Sensing
• Artificial Intelligence

Background:

• Noise interference critically limits the performance and accuracy of gas sensors using tunable diode laser absorption spectroscopy (TDLAS).
• Accurate detection of trace gases like methane (CH4) is essential in various environmental and industrial applications.

Purpose of the Study:

• To develop and validate a novel neural network-based spectral optimization model for enhancing TDLAS measurements of near-infrared methane.
• To improve the signal-to-noise ratio and accuracy of CH4 concentration prediction in the presence of spectral interference.

Main Methods:

• A neural network filter (NNF) employing convolutional and bidirectional long and short-term memory (LSTM) was developed.
• A back-propagation neural network concentration predictor (NCP), enhanced by an adaptive algorithm, was integrated.
• Model training utilized spectral datasets constructed from database parameters, with optimization through experiments using standard gases.

Main Results:

• The proposed NNF demonstrated a 2.58-fold improvement in signal-to-noise ratio enhancement compared to traditional filtering algorithms.
• The NCP achieved an average absolute error of 1.29 ppm and an average relative error of 2.05% for CH4 concentration prediction.
• Allan variance analysis revealed a detection limit of 34.83 ppb for CH4 at an optimal integration time of 406 seconds.

Conclusions:

• The developed TDLAS spectral optimization model effectively mitigates noise interference, significantly enhancing detection performance.
• The neural network-based approach offers a robust solution for high-precision trace gas detection, particularly for methane.
• This study provides valuable insights for advancing optimization algorithms in sensitive gas sensing applications.