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

Noise interference stands as a critical factor that restricts the performance and detection accuracy of gas sensors relying on tunable diode laser absorption spectroscopy (TDLAS) technology. To address this issue, a novel neural network-based spectral optimization model is proposed and applied to near-infrared methane (CH) spectral measurement. The model combines a neural network filter (NNF) with convolutional and bidirectional long and short-term memory coupling and a back-propagation neural network concentration predictor (NCP) improved by an adaptive enhancement algorithm. The experiments utilize database parameters to construct spectral datasets for model training and conduct test experiments with standard gases for model optimization. The experimental results demonstrate that, compared with the traditional filtering algorithm, the NNF proposed in this paper enhances the signal-to-noise ratio improvement effect of spectral signals by 2.58 times. Additionally, the average absolute error and average relative error of CH concentration predicted based on the NCP are 1.29 ppm and 2.05 %, respectively. The results of Allan variance analysis indicate that when the optimal integration time is set to 406 s, the detection limit of CH can reach 34.83 ppb. The TDLAS spectral optimization model proposed in this paper offers significant references for the optimization algorithms of high-precision trace gas detection.