Enhanced TDLAS detection of high-temperature water vapor using neural network-optimized spectral denoising

Abstract

Real-time monitoring of atmospheric water vapor concentrations is critical for combustion optimisation and emission control in high-temperature industrial processes. However, traditional spectroscopic techniques frequently encounter substantial noise interference, thereby impeding the precision of measurements when operating in extreme conditions. This study proposes a novel feedforward neural network (FNN)-assisted Savitzky-Golay collaborative filtering algorithm for noise suppression in tunable diode laser absorption spectroscopy (TDLAS). The proposed method integrates the nonlinear learning capability of FNN with the adaptive smoothing advantages of S-G filtering, achieving superior denoising performance compared to conventional single-filter approaches. Experimental validation using HITEMP-simulated spectra demonstrates a remarkable signal-to-noise ratio (SNR) enhancement from 20.93 dB to 46.60 dB, representing a 25.66 dB improvement 6.25 dB superior to optimal traditional filtering methods. Field tests conducted in a 1010 °C blackbody furnace environment revealed a concentration measurement standard deviation of 17.29 ppm, with Allan deviation analysis confirming a detection sensitivity of 2.2 ppm at 99 s integration time. The system exhibits excellent linear response (R² = 0.97) across 600 continuous measurements, achieving a mean absolute error of 22.46 ppm compared to reference values. This breakthrough in spectral processing technology enables reliable water vapor monitoring under challenging high-temperature conditions, significantly improving process control capabilities in metallurgical and energy-intensive industries.