CH₄ concentration prediction in WMS-TDLAS systems using 2f/1f signal and CNN

Existing deep learning-based wavelength modulation spectroscopy-tunable diode laser absorption spectroscopy (WMS-TDLAS) concentration inversion methods generally use the second harmonic (2f) signal as the model input. However, concentration-related features extracted from the 2f signal are susceptible to non-ideal factors such as laser intensity fluctuations, system drift, and baseline variations, reducing concentration inversion accuracy. To address this issue, a CH₄ concentration inversion method based on the 2f/1f signal combined with a convolutional neural network (CNN) is proposed. The 2f/1f signal is adopted as the model input to reduce the influence of system response variations and enhance the representation of concentration-related spectral information. The CNN automatically extracts nonlinear features from the 2f/1f signal and establishes the mapping relationship with CH₄ concentration. Experimental results show that, on the randomly-split test set, within the CH₄ concentration range of 0-180 ppm, the CNN model with 2f/1f input achieves better prediction performance than that with 2f input. The root mean square error (RMSE) and mean absolute error (MAE) are 0.859 ppm and 0.646 ppm, respectively, and the coefficient of determination (R²) between the average predicted and standard concentrations is 0.99999. Allan deviation analysis indicates a limit of detection (LOD) of 82.96 ppb at an integration time of 37.12 s. The results demonstrate that adopting the 2f/1f signal as input improves the concentration-inversion accuracy within the same CNN framework. Furthermore, a Kalman filter was applied as a post-processing method to suppress random fluctuations in continuous CNN prediction results and improve measurement stability.