A hybrid framework for low-frequency multiple interference fringe suppression in TDLAS based on feature extraction and machine learning

Abstract

To address low-frequency multiple interference fringe noise caused by etalon effects in tunable diode laser absorption spectroscopy (TDLAS), this study introduces a hybrid denoising method that uses a feature informed neural network (FINN). This method innovatively integrates the Gramian angular field (GAF), principal component analysis (PCA), and uniform manifold approximation and projection (UMAP) techniques to efficiently extract both linear and nonlinear features of absorption spectra signals. Additionally, a feature construction module is introduced into a traditional backpropagation neural network (BPNN), significantly enhancing its ability to suppress low-frequency noise. Numerical simulations and experimental results show that the FINN performs excellently in complex noise environments, reducing the maximum absolute error by a factor of 3 compared with that of the BPNN and by >31 times compared with that of the least squares (LS) method. With an integration time of 242 s, the methane (CH₄) detection precision reached 0.331 ppm, fully validating the high accuracy and stability of the proposed approach. This method provides a novel technical approach for low-frequency noise suppression in noncontact optical measurements, with broad application prospects in gas detection, environmental monitoring, and related fields.