Self-Supervised Deep Learning for Denoising in Light-Induced Thermoelastic Spectroscopy Using 2D Homologous 2f Spectral Matrices

This work proposes a masked self-supervised denoising framework for low-SNR second-harmonic (2f) signals based on 2D Homologous 2f Spectral Matrices (HSMs). The core idea is to reorganize the original 1D 2f signals into 2D HSMs, thereby reformulating conventional waveform denoising as a self-supervised structured reconstruction task. In this way, weak signals can be recovered from the intrinsic structural correlations of the data itself rather than from externally constructed pure labels. Experimental results on an acetylene (C₂H₂) light-induced thermoelastic spectroscopy (LITES) sensor show that the proposed framework significantly improves signal quality under low-SNR conditions. The SNR is enhanced by factors of 2.33, 6.27, 17.68, 52.69, and 62.81 at concentrations of 30, 20, 10, 5, and 1 ppm, respectively. After denoising, the concentration-response linearity reaches R² = 0.9999, while the average relative standard deviation (RSD_Mean) of peak amplitude and the relative percentage error (RPE) of 30 ppm peak amplitude are 1.44% and 0.37%, respectively. Corresponding generalization validations were also conducted to verify the robustness of the proposed framework. These results demonstrate that the proposed framework provides an effective self-supervised 1D-to-2D reconstruction strategy for low-SNR signal denoising and offers a practical and extensible route for weak-signal reconstruction under label-limited conditions in the spectroscopy sensing field.