Spectral calibration for atmospheric particles analysis under non-precise focusing conditions using LIBS combined with transfer learning

Abstract

Elemental analysis of atmospheric particulate matter is crucial for air-pollution research. However, the dispersed nature of these particles can lead to variations in the laser-focusing positions, which cause fluctuations in the laser-induced breakdown spectroscopy spectral data. In this study, a transfer-learning approach called transfer component analysis (TCA) is introduced to reduce the impact of spectral-data fluctuations on quantitative-analysis model predictions. First, dictionary learning combined with logistic regression is used to select the effective spectral data for which the laser successfully interacts with the particles. Second, TCA is used to migrate the data of pellet and dispersed atmospheric particles. Using partial least squares regression (PLSR), a TCA-PLSR model is established. The results indicate that the TCA-PLSR model significantly enhances the prediction performance, with test set coefficient of determination (R_P²), root mean square error (RMSE_P), and mean relative error (MRE_P) values of 0.9869, 60.33, and 0.1005, respectively. Compared with the PLSR model, R_P² improved by 55.05 %, RMSE_P decreased by 81.09 %, and MRE_P decreased by 82.68 %. This method effectively detects metal concentrations in atmospheric particles and offers a scientific basis for air-pollution monitoring.