Multivariate methods are applied to the calibration of temperature in the range from 299 to 423 K for the green fluorescence spectra of erbium in lead–fluoride glass doped with 0.5 mol % of erbium and 10 mol % of ytterbium. It is shown that the regression to latent structures using the combination of moving spectral windows is characterized, among the considered methods, by the lowest value (0.2 K) of the root-mean-squared error of prediction of temperature over the test set. Artificial neural networks using two principal components as input variables, the broadband regression to latent structures, the artificial neural network using all the spectral data samples as input variables, and regression to the principal components are inferior in accuracy of the temperature calibration.

To select erbium and ytterbium doped germanate glasses and glass ceramics, which are most suitable as sensitive elements of fluorescent temperature sensors, a multivariate model of temperature calibration has been developed based on principal component analysis, cluster analysis and interval projection to latent structures of up-conversion green fluorescence spectra. The calibration model used 95 spectral variables for the GeO2-Na2O-Yb2O3-MgO-La2O3-Er2O3 glass-ceramic is characterized by the best quality parameters: the root-mean-square error is 0.37 K, the residual prediction deviation for the test subset is greater than 102, and the relative error does not exceed 0.20%.