Study on variable-distance multispectral radiation thermometry based on SABO-GRNN model

To address the challenges of difficult emissivity determination and accuracy affected by measurement distance in traditional radiation thermometry, a multispectral radiation thermometry method was employed based on subtraction-average-based optimizer-generalized regression neural network (SABO-GRNN). A nonlinear mathematical model between target temperature and spectral radiation intensity was established. The principal component analysis (PCA) method was used to extract the principal component containing a large amount of information from the input spectral information and the distance factor which had a significant impact on temperature measurement as input variables of the network model. Then, the GRNN model was improved through fully training on sample data using the SABO algorithm with strong optimization ability and fast convergence speed, and its prediction performance was systematically evaluated. The results showed that using a high-temperature muffle furnace as measurement target at fixed distances, the proposed method achieved an average relative error of 0.51%, representing a 33.1% reduction compared with PCA-extreme learning machine (PCA-ELM) algorithm. Under varying measurement distances, the maximum relative error remained below 1.74%. This study contributes to the application of neural networks and optimization algorithms in radiation thermometry.