A novel combustion diagnostics system based on flame radiation measurement and spectral emissions analysis is presented. The key aspect of this study is the proper estimation of the continuous component or baseline of spectral signals containing spurious discontinuities. We compare both, a PCA based method against an alternative method based on an Artificial Neural Networks model. The accuracy and computational cost for baseline estimation in practical industrial applications is analyzed, in order to incorpore the radiation measurement as a variable suitable for control decisions. Several experiments in industrial facilities demonstrate the suitability of radiation measurement via flame spectrum analysis, in particular the relation between the proposed optical variable and the process operational conditions in the case of an oil boiler for steam production and a ladle furnace preheating process. As results, the methodology based on principal components analysis ( PCA) provides high accuracy for baseline estimation in flame spectra, with metrics of RMSE approximate to 0: 1 . 10(-4) and GFC > 0: 99 with fast computation times, smaller than typical flame flickering. (C) 2016 Elsevier Ltd. All rights reserved.