Flame spectroscopy is extensively used in the analysis of industrial combustion processes. A flame emits energy over a wide spectral region and its associated spectra contains both continuous and discontinuous components. In the literature there are models, based on either Planck's or Wien's law, for representing the flame spectral behavior under different combustion parameters; however, the non-linear nature of these models, the high dimension of the spectral data, and the superposition between the continuous and the discontinuous spectral emissions complicate the theoretical analysis of combustion processes. In this paper exploratory data analysis is used to derive data-driven models for combustion process monitoring. To do so, a database of measured spectra, collected from a real combustion process in the range of 400 to 800 [nm], is used as a priori information about the process. The database contains spectral information about continuous emissions of natural gas, oil, and bio-oil fuels at different combustion conditions. To summarize in a reduced number of terms the whole spectral information contained in the database, traditional as well as probabilistic Principal Component Analysis is conducted in order to create linear data-driven predictive models for combustion process monitoring. The predicting performance of the models is tested using the goodness-of-fit and the root-mean-squared error. The applicability of the models is finally tested by devising a simple automated solution for separating the discontinuous radiation from the continuous emission.