Water Content Compliance of Biodiesel Assessed by Raman Spectroscopy and Machine Learning Classifiers

A combination of Raman spectroscopy and multivariate modeling was applied to build classification models to assess water content in soybean biodiesel. The Raman signature region of water and the characteristic vibrational modes from CH 2 and CH 3 moieties were used to build the models, as they are chemically influenced by the presence of water in biodiesel. Also, through data mining of the biodiesel digital signature region, structural modifications associated with CH 2 , CH 3 , and CC groups were found to be related to the presence of water The fluorescence influence in modeling was also studied, a common Raman interference whose origin in biodiesel can be associated with species such as vitamin E, carotenoids, and chlorophylls. All methods exhibited distinct advantages: Support Vector Machine (SVM) demonstrated superior performance, achieving accuracies above 0.90 with fluorescence data, representing excellent overall performance. Concurrently, discriminant analysis‐based approaches (PLS‐DA and iPLS‐DA) attained specificities higher than 0.90, showing high efficacy in identifying biodiesel non‐compliant with Brazilian regulations (ANP ordinance no. 920/2023), which establishes 200 mg kg −1 as the maximum allowable water content limit. Overall, a consistent improvement in the figures of merit was observed when spectra with a fluorescence background were present and incorporated into the models.