Abstract: In order to detect the content of polar components in used frying oils quickly and non-destructively, we acquired Raman spectra of used frying oil samples collected at different frying times. To create a predictive model for determining the content of polar components in used frying oils with high stability, small errors and high precision, the influence of different spectral preprocessing methods on the performance of predictive models was evaluated, and partial least squares regression analysis regression (PLSR) modelling was carried out to select the optimal spectral preprocessing method. The experimental results showed that the optimal PLSR model was obtained using standard normal transformation with root mean square error of prediction (RMSEP) of 1.18 and correlation coefficient of 0.940 4. Subsequently, the spectral data preprocessed by the standard normal transformation were used to establish a neural network model by error back propagation (BP) and radial basis function training (RBF), separately. The BP neural network model had the best prediction performance with RMSEP of 0.032 6 and correlation coefficient 0.972. The above findings demonstrate that this method can be used for the rapid analysis of polar components of used frying oils.

Key words: polar components of used frying oil; Raman spectroscopy; pretreatment; partial least squares; error back propagation; radial basis function