A nondestructive and faster methodology to quantify mechanical properties of polypropylene (PP) pellets, obtained from an industrial plant, was developed with Raman spectroscopy. Raman spectra data were obtained from several types of samples such as homopolymer PP, random ethylene-propylene copolymer, and impact ethylene-propylene copolymer. Multivariate calibration models were developed by relating the changes in the Raman spectra to mechanical properties determined by ASTM tests (Young's traction modulus, tensile strength at yield, elongation at yield on traction, and flexural modulus at 1% secant). Several strategies were evaluated to build robust models including the use of preprocessing methods (baseline correction, vector normalization, de-trending, and standard normal variate), selecting the best subset of wavelengths to model property response and discarding irrelevant variables by applying genetic algorithm (GA). Linear multivariable models were investigated such as partial least square regression (PLS) and PLS with genetic algorithm (GA-PLS) while nonlinear models were implemented with artificial neural network (ANN) preceded by GA (GA-ANN). The best multivariate calibration models were obtained when a combination of genetic algorithms and artificial neural network were used on Raman spectral data with relative standard errors (%RSE) from 0.17 to 0.41 for training and 0.42 to 0.88% validation data sets.
Keywords: Raman spectroscopy; artificial neural networks; chemometrics; genetic algorithms; mechanical properties; partial least squares; polypropylene; principal component analysis.
© The Author(s) 2016.