Classification of washing powder brands using near-infrared spectroscopy combined with chemometric calibrations

In this study, near-infrared (NIR) spectroscopy is applied for rapid and objective classification of 5 different brands of washing powder. Chemometric calibrations including partial least square discriminant analysis (PLS-DA), back propagation neural network (BP-NN) and least square support vector machine (LS-SVM) are investigated and compared to achieve an optimal result. Firstly, principal component analysis (PCA) is conducted to visualize the difference among washing powder samples of different brands and principal components (PCs) are extracted as inputs of BP-NN and LS-SVM models. The number of PCs and parameters of such models are optimized via cross validation. In experimental studies, a total of 225 spectra of washing powder samples (45 samples for each brand) were used to build models and 75 spectra of washing powder samples (15 samples for each brand) were used as the validation set to evaluate the performance of developed models. As for the comparison of the three investigated models, both BP-NN model and LS-SVM model successfully classified all samples in validation set according to their brands. However, the PLS-DA model failed to achieve 100% of classification accuracy. The results obtained in this investigation demonstrate that NIR spectroscopy combined with chemometric calibrations including BP-NN and LS-SVM can be successfully utilized to classify the brands of washing powder.