Rapid identification between edible oil and swill-cooked dirty oil by using a semi-supervised support vector machine based on graph and near-infrared spectroscopy

Abstract

It is a challenge task to identify the swill-cooked dirty oils from various kinds of edible oils by using near infrared (NIR) spectroscopy. Due to the diversity and deficiency of standard swill-cooked dirty oils samples, the classification model involves complex liner and nonlinear relationships between class label and spectral distribution. Moreover, the small sample problems in the calibration set leads to failure of traditional supervised method such as support vector machine (SVM). A powerful semi-supervised learning method, the semi-supervised support vector machine (GS3VM), is used for classification between swill-cooked dirty oil and edible oil. The GS3VM bases on manifold assumption and approximates the distribution of spectra from both labeled and unlabeled oil samples. Comparing with the PLSDA and SVM, the experimental results show that incorporating unlabeled samples in training process improves the prediction results when insufficient training information is available. Furthermore, excessive numbers of labeled or unlabeled oil samples are helpless for classification performance of GS3VM, which solves the small sample problem and saves the cost of swill-cooked dirty oil samples. Experiment results have established that it is possible to identify the swill-cooked dirty oil from various kinds of edible oils by using the proposed GS3VM approach and NIR data. We hope that the idea of semi-supervise learning obtained in this study will help further investigations in NIR spectra analysis.