Abstract:

When using near-infrared spectroscopy and machine learning methods to quickly identify the quality of Pu-er tea， the spectra collected by medium and low-end near-infrared spectroscopy acquisition equipment have the characteristics of high dimension， overlap and large noise， which seriously affects the accuracy of modeling. This paper proposes a noise-robust feature extraction method， which is combined with support vector machine （SVM） classifier to establish the quality identification method of Pu-er tea. Firstly， the noise-robust feature extraction method， principal component analysis （PCA） and successive projections algorithm （SPA） are used to extract the features from the obtained near-infrared spectral data. Then， SVM is used to train the data after feature extraction to obtain the identification model. The comparison of the identification results of the model shows that for the noiseresidual near-infrared spectral data， the noise robust feature extraction method in this paper can effectively resist the influence of noise and propose feature variables from the high-dimensional spectrum to improve the accuracy of the identification model. The accuracy， recall， specificity， accuracy and F-score predicted by the identification model were significantly higher than those obtained by the other two methods. For the detection of ancient Pu-er tea and non-ancient Pu-er tea， the accuracy and recall predicted by the identification model in this paper have reached 92.06% and 95.38% respectively， indicating that the identification model has good identification ability. The research results provide theoretical reference and basis for accurately judging the quality of Pu-er tea in practical application.

Key words: near-infrared spectroscopy, noise, rapid identification, Pu-er tea, feature extraction, machine learning