The feasibility study of non-destructive detection of cashmere by near-infrared spectroscopy and data driven-based class-modeling

Abstract Cashmere is designated as specialty or luxury fiber with its excellent quality and high economic value. It is common to adulterate cashmere with other animal/synthetic fibers for high profit. Currently, identification of cashmere relies mainly on microscopic examination, which is subject to the experience and skill of the operator. There is a need to develop objective alternative methods. The present work aims to explore the feasibility of combining near-infrared (NIR) spectroscopy with class-modeling and feature selection for cashmere identification. A total of 463 cashmere and non-cashmere samples were collected and prepared, which were divided into three subsets, i.e., the training set, the validation set and the test set. Principal component analysis (PCA) was utilized for exploratory analysis. The Relief algorithm was used for feature selection and only 191 informative features were selected from 1557 original ones for class-modeling. Based on these features, the special class-modeling approach using 8 components built a Relief-based model with satisfactory performance on the independent test set. The optimal full-spectrum model with 9 components performed slightly worst than the Relief-based model. The result confirms that NIR spectroscopy combined with Relief and class-modeling is feasible to carry out the untargeted identification of cashmere.