A stable identification method of wool and cashmere based on localized modeling strategy using NIR spectroscopy

For a long time, accurate identification of cashmere and wool fibers has been a challenge in the textile industry. Traditional chemical and image recognition methods are very complex, time-consuming, and costly. Currently, near-infrared spectroscopy is a new identification method with fast, accurate and non-destructive characteristics. However, there is a dilemma that fibers from different pastures exhibit intra-class differences is always bigger than inter-class differences in spectra, which leads to increased difficulty in identification due to spectral overlap. Therefore, this paper proposes a localized modeling strategy based on the clustering method to balance intra-class and inter-class differences and reduce spectral overlap. The strategy uses the Kennard-Stone (KS) algorithm to select representative samples with a concentrated feature distribution to determine the distribution range of each class, and then calculates the distance between each sample and the representative samples by using the Spectral Angle Mapper (SAM), which divides samples with similar characteristics into the same clusters according to a set distance threshold. This method reduces the spectral overlap rate by re-filtering and dividing the feature distribution of each class of samples into clusters. Experimental results demonstrate that the proposed localized modeling strategy can achieve prediction accuracy of up to 97.8 %, surpassing the 90.7 % accuracy obtained by training a global model with all samples. Therefore, the proposed strategy in this study can effectively reduces the impact of spectral overlap, and improves the recognition accuracy of cashmere and wool.