Quantitative Analysis of Cotton/Polyester/Wool Blended Fiber Content by Near-Infrared Spectroscopy Based on 1D-CNN
HUANG Meng-qiang1, KUANG Wen-jian2, 3*, LIU Xiang1, HE Liang4
1. School of Electronics & Information Engineering, Nanjing University of Information Science & Technology, Nanjing 210044, China
2. School of Physics and Optoelectronic Engineering, Nanjing University of Information Science & Technology, Nanjing 210044, China
3. Texpe Textile Technology Co., Ltd., Suzhou 215159, China
4. Jiangsu Sunshine Group, Jiangyin 214426, China
Abstract：Qualitative and quantitative analysis of fiber composition has always been a research hotspot in textile testing. However, conventional detection methods have problems such as long cycles, complicated processes, and being unfriendly to the environment. Therefore, proposing a fast, non-destructive, and accurate detection method for textile fiber content is particularly important. This study proposes a quantitative calibration model for textile fiber content, which can accurately predict the fiber content of cotton/polyester/wool in textiles. It solves the difficulty that traditional calibration models cannot consider both accurate and multiple fiber predictions. In this study, 645 wool/polyester, cotton/polyester, and wool/polyester/cotton blended samples were taken as the research objects, and an infrared spectrometer collected the near-infrared reflectance spectra of the samples. After the spectral data is preprocessed, the one-dimensional convolutional neural network (1D-CNN) model is used to predict multiple fiber contents simultaneously. The prediction results of three different machine learning algorithms are compared on the same training and test sample sets. The results show that the preprocessing method of linear function normalization and polynomial smoothing filter (SG smoothing, the sliding window is 9, fitting order is 7), combined with the proposed 1D-CNN model has the best effect, and Its model determines that the coefficient R-Squared can reach 0.998, the mean absolute error (MAE) of each content prediction is 0.62, and the root mean square error (RMSE) of prediction is 1.31. At the same time, 138 textile samples that did not participate in the modeling were used to verify the mode's generalization ability. The model's performance on the test set was excellent, with a coefficient of determination R-squared of 0.996, a mean absolute error (MAE) of each content prediction of 0.80, and a predicted mean square of the root error (RMSE) of 2.01. Using the model proposed in this paper, the fiber content in wool, cotton, and polyester blended textiles can be accurately predicted, providing a feasible method for rapid non-destructive testing of textiles and a new idea for the quantitative analysis of other blended fiber content.
