Detection of graft copolymerization weight gain in silk based on hyperspectral imaging technology
LI Henan, WANG Zhenhua, LIU Weihong, HE Haonan, ZHU Chengyan, TIAN Wei
2025, 33(09):
71-78.
DOI: 10.12477/j.att.202411051
Asbtract
(
)
PDF (9587KB)
(
)
References |
Related Articles |
Metrics
Silk has poor wrinkle resistance, so it undergoes modification treatments to increase its drape and initial modulus. This treatment can lead to an increase in the weight of the silk, but an excessive weight gain rate beyond a certain threshold can affect the quality of silk fabrics. Therefore, the detection of the weight gain rate of silk graft copolymerization is of great significance. Hyperspectral imaging technology, originally applied in the field of remote sensing, has now also been successfully used for qualitative and quantitative detection in textile materials.
To investigate the feasibility of using hyperspectral imaging technology to detect the weight gain rate of silk graft copolymerization, this paper produced a total of 109 HEMA-weighted silk samples. The hyperspectral images of the silk samples were collected using a hyperspectral imaging system and corrected with black and white calibration. Subsequently, multiple uniformly distributed and opaque regions on the samples were randomly selected to extract 545 spectral data points within the 1,000–2,400 nm wavelength range, which served as the sample set for subsequent analysis. The sample set data were preprocessed using FD, MSC, and SNV, and then quantitative detection models for silk weight gain rate were established by combining PLS and BP neural networks, respectively. The results showed that MSC and SNV eliminated spectral differences caused by sample inhomogeneity and varying scattering levels while preserving characteristic peaks in the spectral curves. First-order derivative (FD) preprocessing improved the resolution of the spectral curves, increased the number of peaks, and enhanced the separation of overlapping peaks. Using PLS as the classifier, FD-PLS, MSC-PLS, and SNV-PLS prediction models for silk weight gain rate were established, yielding root mean square errors (RMSE) of 0.08,510, 0.03,554, and 0.32,795, respectively. The average RMSE value for the three models was 0.14,953, and their correlation coefficients were 0.67,861, 0.97,098, and 0.9,4003, respectively. This indicates that the models built with PLS as the classifier have a certain degree of fitting effect and generalizability for predicting silk weight gain rate. Using BP neural networks as the classifier, FD-BP, MSC-BP, and SNV-BP prediction models for silk weight gain rate were established, yielding RMSE values of 0.04,943, 0.01,273, and 0.01,200, respectively. The average RMSE value of the three models was 0.02,472, and their correlation coefficients were 0.95,106, 0.99,684, and 0.99,683, respectively. This indicates that the models built with BP neural networks as the classifier generally outperform those built with PLS. Among them, the SNV-BP model achieved the highest prediction accuracy, with an error rate of 1.097%, a test set error of only 1.200%, and a correlation coefficient of 0.99,683. This demonstrates that the BP neural network, as a nonlinear model, has good generalization capability for predicting the weight gain rate of HEMA-grafted silk without overfitting.
The SNV-BP model established in this study, which utilizes spectral data from the 1,000–2,400 nm wavelength range as the raw data, employs SNV as the preprocessing algorithm, and adopts BP neural networks as the classifier, demonstrates exceptionally high accuracy. This confirms the feasibility and accuracy of the method for determining the weight gain rate of HEMA-grafted silk based on hyperspectral imaging technology. Furthermore, it provides a novel approach and foundation for the non-destructive and continuous detection of silk weight gain rates.
