| In this paper, an in-depth study of application of NIRS in textile content was carried out by chemometrics methods on the following topics: the selection of system parameters, the elimination of abnormal samples and the building of regression model.The pure wool, cotton, silk and terylene was collected as the main materials. According to wool content, the 110 different mixed samples were confected, and the wool content was detection index. The NIR spectrum of samples was measured by WQF-400N fourier transform near infrared spectrophotometer. The instrument parameter and accessory were selected through comparison of experiments, and scan times, resolution, measuring times were established. The method of partial least squares regression was adopted for different pretreatment spectrums through derivative, smoothing and multiplication scatter correction. Predictive precision of calibration model was appraised by three statistics which were mean absolute error, maximum absolute error and root mean square error of prediction. The best pretreatment method is the first derivative, 5 point S-G polynomial smoothing and multiplication scatter correction.The abnormal spectrums were selected by mahalanobis distance and leverage algorithm, and partial least squares was applied to regression analysis. The forecast effect was measured with mean absolute error and root mean square error of prediction, and result indicates the method of mahalanobis distance was better. After four abnormal spectrums were eliminated, the spectrum of 106 samples was ensured as calibration set and prediction set. The calibration and predicition samples were differentiated with cluster analysis. The numbers of Calibration samples and predicition samples were 74 and 32 respectively. AKS algorithm was used in calibration samples for further selection, and the spectrum of 65 samples was selected as the best sample set, and the spectrum of the rest of 9 samples was eliminated. The two calibration samples sets were regressed by partial least squares. Their mean absolute error and root mean square error of prediction were reduced from 2．024 and 2．604 to 1．932 and 2．371, which indicated the best samples set had more representation. The rest of 97 samples were regressed by least square support vector machine. The value of Gamma andγwas calculated by three-step search method, its result was [0．00293, 12．785] . Their mean absolute error and root mean square error of regression result were reduced from 2．3230 and 2．7943 to 1．7144 and 2．1317, and three-step search method may save calculation of parameter selection.The upwards analysis indicated that when detecting wool through NIR, the first derivative spectrum is better than the original and the second derivative spectrum, 5 point S-G polynomial smoothing is better than movable and index spectrum, the method of mahalanobis distance is better than leverage algorithm. The selection of AKS algorithm about calibration samples may make calibration samples set have more representation. The forecast samples set was regressed by least square support vector machine. The three-step search method may not only save calculation, but also realize the rational parameter selection. The above contents provide some methods for references for examining quickly wool in the textile. But samples may be measured after being crushed up, which brings some difficulties to site operation. Therefore, a series of problems need to be further researched and soluted.
|
PDF Full Text Request