All of 598 cotton-polyester blend fabrics were supplied by Jinghuan Textile Recycling Handan Co., Ltd. and Zhejiang Green Environmental Protection Co., Ltd. The sample set had a wide range of blended fabric content (from 0% polyester ~100% polyester), various fabric colors and stitch structures. First of all, 598 blend fabrics were selected by FT-MIR (Nicolet Nexus 670, America). From Fig.1 we could see, 100% polyester fabrics have strong absorption in 1 711.7 cm^-1 (ν _C=O) , 1 243.4 cm^-1 (ν _as(C— O) and 1 094.9 cm^-1 ( νsC—O) spectral regions, and 100% cotton fabrics characteristic absorption in 3 332.0 cm^-1 (ν _{O— H}), 1 427.6 cm^-1 (δ _{O— H}), 1 025 cm^-1 (ν _{C— O}), fibers content was obtained by the national standard method (GB/T 2910.11— 2009)^[9], where the cotton fiber in the cotton-polyester blend fabrics was dissolved with 75% H₂SO₄, and the rest of the polyester content was acquired by gravimetric method.

Fig.1

	Figure Option ViewDownloadNew Window
	Fig.1 MIR absorption spectra for cotton, polyester and cotton-polyester fabrics

In the laboratory, temperature and humidity were 25 ℃ and 50%~70% of the conditions. NIR spectra of 252 cotton-polyester blend fabrics were measured on a portable SupNIR-1550 diffuse-reflection NIR spectrometer (Focused Photonies Inc. Hangzhou, China) with co-adding 30 scans ranging from 1 100 to 2 500 nm at 11 nm resolution. Each sample spectrum was produced by three-times-averaging, and the best thickness of tested samples was thicker than 3mm, which could prevent NIR light from passing through the sample and collecting background information. All of 598 cotton-polyester blend fabric samples had been washed to remove surface oil based stains, among which 252 samples were used in calibration set, and the other 346 swatches were used for validation set to verify the accuracy of the model. 100% cotton and 100% polyester fibers spectra were measured and compared. Fig.2 demonstrated that significant spectral differences of 100% cotton and 100% polyester were readily observed in several spectra regions, especially in the 1 400~1 700 and 1 900~2 200 nm, and these distinctive features could be used as a mark of the qualitative analyses of waste cotton-polyester blend fabrics. The NIR quantitative analysis calibration model was established by using PLS method and the model was validated by the parameters of R_C, SEC, R_P and SEP.

Fig.2

	Figure Option ViewDownloadNew Window
	Fig.2 The NIR absorbance spectra for cotton (a) and polyester (b)

The primary NIR spectra (between visible and MIR spectral regions) were generally concerned with the region of 1 100~2 500 nm^{[10, 11]} and were mainly the combinations of band and overtone band of several hydrogen-containing groups (such as C— H, N— H, S— H, O— H). In addition, strong NIR absorption bands were also observed for carbonyl and acylamino organic compounds, almost covering all organic compounds and mixtures. However, as a result of that fact that NIR spectra often had relatively low absorption intensity and sensitivity, contained broad and blunt spectral, these broad-spectrum information needed to be extracted using characteristics to obtain NIR characteristic spectrum of the fabric. Various methods for processing NIR data have been reported in the literature, including: spectral data pretreatment and wavelength selection, such as OSC, net analysis signal (NAS), wavelet transform (WT), PLS, principal component analysis (PCA) etc. For qualitative and quantitative analysis of textiles, fabric components yield distinct absorption at a specific wavelength that could be used to identify and differentiate fabric categories.

The reference value of the content of polyester-cotton blended samples was determined by acid dissolution method, in which cotton-polyester blend fabric was dissolved with 75% sulfuric acid, and the amount of cotton in the sample was calculated from the insoluble polyester. The reaction could be carried out in a water bath at 50 ℃ for about 1 hour, and the content of each sample was measured twice in parallel and the absolute error of the two measurement results did not exceed ± 3%. The average was used as a reference for the sample content established by the model.

The NIR spectra of 252 training samples were shown in Fig.3. As could be seen from the spectrum, which presented two different types, one was normal spectrums, and the other one was slash spectra (spectral features disappear completely). At the same time, a few samples showed different baselines than most of the modeling samples (Fig.3). These baseline drift area was normally in 1 100~1 955 nm. It was possible that these baseline drifts were caused by differences in fabric stitch structure, delusterant or dyeing method. Samples of different stitch structure normally generate slightly difference to NIR spectra, and the baseline drift was usually produced by darker colored fabrics, as shown in Figure 4. Fabric surface effects (e.g. decolorant, delusterant) often led to slash spectra, so did the fine particles, which were attached to the surface of the fiber. A photograph of the microscope was shown in Figure 5, which documented the microstructure of the fiber. As these microscopic photographs showed, the fiber surface of slash spectra samples was rough compared to most of the modeling samples and the fiber surface was distorted, as shown at 1, 2 in Fig.5, which could reduce the light reflection of the fiber due to the loss of their luster. It was much more probable that these features brought about the formation of slash spectra^[9]. While the fiber surface of normal NIR spectroscopy tended to be smooth with no other particles on its surface, as shown at 3, 4 in Fig.5.

Fig.3

	Figure Option ViewDownloadNew Window
	Fig.3 NIR absorption spectra for calibration set

Fig.4

	Figure Option ViewDownloadNew Window
	Fig.4 NIR absorption spectra for different colored cotton fabrics

Fig.5

	Figure Option ViewDownloadNew Window
	Fig.5 Microscope photo for 100% polyester fiber

As could be seen from Figure 3, the baseline drift could not completely affect the position of the main absorption wavelength, however, they could seriously impact the intensity of absorbance in these spectral regions^[12]. To eliminate the baseline drift, 1^st derivative was applied to the original spectra combined with S-G smoothing with second-order polynomials and 13-point windows. During the spectra preprocessing, derivative Math was often used to minimize spectral differences and to eohance small differences between similar samples. These spectra were pretreated by derivative and the results were presented in Fig.6. The results showed that compared with other samples, the extinction effect did not impact the intensity of the derivative spectrum in each spectral region, abd the broad spectra turned into sharp’ s, and the spectrum signal became very strong compared to the original NIR spectrum. In general, For NIR spectroscopy baseline drifts, the influence of fabric stitch structure on the NIR spectra was less significant than that of color, meanwhile, and slash spectra were also emerged features like normal spectrums after preprocessing.

Fig.6

	Figure Option ViewDownloadNew Window
	Fig.6 NIR 1st derivetive spectra for all calibration set

2.3.1 Establishment of quantitative calibration model

As could be seen from Fig.2, the greater spectral differences between cotton and polyester in waste cotton-polyester blend fabrics gave a good basis for establishment of NIR quantitative calibration model. The preprocessing method was an important part to NIR pretreated. Derivative, standardization and signal correction were often used to handle the NIR spectrum for all training sample, and derivative normally was utilized to minimize the spectral difference and baseline drifts, and signal correction generally was used to handle the other disadvantage factors. In a word, preprocessing methods were applied to deduct the non-objected spectra affect on the target one. According to the literature, OSC could remove orthogonal part of spectral matrix X and appropriate value matrix Y. As a newly developed processing method, however, there was an easily ignored disadvantage that spectral noise was not absolutely orthogonal to matrix Y when OSC method was used to handle the NIR data. Therefore, OSC could only eliminate the orthogonal part, but the residuals of noises also could affect robustness of the calibration model seriously, and might be a reason for over fitting. So a series of wavelength selection methods were applied to process NIR data before OSC, and unrelated points to models were removed to eliminate noises interference to some extent. In addition, characteristic wavelength filtered could reduce the cost of computation, and the robustness of calibration model would be improved, and correlation coefficient method was used to select wavelength in this paper. The correlation coefficient method was a correlation calculation for spectral matrix of calibration set, absorbance vector x_i for each wavelength to components underdetermined y_i array of concentrate vector, and obtained a figure of wavelength-correlation coefficient such as Fig.7. The r was obtained by the following formula.

r=n∑i=1nxiyi-∑i=1nx1∑i=1nyin∑i=1nxi2-(∑i=1nxi)2n∑i=1nyi2-(∑i=1nyi)2

Fig.7

	Figure Option ViewDownloadNew Window
	Fig.7 Wavelength-correlation coefficient

The NIR quantitative calibration model was established by fiber content of 252 calibration samples (containing samples with a wide polyester content range and diverse colors), combined with chemometrics.

The spectra were pre-processed by 1st derivative, S-G smoothing, mean centering and other signal correction method in the region of 1 400~2 500 nm, and the results were shown in Table 1. The statistical data indicated that the model established by OSC combined with 1st derivative and mean centre preprocessed showed the best result (prediction accuracy was 97.6%) for the quantitative analysis of cotton or polyester content in waste cotton-polyester blend fabrics.

Table 1

Table 1

Table 1 Calibrated results and predicted results of polyester-content by different pretreated methods Model type Pretreated Method Calibration set Prediction set Prediction Accuracy/% RC RMSECV/% RP SEP/% 1st derivative+S-G smoothing+SNV 0.995 0.053 0.993 2.002 88.0 PLS 1st derivative+S-G smoothing+Mean centre 0.996 0.041 0.995 1.86 93.2 1st derivative +S-G smoothing+Mean centre+OSC 0.998 0.002 0.997 1.121 97.6

Table 1 Calibrated results and predicted results of polyester-content by different pretreated methods

The correlation coefficient plot for LAB and NIR polyester content in best model as described in Fig.8, which had a strong correlation (R² was 0.995 8). From the figure we could see some abnormal samples, which were called “ outlier samples” . There are two kinds of outliers. One was the error of reference value measured or spectrum scanned, which could seriously influence the accuracy of the model, so it must be removed in the optimized model. The solution to this problem was that each sample was re-measured or scanned, and the results were compared with the previous one. The other one was that there was significant difference between outliers and the major modeling samples, and the accuracy of the model would be improved if these samples were removed, but adaptability of model would be reduced simultaneously because of the particularity of their properties. The common method used to solve this problem was establishing a new model, but this was not a practical solution since there was substantial cost and time associated with the development of a robust calibration model. At the same time, It had been proved that the part of singular samples also caused baseline drifts.

Fig.8

	Figure Option ViewDownloadNew Window
	Fig.8 Correlation coefficient plot for polyester content

2.3.2 Validation of quantitative model

In order to make the model applied to the actual analysis work, the accuracy and precision of the best model has been tested. Matching samples t-test applied to the LAB polyester content and NIR values of 346 calibration samples to verify the accuracy of the best model, the results were shown in Table 2. The absolute error of the LAB value and NIR value was all within (-3%, 3%), and the matching samples t-test showed that there were no significant difference between LAB and NIR value with 95% confidence bounds and 5% significance level (t_polyester=0.144< t_{0.05, 345}=1.967). At the same time, changing measurement environment (e.g. temperature, humidity), 6 times spectrum scanning for each sample, and the predicted value of polyester was obtained by quantitative calibration model. The result was shown in Table 3, which presented better repeatability.

Table 2

Table 2

Table 2 T-test of LAB and NIR values Mean value Standard deviation t Sig. LAB-NIR 0.019 2.177 0.144 0.885

Table 2 T-test of LAB and NIR values

Table 3

Table 3

Table 3 Repeatability test of the model Max retries LAB value NIR value Standard deviation RSD /% 6 54.94 53.04~56.79 1.379 0.025 25.45 24.96~26.68 0.685 0.027 …… 71.85 69.02~72.45 1.215 0.017

Table 3 Repeatability test of the model