Swatches of known commercially-available textiles were obtained from the Baylor University Department of Family and Consumer Science. Specimens, approximately one-inch square, were cut from these swatches and placed in acid-free polypropylene sleeves that were labeled with fiber type and a number indicating position in the data matrix. The identity of the textile was provided by the original supplier. NIR spectral analysis was conducted without any other form of sample pretreatment.

Because the textiles used in this study were commercially available and not pretreated prior to analysis, they could have a variety of fiber finishes and dyes. Therefore, the sample population was made as broad as possible so that models generated by SIMCA could account for as many variations within a fiber type as possible. A total of 758 textile samples was cataloged and their diffuse NIR reflectance spectra collected for training and testing. Table 1 shows the textile library used for this study with the number of samples in each fiber category.

NIR analysis was carried out using a dual-channel spectrometer assembled and tested in our laboratory [19] . To acquire diffuse NIR reflectance spectra, the sample beam was directed into an eight-inch diameter barium sulfate (BaSO₄) coated integrating sphere (Model 70451, Oriel) that had a lead sulfide (PbS) detector (Model 7202, Oriel) located in the detector port and mounted 45˚ from the sample window. The integrating sphere was equipped with a baffle and a light trap to reduce specular reflection.

Textile samples were placed in the sample holder at the sample window of the integrating sphere. Because the sensitivity of PbS detectors drops off at wavelength limits of approximately 1000 and 2500 nm, the wavelength

region selected for these studies ranged from 1334 to 1906 nm. Wavelengths outside this region produced signals with significantly reduced signal-to-noise ratios.

Diffuse NIR reflectance scans were recorded with a sampling interval of 2.0 nm, and measurements were reported as, where R is equal to the diffuse reflectance signal. Data were written into ASCII tables by the LabView™ software (version 5.1, National Instruments, Austin, TX) that controls the instrument and converted to Microsoft^® Excel spreadsheets as required for chemometric analysis. The spectral data, contained in the Excel spreadsheets, were imported into the Unscrambler^© software package (versions 9.1, 9.6, and 9.7, Camo, Inc., Woodbridge, NJ). The data were then transposed to list the wavelengths as the x-variables and the reflectance as the y-variables.

Except where stated otherwise, raw spectral data was subjected to Savitzky-Golay smoothing [20] , using a zero-degree polynomial and 11 averaging points, 5 on each side of a given wavelength. The multivariate principal component analysis for each textile category was performed on mean-centered data at the 95% confidence level (p = 0.05 significance level). A mean-centered plot, calculated by the Unscrambler^© statistical package, was obtained by first averaging the individual spectra on a wavelength-by-wavelength basis (i.e., adding the values of each spectrum on a wavelength-by-wavelength basis and dividing each sum by the number of samples). This average spectrum was then subtracted from each individual spectrum on a wavelength- by-wavelength basis to give mean-centered spectra.

Individual principal component analyses were conducted for each fabric type using the smoothed reflectance spectra collected for each group.

As an example, Figure 2 shows the results of the principal component analysis obtained using the spectra in Figure 1 for cotton. The three-dimensional scores plot in Figure 2(a) shows that the samples are segregated into a relatively distinct group with approximately 100% of the relationship being explained by the first and second principal components. Figure 2(b) is a plot of the residual x-variance as a function of leverage. This graph provides a measure of how much the residual x-variance of any individual sample varies compared to other samples, and how distant that sample point is from the model center (leverage). As can be seen in Figure 2(b), with a few exceptions, most samples fit relatively well into the cotton category. Figure 2(c) is a plot of the regression coefficients versus wavelength for the second PC. This plot shows that some wavelengths contribute positively to the model while others contribute negatively. The curve falls gradually, but aside from a small, broad feature between 1600 and 1700 nm, there are few distinguishing characteristics. Figure 2(d) is a plot of residual variance as a function of the number of principal components. Two principal components appear to explain approximately 100% of the spectral.

Using cross-validation, the Unscrambler^© program automatically defaults to a suggested number of principal components that the program deems necessary to explain the variation in the spectral data, which in this case was two (Figure 2(d)). But, close examination of Figure 2(d) shows that the residual variance continues to drop slowly beyond the second principal component. Several different numbers of PCs were tried, and models were computed using both the default number and values up to six [21] -[24] . For cotton as well as the other fabric types studied, six PCs did not appear to over-determine the data and were used in the SIMCA classification studies.

SIMCA classification was performed on the fiber classes of acetate, cotton, polyester, rayon, silk, and wool using the principal component models for each class of fabric. The methodology involved removing a randomly selected number of samples (validation samples in Table 2) in the full set for each textile category (Table 1) and classifying those with a model built from all remaining samples in the same population (calibration samples in Table 2). Validation was performed ten times with a new set of validation samples for each of the six categories, and the percentage of correct classifications (positive result) was calculated at the 95% confidence level. The results are shown in Table 2 and indicate that SIMCA correctly identified the textiles in 89% - 98% of the cases.

SIMCA was also used to check for misclassifications. Each one of the 758 textile samples was individually tested for membership in each of the six textile categories. Tests were performed using PCA models constructed from all members of each textile group (total samples in Table 3). Four results were possible (Figure 3): (A) correctly classified in only one group; (B) classified in more than one group, with one group correct; (C) classified in one or more groups, with no group correct; and (D) not classified in any group. The best outcome is A. Outcome B is still a positive result, but the sample would require further testing. The results (A, B, C or D) are indicated in Table 3 for each textile. The percentage is relative to the total number of samples in the textile class.

As shown in Table 3, acetate, polyester, rayon and wool models performed well, with few misclassifications as other fiber types (columns B and C). Of these four, wool had the most misclassifications with a 5% chance of

^aAverage of ten, separate validation sets randomly selected and analyzed in each category using the indicated number of samples and 6 principal components.

^aSix principal components. ^bAlso classified as polyester, column B. ^cAlso classified as rayon, column B; misclassified as rayon, column C. ^dAlso classified as cotton, column B. ^eAlso classified as rayon, column B; misclassified as rayon, column C. ^fAlso classified as silk, column B; misclassified as silk, column C.

being identified as both silk and wool (3%) or just silk (2%). Both wool and silk are protein based and differ primarily in the relative amounts of amino acids present and the type of hydrogen-bonding holding the protein chains together, making classification difficult.

Cotton samples were correctly classified as cotton in the large majority of cases, but were frequently marked as both rayon and cotton (62%). This result can be anticipated from the similarity in the chemical structures of the two fibers (cellulosic). Silk samples were also correctly classified as silk in the large majority of cases, but were frequently marked as both silk and rayon (24%). The diffuse NIR reflectance spectra of both fibers are broad with few differentiating features. In cases where chemical structures and spectra are very similar, classification should be improved by adding more samples to the PCA models.

Silk is an expensive luxury fabric, and cheaper look-alikes are sometimes substituted for the genuine article. The initial textile library contained one “silk” sample, labeled as Silk-120, that SIMCA flagged as belonging to the polyester category, but not to silk. A visual inspection of the spectrum for Silk-120, compared to representative spectra of polyesters also suggested that the sample was counterfeit. A burn test confirmed that the fabric was polyester.