Abstract
Textiles are an integral part of our everyday lives, e.g., in the form of clothing and furniture. Consequently, their analysis is of great interest in a wide range of application areas including quality control, textile recycling, forensics, and cultural heritage studies. Optical methods offer a large potential for rapid and nondestructive textile inspection. Raman spectroscopy is a particularly promising candidate for this task as it delivers a fingerprint of the sample on the molecular level, thus giving insights into the textile material composition. Currently, a major challenge preventing the widespread use of Raman spectroscopy for textile analysis is the strong fluorescence background originating from additives, e.g., dyes or the fiber material itself. To address this main limitation, this study applies shifted excitation Raman difference spectroscopy (SERDS) for the effective separation of Raman signals from interfering backgrounds. Using a customized Raman setup based on a dual-wavelength diode laser emitting at 785.2 and 784.6 nm, we present investigations for textile analysis and identification on a set of 22 dyed and undyed pure textiles and binary textile mixtures made of cotton, viscose, wool, polyester, and elastane. SERDS in combination with multivariate analysis is demonstrated as a powerful tool for a distinction between different fiber types and fibers of one material colored with various dyes. These results highlight the large potential of SERDS for textile material identification in selected application areas.
This is a visual representation of the abstract.
Keywords
Introduction
Textiles are used for clothing, furniture and carpets, bags, and ropes. A wide range of fields and applications aim to precisely analyze the material of such objects, such as forensics, 1 cultural heritage studies, 2 and the textile industry for the purpose of quality control or finding adequate recycling options after use. 3 However, the material analysis of textiles yields several challenges.
Textiles may be comprised of only one fiber type (e.g., cotton, wool, polyester, etc.) or of a mixture of two or more fiber types. Besides the main textile material, other substances are found on textile samples that have been added in the manufacturing process, so-called additives, including both dyes as well as other finishings such as crease-resistant, flame retardants, and water repellents. Samples from used textile objects may even include further substances to which the textiles have been exposed during washing and use.
Optical methods offer a large potential for rapid and nondestructive textile inspection. Raman spectroscopy is a particularly promising candidate for this task as it delivers a fingerprint of the sample on the molecular level, thus giving insights into the textile material composition. The technique is thus explored for material analysis of textiles in different fields including forensics, 4 cultural heritage studies, 2 and quality control in the textile industry. 5 Besides being a fast and nondestructive method, Raman spectroscopy can provide information about the fiber material as well as other substances on the fiber such as dyes below concentrations of 1% in the textile. 6 The possibility to choose between different excitation wavelengths gives room for further adjustment with regard to the substances of interest.
However, Raman spectra obtained using excitation wavelengths within the visible range are often impossible to analyze due to the strong background caused by fluorescence, which is common for organic samples especially if dyes are involved. 7 Various approaches are thus reported to address the fluorescence issue in Raman spectra. The simplest way is a mathematical subtraction of the fluorescent background from the measured spectrum. This method is effective in correcting the background of Raman spectra with moderate fluorescence, but it is unable to retrieve meaningful Raman spectra from measurements strongly affected by fluorescence.8,9 An experimental approach to reduce the fluorescence intensity is the prolonged exposure of the sample to the excitation laser radiation before the acquisition of the Raman spectra, the so-called photobleaching. 10 Laser-induced photobleaching reduced the fluorescence in a study by Macdonald and Wyeth, 11 but damaged the textile sample as argued by the authors. Furthermore, the required time for the bleaching can be a challenge for applications requiring high sample throughput.
A common approach to measure informative Raman spectra is to analyze a sample with a set of different excitation wavelengths and to find a wavelength that produces spectra with recognizable Raman spectral signatures but little interference from fluorescence.12–15 This approach increases instrumental complexity significantly as it requires several laser sources emitting in different spectral regions as well as matching spectrometers and detectors. Yet, even if a variety of lasers is available, it is not guaranteed that a high-quality Raman spectrum can be acquired for each sample. Instead, it may still prove difficult for some samples to find an appropriate excitation wavelength to achieve a Raman spectrum without a fluorescence background.
A more advanced method to separate Raman signals from fluorescence interference is shifted excitation Raman difference spectroscopy (SERDS). 16 Here, Raman spectra of the sample are recorded consecutively by applying two slightly different excitation wavelengths with a spectral separation on the order of the full width half-maximum (FWHM) of the target Raman signals. 17 The characteristic Raman signals of the sample under investigation shift along with the shift in excitation wavelength, while other contributions, such as fluorescence, remain unaffected. Consequently, subtraction of the two recorded Raman spectra can be used to separate the Raman signals from interfering backgrounds. SERDS can not only address contributions caused by fluorescence but also other backgrounds such as ambient light, which makes the technique particularly interesting for applications outside the laboratory.18,19
This work investigates the potential of SERDS to analyze the composition of selected textile samples. As fluorescence interference from textile fibers generally decreases when near-infrared (NIR) lasers are applied, 4 we chose an excitation wavelength of 785 nm. In the first step, SERDS is used to analyze undyed textile samples made from the fiber materials cotton, viscose (also referred to as rayon), wool, and polyester, and the capability of the method to differentiate different fiber materials is assessed. In the second step, dyed textile samples of the same fiber materials are investigated to study the effect of dyes on the Raman fingerprint. Finally, two samples made from binary textile blends (cotton–polyester and wool–polyester) and a third textile sample with blended material containing polyester and elastane (also known by the brand names Spandex and Lycra) are studied by SERDS. The acquired spectra are interpreted by visual inspection as well as by the multivariate analysis techniques principal component analysis (PCA) 20 and partial least squares discriminant analysis (PLS-DA). 21 This approach enabled the distinction between different textile types as well as the discrimination between identical materials of different colors.
Materials and Methods
Shifted Excitation Raman Difference Spectroscopy (SERDS) Instrumentation
The SERDS experiments are carried out by means of a compact laboratory setup that is described in detail elsewhere 22 and only a brief overview of the main components is given here. An in-house developed 785 nm dual-wavelength diode laser 23 is used as an excitation light source. In the excitation path, the laser radiation is initially passing an optical isolator (FI-780-5TVC, Qioptiq) to prevent back-reflections into the laser cavity and is subsequently focused into a 100 µm core diameter optical fiber (LEONI Fiber Optics). The beam is collimated at the fiber output, passes through two bandpass filters (LL01-785-25, Semrock), and is reflected at a Raman longpass filter (DI02-R785-25 × 36, Semrock) and a silver mirror (Qioptiq). An achromatic lens with a focal length of 30 mm and a diameter of 25.4 mm (Thorlabs) finally focuses the laser light onto the textile specimens generating a spot size of approximately 100 µm. Sample positioning is realized by means of a manual three-axis translation stage (LX30/M, Thorlabs).
In the detection path, the backscattered light from the sample is collected by the same lens used for focusing the excitation radiation. After reflection at a silver mirror and transmission through a set of three Raman longpass filters (DI02-R785-25 × 36 and LP02-785RU-25, Semrock), the Raman Stokes-scattered light is focused into a 200 µm core diameter optical fiber (Thorlabs). This collection fiber then transports the light to the spectrometer (Tornado U1, Tornado Spectral Systems, 4 cm–1 optical resolution) with an attached charge-coupled device (CCD) detector (MityCCD H10141, CriticalLink) that is thermoelectrically cooled down to −10 °C. An in-house written software is used to control the laser operation parameters and to record the Raman spectra.
Sample Material and Measurement Parameters
Because Raman spectroscopy investigates specimens on their molecular level, the primary characteristics of interest for this study are the textile material and the dye. Other properties such as weaving type, thread diameter, or cloth thickness are not expected to influence the relative peak intensities in the Raman spectra. Moreover, the investigation of the potential effects of such physical parameters on the recorded Raman spectra is beyond the scope of the present study. Therefore, the sample set of this study is composed of in total of 22 textiles produced from cellulose-based fibers (cotton, viscose), animal-based fibers (wool), and artificial fibers (polyester and elastane), representing the relevant textile materials marketed today. Besides undyed (white) specimens for each investigated fiber type, a range of differently dyed pure materials as well as chosen binary mixtures of polyester with different fiber types (cotton, wool, and elastane) are investigated. An overview of the specimens, a compilation of individual acquisition times, and the number of averaged spectra for each measurement position are given in Table I. To ease comparison between different samples despite varying amounts of fluorescence interference, the total acquisition time (individual acquisition time used for a single spectrum times the number of averaged spectra) is kept constant at 50 s for each of the two excitation wavelengths used for SERDS. The acquisition times are individually selected for each textile sample in order to avoid saturation of the CCD detector. To account for potential sample heterogeneity, each specimen is probed at 10 different spots using an optical power of 20 mW at the sample position. One exception is the extremely fluorescent blue polyester–cotton mixture that is exemplarily investigated at only three different spots.
Overview of material and color of investigated textile samples as well as individual acquisition times and number of averaged spectra used for SERDS measurements.
Shifted Excitation Raman Difference Spectroscopy (SERDS) Data Analysis
The detailed procedure for SERDS data processing based on an in-house developed algorithm implemented in Matlab (R2017a, The MathWorks, Inc.) is outlined in our previous publication, 22 and only a brief overview is given in the following. At first, the SERDS difference spectrum is calculated from the average Raman spectra recorded at each of the two excitation wavelengths. Subsequently, a baseline correction of the derivative-shaped difference spectrum by means of a cubic spline function followed by a numerical integration is calculated to obtain a reconstructed SERDS spectrum in a conventional form. In the final step, a baseline correction of the reconstructed SERDS spectrum is performed to obtain a straight horizontal baseline.
For PCA, the Matlab function “pca” included in the “Statistics and Machine Learning Toolbox” is used while the PLS-DA is performed using the “Classification toolbox for Matlab (v.5.4)” developed by Milano Chemometrics and QSAR Research Group. 21 Prior to multivariate analysis using PCA and PLS-DA, the SERDS spectra are truncated to the 340–1800 cm–1 spectral region where the majority of characteristic Raman signals of textiles and additives are found. All investigated dyed and undyed pure textiles as well as the polyester (90%)–elastane (10%) and the polyester (40%)–wool (60%) textile mixtures are included in the analysis. Individual spectra are then normalized to their respective maximum value and the data set is mean centered. Cross-validation for PLS-DA is performed using Venetian blinds with 10 data splits. The number of latent variables (LVs) to be included in the PLS-DA model is determined to be five by minimizing the model error.
Results and Discussion
Raman and SERDS Spectra of Undyed Wool
The averaged Raman spectra obtained for the two laser excitation wavelengths (784.6 and 785.2 nm) in the case of undyed white wool are exemplarily displayed in the top part of Figure 1. It becomes obvious that even without the presence of a dye, a pronounced fluorescence interference can be present for natural fibers such as wool. Due to this strong background contribution, virtually no Raman signals of wool can be recognized in the recorded Raman spectra. Only a weak signal around 1000 cm–1 can barely be identified as it shifts with the applied shift in excitation wavelength of 10 cm–1, i.e., 0.6 nm (indicated by dashed vertical lines). Obviously, SERDS can realize an efficient separation of the characteristic Raman signals of wool from the disturbing fluorescence interference, as shown in the bottom curve in Figure 1. The SERDS spectrum thus permits for identification of selected Raman signals of wool, which can be attributed to keratins (a group of proteins) and identified using the literature. 24 This result illustrates the potential of SERDS for the inspection of fibers of natural origin, e.g., wool, even when the material is undyed. Previous reports mention fluorescence interference for undyed fibers mainly with visible excitation wavelengths, e.g., for cotton using 514.5 nm 25 or polypropylene using 632.8 nm. 26 Our study demonstrates that the phenomenon of fluorescence can also occur for 785 nm NIR excitation in the case of undyed wool.
Average of 10 Raman spectra (top curves) excited at 785.2 and 784.6 nm and the corresponding reconstructed SERDS spectrum (bottom curve) obtained from one single measurement position of an undyed wool sample. Numbers in the SERDS spectrum indicate identified major Raman signals of keratins (proteins).
SERDS Spectra of Undyed Textile Specimens
The SERDS spectra of the undyed fiber types are collected to generate a set of reference spectra suitable for fiber type identification in the case of dyed textiles. The average spectra obtained from 10 different measurement spots from white specimens of cotton, viscose, wool, and polyester are presented in Figure 2. The spectra are normalized to their respective maximum and are vertically offset for clarity. For a better visual representation, only the strongest Raman signals of each substance, having an intensity of at least one-quarter of the strongest signal within the investigated spectral range, have their Raman signal position indicated in Figure 2. However, for each textile type, all detected Raman signals that can be identified with the literature are listed in Table II.
Average SERDS spectra of undyed textiles using 10 different measurement positions probed with 5 s individual acquisition time and 20 averaged spectra each. Numbers indicate Raman signal positions (in cm–1) of the strongest signals for each material. The spectra are normalized to their respective maximum and are vertically offset for clarity.
Identified Raman signal positions obtained from average spectra of undyed textiles. Listed are also the identified molecular constituents within each type of material. Bold numbers represent the strongest Raman signals of each substance.
Based on the spectral signatures obtained, the cotton sample has a typical Raman spectrum of cellulose I while the viscose specimen exhibits the characteristic spectral pattern of cellulose II.27,28 Both cellulose modifications can readily be distinguished by specific differences in their spectral fingerprint, e.g., considering the Raman signal positions at 358 and 382 cm–1 (CCC, CO, CCO, and ring deformation) or the relative signal intensity at 897 cm–1 (C–H deformation). 27
For wool, a spectrum of keratins as structural fiber proteins is obtained with the most prominent signals located at 1001 cm–1 (phenylalanine C–C stretch), 1448 cm–1 (CH2 and CH3 bending modes), and 1650 cm–1 (amide I C–O stretch).24,30 The polyester fiber can be identified to be composed of polyethylene terephthalate (PET) showing the strongest Raman signals at 856 cm–1 (ring C–C breathing), 1288 cm–1 (ring-carbonyl stretch, O–C stretch, and ring CH in-plane bend), 1611 cm–1 (ring C=C stretch), and 1722 cm–1 (C=O stretch).32,33 Overall, each investigated fiber type, probed in undyed form, exhibits a unique spectral fingerprint that enables identification of individual materials and discrimination between the different fiber types.
SERDS Spectra of Colored Textiles
For each probed material, dyed and undyed (white) specimens are investigated and their average SERDS spectra are presented in Figure 3. In the case of the cotton samples, the spectra are normalized to the strong cellulose I Raman signal at 382 cm–1 and are vertically offset for clarity (see Figure 3a). It turns out that the characteristic Raman signal at 382 cm–1 (marked with an arrow) can be used as an effective indicator for the presence of cotton as it appears in all investigated cotton specimens, irrespective of their color. The undyed (white) and light-colored (cream) samples show an almost unperturbed spectral signature of cellulose I as given in the reference spectrum in the top part of Figure 2. The red sample exhibits the cellulose I spectrum being overlaid by the spectral pattern of additives that show strong Raman signals predominantly in the spectral range between 1100 and 1650 cm–1. The intensity of these Raman signals is on the same order of magnitude as the Raman signal intensity obtained from the cotton textile matrix.
Shifted excitation Raman difference spectroscopy spectra of undyed and dyed textile samples. The spectra are normalized to the characteristic peak chosen for each textile material and are vertically offset for clarity. Indicator Raman signals for the presence of individual textile types are marked by arrows.
In general, all Raman signals not originating from the fiber matrix itself can be caused by various additives, including dyes and other finishings. As it is very likely that the dyes used to color the textiles are responsible for such strong Raman signals, in the following discussion all signals not attributed to the fiber material are regarded as originating from the respective dyes. It should, however, be noted that small contributions from other additives cannot be ruled out. To clarify this point, detailed investigations on a variety of pure additives commonly used in the textile industry need to be conducted to determine their individual spectral signature but this is far beyond the scope of the present study.
The dark blue specimen shows very strong Raman signals of the dye that are up to one order of magnitude more intense than the cellulose I Raman signals. The Raman signals of the blue dye can be found throughout the entire investigated spectral range up to approximately 1650 cm–1. Nevertheless, an identification of the specimen as cotton can be realized in this case as well based on the characteristic indicator Raman signal of cellulose I at 382 cm–1 that does not overlap with any Raman signals originating from the dye.
Figure 3b displays the normalized SERDS spectra obtained from the viscose samples. While in the case of the two types of undyed (white) viscose the cellulose II Raman spectrum is clearly recognizable, there is some overlap from Raman signals of the dyes present for the colored specimens. An identification of the probed material as viscose is, however, possible for the red and purple textiles using the cellulose II Raman signal at 1095 cm–1 as an indicator (marked with an arrow). The strong and spectrally isolated Raman signal of cellulose II at 897 cm–1 turns out to be less suitable for identification as certain dyes, particularly obvious in the case of the purple dye, show a Raman signal at or around that signal position as well which could lead to some confusion. This fact is easily observed as the intensity ratio of the Raman signal at 1095 cm–1 to the Raman signal at 897 cm–1 is <1 for the purple sample while this ratio is >1 for the white and red viscose specimens. In the case of the red viscose, the Raman signals of the red dye show comparable Raman intensities with respect to the strongest Raman signals from the viscose matrix. For the purple viscose, the strongest Raman signals of the dye are up to a factor of 2.5 more intense compared to the cellulose II signals but material identification is nevertheless possible. In contrast, by means of the simple approach employed using a single indicator Raman signal, an identification of the black viscose specimen cannot readily be achieved. The challenge in this case is due to Raman signals of the dye being more intense than Raman signals of the textile matrix and also showing strong overlap with the spectral fingerprint of cellulose II.
An overview of the SERDS spectra of dyed and undyed wool samples is given in Figure 3c. The yellow sample shows a nearly unperturbed spectral signature of wool thus enabling straight identification, e.g., by means of the indicator Raman band of the amide I band at 1650 cm–1 (marked with an arrow). By comparison with the white wool specimen, a few Raman signals originating from the dye can be recognized but they are mostly of moderate intensity. In contrast, the two red-dyed wool samples show very strong Raman signals caused by the respective dye. These contributions are most pronounced in the 1100–1400 cm–1 spectral range and show intensities up to ninefold more intense than the wool indicator Raman signal. Both samples can, however, visually be identified as wool as the indicator Raman signal is not spectrally overlapping with any dye Raman signals and can clearly be recognized. By comparison of the spectral signatures of the two red dyes, SERDS also reveals that two different dyes have been applied to color these two specimens of wool. In the case of the black wool sample, the very strong Raman spectral signature of the dye is masking the molecular fingerprint of the underlying material. An identification of the textile type, based on the simple approach using one single target Raman signal of wool, is not possible.
The SERDS spectra of polyester are presented in Figure 3d. Here, the Raman signals of the fiber material are dominating the spectra of all probed samples, irrespective of their color. An unambiguous identification of all samples as polyester is clearly possible using the two strong Raman signals at 1611 cm–1 (marked with an arrow) and 1722 cm–1 as material-specific indicators. It is interesting to note that there are virtually no Raman signals originating from dyes identifiable, except for the red polyester sample where closer inspection shows two small Raman signals in the 1530–1590 cm–1 spectral region that are not present in the undyed specimens.
Overall, the simple approach for textile discrimination using the visual identification of characteristic and material-specific Raman signals performs reasonably well. The majority of undyed and dyed specimens made of cotton, viscose, wool, and polyester are successfully identified. Only in the case of two very dark-colored samples (black viscose and black wool), the applied univariate approach does not permit identification of the textile material. In the next section, selected binary textile mixtures, i.e., samples that are composed of polyester and another fiber type, are investigated.
Assessment of Colored Polyester-Containing Binary Textile Mixtures Using SERDS
The first mixed textile specimen is a bright fluorescent yellow composite of 90% polyester and 10% elastane. Visual inspection of the average SERDS spectrum obtained from 10 different measurement spots presented in Figure 4 shows a clear spectral fingerprint of polyester while apparently no obvious Raman signal of elastane or the yellow dye can be recognized. An unambiguous identification of the major compound of the sample as being polyester is thus realized in this case.
Average SERDS spectra of binary textile blends displayed for blue polyester (65%)–cotton (35%), black polyester (40%)–wool (60%), and yellow polyester (90%)–elastane (10%) mixture.
The next specimen is a black mixture of 40% polyester and 60% wool and its average SERDS spectrum is given in the center of Figure 4. Here, an identification of polyester as a minor component is possible based on the four characteristic Raman signals at 631, 856, 1611, and 1722 cm–1. Compared to polyester, wool is a weaker Raman scatterer, and its presence cannot be readily detected in the composite sample. The detected Raman spectroscopic signature of polyester is also superimposed by strong Raman signals of the black dye that are most prominent in the 1050–1350 cm–1 spectral range with intensities 3.4× as strong as the most intense polyester signal at 1611 cm–1.
The last composite textile is made of 65% polyester and 35% cotton and has a blue color. The average SERDS spectrum of three different measurement spots presented as a top curve in Figure 4 confirms the identification of the major polyester phase based on the characteristic two Raman signals located at 1611 and 1722 cm–1. The minor cotton phase cannot be observed in this case due to strong spectral features originating from the blue dye. These contributions are most pronounced in the spectral region up to 1400 cm–1 with intensities up to four times the intensity of the strongest polyester Raman signal located at 1611 cm–1.
It is interesting to note that the black polyester–wool mixture is not the most challenging to probe, despite having the darkest color of the three mixed textiles. In contrast, the polyester–cotton mixture, which does not even have a particularly dark shade of blue, has the strongest fluorescence interference that makes it necessary to reduce the acquisition times for individual spectra to 50 ms to avoid detector saturation. The observation from another study 9 that the darker the textile sample, the more intense the fluorescence interference, therefore does not generally hold true.
In the following sections, the potential of multivariate analysis techniques (PCA and PLS-DA) for distinction between the investigated textile types as well as for the discrimination between identical materials of different colors is evaluated.
Principal Component Analysis for Textile Investigation
The results show that a simple univariate approach using material-specific indicator Raman signals performs well for light-colored specimens but can have some limitations in the case of very dark-colored textiles. For a more detailed analysis of the investigated textile samples, multivariate techniques are applied to the recorded SERDS spectra. The spectral region between 340 and 1800 cm–1 is chosen as the majority of characteristic Raman signals of textiles and dyes are present in this range.
As a first step, PCA 20 is executed, reducing the dimensionality of the data while preserving large amounts of the variation present within the data set. As an unsupervised multivariate analysis method, the PCA algorithm is presented only with the spectral data (but no class or group assignment) and the separation of the samples in the reduced dimensions, called principal components (PCs), originates solely by the variance of the spectral data itself. In the case of clearly distinguishable spectral data, a pronounced separation is obtained while similar spectra are located adjacent to each other.
The score plots involving the first four PCs of the spectra of the textile sample set of this study are presented in Figure 5. As shown in Figure 5a, the main separation along PC 1 (explaining 24.6% of the overall variance in the data set) is between the set of textiles containing polyester and the remaining samples. Due to its intense Raman spectroscopic signature, which is unperturbed by Raman signals of the dyes, all pure polyester samples form a very compact cluster in the score plot, i.e., the textiles made of polyester can clearly be discriminated from textiles produced from all other fiber types. The mentioned cluster also includes the polyester (90%)–elastane (10%) textile mixture. This observation is not unexpected as the sample mainly contains polyester with elastane only being a minor compound. Furthermore, elastane-containing fibers are often core-spun fibers, which means that the main fiber material, in this case polyester, is spun around an elastane core. Hence, the elastane is located in the center of the fiber and thus is less accessible by optical methods. It is also noteworthy that the polyester (40%)–wool (60%) mixture is located between the wool and polyester specimens with respect to PC 1, indicating that potentially even semiquantitative analysis of mixed textile composition is feasible.
Principal component analysis (PCA) score plots including the first four PCs, projection into PC 1–PC 2 plane (a) and projection into PC 3–PC 4 plane (b).
Wool samples show negative score values with respect to PC 2 (18.2% variance explained) and almost exclusively negative scores with respect to PC 1. This quadrant of the score plot, however, also contains other materials, namely blue cotton as well as purple and black viscose. The remaining samples are located in the quadrant with negative PC 1 scores, but positive PC 2 scores. Separate clusters can be recognized for light-colored cotton (two white samples and one cream sample) and white viscose (two samples). Red cotton and red viscose form distinct clusters each as well. For the viscose and cotton specimens, a separation according to the color becomes evident along PC 2. The light-colored textiles (white and cream) show pronounced positive scores while the score values decrease with colors becoming darker. The red samples still show positive scores while all darker cellulose-based textiles (blue, purple, and black) have negative score values according to PC 2.
For further evaluation, Figure 5b presents a second score plot incorporating PC 3 (16.4% variance explained) and PC 4 (9.4% variance explained). The polyester samples and the polyester–elastane mixture form a tight cluster making their identification easy. Along PC 3, a pronounced separation of different textile colors is observed. In the case of wool, the most obvious separation becomes evident. Light-colored wool (white and yellow) has strong positive score values and is clearly separated from all other samples. Wool samples can be further discriminated along PC 4 with the two different red specimens having positive score values whereas the black samples display entirely negative scores.
The two white viscose samples show positive scores in the direction of PC 3 while the dyed viscose specimens have exclusively negative scores. The dyed viscose samples can further be differentiated according to the darkness of their color with respect to PC 4. Here, red specimens show larger negative scores than purple samples while black samples exhibit positive scores. In the case of cotton, light-colored samples (two white samples and one cream sample) display positive score values with respect to PC 3 while red specimens show small negative scores, and blue samples exhibit large negative scores.
The corresponding PCA loadings displayed in Figure 6 indicate that the contributions of Raman signals attributed to the dyes are mainly found in higher PCs. The first PC is primarily separating polyester-containing material from the other textiles while a certain distinction between cellulosic materials (cotton and viscose) and wool is realized by means of the second PC. Nevertheless, already PC 2 contains information from the dyes used to color the cellulosic materials and thus enables a coarse separation between the three groups light-colored (white or cream), red, and dark (blue, purple, and black). This trend becomes even more obvious in the case of PC 3 and PC 4 where the obtained separations between the samples are based on dye Raman signals to a large extent. However, in these cases, Raman signals of polyester can still be identified in the corresponding loadings but this is mainly due to the fact that these bands are very intense in the recorded spectra.
First four loadings of the PCA model (vertically offset for clarity), labels identify characteristic Raman signals of either textile matrices (indicated by C: cotton, V: viscose, W: wool, P: polyester) or dyes (indicated by wavenumbers).
Overall, using an unsupervised multivariate analysis technique with no information on textile materials or dyes, a decent separation of the investigated textiles can be obtained using the first four PCs explaining 68.6% of the overall variance in the data set. The separation observed in the score plots in Figure 5 can be identified to originate from spectroscopic information stemming from the individual textile matrices cotton, viscose, wool, and polyester as well as from selected dyes used to color the textiles. Compared to the univariate approach presented above using single material-specific indicator Raman signals, the multivariate analysis goes one step further by enabling not only a discrimination between different textile materials but also regarding the same material but with different colors. Therefore, SERDS in combination with PCA allows for a more precise identification and characterization of textile samples which could be beneficial for the application in forensic studies.
Discrimination Between Textile Materials Using PLS-DA
The textile samples are effectively separated by PCA based on their textile material and the used dyes. Despite the myriad of different substances in textiles influencing the Raman spectra, discrimination on a specific property, such as the textile material, is advantageous for some applications, e.g., for quality control in industry or for material separation for recycling purposes. To address this point, PLS-DA is applied to the SERDS spectra. The spectra of individual textiles as well as a corresponding class assignment with respect to the textile material are used as input for the algorithm.
In the first attempt, six different classes are used: cotton (class 1), viscose (class 2), wool (class 3), polyester (class 4), polyester–elastane mixture (class 5), and polyester–wool mixture (class 6). Considering five LVs giving a minimum cross-validation error, the algorithm is able to correctly classify all specimens, except for the polyester–elastane mixture where all 10 spectra are incorrectly assigned to the class containing the pure polyester samples. This result is not surprising as the recorded SERDS spectra of the polyester–elastane sample do not show any observable spectral signature of the elastane fraction. As mentioned above, this fact is likely caused by the low elastane fraction (10%) in the mixture and the structure of the fiber where the elastane core is covered by a polyester shell thus making it less accessible to the spectroscopic technique.
Based on this outcome, a second classification attempt is made by combining the pure polyester and the polyester–elastane mixture into one common class that includes all samples with a polyester content of at least 90%. The first class contains 50 spectra obtained from the five cotton samples (two times white, cream, red, and dark blue), the second class comprises 50 spectra recorded on the five viscose specimens (two times white, red, purple, and black), the third class contains 50 spectra obtained from the five wool samples (white, yellow, red, dark red, and black), the fourth class comprises 40 spectra of the four pure polyester specimens (two times white, light red, and light blue) and 10 spectra of the polyester–elastane mixture (yellow), and the fifth class contains the 10 spectra recorded on the polyester–wool mixture (black). Including five LVs within the PLS-DA model results in 100% classification accuracy.
Figure 7a presents a PLS-DA score plot including the first two LVs. Along LV 1 (explaining 24.3% of the overall variance in the data set), a clear separation of samples with at least 90% polyester content (blue circles, large negative score values) from specimens containing the polyester–wool mixture (purple crosses, small negative scores) and all other textile materials (exclusively positive scores) can be realized. LV 2 (17.3% variance explained) enables a distinction between the wool specimens (green triangles) exhibiting negative scores and both cellulose-containing materials cotton (red squares) and viscose (orange diamonds) that have positive scores. In this way, 41.6% of the total variance contained in the data set enables already a distinction between the material types polyester (PET), wool, and cellulose (cotton- and viscose-containing samples).
Partial least squares discriminant analysis (PLS-DA) score plots including the first four LVs from a model including five LVs in total, projection into LV 1–LV 2 plane (a) and projection into LV 3–LV 4 plane (b).
As displayed in Figure 7b, a separation between cotton (red squares) and viscose (orange diamonds) can be realized mostly by LV 3 (11.9% variance explained) but also with a contribution of LV 4 (10.9% variance explained). Overall, the majority of the material-specific distinction is already achieved by the first two LVs while the higher LVs enable the remaining discrimination between cotton and viscose. It should be noted that the inclusion of LV 5 (5.5% variance explained) is also necessary as in a PLS-DA model containing only four LVs, all 10 spectra recorded on the dark blue cotton sample are incorrectly assigned to the class containing the black polyester–cotton mixture (95.2% overall classification accuracy) indicating a residual influence of Raman signals originating from the dye in this case.
The results show that the model including five LVs (explaining 69.9% of the overall variance in the data set) gives an excellent material-specific classification performance. This finding is not unexpected as good model quality is not necessarily related to a very high percentage of explained variance, particularly in the case of complex data sets with thousands of variables, as e.g., for spectroscopic data. This is due to the fact that the primary goal of the classification model is not to describe the majority of the variance in the data but rather to distinguish between the different classes. 21 In our case, the information that is relevant to separate the five classes of textile materials with 100% accuracy is contained in the five LVs that account for approximately 70% of the overall variance in the data set.
The class coefficients representing characteristic spectral patterns for each of the five classes are displayed in Figure 8. Most of the positive contributions for classes 1 to 4 can be identified with characteristic Raman signals of the textile materials cotton, viscose, wool, and polyester using the wavenumber positions of the respective pure materials compiled in Table II. This indicates that the material-specific Raman spectroscopic signatures of the individual textile materials can successfully be extracted by PLS-DA, thus enabling an unambiguous identification of the probed textiles irrespective of their color. Further strong positive signals (not labeled in Figure 8), particularly evident in the class coefficients of cotton and viscose can be attributed to contributions from the dyes in the case of colored textiles. These signals do, however, not compromise the ability for an excellent material classification. In the case of the class containing only the black polyester–wool mixture, the main positive contributions to the class coefficient do neither originate from the polyester nor the wool fraction but can be attributed to Raman signals of the black dye.
Class coefficients of five classes included in the PLS-DA model (vertically offset for clarity), labels identify wavenumber positions of characteristic Raman signals of either textile materials (classes 1–4) or dyes (class 5).
It should be noted that in this pilot investigation the classes with pure materials, i.e., having fractions of the major textile material of at least 90%, contain 50 SERDS spectra obtained from five specimens each while the last class only contains 10 spectra measured on the polyester–wool mixture. It is thus possible that unequal class sizes could cause some bias. To address this point, future SERDS studies for textile material identification would benefit from including a larger variety of textile materials and textile material mixtures of various colors.
In summary, the presented results show that selected textile materials can be identified by SERDS in combination with PLS-DA. Hence, an automated analysis of textile samples is possible, which is relevant for industrial applications, such as quality control of manufacturing processes and recycling.
Conclusion
Fluorescence interference as a major challenge preventing the widespread use of Raman spectroscopy for textile analysis is addressed effectively in this study applying SERDS. The fluorescence does not only originate from dyes as reported in the literature,8,15 but can also be intrinsic to the fiber material itself, as shown in the case of undyed wool. However, SERDS using 785 nm excitation permits for efficient extraction of Raman signals of fibers and dyes from interfering backgrounds in all cases. A univariate approach using a single indicator Raman signal specific for each of the investigated textile types is sufficient for fiber identification in the case of undyed and light-colored textiles made of cotton, viscose, wool, and polyester. Raman spectra of darker textiles show a significant contribution of the dye which makes identification more difficult. However, SERDS in combination with multivariate analysis is demonstrated to be a powerful tool for discrimination between different fiber types including binary fiber mixtures and fibers of the same material colored with different dyes. Identifying the main material of textiles irrespective of the color and dye is accomplished by PLS-DA even for dark-colored textile samples. These results highlight the large potential of SERDS for textile material identification in various application fields, e.g., in industrial areas as well as in forensics and cultural heritage.
Footnotes
Acknowledgments
We are grateful to Maria Krichler (Ferdinand-Braun-Institut (FBH)) for developing the software to control the experimental setup. We would like to thank Marco Jagodzinski (Technical University of Berlin) for helpful discussions concerning the data analysis and comments regarding the manuscript.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the German Federal Ministry for the Environment, Nature Conservation, Nuclear Safety and Consumer Protection (BMUV) within the framework, “KI-Leuchttürme für Umwelt, Klima, Natur und Ressourcen” (AI Lighthouses for Environment, Climate, Nature, and Resources, Grant No. 67KI2016C) and by the Federal Ministry of Education and Research (BMBF) within the projects iCampus (Grant Nos. 16ES1132 and 16ME0425), and Research Fab Microelectronics Germany–FMD (Grant No. 16FMD02).
