Circular economy of post-consumer textile waste: classification through infrared spectroscopy

CIRCULAR ECONOMY OF POST-CONSUMER TEXTILE WASTE: CLASSIFICATION THROUGH 1 INFRARED SPECTROSCOPY 2 Jordi-Roger Riba1*, Rosa Cantero2, Trini Canals2, Rita Puig2 3 1 Universitat Politècnica de Catalunya, Electrical Engineering Department, Rambla Sant Nebridi 22, 4 08222 Terrassa (Spain) 5 2 Universitat de Lleida, Department of Computer Science and Industrial Engineering, Pla de la 6 Massa 8, 08700 Igualada (Spain) 7 *Corresponding author; r[email protected]; Tel. +34 937398365 8 9 ABSTRACT (2000 chars with spaces) 10 The textile and fashion industry is amongst the most resource-intensive and polluting industries, 11 thus impacting the natural environment. During the last decades, there has been an increase in the 12 manufacturing of textiles. Europe consumes large amounts of textiles and clothing due to the 13 current “buy-and-throw-away” culture, so it is crucial to minimize the environmental footprint of 14 the textile and fashion industry. To this end, fashion and textiles should be part of a circular 15 economy, thus extending the life of textiles and clothes, while retaining textile fibers within a closed 16 circuit. There is a need of increasing textile recycling and reuse to minimize the production of virgin 17 textile fibers. However, textiles are mostly sorted manually, thus to process huge volumes of 18 materials and reduce the associated costs, automated sorting systems are required. This paper 19 presents an approach for the sensing and classifying parts of an automatic waste-textile-sorting 20 machine. To this end, the infrared spectra of the textile samples is analyzed and, by applying suitable 21 statistical multivariate methods specially designed to solve classification problems, 100% 22 classification accuracy of unknown fiber samples is reached. The results allow predicting that textile-23 2 fibers can be automatically classified with 100% accuracy at high speed, with no need to apply any 24 prior analytical treatment to the textile samples. 25 Keywords: textile fibers, textile sorting, multivariate analysis, Infrared spectroscopy, classification, 26 pattern recognition 27 28 NOMENCLATURE 29 ATR Attenuated total reflection CV Canonical variate CVA Canonical variate analysis CNN Convolutional neural network ELM Extreme learning machine FT-IR Fourier transform infrared IR Infrared k-NN k nearest neighbors LDA Linear discriminant analysis MLP Multi-layer perceptron NIR Near infrared PC Principal component PCA PLS Principal component analysis Partial least squares SIMCA Soft independent modeling of class SVM Support vector machines 30 3 1. INTRODUCTION 31 Circular economy is a new concept to help the society’s change towards a more sustainable 32 economy. This change needs to re-think and re-design production and consumption patterns to 33 avoid environmental impacts and maintain natural resources as long as possible in the technosphere 34 (Gaustad et al., 2018; Simon, 2019). 35 The current linear model (extraction of resources, production, use and landfilling) is not sustainable, 36 as the resources are limited and there is an ever growing demand (Suárez-Eiroa et al., 2019). 37 Opposite to this linear system, the aim of the circular economy is to provide maximum utility and 38 value of products, components, and materials (The Ellen MacArthur Foundation, 2012). 39 Efforts on deeper implementation of circular economy are nowadays made in several industrial 40 activities, such as packaging (Navarro et al., 2018; Civancik-uslu et al., 2019), agriculture and food 41 (Principato et al., 2019; Teigiserova et al., 2019) or textile (Esteve-Turrillas and de la Guardia, 2017; 42 Yousef et al., 2019). 43 The textile and fashion industry is amongst the most polluting and resource-intensive industries due 44 to the great consumption of water, energy and chemicals, thus affecting the natural environment. 45 The growth of the global population has led to an overall increase in the manufacturing of textiles. 46 European countries consume large amounts of clothes and textiles as a result of the current “buy-47 and-throw-away” culture. Thus, clothing represents the fourth most environmentally harmful 48 consumption area, after housing, transport and food (NCM, 2015). Therefore, this trend should be 49 reversed for the sake of economy and environmental aspects. It is mandatory to minimize the 50 environmental and social footprint related to Europe’s textile production and consumption while 51 improving its sustainability (Roos et al., 2015). To achieve these objectives, much work is required 52 at regional, national, and international levels so that textiles must be part of a circular economy, in 53 4 order to extend product life and preventing hazardous substances. This strategy should allow using 54 textiles again and again as part of a toxic-free cycle (Reichel et al., 2014). 55 According to Shen et al. (Shen et al., 2010), 63% of textile fibers are derived from petrochemicals, 56 thus giving rise to greenhouse gas emissions due to production and use. The remaining 37% includes 57 cotton (24%), a plant requiring large amounts of water (Micklin, 2007) and pesticides (FAO-ICAC, 58 2015), which contribute to toxic pollution (Bevilacqua et al., 2014). Thus, the recycling of cotton is 59 also extremely important (Esteve-Turrillas and de la Guardia, 2017). 60 Processes such as dyeing (Terinte et al., 2014), finishing or printing, produce toxic emissions as well 61 (Swedish Chemicals Agency, 2014), and the manufacturing processes related to textiles usually rely 62 on the use of fossil energy, thus generating greenhouse gas emissions (Roos et al., 2015). According 63 to the Swedish Chemicals Agency, textile production includes around 2,450 different chemicals, 64 1,150 of which being classified as hazardous, so they are of potential risk for the environment and 65 consumers during the use of the textiles (Swedish Chemicals Agency, 2014). 66 As said, water use, greenhouse-gases emissions, toxic chemicals and waste are the main 67 environmental problems that the textile industry needs to face (Allwood, 2006). 68 To significantly reduce the environmental and social footprint of the Europe’s textile industry, 69 radical changes are required, especially in the way in which textiles and clothes are designed, 70 produced, traded, used and recirculated (Sandin and Peters, 2018). Fashion and textiles should be 71 part of a circular economy, thus allowing textiles and clothes life to be extended, to retain textile 72 fibers within a closed circuit, so that they can be used again and again (Dahlbo et al., 2017). 73 Research publications (Hole and Hole, 2019) support the fact that textile recycling and reuse in 74 general reduce environmental impact compared to landfilling and incineration. Therefore, there is 75 a growing regulatory interest to increase textile reuse and recycling, which is consistent with the 76 5 European Union directive on waste (DIRECTIVE 2008/98/EC, 2008). Better reuse and recycling of 77 textiles can lessen virgin textile fibers production (Spathas, 2017). 78 Textile reuse involves different strategies, including trading, swapping, borrowing, renting or 79 inheriting. This can be facilitated by flea markets, second hand shops, garage sales, charities, online 80 marketplaces or clothing libraries among others. 81 Textile recycling usually involves a reprocessing stage of pre- or post-consumer textile waste for 82 being used in new products, both textile or non-textile. Routes for textile recycling can be classified 83 as chemical (depolymerization of polymeric fibers or dissolution of natural fibers), mechanical 84 (pretreatment) or thermal (conversion of PET pellets, chips or flakes into fibers by melt extrusion) 85 (Spathas, 2017). 86 Nowadays low recycling rates are achieved from post-consumer textile waste (Sandin and Peters, 87 2018). Large proportions of used natural or synthetic materials are often discarded as waste, going 88 to landfills instead of processed for reuse or recycling. This is mainly due to lack of specific collection 89 for post-consumer textile waste, the complexity to separate the different discarded textile materials 90 and the costs associated to sorting important volumes (Dahlbo et al., 2017). 91 Currently, textiles are sorted mostly manually. However, this has many drawbacks, including high 92 cost, low speed operation and the impossibility of a full automation, which is required to process 93 huge volumes of materials (Nørup et al., 2019). 94 Although some sorting machines are found in the market, conventional methods and systems for 95 sorting are usually incapable to classify different textile materials, or they require inputs from well-96 trained operators, being time consuming to operate, or excessively expensive to maintain. 97 There is one publication in the literature (Peets et al., 2017) stating that with the spectral data of 98 ATR-FTIR jointly with the application of PCA it was not always possible to distinguish cellulose-based 99 6 fibers (cotton, linen and sometimes viscose) and it was only partly possible to distinguish silk and 100 wool. In another publication (Xing et al., 2019), a system for classifying wool and cashmere fibers 101 based on fractal, parallel-line algorithm, and K-mean clustering algorithms is proposed based on 102 digital photographs of such fibers, obtaining identification rates between 85% and 97.5%. A recent 103 paper (Zhou et al., 2019) identified different types of fibers from the NIR spectrum by applying PCA, 104 SIMCA and LDA with only two classifiers, although it was difficult to distinguish between wool and 105 cashmere fibers. In (Chen et al., 2019) NIR spectroscopy is applied to perform a quantitative 106 determination of fiber components by applying PLS and ELM algorithms, showing that ELM can 107 generate better predictive models than PLS, with a similar computational cost. In (Liu et al., 2019) 108 waste textile fibers are classified from the NIR spectrum by applying SVM, MLP and CNN algorithms, 109 showing that CNN performs better than the others with classification rates between 92% and 98%. 110 The aim of the present work is to contribute in the sensing and classifying parts of an automatic 111 textile-sorting machine. It is done by using a more accurate mathematical modeling based on the 112 data from the IR spectrum, by applying state-of-the-art multivariate methods well suited for this 113 purpose, while improving the robustness of the model by analyzing a large number of textile 114 samples from different origins. The novelty of the method proposed here is the use of ATR-FTIR 115 spectra of the samples for textile recycling purposes (only one previous paper is found in the 116 literature) and the combined statistical multivariate algorithms, which are very powerful supervised 117 models not yet applied to this type of samples. 118 This paper is focused to develop a fast and accurate method for a direct and non-invasive sorting 119 and classification of different textile fibers used for clothing, which include natural, artificial and 120 synthetic fibers, from the spectral data obtained from the FTIR spectra of such samples, with no 121 need of any prior analytical treatment. The results of this paper are focused towards the automation 122 of textile-waste-materials sorting process. For this purpose, textile samples are analyzed by using 123 7 an ATR-FTIR spectrometer, with no previous sample pretreatment, and thus, this system does not 124 need the addition of any chemical or reagent. Therefore, the proposed system is simple and fast to 125 apply. It is known that FTIR spectral data typically includes thousands of data points, one per 126 wavenumber analyzed, and thus, multivariate mathematical methods are required to operate with 127 this large number of points. Such methods include feature reduction algorithms and classifiers, the 128 first ones designed to concentrate the relevant analytical information of the whole data set in a few 129 latent variables, which also let partially removing most of the noise included in the original spectral 130 data (Riba et al., 2020). To calculate the reduced set of latent variables, the principal component 131 analysis (PCA) algorithm is applied followed by the canonical variate analysis (CVA) algorithm. Next, 132 the nearest neighbor (kNN) classifier is applied, this algorithm providing as many output normalized 133 variables within the range 0 - 1 as types of textile fibers or classes defined in the problem, thus 134 assigning an incoming textile sample to the class having the highest output value. 135 This combined methodology (ATR-FTIR spectra and PCA+CVA+kNN mathematical treatment) 136 applied to sorting post-consumer textile-waste is described for the first time in the literature. 137 2. METHODOLOGY 138 This section describes the experimental details and methodology used to prove the accuracy and 139 usefulness of the approach proposed in this paper. 140 2.1. Samples collection and identification 141 This paper deals with 350 textile samples coming from different companies’ catalogs and supplied 142 by Fitex technology center. The whole set of samples includes 200 samples from natural fibers (50 143 cotton, 50 linen, 50 wool and 50 silk samples) and 150 samples from artificial and synthetic fibers 144 (50 viscose, 50 polyamide and 50 polyester samples). Artificial fibers are the ones obtained by 145 transformation of natural products (i.e., viscose comes from cellulose), while synthetic fibers are 146 obtained from oil derivatives. 147 8 With the aim of including the maximum variability in the group of samples studied, different colors 148 (light and dark) and presentations (yarn or fabric) are included. For a quick identification, each 149 sample is coded including catalog origin, color and presentation form (yarn or fabric). 150 To check the performance of the mathematical methods, the whole set of samples was split into 151 two subsets, i.e., the calibration and prediction subsets in the proportion 50%-50%, as shown in 152 Figure 1. 153 Artificial and synthetic fibers (150) - Polyamide (25 calibration + 25 test) - Polyester (25 calibration + 25 test) - Viscose (25 calibration + 25 test) Natural fibers (200) - Cotton (25 calibration + 25 test) - Linen (25 calibration + 25 test) - Wool (25 calibration + 25 test) - Silk (25 calibration + 25 test) 154 155 Figure 1. Summary of the 350 textile samples used in this work. 156 157 2.2. ATR-FTIR methodology 158 Middle infrared electromagnetic radiation, within the wavenumber range 4000 - 400 cm-1, is 159 energetic enough to cause transitions between rotational and vibrational levels of the molecular 160 bonds. Due to the high selectivity of the radiation absorption in the middle infrared because of the 161 molecular bonds, this region of the spectrum is widely used in both qualitative and quantitative 162 analysis. 163 ATR measurements take advantage of the behavior of the IR radiation beam, by passing through 164 two media with different refractive indices. In such systems, the IR beam passes through a crystal, 165 which is transparent to the IR radiation and has a high refractive index, at an angle of incidence 166 9 greater than the critical angle. When the beam reaches the crystal-sample interface, it is almost 167 completely reflected, and only a small fraction of the beam crosses the interface and penetrates the 168 sample slightly. The beam is attenuated in the regions of the infrared spectrum in which the sample 169 absorbs energy. The beam returns to the crystal and leaves at the opposite end of the crystal, and 170 then focus to the detector (McGill et al., 2014). The use of this technique will allow a rapid scanning 171 or acquisition of textile samples without any pretreatment. 172 The FTIR spectra of the textile samples analyzed in this work, were acquired by means of a 173 PerkinElmer Spectrum One (S/N 57458, Beaconsfield, UK) spectrometer equipped with an ATR 174 module. The spectra are recorded in the wavenumber range 4000–650 cm-1, with a resolution of 1 175 cm-1 by averaging four scans to minimize noise effects. Therefore, each original spectral signal 176 includes 3351 spectral points. Subsequently, the spectra are converted to the first and second 177 derivative modes, in order to improve the classification performance of multivariate classification 178 models applied to identify the different textile samples. 179 180 2.3. Mathematical classification approach 181 To solve classification or identification problems from complex datasets, different mathematical and 182 statistical algorithms are available. In such problems, the whole sample set is commonly split into 183 two subsets, i.e., the subsets including the calibration and prediction samples. This approach allows 184 both, calibrating or training the models and to evaluate the behavior and accuracy of the 185 classification model from different samples than those used during the calibration stage (see Figure 186 2). Due to the 3551 wavenumbers constituting the variables measured for each ATR-FTIR spectrum 187 of the textile samples requires to apply appropriate feature extraction/reduction methods. Such 188 algorithms are designed to compress the essential discriminating information included in the raw 189 16 to solve the classification problem, whereas the remaining data constitute the prediction set, which 280 is used to validate the identification procedure, by using different data than that used during the 281 calibration stage. 282 A total of 350 samples are analyzed, 200 corresponding to natural fibers (50 cotton, 50 linen, 50 283 wool and 50 silk samples) and 150 corresponding to synthetic fibers (50 polyamide, 50 polyester 284 and 50 viscose samples). Although viscose is an artificial fiber, for simplification purposes, in this 285 work it is included in the group named synthetic fibers. Therefore, the calibration set includes 175 286 samples (25 cotton, 25 linen, 25 wool, 25 silk, 25 polyamide, 25 polyester and 25 viscose samples), 287 whereas the prediction set includes the remaining 175 samples. The samples are classified by 288 applying the PCA + CVA + k-NN algorithms in this order, obtaining 100% success rate in the 289 classification results provided by the k-NN algorithm, whose results summarized in Table 2 are based 290 on the data shown in Figure 5. 291 CV1 -1.0 -0.8 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 Synthetic, cal Natural, cal Synthetic, pre. Natural, pre. a) 292 CV1 -1.0 -0.5 0.0 0.5 1.0 Synthetic, cal Natural, cal Synthetic, pre. Natural, pre. b) 293 17 CV1 -1.0 -0.5 0.0 0.5 1.0 Synthetic, cal Natural, cal Synthetic, pre. Natural, pre. c) 294 Figure 5. a) Classification of natural versus synthetic fibers from the raw FTIR spectral data by 295 applying the PCA (40 PCs, 99.0% variance) + CVA with 175 calibration and 175 validation samples. 296 b) Classification of natural versus synthetic fibers from the first derivative of the FTIR spectral data 297 by applying the PCA (66 PCs, 99.0% variance) + CVA with 175 calibration and 175 validation samples. 298 c) Classification of natural versus synthetic fibers from the second derivative of the FTIR spectral 299 data by applying the PCA (81 PCs, 99.0% variance) + CVA with 175 calibration and 175 validation 300 samples. 301 302 Table 2. Classification success rate of natural versus synthetic fibers following the PCA + CVA + k-303 NN approach over the 175 prediction samples 304 Preprocessing type k = 3 k = 4 k = 5 k = 6 Raw spectral data 175/175 175/175 175/175 175/175 First derivative of spectral data 175/175 175/175 175/175 175/175 Second derivative of spectral data 175/175 175/175 175/175 175/175 305 Results summarized in Figure 5 and Table 2 show that the PCA + CVA + k-NN approach allow 306 classifying between synthetic and natural fiber samples with 100% accuracy. 307 3.3. Second study. Identification of the different natural fibers 308 Once the unknown incoming samples have been classified successfully as synthetic or natural, this 309 section classifies the unknown natural fibers into four groups, i.e., cotton, linen, wool and silk. As 310 18 explained, both the calibration and prediction set consist of 25 samples of each types, that is, 100 311 samples in total each. 312 The classification results of the natural fibers (cotton, linen, wool and silk) are summarized in Figure 313 6 and Table 3. 314 -0,8 -0,6 -0,4 -0,2 0,0 0,2 0,4 0,6 0,8 -2,0 -1,5 -1,0 -0,5 0,0 0,51,01,5 -1,5 -1,0 -0,5 0,0 0,5 1,0 CV3 CV1 CV2 Cotton (Cal) Linen (Cal) Wool (Cal) Silk (Cal) Cotton (Pre) Linen (Pre) Wool (Pre) Silk (Pre) a)315 -2,0 -1,5 -1,0 -0,5 0,0 0,5 1,0 1,5 2,0 -4 -2 0 2 -2 -1 0 1 2 CV3 CV1 CV2 Cotton (Cal) Linen (Cal) Wool (Cal) Silk (Cal) Cotton (Pre) Linen (Pre) Wool (Pre) Silk (Pre) b) 316 317 318 19 -4 -3 -2 -1 0 1 2 -6 -4 -2 024 -2 0 2 CV3 CV1 CV2 Cotton (Cal) Linen (Cal) Wool (Cal) Silk (Cal) Cotton (Pre) Linen (Pre) Wool (Pre) Silk (Pre) c) 319 Figure 6. a) Classification of the different natural fibers from the raw FTIR spectral data by applying 320 the PCA (31 PCs, 99.0% variance) + CVA with 100 calibration and 100 validation samples. b) 321 Classification of the different natural fibers from the first derivative of the FTIR spectral data by 322 applying the PCA (63 PCs, 99.0% variance) + CVA with 100 calibration and 100 validation samples. c) 323 Classification of the different natural fibers from the second derivative of the FTIR spectral data by 324 applying the PCA (70 PCs, 99.0% variance) + CVA with 100 calibration and 100 validation samples. 325 326 Table 3. Classification success rate of natural fibers (cotton, linen, wool and silk) following the PCA 327 + CVA + k-NN approach over the 100 prediction samples 328 Preprocessing type k = 3 k = 4 k = 5 k = 6 Raw spectral data 100/100 100/100 100/100 100/100 First derivative of spectral data 100/100 100/100 100/100 100/100 Second derivative of spectral data 100/100 100/100 100/100 100/100 329 Results summarized in Figure 6 and Table 3 show that the PCA + CVA + k-NN approach allow 330 classifying between cotton, linen, wool and silk fiber samples with 100% accuracy. 331 332 333 20 3.4. Third study. Identification of the different synthetic fibers 334 Once the unknown incoming samples have been classified successfully as synthetic or natural, this 335 section classifies the unknown synthetic fibers into three groups, i.e., polyamide, polyester and 336 viscose. Both the calibration and prediction set consist of 25 samples of each types, that is, 75 337 samples in total each. 338 The classification results of the synthetic fibers (polyamide, polyester and viscose) are summarized 339 in Figure 7 and Table 4. 340 CV1 -2 -1 01 CV2 -0,6 -0,4 -0,2 0,0 0,2 0,4 0,6 Polyamide (Cal) Polyester (Cal) Viscose (Cal) Polyamide (Pre) Polyester (Pre) Viscose (Pre) a) 341 342 CV1 -2 0 CV2 -1,0 -0,5 0,0 0,5 1,0 Polyamide (Cal) Polyester (Cal) Viscose (Cal) Polyamide (Pre) Polyester (Pre) Viscose (Pre) b) 343 344 21 CV1 -1 0 1 2 CV2 -1,0 -0,5 0,0 0,5 1,0 Polyamide (Cal) Polyester (Cal) Viscose (Cal) Polyamide (Pre) Polyester (Pre) Viscose (Pre) c) 345 Figure 7. a) Classification of the different synthetic fibers from the raw FTIR spectral data by applying 346 the PCA (6 PCs, 99.0% variance) + CVA with 75 calibration and 75 validation samples. b) 347 Classification of the different natural fibers from the first derivative of the FTIR spectral data by 348 applying the PCA (29 PCs, 99.0% variance) + CVA with 75 calibration and 75 validation samples. c) 349 Classification of the different natural fibers from the second derivative of the FTIR spectral data by 350 applying the PCA (39 PCs, 99.0% variance) + CVA with 75 calibration and 75 validation samples. 351 352 Table 4. Classification success rate of synthetic fibers (polyamide, polyester and viscose) following 353 the PCA + CVA + k-NN approach over the 75 prediction samples 354 Preprocessing type k = 3 k = 4 k = 5 k = 6 Raw spectral data 75/75 75/75 75/75 75/75 First derivative of spectral data 75/75 75/75 75/75 75/75 Second derivative of spectral data 75/75 75/75 75/75 75/75 355 Results summarized in Figure 7 and Table 4 show that the PCA + CVA + k-NN approach allow 356 classifying between polyamide, polyester and viscose fiber samples with 100% accuracy. 357 358 3.5. Challenges of this new technique and comparison with the literature 359 360 22 As shown in Table 5, there is only one author (Peets et al.,2017; Peets et al., 2019) using FTIR textile-361 spectra (like in the present study) for identification of different textile fibers and mixtures. 362 Nevertheless, these papers use a very simple mathematical treatment (PCA), thus not being able to 363 differentiate among very similar textile fibers (i.e., cotton/linen/viscose). 364 On the other hand, there are 4 papers in the literature using NIR spectra to classify textile samples, 365 three of them for recycling purposes (Liu et al., 2019; Zhou et al., 2018; Zhou et al, 2019). 366 Nevertheless, only Zhou et al., 2019 are using advanced mathematical algorithms being able to 367 achieve 100% recognition rate (same as the present described technique), but they do not include 368 cotton/linen/viscose (which are the most difficult to distinguish). 369 Table 5. Comparison of results with the previously published in the literature. 370 Reference Types of textile fibers Aim Type of spectrum Mathematic algorithms Recognition rate (%) (Peets et al., 2017) 11 + mixtures Quality control ATR-FTIR PCA No distinction among: cotton/linen/viscose Nor wool/silk (Peets et al., 2019) 16 + mixtures Quality control FTIR PCA No distinction among: cotton/linen/viscose (Chen et al., 2019) 4 + mixtures (wool, polyester, nylon, polyacrilonitrile) Quality control NIR PLS or ELM ELM better predictions (Liu et al., 2019) 2 + mixtures (polyester, wool) Textile recycling NIR SVM, MLP + CNN 92-98% (Zhou et al., 2018) 6 no linen, nor viscose Textile recycling NIR SIMCA 97% (cotton/polyester 90%) (Zhou et al., 2019) 7 no linen, nor viscose Textile recycling NIR PCA, SIMCA, LDA 100% Present paper 7 Textile recycling ATR-FTIR PCA, CVA + k-NN 100% 371 The present technique has shown better results than the described up to now in the literature, thus 372 being a promising option. 373 Nevertheless, further work must be performed before implementation in real sorting machinery, 374 like producing the specific software to be implemented and to make the IR-spectra-database robust 375 23 enough to be able to correctly classify dirty-wet textile-waste entering the recycling system. In 376 addition, after sorting the textile by type of fiber a second sorting by color will be needed (i.e. black- 377 colored cotton-fibers all together), thus reducing additional dyeing. 378 One possible drawback of the present FTIR technique, for its automation at industrial scale, is the 379 contact needed between the sensor and the textile, to register its IR spectrum and compare with 380 the database for classification. A strict maintenance protocol of the sensor would be advisable. 381 4. CONCLUSIONS 382 Today, only a small portion of the textiles is reused or recycled and they are mostly sorted manually. 383 This paper has proposed an automatic sensing and sorting approach focused to increase textile 384 recycling and reuse for minimizing the production and trade of virgin textile fibers which tries to 385 contribute to minimize the environmental problems that the textile and fashion industry is facing. 386 The sorting approach proposed in this work is based on the ATR-FTIR spectrum of the textile 387 samples, which once acquired is processed by means of several algorithms, including the PCA, CVA 388 and k-NN mathematical methods. 389 Experimental results presented in this paper, which are based on 350 textile samples (from 390 companies’ catalogs), have shown that the incoming unknown fiber samples can be automatically 391 classified with 100% accuracy and high speed, with no need to apply any prior analytical treatment 392 to the textile samples. These excellent results prove that the methodology suggested in this work 393 can be a valuable tool for sorting textile fibers for further reuse and recycling. 394 The present promising technique needs further development before its implementation to actual 395 sorting machinery (i.e., software developing, sorting fiber blends, additional sorting by color and a 396 more robust IR database including dirty-wet textiles from postconsumer waste). 397 24 The sorting approach proposed in this paper can be fully automatized for future industrial 398 application, thus allowing to process large volumes of materials and reduce the costs associated to 399 the sorting processes. 400 401 5. ACKNOWLEDGMENTS 402 The authors wish to acknowledge the collaboration of Fitex technological center for providing 403 several catalogs with different samples of textile materials. 404 6. REFERENCES 405 Allwood, J.M., 2006. Well dressed? : the present and future sustainability of clothing and textiles in 406 the United Kingdom. University of Cambridge, Institute of Manufacturing, Cambridge. 407 Bevilacqua, M., Ciarapica, F.E., Mazzuto, G., Paciarotti, C., 2014. Environmental analysis of a cotton 408 yarn supply Chain. J. Clean. Prod. 82, 154–165. https://doi.org/10.1016/j.jclepro.2014.06.082 409 Chen, H., Tan, C., Lin, Z., 2019. Quantitative Determination of the Fiber Components in Textiles by 410 Near-Infrared Spectroscopy and Extreme Learning Machine. Anal. Lett. 411 https://doi.org/10.1080/00032719.2019.1683742 412 Civancik-uslu, D., Puig, R., Voigt, S., Walter, D., Fullana-i-palmer, P., 2019. Resources , Conservation 413 & Recycling Improving the production chain with LCA and eco-design : application to 414 cosmetic packaging. Resour. Conserv. Recycl. 151, 104475. 415 https://doi.org/10.1016/j.resconrec.2019.104475 416 Dahlbo, H., Aalto, K., Eskelinen, H., Salmenperä, H., 2017. Increasing textile circulation—417 Consequences and requirements. Sustain. Prod. Consum. 9, 44–57. 418 https://doi.org/10.1016/j.spc.2016.06.005 419 25 DIRECTIVE 2008/98/EC, 2008. DIRECTIVE 2008/98/EC OF THE EUROPEAN PARLIAMENT AND OF 420 THE COUNCIL of 19 November 2008 on waste and repealing certain Directives (Text with EEA 421 relevance). 422 Esteve-Turrillas, F.A., de la Guardia, M., 2017. Environmental impact of Recover cotton in textile 423 industry. Resour. Conserv. Recycl. 116, 107–115. 424 https://doi.org/10.1016/j.resconrec.2016.09.034 425 FAO-ICAC, 2015. Measuring Sustainability in Cotton Farming Systems: Towards a Guidance 426 Framework. 427 Gaustad, G., Krystofik, M., Bustamante, M., Badami, K., 2018. Circular economy strategies for 428 mitigating critical material supply issues. Resour. Conserv. Recycl. 135, 24–33. 429 https://doi.org/10.1016/J.RESCONREC.2017.08.002 430 Hole, G., Hole, A.S., 2019. Recycling as the way to greener production: A mini review. J. Clean. 431 Prod. 212, 910–915. https://doi.org/10.1016/J.JCLEPRO.2018.12.080 432 Liu, Z., Li, W., Wei, Z., 2019. Qualitative classification of waste textiles based on near infrared 433 spectroscopy and the convolutional network. Text. Res. J. 434 https://doi.org/10.1177/0040517519886032 435 McGill, R.A., Stievater, T.H., Pruessner, M.W., Holmstrom, S.A., Nierenberg, K., McGill, R., Nguyen, 436 V., Park, D., Kendziora, C., Furstenberg, R., 2014. Infrared molecular binding spectroscopy 437 realized in sorbent coated microfabricated devices, in: Next-Generation Spectroscopic 438 Technologies VII. SPIE, p. 910107. https://doi.org/10.1117/12.2050819 439 Micklin, P., 2007. The Aral Sea Disaster. Annu. Rev. Earth Planet. Sci. 35, 47–72. 440 https://doi.org/10.1146/annurev.earth.35.031306.140120 441