A new approach to polymer identification by laser-induced breakdown spectroscopy (LIBS) with adjusting spectral weightings (ASW) was developed in this work aiming at improving the identification accuracy. This approach has been achieved through increasing the intensities of specific characteristic spectral lines which are important to polymer identification but difficult to be excited. Using the ASW method, the identification accuracies of all 11 polymers were increased to nearly 100%, while the accuracies of PE, PU, PP and PC were only 98%, 74%, 90% and 98%, respectively, without using the ASW method.
© 2014 Optical Society of America
With the extensive use of polymers, the classification and recycling of post-consumer polymers have become an urgent problem for environmental protection. Currently, manual sorting according to resin coding system , X-ray fluorescence (XRF) spectrometry, and near-infrared (NIR) spectroscopy [1, 2] are popular methods used for polymer identification. However, manual sorting is not only time-consuming but also error-prone, XRF is especially used to identify polyvinyl chloride (PVC) by detecting chlorine atoms [2–4]. NIR spectroscopy is the most widely used method for polymer identification, which determines the molecular structures of polymers by recording the reflected or transmitted radiation [5–7]. Though it is advantageous with high identification accuracy and speed, it is difficult to detect black polymers and the signal is vulnerable to the impurities of the sample surface [8, 9]. Laser-induced breakdown spectroscopy (LIBS) [10–14] is a novel method used for sorting post-consumer polymers with some intrinsic advantages [15–20, 29]. R. Sattmann et al. applied LIBS to identify five kinds of polymers and achieved a minimum accuracy of 90% . J. M. Anzano et al. performed identification of six kinds of polymers using LIBS combined with simple statistical correlation methods and reached a minimum correct identification rate of 80% . M. Boueri et al. used LIBS setup of a broad band spectrometer (240–820 nm) combined with artificial neural networks (ANNs) to identify eight polymer materials and minimum correct identification rate of 81% was achieved . S. Grégoire et al. carried out a detail analysis of the C2 swan (0, 0) bands in different polymers and found that the C2 signal is stronger for aromatic polymers than that for aliphatic polymers . Although the studies above had presented versatile approaches and obtained meaningful results, the current identification accuracy still cannot meet the requirement and the polymer species used in these works were not sufficient to identify the vast majority of polymers used in our daily lives.
To address these problems, a new approach to adjusting spectral weightings (ASW) for polymer identification was developed in this work, which is a new application of LIBS. Eleven polymers under argon background were successfully identified. The minimum identification accuracy was improved to nearly 100% using this approach.
2. Experiments and methods
2.1 Experimental setup and materials
The schematic diagram of the experimental setup used in this study is shown in Fig. 1. A second harmonic Q-Switch Nd:YAG pulsed laser (wavelength 532 nm, pulse duration 6 ns, repetition rate 10 Hz) was used. The laser beam was reflected by a dichroic mirror, and focused onto the sample surface by a UV-grade quartz lens with a focal length of 150 mm. The plasma spectrum was obtained by an echelle spectrometer (spectral range from 200 to 950 nm with a resolution of λ/Δλ = 5000) coupled with an intensified charge-coupled device (ICCD) camera (Andor Tech, iStar DH-334T). A computer-controlled platform was used to move the samples and the argon gas (Purity 99.999%) was used to shield the plasma from air by a gas pipe with a flow rate of 15 l/min. The samples used in this work include 11 kinds of polymers and their molecular formulas are listed in Table 1.
2.2 Experimental method
The LIBS spectra were produced under the following condition. The laser pulse energy was 44 mJ. The focal point was 1 mm below the polymers surface. The estimated fluence was about 140 J/cm2 (The spot diameter was about 100 μm.). The gate delay and gate width of the ICCD were set to be 1 μs and 1.1 μs, respectively. An integration time was set to be 3 sec per spectrum, corresponding to 30 laser pulses in total. 100 spectra were recorded for each polymer type under argon background. Totally, 1100 spectra for 11 polymers were obtained. The support vector machine (SVM) algorithm [22–25] was used in MATLAB R2011b (MathWorks Corporation, USA) for data processing and obtaining identification accuracy. The kernel function of SVM was radial basis function (RBF) . The penalty parameter “C” of the error term and kernel parameter “g” of RBF have a great impact on the identification results. Therefore, they are optimized by genetic algorithms combined with cross validation method and fixed as 20.021 and 0.0101, respectively. The details can be seen in literature [23, 24]. For each polymer, 100 spectra were divided equally into training set and test set. The former was used to train the SVM model. The later was used to validate the identification performance of the SVM model created by the training set.
2.3 Selection of characteristic spectral lines
The selection of characteristic spectral lines is important to polymers identification. The 12 kinds of main elements and two kinds of molecular bands, namely C, H, O, N, Mg, Ti, Ca, Al, K, Na, Cl, F, C-N and C-C, respectively, were detected by the spectrometer with a broad spectral range. The metallic elements (i.e., Mg, Ti, Ca, Al, K and Na) are commonly used to reduce production costs or to improve physical and mechanical properties [26, 28]. In selecting characteristic spectral lines, the criteria should be complied with the follows: 1) The intensities of spectral lines should reflect the element contents; The higher element content is, the stronger the spectral intensity should be. (2) Characteristic spectral lines that are self-absorbed or self-reversed should not be chosen. Thus 15 spectral lines listed in Table 2 were selected to differentiate and identify the polymers in this work. Before training the SVM model, the spectral intensities for each spectrum were normalized by the carbon spectral line located at 247.86 nm. Principal component analysis (PCA) was used to offer the weightings of normalized characteristic spectral lines in polymers classification .
3. Results and discussions
3.1 Polymers identification without ASW
For the original 100 spectra of each polymer, the former 1th~50th spectra were used to train the SVM model. The other 50 spectra were used to test the performance of the SVM model and to obtain identification accuracy. Figure 2 shows the identification results of the test set spectra by the trained SVM model. The X-axis represents the 550 spectra of the 11 polymers and the Y-axis represents the spectral labels which mark polymer types. Each spectrum has a label. The “○” represents actual label while the “*” represents the predicted label by the SVM model. The successful identification is achieved when the “*” overlaps with the “○”. In Fig. 2, 13 spectra of PU were misidentified as PMMA, five spectra of PP were misidentified as PA, one spectrum of PC was misidentified as ABS, and one spectrum of PE was misidentified as PC. The correct identification rates are all 100% except for PE, PU, PP and PC whose accuracies are 98%, 74%, 90%, and 98%, respectively.
The reason for the misidentification was as follows: Although abundant nitrogen, oxygen, and C2 bands exist in 11 kinds of polymers, the intensities of O I 777.41 nm, C-N (0, 0), C-C (0, 0), and N I 746.87 nm are much lower than that of metal spectral lines and H I 656.28 nm. The stronger the spectral line is, the larger role it plays in identification. Accordingly, only the metal spectral lines and H I 656.28 nm play major roles in the identification, while the spectral lines of O I 777.41 nm, C-N (0, 0), C-C (0, 0), and N I 746.87 nm have not played critical roles and just made a small contribution to the identification. Thus the PE, PU, PP and PC are difficult to be completely identified.
3.2 Polymer identification with ASW
As mentioned above, although the spectral lines of C-N (0, 0), C-C(0, 0), and O I 777.41 nm are essential for further identifying the PE, PU, PP and PC, the spectral line intensities of C-N (0, 0), C-C (0, 0), and O I 777.41nm are too weak to play significant responsible roles in the identification. To further identify PE, PU, PP and PC, and improve the identification accuracy, an approach to adjusting the spectral weightings (ASW) by increasing the classification weightings of O I 777.41 nm, C-N (0, 0), and C-C (0, 0) was proposed. The normalized intensities of O I 777.41 nm, C-N (0, 0), and C-C (0, 0) for all spectra were multiplied by weight factors 4.755, 45.2 and 5.96, respectively. These weight factors were obtained as follows: Each of normalized intensity showed in Fig. 3 is the average value of 550 spectra in the training set. The average normalized intensity of Mg II/C, Ti II/C, Ca II/C, Na I/C, K I/C. and H I/C is 2.222, which is 4.755, 45.2 and 5.96 times larger than the normalized intensities of O I 777.41 nm, C-N (0, 0), and C-C (0, 0), respectively.
After multiplying the normalized intensities of O I 777.41 nm, C-N (0, 0) and C-C (0, 0) for all spectra by three weighting factors, the weighting distribution of normalized characteristic spectral lines has been changed. Table 3 lists the classification weightings of normalized characteristic spectral lines with and without ASW, which was determined by principal component analysis (PCA) algorithm. Obviously, classification weightings of O I/C, C-N (0, 0)/C and C-C (0, 0)/C have been increased to the level of metal spectral lines and H I 656.28 nm. Table 3, the weighting of N I/C was also very small. It is not necessary to increase its weighting because nitrogen in laser-induced polymer plasma exists mainly in the form of C-N molecules rather than nitrogen ions .
A “new” training set was used to train the SVM model. A “new” test set was used to test the performance of the SVM model and to obtain identification accuracy. Figure 4 shows the identification results. The identification accuracy for PU, PP, PE, and PC are all improved to 100% after adjusting the spectral weightings (ASW) by increasing the weightings of O I 777.41 nm, C-N (0, 0), and C-C (0, 0).
To further prove the effect of ASW on identification accuracy improvement, the original 100 spectra for each polymer were marked as #1 to #100 and re-divided into training set and test set in different ways. Another 10 different training sets has been obtained (see Table 4). For each training and test sets, ASW was used in the same way to increase the normalized intensities of three key spectral lines. The identification accuracies were offered by SVM. Table 4 shows the comparison between identification accuracies with and without ASW. The identification accuracies for all polymers have been improved to nearly 100% with ASW.
In summary, a new approach to polymer identification by LIBS with ASW was developed in this work aiming at improving the identification accuracy. This approach was achieved through multiplying the normalized intensities of O I 777.41 nm, C-N (0, 0), and C-C (0, 0) for all spectra by weight factors of 4.755, 45.2, and 5.96, respectively. These factors were extracted from the training set. Through variable validation, the polymer identification accuracies have been improved to nearly 100% with ASW, which promises LIBS applications in polymer identification.
This research was financially supported by the National Special Fund for the Development of Major Research Equipment and Instruments (No. 2011YQ160017), and by the National Natural Science Foundation of China (No. 51128501).
References and links
1. V. Goodship, Introduction to Plastics Recycling (iSmithers Rapra, 2007).
2. E. A. Bruno, “Automated Sorting of Polymers for Recycling,” 2011, http://infohouse.p2ric.org/ref/09/08620.pdf.
3. M. Chanda and S. K. Roy, Plastics Technology Handbook (CRC, 2010).
4. G. R. Gunning, “Applications of ED-XRF technology to on-line analysis,” Adv. X-Ray Anal. 36, 105–109 (1993).
5. W. H. A. M. van den Broek, E. P. P. A. Derks, E. W. van de Ven, D. Wienke, P. Geladi, and L. M. C. Buydens, “Plastic identification by remote sensing spectroscopic NIR imaging using kernel partial least squares (KPLS),” Chemometr. Intell. Lab. 35, 187–197 (1996).
6. W. H. A. M. Van Den Broek, D. Wienke, W. J. Melssen, and L. M. C. Buydens, “Plastic material identification with spectroscopic near infrared imaging and artificial neural networks,” Anal. Chim. Acta 361(1-2), 161–176 (1998). [CrossRef]
7. T. Huth-Fehre, R. Feldhoff, T. Kantimm, L. Quick, F. Winter, K. Cammann, W. van den Broek, D. Wienke, W. Melssen, and L. Buydens, “NIR - Remote sensing and artificial neural networks for rapid identification of post consumer plastics,” J. Mol. Struct. 348, 143–146 (1995). [CrossRef]
8. P. Dinger, “Automatic sorting for mixed polymers,” BioCycle: Journal of Composting & Organics Recycling 33, 80–82 (1992).
9. N. Eisenreich and T. Rohe, “Infrared spectroscopy in analysis of polymers recycling,” Kunststoffe 2, 222–224 (1996).
10. D. A. Cremers and L. J. Radziemski, Handbook of Laser-Induced Breakdown Spectroscopy (John Wiley, 2006).
11. A. W. Miziolek, V. Palleschi, and I. Schechter, Laser-Induced Breakdown Spectroscopy: Fundamentals and Applications (Cambridge University, 2006).
12. X. N. He, W. Hu, C. M. Li, L. B. Guo, and Y. F. Lu, “Generation of high-temperature and low-density plasmas for improved spectral resolutions in laser-induced breakdown spectroscopy,” Opt. Express 19(11), 10997–11006 (2011). [CrossRef] [PubMed]
13. L. B. Guo, C. M. Li, W. Hu, Y. S. Zhou, B. Y. Zhang, Z. X. Cai, X. Y. Zeng, and Y. F. Lu, “Plasma confinement by hemispherical cavity in laser-induced breakdown spectroscopy,” Appl. Phys. Lett. 98(13), 131501 (2011). [CrossRef]
14. L. B. Guo, W. Hu, B. Y. Zhang, X. N. He, C. M. Li, Y. S. Zhou, Z. X. Cai, X. Y. Zeng, and Y. F. Lu, “Enhancement of optical emission from laser-induced plasmas by combined spatial and magnetic confinement,” Opt. Express 19(15), 14067–14075 (2011). [CrossRef] [PubMed]
15. D. A. Rusak, K. D. Weaver, and B. L. Taroli, “Laser-Induced Breakdown Spectroscopy for Analysis of Chemically Etched Polytetrafluoroethylene,” Appl. Spectrosc. 62(7), 773–777 (2008). [CrossRef] [PubMed]
16. I. Moench, R. Sattmann, and R. Noll, “High-speed identification of polymers by laser-induced breakdown spectroscopy,” Proc. SPIE 3100, 64–74 (1997). [CrossRef]
17. R. Sattmann, I. Monch, H. Krause, R. Noll, S. Couris, A. Hatziapostolou, A. Mavromanolakis, C. Fotakis, E. Larrauri, and R. Miguel, “Laser-Induced Breakdown Spectroscopy for Polymer Identification,” Appl. Spectrosc. 52(3), 456–461 (1998). [CrossRef]
18. M. Stepputat and R. Noll, “On-Line Detection of Heavy Metals and Brominated Flame Retardants in Technical Polymers with Laser-Induced Breakdown Spectrometry,” Appl. Opt. 42(30), 6210–6220 (2003). [CrossRef] [PubMed]
19. J. M. Anzano, I. B. Gornushkin, B. W. Smith, and J. D. Winefordner, “Laser-induced plasma spectroscopy for polymer identification,” Polym. Eng. Sci. 40(11), 2423–2429 (2000). [CrossRef]
20. M. Boueri, V. Motto-Ros, W. Q. Lei, Q. L. Ma, L. J. Zheng, H. P. Zeng, and J. Yu, “Identification of Polymer Materials Using Laser-Induced Breakdown Spectroscopy Combined with Artificial Neural Networks,” Appl. Spectrosc. 65(3), 307–314 (2011). [CrossRef] [PubMed]
21. S. Grégoire, M. Boudinet, F. Pelascini, F. Surma, V. Detalle, and Y. Holl, “Laser-induced breakdown spectroscopy for polymer identification,” Anal. Bioanal. Chem. 400(10), 3331–3340 (2011). [CrossRef] [PubMed]
22. B. E. Boser, I. Guyon, and V. Vapnik, “A training algorithm for optimal margin classiers,” in Proceedings of the Fifth Annual Workshop on Computational Learning Theory (ACM, 1992), pp. 144–152. [CrossRef]
23. C. Cortes and V. Vapnik, “Support-vector network,” Mach. Learn. 20(3), 273–297 (1995). [CrossRef]
24. C. H. Wu, G. H. Tzeng, Y. J. Goo, and W. C. Fang, “A real-valued genetic algorithm to optimize the parameters of support vector machine for predicting bankruptcy,” Expert Syst. Appl. 32(2), 397–408 (2007). [CrossRef]
26. M. Boueri, 2010 Ph.D. Dissertation (Lyon:Universite de Lyon) (in France).
27. D. Filmer and L. H. Pritchett, “Estimating Wealth Effects Without Expenditure Data--Or Tears: An Application to Educational Enrollments In States Of India,” Demography 38(1), 115–132 (2001). [PubMed]
28. V. Motto-Ros, A. S. Koujelev, G. R. Osinski, and A. E. Dudelzak, “Quantitative multi-elemental laser-induced breakdown spectroscopy using artificial neural networks,” Journal of the European Optical Society-Rapid Publications 3, 08011 (2008).
29. Y. Yu, Z.-Q. Hao, C.-M. Li, L.-B. Guo, K.-H. Li, Q.-D. Zeng, X.-Y. Li, Z. Ren, and X.-Y. Zeng, “Identification of plastics by laser-induced breakdown spectroscopy combined with support vector machine algorithm,” Acta Phys. Sin. 62, 215201 (2013).