Automated sorting of polymer flakes: fluorescence labeling and development of a measurement system prototype.

Waste Management xxx (2015) xxx–xxx

Contents lists available at ScienceDirect

Waste Management journal homepage: www.elsevier.com/locate/wasman

Review

Automated sorting of polymer flakes: Fluorescence labeling and development of a measurement system prototype S. Brunner ⇑, P. Fomin, Ch. Kargel Institute for Measurement and Automation, Division of Sensor Technology and Measurement Systems, Bundeswehr University Munich, Werner-Heisenberg-Weg 39, 85579 Neubiberg, Germany

a r t i c l e

i n f o

Article history: Received 30 July 2014 Accepted 2 December 2014 Available online xxxx Keywords: Automated plastic sorting Tracer-based sorting Fluorescence marker Fluorescence spectroscopy Multi-class classification

a b s t r a c t The extensive demand and use of plastics in modern life is associated with a significant economical impact and a serious ecological footprint. The production of plastics involves a high energy consumption and CO2 emission as well as the large need for (limited) fossil resources. Due to the high durability of plastics, large amounts of plastic garbage is mounting in overflowing landfills (plus 9.6 million tons in Europe in the year 2012) and plastic debris is floating in the world oceans or waste-to-energy combustion releases even more CO2 plus toxic substances (dioxins, heavy metals) to the atmosphere. The recycling of plastic products after their life cycle can obviously contribute a great deal to the reduction of the environmental and economical impacts. In order to produce high-quality recycling products, mono-fractional compositions of waste polymers are required. However, existing measurement technologies such as near infrared spectroscopy show limitations in the sorting of complex mixtures and different grades of polymers, especially when black plastics are involved. More recently invented technologies based on mid-infrared, Raman spectroscopy or laser-aided spectroscopy are still under development and expected to be rather expensive. A promising approach to put high sorting purities into practice is to label plastic resins with unique combinations of fluorescence markers (tracers). These are incorporated into virgin resins during the manufacturing process at the ppm (or sub ppm) concentration level, just large enough that the fluorescence emissions can be detected with sensitive instrumentation but neither affect the visual appearance nor the mechanical properties of the polymers. In this paper we present the prototype of a measurement and classification system that identifies polymer flakes (mill material of a few millimeters size) located on a conveyor belt in real time based on the emitted fluorescence of incorporated markers. Classification performance and throughput were experimentally quantified using 3 different types of polymers (Polyoxymethylene (POM), Polybutylenterephthalat (PBT) and Acrylonitrile Styrene Acrylate (ASA)) in colored and uncolored form. Overall, 12 classes of plastic flakes were investigated in this study, where 11 classes were labeled with unique binary combinations of 4 fluorescence markers and class 12 includes unlabeled plastic flakes of various colors. From approx. 68,000 investigated flakes it was found that the developed measurement prototype system achieves an average sensitivity (true positive rate) of 99.4% and a precision (positive predictive value) of 99.5%, while being able to handle up to approx. 1800 flakes per second. Ó 2014 Elsevier Ltd. All rights reserved.

Contents 1. 2.

3.

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . State-of-the-art in plastic classification and sorting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1. Physico-chemical and spectroscopic sorting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2. Tracer-based sorting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Plastic identification and classification using fluorescence markers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

00 00 00 00 00

⇑ Corresponding author. Tel.: +49 (0)89 6004 6033. E-mail addresses: [email protected] (S. Brunner), christian.kargel@ unibw.de (Ch. Kargel). http://dx.doi.org/10.1016/j.wasman.2014.12.006 0956-053X/Ó 2014 Elsevier Ltd. All rights reserved.

Please cite this article in press as: Brunner, S., et al. Automated sorting of polymer flakes: Fluorescence labeling and development of a measurement system prototype. Waste Management (2015), http://dx.doi.org/10.1016/j.wasman.2014.12.006

2

S. Brunner et al. / Waste Management xxx (2015) xxx–xxx

4. 5.

6.

7. 8.

Appropriate fluorescence markers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Development of an in-line measurement and classification prototype system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1. Demands and main technical specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2. System design and measurement hardware concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3. Software concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Experimental performance evaluation of the measurement system prototype . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1. Selected polymers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2. Assignment of markers to polymers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3. Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4. Classification performance metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Experimental design and experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1. Introduction Plastic has become the most common material and hardly any other material influences our daily life as much as plastics do. It has been used in almost every area such as packaging, construction, automotive, electronics, and medicine which results in a steadily growing worldwide demand of approx. 288 million tons in the year 2012 (PlasticsEurope, 2013). This extensive demand of polymers in combination with the high durability after the rather short product life phases causes significant environmental and economical impacts because the production of plastics is accompanied by a high energy consumption, a significant discharge of CO2 and it requires a large amount of the fossil resources like crude oil. In order to preserve the world’s limited natural resources and reduce the massive flow of plastic garbage to landfills and plastic debris floating in the world oceans (including micro-plastic entering the food chain), polymer recycling has become a high priority from an environmental, economical and legislative point of view. The numbers are quite impressive: just one ton of recycled plastic can save up to 2,604 liters of crude oil, reduce the energy consumption by 80–90% compared to the production of virgin plastics and help to avoid 22 cubic meters of landfill (BIR, 2014). The potential of polymer recycling is thus enormous. In Europe only 62% (15.6 million tons) of plastics were recovered (Plastic Europe, 2013) in the year 2012 while leaving 9.6 million tons for disposal. Most of the plastic waste still occurs in the field of packaging but the electronics sector and the automotive industry are growing producers of post-consumer plastics waste with rather poor recycling rates of approx. 10% (Plastic Europe, 2008). There is an obvious and urgent need to take a focused and strategic approach towards plastic product and waste management. For example, in order to reduce the many problems with end-oflife products and improve the management of raw resources, the European parliament set up two directives (2000/53/EC and 2002/96/EC) which force manufactures to increase the material recycling rate for end-of-life vehicles to 85% till the year 2015 and for waste electrical and electronic equipment to 80% compared to 2007. In order to assure a high economical value and an adequate quality of components produced from recycled plastics, a high purity of sorted polymer fractions is mandatory. However, it is difficult in practice to fulfill this requirement since waste polymer streams typically consist of a complex mixture of various plastic types which also may include different kinds of dyes, fillers and additives. Impurities caused by unreliable sorting may lead to phase separation and as a result to structural weaknesses in recycled materials. Moreover, state-of-the-art measurement systems fail to identify black and dark-colored polymers which is problematic especially for the automotive and electronics sector since a large

00 00 00 00 00 00 00 00 00 00 00 00 00 00

fraction of polymers used there are black. Mid-infrared (MIR) and laser spectroscopy systems are under development and include rather costly hardware components. A promising way to tackle the existing difficulties in polymer sorting is to make use of a labeling approach based on unique fluorescence dyes (in this article also termed ‘‘markers’’). These are incorporated into virgin polymer resins during the manufacturing process at the (sub) ppm level and provide specific fluorescence signatures (optical spectra) that allow the recognition (classification) of different plastics. In this article we present the developed measurement system prototype able to identify in real time to which a set of categories individual plastic flakes belong. This prototype system was developed within a cooperation project of several partners. Singularization and sorting units were developed by an industrial partner. Polymers were provided by an industrial partner as well. The applied fluorescence markers were produced by an academic research center for organic dyes. We developed the measurement system and signal processing for classification purposes. The classification performance and throughput were experimentally quantified using 12 different classes of plastic flakes (see Fig. 1) of the polymers POM, PBT and ASA in colored and uncolored form. 11 classes were labeled with unique binary code combinations of 4 fluorescent makers and one class did not include any markers at all. This paper is structured as follows: Section 2 briefly reviews existing automated plastic sorting techniques including their advantages and drawbacks. Section 3 explains the basic idea behind the labeling approach and the use of fluorescence markers for highly reliable plastic sorting. Section 4 focuses on the properties and selection of appropriate fluorescence markers. Section 5 presents the developed prototype system including the hardware and software concepts for the in-line measurement of fluorescence spectra emitted from small plastic flakes carried by a conveyor belt. In Section 6 the experimental setup used to evaluate the classification performance is shown before the achieved results and a brief outlook to future investigations are presented in Section 7. 2. State-of-the-art in plastic classification and sorting The economical (and environmental) success of recycling polymer materials at the end of their life cycle highly depends on the reliable sorting into different types and grades in an affordable and fast way such that the value of the recycled product exceeds the incurred costs (Niaounakis, 2013; Cornell, 2007). At present, the majority of plastic sorting is still carried out by hand (Niaounakis, 2013) which is slow, very labor intensive and less efficient even if a labeling system such as the resin identification code is applied (Bruno, 2000). Sorting rates of 50 kg/hour until 100 kg/ hour are reasonable with a purity of 95% (Pascoe, 2000).


3


Class 1

Class 2

Class 3

Class 4

Class 5

Class 6

Class 7

Class 8

Class 9

Class 10

Class 11

Class 12

Fig. 1. Photographs of the 12 classes of plastic flakes used to evaluate the measurement system prototype. Flakes which belong to the classes 1–11 were labeled with unique binary code combinations of 4 fluorescent makers, flakes of class 12 (mixture of differently colored polymer flakes) did not include any markers at all. Table 1 Overview of physio-chemical and spectroscopic sorting technologies. Identification technology

Macro-/ micro-sorting

Application

Drawbacks

Development level

X-ray (Pascoe (2000)) Densimetric (Niaounakis (2013)) Magnetic density separation (Hu et al. (2011)) Triboelectric (Niaounakis (2013))) Optical (Riise et al. (2001)) NIR-spectroscopy (Serranti et al. (2013)) MIR-spectroscopy (BP-Sorting (2013)) Raman spectroscopy (Tsuchida et al. (2009)) Atomic emission spectroscopy (Krieg et al. (1998))

U/s s/U

Beside PVC/PET no other polymers are identified No sorting of complex mixtures; densities must significantly differ from one other Requires non-overlapping mass densities

Industrial Industrial

s/U

Identification of PVC Separation of lighter polyolefins (HDPE) from heavier non-olefins Separation of a mixed flow of polymers

s/U

Low cost sorting

Industrial

U/U U/U

Sorting based on color Fast identification of bottles

No sorting of complex mixtures; particles must be clean and dry Does not recognize the polymer type Not suitable for dark and wet polymers

U/s

Sorting of dark/black polymers

Pilot

U/U

Identification of shredded plastics

Limited measurement speed and optical resolution; expensive hardware Limitations in sorting black polymers

s/U

High speed sorting; identification of black polymers

Problems to identify PP (Polypropylene) and starchbased plastics

Industrial

In order to improve the economical aspect as well as the speed and reliability of plastic recycling, automated sorting systems are used that sort end-of-life polymer products either as larger objects as for instance bottles and containers (macro-sorting) (Bruno, 2000), or in the form of shredded plastic flakes with a few millimeters size (micro-sorting). Automated classification and sorting systems may rely on methods which either make use of the intrinsic physical/chemical/optical properties of polymers or add unique identification features to the polymers during the manufacturing process (tracer-based sorting). 2.1. Physico-chemical and spectroscopic sorting These sorting approaches comprise technologies which make use of intrinsic polymer properties such as density, triboelectric charging behavior or color, and employ optical inspection methods such as Visible-, NIR- (Near infrared), Mid-infrared-, X-ray- or Raman-fluorescence spectroscopy to identify various types of polymers. Table 1 summarizes the most common sorting technologies including their application areas, limitations and level of development. Some brief comments are given in the following.

Industrial

Industrial Industrial

Pilot

X-ray fluorescence technologies are only applicable to separate PVC (Polyvinyl chloride) from PET and may involve higher system complexity and also health risks. Densimetric sorting is a rather low cost and slow method that works efficiently when density differences are clearly pronounced. Triboelectric sorting does not allow the sorting of complex mixtures of polymers and is highly sensitive to surface conditions of particles which must be clean and dry (Niaounakis, 2013). Based on the triboelectric sorting technology the Hamos GmbH invented a commercial system which is capable of sorting up to 750 kg/h (Niaounakis, 2013). Another recently invented sorting technology is based on magnetic density separation whereby mixed flakes of plastic residues are filled into a magnetic liquid which flows through a magnetic field in order to separate flakes based on their mass densities (Luciani et al., 2014). With this sorting method both a sorting purity and accuracy of at least 90% can be achieved (Hu et al., 2011). Optical systems which are based on color imaging (VIS) sensors are usually applied to sort out colored impurities (Riise et al., 2001) but cannot identify the polymer type or grade. In the last few years spectroscopy-based methods in the VNIR (400–1000 nm) (Serranti et al., 2011; Luciani et al., 2013) and NIR (1000–2500 nm) (Serranti


4


et al. (2012, 2013), Ulrici et al., 2013) wavelength range were developed in order to gain a closer look into the chemical composition of different polymer components (Tatzer et al., 2005): polymer flakes to be sorted are exposed to an electromagnetic source with an appropriate spectrum and energy. Characteristic spectra are reflected from the flakes due to the specific absorption properties of different polymers. These optical spectra are acquired and then compared with known reference spectra for identification purposes. A key weakness of this approach arises from the multitude of different dyes, fillers and additives used today which result in a variety of different spectra for the same polymer type and grade. Moreover, NIR spectroscopy cannot be used to identify black and dark-colored plastics since colorants such as the commonly used ‘‘carbon black’’ tend to absorb most of the incoming NIR radiation leading to extremely low reflection intensities and/or featureless spectra. Despite these problems and because of the lack of better solutions, dozens of commercial systems based on NIR spectroscopy are available on the market. For instance, Titech’s PolySortÒ system provides – according to (Nature Works, 2009) – a throughput of up to 5 tons per hour (TITECH, 2014) for flakes with a couple of millimeters size while reaching a sorting purity of 97.5%. According to the manufacturer Steinert GmbH (STEINERT, 2014), the UniSortÒ flake sorting system achieves a throughput of up to 2 tons per hour while providing a sorting purity of at least 90%. Similar to NIR spectroscopy, Raman spectroscopy can be applied to identify polymers based on their molecular structure (Tsuchida et al., 2009). It makes use of the inelastic (Raman) scattering of monochromatic light from a laser source in the visible, NIR, or near ultraviolet range (Merrington, 2011). As an advantage over conventional NIR spectroscopy, H2O and CO2 in the air or on the sample surface cause less negative effects and Raman spectra of polymers exhibit narrow-band peaks (Tsuchida et al., 2009). A prototype system was developed by (Tsuchida et al., 2009) which achieved an accuracy of up to 94%. In order to tackle the difficulties with black plastics, a recent study (BP-Sorting, 2013) focused on the development of an industrial polymer sorting system based on MIR spectroscopy (3000–5000 nm) using a MWIR hyperspectral imaging camera. There are some technical issues that need careful attention for commercial utilization, such as limited identification speed, lower optical resolution (compared to NIR) and low signal-to-noise ratio mainly due to localized melting of polymer materials under the MIR radiation source (Dvorak et al., 2011). Hardware costs are expected to be rather high compared to NIR or VIS spectroscopic systems. According to the report (BP-Sorting, 2013), a prototype system from Steinert GmbH has recently been invented which is supposed to sort 3 tons per hour of plastic flakes with a grain size > 30 mm. Detailed information about the achievable sorting purity could not be found on the project website (date: July 15, 2014). Another approach that has received considerable interest for plastic sorting is based on atomic emission spectroscopy (AES) where the incoming laser light is dispersed by excited molecules. This results in characteristic spectra in the ultraviolet (UV), VIS and NIR range to be acquired and used for identification purposes (Krieg et al., 1998). The Unisensor GmbH developed a granulate sorting system denoted as Powersort 200Ò which achieves a throughput of up to 2 tons/h while measuring up to 1 million spectra/s and achieving a sorting purity of 99.7% (Dvorak et al., 2011). While black samples of PET (Polyethylene terephthalate) and PS (Polystyrene) yield characteristic spectra, PP (Polypropylene) and starch-based plastics are rather difficult to detect (Dvorak et al., 2011). Apart from that, system complexity and costs are expected to be rather high compared with NIR spectroscopy systems.

2.2. Tracer-based sorting Tracer-based sorting approaches address the identification problems associated with dark-colored polymers by adding an additional property that can be measured. The most relevant studies are briefly discussed: In 1992 Bayer disclosed an idea for the identification of polymers by adding small quantities of fluorescent dyes into the virgin plastic resins during the manufacturing process (Luttermann et al., 1992, 1996; Meier et al., 1997). In particular, dyes should be used that significantly differ in terms of their fluorescence emission wavelengths and lifetime (fluorescence decay). Complexes based on rare earth elements such as europium (Eu) or terbium (Tb) are to be preferred due to their long lasting fluorescence. With this approach different plastics can be identified by measuring the emitted fluorescence emission wavelengths and fluorescence lifetime of dyes (or we call ‘‘markers’’) incorporated in the plastics. This approach is currently applied in the biochemical sector (Maris et al., 2012) but has not been put into practice in an industrial sorting system so far due to various problems and formidable challenges (see Sections 3–5). An enterprise supported by the European commission investigated the application of organic and inorganic fluorescence markers for identification and sorting of bottles in the packaging industry (Ahmad, 2004). A prototype measurement system involving four wavelength-selective optical signal detection units was developed and evaluated. Performance tests were carried out using different plastic materials and bottles (without any colorant) including fluorescence markers with concentrations ranging from 0.5 ppm to 20 ppm that emit fluorescence light in the visible band after UV excitation (Simmons et al., 1998; Ahmad, 2004). As a conclusion of this study it was found that multi-layered containers such as milk cartons can be economically viable identified and sorted but black pigments in hosting plastics prevent proper identification. Identification and sorting velocity was found to be limited by singularization and blow ejection performance irregularities (Ahmad, 2004). This system is limited to macro-sorting only and cannot be easily extended to the application of more than 4 markers. In practice, the differentiation of markers that partially overlap with one another and with inherent fluorescence emissions from polymers or additives seems to be impossible with the chosen signal processing approach of simply thresholding the signal intensities. In 2011, another study focused on the appropriate choice of inorganic UV fluorescence tracers and their interaction with certain polymers including additives (Maris et al., 2012). Fluorescence markers in dark-colored polypropylene could be detected at minimum concentrations between 25 ppm and 100 ppm. Mechanical properties of polymers (ABS – acrylonitrile butadiene styrene, PP) were not influenced for concentrations lower than 250 ppm. Measurements were accomplished using a laboratory setup which includes a Xenon lamp with optical filter, a monochromator and a CCD camera. The monochromator represents the bottle neck of this system in terms of measuring rate since it scans through the entire fluorescence spectrum by mechanically rotating a diffraction grating. Hence this system setup cannot provide the high measurement rates required in industry to achieve a reasonable throughput. It might be interesting to note that in 2009 attention was put on the applicability of X-ray fluorescence markers based on rare earth oxides to identify and sort different types of polymers (Bezati et al. (2010, 2011)). Although X-ray tracers provide the opportunity to identify black and dark-colored polymers mostly independent of the object’s surface, problems arise in practice from the rather long measurement durations and considerable marker concentrations


5


Fluorescence emission

0.5

Collection of end-of-life products

4

Type

0

1

A

0

1

0

B

0

0

1

1

C

0

1

0

0

D

0

1

0

1

E

Classification 0

1

1

0

F

0

1

1

1

G

1

0

0

0

H

1

0

0

1

I

1

0

1

0

J

1

0

1

1

K

1

1

0

0

L

1

1

0

1

M

1

1

1

0

N

1

1

1

1

O

500 600 700 800

+ Relative Intensit

3

0

0

Wavelength in nm

Mill

Plastic flake with marker 1 and marker 3

2

0

Measurement Marker 1 system

1

0

Binary coding system 1

1

Marker 3

0,5

Classification of shredded flakes

0

1 0.5 0

500 600 700 800

Wavelength in nm

Measured sum spectrum 500 600 700 800

Wavelength in nm

Marker

Plasc

Sorting unit

Container with sorted flakes

…

Type O

M1 M2 M3 M4

Production of plastic products

Automated sorting for recycling

Identification/classification Excitation light

Type A

Collection and shredding

Relative Intensity

Life phase

Incorporation of fluorescent markers

Relative Intensity

Resin manufacturing

Sorting of shredded plastic waste

Fig. 2. Concept of plastic recycling based upon the automated sorting of plastic waste labeled by fluorescent markers (here M1–M4) incorporated into virgin resins.

ranging from 100 ppm till 1000 ppm required for the reliable identification of polymers. Apart from optical tracers, magnetic markers could be used in principle to label polymers (Mankosa and Luttrell, 2005). Their main advantage is that they are quite insensitive to additives and colorants incorporated in plastic resins. However, at present the application of more than a single magnetic marker seems not to be possible and those markers require rather high concentrations in the plastics in the range of some percent. A restricted use of such ‘‘magnetic plastics’’ might be necessary, e.g. in electronic devices. Given those findings and experiences, the application of fluorescence markers has a significant potential for the highly reliable and economically attractive sorting of plastics including black and dark-colored polymers. It is the main goal of this article to demonstrate how the obstacles and problems of the past that prevented the practical use of such plastic sorting systems can be overcome. The measurement system prototype we present can be applied to micro- and macro-sorting and provides a high sensitivity in the wavelength range of approx. 450–1000 nm, allowing low marker concentrations even at short integration and calculation times (ms) and the application of up to about 10 markers with different center wavelengths. Approximately 300,000 spectra per second can be acquired which results in measurement rates large enough for the classification of a few thousand small plastic flakes per second delivered on a conveyor belt. 3. Plastic identification and classification using fluorescence markers Fig. 2 depicts the basic concept of polymer tracing based on fluorescence markers. In order to establish an optical tracing system, small amounts of fluorophores (fluorescence markers) with unique optical characteristics are incorporated into the material of interest during the compounding step of the resin manufacturing process. Marker concentrations need to be low (at the parts per million (ppm) or even sub-ppm level) for two reasons: (a) the products made from the labeled polymers must neither influence the visual appearance nor the mechanical properties or structural integrity; (b) the costs of the polymers should not be increased much due to the adding of the markers. The lowest marker concentration is essentially defined by the ability of the measurement system to acquire fluorescence emission spectra with a signal-to-noise ratio (SNR) large enough to reliably perform the classification task. The necessary SNR depends upon the desired sorting quality, the marker concentration, the excitation light intensity and the measuring rate, (governed by the sensor integra-

tion time). In the case of black plastics which absorb a large part of the fluorescence light emitted from the markers, higher marker concentrations of up to approx. 100 ppm might be needed. During the life phase of labeled plastic products, fluorescence markers must resist environmental influences as for instance higher temperature and UV radiation. For recycling purposes these plastic products get disassembled and subsequently shredded in dismantling facilities after their life cycle. The plastic fraction first needs to be separated from fractions of other materials like metals, etc. into a stream of a complex mixture of various polymers. Fluorescently labeled polymers can then be identified by acquiring their unique fluorescence emission spectra (optical signatures) and applying sophisticated signal processing and classification algorithms (see identification phase in Fig. 2). Classification results (i.e. the plastic types) are forwarded to a sorting unit which separates the identified polymer flakes accordingly. An important question is how many different fluorescent markers would be needed. The answer clearly depends on the properties of the markers and on the coding scheme to be applied. In order to explain the basic concept, let us assume that a certain number of different markers is available and a simple binary coding scheme is applied. For instance, when combining just n = 4 fluorescence markers with different optical signatures in a binary fashion, up to 24 1 = 15 different polymer types can be labeled (the code ‘‘0000’’ is reserved for unlabeled material, i.e. with no markers included). Each bit of the 4-bit code is defined by the presence (‘‘1’’) or absence (‘‘0’’) of a particular marker. Depending mostly on the optical properties of the fluorescent markers, other coding schemes that increase the number of identifiable plastics might be applicable (Brunner and Kargel, 2011). 4. Appropriate fluorescence markers Tracer-based sorting is only possible with appropriate fluorescence markers which after being homogeneously incorporated into the polymer still emit fluorescence light at the end of the life phase of the plastic product. Fluorescence markers suitable for the plastic sorting application must fulfill several economical, environmental and technical aspects (technical aspects are mainly related to the measurement system and the classification algorithms applied): Markers incorporated into plastics must neither affect the visual appearance nor the mechanical and aging properties of the polymer. Markers should be non-toxic and compatible with all components of the polymer (Maris et al., 2012).


6


Markers must resist environmental influences (UV light, temperature, mechanical stress, etc.) during the life cycle of the hosting plastics. Markers should emit fluorescence light somewhere in the wavelength range between approx. 450 nm and 1000 nm. For this region, adequate and cost effective imaging sensors are available and fluorescence from most plastic additives is not significant. Marker emission spectra should be rather narrow in order to maximize the number of possible markers in a given wavelength range and also increase the classification performance. The spectral signatures emitted from different markers should ideally be orthogonal such that the best possible classification results can be achieved. One must keep in mind though that the marker design is governed by chemical and physical limitations. The intensity of the fluorescence light emissions should be maximized while at the same time the marker concentrations must be as low as possible, not exceeding the sub-ppm or ppm level for economical reasons. As one consequence, the quantum yield (i.e. the probability that an absorbed photon causes the emission of a photon) of the markers should be high and photo quenching also should be avoided (Ahmad, 2004). Markers should resist photobleaching, even when they get exposed to intensive ambient daylight or technical light sources. Photobleaching reduces the fluorescence intensity emitted from the plastics because incorporated marker fluorophores are destructed by a photochemical reaction. Appropriate fluorescence markers must guarantee a fluorescence intensity sufficiently high for classification purposes at the end of the life cycle of the plastic product. It is necessary to note here, that fluorophores highly stable in typical solutions (e.g. chloroform) might show pronounced photobleaching after being incorporated in polymers. The marker concentrations should be adjusted such that the corresponding fluorescence intensities do not differ too much from one another. This significantly improves the classification and smaller error rates can be achieved (Brunner et al., 2012). Overlap of the emission spectra of individual fluorescence markers should be as low as possible and marker emission should be disjunct with intrinsic polymer emission. Marker absorption spectra are preferred to be broad in order to absorb most of the incident light from the excitation source. Alternatively, all markers should absorb the excitation light in the same wavelength band and below approx. 460 nm. This promotes the cost-effective excitation of all fluorophores with the same light source and optical filter arrangement. The markers need to absorb in a wavelength range where commercial illumination sources, especially high power LEDs and laser diodes, are available. Markers need to be incorporated homogeneously into the virgin polymers which is challenging to achieve and needs specific actions during the incorporation process. 5. Development of an in-line measurement and classification prototype system The technical specifications of the measurement prototype system were defined with our partners taking into account the aforementioned limitations of the fluorescence markers and the parameters of the sorting machinery. 5.1. Demands and main technical specifications In many applications – as for our industrial partner – microsorting of waste plastics is required because the plastic parts to

be recycled must be shredded into small pieces in order to process high volumes of material and reduce handling costs: The prototype must be able to process plastic flakes with physical dimensions d between 3 mm and 10 mm. These flake dimensions were derived from flake size histograms of plastic material shredded by typical mills and evaluated using morphological image processing (Gonzalez and Woods, 2008; Jähne, 2005). We found that for all mills used by our industrial partner so far the flake size histograms are symmetric around their mean value with a shape very similar to a Gaussian function. The lower limit of 3 mm is implemented by using a screen for mechanical separation and was chosen because small flakes are increasingly difficult to handle during the sorting process and contribute only little to the (mass) throughput. A throughput of approx. 250 kg of plastic flakes per hour should be achieved with the prototype system. Based on our regression analysis of mill material provided by the industrial partner, 1500 flakes per second need to be handled to achieve this throughput. 15 different types of colored and non-colored polymers need to be sorted with the prototype with an easy extension of the approach to larger numbers. This rather low limit was mostly due to the small number (4) of fluorescence markers available at the project start and due to the costs of the sorting machinery to be build. Our next goal is to apply 10 appropriate markers such that more than 1000 different resins can be uniquely labeled. The spectral imaging sensor should provide a high sensitivity in the wavelength range (450–1000 nm) relevant for appropriate markers while at the same time being inexpensive compared to e.g. intensified systems or cooled detectors. The higher the sensor sensitivity is, the lower the concentration of the markers in plastics (and their costs) can be. 5.2. System design and measurement hardware concept The required measurement rate of 1500 flakes per second or higher results in a low integration time and suggests some kind of parallelization with regard to the overall system design. Furthermore, the commitment to rather cost-effective measurement hardware, especially the spectral imaging sensor, in combination with the short lifetime (a few nanoseconds) of the available fluorescence markers prevents the realization of time-resolved measurement methods. Fig. 3 illustrates the system design and measurement hardware concept put into practice. In order to investigate it in practice, a prototype was constructed and assembled (see Fig. 4). The entire system consists of 3 functional groups: singularization unit, measurement system and sorting unit. Singularization and sorting units were constructed by our industrial partners. The singularization unit receives a mixture of fluorescently labeled and unlabeled flakes of different polymers from a container and generates parallel streams of spatially separated plastic flakes which are forwarded on the conveyor belt to the measurement cell. For initial test purposes a small conveyor belt with a width of 200 mm and a low number of n = 18 parallel streams was used. The belt material and color were accordingly selected. In order to prevent damage of the sensitive optical components, the singularization unit is mechanically decoupled from the conveyor belt and the measurement system. The measurement system is sheltered by a closed encasement which prevents unwanted disturbances and influences from outside e.g. by surrounding stray light. The encasement is painted in black for maximal absorption and a light-absorbing black plate separates the morphology from spectral acquisition system. While


7


Container with a mixture of plasc flakes

Morphology acquision system

Spectral acquision system Frame grabber

Illuminaon unit controller

Synchronizaon unit

Personal computer

Fluorescence excitaon light with opcal filter

Sorng unit SPC

Singularizaon unit Illuminaon source morphology

Conveyor belt

Lens with emission filter

Rotary encoder

Drive motor

Containers with sorted plasc flakes Fig. 3. System design and hardware concept. More detailed information is given in the text.

the morphology system is relatively simple (a white LED line and a fast line-scan camera), the spectral acquisition system is a bit more extensive. A focused high-power blue LED line light with an irradiance of approx. 5000 W/m2 is applied in order to stimulate sufficient fluorescence intensity and thus increase the signal-to-noise ratio of the acquired spectra because fluorescence intensities emitted from the plastic flakes are very weak mostly due to the intentionally low marker concentrations. A customized short-pass optical filter with appropriate optical density, cut-off wavelength and roll-off is mounted in front of this LED line light. The filter is designed to transmit only those optical bands which are absorbed by the fluorophore molecules of the markers. Fluorescence light emitted from the plastic flakes passes an optical emission filter before it gets collected by a lens and enters the imaging spectrograph. The emission filter with appropriate optical density, cuton wavelength and roll-off only transmits optical bands emitted from the fluorophore molecules while the excitation wavelengths are blocked. The spectral acquisition system line-scans the lowlevel fluorescence light emitted from the parallelized flake streams across the conveyor belt. The imaging spectrograph in combination with a fast and highly sensitive industrial CCD camera allows to acquire approx. 2000 spectral channels of the fluorescence emission simultaneously from all flakes on the conveyor belt. Fluorescence light from the scan line on the conveyor belt passes through the entrance slit of the spectrograph, gets collimated and is subsequently dispersed by a prism-grating-prism component into a continuous spectral distribution perpendicular to the captured line. The resulting 2D image (spatial versus spectral dimension) is projected onto the CCD imaging sensor where it is converted into a digital image. Thus one dimension of the CCD sensor represents a spatial line image while the second dimension holds the spectra for every line spatial pixel. This spectral measurement principle has significant advantages in terms of acquisition speed compared to systems based on monochromators or tunable-filters and additionally is less expensive compared to a solution based on n single point spectrometers. This spectral acquisition system based on prism-grating-prism spectrographs and a CCD camera can achieve acquisition rates of up to approx. 300,000 spectra per second with a spectral sampling of approximately 4–5 nm (with spectral binning activated). After the data acquisition, morphological information (flake size and location) and functional information (type/class of polymer) is derived for all flakes on the conveyor belt from the acquired data

by means of signal and image processing as well as classification algorithms and then handed over to the control unit (SPC) of the sorting unit which activates and deactivates the sorting valves to separate and transport the flakes into the right containers. In order to assure accurate timing, both data acquisition systems and the other units are synchronized using the pulse signal provided by a high-resolution rotary encoder mounted on the conveyor belt. All hardware components including cameras, singularization and sorting unit, conveyor belt, LED light controllers and rotary encoder are controlled by a personal computer (Intel i7, 4 physical cores each of which including 2 virtual cores, Windows 7 operating system). On-line processing and analysis of spectral and morphological data are performed in parallel on the same PC. For the maximal spectral acquisition rate of 300,000 spectra per second, one spectrum needs to be analyzed and classified within 27 ls assuming that 8 spectra can be processed in parallel using the 8 available cores. 5.3. Software concept The application software for the prototype including signal and image processing as well as classification was developed in LabVIEWÓ (National Instruments) using the NI-Vision Development Module, the Multicore Analysis and Sparse Matrix Toolkit. It is highly parallelized in order to avoid bottle necks due the multitude of tasks that need to be carried out simultaneously and in order to optimize the memory utilization of the PC and the CPU load. Fig. 5 depicts the implemented modules of the developed software concept: All hardware components including both data acquisition systems, the conveyor belt drive and the sorting unit are synchronized using the hardware synchronization module. This module also includes routines to control the singularization unit and the speed of the conveyor belt. The spectral acquisition module allows to control and adjust all parameters of the spectral acquisition system and grab spectral images in real time. Additional routines to perform the necessary spatial calibration and shading correction are also included. Spectral images from the spectral acquisition module are transferred to the spectral classification module which carries out multiple spectral processing and analyzing steps in order to assign each individual plastic flake to the right class. The results are forwarded to the result database module.


8


Measurement system

Data acquisition system

Container with a mixture of plastic flakes Sync. Unit

+

Illumination units

Sorting unit Rotary encoder Singularization unit

Drive motor Conveyor belt Containers with sorted plastics

Fig. 4. Prototype system for the sorting of fluorescently labeled polymer flakes delivered on a conveyor belt: a singularization unit generates a parallel stream of plastic flakes from an input container comprising a mixture of plastic flakes. The encasing of the measurement system (more clearly depicted as an insert) is opened for demonstration purposes. Flakes are separated into different containers by the sorting unit based on the functional and morphological information determined by the measurement system. More detailed information about the measurement system is provided in the text.

Spectral images acquired by the spectral acquisition module are streamed to the hard disk using the spectral image storage module. In particular, for every scan line (1D) on the conveyor belt the corresponding spectral image (spatial spectral) is stored. If the data are not needed for e.g. optimization and off-line processing later on, this feature can be turned off. The morphology acquisition module includes software routines to setup the parameters of the line-scan camera and allow to grab 2D images line-by-line. In order to maximize the processing speed, the acquired images are stored in a circular buffer and segmented into binary images (flakes and background). An additional routine to perform the necessary spatial calibration is also included. Segmented images are forwarded to the morphological analysis module which performs several preprocessing steps (color segmentation, particle filtering using morphological erosion, morphological closing) (Gonzalez and Woods, 2008; Jähne, 2005) and extracts information (center of gravity and the bounding box) for all flakes using morphological particle analysis. The flake parameters are sent to the result database module. In order to stream images from the line-scan camera to the hard disk, a morphology image storage module was implemented similarly as in the spectral image storage module. If the data are not needed for e.g. optimization and off-line processing later on, this feature can be turned off. Inside the result database module, morphological and functional information (i.e. the classes the flakes belong to) are separately collected, then fused and finally forwarded to the TCP server module. The TCP server module communicates with the TCP client (sorting control application) of the sorting unit (SPC). Flake size, position and class (type) information are forwarded to the client which activates the corresponding valves of the sorting unit at the right time. User interaction, hardware configurations (morphology acquisition, spectral acquisition, hardware synchronization), and result presentation are handled by the user interface module.

6. Experimental performance evaluation of the measurement system prototype In order to evaluate the performance of the developed measurement system prototype in practice, 12 different classes of plastic flakes were used. The small flakes (see Fig. 1) were produced using an industrial extruder. 6.1. Selected polymers For the system evaluation the following 3 technical polymers of different colors were selected, which are highly important in practice: Polyoxymethylene (POM), Polybutylenterephthalat (PBT) and Acrylonitrile Styrene Acrylate (ASA). POM is an intrinsically opaque white thermoplastic which can be characterized by a high hardness, stiffness and strength. It is commonly used in applications for mechanical and electrical engineering, and automotive industry. PBT is a thermoplastic engineering polymer which shows a high strength, stiffness and dimensional stability as well as a low mechanical abrasion and a high resistance against solvents. It is commonly used for chassis and plug housings in electrical engineering. ASA is a terpolymer which is used to produce high-grade glossy and scratch-resistant surfaces. It has a high chemical resistance and a good resistance against atmospheric corrosion and UV radiation. Application fields include the automotive industry as well as constructions and housings used in the outdoor area. The concept and the polymer labeling is certainly not limited to the 3 polymer types used for the performance evaluation in this study, but can be easily extended to a large number of polymers. 6.2. Assignment of markers to polymers Four markers based on perylene derivatives which emit fluorescence light in the wavelength range between approx. 450 nm and


9


Morph. image storage Morph. image Position of conveyor belt Signal from rotary encoder (Conveyor belt)

Morph. aquisition

Morpholog. analysis

Morpholog. information

Hardware synchronization

Results (Code, location and size of flakes) Results

Result database Spectral acquisition

Sorting unit

Spectral classification Functional information

Results

Spectral image Spectral image storage

SPC

TCP client

TCP server

User interaction

Presentation of results

User Interface

Hardware configurations

Signal from rotary encoder (Conveyor belt)

Fig. 5. Software concept and implemented software modules. The developed controls the entire hardware of the prototype and performs the data acquisition, processing and analysis tasks in a parallel fashion. Hardware and user related communications are represented by solid lines. Communications and data exchange between modules are represented by dashed lines and realized through LabVIEWÓ related queue mechanisms. The communication with the sorting unit application (SPC) is carried out via a TCP interface.

excitation light. The marker fluorescence emissions should not interfere with the autofluorescence of the polymers in order to achieve a high classification performance. 6.3. Classification The overall goal of the developed measurement system is to perform multi-class classification (in this study into l = 12 classes). For the results shown in this study, we applied linear spectral unmixing (Chang, 2013) which assumes that a given spectrum with L bands can be represented by the linear superposition of a number of so-called endmembers. When p-endmembers exist in the sample signature r, their relative abundance factors can be calculated using the following linear mixture model (Chang, 2013):

r ¼ Mref a þ n; where r denotes the L-dimensional data vector (fluorescence spectrum). Mref = [m1, m2, . . ., mp] represents the L p signature matrix whose columns are the endmembers (reference spectra of the markers). a = (a1, a2, . . ., ap)T corresponds to the unknown p-dimensional abundance vector and n is the noise vector.

Normalized Emission Intensity

800 nm as shown in Fig. 6 were employed to label the polymers. Since each marker has its own absorption characteristics and quantum yield, the fluorescence intensities are different even if the markers get stimulated by the same excitation source. Fluorescence intensities that differ too much relatively from one another result in a lower classification performance (Fomin et al. (2013)), which is why we experimentally adapted the individual marker concentrations in a way such that the peak intensities of the 4 main lobes of the marker spectra emitted from polycarbonate plates are virtually equal. While many of the desired marker requirements listed in Section 4 are met, the 4 markers have one obvious drawback as shown in Fig. 6: since their fluorescence emissions spectrally overlap, the classification cannot be theoretically optimal and more sophisticated signal processing and classification methods are also required. Some basic considerations with regard to the assignment of certain markers to polymers as shown in Table 2 are advantageous. Intrinsic polymer properties (e.g. color and presence of additional color pigments, autofluorescence, etc.) should be matched to the marker characteristics (emission and absorption spectra). For polymers that host additional color pigments, the assignment of markers and marker combinations should be optimized in order to achieve the highest possible classification performance using a set of given markers. While white, black or gray-shaded plastics simply cause an intensity scaling of the emitted fluorescence spectra, colored plastics reflect incoming light particularly in the spectral range that is associated with their own color and absorb light at the other wavelengths. This results in wavelength-dependent reflection and absorption and hence potentially affects the shape of marker fluorescence emission spectra in different plastics (Fomin et al., 2013). In order to avoid the implementation of appropriate measures and make the signal processing as fast as possible, the wavelength range of a fluorescence marker used for a certain group of plastics should coincide with the spectral range that is associated with the polymer color. In addition, the fluorescence which might inherently be emitted by polymers even without any additives (autofluorescence) (Piruska et al., 2005; Lu et al., 2010) needs to be considered as well. Polymers can show fluorescence emission in the wavelength range from approx. 350 nm to 500 nm, especially when they get stimulated by UV and near UV

1 0.8 0.6 0.4 0.2 0

450

500

550

600

650

700

750

800

Wavelength λ in nm M1 (0.16 ppm)

M2 (1 ppm)

M3 (2 ppm)

M4 (6.6 ppm)

Fig. 6. Measured emission spectra (normalized) of the 4 fluorescence markers incorporated into polycarbonate. The marker concentrations (ppm levels) were adjusted in order to achieve virtually the same emission intensities when using the same excitation light source (here a high-power LED).


10


The signature matrix can be constructed based on the reference spectra of the 4 applied fluorescence markers incorporated into polycarbonate plates. For a measured spectral signature, the ^1; a ^2; a ^3; a ^ 4 Þ to the true marker ^ ¼ ða least-squares estimate (LSE) a ^ ¼ Pr where abundance coefficients in a can be calculated by a P = (MrefT Mref)1 MrefT represents the so called pseudo inverse. In ^ that is closest geometrical terms, the LSE defines a vector ^r ¼ Mref a (in the Euclidean sense) to the measured spectrum r (Chang, 2013). For every investigated flake, several emission spectra are ^g acquired and the expectation value of the abundance vector Efa is used for classification. The classification is carried out by calcu^ g and the LSE reference lating the Euclidean distance between Efa ^ l;ref which are derived for all l classes except abundance vectors a class 12 (since flakes of class 12 do not include any markers). ^ g and a ^ l;ref is conThe class with the minimal distance between Efa sidered the best match (winner takes it all rule). A plausibility check of the classification result based on anomaly detection reduces false positive hits. 6.4. Classification performance metrics In order to evaluate the classification performance of the measurement system in our multi-class problem, the following two metrics were used (Sokolova and Lapalme, 2009): The true positive rate (sensitivity) quantifies how well a classifier can correctly predict examples belonging to the actual class Ci. It is defined by:

TPRi ¼

TP i 100% and TPRM ¼ TPi þ FNi

Pi

i¼1 TPRi

l

;

where TPRi denotes the achieved true positive rate for class Ci and TPRM represents the mean over all i = 1. . .l classes (macro averaging). TPi and FNi are the numbers of true positive and false negative classifications for class Ci, respectively. A large TPRi value (close to 100%) for a class Ci is desired to correctly identify a high percentage of the flakes belonging to that class. TPRi = 100% means that no false negative classifications occurred and all flakes of class Ci were correctly identified. The positive predictive value (precision) measures the probability that the classification of an example as positive is actually true:

PPV i ¼

TPi 100% and PPV M ¼ TP i þ FPi

Pi

i¼1 PPV i

l

;

where PPVi is the positive predictive value achieved by the classifier per individual class Ci, and PPVM the macro average over all l classes. FPi defines the number of negative examples for the class Ci that are

incorrectly classified as positive (false alarms). A large PPVi value (close to 100%) is thus necessary to avoid false positive classifications and achieve a high purity of sorted flakes for class Ci.

7. Experimental design and experimental results The developed measurement system components were fit together with the conveyor belt as well as the singularization and sorting units at our industrial partner’s work floor in order to establish the working prototype. Before the experimental evaluations could start, the properly adjusted data acquisition systems (working distance, focusing, aperture, etc.) were both spatially calibrated in order to guarantee the correct fusion of functional (spectral) and morphological information. The illumination light of both illumination units was focused on the conveyor belt and spatial inhomogeneities (e.g. due to apodization) were measured in order to perform software correction. Reference emission spectra and ^ l;ref including their standard corresponding abundance vectors a deviations required for the anomaly detection where derived for the 11 polymer-marker combinations (class 12 does not include any marker) from representative training sets. We aimed at evaluating the classification performance of the measurement system under the most practical, however still realizable conditions. Statistical metrics like the true positive rate, positive predictive value, etc. are to be derived from a (sufficiently) large number of examples (flakes). As a side note it needs to be pointed out that among other things this number is influenced by the SNR which is different for the different (plastic) classes and varies even from flake to flake. Another experimental difficulty after the sorting of a mixture of flakes has been carried out is how to determine whether all (i.e. thousands of) flakes assigned to a class (i.e. container) actually belong to that class, and how many flakes were put into a wrong container, etc. Even for a low number of different plastics this quickly becomes impractical. Furthermore, for accurate evaluations great care must be taken to completely rule out any cross contaminations between flakes of different classes due to e.g. dust produced by the mills that adheres to flakes of another class. Our experiments for the evaluation of the measurement system’s classification performance were thus designed as follows: At first, 10,030 flakes were investigated for polymer-marker combination number 3 in Table 2 in order to figure out the number of flakes required to achieve statistically significant test results. Since a number of approx. 5000 flakes proved to be sufficient for the given experimental set-up, fluorescence emission spectra from 5000 to 6000 flakes per polymer-marker combination were then separately acquired for the remaining 11 polymer-marker combinations plus an additional group of non-labeled polymers of differ-

Table 2 Assignment of marker combinations (codes) to polymers. The marker concentrations were adjusted such that the peak intensities of the fluorescence emissions do not differ much. Symbols: U marker present; s no marker incorporated. Abbreviations: ASA: Acrylonitrile Styrene Acrylate; PBT: Polybutylenterephthalat; POM: Polyoxymethylene; Delrin is a registered trademark of Dupont; Ultraform is a registered trademark of BASF; Luran is a registered trademark of Styrolution; Hostaform is a registered trademark of Ticona; Pocan is a registered trademark of Lanxess; Tenac is a registered trademark of Asahi Kasei Chemicals. Class nr.

Polymer brand

polymer type

Marker 1

Marker 2

Marker 3

Marker 4

1 2 3 4 5 6 7 8 9 10 11

Delrin YellowÒ UltraformÒ S2320 Nature HostaformÒ C13031 Nature PocanÒ B1508 Nature DelrinÒ Black DelrinÒ Grey TenacÒ 3010 Nature + CPM51049 (1%) DelrinÒ Nature LuranÒ Black HostaformÒSW Grey TenacÒ3010 Nature

POM POM POM PBT POM POM POM POM ASA POM POM

s s s s U U U U U U U

s U U U s s s s U U U

U s U s s s U U s s U

s U U U s U s U s U U


11


Table 3 Confusion matrix calculated from the experimental classification results for l = 12 classes (11 polymer-marker combinations plus class 12 representing non-marked polymers), where the actual classes and the classes predicted using the developed measurement system prototype are shown. Entries along the main diagonal correspond to correct classifications, PPVi denotes the class-related precision, TPRi the achieved true positive rate per class. Overall, the measurement system prototype works highly reliable and provides an average positive predictive value PPVM of 99.5% and an average true positive rate TPRM of 99.4%.

Actual Class

TPRi 1 5352 0 0 0 0 0 0 0 0 0 0 49 2 0 5151 0 0 0 0 0 0 0 72 0 74 3 0 0 10030 0 0 0 0 0 0 0 0 27 4 0 0 0 5120 0 0 0 0 0 0 0 17 5 0 0 0 0 5790 0 0 0 0 0 0 33 6 0 0 0 0 0 6299 0 0 0 0 0 44 7 0 0 0 0 0 0 5154 0 0 0 0 9 8 0 0 0 0 0 0 0 6404 0 0 0 0 9 0 0 0 0 0 0 0 0 5212 0 0 8 10 0 0 0 0 0 0 0 0 0 5869 0 80 11 0 0 0 0 0 0 0 0 0 0 6159 0 12 0 0 0 0 0 0 0 0 0 0 0 7330 PPVi

100

100

100

100

99.9

1

2

3

4

5

100

100

6 7 Predicted Class

100

100

98.8

100

96

8

9

10

11

12

ent colors. After the data acquisition, on-line signal and image processing, fusion of morphological and functional (spectral) information and classification, all flakes of a class were collected in a container and their total mass was determined using a highly accurate laboratory scale. In order to avoid residues and dust from an earlier measurement, the conveyor belt was manually cleaned before each measurement started. Evaluation results for the 12 different classes of plastic flakes (11 fluorescently labeled polymers plus one class of non-labeled polymers) are summarized in the so-called ‘‘confusion matrix’’ (contingency table) illustrated in Table 3. While entries along the main diagonal of the 12 12 matrix show the number of correctly classified flakes for each class (i.e. polymer-marker combination), all other entries in the matrix indicate wrong classifications. The confusion matrix thus quickly provides information about classes that are confused with other classes and how frequently this happens (Brereton, 2009). The bottom row shows the positive predictive value (PPVi) for each class, the very right column shows the true positive rate per class. The parameters of the measurement system and the applied classifier were configured in a way such that the average positive predictive value PPVM and average true positive rate TPRM are equally balanced and non-labeled plastics are not misclassified as labeled ones. Although the system parameters should be chosen depending upon the specific application, the aforementioned tuning is often recommended since it ensures a high purity of the sorted material without contaminations by non-labeled polymers occur while at the same time an acceptably low number of false negative detections (see sum of the column entries for class 12 in Table 3 which results in the elimination of approx. 0.5% of actually labeled polymers from the recycling process). Using the developed measurement system and 12 classes of plastics flakes (11 classes labeled with the 4 fluorescence markers shown in Fig. 6, plus 1 additional class of non-labeled polymers), an average true positive rate TPRM of 99.4% and an average precision PPVM of 99.5% was achieved. According to Table 3 the lowest values for the true positive rate occur for type 2 (TPR2 = 97.2%)

99.1 97.2 99.7 99.7 99.4 99.3 99.8 100 99.8 98.7 100 99.9 99.4 99.5

and type 10 (TPR10 = 98.7%), the lowest values for the positive predictive value appear for type 10 (PPV10 = 98.8%) and type 12 (PPV12 = 96%). A detailed analysis of incorrectly classified flakes showed that their emitted fluorescence intensities are significantly lower in comparison with other flakes of the same class. It seems that in a small number of certain polymer flakes the fluorescence makers are present at a lower concentration level. The marker incorporation process (carried out on an industrial extruder to ensure a high quality) and possible reasons for this behavior are being currently investigated by specialists of our project partner. In addition to a high classification performance, the economical success of the entire concept highly depends on the throughput that can be achieved. With flakes produced by the industrial mill Rapid 1022 and mechanical screening (flake sizes 3–10 mm), 1500 flakes per second must be handled in order to achieve the desired throughput of 250 kg/h. With the current setup of the measurement prototype, up to 1800 flakes per second can be handled using a speed of the conveyor belt of 260 mm/s and the entire width of the conveyor belt of 500 mm (and applying a spectral acquisition rate of approx. 300,000 spectra/s).

8. Conclusions The recycling of plastics is highly desirable to preserve the limited fossil resources, save energy and reduce the amount of CO2 emissions as well as plastic garbage and plastic debris floating in the world oceans. In order to produce recycled polymer parts of high quality and in a cost-effective way, highly reliable automated plastic classification and sorting techniques are a key component. A promising approach to overcome the drawbacks of existing techniques such as NIR spectroscopy is to label plastics by incorporating fluorescence markers at the (sub) ppm level into virgin polymer resins during the manufacturing process and then using the characteristic fluorescence emission signatures for classification purposes.


12


In this study, the prototype of a measurement system was developed which can identify fluorescently labeled polymers and provide the necessary information for the sorting machinery. At current, this prototype system is able to handle approx. 1800 small plastic flakes (mill material) per second which corresponds to a throughput of approx. 300 kg per hour. The classification performance was experimentally evaluated using 3 different technical polymers (3 polymer types, different colors). 11 classes are labeled with unique binary combinations of the 4 applied fluorescence markers and class 12 includes unlabeled polymer flakes of various colors. The measurement system achieves an average sensitivity over all classes of TPRM = 99.4% and an average precision of PPVM = 99.5%. The lowest class-related sensitivity (TPR10 = 98.7%) and precision (PPV12 = 96%) are mainly due to lower marker fluorescence emissions caused by local variations of the marker concentration found in some flakes of two classes. The marker incorporation process is currently being optimized such that the performance metrics of all classes will exceed 99%. Our next research steps will include the investigation of various approaches for the minimization of the marker concentration levels. Acknowledgments The authors thank their industrial partner for supplying the polymers and purchasing the measurement hardware, one of their academic partners for providing the 4 fluorescent dyes which were incorporated by another academic partner into the polymers before small plastic flakes were produced. This work was partly supported by the German Federal Ministry of Economics and Technology as per resolution of the German Parliament. References Ahmad, S., 2004. A new technology for automatic identification and sorting of plastics for recycling. Environ. Technol. 25 (1143–1149), 2004. http:// dx.doi.org/10.1080/09593332508618380. Bezati, F., Froelich, D., Massardier, V., Maris, E., 2010. Addition of tracers into the polypropylene in view of automatic sorting of plastic wastes using X-ray fluorescence spectrometry. Waste Manage. 30, 591–596. Bezati, F., Froelich, D., Massardier, V., Maris, E., 2011. Addition of X-ray fluorescent tracers into polymers, new technology for automatic sorting of plastics: proposal for selecting some relevant tracers. Resour. Conserv. Recycl. 55, 1214–1221. BIR, 2014. Recycling facts. URL: . BP-Sorting, 2013. Overview Report: Black Polymer Sorting – unisort blackeyeÓ. URL: . Brereton, R., 2009. Chemometrics for Pattern Recognition, first ed. John Wiley & Sons Ltd, New York, NY, USA, p. 318. Brunner, S., Kargel, Ch., 2011. Evaluation of potential emission spectra for the reliable classification of fluorescently coded materials. In: Proc. SPIE algorithms and technologies for multispectral, hyperspectral, and ultraspectral imagery XVII. Orlando, Florida, USA. pp. 80480I-80480I-15. Brunner, S., Fomin, P., Zhelondz, D., Kargel, Ch., 2012. Investigation of algorithms for the reliable classification of fluorescently labeled plastics. In: International Instrumentation and Measurement Technology Conference (I2MTC 2012), Graz, Austria. Bruno, E., 2000. Automated Sorting of Plastics for Recycling. URL: . Chang, C., 2013. Hyperspectral Data Processing: Algorithm Design and Analysis, first ed. John Wiley & Sons, New Jersey. Cornell, D., 2007. Biopolymers in the existing postconsumer plastics recycling stream. J. Polym. Environ. 15. http://dx.doi.org/10.1007/s10924-007-0077-0. Dvorak, R., Kosior, E., Moody, L., 2011. Final Project Report: Development of Nir Detectable Black Plastic Packaging. URL: . Fomin, P., Brunner, S., Kargel, Ch., 2013. Investigation of fluorescence spectra disturbances influencing the classification performance of fluorescently labeled plastic flakes. Proc. SPIE 8791, 87911J-87911J-15. doi: 10.1117/12.2020972. Gonzalez, R.C., Woods, R.E., 2008. Digital Image Processing, third ed. Prentice-Hall Inc., Upper Saddle River, NJ, USA.

Hu, B., Giacometti, L., Di Maio, F., Rem, P., 2011. Recycling of WEEE by magnetic density separation. In: Proceedings of the Sixth International Conference on Waste Management and Technology (ICWMT 6), 30 August–1 September 2011, Suzhou, China. Jähne, B., 2005. Digital Image Processing, sixth ed. Springer, Berlin [u.a.]. Krieg, G., Fey, D., Bohleber, J., 1998. Verfahren und Vorrichtung zur Detektion und Unterscheidung zwischen Kontaminationen und Gutstoffen sowie zwischen verschiedenen Farben in Feststoffpartikeln. Lu, B., Zheng, S., Quach, B.Q., Tai, Y.C., 2010. A study of the autofluorescence of parylene materials for [small mu] tas applications. Lab Chip 10, 1826–1834. http://dx.doi.org/10.1039/B924855B. Luciani, V., Serranti, S., Bonifazi, G., Di Maio, F., Rem, P., 2013. Quality control in the recycling stream of PVC from window frames by hyperspectral imaging. In: Proc. SPIE Optics + Optoelectronics 2013, Conference on Optical Sensors, 15–18 April 2013. Prague, Czech Republic. Luciani, V., Bonifazi, G., Rem, P., Serranti, S., 2014. Upgrading of PVC rich wastes by magnetic density separation and hyperspectral imaging quality control. Waste Manage. http://dx.doi.org/10.1016/j.wasman.2014.10.015. Luttermann, K., Claussen, U., Rieß, R., 1992. Verfahren zur Kennzeichnung von Kunststoffen. Luttermann, K., Claussen, U., Orth, P., 1996. Verbessertes Verfahren zur Kennzeichnung von Kunststoffen. Mankosa, M., Luttrell, G., 2005. Plastic Material Having Enhanced Magnetic Susceptibility, Method of Making and Method of Separating. URL: . US Patent 6,920,982. Maris, E., Aoussat, A., Naffrechoux, E., Froelich, D., 2012. Polymer tracer detection systems with {UV} fluorescence spectrometry to improve product recyclability. Miner. Eng. 29, 77–88. doi: (sustainability through Resource Conservation and Recycling). Meier, H., Becker, A., Seng, F., and Luttermann, K., 1997. Polymere Fluoreszenzmarker. Merrington, A., 2011. Recycling of plastics. In: Kutz, M. (Ed.), Applied Plastics Engineering Handbook. William Andrew Publishing, Oxford. Plastics Design Library, pp. 177–192. URL: . NatureWorks, 2009. Using Near-Infrared Sorting to Recycle Pla Bottles. URL: . Niaounakis, M., 2013. Biopolymers: Reuse, Recycling, and Disposal. William Andrew. chapter 5. Physical Recycling. Plastics Design Library, pp. 1151–163. URL: . Pascoe, R.D., 2000. Sorting of Waste Plastics for Recycling. Rapra Technology. volume 11. p. 13. URL: . Piruska, A., Nikcevic, I., Lee, S.H., Ahn, C., Heineman, W.R., Limbach, P.A., Seliskar, C.J., 2005. The autofluorescence of plastic materials and chips measured under laser irradiation. Lab Chip 5, 1348–1354. http://dx.doi.org/10.1039/B508288A. PlasticsEurope, 2008. The Compelling Facts About Plastics 2009. PlasticsEurope. ‘‘B1160 Brussels – Belgium. URL: . PlasticsEurope, 2013. Plastics – the Facts 2013: An Analysis of European Latest Plastics Production, Demand and Waste Data. Riise, B., Allen, L., Biddle, M., Fisher, M., 2001. Value added color sorting of recycled plastic flake from end-of-life electrical and electronic equipment. In: Electronics and the Environment, 2001. Proceedings of the 2001 IEEE International Symposium on, pp. 223–228. doi: http://dx.doi.org/10.1109/ISEE.2001.924530. Serranti, S., Gargiulo, A., Bonifazi, G., 2011. Characterization of post-consumer polyolefin wastes by hyperspectral imaging for quality control in recycling processes. Waste Manage. 31, 2217–2227. Serranti, A., Gargiulo, G., Bonifazi, G., 2012. Hyperspectral imaging for process and quality control in recycling plants of polyolefin flakes. J. Near Infrared Spectrosc. 20, 573–581. Serranti, S., Fraunholcz, N., Di Maio, F., Bonifazi, G., 2013. Recycling-oriented characterization of polyolefin packaging waste. Waste Manage. 33, 574–584. Simmons, B., Overton, B., Ahmad, S., 1998. Fluorescent tracers enable automatic identification and sorting of waste plastics. British Plastics Rubbers, 4–10. Sokolova, M., Lapalme, G., 2009. A systematic analysis of performance measures for classification tasks. Inf. Process. Manage. 45, 427–437. http://dx.doi.org/ 10.1016/j.ipm.2009.03.0020. STEINERT, 2014. UnisortÒ Flake with Nir or Colour Sourting in Fine-Grain Materials. URL: . Tatzer, P., Wolf, M., Panner, T., 2005. Industrial application for inline material sorting using hyperspectral imaging in the NIR range. Real-Time Imag. 11, 99– 107. TITECH, 2014. Sorting Technology. URL: . Tsuchida, A., Kawazumi, H., Kazuyoshi, A., Yasuo, T., 2009. Identification of Shredded Plastics in milliseconds using Raman spectroscopy for recycling. Sensors, 2009 IEEE, vol., no., pp. 1473,1476, 25–28 October 2009, doi: http://dx. doi.org/10.1109/ICSENS.2009.5398454. Ulrici, A., Serranti, S., Ferrari, C., Cesare, D., Foca, G., Bonifazi, G., 2013. Efficient chemometric strategies for PET-PLA discrimination in recycling plants using hyperspectral imaging. Chemometrics Intelligent Lab. Syst. 122, 31–39.


The prototype of a thermoregulatory system for measurement and control of temperature inside prosthetic socket.

Design and Prototype of an Automated Column-Switching HPLC System for Radiometabolite Analysis.

Development and evaluation of a prototype tracking system using the treatment couch.

Development of a polymer based fiberoptic magnetostrictive metal detector system.

BiFCROS: A Low-Background Fluorescence Repressor Operator System for Labeling of Genomic Loci.

Development of an automated fluorescence microscopy system for photomanipulation of genetically encoded photoactivatable proteins (optogenetics) in live cells.

Automated anatomical labeling of abdominal arteries and hepatic portal system extracted from abdominal CT volumes.

bioimpedance measurement system.

Isolation of a gramicidin S hyperproducing strain of Bacillus brevis by use of a fluorescence activated cell sorting system.

Chromosome measurement and sorting by flow systems.

Prototype development of an electrical impedance based simultaneous respiratory and cardiac monitoring system for gated radiotherapy.

A prototype expert system for fishway design.

Design and development of a prototype electrotherapy device.

Development of an automated culture system for laboratory evolution.

Development and validation of a fully automated system for detection and diagnosis of mammographic lesions.

A Method for Automated Cortical Surface Registration and Labeling.

A prototype of a portable TDCR system at ENEA.

Breast Imaging Reporting and Data System (BI-RADS) breast composition descriptors: automated measurement development for full field digital mammography.

Concentration and acid-base controllable fluorescence of a metallosupramolecular polymer.

Automated measurement of urinary proteins.

Development of Prototype Filovirus Recombinant Antigen Immunoassays.

System Design and Development of a Robotic Device for Automated Venipuncture and Diagnostic Blood Cell Analysis.

Development of a [177Lu]BPAMD labeling kit and an automated synthesis module for routine bone targeted endoradiotherapy.

RM 8111: Development of a Prototype Linewidth Standard.