Professional Documents
Culture Documents
L. Madan Kumar I , Boppana Pavan2 ,4 , P.V. Kalyan2 , 3 , Nirmal Savio Paul2 , , R.S. Prakruth2 ,4 and T. Chinnu l
2 Student Trainee, CSIR - Central Electronic Engineering Research Institute, Chennai Centre,
1 CSIR - Central Electronic Engineering Research Institute, Chennai Centre,
Abstract - In this paper we present an embedded system While recycling plastics can be contaminated by the mixing
solution to automate the sorting of different types of plastic by of types and by non-plastic materials [5] .
using the concepts of Near Infrared Spectroscopy (NIRS). Segregating non-plastics, and different types of plastic
Plastics are everyday used non-biodegradable materials when like Poly Ethylene Teraphthalate (PET), High Density Poly
not disposed properly has adverse effects on the environment.
Ethylene (HDPE), Poly Vinyl Chloride (PVC), Low Density
For recycling of plastics different types of plastics (polymers)
have to be identified and segregated. For economic reasons Poly Ethylene (LDPE), Poly Styrene (PS) and Poly
plastics must be identified and sorted instantaneously. The Propylene (PP) from each other are labor-intensive and
NIRS technique has been used for the instantaneous hitherto, not many feasible techniques exist. Society of
identification of plastics. Measurements made by NIRS are Plastic Industries (SPI) has standardized seven codes to
quite accurate and fast. To realize the above, a low cost identify different types of plastic for recycling. Widespread
Raspberry Pi based Control System, for efficient sorting of use of fillers and additives in plastics make them difficult to
waste plastics has been developed. The necessary software recycle because of the difficulty to remove fillers and
required to interface the spectrometer with the Raspberry Pi as additives from plastics. Additives are less widely used in
PET (bottles, containers) & HDPE (milk jugs, shopping bags
well as to process the NIRS data to obtain information on the
polymer category have been written on the general-purpose,
& shampoo bottle) when compared to other plastics. Due to
the ease of recycling most of the recycled plastics is PET &
high-level programming language Python.
EVib = (v + )
and S-H bands observed in NIR spectra can be
2 27r
(k
vii
(1) characteristically attributed to specific polymers, thus
enabling identification of most commonly used plastics. For
where, v i s the vibrational quantum number, h i s Planck's example, the NIR spectra of HDPE exhibit a peak at around
constant, k is force constant and u is reduced mass. 1 200 nm, which is, not present in PET and is only small for
Transitions take place between consecutive energy levels PVC. PET on the other hand exhibits three characteristic
peaks in the region 1 400 -1 700 nm (refer Fig. I).
.,. ---':::"""---':--r--,
r---"'::-.,.----.---:...-
that cause change in the dipole moment. However, the
0.9
harmonic oscillator model fails to explain behavior of actual
molecules, as they don't account for Columbic repulsion
between atoms or dissociation of bonds. As a result, they
!
0,75
more closely represent anharmonic oscillations with
0.7
unequally spaced energy levels, given as [ 1 0] :
l'l
0.65
Ct: (J.S5
Here y is the anharmonicity factor. The anharmonicity can 0.5
result in transitions between vibrational energy. These 0 ...5
transitions between non-contiguous vibrational states yield
absorption bands known as overtones (first and second
I--- HOPE LOPE PET pp pvcl
Wal!! leh (nm)
- -- -.-
overtone, respectively) at, approximately, multiples of the
fundamental vibrational frequency. Also, they are much less Fig. 1 : NIRS spectra of different plastic classes.
likely than the fundamental transitions, so the bands are A further advantage of NIR is that the NIR photo
much weaker (the band for the first overtone is 1 0-100 times detectors namely, germanium, indium arsenide, or indium
weaker than that for the fundamental frequency, depending gallium arsenide, have short response times and high
on the particular bond). These bands appear between 780 nm detector sensitivities. Also, quartz fiber optics with low
and 2000 nm, depending on the overtone order and the bond attenuation and low cost can be used for convenience and for
nature and strength. In polyatomic molecules, two or more remote sensing. NIR spectrometers typically have no
vibrational modes can interact in such a way as to cause moving parts and therefore they are not affected by the
simultaneous energy changes and give rise to absorption vibration or dust/dirt typical of plastic recovery facilities.
bands called combination bands, the frequencies of which Furthermore, NIR instruments require little maintenance and
are the sums of multiples of each interacting frequency. NIR provide excellent reproducibility with negligible
combination bands appear between 1 900 nm and 2500 nm. instrumental drift [8]-[ I 0] ,
The intensity of NIR bands depends on the change in
C. Multivariate Analysis as a vital cog in NIRS analysis
dipole moment and the an-harmonicity of the bond. Because
the hydrogen atom is the lightest, and therefore exhibits the Multivariate Analysis (MVA) [7] is the science of relating
largest vibrations and the greatest deviations from harmonic measurements made on a chemical system to the state of the
behavior, the main bands typically observed in the NIR system via application of mathematical or statistical methods
region correspond to bonds containing this and other light and of designing optimal experiments for investigating
atoms (namely C-H, N-H, O-H and S-H) ; by contrast, the chemical systems/ specimen. The goal of many MVA
bands or bonds such as C=O, C-C and C-CI are much techniques is an efficient production of an empirical or
2
semi-empirical model, derived for data, which is used to between their respective spectra. The above process is
estimate one or more properties. Apart from obtaining a followed for all the available reference signatures. The
model that provides useful predictions, MVA techniques can reference signatures also include that of blank capture (i.e.,
also be used to obtain insight about a chemical system. without any material placed) to recheck for material
availability.
D. From NIRS to information on plastic class
While the NIRS contains valuable clues to the nature of III. COMPUTATIONAL HARDWARE AND SOFTWARE
the polymer, it has to be accompanied by MVA to unravel The MVA algorithms presented above were implemented
information. The data pretreatment and MVA methodology in Python language and ported onto Raspberry Pi boards.
employed to evaluate the plastic polymers are listed below: A. Raspberry Pi
Outlier detection: Outliers are aberrations in the data
Raspberry Pi [ 1 5] - [ 1 6] is an open-source single-board
which occur due to errors in the instruments or in the
microcontroller. The hardware consists of a simple open
acquisition of data. Outliers can significantly alter the
hardware design for the Raspberry Pi board and on-board
characteristics of the data and hence have to be removed.
input/output support. The software consists of a standard
In this work we employ Principal component analysis
programming language compiler and the boot loader that
(PCA) [ I I ] to identify and remove the outliers.
runs on the board. Raspberry Pi hardware uses Linux
B aseline correction: B aseline correction is the process of
kernel-based operating system [ 1 5] - [ 1 6] .
removing background noise in the measurement of the Raspberry Pi can receive inputs from a variety o f sensors
data. In this work we have used the Savitzky Golay ( I I and can control its environment by controlling Halogen
point, I st order differentiation) [ 1 2] technique to correct lights, motors, Air nozzle and other actuators. The Raspberry
the statistical noise in the spectral data. Pi board is programmed using the Python programming
Scatter Correction: Scatter correction is necessary to language.
offset any variations in the measurements due to
differences in particle sizes (for powdery samples), B. Python
effective path length and light scattering. In this effort Python [ 1 5] is a general-purpose, high-level
we use the popular multiplicative-scatter correction programming language and is famous for its code readability.
(MSC) [ 1 3] technique to correct for artifacts related to Raspberry Pi extends excellent support to the Python
scattering. language, some of the other merits of programming with
Pattern recognition, validation and analysis: To build the Python in comparison to other mathematical tool/statistical
model and validate it we have applied the Successive packages such as MATLAB TM or R language. Python has a
Approximation Method (SAM) [ 1 4] . separate GPIO module which allow us to access GPIO pins
on Raspberry Pi. Python supports multiple programming
E. Spectral Angle Mapping (SAM)for Pattern recognition models, including object-oriented, imperative and functional
The spectral Angle Mapping is a simple algorithm for programming styles. Python is a scripting language, but can
pattern correlation and recognition which does not involve also be used for non-scripting.
complex statistical evaluation. It is commonly used for C. Interfacing the embedded hardware and NIRS instrument
directly comparing a given spectra to a known spectra Serial communication is used for communicating between
(usually determined in a lab or in the field with a Raspberry Pi boards and the NIR Spectrometer. Serial
spectrometer). This method uses all the features supplied to communication is the process of sending data one bit at a
it to classify an unknown plastic polymer based on its NIRS time, sequentially, over a communication channel or
spectrum. The spectral data (known as well as unknown) are computer bus. An RS232 serial port operates at 1 2V and
considered as vectors in a multi-dimensional space and the can damage the Raspberry Pi board if connected directly to
angle between the vectors is used to determine their relative the pins of the Raspberry Pi. Hence, we use a RS-232 to
correlation. Taking dot product of the two spectrums does USB converter that would directly go into the USB port of
the projection and finding the angle between them. the Pi to interface the NIR spectrometer and the embedded
Consider that two spectrum A and B , represented by system.
vectors a and b, are under investigation. The two spectrum
are compared by obtaining the angle of inclination between
them using the formula, IV. TECHNICAL DESCRIPTION OF THE SYSTEM
'lAM
-I ab A. Components of the System
=cos ( ), (3)
. The schematic diagram of the plastic sorter is shown in
Ii a II Il b ll
where, the operator stands for the dot product of the Fig. 2. The system consists of the following units:
a conveyor belt assembly for on-line capture of
vectors and the operator I I I I connotes the Frobenius norm.
The classification is based on finding the minimum or spectral signature of plastic materials that are to
smallest angle between the unknown spectrum and the be sorted,
a white light source (halogen lamp) with wide
polymer class reference (library) spectrum. The cosine of the
distance between these two vectors gives the correlation spectral response,
3
an optical assembly to collect rays reflected by the 4. The optical assembly collects the reflectance light
plastic material, and through a multi-mode optical fiber
a NIR based Diode Array Spectrometer (DAS) communicates it to the DAS .
which interprets the optical information and 5 . The embedded system collects the spectrometer data
converts into digital data, performs baseline correction, scatter correction
proximity sensor and pattern recognition of the signature of the
an embedded system to collect digital data from polymer and uses decision supporting system for
the spectrometer and perform necessary classification.
mathematical computations to identify nature of 6. Jet Nozzles are activated based on the command
polymer, signal to eject classified plastics into their
an Air-Jet nozzle which is used to segregate the respective bins.
material of choice. 7 . The data base is updated for all the classified
consumer polymers.
B. System Operation
C. Graphical User lnteiface
The process flow diagram is depicted in Fig. 3 . The
operation of the system is briefly explained below: In order to control and configure the plastic sorter a
I . The singulated consumer/domestic waste polymer graphical user interface (GUI) has been developed. The
samples move on a fixed speed conveyor. GUI is based on python based GUI tool tkinter. Fig. 4
2. The samples are irradiated by ordinary lamps of 250 shows two snapshots of the GUI taken before and during
watts. the sorting of plastics.
3 . Optical device senses the vicinity of samples and
initiates the DAS to capture the signal.
R a s p be r ry P i
Python
RS 2 3 2 to U S B
Co nve rte r
G N U/Li n u x o n GUI
US B
A R M+G PU
GPI O Pins
NIR
....
<4 .j ... ... :
I@ @ @---::
P roxi m ity
... - - - - - - - - - - -.- - - - - - - - - - - - -
Air J et N ozzl es
S p ect ro m et e r
c:::
:::J ......
Se ns o r ...
TFT LC D
H a l oge n
M o n i to r
1
La m ps L-______
I
I
-------------
c===r c===r I
I
I
I
I
I
I
I
I
.!
P rototyp e Co nveyo r B e lt
---
M oto r
Co l l ect i o n B i n s
4
those which involve driving a video display or USB
Data
I n put P rocessing Output
peripherals .This kind of efficiency the system can have a
Tra n sfe r great impact on plastic recycling.
VI. RESULTS AND DISCUSSIONS
A. Description of sample set
A batch of 1 84 samples of 5 maj or plastic varieties was
considered, the details of which are described in Table I.
Each batch contained coloured and transparent plastics.
TABLE I: Description of sample set
PET HDPE PS PP PVC
of
Total
Fig. 3: Process Flow Diagram. Size
54 29 33 37 31
Sample set 1 84
7 Plastic Sorter 1 ,I,@) ii.f:l.l
=
B. Data treatment and Spectral Analysis
Input Functions P rocessin g State Output Count
Inititalize Spectrometer
The collected samples were analyzed using a 1 28 pixel
resolution Microparts STEAG NIR instrument with spectral
Waiting for Wa tting for
Dark Capture
Reference Capture
Wlliti ng for range of 1 000 nm to 1 700 nm. The obtained spectra were
Sorting
B utto n C lick.! Sotrin g to B egi n Sotri n9 to B eg i n
B iI!!g in
pretreated for background, thermal and optical noise using
Stop and Quit the signal processing techniques described in Section II-D.
The treated data was then analysed using Principal
(a) Before start of sorter Component Analysis (PCA) to identify and remove outliers
7J. Plastic Sorter [,= ,I@) in the samples. During pattern recognition samples which
Input Fu nctions P rocl5sing State Output Count covered boundary conditions was used as reference.
Inititarize Spectrometer
Stop and Qu
(b) During the sorting of plastics vectors studied. A value close to 1 .00 implies close match
Fig. 4: GUI for plastic sorter. while a value close to 0 corresponds to poor correlation.
The best match is declared to be the one for which
V. COMPARISON WITH EARLIER ARDUINO SYSTEM cos( BSAM ) is maximum (or alternatively, the BSAM is close
The system was earlier ported onto an Arduino processor to 0). To make the comparison to be effective the calibration
[ 1 7 ] . In relation to the Arduino based system, the Raspberry model is formed with average spectrum of the individual
Pi system offers many advantages. By using a Raspberry PI, class spectra hence the number of comparison can be greatly
this project would make the existing plastic sorting reduced. The class of plastic is then published on the GUI
embedded system more effective. The existing plastic and necessary actuators are activated to sort the plastic in
sorting system runs on Arduino Microcontroller board based their respective bins.
on C language. The Pi has a clear advantage for complex
networked and high performance embedded applications, TABLE II : Tabulated values of cos(BsAM ) for plastic classes.
and those which involve driving a video display or USB
peripherals. PET HDPE PS PP PVC
The Raspberry Pi board runs Python, which is simpler PET 0.993348 0.687779 0.607282 0.660003 0.535888
and more user friendly due to its inbuilt libraries and related HDPE 0.701992 0.997355 0.703522 0.84 1 8 0.656706
functions and it is more effective and gives accurate results.
PS 0.795542 0.645026 0.983352 0.646704 0.552696
As the Pi runs on a LINUX environment interfacing new
devices becomes much simpler compared to Arduino PP 0.7580 1 5 0.636601 0.742697 0.977152 0.699741
interfacing. The Arduino employs an 8-bit ATmega series PVC 0.573234 0.694659 0.53697 0.77783 0.989361
microcontroller whereas the Raspberry Pi is based around a
32-bit ARM processor, and the Arduino is typically clocked
at between 8-1 6MHz and with 2-8kB of RAM available, and D. System performance evaluation
in contrast the Raspberry Pi can be clocked at up to I GHz
The system has undergone rigorous trials to check for its
and may have up to 5 1 2MB of RAM. On top of which the Pi
accuracy and robustness. Regular tests have been conducted
has a GPU and video outputs, Ethernet as standard and USB
to ascertain its repeatability and consistency of performance.
host ports. The Pi has a clear advantage for complex
Table III provides a statistical data of the results obtained
networked and high performance embedded applications, and
over a period of 1 week.
5
TABLE III' Tabulation of classification performances results (particularly for the sorting of PVC and PP items) can
Sample be improved by adding more reference samples to the
Number PET HDPE PS PP PVC
type
of trials
pattern recognition algorithm. However, a penalty may have
to be paid in terms of the execution time.
PET 452 0 0 0 0
452 ACKNOWLEDGEMENT
HDPE 381 0 381 0 0 0 The authors express their heartfelt gratitude to Dr. A S .V
Sarma, Scientist-in-Charge, Mr. R. Sridhar, Chief Scientist
PS 0 0 366 0 0 and Mr. V. Venkatakraman, Senior Principal Scientist (Retd.),
366
of CSIR-CEERI, Chennai, for their guidance, constant
PP 0 0 0 411 0 motivation and support.
411
Some of the key features and advantages of the developed [2] The Self-Sufficiency Handbook: A Complete Guide to Greener Living by
Alan Bridgewater pg. 62--Skyhorse Publishing Inc . . 2007 ISBN
system are listed below, 1 -60239- 1 63-7. ISBN 978 - 1 -60239- 1 63-5
l . Automated, low cost device which can sort five groups [3] " Energy and Economic Value of Non-recycled Plastics and Municipal
of polymers through pattern recognition. Solid Wastes '" at Journalist's Resource. org '".
2. Smart and fast system capable of sorting up to 80 [4] C. Mlgaard. '"Environmental impacts by disposal of plastic from
plastic items per minute. municipal solid waste"', Resources, Conservation and Recycling Volume
I S , Issue I , October 1995, Pages 5 1-63 , Elsevier.
3 . System can handle bulk input (target specification is 4
[5] S.M. AI-Salem, P. Lettieri, J. Baeyens, "'Recycling and recovery routes of
tonnes of plastic per day). plastic solid waste (PSW) : A review"', Waste Management Volume 29,
4. Classification of plastics done to a high degree of Issue 10, October 2009, Pages 2625-2643 , Elsevier.
accuracy. For PET, HDPE and PS materials the [6] G. A. Marques, J. A. S . Ten6rio, '"Use of froth flotation to separate
PVC/PET mixtures '", Waste Management Volume 20, Issue 4, July 2000,
accuracy is close to 1 00% while PP and PVC the
Pages 265-269, Elsevier.
Roman M. Balabin, Ravilya Z. Safieva, and Ekaterina 1. Lomakina (2007).
accuracy is around 95 % . [7]
VI. CONCLUSIONS AND FUTURE RESEARCH TOPICS '"Comparison of linear and nonlinear calibration models based on near
infrared (NIR) spectroscopy data for gasoline properties
In this paper we presented our work on developing an prediction'". ChemomelrLab 1 83-188 .doi: 10. 1 0 1 6/j .chemolab.2007.04.006.
affordable and efficient system that can sort different [8] Isaksson, T., Fearn, T., & Davies, T. (2002). A user-friendly guide to
categories of plastics quickly and accurately through multivariate calibration and classification (Vol. 6). Chichester: NTR
non-destructive method for efficient recycling. The robust publications. Chicago
and indigenously integrated system is capable of handling [9] Osborne, B. G., Fearn, T., & Hindle, P. H. ( 1 993). Practical NIR
the vagaries of environmental conditions. spectroscopy with applications in food and beverage analysis. Longman
The system employs the concepts of NIR spectroscopy & scientific and technical.Chicago
multivariate analysis and has been designed and [ 1 0] Siesler, H. W. , Ozaki, Y., Kawata, S . , & Heise, H. M. (Eds.). (2008).
implemented using Raspberry Pi system. The computational Near-infrared spectroscopy: principles, instruments, applications. Wiley.
corn. Chicago
algorithms and GUI were developed using Python.
[ I I ] S. Wold, K. Esbensen and P. Geladi, "Principal component analysis."
The plastic sorter system can differentiate among 5 types
Chemometrics and intelligent laboratory systems, 2(1), 37-52, 1987.
of plastics, i.e. PET, HDPE, PVC, PS and PP. Not only is the
[ 1 2] A. Savitzky and M. J. E. Golay ( 1 964), "'Smoothing and Differentiation of
system robust, it is also accurate (near 1 00% accuracy for
Data by Simplified Least Squares Procedures '", Anal. Chern., 36,
PET, PS and HDPE and near 95% accuracy for PVC and PP), 1 627- 1 639.
can handle vast quantities of plastic (about 4 tonnes of [ 1 3 ] T. Isaksson and T. Nres, 'The Effect of Multiplicative Scatter Correction
plastic per instrument per day) and sort them quickly (up to (MSC) and Linearity Improvement in NIR Spectroscopy", Applied
80 plastic items per minute). It is worth mentioning here that Spectroscopy, Vol. 42, Issue 7, pp. 1 273- 1 284, 1988.
the accuracy of results obtained match with the requirements [ 14] H.Z.M Shafri, A. Suhaili and S. Mansor, "The Performance of Maximum
of the plastic waste management industry. Likelihood, Spectral Angle Mapper, Neural Network and Decision Tree
The system developed by the authors offers great scope Classifiers in Hyperspectral Image Analysis". J. Comput. Sci. , 3: 4 1 9-423 ,
for further improvement, for instance, the speed of operation. 2007 .
The diode array NIR spectrometer currently used takes about [ 1 5 ] E. Upton &G. Halfacree. Raspberry Pi User Guide. John Wiley, 20 12.
580 milliseconds (ms) for reading the data per scan while the [ 1 6] S . Monk. Programming the Raspberry Pi: Getting Started with Python.
time taken for the execution of computational algorithms is McGraw-Hill, 20 1 3 .
20 ms. State-of-art NIRS devices currently available come [ 17] Madan Kumar Lakshmanan, K . Shankar, Karan H. Shab, T. Chinnu and
V. Venkataraman, "Embedded Wireless-enabled Low Cost Plastic Sorting
with high-speed USB interfaces which take as less as I ms System for efficient Waste Management", INVITED PAPER. presented at
to acquire the spectral data. These fast and accurate NIRS IEEE Global Humanitarian Technology Conference (GHTC), Silicon
devices can be employed to significantly increase the speed Valley, California, USA, October-20 l 3 .
of operation of the system. Further, the accuracy of the