The University of Manchester Research

Classification of Non-ferrous Metals using MagneticInduction SpectroscopyDOI:10.1109/TII.2017.2786778

Document VersionAccepted author manuscript

Link to publication record in Manchester Research Explorer

Citation for published version (APA):OToole, M. D., Karimian, N., & Peyton, A. J. (2018). Classification of Non-ferrous Metals using Magnetic InductionSpectroscopy. IEEE Transactions on Industrial Informatics, 14(8). https://doi.org/10.1109/TII.2017.2786778

Published in:IEEE Transactions on Industrial Informatics

Citing this paperPlease note that where the full-text provided on Manchester Research Explorer is the Author Accepted Manuscriptor Proof version this may differ from the final Published version. If citing, it is advised that you check and use thepublisher's definitive version.

General rightsCopyright and moral rights for the publications made accessible in the Research Explorer are retained by theauthors and/or other copyright owners and it is a condition of accessing publications that users recognise andabide by the legal requirements associated with these rights.

Takedown policyIf you believe that this document breaches copyright please refer to the University of Manchester’s TakedownProcedures [http://man.ac.uk/04Y6Bo] or contact uml.scholarlycommunications@manchester.ac.uk providingrelevant details, so we can investigate your claim.

Download date:11. Sep. 2020

https://doi.org/10.1109/TII.2017.2786778https://www.research.manchester.ac.uk/portal/en/publications/classification-of-nonferrous-metals-using-magnetic-induction-spectroscopy(2f56b261-7385-41e3-80f1-0a04d5fd542e).html/portal/michael.otoole.html/portal/noushin.karimian.html/portal/a.peyton.htmlhttps://www.research.manchester.ac.uk/portal/en/publications/classification-of-nonferrous-metals-using-magnetic-induction-spectroscopy(2f56b261-7385-41e3-80f1-0a04d5fd542e).htmlhttps://www.research.manchester.ac.uk/portal/en/publications/classification-of-nonferrous-metals-using-magnetic-induction-spectroscopy(2f56b261-7385-41e3-80f1-0a04d5fd542e).htmlhttps://doi.org/10.1109/TII.2017.2786778

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 1

Classi�cation of Non-ferrous Metals UsingMagnetic Induction Spectroscopy

Michael D O’Toole, Member, IEEE, Noushin Karimian and Anthony J Peyton

AbstractRecycling automotive, electronic and other end-of-life waste liberates large quantities of metals which can bereturned to the supply chain. Sorting the non-ferrous metalshowever, is not straightforward. Common methods range fromlaborious hand-sorting to expensive and environmentally delete-rious wet processes. The goal is to move towards dry processes,such as induction sensors and vision systems, which can identifyand sort non-ferrous scrap ef�ciently and economically.

In this paper, we present a new classi�cation method usingmagnetic induction spectroscopy (MIS) to sort three high-valuemetals that make up the majority of the non-ferrous fraction -copper, aluminium and brass. Two approaches are investigated:The �rst uses MIS with a set of geometric features returned by avision system, where metal fragments are matched to known testpieces from a training set. The second approach uses MIS only. Asurprisingly effective classi�er can be constructed by combiningthe MIS frequency components in a manner determined byhow eddy currents circulate in the metal fragment. An averageprecision and recall (purity and recovery rate) of around 92%was shown. This has signi�cant industrial relevance, as theMIS-only classi�er is simple, scalable, and straightforward toimplement on existing commercial sorting lines.

Index TermsWaste recovery, Recycling, ElectromagneticInduction, Impedance measurement, Classi�cation algorithms,Spectroscopy.

I. INTRODUCTION

A sustainable future means moving towards a model whereend-of-life products are recycled to become the feedstockof the new. This poses a number of dif�cult question forindustry and legislators; not least, how can the valuablematerials be recovered, and the maximum value be extractedfrom the remainder? A leading example is the automotiveindustry where every year, over 40 million vehicles worldwideare designated end-of-life and marked for disposal [1]. Thesevehicles often contain large quantities of steel, aluminium andother valuable metals. Nearly 8 million end-of-life vehicles aregenerated per year in Europe alone, leading to the EuropeanUnion issuing directive 2000/53/EC, imposing a target of 85%recycle and reuse rate for end-of-life vehicles by 2015 [2]. Themost recently available statistics show that seventeen memberstates had achieved this goal as of 2014 [3].

The disposal of end-of-life vehicles involves the removal ofpollutants (oils, batteries, etc.), then shredding in to smallerfractions from which the metal content is extracted [4], [5].

Manuscript received August 2, 2017; revised October 17, 2017; acceptedDecember 4, 2017.

The authors are with the School of Electrical and ElectronicEngineering, The University of Manchester, M13 9PL, Uk (e-mail:michael.otoole@manchester.ac.uk, noushin.karimian@manchester.ac.uk,a.peyton@manchester.ac.uk)

Ferrous metals are the most easily recovered and are pro�tablyrecycled [5]. The non-ferrous fraction is mainly composedof aluminium alloys (approx. 78%), followed by smallerproportions of brass (12%), and copper (5%) [6], [7]. Thisfraction is much more dif�cult to sort and recover.

One technique is dense-media separation, where �ne parti-cles of magnetite or ferrosilicon are suspended in water tocreate a slurry of controlled speci�c gravity. Low densitymetals, e.g. aluminium, �oat to the surface whereas heavierones sink [8]. Maintaining these slurries, and water treatment,is expensive and has a signi�cant environmental impact [8],[9]. The method is also ineffective at sorting metals where thedifference in density is small, such as copper from brass.

Dry methods are an alternative. The most basic is to sort themetals by hand using colour [9]. This works well where thecolour is distinctive, e.g. white from red metals, but the processis dependent on low labour costs. Kutil et al. [10] demonstratedan autonomous colour sorting system, however accuracy wasdiminished by machine-vibration, variations in ambient light-ing, re�ections and other artefacts generated under industrialconditions. Multi-spectral vision systems have shown signi�-cant promise. These are capable of separating materials withsimilar colour properties, and have reported recovery rates of95% [11] and 98% [12] in the literature. However, they haveyet to be tested under industrial conditions, e.g. with dirtysamples and fast belt speeds, which will most likely degradeperformance. Other methods include laser induced breakdownspectroscopy (LIBS) and X-ray beam absorption [13]. Theseare the gold-standard in metal identi�cation but are expensive,and can be dif�cult to adapt practically for high-throughputand harsh industrial environments.

We propose a method using magnetic induction spec-troscopy (MIS) for classifying non-ferrous metals. MIS isthe change in the magnetic �eld induced in an object, inresponse to different frequency excitation magnetic �elds. Toexplain, consider a conductive sphere in free space, centredat the origin, with radius a, conductivity � and permeability� = �0 = 4� � 10�7 (free space). The sphere is illuminatedby a uniform oscillating magnetic �eld acting along an axisZ. This induces eddy-currents to circulate in the object, whichin turn induces a secondary or scattered magnetic �eld.

Let us take a point z along the Z-axis some distance outsideof the sphere (z > a). Denote Hex and Hs as the complexcomponent in the Z direction of the excitation and secondarymagnetic �eld respectively, at the point z and at frequencyf . The scattered �eld is given by the well-known analytical

2 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

expression [14], [15],

HsHex

= �3a3

z3

�1�2

+13�

cosh�� sinh�

�(1)

where � = (i2�f��)1=2a.

100 102 104 106

Rea

l Com

pone

nt

Imag

inar

y C

ompo

nent

Frequency (Hz)

Increasing conductivity

No change with conductivity

Increasing radius

Increasing radius

-0.6

-0.5

-0.4

-0.3

-0.2

-0.1

0

-0.2

-0.15

-0.1

-0.05

0

Fig. 1: The magnetic induction spectra of a sphere (Hs=Hex),with real (- -) and imaginary components (). Black linesshow the effect of changing conductivity � and grey lines theeffect of changing radius a

Figure 1 shows the real and imaginary component ofHs=Hex over a frequency range f 2

�1; 106

�Hz, for a

sphere with radius a 2 f20; 25; 40g mm and conductivity� 2 f10; 35; 100g MS/m. This plot is the magnetic inductionspectra of the sphere. The real component is shown to be asigmoid shape converging to an asymptote as f ! 1. Theimaginary component shows a curve which decreases to aminima then tend towards zero. This characteristic shape istypical of non-magnetic metal objects - not just spheres - seefor example [16], [17], [18].

Intuitively, the shape of the spectra can be explained interms of the skin-depth effect, i.e. the tendency for eddy-currents to distribute closer towards the surface as frequencyincreases. This reduces the cross-section through which currentcan �ow, and consequently increases resistance in the object.This effect becomes signi�cant when the skin-depth reachessome fraction of the characteristic dimension of the object,resulting in the minima seen on the imaginary component.Increasing the conductivity shifts the curves left as skin-depthbecomes signi�cant at lower frequencies. The asymptote ofthe real component does not change with conductivity. At highfrequencies, the eddy-currents only �ow along the surface ofthe object and thus are unaffected by the conductivity. Chang-ing the radius on the other hand does affect the asymptote.The surface geometry is altered changing the path which theeddy-currents �ow. Increasing the radius moves the asymptotedownwards in this case.

In this paper, we introduce a new method to classify thethree metals predominant in the non-ferrous fraction; copper,aluminium and brass. We use: (1) the magnetic inductionspectra of the metal fragments combined with geometricfeatures returned by a vision system, and (2) the magnetic

induction spectra alone based on an understanding of how theeddy-currents circulate at different frequencies. This work isthe �rst to demonstrate MIS as an information-rich featuresource for use in metal recovery, and �rst to propose aclassi�cation method using this new feature set based on anunderstanding of the physics of magnetic �elds. As a method,it is simple, practical, fast and scalable. It is straightforwardto implement on a commercial sorting line, and capable of thehigh-throughput operation demanded by industry.

II. METHOD

A. ShredderSort MIS + VIA System

The MIS + VIA (Vision Image Analysis) System wasdeveloped by a consortium for the EU project ShredderSort(No. 603676) to demonstrate new sorting algorithms on anindustrial-scale and under realistic operating conditions. Thesystem is shown in �gure 2. It consists of a 1 m wide conveyorwith air ejectors at the discharge-end (Regulator Cetrisa, Spainand Lenz Instruments, Spain), a high speed vision imageanalysis system (VIA) (Joanneum Research, Austria), and amagnetic induction spectroscopy sensor [19] (University ofManchester, UK). Signal processing and operation of thesystem is conducted from the Interface and Control cabinet(Lenz Instruments, Spain).

A metal sample deposited at the in-feed end of the conveyorwill �rst pass the VIA system. This uses structured-laser lightto determine size, shape, position and orientation of the testpieces and return a set of geometric features. The system canreturn up to 24 different features, however in the present work,we will only use the following most useful:

� f1: Area of the test piece (mm2).� f2: Minimum feret diameter (mm).� f3: Maximum feret diameter (mm).� f4: Mean height (mm).

The metal sample then passes over the sensitive zone of theMIS sensor. This returns a six point spectra at frequenciesaround 2, 4, 8, 16, 32 and 64 kHz. Signal processing forthe MIS system is on a NI PXI-1033 chassis with 2x PXI-7853R I/O module (National Instruments, USA) housed in theinterface and control cabinet. This system contains a numberof ADCs and DACs to output the excitation waveforms andmeasure the receive signals, and an FPGA for demodulation.To aid ongoing discussion in this paper, we will designate thedemodulated real and imaginary components of the spectra as,

Z 0p + iZ00p

where the subscript p denotes the frequency point, i.e. p 2P := f2; 4; 8; 16; 32; 64g kHz. Note that this is a relativemeasurement as it is not scaled to give true Hs=Hex. The NIPXI-1033 also reads the belt-encoder on the conveyor systemand controls the air ejectors. The latter were not used in thepresent work. Classi�cation and overall control of the MIS+ VIA system is performed on a desktop PC located in theinterface and control cabinet.

O’TOOLE et al.: CLASSIFICATION OF NON-FERROUS METALS USING MIS 3

Fig. 2: SHREDDERSORT MIS + VIA System

B. Magnetic Induction Spectroscopy (MIS) Sensor

The MIS sensor is shown in �gure 3. It consists of analuminium enclosure (750 mm � 1220 mm � 207 mm) witha stepped acetal lid, mounted underneath a conveyor belt,leaving a gap of approximately 5 mm between belt and sensor.The enclosure contains an array of sixteen solenoids, andpower and instrumentation ampli�ers situated in compartmentsfore and aft of the array. Each solenoid is used to transmit amulti-frequency excitation and detect the resultant secondarymagnetic �eld as the metal targets pass.

The cross-section of a solenoid is shown in �gure 4. Itconsists of, (1) an internal core with a coil to generatethe excitation magnetic �eld, and (2) an outer sleeve withtwo coils to detect the secondary �eld. The internal core isconstructed from an acetal tube with two 6 mm diameter ferriterods (Fair-Rite 4077276011) pressed inside. A 64-turn coilis wrapped in two sections around the tube over the ferritecores using 0.4 mm enamelled wire. This coil is driven bya power ampli�er (LT1210, Linear technologies). The outersleeve is also constructed from an acetal tube with insets cutto a 16 mm diameter. A 600-turn coil of 0.2 mm enamelledwire is wrapped around each inset. The coils are wound inopposing directions and joined to form an axial gradiometer.This can be �nely balanced using a lead-screw at the bottomof the solenoid. The gradiometer is then connected to aninstrumentation ampli�er (AD8029, Analog Devices).

The solenoids are spread evenly over a 400 mm width acrossthe centre of the conveyor in three rows approximately 90mmapart in a diagonal pattern. This creates sixteen 25 mm widechannels, with a solenoid covering each channel. To minimisepotential cross-talk between solenoids, each one transmits anexcitation at a different frequency-set to its neighbours. Thefrequency set P is shifted in the spectrum slightly to createfour different multi-sine waveforms. These are then assigned

Fig. 3: MIS Sensor

to each solenoid in such a way that no two adjacent solenoidshave the same waveform.

Each solenoid is calibrated using a ferrite cylinder (Fer-roxcube, 4B1) with diameter 10 mm and height 20 mm, inthe manner described here [20]. The ferrite is placed on theconveyor-belt coincident with the central axis of a solenoid.

4 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

Fig. 4: Solenoid Schematic

Measurements are calibrated to give Z 0p + iZ 00p = 1 overall frequencies. This calibration procedure accounts for anyphase-shifts or delays present in the system.

In what follows, we only collect data from a single solenoidof the array. A guide is deployed at the infeed end of theconveyor to guide the test pieces over the speci�c solenoid.While the solenoids are essentially duplicates, there may beslight differences in their responses due to calibration ormanufacturing errors. Con�ning our measurements to a singlesolenoid ensures the collection of the most robust and reliabledataset without loss of generality in the approach.

C. Test Pieces and Data CollectionA total of 117 test pieces were used with 36 copper pieces,

44 aluminium, and 37 brass. An example set representative ofthe test pieces is shown in �gure 5. The pieces were cut todifferent sizes and shapes from stock metals, such as bar, rod,channel, etc., provided by a local metal supplier (ManchesterMetals, UK). The metals were common grades or alloy typestypically used in manufacture. The size speci�cation for thesamples is height from 3 mm to 25 mm, and width and lengthfrom 15 mm to 50 mm.

Fig. 5: Sub-set of typical test pieces. A total of 117 test piecesare used in copper (36 pieces), aluminium (44) and brass (37).

Data on each test piece was collected using the followingprocedure: The system is activated and allowed to run for

several minutes to obtain a stable operating temperature. Theconveyor belt is set to run at 1 m/s. Each piece is assigned anumerical designation and is dropped on the conveyor belt inorder. The piece is channelled to run over a speci�c solenoidusing the guide. It passes across the structured laser of thevision system and its geometric features are obtained andrecorded. The piece then passes over the magnetic inductionsensor, where the real and imaginary components at the sixfrequencies are recorded at a 250 Hz sample-rate. This givesa spatial resolution (travel of test piece between samples) of4 mm on a 1 m/s conveyor. A single complex value for eachfrequency is obtained by taking the point where the magnitudeof the response is maximum. Any misses due to the visionsystem not detecting the object, or because it moved outsideof the channel, were marked and the measurement repeated.

D. Classi�cation

We propose classi�cation based on two different featuresets: The �rst uses the magnetic induction spectra combinedwith geometric properties of the test piece returned by thevision image analysis system. The second is simpler. It usesonly MIS, but where the features are chosen based on anunderstanding of how eddy-currents circulate in the test piece.

1) Algorithm 1, MIS + geometric properties: A simplesample matching classi�er similar to a nearest neighbourapproach is proposed. The aim is to �rst match the test-pieceto pieces in the training set with similar geometric features,producing a ‘nearest-match’ for each class. Then, the MIS ofthe test piece is compared to the nearest-matching pieces foreach class and assigned the class that has the most similarspectra. Formally:

1) Recall that Z 0p and Z 00p are the real and imaginarycomponents at frequency p 2 P of a test piece, andff1; f2; :::g its set of geometric features (see sectionII-A). Denote Ẑ 0p and Ẑ 00p and

nf̂1; f̂2; :::

oas the same

for a piece drawn from the training set.2) Normalise the geometric features (test and training) such

that they have zero mean and unity variance [21]. We usethe prime 0 to indicate the feature has been normalised,

f 0j = (fj � �j)=sj ; f̂0j = (f̂j � �j)=sj

where �j and sj is the mean and standard deviation forfeature j of all pieces in the training set.

3) Find amongst the training set of pieces belonging toclass i the �rst K pieces with the nearest matching ge-ometric features to the test piece. The nearest matchingtraining piece is the one with the smallest d,

d =

vuut4X

j=1

�f 0j � f̂ 0j

�2: (2)

We refer to the set of K nearest-matching training piecesfor class i as U iK .

4) Calculate the mean-squared difference between the realand imaginary components of the spectra for each piecein U iK and the test piece, then take the minimum. In

O’TOOLE et al.: CLASSIFICATION OF NON-FERROUS METALS USING MIS 5

this manner we create a function Ei for each class i asfollows,

Ei = minj2U iK

0

@X

p2P

(Z 0p � Ẑ 0p;j)2 + (Z 00p � Ẑ 00p;j)2

jP j

1

A (3)

where jP j is the number of measured frequencies.5) The test piece is assigned a class by evaluating Ei for

each class and,

Class := arg min fECu; EAl; EBrg

2) Algorithm 2, MIS only: To establish the rationale fora purely MIS approach, we make two observations aboutthe characteristic shape of the magnetic induction spectrafor a given non-magnetic conductive test piece following thediscussion in section I:

1) The asymptote of the real component Z 0p as p ! 1is equivalent to treating the test piece as a perfectelectrical conductor, i.e. eddy-currents only circulate onthe boundary. Therefore, Z 0p becomes independent ofconductivity as it approaches the asymptote but retainsits dependence on shape and size.

2) Any point on the curve of the imaginary component Z 00p ,that is not too close to the zero asymptote, is a functionof conductivity, and size and shape of the test piece.

We suggest that by comparing the two MIS components ofthe test piece as described, with one sensitive and the otherinsensitive to conductivity, we may account for the effectsof size and shape of the test piece and classify its materialusing only magnetic induction measurements. This idea wasinitially proposed in previous work by the authors [22]. Thiswork presented some preliminary results for a small set ofmanufactured samples (36 pieces in total) on a laboratory-based test rig. The results were generally positive, but notconclusive given the small sample set. In this work, we expandon this idea and provide more conclusive evidence with a largesample-set and the industrial-scale demonstrator described.

This is a novel approach. Multi-frequency excitation andmeasurement creates a new, information-rich set of featuresfor classi�cation despite modest changes to industrial infras-tructure. Induction sensors are already in use commercially,e.g. Steinert KSS and ISS systems (Steinert GmbH, Germany).Our work extends their use by an inventive analysis of theprocessed signals. We also identify how different parts ofthe spectra express characteristics about the object. In whatfollows, we will compare the two components described anddetermine an algorithm for separating the three metals.

III. RESULTS AND DISCUSSION

A. Design of MIS Classi�er

A typical magnetic induction spectra for one of the testobjects is shown in �gure 6. The �gure shows part of thecharacteristic pro�le described in section I: The imaginarycomponent Z 00p is converging towards zero and the real compo-nent Z 0p converging to an asymptote at around -0.33. Z 064kHzis close to the asymptote and it is therefore realistic to

Fig. 6: Spectral response of a rough aluminium cuboid, di-mensions 47 mm � 24 mm � 6.5 mm, with (- -) the realcomponent Z 0, and () the imaginary component Z 00.

assume that it will be insensitive to conductivity, satisfyingthe assumption discussed in section II-D.

Figure 7 compares the imaginary component at 2, 4, 8 and16 kHz to the real component at 64 kHz, or Z 064kHz (closeto the asymptote). Each plot shows evidence of the samplesbanding into three metal classes in order of conductivity, withthe highest conductivity (copper) banding towards the top andthe lowest (brass) towards the bottom. A measurement at 32kHz is also taken however, the imaginary component is closeto zero at this frequency and thus provides little insight.

The banding holds well for test pieces with smaller Z 064kHzbut appears to break-down, or begin to break-down, afterZ 064kHz < �1:2. This is acute in �gure 7d where both copperand aluminium samples depart from the overall trend. This isdue to poor quality of measurements, most likely clipping orsaturation in the receive electronics caused by too high gain.The coherency of the bands may therefore be improved withgreater dynamic range.

The dashed lines in �gure 7 are logistic-functions of theform,

Z 00p =k1

1 + exp (�k2Z 064kHz)+k12

(4)

where p depends on the �gure, and k1; k2 are constants thatde�ne the shape of the curve. The logistic-function providesa good description of the shape for each band using only twoshape-de�ning parameters. We propose to use equation (4) toconstruct a simple classi�er to determine the metal type of thetest pieces using only spectral information.

In what follows, we examine two logistic-function classi�ersusing Z 004kHz; Z

064kHz and Z

008kHz; Z

064kHz as features. These

two appear to show the clearest banding overall in �gure 7.De�ne a function for each class i 2 fCu;Al;Brg,

Ci =ki;1

1 + exp (�ki;2Z 064kHz)+ki;12� Z 008kHz (5)

The term Ci may be considered a distance measure. Any objectwhich falls on the line de�ned by the logistic-function results

6 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

(a) (b)

(c) (d)

Fig. 7: Comparison of the imaginary component of the spectra Z 00 at frequencies 2, 4, 8 and 16 kHz with the real componentZ 0 at 64 kHz.

in Ci = 0. A test piece is then assigned to the class with thesmallest Ci. Formally,

Class := arg min fCCu; CAl; CBrg

A classi�er for Z 004kHz; Z064kHz is simply obtained following

the same procedure with Z 008kHz replaced with Z004kHz .

The choice of frequency for the imaginary component Z 00is a dif�cult problem which merits some discussion. Returningto the spectra of a sphere in �gure 1, we see evidence of twodifferent electromagnetic effects. The �rst effect shows Z 00 de-creasing towards a minima as frequency increases. This is dueto the relation between frequency and �eld strength predictedby Maxwell’s equations for a linear media. The second effectshows Z 00 increasing and tending towards zero. This is dueto the skin-depth effect, where the eddy-currents progressivelypenetrate less of the object as frequency increases. The minimain the spectra is a point of transition, where one effect becomesdominant over the other.

We desire to take measurements where the same effect isdominant across all the test pieces. Otherwise, the differenteffects may follow different trends when comparing withZ 064kHz . In the present work, measurements are mostly takenfrom where the skin-depth is dominant - as indicated by �gure6. We also desire measurements with a good signal-to-noiseratio. If the frequency for Z 00 is too high, the measurement willbe small and suffer adversely from noise. A suitable frequencyfor Z 00 may therefore be chosen by the following heuristics:

1) A frequency suf�ciently high such that the skin-deptheffect is dominant for Z 00 of the least conductive testpiece with the smallest single dimension.

2) A frequency suf�ciently low such that Z 00 of the mostconductive test pieces can be measured with a goodsignal-to-noise ratio.

How realistic is it to apply this heuristic in scrap metalsorting? Shredded waste is routinely sorted into size, andthe conductivity of the target metals is known. This means

O’TOOLE et al.: CLASSIFICATION OF NON-FERROUS METALS USING MIS 7

a reasonable estimate can be made for the frequency of Z 00which will hold over the fraction to be sorted provided thesize criteria of the fraction is suf�ciently speci�c.

B. Classi�cation PerformanceThe Leave One Out (LOO) method is used to generate

performance statistics on the three classi�cation algorithmsdescribed - the two logistic-function based methods, usingdifferent frequency components as input features, and thesample matching classi�er outlined in section II-D. The LOOmethod works by removing a single test piece from the dataset, then use the remaining pieces to train the classi�er. Theexcluded test piece is then assigned a class, using the newlytrained classi�ers, and returned to the data set. A second testpiece is then removed and the process repeated until each piecehas been classi�ed.

Analysis and data processing is performed in Matlab(Mathworks Inc, USA), with training of the logistic-functions(parameters ki;1; ki;2) using Matlab’s fminsearch optimisa-tion function. Some outliers are excluded when training thelogistic-functions as they signi�cantly skew the best-�t ap-proximation. However, they are retained as test pieces and theirclassi�cation, or mis-classi�cation in this case, is includedin the overall statistics. No outliers are excluded from thematching classi�er’s training sets. The outliers are chosen asany aluminium piece with Z 064kHz < �1:2 and any brass piecewith (Z 004kHz < �0:35) _ (Z

008kHz < �0:25). This excludes a

total of six pieces from the training sets (5 Al., 1 Br.).Classi�cation results are shown in table I. Table headings

are de�ned according to Soklova et al. [23] and are included inthe appendix. Overall the results appear strong, with averageaccuracies for the MIS-only methods close to 95%, witharound 92% precision and recall. The matching classi�ershows relatively poor performance with average accuracy,precision and recall of 83.5%, 75.2% and 75.2% respectively.This is a surprising result. The classi�er works using a twostage matching process: First, a number of training pieces withsimilar shapes and sizes to the test piece are selected fromeach class. The test piece is then assigned the class of theselected piece with the most similar MIS. It was expected thatthis approach would outperform the MIS-only method, as itcompares similar pieces, which should have similar MIS.

C. Geometric Paramaters vs. Z 064kHzA possible cause of the inferior results for the matching

classi�er is that the representation of the geometry of thetest pieces as features (f1; f2; f3; f4) is overly reductive. Thematching classi�er is essentially making a prediction aboutthe magnetic induction spectra for a given set of geometricfeatures and an assumed conductivity. For a test piece withcertain geometric features, we are assuming the spectra hasone form if it is copper, and another if it is brass. However, iftoo much information is lost in condensing the geometry to asmall set of scalar values, then correlation between shape andmagnetic response will be poor, and we can no longer makemeaningful predictions. Other possible causes are the quantityof training data being too sparse to populate the feature space,

Fig. 8: Comparison of the imaginary component of the spectraat 4 kHz (Z 004kHz) compared to f1, the area of the test piece.

or small variations in the path as pieces pass over the solenoid.The same piece travelling close to the solenoid axis wouldhave a slightly different magnetic response than if it wereto traverse off-centre. This might be enough to degrade theperformance of the classi�er, especially for smaller pieces.

It is notable that using Z 064kHz as a proxy for geometry inthe MIS only classi�ers outperforms actual geometric features,as used in the matching classi�er. For instance, let us compareZ 004kHz (a function of both geometry and conductivity) againsta geometric feature. This is shown in �gure 8, using f1 orarea of the test pieces as returned by the vision system. Thereis weak banding of the different metal types, with copperpredominately at the top, aluminium in the middle, and brassmostly at the bottom of the �gure. It is considerably morescattered than �gure 7. The other features f2; f3; f4 showedsimilar or worse results. Their plots are omitted for the sake ofbrevity. We posit that Z 064kHz better accounts for the variablegeometry of the test pieces, or variations in orientation ortravel path as the piece passes over the solenoid, as it is itselfa product of the eddy-current and magnetic �eld interactionstaking place in the test piece.

D. Comparison with other Classi�ersThe performance of the classi�ers compares favourably to

the published literature. Recovery-rates (equivalent of recall)are typically in the range of 90-97% for X-ray techniques, 80-97% for LIBS and 86-95% for optical methods and [13]. X-rayand LIBS sorting is noted for being expensive, in contrast toinduction sensors, and are dif�cult to justify economically.

Kutila et al. [10] reported average precision of 83.5% usinga chromatic plus induction method to sort white (aluminium)from red metals (copper and brass). The precision for sep-arating the same two classes using the MIS-only methodwas approximately 91%. This was under broadly similarexperimental conditions with the following exceptions: (1)Kutila et al. used real scrap metal, as opposed to manufacturedtest pieces in the present work, (2) a larger mass of waste was

8 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

Classi�er Features i Tpi Fni Tni Fpi Acc.i Prec.i Rec.i Acc.� Prec.� Rec.�

Logistic function Z 064kHz , Z004kHz

Cu 36 0 75 6 0.9487 0.8571 1.00000.9487 0.9231 0.9231Al 37 7 71 2 0.9231 0.9487 0.8409

Br 35 2 79 1 0.9744 0.9722 0.9459

Logistic function Z 064kHz , Z008kHz

Cu 36 0 75 6 0.9487 0.8571 1.00000.9430 0.9145 0.9145Al 37 7 70 3 0.9145 0.9250 0.8409

Br 34 3 79 1 0.9658 0.9714 0.9189

Sample MatchingZ 0, Z 00 at all freqs.,

f 1 ; f 2 ; f 3 ; f 4 ;K = 5 (i.e. size of U iK )

Cu 26 10 72 9 0.8376 0.7429 0.72220.8348 0.7521 0.7521Al 31 13 57 16 0.7521 0.6596 0.7045

Br 31 6 76 4 0.9145 0.8857 0.8378

TABLE I: Classi�cation statistics

tested (over 4000 kg), and (3) a faster conveyor speed of 1.5m/s compared to 1 m/s in this work. The use of manufacturedrather than real scrap test pieces may arti�cially in�ate theperformance statistics of the MIS classi�ers, however, thelatter two differences are unlikely to make much difference.Overall, Kutila et al. concluded that induction was not muchuse in sorting these metal types, however we �nd otherwisewhen using a multi-frequency approach.

Hyper-spectral techniques have shown high recall around90-97% [13]. Picon et al [12] reported an average recall of97.63% for separating brass, copper and aluminium. Candianiet al. [11] reported 92-100% for the same metals. In compar-ison, we report average recall around 92% for the MIS-onlyclassi�ers. The main differences in experimental conditionswere: (1) The hyper-spectral studies used a small set of hand-selected real scrap. (2) In Candiani et al., the samples weresigni�cantly smaller - less than 2 mm in size, and (3) thehyper-spectral studies used much slower belt-speeds - 0.33m/s (Picon et al.) and 0.067 m/s (Candiani et al.) versus 1m/s in the present study. Commercial belts typically operatein excess of 2.5 m/s, hence higher belt-speeds are needed forrealistic comparison.

IV. CONCLUSION

We have demonstrated using magnetic induction spec-troscopy and simple geometric features to classify the mainnon-ferrous metals found in end-of-life waste (copper, alu-minium and brass). Our method has the advantages of be-ing fast and scalable with relatively modest measurementhardware. The performance of the MIS-only classi�ers wereparticularly promising, with around 92% precision and recall(equivalent to purity and recovery-rate). This is within industryprescribed thresholds and broadly competitive with reportedstatistics for comparable systems.

A surprising result was the relatively low performance of thecombined vision and MIS classi�er. We speculate that simplegeometric features (area, height, max./min. feret diameter)fail to capture the information required to make meaningfulpredictions about the scattered magnetic �eld - illustratedby poor banding between metals shown in �gure 8. Thecomponent Z 064kHz was found to be a superior geometricfeature in this respect. We do not suggest that vision datahas no value, rather that its interpretation demands a morethorough treatment beyond the scope of the present work.

Induction sensors are distinctly applicable to metal recy-cling: They are low-cost, reliable, practical, and robust. Theyare unaffected by variable lighting or dirty samples, andresistant to harsh environments. However, their performanceis dependent on a degree of homogeneity in the geometry ofthe metal fragments; something dif�cult to achieve in practice.

Our main conclusion is that an effective non-ferrous metalsorting system can be constructed using multi-frequency in-duction sensors and a simple, straightforward classi�cationalgorithm. The most novel aspect is the use of the magneticinduction spectra and the selection of frequency componentsas features, based on the physics of how induced eddy-currentscirculate in metal objects. Two frequency components arechosen as features: (1) A high frequency component, wheremagnetic �eld penetration is negligible, to model the geometryof the metal fragments, and (2) a low frequency component,where skin-depth is signi�cant, which is a function of both ge-ometry and conductivity of the fragments. We have shown that,provided these key criteria are satis�ed, good performancerates can be achieved, and a simple but effective classi�ercan be created distinctly well-suited to industry needs.

ACKNOWLEDGMENTThis research was conducted as part of the EU FP7 funded

project Shreddersort, under grant agreement no. 603676. Wewould especially like to acknowledge Alfred Rinhofer andMalte Jaschik of Joanneum Research for their on-going sup-port with the vision system, and Lenz Instruments, in particularMark Williams, for integration of the industrial prototype.

APPENDIXThe headings for table I, with C := fCu;Al;Brg:� Tpi; Fni; Tni; Fpi: Number of true-positives, false-

negatives, true-negatives, and false-positives for class i.� Acci: Accuracy for class i, i.e. proportion of pieces in

class i, correctly assigned class i (Tp) or not class i (Tn),

Acci := (Tpi + Tni)=(Tpi + Fni + Fpi + Tni)

� Preci: Precision for class i, i.e. the proportion of piecesassigned class i that actually belong to that class,

Preci := Tpi=(Tpi + Fpi)

� Reci: Recall or recovery-rate for class i, i.e. the propor-tion of pieces in class i correctly assigned to that class,

Reci := Tpi=(Tpi + Fni)

O’TOOLE et al.: CLASSIFICATION OF NON-FERROUS METALS USING MIS 9

� Acc�: Average accuracy,Pi2C Acci=3

� Prec�: Av. precision,Pi2C Tpi=

Pi2C (Tpi + Fpi)

� Rec�: Av. recall,Pi2C Tpi=

Pi2C (Tpi + Fni)

REFERENCES[1] S.-i. Sakai et al., An international comparative study of end-of-life

vehicle (elv) recycling systems, Journal of Material Cycles and WasteManagement, vol. 16, no. 1, pp. 120, Feb 2014.

[2] Directive 2000/53/EC of the European Parliament and of the Councilof 18 September 2000 on End-of-Life vehicles, Of�cial Journal of theEuropean Communities, vol. 43, 2000.

[3] Eurostat, 2014 End-of-life vehicles - reuse, recycling and recovery ofELVS, Last update 11/09/2017.

[4] O. Forton, M. Harder, and N. Moles, Value from shredder waste:Ongoing limitations in the uk, Resources, Conservation and Recycling,vol. 46, no. 1, pp. 104 113, 2006.

[5] A. Santini, L. Morselli, F. Passarini, I. Vassura, S. D. Carlo, andF. Bonino, End-of-life vehicles management: Italian material andenergy recovery ef�ciency, Waste Management, vol. 31, no. 3, pp. 489 494, 2011.

[6] Inst. of Scrap Recycling Industries, ISRI Scrap Speci�cations, Circular1994: Guidelines for Nonferrous Scrap, Ferrous Scrap, Glass Cullet,Paper Stock, Plastic Scrap, 1994.

[7] F. Margarido, R. N. Santos, F. Durao, C. Guimaraes, C. Nogueira,P. C. Oliveira, F. Pedrosa, and A. M. Gonc‚alves, Separation of non-ferrous fractions of shredded end-of-life vehicles for valorising itsalloys, in MMME’14-International Conference on Mining, Material andMetallurgical Engineering, 2014.

[8] W. L. Dalmijn and T. P. R. De Jong, The development of vehiclerecycling in europe: Sorting, shredding, and separation, JOM, vol. 59,no. 11, pp. 5256, Nov 2007.

[9] G. Gaustad, E. Olivetti, and R. Kirchain, Improving aluminium re-cycling: A survey of sorting and impurity removal technologies, Re-sources, Conservation and Recycling, vol. 58, pp. 79 87, 2012.

[10] M. Kutila, J. Viitanen, and A. Vattulainen, Scrap metal sorting withcolour vision and inductive sensor array, in International Conferenceon Computational Intelligence for Modelling, Control and Automationand International Conference on Intelligent Agents, Web Technologiesand Internet Commerce (CIMCA-IAWTIC’06), vol. 2, Nov 2005, pp.725729.

[11] G. Candiani, N. Picone, L. Pompilio, M. Pepe, and M. Colledani,Characterization of �ne metal particles derived from shredded weeeusing a hyperspectral image system: Preliminary results, Sensors,vol. 17, no. 5, 2017.

[12] A. Picon, O. Ghita, P. F. Whelan, and P. M. Iriondo, Fuzzy spectral andspatial feature integration for classi�cation of nonferrous materials inhyperspectral data, IEEE Transactions on Industrial Informatics, vol. 5,no. 4, pp. 483494, Nov 2009.

[13] S. P. Gundupalli, S. Hait, and A. Thakur, A review on automatedsorting of source-separated municipal solid waste for recycling, WasteManagement, vol. 60, pp. 56 74, 2017.

[14] J. R. Wait, A conducting permeable sphere in the presence of a coilcarrying an oscillating current, Canadian Journal of Physics, vol. 31,no. 4, pp. 670678, 1953.

[15] K. Sun, K. O’Neill, F. Shubitidze, S. A. Haider, and K. D. Paulsen,Simulation of electromagnetic induction scattering from targets withnegligible to moderate penetration by primary �elds, IEEE Trans. onGeoscience and Remote Sensing, vol. 40, no. 4, pp. 910927, Apr 2002.

[16] J. R. Wait, Some solutions for electromagnetic problems involvingspheroidal, spherical, and cylindrical bodies, Res. NBS (Math. andMathematical Physics) B, vol. 64, pp. 1532, 1959.

[17] M. D. O’Toole, J. L. Davidson, L. A. Marsh, W. Yin, and A. J. Peyton,Evaluation of the thin-skin approximation boundary element method forelectromagnetic induction scattering problems, in 2016 IEEE SensorsApplications Symposium (SAS), April 2016, pp. 16.

[18] O. A. Abdel-Rehim, J. L. Davidson, L. A. Marsh, M. D. OToole, andA. J. Peyton, Magnetic polarizability tensor spectroscopy for low metalanti-personnel mine surrogates, IEEE Sensors Journal, vol. 16, no. 10,pp. 37753783, May 2016.

[19] Uni. of Manchester, Method and apparatus for determining conductiv-ity, UK Patent Application Number GB1 706 951.9.

[20] M. D. O’Toole, L. A. Marsh, J. L. Davidson, Y. M. Tan, D. W. Armitage,and A. J. Peyton, Non-contact multi-frequency magnetic inductionspectroscopy system for industrial-scale bio-impedance measurement,Measurement Science and Technology, vol. 26, no. 3, p. 035102, 2015.

[21] A. K. Jain and R. C. Dubes, Algorithms for Clustering Data. UpperSaddle River, NJ, USA: Prentice-Hall, Inc., 1988.

[22] M. D. O’Toole, N. Karimian, and A. J. Peyton, Fast classi�cationof non-magnetic metal targets using eddy-current based impedancespectroscopy - to be published, in 2017 IEEE SENSORS, Oct 2017.

[23] M. Sokolova and G. Lapalme, A systematic analysis of performancemeasures for classi�cation tasks, Information Processing & Manage-ment, vol. 45, no. 4, pp. 427 437, 2009.

Michael D O’Toole received the M.Eng. (Hons.)degree in integrated engineering from the Universityof Reading in 2006, and the Ph.D. degree from theWolfson School of Mechanical and ManufacturingEngineering, Loughborough University, in 2011. Hewas a Research Associate at The University ofManchester from 2011 to 2016, �rst working in theSchool of Computer Science on simulation and anal-ysis tools for nonsmooth mechanical systems thenin School of Electrical and Electronic Engineeringfrom 2012 working on magnetic induction systems

for non-destructive inspection and characterisation.In 2016, he was awarded a Leverhulme Trust Early Career Research

Fellowship. He is author and co-author of over 20 scienti�c publicationsand has recently �led his �rst patent. His research interests include signalprocessing, and sensor and instrumentation design, with a particular emphasistoward bioimpedance spectroscopy and magnetic induction systems for non-destructive testing.

Noushin Karimian received her Ph.D. degree inElectrical and Electronic Engineering from the Uni-versity of Manchester in 2014. Since then she hasbeen appointed as a Research Associate at this uni-versity on various international projects. Her mainresearch interests are in Electromagnetic sensors forNDT applications and digital signal processing. Shehas been actively involved both as technical andorganisational committee member of internationalconferences, and most recently the European Mi-crowave Conference. She is co-author of over 40

international conference and journal papers and has a registered patent.Dr Karimian is a member of IET, EuMA, BINDT and IOP. She is currently

acting as the Treasurer for the IEEE UK & Ireland Section, and a committeemember of the University of Manchester IEEE Student Branch. She is also acommittee member of the IEEE UK & Ireland Women in Engineering (WiE).

Anthony J Peyton received the B.Sc. degree inelectrical engineering and electronics and the Ph.D.degree in medical instrumentation from UMIST in1983 and 1986, respectively. He was appointed Prin-cipal Engineer at Kratos Analytical Ltd. in 1989,developing precision electronic instrumentation sys-tems for magnetic sector and quadrupole mass spec-trometers, from which an interest in electromagneticinstrumentation developed. He returned to UMIST asa Lecturer and worked with the Process TomographyGroup. He moved to Lancaster University in 1996

as Senior Lecturer. He was promoted to Reader in electronic instrumentationin 2001 and Professor in May 2004. Since December 2004, he has been aProfessor of Electromagnetic Tomography Engineering with the Universityof Manchester. His main research interests currently are in the area ofinstrumentation, applied sensor systems, and electromagnetics. He has been aPrincipal Investigator of numerous national and industry funded projects anda partner of ten previous EU projects, one as a Coordinator. He has been a co-author of over 110 international journal papers, two books, several hundredconference papers, and 12 patents in areas related to electromagnetics andtomography.