Jaepil LEE, Sungho SHIN, Inseok JANG, Seongguk BAE, Sungho JEONG. Method to improve the classification accuracy by in situ laser cleaning of painted metal scraps during laser-induced breakdown spectroscopy based sorting[J]. Plasma Science and Technology, 2025, 27(3): 035502. DOI: 10.1088/2058-6272/ad9bfd
Citation:
Jaepil LEE, Sungho SHIN, Inseok JANG, Seongguk BAE, Sungho JEONG. Method to improve the classification accuracy by in situ laser cleaning of painted metal scraps during laser-induced breakdown spectroscopy based sorting[J]. Plasma Science and Technology, 2025, 27(3): 035502. DOI: 10.1088/2058-6272/ad9bfd
Jaepil LEE, Sungho SHIN, Inseok JANG, Seongguk BAE, Sungho JEONG. Method to improve the classification accuracy by in situ laser cleaning of painted metal scraps during laser-induced breakdown spectroscopy based sorting[J]. Plasma Science and Technology, 2025, 27(3): 035502. DOI: 10.1088/2058-6272/ad9bfd
Citation:
Jaepil LEE, Sungho SHIN, Inseok JANG, Seongguk BAE, Sungho JEONG. Method to improve the classification accuracy by in situ laser cleaning of painted metal scraps during laser-induced breakdown spectroscopy based sorting[J]. Plasma Science and Technology, 2025, 27(3): 035502. DOI: 10.1088/2058-6272/ad9bfd
Method to improve the classification accuracy by in situ laser cleaning of painted metal scraps during laser-induced breakdown spectroscopy based sorting
School of Mechanical and Robotics Engineering, Gwangju Institute of Science and Technology, 123 Cheomdangwagi-ro, Buk-gu, Gwangju 61005, Republic of Korea
Scrap metals are typically covered with surface contaminants, such as paint, dust, and rust, which can significantly affect the emission spectrum during laser-induced breakdown spectroscopy (LIBS) based sorting. In this study, the effects of paint layers on metal surfaces during LIBS classification were investigated. LIBS spectra were collected from metal surfaces painted with black and white paints by ablation with a nanosecond pulsed laser (wavelength = 1064 nm, pulse width = 7 ns). For the black-painted samples, the LIBS spectra showed a broad background emission, emission lines unrelated to the target metals, large shot-to-shot variation, and a relatively low signal intensity of the target metal, causing poor classification accuracy even at high shot numbers. Cleaning the black paint layer by ablating over a wide area prior to LIBS analysis resulted in high classification accuracy with fewer shot numbers. A method to determine the number of cleaning shots necessary to obtain high classification accuracy and high throughput is proposed on the basis of the change in LIBS signal intensity during cleaning shots. For the white-painted samples, the paint peeled off the metal surface after the first shot, and strong LIBS signals were measured after the following shot, which were attributed to the nanoparticles generated by the ablation of the paint, allowing an accurate classification after only two shots. The results demonstrate that different approaches must be employed depending on the paint color to achieve high classification accuracy with fewer shot numbers.
Metal recycling is an efficient solution for reducing energy consumption and greenhouse gas emissions during metal production. For example, the energy consumption for aluminium production can be reduced by 95% when recycled sources are utilized, compared to when the production is based on primary sources [1]. Because recyclable metals are collected as a mixture of various metals, a crucial step in metal recycling is sorting the metals by type. However, sorting metal scraps at high speeds with high classification accuracy is a technical challenge. Typically, ferrous and nonferrous metals are separated using magnetic sorters. The sorting of mixed nonferrous metals is more challenging, and the classification accuracy of existing techniques such as eddy current and X-ray fluorescence (XRF) is unsatisfactory. For example, eddy current separation is significantly influenced by the size, shape, and electrical conductivity of the target metals [2–5], whereas XRF exhibits lower accuracy in measuring light elements such as Mg, Al, and Si [6–8].
In recent years, the sorting of nonferrous metals using laser-induced breakdown spectroscopy (LIBS) has received considerable attention in the field of recycling owing to its high classification accuracy. In LIBS-based sorting, the type of target metal is identified by analyzing the chemical composition, which allows the classification of not only different metals but also metal alloys of similar compositions with high accuracy. In addition, LIBS measurements can be performed with little sample preparation in a short measurement time (on the order of milliseconds) for almost all elements. For LIBS-based metal sorting, Fugane et al [9] reported that commercial aluminium alloys could be sorted with a classification accuracy comparable to that of inductively coupled plasma-optical emission spectrometry when the LIBS signal intensity of the major constituent elements was used for differentiation. Chen et al [10] reported that the classification accuracy of 13 different brands of steel samples exceeded 98% when LIBS technology coupled with a machine learning algorithm was used. Aberkane et al [11] compared four different classification algorithms for LIBS-based sorting of six classes of Zamak alloys (zinc-based alloys) containing 1% copper with varying zinc and aluminium concentration ratios and achieved a high classification accuracy in the range of 92%–100%. In addition to metals, the application of LIBS has been investigated for the classification of other materials, such as plastics [12–14], building materials [15, 16], rare earth elements [17, 18], and electronic wastes [19].
Although the classification of scrap metals with high accuracy is possible using LIBS, the actual implementation of this technique requires careful consideration of several issues to achieve the desired sorting accuracy. To ensure high-accuracy sorting, the irregularity and nonhomogeneity in the shape and size of the scrap metals must be considered. In an industrial-scale LIBS sorting system, scrap metals are often irradiated by a laser beam while moving on a continuously moving system such as a conveyer belt [20]. Because scrap metals are usually of various shapes and are fed to the moving system randomly, the laser beam is not necessarily incident in the direction normal to the sample surface. If the incidence angle is too large, the LIBS signal intensity and signal-to-noise ratio fall abruptly, as reported previously [21, 22]. Alternatively, the sample surface can be outside the focal region of the laser beam if the scrap is too large or too small, which can also cause a sudden drop in the signal intensity. For scrap metals with varying heights, an autofocusing system can be adopted to focus on the sample surface [23]. If the shape and size of the scrap metals change randomly, three-dimensional (3D) sensing and auto-focusing can be used to select an area on the surface at which the laser irradiation will be at a low incidence angle and in the focal region of the laser beam [20, 24].
Another important problem with using LIBS sorting is the presence of contaminants on the target surfaces. Scrap metals are typically covered with various types of contaminants, such as paint, rust, dust, oil, etc. As only the surface layer of a target is ablated and analyzed in LIBS measurements, the presence of contaminants can produce completely different spectra or even no signal from the bulk material if the target surface is heavily contaminated [25]. For example, by monitoring the LIBS spectra of ancient bronze coins for increasing number of laser pulses, Fortes et al [26] reported that the signals of Si, Ca, Mg, and Fe, which were not the constituent elements of bronze, were detected in the beginning; however, the signals of Cu, Sn, and Pb belonging to bronze could be clearly detected as the pulse number increased. On the other hand, for the LIBS measurement of a coated aluminium alloy (30 μm epoxy primer + 60 μm polyurethane topcoat), Yang et al [27] reported that 65 shots of a pulsed nitrogen laser (wavelength = 337.1 nm, pulse width = 3 ns, pulse energy = 90 μJ) were required before the aluminium signal from the substrate was clearly detected. Although the contaminant layer on a target surface could, in principle, be penetrated by firing multiple laser pulses, the number of laser shots that can be practically fired on a piece of moving scrap in an industrial sorting system is limited and depends on the laser parameters and moving speed of the sample. For an industrial sorting system, it is necessary to identify the metal type with the minimum number of shots to achieve high productivity. In addition, from an at-line LIBS measurement of cast steel, Ruiz et al [28] reported that the signals from contaminants were not only detected, but they also interfered with the analytically valuable target signals, causing an increase in the relative standard deviation (RSD).
Several methods have been proposed to address the issues of surface contamination in LIBS-based classification. The most direct approach is mechanically and/or chemically cleaning the surface of scrap metals. For example, Van den Eynde et al [29] cleaned the surface to be irradiated using a Dremel and ethanol for aluminium sorting, which could be time-consuming and unapplicable for automated sorting. For an automated system, Horckmans et al [30] adopted two burst pulses where the first burst cleaned the surface while the second burst produced LIBS signals during refractory sorting. Alternatively, for the classification of nine different types of shredded metals of various sizes and shapes, which were collected from a recycling facility and covered with dirt, Merk et al [31] applied a filter algorithm that rejected a LIBS signal whose intensity was not sufficiently higher than the noise level and treated the target metal scrap as trash (that is, not recycled). Although the classification accuracy of the sorted scrap metals could be increased by filtering, as explained in their research, this also resulted in a significant portion of the samples being classified as trash and not recycled, reaching as high as 39%, depending on the metal type. Park et al [24] proposed an algorithm that suggests a relatively flat (i.e., low incidence angle) and bright (i.e., less contaminated) surface for laser irradiation by measuring the 3D shape and color of each piece of scrap with an optical module. The authors reported that when LIBS analysis was performed by irradiating the algorithm-suggested surfaces of scrap metals, the classificaion accuracy improved by 24.5% compared to when the algorithm was not used. Similarly, Díaz-Romero et al [32] fused LIBS spectra with camera data (3D and RGB images) using a deep-learning model and demonstrated that mixed aluminium scraps could be classified with high accuracy.
The aforementioned methods of measuring the shape and color of scrap metals are effective only when the scrap surfaces have relatively clean areas and the thickness of contaminants is rather thin. However, if these surfaces are uniformly covered with contaminants of finite thickness and the same color, these methods would not work. One example of such a contaminant is paint. A significant portion of scrap metal originates from automobiles or home appliances, which are mostly painted. To sort painted scrap metals using LIBS, the paint layer must be removed first to identify the type of underlying metal. Recently, it was reported that a handheld LIBS analyzer equipped with a cleaning mode was used to sort painted metal scraps, where the effectiveness significantly depended on the operator [25], but the effects of paint layer on LIBS sorting of painted metal scraps are little investigated. In this work, we study the LIBS sorting of painted metals. The signal characteristics of the paint layer and the difference between black and white paints are investigated. A method to determine whether the paint layer is sufficiently removed by cleaning shots is proposed to accomplish highly accurate LIBS classification of painted scrap metals with high throughput.
2.
Experiment
2.1
Samples
Aluminium, copper, and stainless steel were selected as the target metals for sorting. To examine the effects of paint during LIBS measurements, aluminium, copper, and stainless-steel plates purchased from a local metal shop were used as the samples, for which the elemental compositions are shown in table 1. For each metal, the plate was cut into ten blocks with dimensions of 5 cm × 5 cm × 1 cm. Before painting, the sample surfaces were polished using a polishing machine (LUXPOL; R&B, Korea) with #600, #800, and #1,000 SiC sandpaper for 10, 10, and 20 min, respectively.
Table
1.
Elemental composition of the test metal samples.
Among the 10 blocks of each metal, 5 blocks were painted black, whereas the remaining 5 blocks were painted white using spray paint (Il Shin Co., Korea). The constituent elements of the black and white paints are listed in table 2. The samples were painted by spraying them with paint at approximately 30 cm from the sample surface to obtain a uniform coating thickness. Prior to spraying, a narrow tape (approximately 4 mm in width) was attached to the sample surface to measure the thickness of the coated paint. The paint thickness was measured using a confocal microscope (OLS4000-ASG, Olympus, Japan) on a step produced by removing the tape after the paint had completely dried. In addition, the samples were coated to two different thicknesses by changing the spraying time (ts): 5–7 s for six samples (three black and three white) and approximately 10 s for four samples (two black and two white). Figure 1 shows the metal samples before and after painting.
Figure
1.
Metal sample blocks (a) before and (b) after painting.
Figure 2 shows a schematic of the experimental setup for the LIBS measurement. A Q-switched Nd:YAG laser (wavelength = 1064 nm, pulse duration = 7 ns, SL II-10; Continuum, USA) was used to irradiate the target through a lens set-up consisting of a bi-concave lens (f = −75 mm) and a doublet lens (f = 100 mm) for focus control. The back focal length of the lens set-up was calculated to be 780 mm. The intensity of white paint signals was saturated when the laser pulse energy exceeded 19 mJ. Thus, the laser pulse energy was set to 16.3 mJ because the RSD of intensity was the lowest at this value. The plasma emission from the irradiated sample was collected using two plano-convex lenses (f = 50 mm) positioned at an angle of 45° from the incident laser beam and detected using a two-channel spectrometer (Avaspec-ULS2048-2, Avantes, Netherlands). The spectral range of the first channel of the spectrometer was 200–430 nm (resolution: 0.1 nm), and that of the second channel was 415–900 nm (resolution: 0.3 nm). The gate width and delay of the spectrometer were fixed at 1.05 ms and 1.28 μs, respectively.
Figure
2.
Schematic of the experimental setup for LIBS.
The LIBS spectra of a painted target metal were measured in two ways to examine the effect of the paint layer. In the first method, a painted sample was placed at the focal plane of the focusing optics and irradiated repeatedly until the metal signal was detected. At the focus, the spot size estimated by the Gaussian beam equation using the measured crater diameter and laser pulse energy [33] was about 340 μm for which the corresponding fluence became about 17.9 J/cm2. In the second method, the painted sample was first irradiated at a height of 280 mm above the focal plane to remove the paint layer over a wide area (cleaning shots) and then lowered to the focal plane and irradiated again to identify the metal type (analysis shots). The spot size estimated at the cleaning shot position using the Gaussian equation and the corresponding laser fluence were about 600 μm and 5.76 J/cm2, respectively.
3.
Results and discussion
3.1
Effect of paint layer on LIBS measurement
3.1.1
Black paint
The difference between the LIBS spectrum of a bare surface and that of a black-painted surface (ts = 5–7 s, coating thickness ≈ 20 μm) of the aluminium sample for different numbers of shots is shown in figure 3. Each spectrum in figure 3 represents the average of the LIBS spectra measured at three different points for the corresponding number of laser shots. Both samples were placed in the focal plane to obtain these spectra. The LIBS spectra of the bare surface in figure 3(a) show clear Al signals with little noise, and the intensity of each peak remains nearly the same regardless of the shot number.
Figure
3.
Averaged LIBS spectra of the (a) bare and (b) black-painted surfaces for varying number of shots (ts = 5–7 s).
However, the LIBS spectrum of the painted surface in figure 3(b) shows several differences from that of the bare surface. First, the emission from the paint, the first shot signal in figure 3(b), was very weak. An examination of the charge-coupled device (CCD) image of the irradiated surface confirmed that the metal surface was not exposed after the first shot, implying that the emission was caused by paint ablation. Second, a broad background emission appeared in the 400–900 nm range. The continuous emission generally disappears with a sufficiently long gate delay during LIBS measurement of metals as in the bare aluminium case in figure 3(a). Thus, this broad emission is considered to be thermal radiation from the paint. A similar continuous emission was also observed in the LIBS signal of white varnish on aluminium alloy for which the gate delay was 2.5 μs [34] and also during the laser removal of paint from aircraft skin [35]. Finally, the emission lines of elements other than aluminium (C2 516.5 nm, Na I 588.9 and 589.5 nm, Li I 670.7 nm, and K I 766.4 and 769.8 nm) were detected.
Note that Na, Li, and K were not constituent elements of the black paint. The detection of Na signal during LIBS measurement of paint was also reported by other researchers, which was attributed to contamination during sample handling [36]. To check whether these elements were actually contained in the paint or not, although they were not in the nominal composition of the paint provided by the manufacturer, black paint coated on a slide glass was measured by X-ray photon spectroscopy (XPS; NEXSA, Thermo Fisher Scientific, USA). The XPS results in table 3 showed that a small amount of Na and Li were indeed contained in the paint. It is considered that they were included in the paint during manufacturing or sample preparation processes. Thus, they are denoted as ‘paint signals’ hereafter. For the painted surface, the aluminium signal (Al I, 396.1 nm) was not detected until the third shot, at which point a weak signal was observed, as shown in figure 3(b). The appearance of a weak Al signal after the third shot is thought to imply that the paint layer was penetrated by the prior laser shots and the Al surface was ablated, at least partially. Strong paint signals (Na I, Li I, K I, etc.) were detected even after the paint layer had been penetrated. The reason for these persistently strong paint signals could be due to the ablation of the sidewall of the penetration hole in the paint layer, as illustrated in figure 4(a), and/or the residual paint on the ablated surface. To check the possibility of sidewall ablation, the sample was first elevated to the cleaning-shot position in figure 2 and irradiated over an area wider (spot size ≈ 600 μm) than the focal spot diameter (spot size ≈ 340 μm). The laser pulse energy for the cleaning shot was the same as that of the analysis shot. The cleaned sample was then lowered to the focal plane and re-irradiated for the LIBS analysis of the bulk metal composition, as shown in figure 4(b). The ablation patterns and signal intensities of the painted samples with and without the cleaning shot showed significant differences. Figure 5(a) shows the CCD images of the black-painted aluminium surface (ts = 5–7 s) irradiated with an increasing number of analysis shots (ASs) only (no-cleaning shot case) and with an increasing number of cleaning shots (CSs) as well as analysis shots. The CCD images in figure 5(a) show that the irradiated spot appeared bright after two laser shots (two analysis shots, one cleaning shot and one analysis shot, or two cleaning shots).
Table
3.
Atomic composition of the black paint measured by XPS.
Figure
5.
(a) CCD images of the black-painted surface for different numbers of analysis and cleaning shots. α CS + β AS represents that the spot was irradiated first by α number of cleaning shots and then by β number of following analysis shots. Intensity variations of (b) Al I and (c) Na I peaks with respect to the total shot number (ts ≈ 5–7 s). The red and blue circles in (a) represent the approximate boundaries of the laser beams used for the analysis and cleaning, respectively.
The bright reflection from the irradiated area implies that the paint layer was mostly ablated, and the underlying aluminium was exposed. However, one laser shot was insufficient to remove the paint layer, regardless of cleaning or analysis shot. It is considered that the incident laser energy of the first shot was absorbed and consumed for the ablation of black paint itself. The aluminium signal was detected only after the bright surface was observed, that is, from the third shot. Figure 5(b) shows the integrated intensity of the Al I 396.1 nm peak for different numbers of cleaning shots. The error bar represents the standard deviation of the LIBS data from the three measurement points, and the x-axis represents the total shot number (cleaning shots + analysis shots). When the LIBS measurement was conducted without cleaning shots (the AS only case in figure 5(b)), the aluminium signal intensity remained low, even when the shot number increased. In contrast, when the surface was irradiated with two cleaning shots (2 CS case), the aluminium signal intensity for the third shot (i.e., the first analysis shot) was similar to that in the no-cleaning case, but the intensity of the following shot (the second analysis shot) increased significantly. The surface irradiated with three or more cleaning shots (3 CS–5 CS cases) showed strong aluminium signal intensity from the first analysis shot itself. These results demonstrate that an appropriate number of cleaning shots can significantly enhance bulk metal signals. The observed enhancement of the metal signal intensity after two or more cleaning shots may be because the incident laser beam was fully used for the ablation of aluminium, whereas a part of the incident laser beam was consumed for the ablation of paint in the no-cleaning case, as illustrated in figure 4. In addition, in the cleaning-shot cases, the paint signal intensity decreased rapidly as the number of shots increased. Figure 5(c) shows the variation in the Na I signal intensity with respect to the total number of shots. In the no-cleaning case, the Na signal continued to increase with increasing shot numbers and remained at a nearly constant level. However, for the cases with two or more cleaning shots, the Na signal was initially very high but dropped rapidly with the shot number.
In the case of the relatively thicker black paint (ts = 10 s), a bright reflection was observed after the third shot for the no-cleaning case and after the fourth shot for the cleaning shot case. The aluminium signal intensity for the no-cleaning case remained low, as in the ts = 5–7 s sample case, even at high shot numbers.
Owing to the increased paint thickness, a significant jump in the metal signal intensity occurred only after four or more cleaning shots, whereas the intensity after three or fewer cleaning shots was similar to that in the no-cleaning case. The Na I signal for the cases with four or more cleaning shots also showed a rapid decrease as the shot number increased, whereas that in the cases with three or fewer cleaning shots first increased and then remained at a constant level.
3.1.2
White paint
The LIBS spectra of the white-painted samples exhibited different characteristics from those of the black-painted samples. Figure 6(a) shows the CCD images of a white-painted sample irradiated with increasing numbers of laser shots on the same spot (paint thickness ≈ 24.2 μm, analysis shots only). Most importantly, the white paint layer was almost completely removed after the first shot, revealing nearly the same metal surface as the number of shots increased. The paint layer peeled off the aluminium surface in large pieces around the ablation spot. The peeling of the white paint is considered to be due to the transmission of the incident laser light through the paint layer. The transmittance and absorptance of a white paint layer of the same thickness coated on a glass slide were measured to be approximately 38.1% and 4.3%, respectively, at 1064 nm using a spectrophotometer (LAMBDA 950, PerkinElmer, USA), whereas the transmittance and absorptance of a black paint layer were 0.8% and 94.8%, respectively. The differences in transmittance and absorptance between the white and black paints should be attributed to the difference in pigments (table 2). The black paint contained significant portion of carbon black which absorbs strongly, while the white paint had no carbon black but titanium oxides which is highly transmissive at 1064 nm [37]. The absorption of the transmitted laser light by the underlying aluminium possibly led to a rapid increase in temperature and pressure, causing a rupture of the paint layer by the first shot, as observed in figure 6(a). The LIBS spectra of the white-painted samples also exhibited a broad background in the range of approximately 400–900 nm.
Figure
6.
(a) CCD images of the white-painted surface for increasing shot numbers (analysis shots only). (b) Variation in the Al I peak intensity with respect to shot number, and (c) LIBS spectra of the white-painted sample. The red circles in (a) represent the approximate boundaries of the laser beam.
No emissions were observed from the white-painted sample after the first shot. However, after the second shot, a strong Al signal was observed, as shown in figure 6(b), which was significantly stronger than that obtained from the bare surface. The Al signal intensity gradually decreased with increasing shot numbers and became nearly the same as that obtained from the bare surface after the fifth shot. In addition to the strong aluminium signal, several titanium signals along with other paint signals (Na I, Li I, etc.) were detected, as shown in figure 6(c). The titanium signals originated from titanium dioxide, which is one of the main ingredients of the white paint, as shown in table 2. To determine the reason for the unusual increase in the LIBS signal intensity after the second shot, the surface of the white paint sample after the first shot was examined using a field-emission scanning electron microscope (FE-SEM; JSM-7500F, JEOL Ltd.). The surface of the white paint sample after the first shot, shown in figure 7(a), reveals many nanoparticles of varying sizes and shapes. An analysis of these particles (P1, P2, P3, P4, P6) in comparison with the surroundings (P5, P7) by energy-dispersive X-ray spectroscopy (EDS) provided in the FE-SEM confirmed that the nanoparticles had significantly higher titanium and oxygen contents than the surroundings (see table 4), implying that these particles were titanium oxides. When the surface was irradiated by the second shot, the surface morphology showed melting patterns, and the nanoparticles decreased in number and became rounder in shape, as shown in figure 7(b). Therefore, the high signal intensity after the second shot was considered to be closely related to the nanoparticles produced after the first shot. The degree of ionization of plasma and LIBS signal intensity can be significantly enhanced when nanoparticles are deposited on a metal surface, in the so-called nanoparticle-enhanced LIBS [38]. It is considered that the high second shot intensity of the white-painted sample was attributed to both the direct irradiation of incident laser beam on exposed metal surface and the signal enhancement by the nanoparticles produced by the first shot.
Table
4.
Composition of each marked points on the ablated surfaces of the white-painted sample in figure 7.
The training data for classifying the three target metals, aluminium, copper, and stainless steel, were obtained by measuring the LIBS spectra of certified reference materials (CRMs; NIST, Brammer Standard Company) with the elemental compositions listed in table 5. Because a variety of aluminium, copper, and stainless-steel alloys are used in the industry, three types of CRMs per metal with varying compositions were selected to create the training data. From each CRM sample, 150 spectra were collected for training by irradiating with 30 laser shots per point at five different points. Accordingly, a total of 1,350 spectra were collected as the training data, with 450 spectra per metal type.
Figures 8(a)–(c) show the average spectra of the CRMs for each metal. Note that the averaged spectra of black paints (a total of 450 spectra) (figure 8(d)) were also added to the training data as the fourth class (‘contamination’), such that a sample may be considered unrecyclable if only the paint signals were detected. Classification of the training data was performed by principal component analysis (PCA) and, owing to the clear differences between the three metals, the classification accuracy of the training data reached 100% only for PC1–PC2.
Figure
8.
Average spectra of (a) aluminium, (b) copper, (c) STS, and (d) black paint used as training data.
The LIBS data of the black-painted metal samples were collected from samples with two different paint thicknesses obtained by varying the spraying time: three samples for ts = 5–7 s and two samples for ts = 10 s. The classification accuracy was estimated for each shot number because the signal intensity changed significantly with the shot number. For example, the classification accuracy of the ts = 5–7 s samples after being subjected to i shots was estimated using 45 spectra (ith shot spectra/point × five points/sample × three samples/metal × three metals). Similarly, the classification accuracy of the ts = 10 s samples was calculated using 30 spectra (ith shot spectra/point × five points/sample × two samples/metal × three metals). The classification accuracy for each shot was defined as the ratio of the number of correctly identified metal spectra to the total number of spectra. Note that the samples classified as ‘contamination’ were treated as false data.
The paint from the white-painted samples, for both the ts = 5–7 s and ts = 10 s samples, was peeled off after the first shot. Thus, the LIBS data of the ts = 5–7 s and ts = 10 s samples were analyzed together. Accordingly, the classification accuracy of the white-painted samples after being subjected to i shots was calculated by using 75 spectra (ith shot spectra/point × five points/sample × five samples/metal × three metals), as for the black-painted samples.
As shown in figure 3(a), the LIBS spectra of the black-painted samples had a broad background in the 400–900 nm range, and the intensity of the bulk metal signal was much lower than that of the paint signals. Thus, the calculated intensity of the weak metal signal was significantly affected by small changes in the background. The RSD of the integrated intensity values of the Al I 396.1 nm peak collected from three different points on the black-painted samples subjected to 3–10 shots was as large as 45.9%, whereas that of the bare surface was only 7.28%. The signals from each of the three points on the black-painted samples subjected to 3–10 shots were used for the RSD calculation because the Al signal was detected only after the third shot for the black-painted sample. To minimize the influence of the background and large paint signals, background emissions from both the black and white painted surfaces were subtracted using the asymmetric least squares algorithm [39, 40], and line selection was conducted for the emission lines of the three target metals/alloys based on the PCA coefficients of the training data to exclude large paint signals such as those of Na I, Li I, and K I from the classification data. The selected lines for aluminium, copper, and stainless steel are listed in table 6. Finally, each background-subtracted and line-selected spectrum was normalized to its maximum intensity, reducing the RSD of Al I to 5.1%. The spectra after background subtraction, line selection, and maximum normalization were utilized to calculate the shot-to-shot classification accuracy. The classification of test data was performed by using three different algorithms, that is, k-nearest neighbors (kNN), linear discriminant analysis (LDA), and random forest (RF) for comparison.
Table
6.
Selected emission lines for the classification of three metals.
The shot-to-shot classification accuracy of the black-painted samples (ts = 5–7 s) with respect to the increasing shot number for different numbers of cleaning shots is shown in figure 9 for the three classification methods. For the ts = 5–7 s samples, the aluminium signal was clearly detected after the third shot in the no-cleaning case (figure 3(b)), and the aluminium surface was exposed after two cleaning shots, as shown in figure 5(a). Therefore, the calculated classification accuracy was meaningful only after the third shot for both the no-cleaning and cleaning shot cases. Figure 9 shows that the classification accuracies of the kNN and LDA are similar but that of the RF is much worse. The kNN results in figure 9 show that the classification accuracy after the third shot was 73.3% for the no-cleaning (AS only) case and 42.2% for the two cleaning shots (2 CS) case. The relatively poor classification accuracy for the two-cleaning-shot case after the third shot (the first analysis shot) may be due to residual paint on the surface because the fluence for the cleaning shot was much lower than that of the analysis shot. However, after the fourth shot, the classification accuracy of the two-cleaning-shots case increased to 91.1% and 100% for the subsequent shots, whereas that of the no-cleaning case remained at nearly the same value of 75.5% and increased much more slowly. When three cleaning shots were applied, the classification accuracy reached 95.5% for the first analysis shot (the fourth shot) and 100% for the subsequent shots. These results imply that once the metal surface is exposed after the application of cleaning shots, the effect of residual paint can be almost completely eliminated by an additional shot (either analysis or cleaning), and an almost perfect classification can be achieved after the following shot. The classification results by the LDA also showed a similar pattern. However, while the classification accuracy of kNN increased monotonically with shot numbers, that of LDA revealed a slight up and down as the shot number increased.
Figure
9.
The shot-to-shot classification accuracy of the black-painted samples (ts = 5–7 s).
The classification accuracy of the ts = 10 s samples was similar to that of the ts = 5–7 s samples. Note that for these samples, a bright reflection was observed after the third shot for the no-cleaning case and after the fourth shot for the cases with cleaning shots. Thus, the classification accuracy of the cases with cleaning shots was higher than that of the no-cleaning case after the fifth shot and was 100% for the following shots.
These results demonstrate that near-perfect classification can be achieved immediately after the metal surface is exposed when cleaning shots are applied, implying that paint effects can be almost completely eliminated using cleaning shots. However, the classification accuracy of the no-cleaning case remained at a moderate value (approximately 70%–80%) and increased slowly, preventing the classification accuracy from reaching close to 100% even at high shot numbers.
3.2.4
Determination of cleaning shot numbers
As demonstrated by the data in figure 9, a higher classification accuracy could be achieved at a fewer total shot number when the analysis shot was fired after the metal surface was exposed by cleaning shots. However, since the classification accuracy would be high if the shot number is sufficiently large regardless of whether cleaning shots are used, it is more important and practically meaningful to sort the scrap metals with high accuracy at a minimum number of shots. In other words, sorting can be achieved with high throughput by firing as few cleaning shots as possible before an analysis shot.
To enable a high throughput sorting of painted metal scraps by the proposed method, which involves cleaning the paint with a larger spot diameter and then firing an analysis shot in the middle of the cleaned surface with a smaller laser beam, the following two conditions need to be satisfied. First, the change of spot diameter from cleaning shots to analysis shots should be possible in real time without changing the elevation of a sample. Therefore, the optical setup was modified as shown in figure 10(a), which allows both the cleaning and analysis shots to be fired while keeping a sample at the same height. In the modified setup, the spot diameter on the sample surface was adjusted for cleaning and analysis shots by rapidly varying the distance between the focusing lenses L1 and L2 using a voice coil motor. Figure 10(b) shows the computed spot-diameter change on the sample surface (located at a distance of 780 mm from L2) with respect to the gap (S) between L1 and L2. This setup allows a painted scrap on a sorting system to be irradiated by an analysis shot immediately following the desired number of cleaning shots. For a moving scrap as in an industrial scale sorting system, the mirror in figure 10(a) may be replaced by a galvano-scanner so that moving scraps can be tracked and irradiated for cleaning and classification [24].
Figure
10.
(a) Schematic diagram of the modified LIBS system and (b) spot diameter change with respect to the gap between focusing lenses.
Also, cleaning should be stopped as soon as the bulk metal is exposed to ensure a high throughput. Whether the surface was sufficiently cleaned can be determined by monitoring the variation in the LIBS signal of cleaning shots as follows. Figure 11 shows the LIBS spectra and CCD images of the black-painted surface for increasing number of cleaning shots on a copper sample. The LIBS signals for the cleaning shots in figure 11(a) show a broad background and the paint signals (Na I, Li I and K I peaks). Both the broad background and the paint signals are weak for the 1st and 2nd shots but increased significantly for the 3rd shot. The CCD images in figure 11(b) show that paint is mostly removed and the copper surface is exposed after the 2nd cleaning shot. These results indicate that the laser energy for the 1st and 2nd shots was mostly consumed for the vaporization of the paint material. The LIBS signal for the 3rd shot shows a significant increase of the intensities of the broad background and paint signals. Owing to the devoid of the paint layer, the incident laser light must have been strongly absorbed by the metal, causing a higher surface temperature and stronger breakdown of the residual and edge paint. These results imply that the change in the LIBS signal intensity during cleaning shots can be used to determine when the paint layer is mostly removed and an analysis shot can be fired to identify the type of bulk metal.
Figure
11.
(a) LIBS spectra and (b) CCD images of a black-painted copper sample after each cleaning shot (ts = 5–7 s).
Figures 12(a) and (b) show the sum of all pixel intensities of the ts = 5–7 s and 10 s black-painted copper samples, respectively, for an increasing number of cleaning shots. The summation intensity of the ts = 5–7 s sample in figure 12(a) increases sharply at the 3rd cleaning shot because the paint is mostly removed from the sample after the 2nd cleaning shot. Similarly, a sharp increase in the summation intensity occurred for the ts = 10 s sample at the 4th shot, and the corresponding CCD images showed that the paint layer was mostly removed after the 3rd shot (not shown). Therefore, in this work, the sharp increase in the summation intensity was used as the criterion to determine whether the target was ready for an analysis shot.
Figure
12.
Summation intensity of LIBS spectra from the (a) ts = 5–7 s and (b) ts = 10 s black-painted copper samples with respect to cleaning shot numbers.
Since the summation intensity tends to increase for each additional shot until it reaches the maximum, it is necessary to evaluate whether the increase is gradual, as in the case from the 1st shot to 2nd shot (painted surface), or sharp, as in the case from the 2nd shot to the 3rd shot (bulk metal exposed), as shown in figure 12(a). Therefore, the average (Mk) of the summation intensities of the kth shot (Sk) and the preceding shot (Sk−1) is expressed by
Mk=Sk+Sk−12.
(1)
The sum of Mk and Mk−1 becomes
Mk+Mk−1=Sk+Sk−22+Sk−1,k⩾
(2)
When the difference in summation intensities between shots is relatively small such that it can be assumed that {S} _{1}\approx {S} _{2}\approx {S} _{3}\approx \cdots \approx {S} _{k} , corresponding to the case that the surface is fully or partially covered by paint, equation (2) can be written as
{M}_{k}+{M}_{k-1}\approx {2S}_{k} > {S} _{k} .
(3)
In the case that a sharp increase of summation intensity occurred due to the exposure of bulk surface at the kth shot, it can be assumed that {S} _{1}\approx {S} _{2}\approx {S} _{3}{\approx \cdots \approx S}_{k-1}\ll {S} _{k} , and equation (2) can be approximated as
Thus, by comparing the magnitudes of Mk + Mk−1 and Sk after each cleaning shot for k\geqslant 3, it can be determined whether the bulk surface is exposed or not. Specifically, if Mk + Mk−1 < Sk at the kth shot, it is considered that the bulk metal is exposed and an analysis shot can be fired. Conversely, if Mk + Mk−1 > Sk, it is considered that paint removal is incomplete and another cleaning shot needs to be fired and repeat the procedure.
Figures 13(a) and (b) present the test results of the above proposed method to determine when an analysis shot can be fired. The average values of Sk, Mk, and Mk+Mk−1 of five measurement points in figure 12 for increasing shot numbers are shown for the ts = 5–7 s and 10 s samples. It is clear that Mk+Mk−1 < Sk only when a sharp increase of the summation intensity occurred, the 3rd shot for the ts = 5–7 s sample and the 4th shot for the ts = 10 s sample, and Mk + Mk−1 > Sk for all other cleaning shots.
Figure
13.
Variations of the Sk, Mk, and Mk + Mk−1 with respect to cleaning shot numbers for the (a) ts = 5–7 s and (b) ts = 10 s black-painted copper samples.
Table 7 summarizes the classification results with and without the proposed method. In the case of the ts = 5–7 s sample, it was determined that the bulk metal was exposed after the 3rd shot using the proposed method and thus an analysis shot was fired at the 4th shot, resulting in the total number of shots (cleaning shots + analysis shot) becoming four. The classification accuracy was 100% for this case for all of the three methods (Test A). When the same number of analysis shots was fired to the sample without cleaning shots, the classification accuracy dropped to 91.1% and 93.3% for the kNN and LDA, repectively, whereas that of the RF was only 17.8% (Test B). If the target is irradiated by a sufficient number of analysis shots, the residual paint will be eventually removed from the surface and classification accuracy should increase. It took a total of seven shots to achieve 100% accuracy with only the analysis shot for kNN and LDA (Test C), which corresponds to a decrease of throughput by 43% from that of the cleaning shot case. Similarly, the ts = 10 s sample could be classified with 100% accuracy after five shots (4 cleaning shots and 1 analysis shot) by kNN and LDA (Test D). Under the no-cleaning condition, the classification accuracy dropped to 80% and 86.6% for kNN and LDA, respectively, after five analysis shots (Test E) and it took a total of eight shots to achieve 100% classification accuracy (Test F), resulting in approximately 38% less throughput from that of the cleaning case. These results demonstrate that a high classification accuracy can be achieved using a fewer total shot numbers during LIBS sorting of black- painted metals by applying the proposed method to determine the required number of cleaning shots.
Table
7.
Classification accuracy of different cleaning conditions and paint thicknesses.
The shot-to-shot classification accuracies of the white-painted samples are shown in figure 14. Because the white paint layer was peeled off by the first shot (either the analysis shot or cleaning shot) and strong bulk metal signals were detected after the second shot due to nanoparticle effects, the classification accuracy was 97.3% and almost 100% for the no-cleaning and cleaning-shot cases, respectively, from the second shot onward. This slight difference may be due to paint ablation caused by the laser beam edge. Because the classification accuracy for the no-cleaning case was over 95%, cleaning shots were considered unnecessary for the white-painted samples.
Figure
14.
The shot-to-shot classification accuracy of the white-painted samples.
The effects of paint, one of the most common surface contaminants on scrap metals, on the LIBS spectra and classification accuracy were investigated for black and white paints with different coating thicknesses. The LIBS spectra of the painted surfaces showed broad background emission, strong emission peaks unrelated to the target metal, and large shot-to-shot fluctuations, which were mostly resolved by appropriate data processing, such as background subtraction, line selection, and normalization. However, for the black-painted samples, the classification accuracy remained relatively low even after the metal surface was exposed, owing to the ablation of the paint by the laser beam edge. These paint signal effects were almost completely removed by applying cleaning shots prior to the LIBS analysis. A method to determine the number of cleaning shots necessary to obtain high classification accuracy was proposed on the basis of the change of LIBS signal intensity during cleaning shots. It was demonstrated that a much higher throughput can be achieved by applying the proposed method than the no-cleaning case for the same classification accuracy. Since the number of cleaning shot is expected to increase with increasing paint thickness, the use of a higher laser pulse energy could be advantageous for thick paints. For the white-painted samples, the surface of the bulk material was exposed to one laser shot, and the spectral intensity of the second shot was significantly enhanced because of the nanoparticles generated by the first laser shot. Accordingly, high classification accuracy could be readily achieved for white-painted samples, even with no cleaning shots. These results demonstrate that LIBS spectra and classification accuracy are strongly affected by the presence of paint, and different approaches should be adopted depending on paint colors to achieve high classification accuracy with the minimum number of shots.
Acknowledgments
This study was supported by the R&D Center for Valuable Recycling (Global-Top R&D Program) of the Ministry of Environment (No. 2016002250003).
Wens B et al 2010 Environmental impacts of NF-metal recycling - MSW as a resource In: Proceedings of the ISWA World Congress 2010 Hamburg 2010 (https://www.researchgate.net/profile/Thomas-Pretz/publication/228513473_Environmental_Impacts_of_NF-Metal_Recycling-MSW_as_a_Resource/links/5440f77d0cf251bced61cf86/Environmental-Impacts-of-NF-Metal-Recycling-MSW-as-a-Resource.pdf
Green J A S 2007 Aluminum recycling and processing for energy conservation and sustainability ASM International (https://www.sae.org/publications/books/content/b-asm-003/, publisher: ASM International
Noll R et al 2017 Proc. EMC (https://www.researchgate.net/profile/Roland-Ambrosch/publication/321938024_Next_generation_urban_mining-Automated_disassembly_separation_and_recovery_of_valuable_materials_from_elec-tronic_equipment_overview_of_RD_approaches_and_first_results_of_the_European_project_ADIR/links/5a3a1f62a6fdcc34776b36ea/Next-generation-urban-mining-Automated-disassembly-separation-and-recovery-of-valuable-materials-from-elec-tronic-equipment-overview-of-R-D-approaches-and-first-results-of-the-European-project-ADIR.pdf
[20]
Werheit P et al 2011 J. Anal. At. Spectrom.26 2166 doi: 10.1039/c1ja10096c
[21]
Palanco S, Baena J M and Laserna J J 2002 Spectrochim. Acta Part B: At. Spectrosc.57 591 doi: 10.1016/S0584-8547(01)00388-3
[22]
López-Moreno C, Palanco S and Laserna J J 2007 J. Anal. At. Spectrom.22 84 doi: 10.1039/B609705G
Horckmans L et al 2016 Refasort: automated sorting of refractory waste for high value recycling In: Proceedings of the 7th Sensor-Based Sorting & Control 2016 Aachen 2016: 187 (https://publications.rwth-aachen.de/record/668852/files/668852.pdf
DownLoad:
