The material recycling of raw materials is of growing importance in view of finite resources, rising demand, and environmentally damaging extraction and production conditions. This applies in particular to plastics, which are currently manufactured largely from crude oil. The rate of material recycling of plastics worldwide is currently only around 9%. A further 11% is used for thermal recycling, while the remainder is landfilled [1
]. The majority of recycled plastics are also often processed into inferior secondary products. The reason for this is the lack of suitable sensor technologies for classifying plastics, which makes it possible to separate plastic mixtures with sufficient purity (sometimes up to 99.9%). This applies in particular to technical black plastics, which often consist of a complex material mix and are coloured black by carbon fillers. The standard separation technologies used for sorting plastics, such as classification using float-sink methods [2
] or electrostatic separation [3
] or sorting using hyperspectral imaging (HSI) in the near-infrared spectral range [4
], are not suitable for black plastics. Other methods such as X-ray fluorescence (XRF) [6
] are suitable for sorting black plastics but are limited to a small number of plastic combinations.
For the sorting of technical black plastics, some approaches are currently being developed, and some are already being offered commercially. These include hyperspectral imaging in the mid-infrared spectral range [7
] and terahertz spectroscopy [9
]. Both approaches allow the sorting of black plastics but are not optimal due to expensive instrument technology, low achievable throughputs, and large required particle sizes. Another possibility is the use of laser-induced breakdown spectroscopy (LIBS) or Raman spectroscopy [11
]. However, both approaches are not yet widely used in industrial applications for sorting black plastics.
The most promising technology at present is the classification of black plastics on the basis of their fluorescence. Many plastics show a characteristic fluorescence if they are illuminated with intensive laser radiation. The exact causes of fluorescence are not well studied. It is assumed that fluorescence is mainly caused by additives or impurities in the plastics [13
]. With this technology, it is also possible to classify black plastics, and there has already been commercial implementation. The achieved accuracies do not exceed 98% [16
The aim of this work is to improve the technology of fluorescence spectroscopy for the sorting of black plastics and thus to further increase the classification accuracy. For this purpose, fluorescence spectroscopy is to be combined with hyperspectral imaging. In this way, in addition to the fluorescence spectra of the plastic particles, their shape and texture are also obtained. This is especially characteristic for the type of plastic for particles generated by cryogenic grinding [17
]. Convolutional Neural Networks (CNNs), which have achieved great success in the classification of image data, will be used to consider the shape of the plastic particles in the classification [18
For the experiments, about 400 particles each of 14 black plastics in 12 plastic classes were measured with an imaging fluorescence spectrometer. The particles were produced by cryogenic grinding and range in size from 5 mm to 12 mm. The data obtained was then used to train various machine learning algorithms and compare them using statistic methods. The ‘classical’ machine learning algorithms linear discriminant analysis (LDA [21
]), k-nearest neighbour classification (kNN [22
]), support vector machines (SVM [23
]), and ensemble models with decision trees (ENSEMBLE [24
]) were trained. The particle shape was not considered for these algorithms. Convolutional neural networks were trained considering the particle shape [18
]. Therefore, spectral images of the particles were generated, which show the particle shape as well as their fluorescence properties. For plastics without fluorescence, reflectivity in the near-infrared spectral range (NIR) was used to detect the position and shape of the particles. For all trained algorithms, an automatic hyperparameter optimization by random search (RS, [25
]) was performed. The aim was to increase the achievable overall classification accuracy of the models. The hyperparameter optimization by BOA was also tried but did not provide better results than RS and is therefore not shown for reasons of clarity. At the same time, an automated optimization of the hyperparameters of the classification algorithms allows for a simple training of models for the classification of new plastics. This reduces the demands placed on plant operators during subsequent industrial use, since the user does not need to have any prior knowledge of machine learning. The obtained classification accuracies of the models were examined by means of ANOVA with subsequent Tukey test with regard to statistically significant differences. Models for the simultaneous classification of all 12 plastic classes and for the classification of 41 industrially relevant mixtures of two to three plastic classes were examined.
It was found that CNNs using the spectral and shape information can achieve statistically significantly and better overall classification accuracy for the classification of all 12 plastics than classical machine learning algorithms using only the spectral information. For the classification of the mixtures, there were no differences between the three considered algorithms—ENSEMBLE, SVM, and CNN. Furthermore, the automatic optimization of the hyperparameters by random optimization proved to be a very good possibility to improve the overall classification accuracy of the models. In total, the desired overall classification accuracy of 99.9% could be achieved for 18 of the 41 plastic mixtures with two or three classes.
In this study, an imaging fluorescence spectrometer with additional illumination in the NIR spectral range was used to classify technical black plastic particles after cryogenic grinding. These plastics are arising in particular during the recycling of plastic components from the automotive or electronics industries. Since these are often composite components, the waste is ground to small particles before recycling and must then be sorted with high purity (99.9%). Even the smallest impurities can reduce the quality of the recycled material and thus the achievable price. A sorting of technical black plastic particles is not yet possible with this purity. The aim was therefore to measure the fluorescence of the black plastics after excitation with a 450 nm laser and to classify them with high overall accuracy using machine learning models. In addition, attempts have been made to use the shape of the plastic particles for classification by using CNNs, thereby increasing the achievable overall classification accuracy.
A total of around 400 particles were measured from 14 plastics in 12 plastic classes. The classification was carried out using the algorithms discriminant analysis, k nearest neighbour classification, support vector machines, classification ensembles with decision trees, and convolutional neural networks. It was also attempted to find optimal model parameters, which can significantly increase the overall classification accuracy of the models. These parameters were determined by an automatic hyper parameter optimization by random search.
The experiments with the total data set of all plastics showed that the best results could be achieved using CNNs, kNN, and ENSEMBLE algorithms. The highest overall classification accuracy was 93.5% for the CNNs. Hyperparameter optimization led to a statistically significant improvement in overall classification accuracy for most algorithms.
When considering 41 plastic mixtures with two to three plastic per mixture, the desired overall accuracy of at least 99.9% was achieved for 18 of the plastic mixtures. Here, the use of CNNs showed no improvement compared to ENSEMBLE and SVM algorithms. In the future, an industry-oriented demonstrator for the classification of technical black plastic particles using imaging fluorescence spectroscopy will be developed, and more data is to be recorded for the training of better models.
Overall, the method presented seems to be a promising approach for the classification of black plastics and could contribute to an increase in the recycling of plastic waste.