@phdthesis{,
    author = {Fomin, Petr},
    title = {Time-gated spectroscopy for the classification and recycling of fluorescently labeled waste plastics},
    editor = {},
    booktitle = {},
    series = {},
    journal = {},
    address = {},
    publisher = {},
    edition = {},
    year = {2020},
    isbn = {},
    volume = {},
    number =  {},
    pages = {},
    url = {},
    doi = {},
    keywords = {plastic waste; recycling; spectroscopy; fluorescence; time-gated; classification; sorting},
    abstract = {Plastics have long become indispensable materials in modern society. In light of the ever-growing production volumes of plastics, recycling of end-of-life plastic products becomes increasingly important every year. The recycling of plastics waste is attractive from both the environmental and economic points of view. It reduces the amounts of waste buried in landfills and the CO2 emission associated with the production of new plastic resins and helps reduce the usage of strategic fossil resources such as crude oil. One of the main difficulties associated with the recycling of plastics is the need for mono-fractional sorting of plastics of different types and grades. If different types of plastics are mixed and used together for recycling, the resulting product has low quality and cannot compete with products made from brand new plastics. This is especially important in demanding applications where high-quality material properties are required. Thus sorting waste plastics is key to the success of recycling. The goal of this dissertation is the development and practical evaluation of a spectroscopic measurement and classification system for the automated sorting of shredded flakes of different plastic types based on their fluorescence spectra. In contrast to the state-of-the-art spectroscopic approaches which rely on the intrinsic properties of the plastics, this work investigates the concept of “fluorescent labeling”. The idea is to incorporate small amounts (at ppm concentration levels) of appropriate fluorescent tracers (or “markers”) into the raw plastics during the manufacturing process thus generating unique fluorescence spectra emitted by the plastics. These fluorescence spectra can then be measured using a dedicated spectroscopic instrument and used for the plastics classification. Markers are incorporated into the plastics according to a certain (e.g. binary) coding scheme in order to increase the overall number of different plastics that can be labeled. Fluorescent markers can be especially helpful in the case of dark and black plastics with their normally flat, non-characteristic reflectance and/or fluorescence spectra by adding specific features to these spectra. Additionally, the use of fluorescence markers does not only allow the classification of plastics of different types (i.e. with different chemical structures), but also plastics of the same type that are produced by different manufacturers or sold to different customers. In fact, virtually any information can be encoded into plastics through fluorescence labeling. For example, fluorescent markers can help distinguish original plastic parts from counterfeit ones. The idea of fluorescent labeling was first proposed about 20 years ago but has not yet been put into wide industrial practice. Some commercially available systems carry out macro-sorting of plastics, i.e. the sorting of large objects such as for example plastic bottles. For complex assemblies consisting of many parts made of different plastic types, a prior disassembling is thus necessary. In contrast, micro-sorting deals with small (in the millimeter range) flakes of shredded plastics, does not require the labor-intensive prior disassembling, and is, therefore, more flexible and attractive for the recycling of plastic waste. This dissertation focuses on the micro-sorting approach. Due to the chemical structure of the plastics and/or various additives such as brightening components or for UV-protection, many plastics exhibit the so-called autofluorescence. The autofluorescence spectrally overlaps with the fluorescence of the incorporated markers and may even completely mask it. A strong autofluorescence can make correct and reliable classification of fluorescently labeled plastics problematic or even impossible in practice. To combat the negative influence of the autofluorescence, a method called the time-gated fluorescence spectroscopy (TGFS) is proposed in this work. This method dwells on the fact that the fluorescence emission decays exponentially when the excitation light is turned off and that the fluorescence decay time constants of inorganic markers are usually orders of magnitude larger than those of the typical autofluorescence. Using a pulsed excitation light in combination with time-gated acquisition of the fluorescence emission makes it possible to (almost) completely avoid the presence of autofluorescence in the measured spectra. The obvious downside of this method is a decreased signal intensity and thus lower signal-to-noise ratio (S/N ratio) of the acquired fluorescence spectra in comparison to spectra acquired without pulsing and time-gating. Low fluorescence intensities and S/N ratios are disadvantageous for the classification performance and thus must be maximized. In the context of TGFS, the intensity of the acquired fluorescence spectra mainly depends on the fluorescence decay time constants of the markers and on the acquisition parameters of the time-gating. The former are governed by chemical/physical laws and are difficult to modify. The latter, however, can be easily varied in order to achieve the highest possible S/N ratio and in turn the best classification. For this purpose, a mathematical model for the fluorescence intensity of the TGFS spectra is proposed and investigated in this dissertation. This model is then used to find the best suitable acquisition parameters for TGFS. In order to achieve the best possible classification performance, various approaches to the classification of the fluorescence spectra emitted from the labeled plastics are investigated including the relatively simple and thus numerically efficient naive Bayes’ methods and spectral similarity measures, as well as more complex ones such as neural networks (NN), support vector machines (SVM) and random forests (RF). The classification algorithms are implemented in a simulation framework developed to allow the modeling of marker fluorescence spectra corrupted with different disturbances important in practice, such as the measurement noise, autofluorescence, etc. Classifiers are evaluated using computer simulations with respect to these disturbances. In order to quantify the classification performance of a TGFS system for fluorescently labeled plastics in practice, a prototype system was built. The prototype was developed with a particular focus on an industrial environment in a typical recycling facility and designed to process shredded plastic flakes with sizes between approx. 3 mm and 10 mm on a 500 mm wide conveyor belt in 50 parallel channels. Six fluorescent markers were used in binary combinations allowing to label up to 26 − 1 = 63 different plastics. The prototype was thus capable to classify and sort flakes of up to 63 different fluorescently labeled plastics simultaneously with a mass throughput of up to about 250 kilograms per hour. Extensive experimental measurements with approximately 140 000 shredded flakes of different fluorescently labeled plastics and ∼10 000 unlabeled plastic flakes were carried out to evaluate the performance of the prototype. A very high classification performance was achieved: an average sensitivity (i.e. true positive rate, TPR) of 99.76% and an average precision (i.e. positive predictive value, PPV) of 99.88%. Additionally, measurements with fluorescently labeled black plastic flakes were carried out yielding virtually the same high performance: TPR of 99.76% and PPV of 99.60%. The results are not 100% perfect since some misclassifications occurred due to the low S/N ratio of the spectra measured from very small flakes (smaller than approx. 2 mm). The majority of the misclassifications are due to the unequal intensities of individual markers in the marker combinations which led to some “stronger” markers masking the presence of the “weaker” markers in the measured fluorescence spectra. Nevertheless, out of 150000 experimentally investigated flakes, only 338 were misclassified, which only is 0.23%. The investigations in this dissertation show that a highly reliable classification of fluorescently labeled plastics is possible in practice. This work proves that the principle of fluorescent labeling is applicable not only for the macro-sorting of large plastic objects but also for the more versatile micro-sorting of small shredded plastic flakes. Moreover, this approach can be successfully implemented in an industrial environment. Clearly, certain adaptation and optimization steps must be taken especially with respect to achieving a higher mass throughput (several tons per hour) for an industry-scale operation. Using a larger number of markers is also possible and would allow labeling and classifying more plastics simultaneously. Overall, this dissertation demonstrates a promising way to make the recycling of waste plastics flexible, economically attractive, and successful.},
    note = {},
    
    school = {Universität der Bundeswehr München},
    
}