Method and device for identifying and classifying single-material vs multi-layer plastics for recycling applications

Abstract

The present invention relates to a device for identifying and classifying samples of single-material versus multi-layer plastics for recycling applications that comprises a THz time-domain spectroscopy system and a processor. The spectroscopy system is provided with at least one optical source, a THz emitter, a THz receiver, and a beam splitter for directing the signal reflected by the sample towards the receiving antenna. The processor is provided with a machine learning classifier that takes the THz temporal trace of the signal reflected by the sample and measured by the receiving antenna as baseline data to classify the sample as single-material or multi-layer.

Description

[0001] PROCEDURE AND DEVICE FOR THE IDENTIFICATION AND CLASSIFICATION OF MONOMATERIAL VS. MULTILAYER PLASTICS FOR RECYCLING APPLICATIONS

[0002] DESCRIPTION

[0003] TECHNICAL SECTOR

[0004] The invention addresses the problem that packaging composed of layers of different materials ("multilayer") introduces in the pre-recycling sorting process of plastics obtained from urban waste. The present invention proposes a system and method for identifying whether a package is mono-material (a single material, which usually implies a single layer) or multi-layer (formed by several layers, at least two of which are made of different materials), primarily based on non-contact measurement of the overall thickness of the package.

[0005] STATE OF THE ART

[0006] A first step in waste separation is sorting by composition for the subsequent reuse of materials in the manufacture of new products. Currently, packaging sorting in urban waste sorting plants is primarily based on two technologies: machine vision or NIR. In both cases, the packaging is illuminated with radiation from these bands, and the reflected signal from its surface is recorded.

[0007] The sorting process for single-material packaging (composed of a single polymer) is quite advanced, achieving good separations by material, color, and material. However, many packages require multiple layers of different materials to achieve a better barrier effect and thus maintain product properties for the specified time. In this case, NIR systems sort the packaging based on the outermost layer facing the sensor. This results in contamination of the fraction, which significantly reduces the value of the recovered product.If there is a significant number of multi-layer packaging in the fraction, the purity may fall below the required purity threshold, so the bale of packaging must be sent to landfill as its reuse is not viable, since the different melting temperatures of the different materials would cause failures in the extrusion machines, which would entail not only the loss of production but also a shutdown for cleaning.

[0008] The current approach to this problem is, on the one hand, manual disposal of multi-layer packaging. Operators know which common packaging is multi-layered, and they are eliminated. On the other hand, efforts are being made to reduce the presence of multi-layer packaging at the regulatory level, but the performance offered by single-material packaging is not the same as that of multi-layer packaging, and therefore multi-layer plastics continue to be used, which end up incinerated or in landfills. The alternative is food waste, since suboptimal packaging deteriorates more quickly, with the resulting environmental and economic burden for both the distribution chain and society. For these reasons, the sorting process would benefit at several stages if methods existed to separate multi-layer packaging and thus improve the purity of the recovered fractions.

[0009] Additionally, there is significant activity in the development of delamination solutions that enable efficient recycling of the most common multi-layer packaging. The efficiency of these multi-layer recycling systems would also benefit from a sorting system to detect the presence of single-material packaging, which consumes reagents and resources but goes unused, thereby affecting the economic efficiency of these plants.

[0010] NIR systems, as indicated, cannot detect whether a package is multilayered, since the reflected signal comes from the layer facing the sensor. These systems are based on analyzing the spectrum of the reflected signal produced by Fresnel's laws, according to which the reflected signal depends on the difference in refractive indices between the two media at the interface.

[0011] Machine vision systems are another method based on a different physical principle. Specifically, they rely on a library of images of previously recorded and classified containers. The camera compares the container on the conveyor belt and identifies it as one from the library. This system is especially problematic when the containers are compressed or broken.

[0012] SUMMARY OF THE INVENTION

[0013] The system and method of the invention resolve the aforementioned drawbacks by using a THz time-domain spectroscopy (THz-TDS) system combined with a machine learning classifier that allows the packaging to be classified as mono-material or multi-layer. This classifier can optionally be combined with two other classifiers to provide additional information about the materials that make up the structure.

[0014] More specifically and advantageously, the waves used are in the range of 0.1 THz to 15 THz to discriminate whether the packaging is mono- or multi-layered. THz waves, due to their wavelength, provide volumetric information. Unlike visible or infrared radiation, plastics, regardless of color, are semi-transparent in THz. Therefore, the waves pass through the packaging, and information about the set of layers can be obtained. Because the typical thicknesses of consumer packaging are typically similar to the wavelength used in THz, it is not obvious to identify whether the material is mono-material or not. Therefore, the system must be complemented with machine learning-based processing that allows the captured traces to be classified with previous examples (obtained experimentally or created artificially from a model of the electromagnetic interaction of the structure).

[0015] Specifically, the invention is a device and method for identifying and classifying samples of mono-material versus multi-layer plastics for recycling applications comprising: a time domain THz spectroscopy system and a processor, the spectroscopy system being provided with at least one optical source (which may be a pulsed femtosecond laser), at least one THz emitter (which may be a photoconductive antenna or a non-linear crystal for converting the laser pulses into THz pulses), at least one THz receiver (which may be another photoconductive antenna or a non-linear crystal), optical elements for directing the THz beam towards the sample and from the sample to the receiver, a beam splitter for directing the signal reflected by the sample towards the receiving antenna,where the processor is provided with a machine learning classifier that takes as starting data the THz time trace of the signal reflected by the sample and measured by the receiving antenna. Advantageously, the transmitter transmits in the range of 0.1 THz to 15 THz, the classifier is of the KNN type, and the classifier may be provided with program means for selecting a set of points from the trace and sending the selection to the classifier, which then obtains the thickness.

[0016] BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to assist in a better understanding of the characteristics of the invention and to complement this description, the following figures are attached as an integral part thereof, which are illustrative and not limiting in nature:

[0018] Figure 1 is a histogram showing the principle on which the invention is based.

[0019] Figure 2 is a block diagram of the THz system according to an implementation of the system of the invention.

[0020] Figure 3 is a flow diagram of the process of the invention.

[0021] DETAILED DESCRIPTION

[0022] Plastics are generally semi-transparent in the THz and sub-THz range, so it is possible to use waves in this region of the spectrum to characterize samples by their thickness without the need for physical contact.

[0023] The system of the invention uses a high-speed THz temporal spectrometer (THz-TDS, Terahertz Time Domain Spectroscopy) in a reflection configuration with illumination perpendicular to the sample through a beam splitter to determine the thickness without contact.

[0024] The method is based on the recognition that mono-material packaging systematically has lower overall thicknesses than multi-layer packaging (Fig. 1) due to the commitment to reducing the amount of resin used while maintaining mechanical properties, while multi-layer packaging must provide special barrier characteristics against various elements, which implies greater thicknesses. This has been verified by a multitude of food packaging suppliers, which, in the case of multi-layer packaging, are largely PET-PE and PE-EVOH-PET.

[0025] A preferred implementation of the system is shown in the diagram in Fig. 2. The laser block, preferably a femtosecond pulsed laser (<100 fs), is used to illuminate photoconductive antennas (PCAs). The connection, indicated by the solid lines, is via 1550 nm single-mode optical fiber, where the fibers are a combination of standard and dispersion-compensating single-mode fibers, preferably of the polarization-maintaining type in both cases. Two optical delay lines (ODLs) are used.

[0026] The generation and detection of THz pulses is carried out using photoconductive antennas (one transmitter and one receiver) and two optical delay lines: one slow (ODL) and one fast (FODL). The transmitter antenna generates the THz pulse, which is radiated toward the sample through the beam splitter, which in turn directs the signal reflected by the sample toward the receiver antenna, which samples the received THz signal to generate a photocurrent proportional to the field strength of the THz signal coming from the sample. Without any further modifications, only one measurement point would be obtained, so the sampling point must be changed (obtaining a measurement point at each time instant within a window). To do this, it is necessary to continuously delay the optical path of one of the pulses relative to the other.The slow ODL specifies the measurement window, while the fast ODL allows the delay to be performed continuously on the reference optical pulse, thus obtaining the pulse / signal.

[0027] Signal processing

[0028] This equipment must be complemented with signal processing based on machine learning tools.

[0029] Optionally, the procedure includes applying preprocessing to both the training sample set and the measurements, in order to minimize the noise present in the measurements due to the deterministic amplitude oscillations inherent in the antenna and the noise induced by the vibration of the conveyor belt of the containers to be sorted. To do this, the reflectance is measured with the belt in motion in the absence of a sample in order to obtain the system reference under operating conditions. To obtain this reference, for example, hundreds or thousands of traces are obtained (since the system provides at least dozens of measurements per second, a high number of traces can be taken), the average is calculated, and the traces are stored in the internal memory of the measurement system.When measuring reflectance with a container on the belt, the reference is removed by subtraction, thus removing the effect of deterministic antenna noise and belt vibrations, leaving the measurement vertically centered at 0.

[0030] As part of the preprocessing, the THz trace corresponding to the signal reflected from the sample is segmented. This segmentation consists of selecting a set of points from the trace to pass to the classifier (i.e., time windowing). This is necessary to ensure that all signals passed to the classifier have the same number of points. This windowing allows the classifier to be independent of the position of the pulse within the trace associated with the position of the container in the zone illuminated by the THz signal.

[0031] After the data has been segmented, the Standard Normal Variate (SNV) technique is applied so that the data are standardized with a mean of 0 and unit variance, understanding the characteristics as the temporal moments.

[0032] Sorter

[0033] Following these steps, the THz time signature data measured by the receiving antenna are suitable for use in the classifier that determines the structure of the containers, whether single-layer or multilayer. This classifier determines the structure type primarily based on thickness. To confirm this, a principal component analysis was performed on the measurements from the training and validation sets used in the classifier model, extracting 143 principal components and coloring them based on structure type, sample thickness, and material combination. Score analysis determines a classification based on sample thickness, and representation by material combination reveals a potential differentiation by material type for multilayer containers, where the refractive index of the material could interfere with classification.

[0034] To select the optimal classifier, different classifiers were modeled. First, the hyperparameters on which the classifier is based were optimized by applying the Asynchronous Successive Halving Algorithm (ASHA). This algorithm randomly selects several models with different hyperparameter values ​​and trains them on a small subset of the training data. If the performance of a particular model is promising, it is promoted and trained with a larger amount of training data. This process is repeated, and successful models are trained on increasingly larger amounts of data. By default, the model with the lowest cross-validation classification error is chosen at the end of the optimization, resulting in a single model optimized for the initial training set.Within the ASHA algorithm, it has been established that the main classifiers to be modeled are primarily classifiers based on the maximum probability of an individual's belonging to a group. Therefore, it would be possible to use different classifiers of this type. Specifically, the following have been tested: classification trees, k-nearest neighbor (KNN), Random Forest (RF, massing of classification trees), as well as classifiers based on the discriminant power of the data, such as Support Vector Machines (SVM) and discriminant analysis classifiers (both linear and nonlinear). Table 1 shows a summary of the conventional metrics obtained for different classifiers.

[0035] Table 1. Comparison of classifiers.

[0036] To avoid overfitting problems and improve the stability and fit of the models, the technique known as Bagging has been applied in order to obtain robust results in terms of the efficiency of the different classifiers. This technique consists of randomly scanning the training and validation sets and obtaining the optimal model for those training and validation sets. This process is repeated n times (in a preferential case, 60 iterations), with an optimal model obtained for each iteration. At the end of the entire process, the result is filtered based on the Kernel model that is repeated the most times, and within that type of classifier, the model that maximizes the result is chosen based on a series of selection parameters such as accuracy (R2), F1 score, precision, recall, or the Matthews correlation coefficient.

[0037] The modeling process of the different classifiers can be summarized as follows:

[0038] ■ Data processing, training set and validation.

[0039] ■ Modeling & Validation (using ASHA and selection parameters).

[0040] ■ Best model selection (Bagging).

[0041] ■ Use of the best model for prediction.

[0042] The classification efficiency results for each of the classifiers and for each of the classes being determined are shown in Table 2, along with the model used to determine the structure and the material it is composed of. Table 2. Experimental results obtained with a classifier that determines the structure of the container using the KNN model with K=1 with the containers in motion.

[0043] Target R2 (%) Classification efficiency

[0044] Monomaterial 92.02 0.89 vs multilayer

[0045] Monomaterial 74.00 0.69

[0046] Multilayer 74.87 0.61

[0047] In the case of determining whether the structure is monomaterial or multilayer, experimental tests show an efficiency of 89.07%.

[0048] Processing summary

[0049] The general scheme of the process carried out is shown in Fig. 3

[0050] The previous explanation describes the model training process. When a new measurement arrives at the THz-based system, classification is performed following the same steps used in training. The signal is processed in the same way as the training process, but with the models already loaded, allowing for faster prediction of new bins.

[0051] A single container can correspond to multiple measurements (several traces). The detection of a container's structure type (monomaterial or multilayer) is done by voting. That is, a series of measurements are collected from the sample (the number depending on the sample length and speed) and a vote is taken on the model's predictions for each sample point, resulting in the most accurate one.

[0052] In view of this description and figures, the person skilled in the art will understand that the invention has been described according to some preferred embodiments thereof, but that multiple variations can be introduced into said preferred embodiments, without exceeding the object of the invention as it has been claimed.

Claims

CLAIMS 1. Device for the identification and classification of samples of monomaterial versus multilayer plastics for recycling applications, comprising: a time domain THz spectroscopy system and a processor, wherein the spectroscopy system is provided with at least one optical source, at least one THz emitter, at least one THz receiver, optical elements for directing the signal from the emitter to the sample and from the latter to the detector, and at least one beam splitter for directing the signal reflected by the sample towards the receiving antenna, wherein the processor is provided with a machine learning classifier that takes as starting data the THz time trace of the signal reflected by the sample and measured by the receiving antenna to classify the sample as monomaterial or multilayer.

2. Device for the identification and classification of samples of monomaterial versus multilayer plastics according to claim 1, characterized in that the emitter emits in the range of 0.1 THz to 15 THz.

3. Device for the identification and classification of samples of monomaterial versus multilayer plastics according to claims 1-2, characterized in that the classifier is of the KNN type.

4. Device for the identification and classification of samples of mono-material versus multi-layer plastics according to claims 1-3, characterized in that the processor is provided with program means for selecting a set of points from the trace and sending the selection to the classifier.

5. Procedure for identifying and classifying samples of mono-material versus multi-layer plastics for recycling applications, which includes the following steps: - emit a signal in the THz range towards the sample; - collect the signal reflected by said sample; - take the time trace of the reflected signal as starting data and send it to a machine learning classifier that classifies the sample as monomaterial or multilayer.

6. Procedure for the identification and classification of samples of monomaterial versus multilayer plastics according to claim 5, characterized in that the emitter emits in the range of 0.1 THz to 15 THz.

7. Procedure for the identification and classification of samples of monomaterial versus multilayer plastics according to claims 5-6, characterized in that the classifier is of the KNN type.

8. Procedure for the identification and classification of samples of monomaterial versus multilayer plastics according to claims 5-7, characterized in that a set of points are selected from the trace and sent to the classifier.

Images