Method and system for training machine learning models for classifying components of material streams

Abstract

A method and system for training a machine learning model configured to characterize components in a material stream containing multiple unknown components includes determining a training reward associated with each unknown component among the multiple unknown components included in the material stream and physically separating at least one unknown component from the material stream using a separator unit based on the training reward. The separator unit is configured to move the selected at least one unknown component to another accessible partition. The separated at least one unknown component is analyzed to determine its ground truth label, and the determined ground truth is used to train incremental versions of the machine learning model.

Description

[Technical field]

[0001] The present invention relates to a method and system for training a machine learning model configured to characterize components in a material stream containing multiple unknown components. The present invention further relates to a computer program product. [Background technology]

[0002] Effective data classification plays an important role in many applications. For example, computer vision applications may apply classifiers and statistical models (e.g., machine learning models, regression models) to captured images and video streams to recognize components or objects. To ensure the reliability of the classifier, it is necessary to train it using multiple labeled examples. Such systems often manually label data and rely on human labor.

[0003] Important applications for characterizing one or more components in a material stream include, for example, recycling processes, agricultural processes, food production processes, etc. Characterization can be used, for example, in quality control, valuation, and process engineering and control. For example, in waste treatment, many waste streams are traditionally recycled in suboptimal ways due to lack of characterization data. To improve, suitable characterization techniques for heterogeneous material streams (e.g., bulk solid waste streams) are needed.

[0004] Typically, characterization of material streams involves manual inspection of objects by humans, e.g., factory personnel working in special facilities. This method is time-consuming, subjective, expensive, and ultimately provides little information about the particles in the material stream. In another conventional method, samples are taken and tested / analyzed, e.g., in a laboratory. This process can be time-consuming (chemical analysis can take days, weeks or months) and increases costs. Furthermore, only a small portion of the total amount of components / substances / objects in the material stream is characterized. Typically, many material streams are not optimally identified due to the difficulty of measuring the quality of the material. Fast, objective and / or automated methods that provide data at a more detailed level are needed. One example of material stream characterization is waste characterization.

[0005] Machine learning models are statistical classifiers that can be trained using large amounts of data that can be labeled by humans and / or experiments. Such labeling can be a labor-intensive and / or expensive process. One of the bottlenecks in building accurate statistical systems is the need to spend a long time (manually) labeling to obtain high-quality labeled data. Typically, the samples to be labeled (referred to as new data points) are selected randomly such that the training data matches the test set.

[0006] Thus, determining the ground truth in training a machine learning model is laborious and time-consuming in various cases and applications, and there is a strong demand for effectively obtaining a well-trained predictive model while reducing the required labor and cost. It is desirable to obtain such a predictive model more efficiently. Summary of the Invention

[0007] It is an object of the present invention to provide a method and system which avoids at least one of the above mentioned drawbacks.

[0008] Additionally or alternatively, it is an object of the present invention to improve the characterization of components in a material stream that contains multiple unknown components.

[0009] Additionally or alternatively, it is an object of the present invention to improve the efficiency of training machine learning models, such as component label prediction models.

[0010] Additionally or alternatively, it is an object of the present invention to provide improved waste disposal.

[0011] The present invention provides a method for training a machine learning model configured to characterize components in a material stream comprising a plurality of unknown components, the method including the steps of: scanning the material stream with a sensor system configured to image the material stream comprising the plurality of unknown components; predicting, for each of the plurality of unknown components in the material stream, one or more predicted labels and associated label prediction probabilities with a machine learning model configured to receive as input the images of the material stream and / or one or more features of the plurality of unknown components extracted from the images of the material stream; determining a training reward associated with each unknown component of the plurality of unknown components in the material stream; and determining at least one of the plurality of unknown components from the plurality of unknown components in the material stream based at least in part on the training reward associated with the plurality of unknown components. selecting another unknown component, wherein the selected at least one unknown component is physically separated from the material stream by a separation unit, the separation unit configured to move the selected at least one unknown component to another partition where it is accessible; analyzing the separated at least one unknown component to determine a ground truth label of the separated at least one unknown component, wherein the determined ground truth label of the separated at least one unknown component is added to a training database; and training an incremental version of the machine learning model using the determined ground truth label of the physically separated at least one unknown component.

[0012] The training reward may be a prediction of an improvement in the performance of a (machine learning) model / classifier by adding one or more ground truth labels bound to multiple components of the material stream to the database. The training reward may be a prediction of an increase in the performance of the machine learning model as indicated by a performance metric. For example, depending on the application, different performance metrics can be used. For example, the performance metric can be accuracy, purity, yield, etc. Many different performance metrics or scores can be used.

[0013] By isolating the selected components from the material stream, an analysis can be performed to determine ground truth labels. The obtained information can be fed back to the machine learning model during the training process. In this way, the accuracy of the progressively trained machine learning model can be improved. One or more components in the material stream that, if labeled for ground truth determination, would maximally improve the performance and / or accuracy of classification (see labeling prediction) by the machine learning model may be selected. The system may select and physically isolate the components that would result in improved model learning over other components in the material stream.

[0014] The manual / experimental effort for training a machine learning model can be effectively reduced by combining active learning, semi-supervised learning, and unsupervised learning techniques. The system includes a separation absorption unit for physically separating one or more unknown components for further analysis. The selection of one or more unknown components may be performed using confidence scores, predicted probabilities, entropy, density in feature space, etc.

[0015] The machine learning model may employ active learning for training. The machine learning model may be viewed as a learner that can actively select training data. Physical active learning may include means for physically isolating selected components in the material stream to be analyzed to provide selected training data. In active learning, the cycle of experimentation, learning the results, and selecting components for the next experiment is repeated, so that the total amount of experiments can be reduced. The learning of the results and the selection of components for the next experiment are performed by a computer. The system includes a separator unit for physically isolating the selected components from the material stream. More results can be obtained with fewer experiments or fewer amounts of experiments. This physical active learning can be utilized in experiment design to appropriately design experiments for analyzing components that are costly, laborious, and time-consuming.

[0016] The machine learning model may be configured to receive as input one or more user-defined features of the plurality of unknown components extracted from the images of the material stream.

[0017] The selection in the step of selecting the at least one unknown component may employ user-created selection criteria.

[0018] The separation unit may include multiple sub-units employing different separation techniques.

[0019] The separation unit may have at least a first sub-unit and a second sub-unit, and one of the first sub-unit and the second sub-unit may be selected for physical separation of the selected at least one unknown component based on the one or more features of the plurality of unknown components extracted from the image of the material stream.

[0020] Depending on certain characteristics of the unknown components, an appropriate sub-unit of the separation unit can be used to separate the unknown components from the material stream. For example, depending on the mass, size, etc. of the unknown components, different separation techniques may be required. For example, paper may be better separated using a fluid spraying means, whereas metal chunks may be better separated using mechanical means. As one or more features of the unknown components are extracted from the image of the material stream, this data can be obtained and advantageously used to select the appropriate sub-unit.

[0021] The first sub-unit may be used for the physical separation of smaller and / or lighter components in the material stream and the second sub-unit may be used for the physical separation of larger and / or heavier components in the material stream.

[0022] In some examples, the machine learning model can further control which separation technique is most appropriate to physically separate the selected unknown component from the material stream.

[0023] The first subunit may be configured to direct a fluid jet towards the components to separate the components in order to blow the components into another accessible compartment, and the second subunit may be configured to separate the components using a mechanical manipulation device.

[0024] The mechanical manipulation device of the second sub-unit may include at least one robotic arm.

[0025] Data indicative of mass may be calculated for each of the plurality of unknown components in the material stream.

[0026] Optionally, the components in the material stream may be scanned by a sensor system including an X-ray sensor configured for multi-energy imaging to obtain at least a low-energy X-ray image and a high-energy X-ray image. The image obtained by the sensor system may be segmented to isolate one or more distinguishable objects in the image, and data indicative of the area of ​​the segmented objects may be determined. For each of the segmented objects, data indicative of the area density and data indicative of the atomic number may be determined by analysis of the low-energy X-ray image and the high-energy X-ray image. The data indicative of the area density and the atomic number may be determined by a model that is calibrated by performing multi-energy X-ray imaging using a plurality of materials with different known area densities and atomic numbers. Data indicative of the mass of each of the segmented objects may be calculated based on the data indicative of the area density and the data indicative of the area.

[0027] A force generated by the fluid jet may be adjusted based on the mass of the selected at least one unknown component.

[0028] A value indicative of a difficulty of physically separating the unknown component from the material stream may be determined by the separation unit, and the difficulty value may be associated with each of the plurality of unknown components. The at least one unknown component may be selected from the plurality of unknown components in the material stream based on the value indicative of a difficulty of physically separating the unknown component from the material stream.

[0029] Ranking the difficulty of separation may greatly improve the efficiency of learning the machine model. Separating a component that is difficult to separate from the material stream may result in failure of the ground truth analysis. In such a case, the learning performance may be degraded as a result of not selecting other unknown component candidates even though they could have been selected. In the present invention, such a situation can be effectively prevented by considering the difficulty of separation as well. For example, when it is difficult to physically separate an unknown component, the extent to which the unknown component is obstructed (e.g., by surrounding objects) may play an important role. In addition, other components may be attached, which may affect the analysis. In some examples, a prediction or estimate of separation accuracy and / or separation purity is determined based on which at least one unknown component is selected from the multiple unknown components in the material stream for physical separation.

[0030] A plurality of top unknown components may be selected from the plurality of unknown components in the material stream based on the training rewards associated with the plurality of unknown components, and a subset of the plurality of top unknown components may be selected for physical separation based on a value indicative of the difficulty of performing physical separation by the separation unit.

[0031] If desired, a training reward may be calculated based on one or more reward indicators. It is also contemplated that a training reward may be provided by a user (e.g., an estimate based on expert knowledge).

[0032] A machine learning model or learning machine may be understood as a computational entity that relies on one or more machine learning algorithms to perform tasks that it is not explicitly programmed to perform. In particular, a machine learning model can adapt its behavior to the environment. In the context of characterization and detection of components in material streams, this adaptability is crucial since material streams often face changing conditions and requirements. The system may be configured to take in new incoming data to operate in real time. At any point in time, the knowledge of the machine learning model can be increased by adding new data points. In batch mode, large datasets can be collected and the entire dataset can be processed at once. In incremental mode, the machine learning model can be augmented with new data at any time (lightweight and highly adaptive).

[0033] Incremental versions of the machine learning model may be periodically trained using data obtained periodically from the analysis of the ground truth labels of the isolated at least one unknown component.

[0034] According to the present invention, by employing active learning, the number of training examples to be labeled can be significantly reduced. Thus, the most informative examples in relation to a given cost function are selectively sampled for human (e.g., at least partially manual) and / or machine labeling to explore unlabeled examples. Advantageously, active learning algorithms can provide a way to effectively select examples for (physical) labeling that provide the greatest performance improvement.

[0035] In some examples, the next component to be analyzed is selected based on its (cluster's) distance from other components in feature space. However, other techniques can be used. For example, statistical techniques can be employed in which components are selected and isolated for analysis such that at least one statistical property (e.g., learner variance) of the future machine learning model is optimized. In some examples, the selection of components is performed based on the level of discrepancy of the ensemble of classifiers. It will be appreciated that other techniques are also contemplated for determining the training rewards associated with each of the identified components in the material stream.

[0036] The plurality of unknown components may be divided into one or more clusters, each of the clusters including components having similar features and / or characteristics, and the identified unknown components may be assigned the training reward based at least in part on their distance from the one or more clusters.

[0037] The present invention allows for the control of the training reward by a certainty prediction in a machine learning model, e.g. having a neural network implementation. However, it will be appreciated that the training reward may be determined based on the uncertainty / confidence of the prediction of the predicted label by the machine learning model. Other implementations are also envisioned. In some cases, a component of a material stream may be observed that has not been previously processed by a machine learning model. In such cases, the model can determine with a fairly high degree of certainty that it belongs to a certain classification (predicted label), but in reality, the component belongs to a classification that the model has not yet observed. This may be overcome by analyzing the clusters in the feature space rather than the component's uncertainty. If a component is far from all currently observed clusters in the feature space, a selection can be made based on the component's location or distance to the cluster. This can be considered as anomaly or outlier detection.

[0038] Advantageously, diversity can be effectively taken into account in training a machine learning model. For example, a physical active machine learning system may select new unknown components (i.e., unlabeled observations) that are diverse across all observed unknown components (i.e., unlabeled observations). In this way, the (physical) active machine learning system can assemble a higher quality training set.

[0039] Clustering algorithms may be used to distinguish between different clusters and to ascertain whether the clusters are essentially different from each other. Components identified as being located deep within a cluster may have low uncertainty, while particles far from the core or between several clusters may have high uncertainty. Training rewards may be tied to the clusters to more accurately determine the clusters and their boundaries in one or more dimensions.

[0040] The training reward may be based at least in part on the confidence score.

[0041] A training score can be determined based on the uncertainty and diversity of each unlabeled data point associated with the identified component in the material stream. In some examples, the data points with the top n scores are selected in a batch, where n corresponds to a batch size. The batch size can be understood as the number of data points (see components) selected from the unlabeled material stream. The selected components may be separated from other unlabeled components in the material stream for manual and / or experimental annotation.

[0042] Physical active learning can be used to train machine learning models more efficiently: models can be trained on a selected set of multiple unknown components in the material stream using experimentally labeled data (e.g., manual judgment, automated judgment) rather than manually / human labeled data.

[0043] In an active learning process, the machine learning model may first be trained using a first set of ground truth data. This first set may be, for example, a small set generated manually or automatically by the model. A sensor system may be used to recognize one or more unknown components from the multiple unknown components as candidates for providing training data. For example, a training reward (e.g., a confidence measure) may be employed to predict candidates that are currently incorrectly recognized by the machine learning model. For example, the selected one or more unknown components may correspond to cases that are likely to have a recognition error by the used (trained) machine learning model. The one or more unknown components are then physically separated by a separation unit (e.g., a robotic configuration having one or more sensors for performing a separation task) to allow further analysis to determine the ground truth. For example, a human may manually verify the separated selected one or more unknown components. Additionally or alternatively, a machine and / or other sensor device may be used to experimentally determine the separated selected one or more unknown components.

[0044] Optionally, multiple components may be selected simultaneously and separated one instance at a time. This is relevant when there is a weak correlation between the feature space and the target (label / dependent variable) space. Optionally, selecting multiple components for separation may be based on the labels predicted by a machine learning model.

[0045] Optionally, the multiple identified unknown components may be ordered based on the training reward as candidates for selection to derive a selectively sampled order. The top identified unknown components following the selectively sampled order are separated and analyzed to determine their ground truth labels. Incremental versions of the machine learning model may be trained based on the ground truth labels.

[0046] Optionally, the machine learning model may be configured to employ pool-based active learning, where the machine learning model is exposed to a pool of unlabeled data points that are linked to identified components in the material stream. The machine learning model may be configured to iteratively select one or more components of the plurality of components in the material stream for at least partially manual and / or at least partially automatic (e.g., using a measurement device) annotation to determine a ground truth.

[0047] The accessible separate compartment may enable manual removal of the isolated unknown component, and an indication of internal references of the machine learning model may be provided for the isolated unknown component located within the accessible separate compartment, and the analysis of the at least one selected unknown component may be performed, at least in part, by human annotation.

[0048] Optionally, the machine learning model may be sequentially query-based, where one component at a time is selected and isolated for further analysis.

[0049] If desired, the machine learning model may be batch mode based, where a batch of components are selected and isolated for analysis (e.g., simultaneously) prior to updating the machine learning model.

[0050] The separated unknown components may be analyzed in an analytical device.

[0051] To determine the ground truth label based on a characterization, the analysis unit may be arranged to automatically perform the characterization of the isolated unknown component in the accessible separate partition. In some examples, the isolated unknown component may be automatically analyzed by the analysis unit.

[0052] The analysis unit may be configured to perform a chemical analysis of the separated components to determine the ground truth labels based at least in part on a chemical analysis of the separated components.

[0053] The analysis unit may be configured to perform destructive measurements of the separated components to determine the ground truth labels based at least in part on the destructive measurements of the separated components.

[0054] The analysis unit may be configured to perform at least one of energy or wavelength dispersive X-ray fluorescence spectroscopy (XRF), assays, inductively coupled plasma optical emission spectroscopy (ICP-OES), inductively coupled plasma atomic emission spectroscopy (ICP-AES), inductively coupled plasma mass spectrometry (ICP-MS), laser induced breakdown spectroscopy (LIBS), (near) infrared (NIR) spectroscopy, hyperspectral spectroscopy, X-ray diffraction analysis (XRD), scanning electron microscopy (SEM), nuclear magnetic resonance (NMR) and Raman spectroscopy. Also, measurement techniques may be combined to determine ground truth.

[0055] The analysis unit may be configured to perform measurements in an offline mode with respect to the sensor system. The analysis unit may also be configured to operate in a batch processing mode to determine ground truth labels of isolated objects. In some examples, the analysis unit is configured to perform measurements in near real-time with a time delay (e.g., several minutes). It will be appreciated that in some examples, the analysis unit may also be configured to provide relatively fast feedback, e.g., operating in real-time or quasi-real-time (e.g., online measurements).

[0056] In some examples, the analysis unit may be configured to perform delayed measurements (eg, non-real-time).

[0057] Optionally, the analysis unit is configured to perform discontinuous, periodic and / or intermittent measurements for determining a ground truth of the selected object. The measurement techniques performed by the analysis unit may require extended, continuous or relatively long measurement processes.

[0058] One or more measurement techniques employed by the analysis unit may require preparatory steps that may be time consuming and / or at least partially destructive. In some examples, one or more measurement techniques are not performed in real time. The measurement techniques employed may be relatively expensive and / or require manpower.

[0059] Optionally, the analysis unit performs non-imaging measurements. In some examples, the analysis unit does not perform (optical) imaging techniques, e.g. does not generate images. For example, the analysis unit may be configured to perform measurements based on chemical analysis.

[0060] Optionally, the sensor system may include an X-ray sensor configured to perform multi-energy imaging to obtain at least a low-energy X-ray image and a high-energy X-ray image. The image obtained by the sensor system is segmented to isolate one or more different unknown components in the image. Data indicative of an area of ​​the segmented object is determined. Data indicative of an area density and data indicative of an atomic number and / or chemical composition for each segmented unknown component may be determined by analysis of the low-energy X-ray image and the high-energy X-ray image. The data indicative of the area density and atomic number are determined by a model that is calibrated by performing multi-energy X-ray imaging using a plurality of materials with different known area densities and atomic numbers. Data indicative of a mass for each segmented unknown component may be calculated based on the data indicative of the area density and the data indicative of the area of ​​the segmented object.

[0061] The X-ray sensor may be a dual energy X-ray sensor.

[0062] The sensor system further includes a depth imaging unit for determining data indicative of a volume of the segmented object.

[0063] The depth imaging unit includes at least one of a 3D laser triangulation unit or a 3D camera.

[0064] The sensor system may further include a colour image processing unit configured to capture colour images of the segmented object.

[0065] Prior to determining features for each of the one or more segmented objects, data from different subsystems of the sensor system may be aligned.

[0066] For each of the one or more segmented objects, features relating to at least one of volume, size, diameter, shape, texture, color, or eccentricity may further be determined.

[0067] A material stream is transported by a conveyor and the material stream is scanned by a sensor system to characterize objects in the material stream.

[0068] Features of one or more segmented objects may be preserved to construct a digital twin model.

[0069] The material stream is characterized prior to transportation to determine a first digital identification marker and then transported to a remote location where the material stream is characterized to determine a second digital identification marker, and the first and second digital identification markers are compared to each other to determine changes in content during transport.

[0070] The material stream may be heterogeneous.

[0071] The material stream may be selected from the group consisting of solid waste, produce, agricultural products, or batteries.

[0072] Typically, the conventional approach to characterization of objects / components in material streams is manual human inspection of objects. This is often done in waste streams. Moreover, this is often done by superficial visual inspection of samples that are too small and therefore unrepresentative. Not only is this work tedious and time-consuming, but due to its subjective nature, the resulting conclusions are not always reliable. This work hinders the transition to a circular economy, since quality control is a key objective of characterization of material stream (e.g. waste stream) components, and fluctuations in the quality of second-hand goods reduce market interest. The present invention provides a fast, objective and accurate automated method that utilizes a more detailed level of data. The automated inspection is driven by artificial intelligence (AI), moving material stream component characterization to a data-driven and automated approach.

[0073] The machine learning model may be an online or continuous learning model, configured to be updated with each new analysis of the selected isolated unknown components (see samples). The analysis may be performed, for example, by classification by a user (e.g., at least partially manual) or fully automated using an analysis unit (e.g., experimentally determined). A combination of automated and manual analysis to determine ground truth labels is also envisioned.

[0074] If desired, a deep learning machine learning model is employed. Deep learning is a class of machine learning techniques that employs representation learning methods, allowing a machine to be fed with raw data and determine the representations required for data classification. Deep learning can identify structures in a dataset using a backpropagation algorithm that is used to modify the internal parameters (e.g., node weights) of the deep learning machine model. Deep learning machines may utilize a variety of multi-layer architectures and algorithms.

[0075] A deep learning neural network environment may contain many interconnected nodes called neurons. Externally activated input neurons activate other neurons based on their connections to other neurons governed by the parameters of the neural network. A neural network can behave in a certain way based on its own parameters. Training a deep learning model refines the model parameters that represent the connections between neurons in the network so that the neural network behaves in a desired way (e.g., performs better at the intended task, such as classifying components in a material stream).

[0076] Deep learning operates on the understanding that many datasets contain a hierarchy of features, from low-level features (e.g., edges) to higher-level features (e.g., patterns, objects, etc.). For example, when examining an image, instead of looking for objects, the model will look for edges that form motifs that form the objects being looked for. Learned observable features are the objects and quantitative regularities that the machine learning model has learned. Given a large amount of well-classified data, a machine learning model will be able to distinguish and extract the features needed to successfully classify new data.

[0077] The machine learning model may utilize a convolutional neural network (CNN). In some examples, deep learning can utilize segmentation of a convolutional neural network to find and identify learned observable features in the data. Each filter or layer of the CNN architecture can transform the input data to improve the (feature) selectivity and robustness of the data. Abstracting the data in this way allows the machine to focus on the features in the data it is trying to classify and ignore irrelevant background information. Deep learning machine models using convolutional neural networks (CNN) can be used for image analysis.

[0078] According to another aspect of the present invention, there is provided a system for training a machine learning model configured to characterize components in a material stream including a plurality of unknown components.The system includes a processor, a computer readable storage medium, a sensor system, and a separate unit, the computer readable storage medium storing instructions that, when executed by the processor, cause the processor to: operate the sensor system to scan the material stream to image the material stream including the plurality of unknown components; predicting one or more labels and associated label probabilities for each of the plurality of unknown components in the material stream by a machine learning model configured to receive as input the image of the material stream and / or one or more features of the plurality of unknown components extracted from the image of the material stream; determining a training reward associated with each unknown component in the plurality of unknown components in the material stream; and calculating a training reward at least equal to the training reward associated with the plurality of unknown components. and operating the separation unit to physically separate the selected at least one unknown component from the plurality of unknown components in the material stream, the separation unit moving the selected unknown component to another accessible partition. The separation unit further includes instructions to execute the steps of: selecting at least one unknown component from the plurality of unknown components in the material stream based at least in part on the selected unknown component; operating the separation unit to physically separate the selected at least one unknown component from the material stream, the separation unit moving the selected unknown component to another accessible partition; receiving, for the separated at least one unknown component, the ground truth label determined by performing an analysis, the receiving step adding the determined ground truth label for the separated at least one unknown component to a training database; and training an incremental version of a machine learning model using the ground truth label determined for the physically separated at least one unknown component.

[0079] The method for selecting specific components of the material stream to include in the training set may be based on an estimate of the "reward" (estimated performance increase) obtained by including each identified component in the training set. The selected specific components of the material stream may then be analyzed in isolation to further train the machine learning model. The reward may be based on the uncertainty associated with unlabeled components of the material stream. However, the training reward may also be based on the identified clustering of components in the feature space.

[0080] Active learning is a specific field of machine learning where an algorithm can interactively query an information source to obtain a desired output (e.g., at least one of material properties, material type, material features, chemical analysis, color, shape properties, mass, density, etc.) for new data points. In the physical active learning provided by the present invention, a separator unit is used to physically separate one or more selected unknown components in a material stream for further analysis that provides one or more new data points. The physical active learning model can determine the measurement to be made according to a training reward (e.g., a weighted score) that indicates the "optimality" of the input data points. In some examples, this training reward may be determined and / or calculated using only the input data information.

[0081] A user (e.g., an expert or operator) may impose additional criteria on the selection of components for separation and analysis. In some examples, a training reward is not always calculated, but may be assumed by the user. Components may be selected, separated and analyzed to determine a ground truth for training a machine learning model based on predetermined (e.g., experience / knowledge-based) assumptions. Physical separation of components in a material stream can be performed, for example, based at least in part on the characteristics (e.g., shape, density, ...) of the components.

[0082] In some examples, the sensor system includes one or more imaging devices, such as a 2D-camera, a 3D-camera, an X-ray imaging system, etc. Other imaging devices can also be used, such as a computed tomography (CT) system, a magnetic resonance imaging (MRI) system, etc. A combination of imaging devices can also be used, such as a 3D camera system in combination with an X-ray system.

[0083] According to one aspect, the present invention provides a system and method having means for selectively isolating one or more components in a material stream for further analysis to determine a ground truth label. One or more components are selected based on their particular characteristics, and the ground truth label is used for active learning training of a machine learning model used to obtain a predictive label for each of the components in the material stream. In some examples, the selection of one or more components for separation and further analysis for ground truth determination may be based, for example, on selection criteria provided by a user (e.g., selection of components with high density, components with certain visual characteristics such as color, components with a particular shape, etc.). For example, an expert may estimate whether determining the ground truth for the selected components will provide a significant training reward for the model. Advantageously, a separation device may be obtained with a sensor that provides a physical separation of the selected components used to train the machine learning model.

[0084] Instead of learning from a random set of examples (passive learning), machine learning models can be made to operate on examples that are labeled, which can be considered active learning. With active learning, better performance can be obtained using a subset of the training data. The present invention employs physical active learning, where a separator unit is provided that is arranged to physically separate components from the material stream.

[0085] It will be appreciated that a machine learning model may use the processing power of a computer to run algorithms to learn predictors of data behavior or characteristics. Machine learning techniques may run algorithms on a set of training samples (training set) with known classifications or labels, such as a set of components known to exhibit certain characteristics / features, to learn features that predict the behavior or characteristics of unknowns, such as whether an unknown component belongs to a particular classification or group.

[0086] It will be appreciated that labeling may be performed in a variety of ways and by a variety of entities, for example, labeling may be performed by a machine learning model (i.e., providing predicted labels), whereas labeling may also be performed to determine ground truth labels (e.g., performed by an analyst, a human annotator, an experimental setup, etc.).

[0087] It will be appreciated that the training reward can be considered as a training reward or an active reward. It can be understood as a prediction and / or suggestion of how much its performance would improve if labeled with ground truth and used to train a machine learning model. The training reward may indicate an improvement of the machine learning model by training it using the determined ground truth labels associated with the selected components of the material stream. The training reward can be understood as a learning reward in the machine learning process.

[0088] It will be appreciated that a variety of active learning techniques can be implemented. The active learning technique may be configured to select the action that provides the greatest gain in knowledge or "know-how" when selecting the training set. Active learning techniques may vary with respect to the way in which "knowledge" and knowledge gain are quantified. There may also be a variety in the way in which it is determined which action is likely to lead to the greatest knowledge gain. A variety of implementations are possible.

[0089] It will be understood that any aspect, feature and option described in terms of the method applies equally to the system of the invention and the described recycling device. It will also be apparent that any one or more combinations of the above aspects, features and options are possible. [Brief description of the drawings]

[0090] The invention will be further explained on the basis of exemplary embodiments shown in the drawings, which are given by way of non-limiting example, and it should be noted that the drawings are only schematic representations of embodiments of the invention and are given by way of non-limiting example.

[0091] [Figure 1] 1 shows a schematic diagram of one embodiment of a system. [Diagram 2] 1 shows a schematic diagram of one embodiment of a system. [Diagram 3] 1 shows a schematic diagram of one embodiment of the method. [Figure 4] An example of a feature space is shown. [Diagram 5] 1 shows the distribution of features for several different component types. [Figure 6] 1 illustrates an exemplary learning process indicator. [Figure 7] A schematic diagram of the system is shown. [Figure 8] A schematic diagram of the method is shown. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0092] In supervised machine learning, a model is trained on a (large) material stream where each object is labeled. The labels may indicate the type of material (metal, wood, glass, ceramic, etc.) of each component / object identified in the material stream and can be used to teach the machine learning model how to correctly classify the components / objects in the material stream. Since painstakingly selecting thousands of components or particles from a stream of multiple heterogeneous materials is a time-consuming and costly task, determining and / or preparing labeled data is often the bottleneck in the training process. While unlabeled data is abundant and easy to obtain in a material stream, labeled data may be scarce and difficult to obtain. Moreover, the entire labeling process may have to be repeated from start to finish every time a new material stream is considered. The present invention employs a data-driven characterization of components in a material stream that can significantly reduce labeling costs while essentially achieving the same accuracy as a supervised model that uses the entire training dataset. By employing active learning, the machine learning model itself is able to select an optimal small subset of components (objects, particles, etc.) in the material stream that need to be labeled. Training a machine learning model using only this small labeled subset results in a model with performance comparable to that of a model trained using all of the components in the material stream.

[0093] 1 is a schematic diagram of an embodiment of a system 1 for training a machine learning model configured to characterize components in a material stream 3 that includes an unknown component 3i. The system 1 includes a processor, a computer-readable storage medium, a sensor system 5, and a separator unit 100. The computer-readable storage medium stores instructions that, when executed by the processor, cause the processor to perform the following steps: operate the sensor system 5 to scan the material stream 3 to image the material stream 3 having a plurality of unknown components 3i; predict one or more labels and associated label probabilities for each unknown component 3i in the material stream 3 by a machine learning model configured to receive as input an image of the material stream 3 and / or one or more features of the unknown components extracted from the image of the material stream 3; determine a training reward associated with each component 3i of the plurality of unknown components 3i in the material stream 3; and characterize at least one unknown component 3i from the plurality of unknown components 3i in the material stream 3 based at least in part on the training reward associated with the unknown component 3i. The method includes instructions to perform the steps of: selecting a component; operating the separator unit 100 to physically separate the selected at least one unknown component from the material stream 3, where the separator unit 100 moves the selected unknown component to another accessible partition 101; receiving a ground truth label determined for the separated at least one unknown component by performing the analysis, where the ground truth label determined for the separated at least one unknown component is added to a training database; and training an incremental version of a machine learning model using the ground truth label determined for the physically separated at least one unknown component.

[0094] In one example of this embodiment, the separator unit comprises a robotic arm that automatically separates selected components into compartment 101. It will be appreciated that other means may be employed to selectively move selected components from the material stream 3 into compartment 101 for further analysis with respect to the ground truth determination. This may be performed in a variety of ways, including, for example, robotic means that physically perform the separation. A variety of other techniques may also be employed. For example, ejection of selected components from the material stream may be accomplished by air jets (e.g., using air nozzles). Also, a combination of techniques may be used (e.g., depending on the size of the parts to be separated / isolated from the material stream). For example, larger parts may be physically separated using a robotic arm, and smaller parts may be separated by fluid jets using fluid nozzles.

[0095] Due to the large number of components contained in the material stream 3, it may be impractical for a human being to manually label each component (large data set). To optimize the labeling task involved in training a data classifier, an active learning method is employed to select only promising and exemplary components for manual labeling. The selected components in the material stream are automatically physically separated by a separator unit 100, which in this embodiment is deployed by a robotic arm. However, as mentioned above, one or more other means may also be employed.

[0096] The machine learning model may be an active learner that applies a selection function to physically separate components for labeling. Based on the selection, components may be separated from the material stream 3 into a separate accessible compartment 101 for manual and / or experimental labeling to determine the ground truth. The machine learning model (see classifier) ​​may be retrained with the newly labeled data, and this process may continue, for example, until a predefined stopping criterion is met. Since the components to be labeled for training the machine learning model are selected and separated based on the training reward, the time-consuming process of retraining the classifier on new data points may be avoided. Thus, the machine learning model may be trained more efficiently.

[0097] FIG. 2 is a schematic diagram illustrating an embodiment of the system 1 similar to the embodiment shown in FIG. 1. In this embodiment, the separator unit 100 includes an operable lid 103 disposed in the path of the material stream 3. For example, the material stream may be transported by a conveyor belt on which the operable lid 103 is disposed. The system may be configured to selectively open the lid 103 to separate one or more components in the material stream 3. An optional optical unit 105 (e.g., a camera) may be used to detect when the lid 103 is opened to separate one or more components from the material stream 3. It will be appreciated that other variations are possible, for example, not using the optical unit 105. The optical unit 105 may be optional, for example, in some exemplary embodiments, data from the sensor system 5 may be used to detect when the lid 103 is opened to separate one or more components from the material stream 3. In some examples, the optional optical unit 105 may also be disposed upstream to increase the reaction time when the lid 103 is opened.

[0098] The most relevant data points associated with the multiple identified components in the material stream can be selected to isolate, manually and / or experimentally label the components and determine the ground truth. The resulting ground truth can then be used to further train the machine learning model. The selection is based on the training reward, thereby obtaining maximum generalization ability of the machine learning model with minimal human labeling effort.

[0099] 3 is a schematic diagram of one embodiment of method 20. The method employs active machine learning such that a set of samples of material stream 3 for which it is desired to receive training data is selected, rather than passively receiving samples selected by an external entity. For example, this allows the machine learning model to select the samples it determines are most useful (relevant for training) for learning, rather than relying solely on an external human expert or external system to identify and provide samples for the machine learning model to learn from.

[0100] The pool-based active learning cycle is illustrated in FIG. 3. A machine learning model 23 may be trained using a labeled training set 21. The machine learning model may be presented with an unlabeled pool 25. The machine learning model may predict labels and training rewards to be associated with components in the material stream. Queries may then be selected for analysis 27 (human annotation and / or experimentation). The selection may be based on the training reward. The selected components may be physically separated for labeling. The results of the analysis / labeling may be used as a further training set (see labeled training set 21) for the machine learning model.

[0101] Active learning or query learning can overcome the labeling bottleneck in the training process by using queries as unlabeled examples that are to be labeled by an oracle, i.e. a human annotator and / or an automated analyzer. In this way, an active learner aims to achieve high accuracy using as few labeled instances as possible, thereby minimizing the cost of obtaining labeled data. Many query strategies exist. For example, the so-called pool-based active learning is employed, where the training data is split into a (small) labeled dataset on the one hand and a large pool of unlabeled instances on the other hand. The active learner may operate greedily. It may select the samples to query the annotator by simultaneously evaluating all instances in the unlabeled pool. The components (samples) that maximize a certain criterion are sent to the oracle for annotation and added to the labeled training set. Updated results sent by the model allow the active learner to make new choices of queries for the human annotators.

[0102] The active learner can employ one or more criteria to select new components to separate and analyze for annotation. There are various approaches to this. The query strategy employed in some advantageous embodiments is based on uncertainty sampling. The active learner queries the instances of the unlabeled pool for which it is most uncertain how to label. In the following equation, let x be a feature vector describing a component present in the unlabeled pool of components of the material stream. Under the model θ, the material classification, i.e., the label of the particle, can be predicted as the class with the highest posterior probability among all classes y.

number

[0103] One example of a query strategy is to select the component with the lowest prediction confidence by computing equation (1) above for all components in the unlabeled pool and then selecting one according to equation (2).

number

[0104] This criterion is equivalent to the machine learning model choosing the sample that is most likely to mislabel x, i.e., the most likely sample to be labeled is the least likely sample among the unlabeled components available for the query. The drawback is that the machine learning model only considers information about the most likely label, discarding information about the rest of the posterior distribution.

[0105] An alternative sampling strategy that overcomes the above drawbacks is to use the Shannon entropy as an uncertainty measure.

number

[0106] An example of a feature space is shown in Figure 4. The present invention employs physical active learning where only those components / objects in the material stream are selected and isolated to determine ground truth labels that are then used for further training of the machine learning model that was used to predict the labels associated with the components / objects in the material stream. In this way, automated analysis of the specific selected and isolated components / objects can be used to more effectively train a machine learning model (e.g., a classification model).

[0107] Thereafter, determining the ground truth label for each component individually may be too laborious. Advantageously, models can now be trained very well with less data. The system can automatically select and isolate components in a material stream for further analysis to determine the ground truth label. This is highly useful, for example, for waste treatment processes involving one or more waste streams. For example, the system may be configured to perform waste characterization, which allows for efficient further training of the employed machine learning model. Furthermore, in some examples, the system may be configured to perform material sorting based on the waste characterization. It will be appreciated that the present invention can be used in other applications, such as characterizing other material streams.

[0108] The determination of the ground truth may be established in a variety of ways, including, for example, partially manual labeling (e.g., analysis is performed at least partially by a human). It may also be determined automatically, for example, by performing chemical experiments. A combination of approaches may also be employed, for example, where different techniques are required, such as when determining different characteristics to derive the ground truth label. Different characteristic parameters may be determined to determine the ground truth (e.g., mass, chemical reaction, weight, geometric characteristics, etc.).

[0109] A material stream may be a flow of heterogeneous substances or components. Different algorithms and techniques can be used to determine which particles contribute most to training the machine learning model. For this purpose, different active learning methods can be applied.

[0110] Different strategies can be adopted (e.g., by analysis) to select the next points for ground truth labeling. In the example shown in Fig. 4, the system is configured to operate the separator unit 100 based on the distance to the cluster to separate one or more components selected for ground truth labeling (see sampling). For example, various approaches can be used, such as: -Select the sample that is located farthest from all cluster centers, which allows to discover potential new (sub)classes. - Select samples located between clusters (e.g. samples that are equidistant to two clusters), which allows refining the inter-class decision. Select samples that are located farthest from the majority of samples / clusters (i.e., isolated samples). This allows identifying outliers / anomalies that may represent new (sub)classes.

[0111] Combinations of the above techniques may also be used, and it will be appreciated that other selection strategies may also be employed.

[0112] FIG. 5 shows the distribution of features for different component classes. Components 3i in material stream 3 can be classified into different classes, e.g., paper, wood, glass, stone, ferro, and non-ferro. Exemplary classes are shown in FIG. 5. The machine learning model may be a classification model trained to distinguish between these different classes. The graphs in FIG. 5 show the univariate and bivariate distributions of four features. As expected, some features are better at identifying certain materials than others. For example, atomic number is good at identifying paper and non-ferro, but not stone and glass. However, the opposite is true for average density. In this example, a total of 31 features are combined to train the classifier, maximizing learning capacity.

[0113] The diagonal graphs show kernel density estimates of the distribution of the four features from the dataset. The other graphs are scatter plots of the individual features, with the mean atomic number <z>, the logarithm of the mean density 〈ρ〉, the logarithm of the standard deviation of the height 〈σ〉, and the logarithm of the perimeter of the component.

[0114] Selection and isolation of components in the material stream for ground truth label analysis can be based on the confidence of the current machine learning model (see classifier) ​​in the identified unlabeled components in the material stream.

[0115] Figure 6 shows various graphs of the learning process indicators. In the top part of Figure 6, learning curves are shown for several example models employing different criteria for selecting new components to be separated and analyzed for annotation (random sampling and uncertainty sampling based on minimum confidence and entropy criteria, see equations (2) and (3) respectively). These curves show how the test performance of each model changes as a function of the number of queried (selected, separated, analyzed for classification) components in material stream 3 or as a function of the size of the labeled training set. In this example, the sample size is incremented by 1 and a support vector machine (SVM) with a radial-based kernel is used as the classifier.

[0116] We compare uncertainty sampling based on the reliability criterion in equation (2) with random sampling based on the entropy criterion in equation (3), where in the latter case, material stream components are queried completely randomly rather than based on some uncertainty criterion.

[0117] It is generally believed that models perform better with larger sample sizes because more labeled data contains more information. However, this does not happen at the same speed for all models. The graph shows that while entropy- and confidence-based sampling techniques perform comparable, random sampling is clearly inferior in classifying the components of the material stream. In the limit of large sample sizes, all model performance converges to the “optimum” value of the model using the entire training dataset. This performance is what active learning models must compete against, and is shown in Figure 6 as the baseline accuracy of 0.988. It is clear from Figure 6 that as the performance of the active learner approaches the baseline accuracy, more samples are required to achieve small absolute gains. For example, uncertainty sampling requires 77 labeled instances to reach 99% of the baseline accuracy and 195 labeled samples to reach 99.9% of the baseline accuracy. In other words, the number of labeled samples must more than double to achieve an accuracy improvement of only 0.9 percentage points. This raises an important practical question: what is the trade-off between the cost of labeling and the potential cost of making misclassifications?

[0118] The curves in the graph show the average results for 250 different random initial conditions, with the boundaries of the shaded region defined by the 10% and 90% quantiles. Furthermore, a cross-section of the feature space defined by the mean atomic number Z and density ρ at three different stages of the learning process is shown. The first column shows which samples have been queried up to that point. The second and third columns show the behavior of the minimum confidence and entropy measures in this two-dimensional cross-section of the feature space. The remaining unlabeled samples are also shown, with the ones with the most uncertainty marked with a cross. This is the next component that should be analyzed in isolation (e.g., by human annotators or experiments).

[0119] The other graphs in Figure 6 show three positions on the learning curve that have been further explored. They show a cross-section of the 31-dimensional feature space, expressed as the mean atomic number Z and the logarithm of the density ρ. The first column shows which samples have been labeled up to that point, for each of the three stages of the learning process. At the start of active learning, the samples are roughly evenly depicted in the space, which is why the difference with the random model is not very large at this point.

[0120] However, as more data becomes available, the active learner starts to recognize the boundary regions between different material classes and primarily queries samples that are in the immediate vicinity of these class boundaries. This phenomenon can be observed in the second and third columns of Figure 6, which represent the behavior of the confidence and entropy sampling criteria for a 2D model, respectively. As the labeled training data increases, the boundaries separating different material classes become more pronounced as regions with higher uncertainty. Although the class boundaries appear smoother for the entropy criterion than for the confidence criterion, both have roughly the same shape. This explains why similar samples are queried and the performance of both is roughly the same.

[0121] In general, the optimal choice of uncertainty metric strongly depends on the dataset at hand. However, if classification is performed simply by majority voting, a confidence criterion is probably appropriate: a component is assigned to the class with the highest posterior probability. However, if more complex rules are used (e.g., in the case of imbalanced datasets), entropy is undoubtedly the more obvious choice.

[0122] FIG. 7 shows a schematic diagram of an embodiment of the system 1. In this example, at least one of an optional (color) camera 7 and an optional 3D laser triangulation unit 9 are arranged to determine additional characteristics attached to each of the compartmentalized objects. Thus, in some examples, it is also possible to distinguish the identified and / or segmented objects based on at least one of size, shape, color, texture, visual insight, etc., next to features / characteristics related to material type, mass, etc. Such information may also enable virtual experiments. In this example, the sensor unit 5 includes an X-ray sensor 11 having two X-ray subunits 11a, 11b for performing dual energy X-ray imaging. Furthermore, the camera 7 and the 3D laser triangulation unit 9 are integrated with the sensor unit 5. In this way, the sensor unit 5 provides multiple images that can be aligned and / or combined by, for example, a computer unit 13. Alignment and / or combination of image data obtained from different cameras / detectors allows for better determination of features / characteristics of the segmented objects. One or more materials are segmented and the individual segmented objects 3i are analyzed to determine their relevant features / characteristics. In this example, features 15 such as density, material, shape, size and mass are determined for each segmented object. It will be appreciated that other sets of features are possible. It is also possible to derive the relative weight percentage of each segmented object from the data.

[0123] The system according to the present invention can provide important advantages in waste characterization applications, as it can characterize one or more materials faster and more autonomously while requiring less human (labor-intensive) input.

[0124] To develop a model that can recognize images of waste and classify them into different categories, a machine learning model can be trained by showing it many images and labeling each image with a label describing what is in it. This traditional approach, where all data is pre-labeled, is called supervised learning. Labeled data fuels the machine learning algorithm. In waste characterization techniques, labeled data is typically generated by scanning a physical "pure" single-material stream, which is often prepared manually by painstakingly selecting thousands of individual particles from heterogeneous waste streams.

[0125] In the recycling industry, waste characterization has several important applications: It can be used for valuation: Rapid and reliable characterization of the complete material stream reduces the exposure to commodity stock market fluctuations; It can be used for quality control: In a circular economy, it is desirable to have guaranteed quality of recycled products; Characterization techniques help to establish market confidence; It can be used for process engineering: The technical and economic feasibility of waste recycling processes can be evaluated as well as new process designs by virtual experiments; It can be used for online process optimization: Sorting processes can be measured, controlled and optimized in situ.

[0126] In some instances, direct in-line characterization techniques can be provided to assess materials qualitatively (material type, chemistry, purity, etc.) and quantitatively (mass balance, physical properties, etc.). Such in-line characterization systems can be configured to fully assess heterogeneous and complex material streams, eliminating the need for subsampling. Furthermore, mass balances can be created on the fly. In fact, for each material object, a digital twin can be created and virtually evaluated.

[0127] The present invention provides data-driven material characterization using physical active learning that can significantly reduce the labeling effort in collecting training data. Traditional machine learning algorithms require large, fully labeled data sets for training, but it has been shown that only a small portion of this data is needed to make good predictions. Active learning allows models to be trained on a small subset selected by the algorithm, and can achieve the same accuracy as if the model was trained on the full data set. In some examples, active learning can reduce labeling costs by 70% while maintaining 99% accuracy of training on a fully labeled data set.

[0128] It will be appreciated that the systems and methods according to the present invention can be used with a variety of material streams. In some examples, material streams include construction and demolition waste, however, other waste streams can also be used.

[0129] FIG. 8 is a schematic diagram of the method 30. In a first step 31, one or more material objects or components are identified and segmented. This can be performed by object detection algorithms and / or segmentation algorithms. Images are acquired using the sensor unit 5. The segmented images may also be acquired, for example, after performing an alignment and / or a synthesis of different images provided by different sensors or sub-units of the sensor unit 5. In this example, a box 20 is placed around the segmented objects 3i. In a second step 33, properties / features 15 are determined for each of the segmented objects 3i. In this example, mass, volume, atomic number are determined. In a third step, a label may be predicted by a machine learning model. As shown in step 37, this can be done by providing the data as input to a neural network 25 trained to obtain a (predicted) label 17 as output. In this example, the trained neural network is a deep learning model. However, other machine learning models can also be used, such as, for example, support vector machines (SVMs), decision tree-based learning systems, random forests, regression models, autoencoder clustering, nearest neighbor (e.g., kNN) machine learning algorithms, etc. In some instances, instead of artificial neural networks, alternative regression models are used.

[0130] The invention provides for more efficient training of the machine learning model (e.g., deep neural network) used. Active learning allows for a reduction in the number of training samples to be (manually) labeled by selectively sampling a subset of unlabeled data (in the material stream). This can be reduced by inspecting the unlabeled samples and selecting the most informative with respect to a given cost function for human and / or experimental labeling. The active learning machine learning model can select the samples that can provide the greatest improvement in performance, thereby reducing the effort of human and / or experimental labeling. Selectively sampling components from a plurality of components in the material stream assumes that there is a pool of candidate components to label among the plurality of components. Since there may be a constant stream of new and relatively unique components in the material stream, the stream provides a source for continuously and effectively improving the performance of the machine learning model. Advantageously, the selected components may be automatically separated by the system using a separation unit. The active learning machine model may derive a small subset of all components collected from the material stream for human and / or experimental labeling.

[0131] For example, a classified data set obtained by human annotation can be used to train an initial deep learning neural network. This data set builds the initial parameters of the neural network, which is the supervised learning stage. During the supervised learning stage, the neural network can be tested to see if the desired behavior is achieved. Once the desired neural network behavior is achieved (e.g., once the machine learning model is trained to operate according to a specified threshold), the machine learning model can be deployed for actual use (e.g., testing the machine on "real" data). During operation, the classifications made by the neural network can be confirmed or denied (e.g., by an expert user, an expert system, a reference database, etc.) to continue to improve the operation of the neural network. The neural network in this example is then said to be in a state of transfer learning, since the parameters for the classifications that determine the behavior of the neural network are updated based on ongoing interactions. In some examples, the neural network of the machine learning model can provide direct feedback to another process, for example, to change the control parameters of a waste recycling process. In some examples, the neural network outputs buffered and verified data (e.g., via the cloud, etc.) before providing it to another process.

[0132] Data acquisition can be done in a variety of ways. A sensor system may include a variety of sensors. As an example, data on material properties of particles in a material stream (e.g., waste stream) is collected by a multi-sensor characterization device. First, dual energy X-ray transmission (DE-XRT) may be used to "see through" the material and determine certain material properties, such as average atomic number and density. The advantage of this is that the entire volume of the part can be inspected, not just the surface (e.g., waste is often dirty and surface properties are not necessarily representative of the bulk of the material). Next, additionally or alternatively, a 3D laser triangulation unit may be utilized to measure the shape of the object at high resolution (e.g., sub-millimeter accuracy). This can provide additional information, such as three-dimensional shape and volume, that complements the information collected from the DE-XRT. Then, additionally or alternatively, an RGB detector may be used, which allows components contained in the material stream to be differentiated in terms of color and shape. In some examples, multiple sensors as mentioned above are used together. If necessary, image processing may be used to segment the image into individual components. From these segmented images, various features describing the shape of the object may be calculated. For example, component area, eccentricity, perimeter, etc. In some examples, this may be done for all images obtained from all sensors.

[0133] Various neural network models and / or neural network architectures can be used. Neural networks have the ability to process, e.g., classify, sensor data and / or pre-processed data (e.g., feature characteristics determined for segmented objects). Neural networks may be implemented in computerized systems. Neural networks may serve as a framework for various machine learning algorithms to process complex data inputs. Such neural network systems are generally not programmed with task-specific rules, but can "learn" to perform a task by considering multiple examples. Neural networks are based on a collection of connected units or nodes called neurons. Each connection can transmit a signal from one neuron to another in the neural network. A neuron that receives a signal can process the signal and send it to further neurons connected to it (see activation). The output of each neuron is usually calculated by a nonlinear function of the sum of the inputs. The multiple connections can have respective weights, which are adjusted as learning progresses. There may also be other parameters, such as biases. Typically, neurons are aggregated into layers. Different layers can perform different types of transformations on their inputs, forming deep neural networks.

[0134] A deep learning neural network can be considered as a representation learning method with multiple levels of representation. This representation learning method is obtained by composing simple but non-linear modules that convert one level of representation starting from the raw input into a higher, somewhat abstract level of representation, respectively. Neural networks can identify patterns that are difficult to identify using traditional methods. Thus, instead of writing custom code specific to the problem of printing a structure in certain printing conditions, a network can be trained to handle different and / or changing printing conditions of the structure, for example using a classification algorithm. The neural network may be fed with training data so that it can determine a classification logic for efficiently controlling the printing process.

[0135] When a step of a method is described as being performed after another step, it means that the other step may be followed directly by the other step or that one or more intermediate steps may be performed before the step is performed, unless otherwise specified. Similarly, when a connection between components such as neurons of a neural network is described, it will be understood that this connection may be established directly or through intermediate components such as other neurons or logical operations, unless otherwise specified or excluded by context.

[0136] It is clear that the term "label" can be understood as both a categorical variable (e.g., using a neural network) and a continuous variable (e.g., using a regression model). For example, a continuous variable may have uncertainty (e.g., a chemical analysis variable).

[0137] It will be appreciated that the method may include computer-implemented steps. All steps described above may be computer-implemented steps. The embodiments may include a computer device for carrying out the process. The invention also extends to a computer program adapted for carrying out the invention, in particular a computer program on or in a carrier. The program may be in the form of source code or object code or in any other form suitable for use in implementing the process according to the invention. The carrier may be any entity or device capable of carrying a program. For example, the carrier may be a storage medium such as a ROM, for example a semiconductor ROM, or a hard disk. Furthermore, the carrier may be a transmissible carrier, such as an electrical or optical signal, and may be conveyed by means such as electrical or optical cable, wirelessly, for example via the internet or the cloud.

[0138] Some embodiments may be implemented, for example, using a machine or tangible computer readable medium or element capable of storing instructions or sets of instructions that, when executed by a machine, can cause the machine to perform methods and / or operations in accordance with the embodiments. Various embodiments may be implemented using hardware elements, software elements, or a combination of both. Examples of hardware elements include processors, microprocessors, circuits, application specific integrated circuits (ASICs), programmable logic devices (PLDs), digital signal processors (DSPs), field programmable gate arrays (FPGAs), logic gates, registers, semiconductor devices, microchips, chipsets, etc. Examples of software may include software components, programs, applications, computer programs, application programs, system programs, machine programs, operating system software, mobile apps, middleware, firmware, software modules, routines, subroutines, functions, computer implemented methods, procedures, software interfaces, application program interfaces (APIs), methods, instruction sets, computing code, computer code, etc.

[0139] The present invention is described herein with reference to specific examples of each embodiment of the present invention. However, it will be apparent that various changes, modifications, substitutions, and alterations can be made in the present invention without departing from the essence of the present invention. Although features are described herein as part of the same or separate embodiments for clarity and concise description, alternative embodiments having all or a partial combination of the features described in these separate embodiments are also assumed and understood to be within the framework of the present invention outlined by the claims. Therefore, each specification, drawing, and example should be interpreted in an illustrative sense, not a restrictive sense. The present invention is intended to embrace all alternatives, modifications, and variations that are within the scope of the appended claims. Also, many of the described elements are functional entities that can be implemented as separate or distributed components, or in combination with other components, in any suitable combination and location.

[0140] In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of other features or steps than those listed in the claim. Moreover, the words "a" and "an" shall not be construed as being limited to "only one" but are used to mean "at least one" and do not exclude a plurality. The mere fact that certain measures are recited in mutually different claims does not indicate that a combination of these measures cannot be used to advantage.< / z>

Claims

1. 1. A method for training a machine learning model configured to characterize components in a heterogeneous material stream containing a plurality of unknown components, the method comprising: scanning the material stream with a sensor system configured to image the material stream including the plurality of unknown components; predicting, for each of the plurality of unknown components in the material stream, one or more predicted labels and associated label prediction probabilities by a machine learning model configured to receive as input the images of the material stream and / or one or more features of the plurality of unknown components extracted from the images of the material stream; determining a training reward associated with each unknown component in the plurality of unknown components in the material stream; selecting at least one unknown component from the plurality of unknown components in the material stream based at least in part on the training reward associated with the plurality of unknown components; determining a ground truth for the at least one unknown component requires analysis in physical separation, and a selected at least one unknown component is physically separated from the material stream by a separation unit, the separation unit configured to move the selected at least one unknown component to another compartment where it is accessible; The method further comprises: analyzing the at least one isolated unknown component to determine a ground truth label for the at least one isolated unknown component; training an incremental version of the machine learning model using the determined ground truth label of the at least one physically separated unknown component; A method comprising: performing a chemical analysis of the at least one unknown component separated from the material stream to determine the ground truth label based at least in part on a chemical analysis.

2. 2. The method of claim 1, wherein the machine learning model is configured to receive as input one or more user-defined features of the plurality of unknown components extracted from the images of the material stream, and employ user-created selection criteria for selection in the step of selecting the at least one unknown component.

3. 2. The method of claim 1, wherein the separation unit includes a plurality of sub-units employing different separation techniques, the separation unit having at least a first sub-unit and a second sub-unit, and one of the first sub-unit and the second sub-unit is selected for physical separation of the selected at least one unknown component based on the one or more features of the unknown components extracted from the image of the material stream.

4. 4. The method of claim 3, wherein the first sub-unit is used for the physical separation of smaller and / or lighter components in the material stream and the second sub-unit is used for the physical separation of larger and / or heavier components in the material stream.

5. 4. The method of claim 3, wherein the first subunit is configured to separate the components by directing a fluid jet towards the components to blow the components into the accessible separate compartment, and the second subunit is configured to separate the components using a mechanical manipulation device.

6. The method of claim 1 , further comprising calculating data indicative of a mass for each of the plurality of unknown components in the material stream.

7. The method of claim 5 , wherein the force generated by the fluid jet is adjusted based on the mass of the selected at least one unknown component.

8. 2. The method of claim 1, wherein a value indicative of a difficulty of physically separating the unknown component from the material stream by the separation unit is determined, the difficulty value is associated with each of the plurality of unknown components, and the step of selecting at least one unknown component from the plurality of unknown components in the material stream is performed based on the difficulty value.

9. 9. The method of claim 8, wherein a plurality of top unknown components are selected from the plurality of unknown components in the material stream based on the training rewards associated with the plurality of unknown components, and a subset of the plurality of top unknown components is selected for physical separation based on a value indicative of the difficulty of performing physical separation by the separation unit.

10. the accessible separate compartment allows for manual removal of the isolated unknown component; An indication of an internal reference of the machine learning model is provided for the isolated unknown component located within the accessible separate partition; The method of claim 1 , wherein the analysis of the at least one selected unknown component is performed at least in part by human annotation.

11. The method of claim 1 , wherein the isolated unknown component is analyzed in an analysis unit, the analysis unit being arranged to automatically perform the characterization of the isolated unknown component in the accessible separate partition to determine the ground truth label based on a characterization.

12. The method of claim 11 , wherein the analysis unit is configured to perform destructive measurements of the isolated components to determine the ground truth labels based at least in part on the destructive measurements of the isolated components.

13. 12. The method of claim 11, wherein the analysis unit is configured to perform at least one of energy or wavelength dispersive X-ray fluorescence spectroscopy, assaying, inductively coupled plasma optical emission spectroscopy, inductively coupled plasma atomic emission spectroscopy, inductively coupled plasma mass spectrometry, laser induced breakdown spectroscopy, infrared spectroscopy, hyperspectral spectroscopy, X-ray diffraction analysis, scanning electron microscopy, nuclear magnetic resonance, and Raman spectroscopy.

14. The method of claim 1 , wherein the one or more characteristics relate to at least one of volume, size, diameter, shape, texture, color, and eccentricity.

15. 1. A system for training a machine learning model configured to characterize components in a heterogeneous material stream containing a plurality of unknown components, the system comprising: The system includes a processor, a computer-readable storage medium, a sensor system, and a separate unit, the computer-readable storage medium storing instructions: The instructions, when executed by the processor, cause the processor to: operating the sensor system to scan the material stream to image the material stream including the plurality of unknown components; predicting, for each of the plurality of unknown components in the material stream, one or more labels and associated label probabilities by a machine learning model configured to receive as input the images of the material stream and / or one or more features of the plurality of unknown components extracted from the images of the material stream; determining a training reward associated with each unknown component in the plurality of unknown components in the material stream; selecting at least one unknown component from the plurality of unknown components in the material stream based at least in part on the training reward associated with the plurality of unknown components, the selecting step requiring analysis in physical separation to determine a ground truth for the at least one unknown component; operating the separation unit to physically separate the selected at least one unknown component from the material stream, the separation unit moving the selected unknown component to another accessible compartment; receiving a ground truth label determined for the at least one isolated unknown component by performing the analysis, and adding the ground truth label determined for the at least one isolated unknown component to a training database; training an incremental version of the machine learning model using the determined ground truth label of the at least one physically isolated unknown component; The system is configured to perform a chemical analysis of the at least one unknown component separated from the material stream to determine the ground truth label based at least in part on a chemical analysis.

Images