WO2025125332A1

WO2025125332A1 - Supervising a particle quality

Info

Publication number: WO2025125332A1
Application number: PCT/EP2024/085691
Authority: WO
Inventors: Markus RAUTENSTRAUCH; Tobias Reichel; Benjamin CHIKHI
Original assignee: BASF SE
Current assignee: BASF SE
Priority date: 2023-12-15
Filing date: 2024-12-11
Publication date: 2025-06-19
Anticipated expiration: 2026-06-15
Also published as: TW202532829A

Abstract

A method (M) for supervising at least one particle quality of particles (102) produced by a particle production process at an industrial plant (100), the method (M) including: providing (S26) a package identifier indicative of a particle package (110) of particles (102) produced by the production process; generating (S12) a particle quality classification from particles (102) of the particle package (110); generating (S36) and outputting (S40) a classification data set including at least the package identifier and the particle quality classification of the indicated particle package (110); and generating (S38) and outputting (S42) a classification rating based on the particle quality classification and a classification rating criterion. A control device (118) for supervising at least one particle quality of particles (102) produced by a particle production process at an industrial plant (100).

Description

Supervising a particle quality

This disclosure relates to a method for supervising at least one particle quality of particles produced by a particle production process at an industrial plant. This disclosure also relates to a computer program product, and to a control device for supervising at least one particle quality of particles produced by a particle production process at an industrial plant.

Document WO 2022/058414 A1 shows a method for digitally tracking a chemical product manufactured at an industrial plant having an equipment, the product being manufactured by processing an input material using a production process via the equipment. For example, process data received from the equipment and indicative of a process parameter and/or an equipment operating condition may be appended to an object identifier indicative of a property of the input material. This way, a so-called digital twin of the chemical can be generated, which can provide a precise description of a technical and/or chemical property of a specific and/or identifiably product quantity in a trackable and/or exchangeable manner.

Document US 2023 / 350 395 A1 relates to a method for correlating manufacturing parameters to at least some properties of a product for ensuring product quality or product stability. An input material being processed by a processing equipment is divided into package objects, which are real-world packages. Subsequent processing of such package objects is managed by corresponding object identifiers, which are data objects. A ML model is trained based on data from one or more historical upstream object identifiers and current performance parameters measured from recent samples of the chemical product. For example, quality data from image analysis is used, and the analysis results done on the sample are appended at the sampling object identifier. The ML model and/or upstream object identifier can be used for correlating one or more performance parameters of the chemical product to the specifics of the production process.

Document WO 2022 / 103 337 A1 relates to calculating material packaging porosity and using machine learning to classify particulate material. The disclosed method has: inputting images of multiple particles, e.g. of a rice pile, in a first step, identifying individual particles using segmentation in another step, extracting surface features (color, texture) in another step, recognizing a particle classification by inputting particles into a convolutional neural network in another step, and respectively calculating an index indicating the quality of the particulate material in another step. It is an object of the present disclosure to increase a utility of the supervision of a particle production.

According to one aspect of this disclosure, a method for supervising at least one particle quality of particles produced by a particle production process at an industrial plant is suggested. The suggested method includes: Providing a package identifier indicative of a particle package of particles produced by the particle production process; generating a particle quality classification from particles of the particle package; generating and outputting a classification data set including at least the package identifier and the particle quality classification of the indicated particle package; and generating and outputting a classification rating based on the particle quality classification and at least one classification rating criterion.

That is, the suggested method results in both a classification data set and a classification rating. The output classification data set describes one or more properties of the particle package by means of at least one class information. The classification data set thus may be used during further processing steps of the particles of the particle package, for example as a digital twin or within a digital twin of the particle package. The output classification rating on the other hand describes a rating (in other words: an assessment) of the classification, i.e. , it provides context of the classification. The classification rating thus may be used during further processing steps of the particles of the particle package, too, for example as a digital twin or within a digital twin of the particle package. In addition thereto, the classification data set and the classification rating may also be used as feedback on the particle production process.

For illustration purposes, a production of an example product, such as shoe sole, is considered to have a first process of making many single foamed particles from input particles and then a second process of making one shoe sole from many foamed particles. If the suggested method would be applied to the first process of this simplified and exemplary production, then for example the classification data set may be used for optimizing parameters of the second process (i.e. making a better sole from many well-known particles) while the classification rating may be used for optimizing parameters of the first process (i.e. making better particles by using the an- notation/context information of the classification rating).

That is, the rating of the classification based on the at least one classification rating criterion increases the available information, and thus it increases the utility of the supervision of a particle production. In other words, the method's output is usable in both a forward direction and a backward direction of a production process. A particle may mean any type or kind of particle, granule and/or powder. One may understand a single particle as a distinct structural entity. According to a preferred option, the particles may preferably be chosen from at least one particle category from a group consisting of these particle categories: synthetic particles, polymer particles, compact particles, expanded and/or foamed particles, coated particles, particles measuring at least 0.5 mm in diameter, and particles measuring up to 15 mm in diameter.

A particle quality may mean any measurable and/or quantifiable property of a particle. A particle quality may be chosen from a group of properties consisting of a chemical property, a physical property, an optical property, an electrical property, a magnetic property, and a mechanical property.

A particle production process may mean any process step or set of process steps during preparing and/or manufacturing a particle including, for example, a process selected from the following list consisting of: a particle forming process, a particle coating process, a particle foaming process, a particle finishing process, and a particle inspection process.

Providing the package identifier may mean that an identifiable and/or discriminable element, such as a title, a code, a number, an alphanumeric sequence, a binary sequence, and/or the like, is provided. The package identifier may be unique. Preferably, each package identifier unambiguously identifies one particle package. A package identifier may be understood as being indicative for the particles of the particle package identified by this package identifier.

A particle package may mean any quantity of one or more particles. Typically, a particle package may be dimensioned to fit into a container, such as a so-called octabin, for example. A particle package may be a physically limited unit, such as particles within a container of sorts. For example, the method may be applied at a quality check for incoming goods at an industrial plant. A particle package may be a virtually separated and physically separatable unit, such as particles on/within a conveyor, like atop of a belt conveyor. For example, the method may be applied to a stream of particles output from a particle production device, such as a particle coating apparatus or the like. A particle package may be defined by its size, for example 10 kg or 50 kg or 0.25 m³ or 100 liters or 500.000 particles. A particle package may be originally defined by a production period of the particle production process and/or performed by a recent particle production device.

Generating a particle classification may mean to classify the particles of the particle package.

This step may have: classifying all particles of the particle package. This step may have: classi- fying a part-quantity of particles from all of the particles of the particle package. The small quantity preferably is a quantity prepared to represent all particles of the particle package. The classifying or classification preferably is based on the particle quality of the particles.

Outputting the classification data set may for example mean: providing the classification data set for retrieval by a client. Outputting the classification data set may for example mean: broadcasting the classification data to a channel. Outputting the classification data set may for example mean: recording the classification data set into a record (such as a structured record like a file and/or data base). Outputting the classification data set may for example mean: storing the classification data set into a storage (such as a data base, a data cloud and/or a hard drive). Outputting the classification data set may for example mean: sending the classification data set to one or more recipients. The same meanings apply to outputting the classification rating.

The classification data set is preferably configured to interrelate and/or connect the package identifier with the generated classification of the particles of the package identified by this package identifier. For example, the step of generating the classification data set may be or include appending and/or connecting a classification of the particles of the particle package with I to the package identifier identifying this particle package. The classification data set is preferably configured and/or usable for indicating one or more quality classes of each one or more particle qualities of the particles of the identified particle package.

A classification of particles may be generated by performing for each of multiple particles: detecting a quality amount of a quantifiable particle quality of the respective particle and selecting a particle quality class corresponding to this quality mount from a particle quality classification scheme (e.g., a table), which assigns different particle quality classes to different amounts of this particle quality. The classification of particles may be generated by performing for multiple particles: detecting a quality amount of a quantifiable particle quality of each of the multiple particles and determining a statistically descriptive amount / value describing the multiple quality amounts together.

Generating a classification rating may mean to rate the generated classification. This may include to provide a rating scheme and to select a rating value according to the generated classification from the rating scheme.

Generating a classification rating may mean that a classifier device, which performs the step of generating a classification, outputs or returns at the same time for the same particle/-s a class indicator and a confidence indicator. For example, when a classification is performed by an im- age pattern recognition model, like a trained image pattern recognition model, the model may be configured and/or trained to evaluate and output its pattern recognition confidence together with its pattern recognition result.

Preferably, the classification rating is output in combination with the package identifier and/or a time stamp. These are two exemplary ways for enabling an operator of a production device of the classified particles to relate the rating to the according production (and thus production parameters).

The output classification data set and the output classification rating may be separate from each other, like separate data sets. According to an option, the classification rating may be included in the classification data set. There may be output two data set, which are the classification data set and a rating data set including at least the package identifier, the particle quality classification of the indicated particle package, and the classification rating of the particle quality classification of the indicated particle package. If the classification rating is included in a data set, this data set may include the classification rating criterion, a description of the classification rating criterion, and/or a classification rating criterion identifier indicative of the classification rating criterion. That is, the classification rating can be output in one or more of many forms. The form of the output classification rating may for example depend on a role of a recipient of the output classification rating (e.g., a customer may have different needs, such as a degree of precision, than an operator of the production process).

According to an option, the particle quality classification may include a low confidence class indicative of a quantity of particles, wherein a confidence of a quality classification of these particles has previously been below a confidence threshold. The classification rating criterion may include a low confidence threshold. The generated classification rating may indicate whether the quantity of particles, which is indicated by the low confidence class, exceeds the low confidence threshold. For example when a particle is not classified with high certainty, for example because its target shape is not well-known or because a quantifiable deviation of its actual shape from its target shape is not characteristic for a pre-defined shape feature class, this particle may be assigned to a probable class while a certainty coefficient of this particle is set to a low value. This option has the benefit that an operator or the like is warned before relying on a classification, which is not reliable enough.

In the above option, there are two appearances of "confidence". First, there is quantity of particles of which particles the confidence of the quality classification has previously been below the confidence threshold. Here, the threshold relates to the confidence level. This confidence level is or corresponds to a probability value of a correct classification of a certain particle to a certain particle quality class. By means of this confidence threshold, it can be defined when a classification of a particle (such as a classification of an image of the particle) will be assessed "not reliably identified". For example, this confidence threshold may be set to 90%, such that a classifying agent (e.g. a pattern recognition model) must be at least 90% sure (e.g. it must assess a reliability of the classification to be at least 90% reliable), that a given particle is of a specific particle quality class, in order to classify said particle into said specific particle quality class. For example, if a pattern recognition model is 90% sure, that an image shows a particle having a tail, then the pattern recognition model classifies this particle into a class "particle with tail", else it classifies this same particle into the low confidence class instead.

The second appearance of “confidence” relates to the low confidence threshold. This low confidence threshold defines a maximum quantity of particles per particle package, e.g. a maximum percentage of particles, to be classified into the low confidence class. For example, if the percentage of particle classified into the low confidence class exceeds this maximum percentage, then the classification rating indicates that too many particles have not been reliably classified. If this low confidence threshold is exceeded, an operator or the method itself may for example decide on a need for a further / deeper training of the classification model.

According to an option, the particle quality classification may be generated by providing a particle information to a trained model, which particle information is characteristic for a/the particle quality. A trained model, such as a trained pattern recognition model, enables an automation of the classification step, which in turns enables a high throughput of particles.

In embodiments, the steps of generating and outputting the classification data set and/or the classification rating is executed by a trained model, e.g. a classifier in terms of an artificial neural network or any other computer-implemented algorithm capable of dividing the received data into predefined particle quality classes. The trained model may be implemented to generate and output the classification rating.

Preferably, the suggested method further includes: initiating a further training of the model based on the classification rating, preferably if the classification rating is outside a predetermined rating range. For example, when too many particles are classed into a low confidence class, then the pattern recognition may be considered as being not reliable. Thus, the trained model may need a further training. A further training may mean a complete, new training with additional training data. However, the further training may preferably mean an additional training of the already trained model such that the trained model improves. According to an option, said initiating of the further training includes at least one of: providing a particle quality classification generated from a particle package by the trained model; providing a reference classification of the same particle package; and generating an accordance value indicative of an accordance of the provided particle quality classification with the provided reference classification. Thus, by providing the reference classification from the same particle package as a model-generated classification, the model can be evaluated for classification precision. Then, a previous or a later particle classification of particles of an unknown particle quality can be assessed based on the classification of the reference particles of the known particle quality. The accordance between the provided reference classification with the particle quality classification generated from the same package may serve as a measure for assessing a classification reliability. Preferably, the particle quality classification and the reference classification implement the same / one common classification scheme / set of classes. One may understand the accordance value as value indicating a fit of the particle quality classification to the reference classification. The accordance value may inform an user / an operator where the trained model differs from a reference.

Preferably, a reference package is provided, wherein the reference classification is a classification of the particles of the reference package. Thus, the trained model generates this particle quality classification from the provided reference package. Preferably, the reference classification is generated externally, such as by one or more human experts.

According to an option, the particle quality classification and the reference classification each indicate for multiple particle quality classes each a quantity of particles having a particle quality associated to this particle quality class. In other words, the particle quality classification is preferably not (necessarily) a single particle package quality class for all particles of the particle package, but (in most cases) it indicates the quantity of particles of the particle package per particle quality class. Thus, a recipient of the classification data set (such as an operator of a previous / subsequent production device) gains deeper knowledge on the particle quality, such that production parameters can be optimized.

According to an option, the accordance value is a confidence matrix, which indicates for each quality class of the particle quality classification and of the reference classification a quantity of particles correctly classed into this quality class and/or a quantity of particles wrongly classed into each other of the quality classes. That is, an information triple is suggested, which indicates how many particles are there of each class, and how reliable is each class identified. One may understand a confidence matrix as a means to compare on a class-to-class basis an accord- ance between a particle quality classification generated by this same rained particle classification model and a reference classification.

In embodiments, training and/or further training includes annotating training data for generating reference classification data, wherein a training data set for a training particle includes at least one captured image of the training particle and an assigned particle quality classification, i.e. an identifier for a quality class into which the training particle, and/or the captured image is classified.

The suggested method may include: providing the generated classification rating and the generated classification data set, preferably when the classification rating is outside a/the predetermined rating range; wherein the generated classification data set includes the particle quality classification, on which particle quality classification said classification rating is based. According to an option, the suggested method may include: providing the generated classification rating and the generated classification data set, based on the classification rating, preferably when the classification rating is outside a/the predetermined rating range; wherein the generated classification data set includes the particle quality classification, on which particle quality classification said classification rating is based. Thus, an easy inspection of the data is enabled when the classification is rated outside the given rage. For example, the predetermined rating range may be set to "a low confidence class below 2% of the particles"; if e.g. 3% of the particles are classified with a low classification confidence, then the classification rating and the underlying particle quality classification are output I provided in a way, that is configured for an inspection.

Preferably, the particles may be chosen from at least one particle category from a group consisting of these particle categories: synthetic particles, polymer particles, compact particles, expanded and/or foamed particles, coated particles, particles measuring at least 0.5 mm in diameter, and particles measuring up to 15 mm in diameter. Preferably, a particle package may mean any quantity of one or more particles. Preferably, a particle identifier indicative of a classified particle and a package identifier indicative of a particle package of the classified particle may be included in the classification data set. In other words, the classification data set may include a particle identifier indicative of a / each classified particle from the particle package indicated by the package identifier.

According to an option, the particle quality classification may be generated by performing an image-based pattern recognition based on captured images of the particles of the particle package. This may be referred to as image-based classification. Image-based classification is versa- tile, as it is applicable to different particle types, and it is easily trainable and controllable, as a human can visually supervise the classification.

In embodiments, the method includes capturing an image of a respective particle for generating an image data set.

Preferably, the method may include: creating and storing an image data set from each captured image, which image data set includes the image and a particle identifier indicative of the particle classified based on this image. That is, particle images are archived. Thus, a particle quality can be assessed and classified from ex post. The image data set preferably includes the package identifier indicative of the particle package of this particle indicated by the particle identifier. Thus, a quality classification of a package can be made even when the particle package is no longer available (e.g., sold to a customer).

The classification data set preferably includes the particle identifier indicative of the I each particle from the particle package indicated by the package identifier. Thus, a re-classification is easily possible, e.g., when an improved classification model is available.

Preferably, an image / each image, on the basis of which the pattern recognition is performed, is captured while the imaged particle is falling freely through a field of view of a camera device. In other words, said step of generating the particle quality classification may include: repeatedly capturing an image of a respective particle of the particle package, which particle is falling freely through a field of view of a camera device during said capturing, and generating the particle quality classification from the images of the particles of the particle package. This option provides singularized particles, such that each particle can be separately captured. Further optional, the pattern recognition may be based on a sequence of captured images of each of the particles. In other words, while falling freely through said field of view, several images may be captured per particle, which several images thus form the sequence of images. Thus, it is highly likely that each particle is shown from different angles. This improves a shape feature detection, for example. Further optionally, each image of this sequence may be linked to the particle identifier. Thus, the images showing a same particle are easily identifiable, e.g., from a database.

According to another preferred option, an image / each image is captured while the imaged particle is floating in a suspension, on the basis of which image/images the pattern recognition is performed. According to still another option, an image / each image is captured while the imaged particle is arranged on a vibrating or vibrateable plate, on the basis of which image/images the pattern recognition is performed. That is, there are three preferred way of singularizing the particles before capturing images thereof, which are making the particles falling freely through a field of view of a camera device, floating the particles within a field of view of the camera device, and/or exposing the particles to a vibrating excitation. Which way of singularization is selected, may depend on properties of the particle, such as a density, a hardness and/or a chemical stability, and/or on the on the particle quality to be captured.

According to a further aspect, a computer-readable medium storing computer program instructions, wherein the computer program instructions, when executed by a computer or computerized device, such as a control unit, cause this computer / computerized device to perform operations comprising the suggested method for supervising at least one particle quality of particles produced by a chemical particle production process at an industrial plant. The computer- readable medium is, in particular, a non-transitory computer-readable medium.

According to another aspect of this invention, a computer program product is suggested. The suggested computer program product comprises a program code for executing the suggested method for supervising at least one particle quality of particles produced by a chemical particle production process at an industrial plant by a computerized device when run on at least one computerized device. A computer program product, such as a computer program means, may be embodied as a memory card, USB stick, CD-ROM, DVD or as a file which may be downloaded from a server in a network. For example, such a file may be provided by transferring the file comprising the computer program product from a wireless communication network.

According to another aspect of the invention, a control device is suggested, that is configured for supervising at least one particle quality of particles produced by a particle production process at an industrial plant. The suggested control device is configured for performing the suggested method for supervising at least one particle quality of particles produced by a particle production process at an industrial plant. The suggested control device thus realizes the advantages and features of the suggested method.

In embodiments, the method includes amending, adapting, tuning and/or changing the particle production process as a function of the classification rating and/or the particle quality classification. Supervising the particle quality may thus facilitate monitoring and/or controlling the particle production process.

According to another aspect of the invention, a device is suggested, that is configured for supervising at least one particle quality of particles produced by a particle production process at an industrial plant. The suggested device includes: a particle supplying means configured for supplying particles from a particle package; and the suggested control device. The suggested device thus has the advantages and features of the suggested method.

Further possible implementations or alternative solutions of the invention also encompass combinations - that are not explicitly mentioned herein - of features described above or below in regard to the embodiments. The person skilled in the art may also add individual or isolated aspects and features to the most basic form of the invention.

Further embodiments, features and advantages of the present invention will become apparent from the subsequent description and dependent claims, taken in conjunction with the accompanying drawings, in which:

Fig. 1 schematically shows a particle production system at an industrial plant, the industrial plant being not further illustrated, the particle production system having a particle production device and a device for supervising at least one particle quality according to an embodiment of the invention;

Fig. 2 schematically shows particles having different shapes illustrating different visually detectable shape feature types;

Fig. 3 schematically shows a process flow diagram illustrating a method for supervising at least one particle quality according to an embodiment of the invention;

Fig. 4 schematically shows an exemplary time series of a classification rating output from the method for supervising at least one particle quality illustrated in Fig. 3.; and

Fig. 5 schematically shows an exemplary confidence matrix, which may be generated by performing the method for supervising at least one particle quality illustrated in Fig. 3.

Next will be described a particle production system at an industrial plant 100 for producing particles 102. The industrial plant 100 is illustrated in Fig. 1. The particles 102 are foamed particles 102 for example, but the invention is not restricted thereto. Each particle 102 has a particle shape 104, which may be considered a normal shape feature (a.k.a. target shape) or a shape feature deviation. Particle shapes 104 are exemplarily shown in Fig. 2.

It shall be noted however that particle shape is merely an example for a particle quality and that shape features are merely an example for detectable particle features, which are chosen for this description as they are very visual. However, a suggested method M for supervising at least one particle quality of particles 102 produced by a particle production process at an industrial plant is not restricted to supervising particle shapes.

The particles 102 may preferably be selected from these materials: polystyrene or PS, expandable polystyrene or EPS, polyethylene or PE, polypropylene or PP, expanded polypropylene or EPP, polyurethane or PU, expanded polyurethane or EPU, thermoplastic polyurethane or TPU, expanded thermoplastic polyurethane or E-TPU, polylactic acid or PLA, and/or polybutylene succinate or PBS. Preferred materials are available under names, like trademarks, such as "Styropor", "Neopor", "S-Neopor", "X-Neopor", "Neopolen", "Neopolen-P", "Neopolen-E", "Infin- ergy", "ecovio", "ecovio EA", "Styrodur", "Kerdyn", and/or "E-por".

In the order of a flow or stream of the particles 102, the industrial plant 100 has a particle production device 106. The particle production device 106 performs and/or represents the supervised particle production process.

Next is a particle conveyance device 108, which conveys a particle package 110 of the particles 102 produced by the particle production process of the particle production device 106. Then, the particles 102 having the particle shapes 104 pass through a field of view 112 of a camera device 114. Then, the particles 102 are conveyed by means of another conveyance device 108 in the particle package 110 to a container 116. The camera device 114 is connected to a control device 118. The control device 118 is configured for performing said method M for supervising the at least one particle quality of particles 102 produced by the particle production process at the industrial plant 100.

The particle production device 106 performs any particle production process. For example, the particle production device 106 may be supplied with granules and may be configured to foam those granules into foamed particles 102. The particles 102 from the particle production device 106 may for example be provided in a particle stream, which is separated on a time-basis into several particle packages 110. The particles 102 may for example also be provided by the particle production device 106 in physically separated particle packages 110, e.g., by gaps between the particle packages 110.

In the production of particles 102, such as compact granules and particle foam, it may be important to ensure a consistent particle quality of the particles 102 and/or to minimize the number of particles 102 having shape feature deviations. The shape feature deviations (product deviations) may for example affect an external, morphological appearance of individual, foamed par- tides 102. The shape feature deviating partides 102 may otherwise be part of one or more batches delivered to a customer as a mixed form and may be objected. Also, a quality of the produced particles 102 should not depend on the production site or the operator but should be consistent.

There are commercially available particle analyzer systems such as PartAn3D or CAMSIZER from Microtrac Retsch GmbH, by means of which a geometric dimension (width, length, thickness) can be recorded, for example. Such particle analyzer systems do not provide information about shape feature types or statistics about different shape feature types or their frequency. When such a shape feature type analysis should be done by trained human analyzer personnel, samples cannot be taken continuously but only periodically from an industrial scale particle stream. Such manual shape feature -finding would thus be ineffective: it requires manpower, is slow and leads to a high raw material consumption because it takes a while to find out that the product quality is not within the specifications.

By providing an online classification and quantification of granules and/or particles that deviate externally from a given standard particle or particle standard during a preparation/production process, optical quality data may directly be allocated to the corresponding process data. A particle analyzer system as described above may be used as a multidimensional measuring point that delivers one or more pictures of granules/particles and/or their geometric dimension. The pictures of the particles 102, that are obtained from the particle analyzer system, are classified into a particle quality classification by an algorithm, and a classification rating based on the classification is generated. The particle quality classification may have one or more previously defined shape feature classes. An example of the classification rating may be statistics on the frequency of occurrence of a shape feature type in a sample.

The classification is provided in the form of a classification data set including the particle quality classification and a package indicator, which identifies the particle package whose particles 102 are classified.

The classification rating may mean a rating of the particles 102 of the particle package 110, which rating relates the particle quality classification to a provided classification rating criterion. As it is based on the classification rating criterion, the classification rating is adaptable by providing an adapted classification rating criterion, for example when different customers provide individually adapted rating criteria. For example, a shape feature information and an outer geometrical data of the gran- ules/particles 102 may be provided to the control device 118 and/or to a control station of the production device 106. Each shape feature type may be given a predefined upper limit of occurrence; exceeding the upper limit triggers an alarm signal that enables the controller to change process parameters accordingly while the process is running. This allows for a fast response if the product quality starts to drift out of the specifications. In this way, a total number of shape feature deviating particles 102 in a batches can be reduced, which can increase customer satisfaction. Additionally, a fast adjustment of one or more production conditions can reduce raw material consumption.

The control device 118 generates and outputs a classification data set, which includes at least the package identifier and the one or more generated particle quality classification of the particles 102 of the particle package 110 indicated by this same package identifier. The particle production process implemented by industrial plant 100 can be included in a downstream process for manufacturing an article as discloses in WO 2022/058414 A1. The package identifier can be generated and managed according to the methods with respect to the object identifier as disclosed in WO 2022/058414 A1 which is hereby incorporated by reference..

The shape feature information such as shape feature statistics may thus be linked to a virtual process package that relates to the granules/particles in question. Such a virtual process package may mean a digital twin of a particle package 110 of material with a certain size, that is followed through the production process. The virtual process package contains information about the raw material as well as process parameters of each processing step the package goes through. By linking the shape feature information to the virtual process package, shape feature deviations and product quality can directly be connected to process conditions. Additionally, if several process packages of a certain size (for example 10 or 50 kg) are combined to bulk boxes, their quality level can be determined because the information of each package within the bulk box is known.

Applying this technique to an industrial scale particle production process may produce at least one of the following advantages: A total and/or relative number of shape feature deviations may be reduced. A particle quality may be improved. A customer satisfaction may be improved. A continuity of a particle quality may be increased. A particle quality may be quantified, especially between different production lines and/or production sites. A particle quality may be harmonized between different production lines and/or even between different production sites. An amount of information and/or knowledge about the particle production process (such as: what kinds of shape feature appear, how many shape feature deviations appear per shape feature type [may also be referred to as a shape feature kind or shape feature class], and/or particle size distribution) may be generated and/or increased, for example through connecting the shape feature information (particle quality classification) and process information. A process control of the particle production process may be improved. An amount of raw material may be reduced. A particle quality may be tailored and/or selected to specific needs of an individual customer. The classification rating may be provided to a succeeding production process, thus enabling an operator and/or control device thereof to adapt a parameter of this succeeding process in order to improve a quality of an end product.

Next, there will be highlighted some aspects of the suggested method and/or device for supervising at least one particle quality of particles produced by a particle production process at an industrial plant.

According to one preferred option, only a sample may be taken and/or separated from the main product stream (e.g., from the particle package 110 on the particle conveyance device 108) and made available to the field of view 112 of the camera 114, that is to the optical particle analyzer device 112, 114, 118. This option has the effect that a total output per time from the particle production device 106 can permanently be significantly larger than a maximum throughput of the optical particle analyzer device 112, 114, 118. However, according to another option, all particles 102 may be scanned by the camera 114 to guarantee very strict particle quality requirements.

In this embodiment, the detection of the particles 102 takes place in free fall. During free fall, multiple images of a particle may be taken while the particles 102 are in a natural rotation. For example, 5 to 8 images per particle 102 may be possible. That is, the free fall has the function of singularizing the particles 102 and rotating the particles 102 before the camera 114. However, the singularizing function or both functions may alternatively be provided by arranging the particles 102 atop of a vibratable plate / a vibration device and/or by arranging the particles 102 within a fluid, especially a liquid.

The camera device 114 generates images of the particles 102 in the form of image files.

Here is described an optional way of image saving: The image files of the multiple images for a respective particle 102 may comprise in a file name thereof a sample number indicative of the sample from the particle package, on which sample the image is based I from which it is captured. The sample number is configured to act as a unique identifier for linking the process package of the digital twin and sample that was analyzed with the optical particle analyzer. Each image is assigned to a GUID (General Unique Identification Number, which maps to a sample and thus to a particle package 110). The mapping remains the same for all following analyses. For example: A particle package identifier might read: "ABC-987654321-842", wherein "ABC" indicates a particle production device and/or a production site, "987654321" indicates an order number such as a process order number, and "842" indicates a unique counter within said order. In this case, a particle 102 being assigned with a particle number "1236" might be depicted on images of images files having the exemplary image file names: "ABC-987654321 - 842_1236_01.jpg", "ABC-987654321-842_1236_02.jpg", "ABC-987654321-842_1236_03.jpg", and so forth.

The measured particle geometries incl. statistical processing may be evaluated. The results may include direct geometrical information such as width, length, and thickness of the particles 102, as well as characteristics, that relate to particle geometries, such as an index, which characterizes the particle size distribution or particle volume. Such obtained measurement results (raw pictures of the samples, statistics resulting from an evaluation - such particle size distribution, length, width, thickness, geometrical data etc., and the number of particles of each sample) may be assigned to a corresponding digital twin process package. A storing of such data may for example take place in a graph database, which may mean an object-oriented approach / database structure. Preferably, all characteristics are assigned/linked to a sample. The optical particle analyzer 112, 114, 118 thus may act as a multidimensional measuring point.

The image files archived by the optical particle analyzer 112, 114, 118 may be analyzed by an algorithm from the artificial intelligence type(s). The algorithm is preferably performed by the control device 118. The algorithm preferably follows a structure of a neural network and is preferably configured to assign the images of the particles 102 to previously defined image classes. The main classes are for example: a class for particles 102 having a normal shape feature, several defined shape feature deviation classes (see figure below), and a low-confidence class for low-confidence images, which is a class for image files which are not assignable because the algorithm cannot classify them with sufficient accuracy.

Said assignment to the different classes may preferably be based on a previously trained model (sorting process). The model preferably is a trained pattern recognition model. For example, statistics about a frequency may additionally be generated, with which frequency a shape feature type occurs in a sample (in other words: a percentage of particles 102 having a specific shape feature type among all particles 102 captured by the camera 114). In addition to shape feature classification, for example, a particle weight can be calculated based on particle volume information from the camera 114 / control unit 118. Any obtained shape feature and particle information, such as classification results of the particle packages, level of confidence (i.e. how high does the algorithm rate the reliability of the classification), number of particles and images of particles of said particle package, and/or particle images saved in each result class, may be assigned to the respective classification data set by means of the package identifier indicative of the particle package / sample. These statistics can thus also be linked to the respective particle package. Like in the previous step, a storing of the generated classification data sets preferably takes place in a graph database.

Fig. 2 displays several particle shape patterns P1 to P5, each belonging to a different particle 102 having a different particle shape 104. The patterns P1 to P5 all relate to particles 102 produced by a production process for producing foamed particles 102, for sake of a preferred example. The pattern P1 shows a normal foamed particle 102 in a target particle shape 104. Patterns P2 to P5 show particle deviations of an elliptic kind of foamed particle 102. The pattern P2 shows two or more agglomerated foamed particles 102. The pattern P3 shows a particle 102 having a small tail. The pattern P4 shows a wrinkled foamed particle 102. The pattern P5 shows an edgy foamed particle 102. The above listing is not complete but merely gives a few examples for foamed particle shape features. The patterns like the particle shape patterns P1 to P5 may preferably be used for training a pattern recognition model for automatically generating a particle classification.

All data (such as particle images, particle information, package identifiers, classification data sets, and/or classification rating) may preferably be saved on a local raw data storage. Additionally, all or part of the data may also be transferred into a cloud storage. For example, the classification data sets may be made available in the cloud storage for further analysis. Additionally, an interface for the transfer of raw images may be provided.

Thus, a direct assignment of the optical quality data with the corresponding process data is now possible.

Preferably, the shape feature and particle information, such as the package identifier and the particle quality classification, may be provided to a control room of the production device 106. An alarm may for example be triggered in case a predefined upper limit of occurrence of one or more shape feature types is exceeded. Corresponding countermeasures can then be taken, for example by an operator or according to pre-defined operation patterns or by a trained operation model. The alarm may be connected to a concrete suggestion of a countermeasure. For exam- pie, an exceeding of a threshold value for too small beads may directly trigger a prompt to flush one or more die plates.

According to a preferred aspect, the classification model, such as the trained pattern recognition model trained for generating the particle quality classification, may be continuously optimized by an ongoing training thereof. Those particle images, that could not be assigned to any of the shape feature classes or to the shape feature deviation-free particle class, are preferably assigned to a low confidence class. The number of low-confidence images is monitored, and an alarm or the like may be triggered in case this number exceeds a predefined threshold value. The threshold value can be freely chosen; it can be 5 %, for example. The threshold value and the alarm, which is triggered when the value is exceeded, may be configured to signalize that a further training of the classification model may deemed necessary. The images in the low- confidence class are then classified manually, for example by using a software configured for generating training data. Those newly classified images are uploaded into the classification model as new training data and the model is updated.

Next, an example process of the method M for supervising at least one particle quality of particles produced by a particle production process at an industrial plant is described with reference to its flow chart in Fig. 3.

In a step S10, the particles 102 of a particle package 110 produced by the production process of the particle production device 106 are supplied by the particle conveyance device 108.

In a step S12 a particle quality classification is generated. This step S12 has for example the sub-steps S14 to S22 and S34 described below.

In a step S14, the particles 102 are singularized. For example, in a step S16 the particles 102 are made to fall freely through the field of view 112 of the camera device 114. Alternatively, in a step S18, the particles 102 are made to float in a suspension. Again alternatively, in a step S20, at least one particle 102 is arranged on a vibrating or vibrateable plate device.

In a step S22, one or more images of each of the particles 102 are captured by the camera device 114. Especially if the particles 102 are falling freely through the field of view 112, a sequence of images is captured in step S22 by the camera device 114.

In a step S24, a particle identifier is provided for each particle 102 captured in an image. In a step S26, a package identifier indicative of the particle package 110 is provided.

In a step S28, an image data set from each captured image is stored. The image data set includes the captured image, the particle identifier, and the package identifier, for example.

In a step S30, a classification scheme is provided. For example, in a sub-step S32, a trained pattern recognition model is provided.

Then, in a step S34, for each particle 102, a classification is selected based on the captured image/s of the particle 102 by means of the trained pattern recognition model. The classification preferably follows a classification scheme of pre-defined particle quality classes (see the above explanation relating to Fig. 2). One may consider the classification scheme as a definition and/or a defining ruleset of the particle quality classes.

For example, each particle 102 is classed/classified in step S12 into one of seven classes. In this example, the first class corresponds to normal shaped particles, and the classes two to six correspond each to one of the patterns P1 to P5 explained above in connection with Fig. 2. The seventh class is a low-confidence class. The trained pattern recognition model, which is provided in S32, is configured such that is classifies each particle 102 into one of the classes of patterns P1 to P5 only if it (the pattern recognition model) only if a confidence of this classifying is above a classifying confidence threshold. Otherwise (i.e. when the pattern recognition model is not able to confidently class this particle into one of the classes of patterns P1 to P5) this particle 102 is classed into the low-confidence class.

In a step S36, a classification data set is generated including the particle quality classification and the package identifier. The classification data set preferably has the form of a digital classification data set.

In a step S38, a classification rating is generated on the basis of the particle quality classification generated in step S12 and a classification rating criterion. Said classification rating criterion may be provided separately.

For example, a total number of low-confidence classed particles per particle quantity and/or a percentage of low-confidence classed particles per particle quantity is compared to a low confidence threshold value. The result of this comparison, for example a classification rating flag, is then preferably stored as a separate data set or as a separate data item within the classification data set generated in S36. In a step S40, the classification data set is output. In a step S42, the classification rating is output. The classification data set and rating can be generated and output in one data structure for further processing.

According to one option, the classification data set includes only the particle quality classification and the package identifier. This is the classification data set generated in step S36. According to this option, the classification rating is not included in the classification data set. Instead, the classification rating may have the form a data set separate from the classification data set.

According to another option, there is output a single data set including the package identifier, which identifies the particle package 110, the particle quality classification, which indicates the particle quality of the particles 102 of the identified particle package 110, and the classification rating, which indicates a conformity with the classification rating criterion. Under this second opinion, steps S40, S42 are included or merged into a single step S40, S42.

Thus, a controller of the particle production device 106 may receive the particle classification data set and/or the classification rating. As each of the particle classification data set and the classification rating supervises a particle quality of the particles 102 from the particle production device 106, the controller can monitor the particle quality. Optionally, the controller may alter production parameters of the particle production device 106 to improve the particle quality in specific particle packages.

Fig. 4 is an exemplary graph showing a time series of a value v over a time t. The value v is, for example, a ratio of agglomerated foamed particles 102 (pattern P2 of Fig. 2) to all particles 102 the particle package 110. It can be seen that, within the whole depicted period from a start time t_0 until now, it is only at a time_1 , where the value v exceeds a threshold v_th. This may raise an alarm. After a very short period the ratio of agglomerated foamed particle 102 falls below the threshold value v_th. This short period may reflect a lag between a production of a particle package until an outputting of the classification data set regarding this particle package. That is, in this depicted case, on operator reacted quickly to ensure a high particle quality.

In a step S44, a further training of the trained pattern recognition model, which is provided in S32, is initiated based on the classification generated in S38. For example, there may be a rule that a further training is needed when at least 1% or even at least 0.5% of all classified particles are classified into the low confidence class. Further examples may include a customer feedback, wherein customer may refer not only to a commercially defined customer but to any instance receiving the particle package, including another production process at the same indus- trial plant. Further, any other quality check process may provide a feedback value, which can be taken into consideration for deciding on the initiating of the further training. For example, a supervision process after a production process, which uses the particle package 110, may report a quality characteristic value, which can be interrelated to the particle package by means of the package identifier.

This step S44 may have one or more of sub-steps S46 to S52 explained here.

In a sub-step S48, a particle quality classification is generated from a particle package 110 by said trained pattern recognition model. Further, in a sub-step S50, a reference classification of the same particle package 110 is provided. Then, in a sub-step S52, an accordance value indicative of an accordance of the particle quality classification provided in S48 with the reference classification provided in S50, is generated. This accordance value provides information to a user I an operator where the trained model differs from a reference. Preferably, there is generated an accordance value for each class of the classification scheme.

In a first variant, the particle package 110 of step S44 is a previously used particle package 110. Here, the reference classification may be generated by one or more experts from the image data sets of the particles 102 of this particle package 110, which were stored in S28 and may be retrieved by means of the particle identifier.

In another variant, the particle package 110 of step S44 is a reference package. This reference package is provided in a step S46, before, in said step S48, the particle quality classification is generated from the particles of the reference package. Thus, in step S48, generation of the particle quality classification is performed the same way as with the particles 102 from the particle package 110 in step S12. Further, the reference classification provided in S50 fits this reference package provided in S46. For example, the reference classification may preferably be generated manually from the particles 102 of the reference package.

For example, said accordance value may be a confidence matrix, Fig. 5 is a table or map, illustrating a confidence matrix, which may be drawn or created in steps S44 to S52. The reference particles I reference package(s) may be separated from any training particles / training packages before a training of the classification model. After the training of the classification model, the reference particles are provided to the model for being classified. That is, a predicted label may be determined. The goal is that the particle quality classification of the reference particles by the trained classification model corresponds exactly to the reference classification, which may have previously been determined by a human specialist. If all particles 102 are assigned to their cor- rect particle quality class, the confidence matrix would have the value of 1 .0 or 100% in its main diagonal and the value of 0.0 or 0% in all other fields. However, if a reference particle 102 is not correctly classified, then the values in matrix fields accordingly deviate (i.e., lower in the main diagonal and higher in other fields).

In the example of Fig. 5, only the reference particles 102 of the classes "normal" and "agglomerated" are correctly classified, whereas only two out of three reference particles 102 of the class "edgy", one out of ten reference particles 102 of the class "tail", and none reference particles 102 of the class "wrinkled" are correctly classified. Further, the generated particle quality classification of class "normal" contains false-positives of particles 102 from the classes "tail", "wrinkled", and "edgy". Thus, the confidence matrix provides a simple-to read overview or survey on the power or quality of the trained classification model.

The confidence matrix transparently shows which classes need a further training. In the example of Fig. 5, only the training of class "agglomerated" seems to be satisfyingly completed, whereas a further training of the other classes seems necessary.

It is understood that in this disclosure supervising may be interpreted as monitoring how a particle package evolves in a production process with regard to its quality and/or bookkeeping the quality of particles contained in respective particle package.

Reference signs:

100 industrial plant

102 particle

104 particle shape

106 particle production device configured to perform a particle production process

108 conveyance device

110 particle package

112 field of view of a camera device

114 camera device

116 particle container

118 control device

P1 to P5 particle shape pattern t time t_0 start time t_1 event time v value v_th threshold value

M method for supervising at least one particle quality of particles produced by a particle production process at an industrial plant

S10 Supplying particles of a particle package produced by the production process

S12 Generating a particle quality classification

S14 Singularizing the particles

S16 Making the particles fall freely through a field of view of a camera device

S18 Making the particles float in a suspension

S20 Arranging at least on particle on a vibrating or vibrateable plate device

S22 Capturing one or more images of each of the particles

S24 Providing a particle identifier for each captured particle

S26 Providing a package identifier indicative of the particle package

S28 Storing an image data set from each captured image including the image, the particle identifier, and the package identifier

S30 Providing a classification scheme

S32 Providing a trained pattern recognition model

S34 Selecting a classification based on the captured image/-s of the particle by means of the trained pattern recognition model

S36 Generating a classification data set

S38 Generating a classification rating

S40 Outputting the classification data set

S42 Outputting the classification rating

S44 Initiating, when the classification rating is outside a predetermined rating range, a further training of the model

S46 Providing a reference package of particles

S48 Providing a particle quality classification generated from a particle package by the trained model, which preferably is the reference package

S50 Providing a reference classification of the same particle package

S52 Generate an accordance value indicative of an accordance of the provided particle quality classification with the provided reference classification

Claims

1. A method (M) for supervising at least one particle quality of particles (102) produced by a particle production process at an industrial plant (100), the method (M) including: providing (S26) a package identifier indicative of a particle package (110) of particles (102) produced by the production process; generating (S12) a particle quality classification from particles (102) of the particle package (110); generating (S36) and outputting (S40) a classification data set including at least the package identifier and the particle quality classification of the indicated particle package (110); and generating (S38) and outputting (S42) a classification rating based on the particle quality classification and a classification rating criterion.

2. The method of claim 1, wherein the particle quality classification includes a low confidence class indicative of a quantity of particles (102), wherein a confidence of a quality classification of these particles (102) has previously been below a confidence threshold; wherein the classification rating criterion includes a low confidence threshold; and wherein the generated classification rating indicates whether the quantity of particles (102) indicated by the low confidence class exceeds the low confidence threshold.

3. The method of any one of claims 1 - 2, wherein the particle quality classification is generated by providing a particle information to a trained model, which particle information is characteristic for the particle quality; and wherein the method (M) further includes: initiating (S44), based on the classification rating, a further training of the model.

4. The method of claim 3, wherein said initiating of the further training includes: providing (S48) a particle quality classification generated from a particle package (110) by the trained model; providing (S50) a reference classification of the same particle package (110); and generating (S52) an accordance value indicative of an accordance of the provided (S48) particle quality classification with the provided (S50) reference classification.

5. The method of claim 4, wherein the particle quality classification and the reference classification each indicate for multiple quality classes a quantity of particles having a particle quality associated to this quality class; and wherein the accordance value is a confidence matrix, which indicates for each quality class of the particle quality classification and of the reference classification a quantity of particles correctly classed into this quality class and/or a quantity of particles wrongly classed into each other of the quality classes.

6. The method of any one of claims 1 - 5, further including: providing (S44), when the classification rating is outside the predetermined rating range, the generated classification rating and the generated classification data set; wherein the generated classification data set includes the particle quality classification, on which particle quality classification said classification rating is based.

7. The method of any one of claims 1 - 6, wherein the particles (102) are chosen from at least one particle category from a group consisting of these particle categories: synthetic particles, polymer particles, compact particles, expanded and/or foamed particles, coated particles, particles measuring at least 0.5 mm in diameter, and particles measuring up to 15 mm in diameter.

8. The method of any one of claims 1 - 7, wherein the particle quality classification is generated by performing an image-based pattern recognition based on captured images of the particles (102) of the particle package (110).

9. The method of claim 8, further including: creating and storing (S28) an image data set from each captured image including the image and a particle identifier indicative of the particle (102) classified based on this image.

10. The method of any one of claims 8 - 9, wherein an image, on the basis of which the pattern recognition is performed, is captured (S22) while the imaged particle is falling freely (SS16) through a field of view (112) of a camera device (114).

11. The method of claim 10, wherein the pattern recognition is based on a sequence of captured images of each of the particles (102).

12. The method of any one of claims 8 - 9, wherein an image, on the basis of which the pattern recognition is performed, is captured (S22) while the imaged particle is floating (S18) in a suspension.

13. The method of any one of claims 8 - 9, wherein an image, on the basis of which the pattern recognition is performed, is captured (S22) while the imaged particle is arranged (S20) on a vibrating or vibrateable plate.

14. The method of any one of claims 1 - 13, wherein said generating (S38) the classification rating includes: Providing a rating scheme, and Selecting a rating value according to the generated classification from the rating scheme.

15. The method of any one of claims 1 - 14, wherein the classification rating is included in the classification data set.

16. A computer program product comprising a program code for executing the method (M) for supervising at least one particle quality of particles (102) produced by a particle production process at an industrial plant (100) according to any one of claims 1 - 15 by a computerized device when run on at least one computerized device.

17. A control device (118) for supervising at least one particle quality of particles (102) produced by a particle production process at an industrial plant (100), the control device (118) being configured for performing the method (M) for supervising at least one particle quality of particles (102) produced by a particle production process at an industrial plant (100) according to any one of claims 1 - 15.