Method for segment-by-segment parameter assignment to a component, method for generating manufacturing data using a method of this type, and production of a component using manufacturing data of this type

Abstract

The invention relates to a method for segment-by-segment parameter assignment to a component, comprising the steps of: a) reading a three-dimensional component dataset describing a geometry of a component, which describes the component by means of three-dimensional data of a surface of the component, into a computer; b) virtually segmenting the component dataset into multiple component portions, and assigning the segmented component portions to one type of at least two different component portion types by means of a computing model formed by machine learning; wherein by means of the computing model, c) a first parameter is assigned to a first component portion of the multiple component portions. The invention is characterized in that the assigning of the first parameter in step c) is a computer-assisted determination of a first wall thickness in the region of the first component portion from data of the component data set describing the first component portion.

Description

[0001] Eisenführ Speiser

[0002] Hamburg, January 15, 2026

[0003] Our reference: BH 2757-02 WO LBI / akp / bki / rof

[0004] Applicant / Owner: BEGO Medical GmbH

[0005] Official file number: New registration

[0006] BEGO Bremer Goldschlägerei Wilh. Herbst GmbH & Co. KG

[0007] Wilhelm-Herbst-Straße 1, 28359 Bremen

[0008] Methods for assigning parameters to a component segment by segment, methods for generating manufacturing data using such a method, and manufacturing a component using such manufacturing data.

[0009] The invention relates to a method for assigning parameters to a component segment by segment. A further aspect of the invention is the generation of a manufacturing data set using segment-by-segment parameter assignment to a component and the production of a component using such a manufacturing data set.

[0010] 5. Many commonly used manufacturing processes are controlled by a manufacturing data set, which controls a production system based on three-dimensional data describing the component to be manufactured. Such manufacturing processes can be broadly divided into additive (building up) and subtractive (removing) manufacturing processes. Additive manufacturing processes operate on the principle that a component is produced point by point, line by line, or layer by layer by selectively adding and curing material. Examples include selective laser melting, 3D printing, stereolithography, and numerous other methods and variations of such additive manufacturing processes. Subtractive manufacturing processes produce a component from a prefabricated semi-finished product by selectively removing material. Examples include computer-aided milling and electrical discharge machining (EDM).Furthermore, hybrid processes are used in which an additive manufacturing process is combined with a subtractive manufacturing process, in which a component is first produced using additive manufacturing and then subsequently post-processed by subtractive machining.

[0011] A crucial step preceding the actual manufacturing process in these production methods is the creation of a data set that can be used to produce the component and thus serves to control the production plant(s). This data set, often referred to as a manufacturing data set or production data set, contains control instructions for the production plant. These instructions are derived from the three-dimensional data of the product to be manufactured, possibly taking into account corrections required due to manufacturing-related shrinkage or foreseeable warpage, and include further control parameters that define, for example, the direction of layer build-up or manufacturing accuracy, such as layer thickness or surface quality.

[0012] The planning of such a component is typically computer-aided, usually in design software known as CAD software. This gives the user considerable design freedom, as many of these data-driven manufacturing processes have few manufacturing-related design restrictions on the components to be produced. When generating such a data set, the user can often access standard elements stored in a library, which can be integrated into their design to make the computational planning effort more efficient. The result of such planning in the design software is a 3D data set that describes the entire component.

[0013] Due to the considerable design freedom, a common problem is that specific properties often need to be assigned to individual sections of the overall component, properties that must be considered during manufacturing, or that individual sections of the overall component must meet specific design requirements, such as accuracy specifications, minimum or maximum wall thicknesses, or surface qualities. If the data set is not defined in such a way that these required properties are correctly maintained and defined for each component section, the result of manufacturing the component in the data-driven manufacturing process may be of inferior quality, constitute scrap, or fail in later use.

[0014] A typical example of such a component is a dental prosthesis in the form of a so-called cast partial denture. A cast partial denture is a framework attached to the lower or upper jaw at anchor points such as existing teeth or implants. Dental components such as crowns or bridges can be attached to this framework to replace missing teeth. The cast partial denture must be formed into various components, such as clasps (retention and support elements) designed as partial clasps to anchor the denture to existing teeth, a denture base, and support surfaces that hold the dental prostheses. For some of these elements, minimum wall thicknesses must be maintained to ensure sufficient strength and load-bearing capacity.Furthermore, such or other sections of a cast metal partial denture must be manufactured with a defined accuracy and a defined surface quality.

[0015] Before manufacturing begins, the data set must therefore be assigned a corresponding parameter with regard to its data on specific component sections, for example a corresponding wall thickness in such areas or a corresponding manufacturing accuracy or surface quality.

[0016] This allocation and verification of compliance with a required wall thickness is regularly very time-consuming using the data set generated from the planning process for the entire component. It requires manually identifying the different component sections using a visualization of the component, precisely marking them, and then checking whether the corresponding target wall thicknesses are being met. Similarly, specific manufacturing parameters, such as those relating to a certain manufacturing accuracy or surface quality, must be assigned precisely to the data in the data set that pertain to the corresponding component section. If necessary, a correction and re-verification must be made to the data set to achieve or maintain a minimum wall thickness.This verification, correction and allocation process for the required different wall thicknesses or manufacturing parameters in the respective component sections is, on the one hand, time-consuming, and on the other hand, planning errors can arise due to incorrect allocations or inaccurate differentiations of the component sections, which lead to a model cast prosthesis of insufficient quality.

[0017] The invention is based on the objective of providing a method that makes the planning of manufacturing data for a data-driven manufacturing process more efficient and safer against planning errors. This objective is achieved according to the invention by a method for segment-wise parameter assignment to a component, comprising the following steps:

[0018] a. Reading a three-dimensional component data set describing the geometry of a component, which describes the component using three-dimensional data of a surface of the component, into a computer,

[0019] b. virtual segmentation of the component data set into several component sections and assignment of the segmented component sections to a type of at least two different component section types using a computational model trained by machine learning, wherein the computational model

[0020] c. a first parameter is assigned to a first component section of the multiple component sections, and

[0021] d. a second parameter is assigned to a second component section of the multiple component sections,

[0022] e. Create, from the component data set with the first and second parameters, a manufacturing data set that is designed to control the production of the component in a machine-based manufacturing process.

[0023] The method according to the invention achieves automated segmentation of a data set describing an entire component into several component sections, with the possibility of assigning parameters to individual component sections. This segmentation and parameter assignment is computer-aided. In principle, for the purposes of the invention, a computer can be a single computer; however, the invention further envisions the computer as a network in which several computers are interconnected for data communication, and individual process steps of the method according to the invention are executed on different computers that constitute this network.

[0024] In a first step (a), a component data set describing a component with respect to its geometry is read into the computer. This component data set can, for example, be a 3D data set consisting of nodes representing the geometry of the three-dimensional component and node connections defining the edges that define the neighborhood relationships between the nodes. Alternatively, the component data set can also define the 3D data as voxels or as Signed Distance Fields (SDFs). Often, raw data resulting from a CAD design process for a component must be further processed by various computational steps. In the context of the invention, such processing of the raw data set is understood to encompass the creation of the component data set that serves as input for the inventive method.In the inventive method, the component data set, which describes the entire component, is segmented virtually using a computer and divided into at least two, and usually several, different component sections, each described by a part of the component data set. These segmented component sections are each assigned to a type, with the component typically containing different types of component sections, such as, for example, a base, several retention sections, several support sections, and the like in the case of a cast partial denture. According to the invention, this segmentation is performed using a computational model developed through machine learning, i.e., a computational model with artificial intelligence.The result of this virtual segmentation using the Kl computational model is that certain nodes or each node of the component data set are assigned an affiliation to a specific component section or a type of component section.

[0025] The computational model is specifically trained for this purpose. Training the AI ​​model using supervised learning has proven particularly effective. This training is based on numerous component datasets in which a user has manually assigned component segment types to specific nodes or to each node. This type assignment forms the ground truth against which the AI ​​model's predictions are compared to train the AI ​​model and improve its accuracy. The result of this virtual segmentation using the AI ​​model is a subdivision of the component defined by the component dataset into multiple component segments, each assigned to a specific type. Multiple component segments can be assigned to the same type if component segments of the same type are present multiple times within the component.Component sections can also remain untyped, i.e., not assigned to any type, for example, if such component sections do not fall under any specific qualitative requirements and consequently do not require the parameter assignment explained below.

[0026] After the virtual segmentation of the component data set, a first and a second parameter are assigned to the first and second component sections, respectively. Assigning a parameter in this context means that a manufacturing parameter is assigned to the component section, or a property is calculated from the component section, such as wall thickness or wall thickness distribution within the component section, and this property is assigned as a parameter. Therefore, assigning a parameter can consist of either assigning a specific parameter relevant for subsequent manufacturing to the component section, such as a manufacturing accuracy or surface quality that must be maintained in this component section, or calculating a property from the geometric data of this component section, such as wall thickness.

[0027] In a subsequent step, a manufacturing data set is created from the component data set containing the first and second parameters. This manufacturing data set, which can also be created in several steps if necessary (for example, when multiple components are combined for joint production), then serves to control the data-driven manufacturing system for the additive or subtractive manufacturing of the component using the machine-based manufacturing process. In creating this manufacturing data set, the component data set is used, as well as, directly or indirectly, the first and second parameters, and potentially other parameters assigned to different component sections.This can therefore include the fact that the wall thicknesses calculated from the component data set and assigned as parameters are implemented as information inherent in the component data set when creating the manufacturing data set; furthermore, parameters explicitly assigned to a component section, which describe, for example, a manufacturing accuracy or a surface quality or manufacturing settings that cause this component section, can be incorporated into the manufacturing data set when it is created in order to effect corresponding control in the subsequent manufacturing process.

[0028] The method according to the invention enables the improved production of a manufacturing data set for data-driven manufacturing of a component in a machine-based manufacturing process such as an additive, subtractive, or hybrid manufacturing process of the type described above. Advantageously, a component data set is first virtually segmented, so that visualization and / or optimization of individual, subdivided component sections of the component data set becomes possible.The component sections differentiated during segmentation can be checked with regard to their wall thickness, modified if necessary, and compared with target wall thicknesses specified for the component section type. Alternatively, specific parameters relevant to production control, such as the geometric accuracy or surface quality of the component section, can be assigned to the segmented component sections and implemented accordingly in the manufacturing data set. This method thus enables a more efficient and less error-prone generation of manufacturing data sets for the production of components in a data-driven, automated manufacturing process.

[0029] In principle, the AI-supported computational model according to the invention can be implemented with an architecture and an algorithm as described, for example, in EP 3835983B1 or EP 391255B1. However, in contrast, the input variables consist of numerous component data sets from different components, with a user manually segmenting the components into sections and assigning a type as ground truth. The output variables are a segmentation of the component, i.e., for example, an assignment of the component's nodes to different component sections and an assignment of a type. The training method can then be executed with these data using algorithms and parameters as described, for example, in EP 3835983B1, EP 391255B1, or EP 4343694B1.Particularly preferred for the purpose of high prediction accuracy of the Kl model is the use of the weighting explained below to avoid a shift of the Kl model towards frequently occurring node types and the structure of the Kl model explained below as preferred embodiments.

[0030] According to a first preferred embodiment of the method, the assignment of the first parameter in step c) is a computer-aided determination of a first wall thickness in the area of ​​the first component section from the data describing the first component section in the component data set, and the assignment of the second parameter in step d) is a computer-aided determination of a second wall thickness in the area of ​​the second component section from the data describing the second component section in the component data set.

[0031] f. the first wall thickness is compared, using computer-aided calculation, with a first target wall thickness that is assigned to the component section type of the first component section, and

[0032] g. the second wall thickness is compared using computer-aided calculations with a second target wall thickness that is assigned to the component section type of the second component section,

[0033] h. the component data set is modified computer-aided by manual input from a user or by means of a pre-programmed algorithm if the computer-aided determined error value, which is formed from a deviation of the first wall thickness from the first target wall thickness, exceeds a predetermined permissible first error value, and

[0034] i. the component data set is modified computer-aided by manual input from a user or by means of a pre-programmed algorithm if a computer-aided determined error value, which is formed from a deviation of the second wall thickness from the second target wall thickness, exceeds a predetermined permissible second error value,

[0035] - the manufacturing data set is created from the component data set if, in each of the component sections, the error value determined by computer from the component data set, or the possibly modified component data set, does not exceed the permissible error value for the respective component section.

[0036] According to this training method, when assigning the first or second parameter, a wall thickness measurement is performed within the respective component section, and the corresponding parameter is assigned. It should be understood that this assignment can encompass a single wall thickness in a specific area of ​​the component section, multiple wall thicknesses (referring to a wall thickness profile) within a specific area of ​​the component section, or wall thicknesses across the entire component section. This determined wall thickness is then compared to a target wall thickness using computer software. This target wall thickness is assigned to the component section based on its previously established classification and can, for example, be stored in a lookup table within the computer and retrieved accordingly.The target wall thickness can be, for example, a minimum wall thickness, a maximum wall thickness, or a target wall thickness range, which is used as a specification for the computer-generated wall thickness. If a deviation is detected during this computer-generated comparison of the actual virtual wall thicknesses in the component section and the target wall thickness, an error value is calculated that characterizes the extent of this deviation and is compared with a permissible error value. If the extent of the deviation is so large that the computer-generated error value exceeds the permissible error value, this means that the wall thickness in the component section does not meet the quality requirements for the entire component and, consequently, a modification of the component data set is required before the manufacturing data set can be created from it.It is important to understand that the calculation of the error value and its comparison with a permissible error value can be performed in various ways. For example, the wall thickness can be determined at a specific location or at any point within the component section, and any deviations from or exceeding of the target wall thickness can be identified for each of these individual wall thicknesses. Furthermore, average wall thicknesses can be calculated and compared with corresponding target wall thicknesses to determine an error value, or an error calculation can be performed in another way to generate a representative value for the quality of the wall thickness in the component section.In particular, areas of the component section where the wall thickness is critically undershot or exceeded can be weighted so highly in the evaluation and comparison with the permissible error value that such areas, even if they only constitute a small dimension of the component section, trigger an exceedance of the permissible error value and consequently lead to a necessary modification of the component data set.

[0037] This modification can be automated using a pre-programmed algorithm, for example, by modifying the component data set so that areas where the wall thickness in the component section does not correspond to the target wall thicknesses are made thicker or thinner accordingly, in order to be above a minimum wall thickness, below a maximum wall thickness, or within a wall thickness range. Alternatively, the component data set can also be modified manually by a user, for example, by the user making corrections to the virtual component via suitable user interfaces, which are then implemented as changes in the component data set to alter the wall thicknesses so that they fall within the target range.

[0038] Only after these steps have been completed and the component data set has been modified accordingly, so that the computer-generated error value in the component sections is below the permissible error value for each component section type, is the manufacturing data set created from the component data set. This ensures that the manufacturing data set controls the production system in such a way that no deviation in wall thickness from the target wall thickness specified for each individual component section occurs.

[0039] Basically, it should be understood that with this modification of the component data set, the segmentation of the component data set and corresponding assignment of the parameter in the form of determining the wall thicknesses can be carried out again after each modification of the component data set in order to perform a corresponding check of the actual virtual wall thicknesses with the target wall thicknesses in each component section and in this way to check whether the component data set meets the quality requirements in all component sections.

[0040] According to another preferred embodiment, it is provided that

[0041] j. a deviation of the determined error value from the permissible error value is displayed to a user in a virtual component representation via a computer's graphical user interface and the computer is programmed to receive user input in step h) or i) by means of which a wall thickness change in the virtual component representation is determined,

[0042] k. the modification of the component data set in step h) or i) into a modified component data set from an application of the specified wall thickness change to the component data set, and

[0043] that

[0044] l. steps c) and f) or d) and g) are repeated with the modified component data set, and

[0045] m. a deviation of the error value determined in step I) from the permissible error value is displayed to a user of the virtual component representation via the computer's graphical user interface.

[0046] According to this training, the deviation of the measured error value from the permissible error value is displayed to the user via a computer's graphical user interface. This can be done, for example, by presenting the user with a three-dimensional virtual model on a screen, where the corresponding deviations of the measured error value from the permissible error value are indicated by appropriate markings, such as color coding. The user can then use this virtual three-dimensional model of the component to identify in which sections and areas of the respective sections the wall thickness does not correspond to the target wall thickness.Using appropriate computer-aided virtual tools, the user can then modify the wall thickness in such areas, for example, by adding or removing excess material in a specific section of the component, thus conveniently achieving an optimized component geometry. It is important to understand that after each manual modification of the component data set by the user, the graphical representation can be updated. For this purpose, the error value can be recalculated and compared again with the permissible error value to show the user whether the modification has changed the wall thickness to meet the requirements for the specific type of component section.If necessary, the AI-based computational model-supported segmentation can also be carried out again after a corresponding modification of the data set, since such modifications can also shift the boundaries of component sections and consequently an area of ​​the component data set can fall under a different component section and consequently a different type, which can change the requirements for this area.

[0047] It is even more preferred if the error value determined in steps h) and i) is calculated by computer-aidedly classifying the wall thicknesses assigned in each component section into at least two different categories, wherein a first category defines a permissible deviation of the wall thickness and a second category defines an impermissible deviation of the wall thickness, an overall assessment of the component section is computer-aidedly determined from the dimensions of the areas of a component section and a weighting of the respective category of the area of ​​the component section, in particular by a mean calculation formed with the weighting, and by comparing the overall assessment of the component section with a predetermined quality score, a fulfillment or non-fulfillment of a quality requirement for the component section is determined.

[0048] According to this embodiment, several areas within a component section are assigned wall thicknesses by computer, and these wall thicknesses are included in the calculation of the error value. The resulting deviation of the wall thickness from the target wall thickness is divided into at least two categories, one defining an acceptable deviation and the other an unacceptable deviation. Optionally, further categories can be added, for example, a category defining a deviation from the target wall thickness that can be compensated for by other areas of the component section if these have a sufficient or greater wall thickness. From the computer-determined category assignments for the several areas of the component section, an overall assessment of the component section is then calculated.This combined assessment of all areas of the component section serves to develop a differentiated overall judgment of the component section in order to determine whether or not it meets a quality requirement. For this purpose, the relevant categories of the component section's areas, possibly taking into account their volume, can be included in an average calculation, which can then be weighted. For example, areas that deviate critically from a target wall thickness can be weighted highly, up to such a high weighting that it directly results in the entire component section failing to meet a quality requirement.

[0049] This overall assessment, determined in this way, can then indicate to the user whether a component section has a wall thickness profile that meets the quality requirements for that type of component section, or whether this component section fails to meet these quality requirements, for example, because it has a consistently insufficient wall thickness or because it has a significantly insufficient wall thickness in one area. The fulfillment or non-fulfillment of the quality requirements can, in turn, be displayed to the user via a graphical user interface, thus enabling a particularly quick and efficient modification of the component data set to generate a data set that meets the quality requirements in each of the segmented component sections.

[0050] According to a further preferred embodiment, the assignment of the first parameter in step c) is an assignment of a first manufacturing parameter, the assignment of the second parameter in step d) is an assignment of a second manufacturing parameter, wherein the first and second manufacturing parameters, respectively, define in particular a manufacturing accuracy, a surface quality or a material density.

[0051] According to this embodiment, a first and a different second manufacturing parameter are assigned to the component sections differentiated using the computer-aided process. These manufacturing parameters, along with the component data set, serve as the basis for creating the manufacturing data set. This makes it possible to assign different surface qualities, manufacturing accuracies, or material densities to the various component sections and implement them in the subsequent manufacturing process of the component based on the manufacturing data set. The assignment of the manufacturing parameters can be done manually, for example, by a user checking the previously computer-segmented component sections and making the corresponding assignment of manufacturing parameters through appropriate data input.Alternatively or in combination with this, the assignment of manufacturing parameters can also be done by a calculation rule, for example by regularly assigning certain types of component sections to a specific manufacturing parameter, or by assigning such a manufacturing parameter assignment through the computer-aided segmentation of the component into the different component sections using machine learning.

[0052] According to another preferred embodiment, it is provided that in steps b) to d) the component data set

[0053] i. is transformed in an input multilayer perceptron, which uses node data of the component data set as input data,

[0054] ii. in a graph convolutional network, preferably a feature-steered graph convolutions for 3D analysis based on features of edges formed between the nodes of the component dataset, the three-dimensional shape of the component is analyzed, and

[0055] iii. in an output multilayer perceptron, which uses as input data features of the nodes of the component data set enriched in steps i) and ii), a type is assigned to each or groups of nodes.

[0056] In this implementation, the machine learning-based artificial intelligence model is comprised of three main components specifically designed for processing graph structures and three-dimensional data. The input Multilayer Perceptron (MLP) transforms the raw input data, which contains the features of the nodes. The Graph Convolutional Network (GCN) analyzes the 3D shape, a process that can be performed particularly efficiently using a Feature-Steered Graph Convolutions for 3D Shape Analysis (FeaStNet). FeaStNet differs from conventional GCNs in that the convolution kernels are dynamically controlled based on edge features. This type of GCN is particularly well-suited for component datasets in the form of nodes and edge definitions. The GCN can be adapted to other component dataset formats such as voxels or SDN.This achieves a particularly effective adaptation for the analysis of three-dimensional shapes, as the topological and geometric relationships between the nodes are taken into account. In the initial MLP, the node characteristics are further processed after processing by the GCN, and the final type assignment is made, in the sense of assigning each node to a component segment and its type. This MLP uses the characteristics enriched by the previous building blocks and assigns a type to each node, so that all nodes of a component segment differentiated from the component by segmentation are assigned to such a type.

[0057] According to a further preferred embodiment, the computational model trained by machine learning is trained using data comprising a multitude of component data sets of a multitude of different components, each component data set containing a user-managed subdivision into component sections and an assignment of a component section type to each component section, wherein a loss function used in the machine learning process is corrected by means of weighting factors that compensate for a more frequent occurrence of data from component sections of one type compared to a less frequent occurrence of data from component sections of another type.

[0058] According to this embodiment, the computational model trained through machine learning is trained with data relating to components previously subdivided into component sections, each of which has been assigned a type. The data used for machine learning therefore includes, on the one hand, component datasets describing the geometry of the components, and on the other hand, data describing how these components were previously manually subdivided into component sections by a user and which type was assigned to the component sections defined by this subdivision.This data can be implemented, for example, in such a way that each node in the component data set is assigned a type, so that a node, in addition to three coordinates that describe its position in a coordinate system, has a further specification that describes the type of the node, so that a component section of a certain type is defined by a corresponding large number of connected nodes.

[0059] When processing this data in the machine learning process, it is preferable to correct any loss function used by applying weighting factors. These factors compensate for the fact that nodes of some component section types occur more frequently than others, for example, because one component section type occupies more volume than another, or because one type of component section is present more often in the component than another. These weighting factors prevent the computational model in the machine learning process from favoring the more frequently occurring nodes of component section types as majority classes and neglecting the less frequent nodes as minority classes, which would reduce the accuracy of the assignment of types to component sections.The weighting factors are adjusted in such a way that, during the learning process, they increase the importance of minority classes and weaken the dominance of majority classes. This results in the computational model achieving a more balanced performance across all component section types.

[0060] It is even more preferred if the computational model trained by machine learning is trained using data comprising a multitude of component datasets of a multitude of different components, each component dataset with a user-defined manual subdivision into component sections as a ground-truth label and the assignment of a component section type to each component section, wherein the computational model trained by machine learning comprises an overarching meta-computing model and several subordinate specialized computational models, wherein a first specialized computational model is trained by machine learning to have optimized accuracy for recognizing a component section of a first type, a second specialized computational model is trained by machine learning to have optimized accuracy for recognizing a component section of a second type, and the meta-computing model is trained.to weight the calculations of the component sections of the first and second types, derived from the first and second special calculation models, relative to each other.

[0061] According to this embodiment, several different computational models trained through machine learning are used to perform the subdivision of the component into component sections. These computational models comprise at least two, preferably several, specialized computational models, each optimized in its accuracy to recognize a specific type of component section. Each of these specialized computational models can consist of or include the three main building blocks described above: input MLP, GCN, and output MLP. The result of such an analysis of a component dataset by such a specialized computational model is a subdivision of the component into several component sections and a type assignment for each component section, with particularly high accuracy with respect to a specific type of component section.The first special computational model is specialized for a first type and has optimized accuracy for this type, the second special computational model has optimized accuracy for another, second type of component section, and several further special computational models, each optimized for a specific type of component section, can be used if necessary.

[0062] The results of all specialized computational models are then combined by a higher-level meta-model, which, taking into account the respective accuracies of the specialized models, calculates a final subdivision of the component into its various sections and assigns the types to these sections. Because this meta-model can draw on the high accuracies of the individual type assignments of the specialized models and integrates them into a comprehensive overview and classification of the component sections according to the optimized accuracies considered in the meta-model, a particularly high level of accuracy is achieved in both the subdivision into the component sections and the type assignment. The meta-model can be a programmed algorithm, but preferably it is a machine learning-based model.

[0063] It is particularly preferred if the meta-computing model is trained by machine learning from the results of the specialized computational models and the user-defined component sections, along with the type assignments to these sections as ground-truth labels. According to this embodiment, the meta-computing model is trained by machine learning, using as training data the results of the specialized computational models and datasets containing user-defined components and manually assigned type assignments to these sections as ground-truth labels. This allows the model to learn an optimized consideration of the results of the specialized computational models with regard to the appropriate type assignments of the component sections.

[0064] Another aspect of the invention is the use of the previously described method for generating a manufacturing data set for medical products, wherein the various component sections are preferably assigned parameters such as a target wall thickness, surface quality, manufacturing accuracy, support density, and / or post-processing. In principle, AI-supported segmentation of a component into multiple component sections is particularly well suited for medical products such as implants, bone replacement prostheses, dental prostheses, or exoprostheses, since these often have a geometrically complex three-dimensional structure with few surfaces that can be described by simple geometric references. Furthermore, such medical products are frequently subject to legal or other regulatory requirements that prescribe their quality in various respects.Accordingly, it is particularly advantageous, for example, if a target wall thickness, surface quality, manufacturing accuracy, support density, and / or post-processing can be individually assigned to the AI-supported segmented component sections. It should be understood that the assignment of the target wall thickness is used in the planning process, for example, to adjust the actual wall thickness by modifying the component data set accordingly—that is, an actual geometric change of the component is defined in the virtual planning process by modifying the component data.In contrast, parameter assignments that affect surface quality or manufacturing accuracy are generally not associated with a change to the component data set, but rather correspond to the assignment of manufacturing parameters, such as a layer thickness, a traverse speed during material deposition or selective curing, a focal point size or laser beam intensity for material curing in the SLM process, a material deposition thickness in 3D printing, a milling cutter type or cutter dimensions in CAM milling, or the like. Here, the respective manufacturing specifications for the selected additive or subtractive manufacturing process must be set so that they correspond to or implement the assigned manufacturing parameter.

[0065] A support density can also be set using such manufacturing parameters, for example, when, during compilation, supports are selected from a library based on a support density assigned to a component section and their number is chosen according to the support density. This can then be parameterized and implemented in an adjustment of the component data in the production data set to also describe the geometry of these supports. In other pre-planning methods, the component data set can also be supplemented with this support data and then converted into the production data set.

[0066] AI-supported segmentation can also be used to automatically identify areas requiring additional rework. This might be necessary, for example, due to increased surface finish requirements. To avoid having to rework the entire component in this case, the segmentation process defines these previously segmented sections of the component by assigning the necessary rework as manufacturing parameters. Such rework can involve adding a manufacturing allowance to these sections by modifying the data accordingly. This allowance is then removed with high precision in a subsequent mechanical machining process, such as milling, turning, or grinding.Preferably, a distinction can also be made between a "top" and a "bottom" as separate component sections, so that post-processing is only carried out on one of the two sides. Preferably, the component is then produced in a first forming process using an additive manufacturing process. Subsequently, the surface finish is achieved on the previously defined areas of this component using a subtractive manufacturing process. For this purpose, it is advantageous if the areas to be post-processed have been virtually thickened beforehand in the component data set.

[0067] In principle, the desired segmentation can be performed once, thereby separating different component sections. Preferably, two or more segmentations of the component can also be carried out sequentially (serially) or simultaneously (in parallel). This makes it possible to define a first set of different component sections for one parameter and a second set of different component sections for a second parameter. For example, the first parameter can be used to check the minimum wall thickness of the component sections, and the second parameter can be used to assign different surface qualities to the different component sections.The component sections resulting from the first segmentation can differ from the component sections of the second segmentation, i.e., they can represent further subdivided component sections, combined component sections, or partially overlapping component sections.

[0068] In all embodiments of the invention, a distinction can preferably be made between the "top" and the "bottom" of the component section, so that the parameter to be applied can be applied to the entire component section (e.g., for the parameter assignment "wall thickness") or only one side of the component section (e.g., for the parameter assignment "surface quality," "support density," or "rework"). Preferably, in all embodiments of the invention, two or more segmentations of the component can also be carried out sequentially or simultaneously, as explained above.A further aspect of the invention is the use of the previously described method for generating a manufacturing data set for dental prostheses, wherein a first type of component section is assigned a first minimum wall thickness as the target wall thickness, a second type of component section is assigned a second minimum wall thickness as the target wall thickness, which differs from the first minimum wall thickness, and the assignment of the first or second parameter comprises determining a wall thickness profile in the respective component section, comparing the wall thickness profile in a component section with the minimum wall thickness assigned to the type of component section, and modifying the component data set computer-aided by means of manual input by a user or by means of a pre-programmed algorithm when a computer-aided determined error value, which results from a deviation of the first or second wall thickness from the first or second, is obtained.The second minimum wall thickness is formed, a predetermined permissible error value is exceeded, the manufacturing data set is created from the component data set, if in each of the component sections the error value determined by computer from the component data set, or the possibly modified component data set, does not exceed the permissible error value for the respective component section.

[0069] According to this aspect of the invention, the inventive method is used in a specific way to check dental prostheses with regard to their mechanical quality by determining the wall thicknesses of the dental prosthesis in different component sections and comparing them with the respective specific minimum wall thicknesses for the respective component sections.For this purpose, the machine learning-based computational model divides the dental prosthesis into several component sections, each of which is assigned a type. For example, in the data set for a cast partial denture, the clasps (retaining and supporting elements), the base, and the support surfaces are segmented as component sections and assigned to their respective types. The minimum permissible thicknesses for these component sections are then compared with the actual wall thicknesses that can be determined from the component data set. In a simplified manner, this can be done by ensuring that the minimum wall thickness in any component section is never undercut to guarantee sufficient quality and allow the manufacturing data set to be generated accordingly.In improved models, an error value can be generated from local instances where the minimum wall thickness is not met. This value depends on the size of the area that falls below the minimum wall thickness and the difference between the actual wall thickness and the minimum wall thickness in that area. This allows areas with a wall thickness below the minimum to be partially compensated for by adjacent areas with a wall thickness above the minimum. Furthermore, if the resulting error value is below a permissible threshold, a dental prosthesis can still be manufactured from the generated production data set, even if a specific minimum wall thickness is not met locally in one or more component sections.In this context, a cast metal partial denture is understood to be a dental prosthesis, but other dental prostheses used by the dentist to treat the patient, such as dental abutments, crowns or bridges, can also be tested and manufactured using the invention.

[0070] Another aspect of the invention is a device for assigning parameters to a component segmentally, comprising: a. electronic means for reading a three-dimensional component data set describing the geometry of a component, which describes the component by means of three-dimensional data of a surface of the component, into a computer; b. electronic means for virtually segmenting the component data set into several component sections and assigning the segmented component sections to a type of at least two different component section types by means of a computational model developed by machine learning, wherein the electronic means are configured to assign a first parameter to a first component section of the several component sections by means of the computational model; and d. assign a second parameter to a second component section of the several component sections; e.electronic means that are designed to create a manufacturing data set from the component data set with the first and second parameters, which is designed to control a data-controlled machine manufacturing device for the production of the component.

[0071] The device according to the invention can comprise a computer in the form of a single local machine on which the means according to the invention are implemented by appropriately programmed hardware. However, the device can also be configured as two or more computers or servers connected via a network, which execute one or more of the computational steps that must be performed by the device at locally different locations and communicate the corresponding result data to each other as input data for the next computational step.Finally, another aspect of the invention is a device for manufacturing a component in a data-controlled additive or subtractive manufacturing process, comprising the device described above and a manufacturing device configured to receive the manufacturing data set and further configured to manufacture the component electronically controlled by the manufacturing data set.

[0072] According to this embodiment, the device according to the invention further comprises a manufacturing device that can produce the component defined by the optimized component data set and the manufacturing data set generated therefrom, based on the manufacturing data set. Such a manufacturing device can, for example, be a laser melting system in which the component is produced layer by layer from powder material by subjecting it to selective local melting and subsequent solidification in each layer. Other rapid prototyping manufacturing systems, such as stereolithography, 3D printing, or the like, can also be used. Furthermore, subtractive manufacturing processes, i.e., material removal processes such as computer-aided three-dimensional milling, can be used to produce the component from a blank by selectively removing material.Finally, combined manufacturing processes, which are carried out in one or more different manufacturing devices, can also be used to produce the component using the manufacturing data set.

[0073] A preferred embodiment is explained with reference to the accompanying figures. These show:

[0074] Figure 1 shows a schematic sequence of an embodiment of a method according to the invention,

[0075] Figure 2 shows a top view of a cast partial denture with eleven component sections, which fall into four different types,

[0076] Figure 3 shows a graphic representation of a cast partial denture which has such thin walls that a tolerance value in relation to a minimum wall thickness is exceeded, and

[0077] Figure 4 shows a schematic representation of a possible operating mode of the machine-trained computational model. Referring initially to Figure 1, a method according to the invention begins with the reading in of a component data set 10 from a CAD design of a component, such as a cast partial denture. This component data set consists of the coordinates of a plurality of interconnected nodes, whose spatial position is defined by the coordinates and whose relationship to each other is defined by edge profile information contained in the component data set. This plurality of nodes and edges defines the surface or the entire volume of a component.

[0078] In a second step, the component data set is passed on to several specialized computational models 20a, b, c, which analyze this component data set using artificial intelligence trained through machine learning. Each of these specialized computational models 20a-c has been trained using the principle of supervised learning with a large number of component data sets. These component data sets have been manually subdivided into component sections, and each section has been assigned a type as truth by a user. Frequently occurring component section types are weighted down, while rarely occurring types are weighted up.The respective special calculation models 20 a, b, c have been optimized such that the first special calculation model 20a can recognize clasps as a component section on a cast partial denture in a special way, that the second special calculation model 20b can recognize support surfaces on a cast partial denture in a particularly optimized way with high accuracy, and that the third special calculation model 20c can recognize a base as a component section on a cast partial denture with high accuracy and can assign a corresponding type.

[0079] Each of these special computational models 20 a, b, c comprises, as the main components of the Kl computational model, an input MLP 20a', b', c', a FeaStNet 20a", b", c", and an output MLP 20a"', b"', c"', and delivers as a result a subdivision of the component into several component sections and an assignment of each component section to a specific type. The accuracy of each of the special computational models 20a, b, c is optimized for a specific type of component section.

[0080] With reference to Fig. 4, the inventive method can, for example, proceed as follows: A computer model based on PyTorch, consisting of several components, is used to analyze and classify the nodes from a 3D dataset 400. First, an input MLP 410 is used to process the input data (e.g., the features of the nodes). This is followed by a Graph Convolutional Network (GCN) 420, based on a FeaStNet variant. This variant dynamically controls the convolution kernels based on the edge features, which is particularly advantageous for evaluating 3D shapes. Finally, the features processed by the GCN are passed to an output MLP 430, which performs the classification and, as a final step, assigns a label (design element) to each node in an output 440, thus assigning each node the type of component section to which it belongs.

[0081] A particularly preferred architecture is one in which the AI ​​model consists of precisely these three main components and explicitly focuses on the topological and geometric relationships within the 3D dataset. The connection information (edge ​​index) between the nodes is taken into account to fully exploit the neighborhood relationships.

[0082] Figure 4 shows a possible embodiment of such a neural network structure, consisting of the input MLP 410, the FeaStNet-based GCN 420, and the output MLP 430. The first two layers, 410 and 420, can optionally contain bias terms to allow flexible adaptation to different 3D geometries. The input data 400 consists of nodes and edge connections (vertices and edge indices).

[0083] The described AI model was implemented using PyTorch and supervised learning was applied, with each node in the 3D dataset having a label as the "ground truth". Due to the varying frequencies of the design elements in the training data, a weighted cross entropy was used as a loss function. This weighting is intended to ensure that even rare classes are adequately considered and not overshadowed by frequently occurring classes.

[0084] The Adam optimization algorithm was used. The number of epochs was approximately 250, with early stopping employed to prevent overfitting. The learning rate was adjusted based on the data characteristics in several trials and was mostly in the range of 10⁻¹⁵. 3 up to 10" 4The minibatch size depended on the available GPU memory and was chosen to achieve a balanced ratio between training accuracy and efficiency.

[0085] During training, common metrics such as loss and accuracy were monitored. As soon as the validation loss (and thus the performance on the validation data) no longer improved over several consecutive epochs, early stopping was triggered to prevent overtraining.

[0086] In parallel computation with multiple specialized models (ensemble network), a simple, fully connected feedforward network is used, in which the predictions of several specialized models are combined into a single overall result. Each of these specialized models focuses on a specific design element or subtask and delivers correspondingly tailored predictions. These individual results are then integrated as inputs into the feedforward network so that they can be weighted and combined to arrive at a final classification.

[0087] The Adam optimizer is used to train this ensemble network, with Weighted Cross Entropy serving as the loss function. This weighting takes into account imbalances in the class distribution, ensuring that no minority class is suppressed by more frequently occurring classes. Targeted aggregation of the specializations of the individual models results in a robust and generally more precise classification. From this, an output 440 is generated for each node, assigning a type that classifies the respective component section.

[0088] Numerous component datasets from various parts serve as input data for training the AI ​​model. Each of these parts was manually segmented into component sections by a user, and these sections were assigned a type. This subdivision and type assignment forms the ground truth for the learning process.

[0089] After training, the AI ​​model receives component data sets from components that have not yet undergone segmentation. As input data, the AI ​​model generates a segmentation of the component, assigning, for example, each or some nodes of the component to component sections and thus defining multiple component sections. Each component section is then assigned a type, which is also stored in the AI ​​model's input data.

[0090] The results of all specialized computational models 20a, b, c are passed to a meta-computing model 30. This meta-computing model 30 is trained through machine learning using a multitude of results from specialized computational models on component datasets and user-defined segmentation and typing of the component sections, and therefore also represents a computer-based computational model. Using input MLP 30', intermediate MLP 30", and output MLP 30"', with appropriate weighting of the results of the specialized computational models 20a, b, c with respect to accuracy for the individual component sections, it performs a final segmentation of the component into its constituent sections and assigns the types to these sections. The result of this process is a subdivision of the component into several sections and a highly accurate typing of each of these sections.

[0091] Figure 2 shows such a subdivision of the cast partial denture with four clasps 110a-d, two support surfaces 120a, b, one base 130, and four supports 150a-d. As can be seen, the component sections of type clasp, type 2 clasp, and type base occur in different sizes and with different numbers of nodes on such a typical cast partial denture. If such component data is used for machine learning of a computational model and the corresponding subdivisions into component sections and type classifications of the component sections are entered manually as ground truth labels, the problem arises that the nodes of the base occur more frequently in the learning model than the nodes of the clasps, and the model could therefore learn an undesirable shift in accuracy towards the more frequently occurring nodes of the respective component section types.

[0092] This can be compensated for by appropriate weighting factors incorporated into the loss function to achieve high accuracy in the computational model for each component section and its classification. These weighting factors consequently emphasize the component section with the bracket type more strongly and the component section with the base type less strongly, in order to learn a balanced computational model with high accuracy.

[0093] In a subsequent step 40, wall thicknesses of the respective component sections of the component are calculated from the component data set and compared with pre-stored minimum wall thicknesses, which are specified for each of the identified and typed component sections.

[0094] These calculated wall thicknesses are compared in step 50 with the minimum wall thicknesses pre-stored for the respective standardized component section, and an error score is calculated. This score takes into account in which areas the minimum wall thickness is undercut in the respective component section, by how much, and how large this area is. The error score is then compared with a permissible error score, which can be the same for all component sections or can be individually predefined for each component section.

[0095] If the calculated error score of a component section exceeds the permissible error value for that component section, this is indicated by a corresponding coloring of the critical, too thin areas 210 ad in a graphical visualization of the component and gives the user the opportunity, based on this graphical error representation, to change the geometry of the component in a step 60 in differentiation from the non-critical area 220, as shown in Figure 3, for example by adding additional material in this too thin area, or if necessary also by reducing too thick areas by removing material there.In Figure 3, for the purpose of a black and white representation, this differentiation is shown by different types of hatching; an actually realized virtual representation may of course differ from this and, for example, use a green-red distinction or represent critical areas by flashing or other highlighting signaling.

[0096] This virtual planning with material application and material removal can be checked by repeatedly feeding 61 the modified component data set into a recalculation with each determination of the wall thicknesses, subdivisions into component sections and typing of the component sections on the virtual model, until the user has planned a high-quality cast partial denture through the virtual material application and material removal and from this cast partial denture, which does not exceed the respective permissible error value in any of the component sections, a manufacturing data set is created in a subsequent step 70, with which a manufacturing plant can then be controlled to produce the component.

[0097] Basically, it should be understood that instead of manually optimizing the geometric design of the cast partial denture by the user, automatic optimization of the cast partial denture can also be carried out by a corresponding algorithm within the computer-aided planning, for example in such a way that in areas that have been identified as too thin, computer-aided material is automatically added to the virtual model to increase the wall thickness until the corresponding error value is undercut and the quality is consequently sufficient to produce the manufacturing data set and manufacture the component.

Claims

Claims 1. Method for segment-wise parameter assignment on a component, comprising the following steps: a. Reading a three-dimensional component data set describing the geometry of a component, which describes the component using three-dimensional data of a surface of the component, into a computer, b. virtual segmentation of the component data set into several component sections and assignment of the segmented component sections to a type of at least two different component section types using a computational model trained by machine learning, wherein the computational model c. a first parameter is assigned to a first component section of the several component sections, and d. a second parameter is assigned to a second component section of the multiple component sections, e. Create, from the component data set with the first and second parameters, a manufacturing data set that is designed to control the production of the component in a machine-based manufacturing process.

2. Method according to claim 1 , characterized by the fact that Assigning the first parameter in step c) is a computer-aided determination of a first wall thickness in the area of ​​the first component section from the data of the component data set describing the first component section. Assigning the second parameter in step d) is a computer-aided determination of a second wall thickness in the area of ​​the second component section from the data of the component data set describing the second component section. f. the first wall thickness is compared, using computer-aided calculations, with a first target wall thickness assigned to the component section type of the first component section, and g. the second wall thickness is compared, using computer-aided calculations, with a second target wall thickness assigned to the component section type of the second component section. h. the component data set is modified computer-aided by manual input from a user or by means of a pre-programmed algorithm if a computer-aided determined error value, which is formed from a deviation of the first wall thickness from the first target wall thickness, exceeds a predetermined permissible first error value, and i. the component data set is modified computer-aided by manual input from a user or by means of a pre-programmed algorithm if a computer-aided determined error value, which is formed from a deviation of the second wall thickness from the second target wall thickness, exceeds a predetermined permissible second error value, The manufacturing data set is created from the component data set if, in each of the component sections, the error value determined by computer from the component data set, or the possibly modified component data set, does not exceed the permissible error value for the respective component section.

3. Method according to claim 2, characterized by the fact that j. a deviation of the determined error value from the permissible error value is displayed to a user in a virtual component representation via a computer's graphical user interface and the computer is programmed to receive user input in step h) or i) by means of which a wall thickness change in the virtual component representation is determined, k. the modification of the component data set in step h) or i) into a modified component data set from an application of the specified wall thickness change to the component data set, and that I. steps c) and f) or d) and g) are repeated with the modified component data set, and m. a deviation of the error value determined in step I) from the permissible error value is displayed to a user in the virtual component representation via the computer's graphical user interface.

4. Method according to claim 3, characterized in that the error value determined in steps h) and i) is formed by The wall thicknesses assigned in each component section are computer-aided to be classified into at least two different categories, where a first category defines a permissible deviation of the wall thickness and a second category defines an impermissible deviation of the wall thickness. From the dimensions of the areas of a component section and a weighting of the respective category of the area of ​​the component section, an overall assessment of the component section is determined using computer support, in particular by an average calculation formed with the weighting, and by comparing the overall assessment of the component section with a predetermined quality score, a fulfillment or non-fulfillment of a quality requirement for the component section is determined.

5. Method according to claim 1 , characterized by the fact that Assigning the first parameter in step c) is an assignment of a first manufacturing parameter, Assigning the second parameter in step d) is an assignment of a second manufacturing parameter, where the first or second manufacturing parameter defines in particular a manufacturing accuracy, a surface quality, a material thickness, a number of additively manufactured support elements or a post-processing.

6. Method according to one of the preceding claims, characterized in that in steps b) to d) the component data set i. In an input multilayer perceptron, which uses node data of the component dataset as input data, the component is transformed; ii. In a graph convolutional network, preferably a feature-steered graph convolutions for 3D analysis, the three-dimensional shape of the component is analyzed based on features of edges formed between the nodes of the component dataset; and iii. In an output multilayer perceptron, which uses as input data features of the nodes of the component data set enriched in steps i) and ii), a type is assigned to each or groups of nodes.

7. Method according to any of the preceding claims, characterized by the fact that the computational model trained through machine learning is trained using data which a large number of component data sets from a large number of different components, Each component data record includes a user-defined manual subdivision into component sections and the assignment of a component section type to each component section. include, where a loss function used in the machine learning process is corrected by means of weighting factors that compensate for a more frequent occurrence of data from component sections of one type compared to a less frequent occurrence of data from component sections of another type.

8. Method according to any one of the preceding claims, characterized by the fact that the computational model trained through machine learning is trained using data which a large number of component data records of a large number of different components, each component data record with a ground truth label, a manual subdivision into component sections by a user and assignment of a component section type to each component section include, wherein the computational model developed through machine learning comprises an overarching meta-computing model and several subordinate specialized computational models, wherein A first specialized computing model is trained through machine learning to have optimized accuracy for recognizing a component section of a first type, a second specialized computing model is trained through machine learning to have optimized accuracy for recognizing a component section of a second type, The meta-calculation model is designed to weight the calculations of the component sections of the first and second types, which were performed using the first and second special calculation models, relative to each other.

9. Method according to the preceding claim, characterized by the fact that the meta-computing model is trained by machine learning from the results of the special computational models and the ground truth label manually assigned by the user.

10. Use of the method according to one of the preceding claims for generating a manufacturing data set for medical products, wherein the various component sections preferably serve as parameters • a target wall thickness, • a surface quality, • manufacturing accuracy, • a support density, and / or • post-processing 11. Use of the method according to one of the preceding claims for generating a manufacturing data set for dental prostheses, wherein preferably a first minimum wall thickness is assigned as a target wall thickness to a first type of component section, a second type of component section is assigned a second minimum wall thickness as the target wall thickness, which differs from the first minimum wall thickness, and Assigning the first or second parameter involves determining a wall thickness profile in the respective component section. the wall thickness profile in a component section is compared with the minimum wall thickness assigned to the type of component section, and The component data set is modified computer-aided by manual input from a user or by means of a pre-programmed algorithm if a computer-aided determined error value, which is formed from a deviation of the first or second wall thickness from the first or second minimum wall thickness, exceeds a predetermined permissible error value. The manufacturing data set is created from the component data set if, in each of the component sections, the error value determined by computer from the component data set, or the possibly modified component data set, does not exceed the permissible error value for the respective component section.

12. Device for segment-wise parameter assignment on a component, comprising: a. Electronic means for reading a three-dimensional component data set describing the geometry of a component, which describes the component using three-dimensional data of a surface of the component, into a computer, b. Electronic means for virtually segmenting the component data set into several component sections and assigning the segmented component sections to a type of at least two different component section types by means of a computational model trained by machine learning, wherein the electronic means are designed to assign a first parameter to a first component section of the several component sections by means of the computational model, and d. to assign a second parameter to a second component section of the multiple component sections, e. electronic means which are designed to create from the component data set with the first and second parameters a manufacturing data set which is designed to control a data-controlled machine manufacturing device for the production of the component.

13. Device for manufacturing a component in a data-controlled additive or subtractive manufacturing process, comprising a device according to claim 12 and a manufacturing device configured to receive the manufacturing data set and further configured to manufacture the component electronically controlled by the manufacturing data set.

Images