Method of generating a control sequence for a computer aided manufacturing facility

Abstract

A method of generating a control sequence for a computer aided manufacturing (CAM) facility comprises storing geometry specifications of previous parts together with a dataset of a computer aided man

Description

The present invention relates to a Method of generating a control sequence for a computer aided manufacturing facility, the method comprising: (a) digitally storing geometry specifications of previous parts together with a dataset of a computer aided manufacturing code previously produced; (b) digitally storing a new part geometry specification. Furthermore, the invention relates to an automated production machine or a computerized numerical control machine comprising at least one processor made and prepared to carry out said method. Computer-Aided Manufacturing [CAM] programming is the process of using computer software to generate instructions and code that control and automate manufacturing machines, such as Computer Numerical Control [CNC] machines - the instructions are sometimes named NC-code. CAM programming translates a design or model, typically created in Computer-Aided Design [CAD] software, into instructions that guide the machine's tools and movements to manufacture the desired product. The CAM programming leading into NC-code generation, is the main effort in the entire CAM programming process. CAM programming may be very time consuming depending on the complexity of the part or product being manufactured, the proficiency of the programmer, and the capabilities of the CAM software being used. CAM programming involves creating and optimizing tool paths, defining machining parameters, and simulating the machining process to ensure accuracy and efficiency. CAM programming requires experience and familiarity with the software used. It is one object of the invention to reduce the time effort and expertise requirement of the CAM-programmer for generating the CNC-code. The conventional CAM programming requires time, money and expertise, although the features to be manufactured might have already programmed for same or very similar features in different parts. The standard conventional approach to obtain CNC-code is a manual process. Basically, a machinist would refer to their own expertise and knowledge from programming previous parts that are most like the one currently at hand. This process relies on their capabilities to address new challenges and recall their previous experiences. The main limitation of this process is the human expertise which is a risk for each company. Once, an expert machinist leaves the company the knowledge may be lost. Another conventional approach is a rule-based CAM-system. Modern systems may comprise such support mechanisms for the CAM programmer. These algorithms are limited by rule-based methods, which focus on the part directly without taking the history of CNC programming of similar parts into consideration. Currently, the most common approach is divided into two steps. The first step is a feature extraction from the CAD model. In most cases it starts with rule-based feature detection. This step is followed by a parameter extraction from these detected features like height, diameter, largest and smallest radius, or other parameters. For the second step two options are common practice. One option focuses on out-of-the-box parametric tools and operation suggestion, often pre-defined feature-based machining [FBM] rules. The feature and its parameters go through pre-defined rules and the assumingly best possible tool and operation is suggested in the end. The other possibility is to generate own rules for the detected features to follow. The generation of own rules is a difficult process and needs additional training and even more qualified personal. The generated rules are very specific for very specific features and need adjustment in case of small differences. This process is time-intensive and expensive which is motivating the user to avoid this option. One commercially available example of a CAM system incorporating some rule-based elements is Siemens NX CAM (formerly known as Unigraphics NX CAM; https: / / plm.sw.siemens.com / en-US / nx / manufacturing / ). A commercially available rule-based system is Mastercam (https: / / www.mastercam.com / ) which is a CAM software that incorporates rule-based programming techniques to automate the generation of tool paths and machining instructions. It provides a range of predefined rules and strategies for various machining operations, such as milling, turning, and wire Electrical Discharge Machining [EDM], As lined out, the existing solutions for tools and operations suggestion come with difficult problems. Manual processing relies on a machinist’s capability to address new challenges and use previous experiences, wherein the knowledge requires a long training and may easily be lost for the investing company. The main issue with feature-based machining [FBM] rules is that they are not easily adaptable for new part geometries and customer specific challenges. This means that FBM is not scalable. The invention aims to generate a support system for the CAM programmer, which suggests tools and operations for the CAM programming of the features based on the previous parts and therefore reducing time and costs in CAM programming. This invention combines previous customer data and public machining knowledge, to overcome the shortcomings of rule-based systems and manual processing, while resolving the need for specialist-machinists. The invention solves the above discussed drawbacks and problems by a method of the incipiently mentioned type comprising the additional steps of: (c) Generating graph-structured data from said geometry specification of the previous parts and storing them in a database respectively linked to said dataset of the previously produced computer aided manufacturing code; (d) Generating graph-structured data from said geometry specification of said new part; (e) Identifying geometrical features of the new part as sub-graphs; (f) Performing a sub-graph search in said database for every geometrical feature of the new part to obtain sub-graph search results; (g) Selecting a previous part feature for each new part feature based on the sub-graph search results; (h) Extracting said computer aided manufacturing code from said dataset corresponding to the previous part features selected; (i) Identifying parameters of the extracted computer aided manufacturing-code to be adjusted to the specific feature geometry of the new part; (j) Adjusting parameters of the extracted computer aided manufacturing-code to the specific feature geometry of the new part; (k) Assembling the computer aided manufacturing code extractions to a computer aided manufacturing code for machining the new part. Herein, a graph neural network [GNN] is a type of neural network designed to process and analyze structured data represented as graphs. Herein the correct term for a GNN is "graph neural network" and not "graphical neural network." The term "graph" in GNN refers to the mathematical representation of data as a collection of nodes and edges, herein called “graph-structured data”, while "neural network" refers to the computational model inspired by the human brain's neural networks. Graph-structured data may also be called graph neural network data as this data format is suitable to be processed by graph neural network respectively by a GNN. In such graph, data entities are represented as nodes, and the relationships between these entities are represented as edges. GNNs are specifically designed to capture the dependencies and relationships between nodes and utilize them for tasks such as node classification, link prediction, and graph classification. According to one embodiment of the invention it is proposed that the graph-structured data comprises nodes and edges, wherein the graph-structured data is structured such that the node attributes are: - surface type, - surface area, - centroid, - length and / or - ratio of radial distance vector direction; the edge attributes are: - dihedral angle between two neighboring surfaces, - curve type of the edge between two neighboring surfaces, - edge length of the edge between two neighboring surfaces, - normal vector. The invention benefits from the fact that GNNs can handle irregular and non-Euclidean data structures. The GNNs propagates information between connected nodes in a graph, allowing each node to aggregate and update its representation based on the information from its neighbors. This propagation process is performed through multiple iterations or layers, enabling the GNN to capture increasingly complex patterns and relationships in the graph data. Preferably, the invention uses a graph neural network [GNN] based search or a graph querying tool for the step of searching respectively finding of similar features when comparing the part to be CAM-processed and known part specifications stored in said database (step of performing a sub-graph search in said database). Using GNN gives flexibility with regard to the input geometry as according to the experience underlying the invention. The GNN-based search behaves stable through over a wide range of geometries even when these are unsimilar to the known part specifications. According to the invention, when comparing the rule-based feature detection or conventional feature comparison algorithms GNN-based search appears to be superior. Known tools, for example specific search engines may be used to find similarities in the GNN data according to the invention. Some examples for such tools are: Graph querying tool of graph database engines like Neo4j, Amazon Neptune, or Microsoft Azure Cosmos DB. These tools are designed to store and query graph-structured data efficiently. Such graph querying tools use Graph Query Language (GQL), wherein GQL is a standardized language specifically designed for querying and manipulating graph data. It provides a set of syntax and operations to retrieve specific data from a graph database or perform complex graph traversal and pattern matching. Some examples of GQLs are Cypher (used in Neo4j), Gremlin (used in Apache TinkerPop), and GraphQL (a general-purpose query language that can be used with various graph databases). These GQLs enable users to express their queries and operations in a graph-oriented manner. Graph Analytics Platforms: Platforms like Apache Giraph, Apache Flink, or Apache GraphX provide graph processing capabilities and algorithms for analyzing large-scale graph data. These platforms can be used to find similarities in GNN data by applying graph algorithms like community detection, similarity measures, or graph pattern matching. Graph Visualization Tools: Visualization tools like Gephi, Cytoscape, or NetworkX can help visualize the graph structure and identify similarities in GNN input or output data. These tools provide interactive visualizations and clustering algorithms to explore and analyze the graph data. Machine Learning Libraries: Popular machine learning libraries like TensorFlow, PyTorch, or scikit-learn offer various algorithms and techniques for finding similarities in GNN data. These libraries provide functionalities for feature extraction, dimensionality reduction, clustering, or similarity measures that can be applied to GNN data. In the context of searching GNN-data using GNN-based search or graph querying tools, there is a Graph Neural Network (GNN) working in the background. This GNN underlying the GNN-based search or a graph querying tool may be trained during a training phase and may be operative during an operative phase. The GNN may take graph-structured data as input and may produce graph-structured output. The input graph may consist of nodes, edges, and their attributes, while the output graph may represent predictions or modified versions of the input graph. More preferred the input may comprise: Node Features: Each node in the graph may be associated with a feature vector that represents its attributes or characteristics. These features can be numeric values, categorical variables, or embeddings. Herein, embeddings refer to low-dimensional vector representations of nodes or edges in a graph. These embeddings capture the structural and relational information of the graph and are learned through the GNN model. Edge Features: In some cases, the edges between nodes may also have associated features, such as the strength of the connection or the type of relationship. Adjacency Matrix or Graph Structure: This matrix represents the connectivity information of the graph, indicating which nodes are connected to each other through edges. The output may depend on the specific task being performed by the GNN. The output format may include: Node-level Predictions: GNNs can predict labels or properties for each individual node in the graph, such as classifying nodes into different categories. Graph-level Predictions: GNNs can make predictions or classifications about the entire graph as a whole, such as determining whether a graph represents a certain concept or category. Edge-level Predictions: GNNs can predict properties or relationships between pairs of nodes, such as predicting missing edges or identifying the strength of connections. The output format may depend on the specific problem and the design of the GNN architecture. The GNNs may be designed flexible and may be able to adapt to the input based on the requirements of the task at hand. When training said GNN graph-structured training data may be used. This means that the training samples are represented as graphs, where each node and edge may have associated features or labels. The GNN may be trained to learn the patterns and relationships within the graph data, enabling it to generalize and make predictions on unseen graphs during the testing phase. The GNNs may also be applied to non-graph-structured data. In such cases specific modules of the GNN may be used to convert non-graph data into a graph representation which can be more easily acquired by the GNN. These modules capture the underlying relationships and dependencies in the non-graph data and represent them in a graph structure. The GNN may first convert the data into graph representations. For example, for problems involving sequential data (such as natural language processing or time series analysis), the data may be transformed into a graph structure where the nodes represent individual elements in the sequence, and the edges represent the temporal or contextual relationships between them. According to the understanding of the invention GNN and an LLM are both types of Al models but serve different purposes and operate on different types of data. While GNN are explained above in the context of GNN-query tools, LLMs primarily operate on sequential or textual data, learning patterns, grammar, and semantics of language. LLMs are commonly used in natural language processing tasks, including text generation, sentiment analysis, machine translation, and question answering. LLMs generate human-like language output, which can be used for tasks like text completion or summarization. As a specific embodiment in the field of LLMs specific Generative Pre-trained Transformer [GPT] models are known. While LLM is a broad term that refers to any model designed to understand and generate human-like language, GPT models specifically refer to a type of language model developed by OpenAI. GPT models are pre-trained on a massive corpus of internet text to develop a deep understanding of language. GPT models are therefore already powerful because they are well trained and are more focused on knowledge retrieved from the internet. GPT models can be fine-tuned on specific tasks or domains to improve performance. According to a preferred embodiment, the method may involve the using of a large language model [LLM] - in particular a GPT-module - for: (i) Identifying parameters of the extracted computer aided manufacturing-code to be adjusted to the specific feature geometry of the new part and / or (j) Adjusting parameters of the extracted computer aided manufacturing-code to the specific feature geometry of the new part and / or (k) Assembling the computer aided manufacturing code extractions to a computer aided manufacturing code for machining the new part. According to a preferred embodiment, using the LLM, may comprise that the extracted manufacturing information is used as input into the prompt template of the LLM. According to one embodiment of the invention it is proposed that the method includes the additional step of: - Training a large language model with training data to prepare the large language model for generating machine specific computer aided manufacturing code. The LLM may be additionally prepared or trained by uploading a manufacturing handbook and / or any other relevant available information about the tools and operations preferably as documents in the LLM. Preferably, specific tool libraries of available tools and machining capabilities of available machines may be uploaded and more preferably may be matched through a minimization optimization algorithm, searching for the best tool and / or machining operation given the LLM suggestion. This beneficial embodiment of the invention enables a generic Al algorithm to adapt to known best practices incorporating available knowledge seamlessly with least effort into generating computer aided manufacturing code for a new part. According to one embodiment it is proposed that the method includes the additional step of: - Feeding said extracted computer aided manufacturing-code from the computer aided manufacturing-codes of the previous parts into large language model and instructing the large language model to generate the computer aided manufacturing code for generating the new part. Using an LLM enables outputting of tool and operation suggestions which are not limited to a predefined set of rules and the output may significantly deviate from known part geometry used in potential training phases. The resulting system may be considered to be more creative. The resulting trained LLMs are able to suggest the correct tools and operations without the need of a specialist. According to the invention each feature (like hole, groove, chamfer etc.) on the geometry model is represented as a subgraph. According to a preferred embodiment of the invention a GNN may be trained to learn and extract relevant information from the given attributes. It was found that a dedicated training may not be necessary to obtain reasonable results but a dedicated training may enhance the result accuracy. To enrich the database a GPT model may be used for a document search, wherein said documents may include computer aided manufacturing code. The database may be enriched by additional information, such that features that are considered to be similar are specifically marked in the dataset to be linked to each other by similarity. This additional information may accelerate the graph neural network-based search finding subgraphs which are similar to the feature that needs to be manufactured. When searching for similar features an acceptance rate for similarity or similarity score threshold may be set e.g., by the user, first, that is decisive which subgraphs are acceptable when checking the similarity score. This search may result in a single similar feature that is found in respectively the previous parts stored in the database - which may be termed the historic data, but also it can return multiple similar historic features. One approach to select only one single historic feature of several search hits may be to choose the one with the best similarity score. If no historic data is found during the similarity search qualifying for extraction and usage for the new part the user may be informed that a manual addition of the respective feature may be necessary. One step of the method according to the invention is identifying parameters of the extracted computer aided manufacturing-code to be adjusted to the specific feature geometry of the new part. Subsequently an adjusting of parameters of the extracted computer aided manufacturingcode to the specific feature geometry of the new part may be based on an LLM. After the determination of similar features, which have a sub-graph similarity score larger than the defined threshold, the tools and operations that are used to manufacture these features are extracted from historic data. In detail the tool, tool type and its geometrical parameters (like dimensions) and its capabilities (feeds and speeds), the operation type and its parameters (selected faces, cutting areas etc.) are extracted. The extracted manufacturing information may be used as a direct input or an assisting prompt for the LLM and may be encoded in the prompt template. Most preferably a manufacturing handbook and / or any other relevant available information about the tools and operations is loaded as documents in the LLM. According to another preferred embodiment machine-specific tool libraries and machining capabilities may be matched through a minimization optimization algorithm, searching for the best tool / operation given the LLM suggestion. This option may be most beneficial as it allows a generic Al algorithm of the LLM to adapt to machine specific attributes seamlessly. Herein lies another advantage of the invention as the historic data incorporates company specific or company-customized rules. Conventional rule-based algorithms result in high generation costs and maintenance costs to keep the rules up to date in case of any significant part design changes. The invention solves the generation and maintenance problem of the company specific rules since it automatically extracts them from their historic data with high accuracy, and the newly generated CAM-code is directly adaptable to the available tool library. According to the invention the company specific or company-customized rules are basically inherited from recycling the previous work. Additional knowledge to manage or administer the method is not necessary since the invention’s method is able to extract information from similar features. This searched, selected, and extracted result may be used as a suggestion mechanism for the CAM programmer, suggesting tools and operations, which are generated automatically based on the company’s functionalities such as historic data and available tools and machine capabilities in the factory as well as the industry’s best practices. According to one embodiment it is proposed, the method includes the additional step of: - Machining said new part using the computer aided manufacturing code generated for machining on an automated production machine or a computerized numerical control machine. Another teaching according to a preferred embodiment of the invention, which is of equal importance, relates to an automated production machine or a computerized numerical control machine comprising at least one processor made and prepared to carry out a method according to the invention. All explanations given with regard to the Method are fully applicable. According to one embodiment it is proposed, the method includes the additional steps of: - identifying parameters of the extracted computer aided manufacturing-code to be adjusted to the specific feature geometry of the new part and - assembling the computer aided manufacturing-code extractions to a computer aided manufacturing-code for machining the new part. According to one embodiment it is proposed that said similarity is determined as cosinesimilarity or L2-similarity. Cosine similarity measures how similar two vectors are by calculating the cosine of the angle between them. It ranges from -1 to 1, where 1 means the vectors are very similar, 0 means no similarity, and -1 means they are completely dissimilar. L2 (Euclidean) Similarity: L2 similarity measures similarity between vectors by considering their Euclidean distance. It takes into account both the magnitude and direction of the vectors. Both cosine similarity and L2 similarity are commonly used in various fields to assess similarity between vectors or documents. They provide valuable insights into the degree of similarity between different items. Brief description of the drawings The properties, features and advantages of this invention described above, as well as the manner they are achieved, become clearer and more understandable in the light of the following description and embodiments, which will be described in more detail in the context of the figures. This following description does not limit the invention on the contained embodiments. Same components or parts can be labeled with the same reference signs in different figures. In general, the figures are not for scale. It shall be understood that a preferred embodiment of the present invention can also be any combination of the dependent claims or above embodiments with the respective independent claim. The figures show: Fig. 1: illustrates a flow diagram of an example methodology, wherein the flow diagram is focused to illustrate the stepwise structure of the method; Fig. 2: illustrates a flow diagram of an example methodology with a different structuring focused on the input into the large language model; Fig. 3: shows details of the search for similar sub-graphs; Fig. 4: shows generating graph-structured data from a geometry specification of the previous parts. Detailed description of the drawings Figure 1 illustrates a method according to the invention in a workflow diagram. The method deals with generating a control sequence CTS for a computer aided manufacturing CAM facility. During a first step (a) geometry specifications GMS of previous parts PPT are digitally stored together with a dataset DTS of a previously produced computer aided manufacturing CAM code. As a second step (b) a new part NPT geometry specification GMS is stored. The invention proposes to use graph-structured data in the CAD-CAM domain. As an essential step (c) of the method graph-structured data DTT from said geometry specification GMS of the previous parts PPT are Generated and stored in a database DBS respectively linked to said dataset DTS of the previously produced computer aided manufacturing CAM code. Furthermore, as a step (d) graph-structured data DTT are Generated from said geometry specification GMS of said new part NPT. The graph-structured data DTT is structured such that the node NDE attributes comprise surface SFC type TYP, surface SFC area ARE, centroid CTD, length LGT and / or ratio RTO of radial distance vector direction DRC. The graph-structured data DTT is structured such that the edge EDG attributes comprise dihedral angle AGL between two neighboring surfaces SFC, curve type TYP of the edge EDG between two neighboring surfaces SFC, edge EDG length LGT of the edge EDG between two neighboring surfaces SFC, normal vector. The use of use graph-structured data enables in step (e) to Identify geometrical features FTR of the new part NPT as sub-graphs SGP and to perform in step (f) a sub-graph SGP search in said database DBS for every geometrical feature FTR of the new part NPT to obtain sub-graph SGP search results RST. The sub-graph SGP search is a graph neural network GNN based search or may be done by a graph querying tool GQT During the sub-graph SGP search for every geometrical feature FTR of the new part NPT similarity to geometrical features FTR of the previous part PPT is determined and a similarity score SSC for pairs of previous part PPT features FTR and new part NPT features FTR is respectively assigned. Preferably said similarity may be determined as cosine-similarity CSL or L2-similarity L2S. To guarantee that only suitable sub-graph SGP features FTR are selected a similarity score threshold TRS is provided and said selecting of a previous part PPT feature FTR is done under the condition that the similarity score SSC for pairs of a previous part PPT feature FTR and a new part NPT feature FTR exceeds said similarity score threshold TRS. From these search results, which are significantly similar to geometrical feature FTR of the new part NPT in step (g) a previous part PPT feature FTR is selected for each new part NPT feature FTR based on the sub-graph SGP search results RST. During step (h) each dataset of each previous part PPT feature FTR selected is evaluated by extracting said computer aided manufacturing CAM code from said dataset DTS. As the previous part feature is similar but, in most cases, not identical to the new part NPT feature FTR in step (i) parameters of the extracted computer aided manufacturing-code are identified to be adjusted to the specific feature FTR geometry of the new part NPT. In subsequent step (j) parameters of the extracted computer aided manufacturing-code are adjusted to the specific feature FTR geometry of the new part NPT. After the sub-structuring in subgraphs and before manufacturing the new part step (k) of the method provides assembling the computer aided manufacturing CAM code extractions to a computer aided manufacturing CAM code for machining MCH the new part NPT. Preferably the assembling and / or adjusting of extracted computer aided manufacturing CAM code is done using a large language model LLM for assembling and / or adjusting machine MCH specific computer aided manufacturing CAM code CCD. Subsequently said extracted computer aided manufacturing-code from the computer aided manufacturing-codes of the previous parts PPT may be fed into a large language model LLM and the large language model LLM may be instructed to generate the computer aided manufacturing CAM code CCD for generating the new part NPT. To obtain highest accuracy and efficiency in the adjusting and assembling of feature specific sub-graphs SBG said large language model LLM may be trained with training data TDT to prepare the large language model LLM for assembling and / or adjusting and / or generating machine MCH specific computer aided manufacturing CAM code CCD. Preferably said training data TDT to prepare the large language model LLM for generating machine MCH specific computer aided manufacturing CAM code CCD may comprise a machine MCH specific manufacturing handbook MAN and / or information about the tools TOL and operations OPT and tool TOL libraries and machining MCH capabilities CPB available in the shop or company using the method according to the invention. Finally said new part NPT may be produced or machined using the computer aided manufacturing CAM code CCD generated for machining on an automated production machine MCH or a computerized numerical control machine MCH. The complete method as defined herein may be implemented and run on an automated production machine MCH or a computerized numerical control machine MCH comprising at least one processor CPU made and prepared to carry out a method according to the invention or one of its preferred embodiments. Figure 2 illustrates a flow diagram of an example methodology with a different structuring focused on the input into the large language model LLM the generation of graph-structured data DTT isn’t shown in figure 2. During the search in step (f) the similarity score of previously machined features FTR is determined in said database DBS for every geometrical feature FTR of the new part NPT. This illustration shows as step (f) said sub-graph (SGP) search to obtain sub-graph (SGP) search results (RST). The database DBS contains geometrical features FTR -as GNN-data - including details like diameters of boreholes etc.. Furthermore, the database DBS comprises the CAM code like tool TOL specifications and the details of the tool operations OPT. All this input is fed into the large language model LLM for assembling and / or adjusting machine MCH specific computer aided manufacturing CAM code CCD. The LLM further receives training data or preparatory data like a machine (MCH) specific manufacturing handbook (MAN) and / or information about the tools (TOL) and operations (OPT) and tool (TOL) libraries and machining (MCH) capabilities (CPB). The large language model LLM outputs an operation suggestion including tool path PTH and CAM-code CCD. Furthermore, the large language model LLM outputs tool parameter suggestions TPS (e.g., tool type, length, diameter, number of flutes ...), which may be used for matching with the library LBR of the available tools and finally may be used to select the best option of several options of tool path PTH and CAM-code CCD suggestions. Figure 3 shows details of step (f) - of performing a sub-graph SGP search in said database DBS for every geometrical feature FTR of the new part NPT to obtain sub-graph SGP search results RST. In Figure 3: - Ni means the number of nodes N in a sub-graph i. For every sub-graph SBG of said new part NPT; - SB means the number of sub-graphs SBG in a new part NPT; - m means the number of previous parts PPT; - esi means an embedding, which refers to low-dimensional vector representations of a node or an edge in said graph-data DTT. For every sub-graph SBG in a new part NPT a similarity score SSC based on the L2 similarity to the previous part features FTR in the database DBS is determined. In the next step the similarity scores SSC are compared to a similarity score threshold TRS and a best match is selected. Subsequently the new part NPT feature FTR and the previous part PPT feature FTR are compared, and the extracted CAM-code CCD is adjusted accordingly. The final output is tools TOL and operations OPT as CAM-code CCD. Figure 4 illustrates generating graph-structured data DTT from a geometry specification GMS of the previous parts PPT. After generating the graph-structured data DTT geometrical features FTR of the new part NPT are identified as sub-graphs SGP. These sub-graphs SGP are stored in said database DBS for further processing as described herein. Independent of the grammatical term usage, individuals with male, female or other gender identities are included within the term. List of reference numbers AGL angle ARE area CAM computer aided manufacturing CCD code CPB capability CPU processor CTD centroid CTS control sequence DBS database DRC direction DTS dataset DTT data EDG edge FTR feature GMS geometry specification GNN graph neural network LGT length LLM large language model MAN handbook MCH machine NDE node NPT new part OPT operation PPT previous part RST result RTO ratio SFC surface SGP sub-graph SSC similarity score TRS threshold TOL tool TYP type

Claims

1. Computer implemented method of generating a control sequence (CTS) for a computer aided manufacturing (CAM) facility, the method comprising:(a) Digitally storing geometry specifications (GMS) of previous parts (PPT) together with a dataset (DTS) of a computer aided manufacturing (CAM) code previously produced;(b) Digitally storing a new part (NPT) geometry specification (GMS);characterized by the additional steps:(c) Generating graph-structured data (DTT) from said geometry specification (GMS) of the previous parts (PPT) and storing them in a database (DBS) respectively linked to said dataset (DTS) of the previously produced computer aided manufacturing (CAM) code;(d) Generating graph-structured data (DTT) from said geometry specification (GMS) of said new part (NPT);(e) Identifying geometrical features (FTR) of the new part (NPT) as sub-graphs (SGP);(f) Performing a sub-graph (SGP) search in said database (DBS) for every geometrical feature (FTR) of the new part (NPT) to obtain sub-graph (SGP) search results (RST);(g) Selecting a previous part (PPT) feature (FTR) for each new part (NPT) feature (FTR) based on the sub-graph (SGP) search results (RST);(h) Extracting said computer aided manufacturing (CAM) code from said dataset (DTS) corresponding to the previous part (PPT) features (FTR) selected;(i) Identifying parameters of the extracted computer aided manufacturing-code to be adjusted to the specific feature (FTR) geometry of the new part (NPT);(j) Adjusting parameters of the extracted computer aided manufacturing-code to the specific feature (FTR) geometry of the new part (NPT);(k) Assembling the computer aided manufacturing (CAM) code extractions to a computer aided manufacturing (CAM) code for machining (MCH) the new part (NPT).

2. Method according to claim 1, wherein the graph-structured data (DTT) comprisesnodes (NDE) and edges (EDG), wherein the graph-structured data (DTT) is structured such that the node (NDE) attributes are:- surface (SFC) type (TYP),- surface (SFC) area (ARE),- centroid (CTD),- length (LGT) and / or- ratio (RTO) of radial distance vector direction (DRC);the edge (EDG) attributes are:- dihedral angle (AGL) between two neighboring surfaces (SFC),- curve type (TYP) of the edge (EDG) between two neighboring surfaces (SFC),- edge (EDG) length (LGT) of the edge (EDG) between two neighboring surfaces (SFC), - normal vector.

3. Method according to one of the preceding claims, wherein the sub-graph (SGP) search is a graph neural network (GNN) based search and / or done by a graph querying tool (GQT).

4. Method according to one of the preceding claims, wherein the sub-graph (SGP) search comprises analyzing for every geometrical feature (FTR) of the new part (NPT) similarity to geometrical features (FTR) of the previous part (PPT) and assigning a similarity score (SSC) for pairs of previous part (PPT) features (FTR) and new part (NPT) features (FTR).

5. Method according to preceding claim 4, wherein said similarity is determined as cosinesimilarity or L2-similarity.

6. Method according to preceding claim 4 or 5, wherein a similarity score threshold (TRS) is provided and said selecting of a previous part (PPT) feature (FTR) done under the condition that the similarity score (SSC) for pairs of a previous part (PPT) feature (FTR) and a new part (NPT) feature (FTR) exceeds said similarity score threshold (TRS).

7. Method according to one of the preceding claims, wherein the assembling and / or adjusting of extracted computer aided manufacturing (CAM) code is done using a large language model (LLM) for assembling and / or adjusting machine (MCH) specific computer aided manufacturing (CAM) code (CCD).

8. Method according to the preceding claim 7, wherein the method includes the additional step of:- Training a large language model (LLM) with training data (TDT) to prepare the large language model (LLM) for assembling and / or adjusting and / or generating machine (MCH) specific computer aided manufacturing (CAM) code (CCD).

9. Method according to at least previous claim 8, wherein said training data (TDT) to prepare the large language model (LLM) for generating machine (MCH) specific computer aided manufacturing (CAM) code (CCD) comprises:- a machine (MCH) specific manufacturing handbook (MAN) and / or- information about the tools (TOL) and operations (OPT) and tool (TOL) libraries and machining (MCH) capabilities (CPB).

10. Method according to one of the preceding claims, wherein the method includes the additional step of:- Feeding said extracted computer aided manufacturing-code from the computer aided manufacturing-codes of the previous parts (PPT) into large language model (LLM) and instructing the large language model (LLM) to generate the computer aided manufacturing (CAM) code (CCD) for generating the new part (NPT).

11. Method according to one of the preceding claims, wherein the method includes the additional step of:- Machining said new part (NPT) using the computer aided manufacturing (CAM) code (CCD) generated for machining on an automated production machine (MCH) or a computerized numerical control machine (MCH).

12. Automated production machine (MCH) ora computerized numerical control machine (MCH) comprising at least one processor (CPU) made and prepared to carry out a method according to one of the preceding method claims.

Images