CAM programming algorithm and system based on generative artificial intelligence large model

Abstract

The invention relates to the technical field of CAM intelligent programming, and discloses a CAM programming algorithm and system based on a generative artificial intelligence large model. The algorithm comprises the steps of analyzing a CAD document, fusing geometric features, tolerances and material information through a generative artificial intelligence large model, and generating a structured semantic skeleton. Then, the semantic skeleton is matched with a process knowledge base, a preliminary processing instruction cluster is generated, and a dynamic processing graph with a node dependency relationship is constructed; secondly, receiving an editing intention of a user for the graph, analyzing the intention by using a generative artificial intelligence model, and automatically reconstructing node logic and connection of the graph to form a corrected processing logic network; and finally, before execution, simulating a network execution process by using a generative artificial intelligence model, and identifying potential logic and resource conflict nodes. According to the invention, intelligent conversion from design to processing instructions, and intent-driven process chain dynamic optimization and conflict pre-verification are realized.

Description

Technical Field

[0001] This invention relates to the field of CAM intelligent programming technology, specifically to algorithms and systems for CAM programming based on generative artificial intelligence large models. Background Technology

[0002] In the field of computer-aided manufacturing (CAD), current mainstream CAM automatic programming technologies primarily rely on rule-based expert systems and geometric feature recognition algorithms. Existing methods first perform geometric analysis on the CAD model, extracting discrete features. Then, based on a predefined, fixed process rule library, these features are mapped one by one to specific machining operation steps and parameters. This technical approach can effectively handle parts with well-defined structures and standard features; its essence is the matching and application of pre-defined logic to known patterns. However, when faced with parts with complex designs, tight process coupling between features, or rich tolerance and material information, existing technologies struggle to deeply understand the inherent manufacturing relationships between geometric morphology, dimensional accuracy requirements, and material properties. A significant semantic gap exists in the process planning process, from design information to executable manufacturing instructions, making it difficult to automatically generate globally collaboratively optimized process solutions.

[0003] When modifications to an existing processing sequence are required, existing systems typically require users to directly manipulate specific nodes or parameters on the process tree. After such local modifications, the system usually lacks the ability to automatically and coherently adjust the entire process chain. The input conditions, state dependencies, and corresponding updates to resource sequences for subsequent processes heavily rely on manual checks and re-coordination by process engineers. This modification model is inefficient and highly susceptible to human error in complex process connections and resource transfers, leading to logical contradictions or resource conflicts throughout the processing sequence, increasing the time required for process verification and the risk of solution failure. Summary of the Invention

[0004] The purpose of this invention is to provide an algorithm and system for CAM programming based on a generative artificial intelligence large model, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides an algorithm for CAM programming based on a generative artificial intelligence large model, the algorithm comprising: It receives and parses computer-aided design documents, extracts geometric features, tolerance annotations and material property information, and fuses the information to generate a structured semantic skeleton based on a generative artificial intelligence big data model; The structured semantic skeleton is input into a preset process knowledge base for matching and mapping. Based on the mapping relationship between processing strategies and parameters stored in the process knowledge base, a preliminary processing instruction set is generated. The initial processing instruction cluster is expanded into a series of directed and related processing step nodes. Input conditions and output states are defined for each processing step node. Based on the dependency relationship between the input conditions and output states, multiple processing step nodes are connected into a dynamic processing graph. The system receives the user's editing intent for the dynamic processing graph, which includes instructions to modify, add, or delete nodes of specific processing steps. The system then analyzes the editing intent using a generative artificial intelligence model and reconstructs the node logic and connection relationships of the dynamic processing graph accordingly to generate a corrected processing logic network. Before the modified processing logic network is executed, the topology and node content of the modified processing logic network are input into the generative artificial intelligence large model. The generative artificial intelligence large model simulates the network execution process to identify potential logical conflict nodes and resource conflict nodes.

[0006] Preferably, the step of receiving and parsing the computer-aided design document, extracting geometric features, tolerance annotations, and material property information from it, and fusing the information to generate a structured semantic skeleton based on a generative artificial intelligence model includes the following steps: Generative artificial intelligence large model is used to perform semantic segmentation on the overall descriptive text of computer-aided design documents to identify feature description paragraphs related to manufacturing; Entity and relation extraction is performed on the feature description paragraphs to identify key geometric entities, precision constraint entities, and material entities, and spatial and logical constraint relationships between entities are established. The identified key geometric entities, precision constraint entities, material entities, and the spatial and logical constraint relationships between entities are reorganized according to a preset manufacturing feature framework to form the structured semantic skeleton with manufacturing features as the core and associated with constraints and material properties.

[0007] Preferably, the step of inputting the structured semantic skeleton into a preset process knowledge base for matching and mapping includes the following steps: Each manufacturing feature unit in the structured semantic skeleton is compared with the feature template in the process knowledge base; For each manufacturing feature unit, a feature template with a similarity exceeding a threshold is selected, and the set of candidate processing strategies associated with the feature template is further filtered with the precision constraint attributes and material properties of the manufacturing feature unit. From the set of candidate processing strategies after secondary filtering, a processing strategy is selected according to the strategy priority rule, and the specific parameters of the manufacturing feature unit are bound to the selected processing strategy to generate basic processing instructions for the manufacturing feature unit. The basic processing instructions corresponding to all manufacturing feature units are collected to form the preliminary processing instruction cluster.

[0008] Preferably, the secondary filtering of the candidate processing strategy set associated with the feature template and the accuracy constraint attributes and material properties of the manufacturing feature unit includes the following steps: Obtain the set of candidate processing strategies associated with the feature template from the process knowledge base, and read the applicable accuracy range and material type of each candidate processing strategy; Extract the precision constraint attributes of the manufacturing feature units, including dimensional tolerances and geometric tolerances, and extract the material properties of the manufacturing feature units, including material hardness and toughness parameters; The applicable accuracy range of each candidate machining strategy is compared with the accuracy constraint attributes of the manufacturing feature unit, and candidate machining strategies whose applicable accuracy range covers the accuracy constraint requirements of the manufacturing feature unit are selected. The applicable material types of the selected candidate processing strategies are matched with the material properties of the manufacturing feature units. Candidate processing strategies with mismatched material types are eliminated to obtain a set of candidate processing strategies after secondary filtering.

[0009] Preferably, the step of expanding the initial processing instruction cluster into a series of directed and related processing step nodes, defining input conditions and output states for each processing step node, and connecting multiple processing step nodes into a dynamic processing graph based on the dependency relationship between the input conditions and output states includes the following steps: Each basic machining instruction in the initial machining instruction set is analyzed, decomposed into indivisible basic machining operations, and each basic machining operation is defined as a machining step node. Analyze the input conditions for each machining step node, including the blank state, fixture state, and tool positioning state; Analyze the output state of each processing step node, including the geometric state and surface state of the workpiece. Establish matching rules between output states and input conditions. If the output state of one processing step node can satisfy the input conditions of another processing step node, then establish a directed connection between the two processing step nodes, and finally form the dynamic processing map.

[0010] Preferably, the step of parsing each basic processing instruction in the preliminary processing instruction set, decomposing it into indivisible basic processing operations, and defining each basic processing operation as a processing step node includes the following steps: Read each basic machining instruction in the preliminary machining instruction set and identify the sequence of machining actions contained in the basic machining instructions; According to the atomicity rule of machining actions, each machining action sequence is decomposed into the smallest unit of machining operation. Each smallest unit of machining operation cannot be further divided and corresponds to an independent machine tool movement or tool operation. Assign a unique identifier to each smallest unit of processing operation, and define the input parameters and output results required for the processing operation; Each smallest unit of processing operation is encapsulated as a processing step node, and the input parameters and output results are recorded in the processing step node.

[0011] Preferably, receiving the user's editing intent for the dynamic processing graph, the editing intent including instructions to modify, add, or delete nodes of specific processing steps, and parsing the editing intent using a generative artificial intelligence model to reconstruct the node logic and connection relationships of the dynamic processing graph accordingly, includes the following steps: Receive the user's editing intent and mark the processing step node targeted by the editing intent as the change center node; Starting from the change center node, traverse the dynamic processing graph to find all upstream and downstream nodes on the logical path that are affected by the change center node, forming a set of affected nodes. Input the changes to the central node and the set of affected nodes into the generative AI model, and request the generative AI model to generate a logically consistent subgraph replacement scheme. The original logical structure from the change center node to the affected node set in the dynamic processing graph is replaced by the subgraph replacement scheme to complete the reconstruction and obtain the modified processing logic network.

[0012] Preferably, the step of inputting the changes to the central node and the set of affected nodes into the generative AI model, and requesting the generative AI model to generate a logically consistent subgraph replacement scheme, includes the following steps: Extract the changes to the change center node, including the modified processing operation description, newly added processing steps, or deleted node information; Obtain the set of affected nodes from the dynamic processing graph, including all upstream and downstream nodes that directly or indirectly depend on the change center node; Input the structured data of the changes and the affected node set into the generative AI model, and request the generative AI model to generate a subgraph replacement scheme based on processing logic constraints and resource constraints; Generative artificial intelligence large model outputs a subgraph replacement scheme, which includes a new node sequence and connection relationship to ensure the logical coherence and resource consistency between the new subgraph and the original graph.

[0013] Preferably, identifying potential logical conflict nodes and resource conflict nodes includes the following steps: Define a resource requirement list for each machining step node in the modified machining logic network. The resource requirement list includes the required machine tool axes, tool numbers, and fixture locations. During the execution of the generative artificial intelligence large model simulation network, the processing step nodes are activated sequentially in logical order, and it is checked whether the resource requirement list of the activated node overlaps with the currently occupied resources. If resource overlap occurs, the currently activated node is determined to be a resource conflict node; At the same time, check whether the input conditions of the activated node have been fully satisfied by the output states of all its upstream nodes. If not, the currently activated node is determined to be a logical conflict node.

[0014] Preferably, the present invention also includes a CAM programming system based on a generative artificial intelligence large model, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the CAM programming algorithm based on the generative artificial intelligence large model as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By leveraging generative artificial intelligence large-scale models to process multi-source design information and utilizing their deep semantic understanding and generative capabilities, a structured semantic skeleton integrating geometric, tolerance, and material properties is constructed. This skeleton is not a simple set of features, but rather expresses the inherent manufacturing logic relationships between design elements, forming a consistent high-level semantic blueprint that serves process planning. This allows the derivation of initial process schemes to be based on a comprehensive understanding of the design intent, overcoming the semantic limitations of traditional rule-based methods when dealing with complex and implicit process constraints.

[0016] By directly analyzing the user's advanced editing intent for the processing map through a generative artificial intelligence model, the model drives the adaptive reconstruction of the topology and node content of the entire processing logic network. The model can infer the impact of editing actions on the global network state, automatically adjusting the input and output conditions of relevant nodes and the dependencies between nodes to ensure that the generated new network is logically self-consistent. This achieves a shift from local manual adjustments to intelligent revision driven by global intent, simplifying the iteration process of process solutions and avoiding chain-like logical errors caused by local modifications.

[0017] Before execution, a generative artificial intelligence model is used to simulate and extrapolate the complete processing logic network. Based on its understanding of the manufacturing process and physical constraints, the model simulates the state evolution and resource consumption of each node during execution, and can proactively identify potential logical and resource conflict nodes, including sequence contradictions and resource contention. This pre-verification capability in a virtual environment provides intelligent and forward-looking assurance for the reliability and safety of the processing scheme. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the CAM programming algorithm based on a generative artificial intelligence large model as described in this invention. Figure 2 A flowchart for generating a structured semantic skeleton; Figure 3 A flowchart for generating the initial processing instruction set; Figure 4 This is a diagram showing the status analysis of each node in the processing steps. Figure 5 This is a diagram illustrating the impact of editing operations. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1This invention provides an algorithm for CAM programming based on a generative artificial intelligence (AI) large-scale model. The algorithm includes: receiving and parsing a computer-aided design (CAD) document; extracting key information such as geometric features, tolerance annotations, and material properties from the document using the generative AI large-scale model; and fusing these to generate a structured semantic skeleton, which serves as an intermediate representation connecting design and manufacturing. This structured semantic skeleton is then input into a pre-defined process knowledge base for matching and mapping. The process knowledge base stores a large number of feature templates and their corresponding processing strategies and parameters. A preliminary processing instruction set is generated through matching and mapping. This preliminary processing instruction set is expanded into a series of directed and related processing step nodes. For each node, the input conditions required for execution and the output state reached after execution are clearly defined. Based on the dependencies between these input and output states, all processing step nodes are connected into a dynamic processing graph, which visualizes the logical flow of the entire processing technology. The system allows users to interactively edit the generated dynamic machining atlas. It receives user requests to modify, add, or delete nodes for specific machining steps within the atlas. A generative artificial intelligence (AI) model analyzes these requests and intelligently reconstructs the node logic and connections of the dynamic machining atlas, generating a revised machining logic network. Before this revised network is finally converted into machine tool executable code, virtual verification is performed. The network topology and detailed information for each node are input into the generative AI model, which simulates the entire network's execution process to identify potential logical and resource conflict nodes, ensuring the feasibility and safety of the machining plan.

[0021] Example 1: See Figure 2 In practical implementation, the computer-aided design documents received by the system typically include three-dimensional geometric model data, along with associated drawing annotations and technical requirement text. A generative AI large-scale model performs semantic segmentation on the overall descriptive text of the computer-aided design document. This semantic segmentation operation, based on natural language processing technology, identifies descriptive paragraphs in the document that are directly related to the manufacturing process. By analyzing syntactic structure and contextual semantics, the generative AI large-scale model distinguishes text blocks describing manufacturing-related information such as part functions, assembly relationships, and process requirements, thereby filtering out general instructions or management information irrelevant to manufacturing. In some embodiments, the semantic segmentation process utilizes the generative AI large-scale model to encode and classify the text sequence, labeling each semantic unit as either manufacturing-related or manufacturing-irrelevant, thereby accurately extracting descriptive paragraphs containing processing features, precision requirements, and material specifications.

[0022] In practical implementation, entity and relation extraction from feature description paragraphs is the core step in constructing a structured semantic skeleton. Generative AI large-scale models utilize their pre-trained language understanding capabilities to identify various types of named entities from natural language descriptions. Key geometric entities include references to specific geometric shapes such as planes, cylindrical surfaces, holes, slots, chamfers, and threads. Precision constraint entities include dimensional tolerances, such as diameter and length tolerances, geometrical tolerances, such as parallelism, perpendicularity, and position, as well as surface roughness requirements. Material entities are explicitly defined as material grades or names, such as "45# steel" and "Aluminum alloy 6061". Generative AI large-scale models not only identify discrete entities but also further analyze and establish spatial and logical constraint relationships between them. Spatial constraint relationships describe the orientation and topological connections between geometric entities, such as "plane A is parallel to plane B" and "hole C is located at the center of plane D". Logical constraint relationships bind precision constraint entities, material entities, and geometric entities, such as "surface roughness Ra≤1.6 requirement applies to plane E" and "material hardness requirement HBW220-250 is associated with the entire part". It is understandable that the entity and relation extraction process transforms unstructured natural language text into a structured entity and relation network, providing a data foundation for subsequent information reorganization.

[0023] In practice, the identified key geometric entities, precision constraint entities, material entities, and the spatial and logical constraints between entities are reorganized according to a predefined manufacturing feature framework. This predefined manufacturing feature framework is a predefined data model with manufacturing features as the core organizational unit. The reorganization process first categorizes the identified key geometric entities into the corresponding feature categories defined by the manufacturing feature framework. For example, a set of associated cylindrical surfaces and an end face might be categorized as a "stepped shaft" feature. The reorganization process then attaches the precision constraint entities and material entities associated with these geometric entities as attributes to the corresponding manufacturing features. In practice, the spatial and logical constraints between entities are also transformed into representable relationships within the manufacturing feature framework, such as "adjacent" or "dependent" relationships between two manufacturing features. A generative artificial intelligence model drives the entire reorganization process, ensuring that information extracted from the original text is accurately and unambiguously mapped to the corresponding fields and connections of the structured semantic skeleton. The structured semantic skeleton is represented using a graph or tree structure, where nodes represent manufacturing features and their attributes, and edges represent relationships between features. Understandably, the final structured semantic skeleton completely strips away the original natural language form, transforming into a machine-readable and semantically clear data structure centered on manufacturing features and incorporating all constraints and material properties. This structured semantic skeleton serves as the standardized input for subsequent process knowledge base matching and mapping.

[0024] Example 2: See Figure 3In practice, the structured semantic skeleton contains multiple manufacturing feature units, each describing the specific geometry, precision constraints, and material properties of a part. The first step in the matching and mapping process is to compare the similarity of each manufacturing feature unit in the structured semantic skeleton with a predefined feature template in the process knowledge base. The process knowledge base is a database storing a large number of typical manufacturing feature models and their processing experience. A feature template is an abstract description of a class of manufacturing features with common geometric and process characteristics, such as "through hole," "blind hole," "countersunk hole," "rectangular groove," and "stepped surface." Each feature template is associated with one or more candidate processing strategies. The similarity comparison calculates the degree of matching between the current manufacturing feature unit and each feature template in the process knowledge base across multiple dimensions. In practice, the similarity comparison function comprehensively considers geometric shape similarity, topological similarity, and size ratio similarity. It can be understood that the calculation process is not a simple keyword matching, but rather a deep metric learning based on feature description vectors to ensure matching accuracy. The final similarity score between the manufacturing feature unit and the feature template is then calculated. It can be calculated using the following formula:

[0025] Where: represents a manufacturing feature unit With feature template The overall similarity score, This represents the geometric similarity score. This represents the topological similarity score. The score represents the similarity in size proportions. , and These are the weighting coefficients of the scores for each dimension, and For each manufacturing feature unit, select those with similarity scores exceeding a preset threshold. The feature template that satisfies The feature templates are then used in subsequent processing steps.

[0026] In practical implementation, after selecting feature templates with similarity exceeding a threshold for each manufacturing feature unit, a secondary filtering process is required, involving the set of candidate processing strategies associated with the feature templates and the precision constraint attributes and material properties of the manufacturing feature unit. The set of candidate processing strategies associated with the selected feature templates is retrieved from the process knowledge base. Each candidate processing strategy stores its metadata in the process knowledge base, explicitly including the applicable precision range that the strategy can reliably achieve and a list of recommended material types. The applicable precision range is typically represented by dimensional tolerance grades and geometrical tolerance grades, while the material type list includes specific material grades or material categories. The precision constraint attributes inherent in the manufacturing feature unit itself are extracted. These attributes include dimensional tolerances, such as the vertical deviation of the hole diameter, and geometrical tolerances, such as the positional tolerance of the hole's axis relative to a datum. Simultaneously, the material properties of the manufacturing feature unit are extracted, which may include the material's hardness value, toughness parameters, or heat treatment state. The first step of the secondary filtering is to compare the applicable precision range of each candidate processing strategy with the precision constraint requirements of the manufacturing feature unit, selecting those candidate processing strategies whose applicable precision range completely covers the precision constraint requirements of the manufacturing feature unit.

[0027] In practical implementation, a processing strategy is selected from the candidate processing strategy set after secondary filtering, according to a preset strategy priority rule. The strategy priority rule is a decision logic predefined in the system, which may comprehensively consider multiple optimization objectives, such as maximizing processing efficiency, minimizing processing cost, and ensuring the most stable processing quality. In some embodiments, the strategy priority rule can be a multi-objective decision function, where the input is the performance indicators of the candidate processing strategies, and the output is a comprehensive priority score. Optionally, the strategy priority rule can also be a simple rule-based ranking, such as prioritizing single-process strategies and then selecting composite-process strategies. From the candidate processing strategy set after secondary filtering, the processing strategy with the highest comprehensive priority is selected as the final choice according to the strategy priority rule. After selecting a processing strategy, the specific parameters of the manufacturing feature unit are bound to the selected processing strategy. The specific parameters of the manufacturing feature unit include the precise geometric dimensions, position coordinates, and direction vector of the feature. The selected processing strategy defines the macroscopic method and sequence of steps for processing. The binding process involves filling the specific parameter values ​​into the variable slots of the processing strategy, generating an executable, parameterized basic processing instruction for that manufacturing feature unit. In some embodiments, for a blind hole feature with a diameter of 10mm, a depth of 20mm, a tolerance of H7, and a material of 45 steel, after matching, filtering, and priority selection, the machining strategy of "drilling-reaming-reaming" may be selected. The binding process fills the diameter parameter of 10mm and the depth parameter of 20mm into the diameter and depth variables of the "drilling" step to generate specific drilling instructions; and fills the final target diameter of 10H7 into the variables of the "reaming" step to generate reaming instructions, thereby forming the basic machining instructions for this blind hole feature.

[0028] Example 3: In specific implementation, the preliminary machining instruction set contains a series of basic machining instructions for different manufacturing features. Each basic machining instruction in the preliminary machining instruction set is read; each basic machining instruction describes the macroscopic machining intent for a manufacturing feature. The parsing process requires identifying the sequence of machining actions contained within each basic machining instruction. A basic machining instruction such as "finish mill the upper surface to a height of 20±0.1 mm, surface roughness Ra3.2" is not an atomic operation. The system decomposes the machining action sequence based on predefined machining action atomicity rules, which define what granularity of operation is considered the smallest indivisible unit machining operation. The smallest unit machining operation corresponds to an independent, basic machine tool motion control instruction or tool state switching instruction. In specific implementation, a basic machining instruction of "finish mill the upper surface" may be decomposed into a series of smallest unit machining operations such as "rapidly position to the safety plane above the machining start point," "linear feed down to a depth of -0.5 mm," "mill one revolution of the contour in a clockwise direction," and "lift the tool to the safety plane." The decomposition is based on the logical independence of the operations and the independence of the machine tool execution. Each smallest unit machining operation has a clear start and end state during execution and usually corresponds to a program segment or a group of consecutive program segments in the CNC code. Each decomposed smallest unit machining operation is assigned a globally unique identifier within the current programming session. This identifier is used to uniquely identify this operation node in subsequent processes. Simultaneously, all input parameters required for the execution of each smallest unit machining operation and the expected output results after successful execution are clearly defined. Input parameters include the spatial coordinates of the machining start point, the feed rate of the cutting process, the spindle rotation speed, and the tool number information used. Output results are the changes the operation makes to the workpiece or environmental state, such as "0.5 mm of material was removed from the upper surface of the workpiece," "A hole with a diameter of 5 mm was formed at position (X1, Y1)," and "Tool No. 3 was mounted on the spindle."

[0029] In practical implementation, each smallest unit machining operation is encapsulated into a machining step node. This node is a data structure that records the operation's unique identifier, type description, a list of all required input parameters, and a description of the output result. The encapsulation process transforms natural language or parameterized operation descriptions into node objects with a strict format. For example, for the smallest unit machining operation "milling one circle of contour in a clockwise direction," the encapsulated machining step node will store: identifier "OP_0021," operation type "contour milling," input parameters including "starting point coordinates (Xs, Ys, Zs)," "feed rate F200," "spindle speed S3000," and "tool number T02," and an output description of "generating a closed contour path from the starting point to the ending point, with a path width equal to the tool diameter, and material removed." Once all basic machining instructions have been parsed and encapsulated into independent machining step nodes, the system obtains a set containing numerous nodes. At this point, no execution order association has been established between these machining step nodes; the construction of the dynamic machining map requires defining input conditions and output states for each machining step node. Input conditions are the workpiece or resource states generated by upstream operations that must be satisfied before a machining step node begins execution. Input conditions are usually more abstract than input parameters and include the initial or intermediate geometry of the workpiece, the clamping status of the fixture, whether a specific tool is installed and in place, and whether the coolant is on. Output states are the new states reached by the system after a machining step node is successfully executed. Output states are a summary and abstraction of the output results and include the new three-dimensional geometry of the workpiece, the surface roughness value of a specific surface, and the tool number currently installed on the machine tool spindle. In practical implementation, formally defining the input conditions and output states of machining step nodes is a prerequisite for establishing dependencies. The system establishes matching rules between output states and input conditions, and these matching rules determine the connection logic between nodes. The core of the matching rules is to check whether the output state set of a machining step node A can completely satisfy one or more of the input condition sets of another machining step node B. If the conditions are met, a directed connection is established between processing step node A and processing step node B, pointing from A to B. This directed connection indicates that processing step node B depends on the state generated by processing step node A, and processing step node A must be executed before processing step node B. The determination of state matching can be based on a state matching degree function. Perform the calculation:

[0030] in: This represents the degree of matching between the output state of processing step node A and the input conditions of processing step node B. This represents the set of output states for processing step node A. This represents the set of input conditions for processing step node B, and the operators. Indicates the number of elements in the set. This represents the intersection operation of sets. When the matching degree... Greater than a preset threshold When the system determines that the dependency relationship is valid, a directed connection is established. By traversing all processing step node pairs and establishing connections according to matching rules, the system dynamically links multiple originally discrete processing step nodes into a directed network structure based on the dependency relationship between their input conditions and output states.

[0031] See Figure 4 This chart presents a comprehensive analysis of the state distribution and dependency matching degree of each machining step node. The left side uses a color-coded bar chart to show the quantity distribution of various machining operations in the overall process. Different colors distinguish typical manufacturing operation types such as contour milling, planar milling, and drilling. The bar height clearly reflects the importance and execution frequency of each operation in the overall machining task. The right side, accompanied by a synchronously changing line chart, clearly depicts the fluctuation trend of the state matching degree throughout the machining sequence. This indicator uses a mathematical formula to quantify the degree of fit between input conditions and output states between nodes. The line trend intuitively reveals the coherence of the machining logic and the tightness of node dependencies. The chart design employs a professional color scheme and eye-catching annotations, enabling observers to quickly grasp the structural characteristics and logical rationality of the machining process, providing a visual basis for evaluating the overall optimization level of the manufacturing solution.

[0032] Example 4: In a specific implementation, the system presents a dynamic processing graph generated from the initial processing instruction set. The dynamic processing graph graphically displays processing step nodes and their directed connections. Users submit editing intentions for the dynamic processing graph through an interactive interface. These intentions include modifying the operation content of specific processing step nodes in the graph, inserting new processing step nodes at specified positions in the dynamic processing graph, or deleting existing specific processing step nodes. The system receives the user's editing instructions and marks the processing step node in the dynamic processing graph directly targeted by the editing intention as the change center node. The change center node is the specific location where the user wishes to perform add, delete, or modify operations, and it serves as the starting point for subsequent logical impact analysis. In a specific implementation, the system uses the change center node as the starting point and traverses the directed network structure of the dynamic processing graph simultaneously in two directions. The traversal operation proceeds along the directed connection edges, searching upstream for all processing step nodes whose output state is directly or indirectly dependent on the change center node, and downstream for all processing step nodes whose output state is directly or indirectly dependent on the change center node. By traversing the system, the system collects all processing step nodes along the logical path that will be affected by the edit operation of the change center node. These processing step nodes together constitute a set of affected nodes. The set of affected nodes includes not only adjacent nodes that are directly connected to the change center node, but also indirect adjacent nodes that may be affected due to dependency propagation.

[0033] In practice, the changes to the central node and related information about the affected node set are input into the generative AI model. The changes include the user-modified processing operation description text, the complete definition of newly added processing step nodes, or descriptions of logical gaps resulting from user-requested node deletions. Information about the affected node set is provided in structured data format, including attribute descriptions of each processing step node, current connections between nodes, input conditions, and output states for each node. The system sends a request to the generative AI model, asking it to generate a logically consistent new subgraph replacement scheme for the affected local graph region based on its understanding of processing logic constraints and manufacturing resource constraints. Logical constraints ensure that the sequence of processing steps conforms to the physical laws of geometric shape generation, while resource constraints ensure that the use of resources such as tools and fixtures does not conflict. After receiving the changes and the affected node set data, the generative AI model performs inference analysis to determine how the new state introduced by the changes connects with the existing states of nodes in the affected node set. Generative AI large-scale models may need to adjust the original local process sequence, break down a complex operation into multiple finer-grained processing step nodes, merge some operations, or insert new transitional processing step nodes to ensure the continuity of state transitions. The generative AI large-scale model ultimately outputs a subgraph replacement scheme, which includes a series of new processing step node sequences and directed connections between these new processing step nodes. Logically, the subgraph replacement scheme can smoothly connect with the unaffected rest of the dynamic processing graph. In some embodiments, the generative AI large-scale model evaluates a quantifiable function for the internal logical coherence of the subgraph replacement scheme. It can be represented as:

[0034] in: Subgraph replacement scheme The internal logical coherence score, Represents the subgraph replacement scheme The evaluation value of the logical rationality of each processing step node itself. Represents the subgraph replacement scheme The evaluation value for the reasonableness of the dependencies of all directed edges in the equation. and These are the weighting coefficients used to balance the two evaluation values. Generative AI large models optimize this when generating subgraph replacement schemes. The value. The mapping relationship between the user's editing intent and the system's internal representation can be illustrated by a table: Table 1: Examples of mapping between editing intent and structured representation

[0035] In implementation, the system replaces the original logical structure of the dynamic processing graph—from the change of the central node to the affected node set—with a subgraph replacement scheme generated by a generative artificial intelligence model. The replacement operation is a structured graph operation, which may include removing a batch of old processing step nodes and directed edges from the original dynamic processing graph, inserting a batch of new processing step nodes and directed edges into the dynamic processing graph, and re-connecting the boundary nodes of the new subgraph replacement scheme with the nodes in the original dynamic processing graph that were not replaced. After the replacement operation, the dynamic processing graph is logically reconstructed, resulting in a revised processing logic network. This revised processing logic network incorporates the user's professional editing intentions, while the intervention of the generative artificial intelligence model ensures the network's overall logical consistency and technological rationality.

[0036] See Figure 5 This system systematically presents the structural impact and logical coherence evaluation results of editing operations on dynamic processing diagrams. Grouped bar charts clearly compare the degree of disturbance to the original processing network caused by three types of editing operations: modification, addition, and deletion. Two sets of parallel bars respectively display the range of affected nodes and the number of newly inserted nodes, intuitively reflecting the depth and breadth of process transformation caused by different editing types. The line chart tracks the change trajectory of the logical coherence score of the subgraph replacement scheme after editing. This score comprehensively considers two dimensions: the logical rationality of the nodes themselves and the dependence of connection relationships. The line trend directly reflects the performance level of editing operations in maintaining the self-consistency of the processing flow. Through the organic connection of multi-level visual elements, the charts fully demonstrate how the user's editing intent is transformed into a stable and reliable manufacturing logic network through an intelligent reconstruction mechanism, providing an intuitive quantitative reference for process optimization decisions.

[0037] Example 5: In specific implementation, the revised machining logic network includes a complete sequence of machining steps and their dependencies after editing and reconstruction. Logical and resource feasibility verification must be performed before physical machining. A detailed resource requirement list is defined for each machining step node in the revised machining logic network. The resource requirement list explicitly lists the manufacturing resources required to execute the machining step node in structured data form. The resource requirement list typically includes the machine tool motion axes required for the operation, such as the X-axis, Y-axis, Z-axis, and C-axis; the tool number and type used for cutting; the specific positioning and clamping points of the fixture used to fix the workpiece on the machine tool table; and may also include auxiliary resources such as coolant nozzle channels and tailstock center positions. The definition of the resource requirement list is based on the analysis of the operation content of each machining step node. For example, a milling operation requires the X-axis, Y-axis, and Z-axis to be linked, requires the use of a face milling cutter with a specific number, and requires the use of fixture clamping points. The completeness of the resource requirement list is the basis for subsequent conflict detection.

[0038] In practical implementation, the topology information of the modified processing logic network and the detailed information of each processing step node, including the node's resource requirement list, input conditions, and output status, are input into the generative AI large-scale model. The generative AI large-scale model simulates the execution process of the entire modified processing logic network. The simulation does not actually drive the machine tool, but rather logically deduces the execution and state transitions of each node in the network within a virtual, digital manufacturing environment. At the start of the simulation, the generative AI large-scale model initializes a virtual resource status table, recording the current occupancy status of all critical resources. Initially, all resources are marked as "idle." Based on the topological order of the modified processing logic network, the generative AI large-scale model logically activates processing step nodes where all input conditions are met and the required resources are available. Before attempting to activate a processing step node, the generative AI large-scale model performs a resource conflict check. This check queries the resource requirement list of the current processing step node and compares it with the resources currently occupied by upstream activated nodes recorded in the virtual resource status table to check for resource overlap. Resource overlap means that the same resource is requested by multiple processing step nodes at the same time.

[0039] In its implementation, the generative AI large-scale model performs logical conflict checks concurrently with the simulation execution. For each processing step node to be activated, the logical conflict check verifies whether all its input conditions have been fully satisfied by the output states of all its upstream processing step nodes. Upstream processing step nodes refer to all nodes in the network topology that can be reached through directed connections; the output states of these nodes are prerequisites for the current node's execution. During the simulation, the generative AI large-scale model maintains a global set of workpiece and system states. This set records the cumulative effect of the output states generated by all executed nodes up to the current simulation time. When the generative AI large-scale model attempts to activate a processing step node, it compares the node's list of input conditions with the current global state set. If at least one input condition cannot find a corresponding state item in the current global state set that meets the requirements, the input conditions of the current processing step node are determined not to be fully satisfied, and the generative AI large-scale model marks the currently attempted processing step node as a logical conflict node. Logical conflict nodes indicate that, under the revised processing logic network definition, the execution prerequisites for this processing step node have not yet been established. This may be because the upstream node it depends on has not yet been executed, or because the output state of the upstream node does not meet its input requirements. This exposes contradictions in the process sequence or state definition.

[0040] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0041] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A CAM programming algorithm based on a generative artificial intelligence large model, characterized in that, The method comprises: receiving and parsing a computer-aided design document, extracting geometric features, tolerance labels and material attribute information therefrom, and fusing the information based on a generative artificial intelligence large model to generate a structured semantic skeleton; inputting the structured semantic skeleton into a preset process knowledge base for matching and mapping, generating a preliminary processing instruction cluster according to the mapping relationship between the processing strategies and parameters stored in the process knowledge base; expanding the preliminary processing instruction cluster into a series of directed and associated processing step nodes, defining input conditions and output states for each processing step node, and connecting multiple processing step nodes into a dynamic processing graph according to the dependency relationship between the input conditions and output states; receiving a user's editing intention for the dynamic processing graph, the editing intention including modification, addition or deletion instructions for specific processing step nodes, analyzing the editing intention through a generative artificial intelligence large model, and reconstructing the node logic and connection relationship of the dynamic processing graph accordingly to generate a revised processing logic network; before the revised processing logic network is executed, inputting the topology structure and node content of the revised processing logic network into a generative artificial intelligence large model, simulating the network execution process by the generative artificial intelligence large model to identify potential logic conflict nodes and resource conflict nodes therein.

2. The generative artificial intelligence large model based CAM programming algorithm according to claim 1, wherein, The receiving and parsing of the computer-aided design document, extracting geometric features, tolerance labels and material attribute information therefrom, and fusing the information based on a generative artificial intelligence large model to generate a structured semantic skeleton comprises the following steps: using a generative artificial intelligence large model to perform semantic segmentation on the overall description text of the computer-aided design document, and identifying feature description paragraphs related to manufacturing; performing entity and relationship extraction on the feature description paragraphs, identifying key geometric entities, precision constraint entities and material entities, and establishing spatial and logical constraint relationships between entities; reorganizing the identified key geometric entities, precision constraint entities, material entities and spatial and logical constraint relationships between entities according to a preset manufacturing feature framework to form the structured semantic skeleton with manufacturing features as the core and associated constraints and material attributes.

3. The generative artificial intelligence large model based CAM programming algorithm according to claim 2, wherein, The inputting of the structured semantic skeleton into a preset process knowledge base for matching and mapping comprises the following steps: comparing the similarity of each manufacturing feature unit in the structured semantic skeleton with the feature templates in the process knowledge base; selecting a feature template with a similarity exceeding a threshold for each manufacturing feature unit, and performing secondary filtering on the candidate processing strategy set associated with the feature template and the precision constraint attribute and material attribute of the manufacturing feature unit; selecting a processing strategy from the secondary filtered candidate processing strategy set according to strategy priority rules, and binding the specific parameters of the manufacturing feature unit with the selected processing strategy to generate a basic processing instruction for the manufacturing feature unit; collecting the basic processing instructions corresponding to all manufacturing feature units to form the preliminary processing instruction cluster.

4. The generative artificial intelligence large model based CAM programming algorithm according to claim 3, wherein, The secondary filtering of the candidate machining strategy set associated with the feature template from the precision constraint attribute and the material attribute of the manufacturing feature unit comprises the following steps: Obtaining the candidate machining strategy set associated with the feature template from the process knowledge base, reading the applicable precision range and material type of each candidate machining strategy; Extracting the precision constraint attribute of the manufacturing feature unit, including the size tolerance and geometric tolerance requirements, and extracting the material attribute of the manufacturing feature unit, including the material hardness and toughness parameters; Comparing the applicable precision range of each candidate machining strategy with the precision constraint attribute of the manufacturing feature unit, and screening out the candidate machining strategies whose applicable precision range covers the precision constraint requirements of the manufacturing feature unit; Matching the applicable material type of the screened candidate machining strategy with the material attribute of the manufacturing feature unit, eliminating the candidate machining strategies whose material types do not match, and obtaining the secondary filtered candidate machining strategy set.

5. The generative artificial intelligence large model based CAM programming algorithm according to claim 1, wherein, The expansion of the preliminary machining instruction cluster into a series of directed associated machining step nodes, the definition of input conditions and output states for each machining step node, and the connection of multiple machining step nodes into a dynamic machining graph according to the dependency relationship between the input conditions and the output states comprise the following steps: Parsing each basic machining instruction in the preliminary machining instruction cluster, decomposing it into indivisible basic machining operations, and defining each basic machining operation as a machining step node; Analyzing the input conditions of each machining step node, including the blank state, the fixture state, and the tool positioning state; Analyzing the output state of each machining step node, including the geometric state and the surface state of the workpiece reached; Establishing matching rules between the output state and the input condition, and establishing a directed connection between two machining step nodes if the output state of one machining step node can meet the input condition of another machining step node, and finally forming the dynamic machining graph.

6. The generative artificial intelligence large model based CAM programming algorithm according to claim 5, wherein, The parsing of each basic machining instruction in the preliminary machining instruction cluster, the decomposition of it into indivisible basic machining operations, and the definition of each basic machining operation as a machining step node comprise the following steps: Reading each basic machining instruction in the preliminary machining instruction cluster, and identifying the machining action sequence contained in the basic machining instruction; According to the atomicity rule of machining action, decomposing each machining action sequence into the smallest unit of machining operation, each smallest unit of machining operation being indivisible and corresponding to an independent machine tool movement or tool operation; Assigning a unique identifier to each smallest unit of machining operation, and defining the input parameters and output results required by the machining operation; Encapsulating each smallest unit of machining operation as a machining step node, and recording the description information of the input parameters and the output results in the machining step node.

7. The generative artificial intelligence large model based CAM programming algorithm according to claim 1, wherein, The receiving user's editing intention for the dynamic machining map includes modification, addition or deletion instructions for a specific machining step node. The editing intention is analyzed by a generative artificial intelligence large model, and the node logic and connection relationship of the dynamic machining map are reconstructed accordingly, including the following steps: Receiving the user's editing intention, marking the machining step node targeted by the editing intention as a change center node; Starting from the change center node, traversing all upstream and downstream nodes affected by the change center node in the logical path in the dynamic machining map to form an affected node set; Input the change content of the change center node and the affected node set into the generative artificial intelligence large model, and request the generative artificial intelligence large model to generate a logically consistent subgraph replacement scheme; Replace the original logical structure from the change center node to the affected node set in the dynamic machining map with the subgraph replacement scheme to complete the reconstruction and obtain the modified machining logic network.

8. The generative artificial intelligence large model based CAM programming algorithm according to claim 7, wherein, The step of inputting the change content of the change center node and the affected node set into the generative artificial intelligence large model and requesting the generative artificial intelligence large model to generate a logically consistent subgraph replacement scheme includes the following steps: Extract the change content of the change center node, including the modified machining operation description, the added machining step or the deleted node information; Obtain the affected node set from the dynamic machining map, including all upstream and downstream nodes that directly or indirectly depend on the change center node; Input the change content and the structured data of the affected node set into the generative artificial intelligence large model, and request the generative artificial intelligence large model to generate a subgraph replacement scheme based on machining logic constraints and resource constraints; The generative artificial intelligence large model outputs a subgraph replacement scheme, which includes a new node sequence and connection relationship, ensuring the logical coherence and resource consistency of the new subgraph and the original graph.

9. The generative artificial intelligence large model based CAM programming algorithm according to claim 1, wherein, The step of identifying potential logical conflict nodes and resource conflict nodes includes the following steps: Define a resource requirement list for each machining step node in the modified machining logic network, including the required machine tool axis, tool number, and fixture point; During the simulation network execution process of the generative artificial intelligence large model, activate the machining step nodes in logical order and check whether the resource requirement list of the activated node overlaps with the currently occupied resources; If there is a resource overlap, the currently activated node is determined to be a resource conflict node; At the same time, check whether the input conditions of the activated node have been fully met by the output states of all its upstream nodes. If not, the currently activated node is determined to be a logical conflict node.

10. A CAM programming system based on a generative artificial intelligence large model, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor, when executing the computer program, implements the steps of the CAM programming algorithm based on the generative artificial intelligence large model in any one of claims 1 to 9.

Images