Polishing control method and system based on AI large model and knowledge graph

Abstract

The application discloses a polishing control method and system based on an AI large model and a knowledge graph, relates to the field of metal additive manufacturing post-processing, and comprises the following steps: acquiring three-dimensional model data and three-dimensional scanning data of a workpiece to be processed and performing registration, and extracting support structure information to be processed on the surface of the workpiece; inputting the support structure information to be processed into an AI large model, outputting a support type identification result, and matching a removal process and initial process parameters based on a support feature rule base and a removal process matching knowledge graph; inputting the identification result, the initial parameters and a tool state into a process parameter optimizer, outputting dynamic optimization parameters and generating a machining path; performing machining, collecting sensing data to update surface information, and recursively removing residual support; and polishing the removal area until the surface roughness meets preset requirements. The application solves the problems of low automation degree and difficulty in self-adaptive processing of complex support structures in the prior art.

Description

Technical Field

[0001] This invention relates to the field of post-processing in metal additive manufacturing, specifically to a grinding control method and system based on AI large models and knowledge graphs. Background Technology

[0002] Metal additive manufacturing (3D printing) technology, especially laser powder bed fusion (LPBF), is widely used in high-value-added fields such as aerospace and medical implants due to its ability to create complex geometries and lightweight structures. However, in existing technologies, metal additive manufacturing requires the addition of support structures to prevent deformation or collapse when printing parts with overhanging structures, large spans, or complex internal cavities. After printing, these support structures must be precisely removed, and the contact areas must be post-processed to achieve the required dimensional accuracy and surface quality of the parts. The handling of these support structures has become one of the key bottlenecks restricting the large-scale, industrial application of metal additive manufacturing technology.

[0003] In existing technologies, the removal and post-processing of metal support structures still mainly rely on manual operation and traditional machining methods. Among them, manual operation has problems such as low efficiency, poor consistency of processing quality, high labor intensity and safety hazards, making it difficult to meet the needs of mass production; while traditional machining methods are limited by poor tool accessibility, difficulty in precise control of cutting force, and the risk of part damage and tool breakage.

[0004] In recent years, new technologies such as robot-assisted automated processing have emerged in the industry, which have improved operational efficiency to some extent. However, these technologies still struggle to identify complex support structures and can only handle simple supports with regular shapes or preset paths. They lack the ability to identify and process complex support types such as tree-like, grid-like, and lattice-like structures, as well as supports located inside parts or residual supports. They still rely on manual intervention for programming and adjustment, and cannot achieve truly adaptive intelligent processing.

[0005] Therefore, the lack of automation and adaptability in existing technologies when dealing with complex support structures has become a key bottleneck restricting the large-scale and industrial application of metal additive manufacturing technology. Furthermore, for metal additive manufacturing parts requiring high surface quality, the grinding and polishing process after support removal also relies on manual operation or automated equipment with fixed parameters. Existing automated grinding solutions lack the ability to dynamically optimize process parameters such as contact force, abrasive belt grit size, feed speed, and oscillation frequency. During grinding, improper control of contact force can easily lead to over- or under-grinding of the workpiece surface, and the process parameters cannot be adaptively adjusted according to the abrasive belt wear state, making it difficult to guarantee the consistency and surface quality of batch processing. Summary of the Invention

[0006] This application provides a polishing control method and system based on AI large model and knowledge graph to solve the problems in the prior art where the removal of support structure depends on manual operation or traditional mechanical processing, has a low degree of automation, and the existing robot processing schemes have difficulty in identifying complex support structures, cannot dynamically adapt to multiple processes and parameters, and lack closed-loop feedback, resulting in poor processing quality consistency.

[0007] In view of the above problems, the technical solution provided in this application is as follows:

[0008] Firstly, this application provides a refinement control method based on large AI models and knowledge graphs, characterized in that the method includes:

[0009] Acquire the three-dimensional model data and three-dimensional scan data of the workpiece to be processed, register the three-dimensional model data and the three-dimensional scan data, and extract the information of the support structure to be processed on the surface of the workpiece.

[0010] The information of the support structure to be processed is input into the AI ​​big model, and the AI ​​big model outputs the support type identification result. Based on the support feature rule base and the knowledge graph of removal process matching, the corresponding removal process and initial process parameters are matched for each type of support area.

[0011] The support type identification result, the initial process parameters, and the current tool status are input to the process parameter optimizer, which outputs dynamically optimized process parameters and generates the machining path for the current round based on the dynamically optimized process parameters.

[0012] The support removal process is performed according to the processing path, and the sensing data during the processing is collected. The workpiece surface information is updated according to the sensing data, and the residual support is recursively removed until it is determined that there is no support to be processed.

[0013] Secondly, this application provides a polishing control system based on AI large models and knowledge graphs, the system comprising:

[0014] The data acquisition and registration unit is used to acquire the three-dimensional model data and three-dimensional scanning data of the workpiece to be processed, register the three-dimensional model data and the three-dimensional scanning data, and extract the information of the support structure to be processed on the surface of the workpiece.

[0015] The support identification and process matching unit is used to input the support structure information to be processed into the AI ​​big model, output the support type identification result through the AI ​​big model, and match the corresponding removal process and initial process parameters for various support regions based on the support feature rule base and the removal process matching knowledge graph.

[0016] The parameter optimization and path planning unit is used to input the support type identification result, the initial process parameters and the current tool status to the process parameter optimizer, output dynamically optimized process parameters, and generate the machining path for the current round based on the dynamically optimized process parameters.

[0017] The execution and feedback unit is used to perform support removal processing according to the processing path, collect sensing data during the processing, update the workpiece surface information according to the sensing data, recursively remove residual supports, until it is determined that there are no supports to be processed.

[0018] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0019] First, by acquiring the 3D model data and 3D scan data of the workpiece to be processed, registering the two and extracting the information of the support structure to be processed, the boundary area between the support and the workpiece can be accurately located, providing a reliable data foundation for subsequent identification and processing. This solves the problems of existing technologies that are difficult to accurately extract support structure information and are easily affected by occlusion and reflection interference.

[0020] Furthermore, the information of the support structure to be processed is input into the AI ​​large model to output the support type identification result. Based on the support feature rule base and the knowledge graph of removal process matching, the corresponding removal process and initial process parameters are matched, realizing the automatic identification and process adaptation of various complex support types such as sheet-like, tree-like, block-like, and grid-like structures. This solves the problem that the existing technology can only handle rule-based support, has a single process mode, and relies on manual settings.

[0021] Furthermore, the support type identification results, initial process parameters, and current tool status are input into the process parameter optimizer, which outputs dynamically optimized process parameters and generates machining paths accordingly. This enables the model to adaptively adjust key parameters such as rotational speed, feed rate, and contact force based on real-time working conditions such as tool wear and support chatter, thus solving the problem that existing technologies have fixed process parameters and cannot dynamically respond to changes in machining status.

[0022] Finally, the support removal process is performed according to the processing path, and the sensing data during the processing is collected to update the workpiece surface information. The residual support is recursively removed until there is no support to be processed, forming a real-time feedback and closed-loop control mechanism for the processing process. It can dynamically identify residual support and automatically perform multiple rounds of removal, solving the problem that the open-loop control of the existing technology leads to incomplete support removal and difficulty in ensuring the consistency of processing quality.

[0023] In summary, the technical solution of this application realizes intelligent closed-loop control of the entire process of support structure from identification, process matching, parameter optimization to execution feedback, which effectively improves the automation level and processing quality consistency of post-processing in metal additive manufacturing. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the polishing control method based on AI large model and knowledge graph provided in the embodiments of this application.

[0025] Figure 2 This is a schematic diagram of the polishing control system based on AI large model and knowledge graph provided in the embodiments of this application.

[0026] The components represented by each number in the attached diagram are explained as follows: 11 is the data acquisition and registration unit, 12 is the support identification and process matching unit, 13 is the parameter optimization and path planning unit, and 14 is the execution and feedback unit. Detailed Implementation

[0027] This application provides a polishing control method and system based on AI large models and knowledge graphs, specifically addressing the technical problems of existing technologies, such as reliance on manual operation or traditional mechanical processing for support structure removal, low automation levels, difficulty in identifying complex support structures, inability to dynamically adapt to multiple processes and parameters, and poor processing quality consistency due to lack of closed-loop feedback. Specific embodiments are as follows:

[0028] It should be noted that in the following embodiments, the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0029] Example 1, as Figure 1 As shown, this application provides a polishing control method based on AI large models and knowledge graphs, the method including:

[0030] S100: Acquire the three-dimensional model data and three-dimensional scan data of the workpiece to be processed, register the three-dimensional model data and the three-dimensional scan data, and extract the information of the support structure to be processed on the surface of the workpiece.

[0031] In the field of post-processing technology for support structures in metal additive manufacturing, existing technologies typically rely on manual identification or traditional visual inspection methods to locate the support structure. Manual methods require operators to judge the support position based on experience, which is inefficient and prone to omissions. While traditional visual inspection methods can acquire images of the workpiece surface, they struggle to accurately distinguish the support structure from the workpiece itself, especially when the support and workpiece are made of the same material and have similar colors. Visual inspection is easily affected by reflections and occlusions, making it difficult to accurately extract the boundary information between the support and the workpiece. Furthermore, existing technologies lack effective means to compare the design model with the actual workpiece, making it difficult to identify deformations or residual supports generated during the printing process. Therefore, accurately obtaining the position, shape, and connection relationship between the support structure to be processed and the workpiece becomes the primary technical problem to be solved in this step.

[0032] Step S100 in the method provided in this application embodiment includes:

[0033] Obtain the design 3D model data of the workpiece to be processed;

[0034] The actual three-dimensional scanning data of the surface of the workpiece to be processed is acquired using a three-dimensional scanner.

[0035] The three-dimensional model data is registered with the three-dimensional scan data, and the deviation area between the two is calculated.

[0036] The regions within the deviation area that belong to the support structure are taken as the support structure information to be processed. The specific implementation method is as follows.

[0037] In this embodiment, the three-dimensional model data refers to the computer-aided design model of the workpiece to be processed. This model records the theoretical geometry of the workpiece, including complete design information of the part body and supporting structure.

[0038] 3D scan data refers to point cloud or mesh data obtained by scanning the actual printed workpiece with a 3D scanner, which reflects the true geometric shape of the workpiece.

[0039] Registration refers to the process of aligning 3D model data with 3D scan data in a spatial coordinate system. By using registration algorithms such as the iterative nearest point algorithm, the optimal spatial transformation relationship between the two data is calculated, so that the two can achieve accurate matching in the same coordinate system.

[0040] Deviation areas refer to regions where the spatial locations of the 3D model data and 3D scan data differ after registration. These differences may arise from thermal deformation during the printing process, residual support structures, or surface defects on the workpiece.

[0041] In this step, the first step is to acquire the design 3D model data of the workpiece to be processed. This model is usually stored in formats such as STL and STEP, and contains complete geometric information of the workpiece body and supporting structure, serving as a theoretical benchmark.

[0042] Furthermore, a 3D scanner is used to perform a full-range scan of the workpiece to collect actual 3D scan data of the workpiece surface. During the scanning process, the workpiece is fixed on the worktable, and the scanner acquires point cloud data of the workpiece surface from multiple angles. After noise reduction and fusion processing, a complete 3D model of the workpiece surface is formed, reflecting the true geometric shape of the workpiece.

[0043] Furthermore, a registration algorithm is used to spatially align the 3D model data with the 3D scan data. The registration process first uses coarse registration to roughly align the two data sets, and then uses fine registration (such as the iterative nearest point algorithm) to finely adjust the data, minimizing the distance between corresponding points. After registration, the 3D model data and the 3D scan data are in the same coordinate system, allowing for point-by-point comparison.

[0044] Furthermore, the deviation region between the two registered data sets is calculated. This deviation region falls into two categories: one is areas present in the model but not in the actual workpiece, i.e., areas where supports have been removed but the model still shows their presence; the other is areas present in the actual workpiece but not in the model, i.e., areas where supports were not completely removed or where printing residue remains. The areas belonging to the support structure within the deviation region are extracted as support structure information to be processed, used for subsequent identification and processing.

[0045] For example, taking a metal 3D printed part as an example, the 3D model of the part includes the main body and a tree-like support structure, and the model file is stored in STL format. After printing, a structured light 3D scanner is used to scan the workpiece, obtaining 3D scan data containing 2 million point cloud data. The iterative nearest point algorithm is used for registration. After coarse registration and fine registration, the average registration error between the model and the scan data is 0.02mm. After registration, the deviation area is calculated, and a protrusion area with a height of about 3mm is found under a certain overhanging structure. This area has a support design in the model but appears as a residual structure in the scan data. Therefore, this area is extracted as the support structure information to be processed, including its geometry, position coordinates, and connection boundary with the workpiece body.

[0046] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0047] In summary, this step accurately extracts the support structure information by registering and comparing the designed 3D model with the actual 3D scanning data, overcoming the shortcomings of existing technologies that rely solely on visual inspection and cannot distinguish between the support and the workpiece body. The registration technology eliminates the positional deviation caused by printing deformation, ensuring that the extracted support information has an accurate spatial relationship with the workpiece body. The extracted support structure information includes complete geometric features such as shape, position, and boundaries, providing high-quality data input for subsequent AI large-scale model recognition and process matching.

[0048] S200: Input the information of the support structure to be processed into the AI ​​big model, output the support type identification result through the AI ​​big model, and match the corresponding removal process and initial process parameters for each type of support area based on the support feature rule base and the removal process matching knowledge graph.

[0049] After acquiring the information of the support structure to be processed, how to transform this geometric data into executable process decisions becomes the technical problem to be solved in this step. In the existing technology, the identification of support structures mainly relies on human experience or traditional image processing algorithms. Human identification is inefficient and highly subjective, while traditional algorithms are difficult to handle complex geometric shapes such as tree-like and lattice-like structures, and cannot accurately obtain key features such as the connection method and spatial distribution between the support and the workpiece. At the same time, even if the support type is identified, existing technologies lack a unified decision-making mechanism for matching the most suitable removal process and initial parameters for different types of supports. They often use a uniform process to process all supports, resulting in incomplete removal of complex supports or damage to thin-walled structures. Therefore, a method is needed that can automatically identify complex support types and intelligently match the optimal process to lay a solid foundation for subsequent parameter optimization and path planning.

[0050] Step S200 in the method provided in this application embodiment includes:

[0051] The geometric features of each support region in the support structure information to be processed are input into the AI ​​large model. The geometric features include at least the shape features, size features, spatial distribution features and connection method with the workpiece body of the support region.

[0052] The AI ​​model extracts and classifies the geometric features, and outputs the support type identification results for each support region. The support types include at least sheet-like support, tree-like support, block-like support, grid-like support, lattice-like support, and point column support.

[0053] In this embodiment, geometric features refer to the morphological description parameters of the support area in three-dimensional space. Shape features include the aspect ratio, curvature, and contour of the support; dimensional features include the thickness, height, and cross-sectional area of ​​the support; spatial distribution features include the density, arrangement, and distance from the workpiece surface of the support; and the connection method with the workpiece body includes surface contact, line contact, and point contact. These features collectively determine the ease of support removal and the required process.

[0054] AI large-scale models refer to deep learning models pre-trained on massive amounts of data, specifically designed for processing 3D point cloud data. Their training consists of two phases: First, self-supervised pre-training is performed, training the model to perform occlusion prediction tasks on a vast array of general-purpose 3D models, allowing the model to reconstruct randomly occluded areas and master the ability to extract general geometric features such as edges, planes, and curved surfaces. Then, supervised fine-tuning is performed, collecting 3D scan data of actual printed parts, manually labeling the segmentation masks and type labels of the support areas, and learning to classify six types of support structures, including sheet-like and tree-like structures. After training, the model can automatically receive information about the support structures to be processed and accurately output the type labels of each support area, providing a reliable basis for subsequent process matching.

[0055] Support type identification results refer to the labels obtained after classifying each support area. Different types of supports have significant differences in structural strength, removal difficulty, and bonding force with the workpiece, requiring different removal processes.

[0056] In this step, the geometric features of each support region are first extracted from the support structure information to be processed. The extraction process includes: obtaining the boundary contour of the support region through 3D registration results and calculating its shape feature parameters; obtaining the dimensional features of the support such as thickness and height through scanning data; analyzing the distribution density and arrangement pattern of the support on the workpiece surface to obtain spatial distribution features; and determining the connection method based on the contact area and contact angle between the support and the workpiece body.

[0057] Furthermore, the extracted geometric features are input into the AI ​​large-scale model. The AI ​​large-scale model contains multi-layer neural networks that use a self-attention mechanism to capture the relationships between features for feature extraction and classification reasoning. The model has been trained on a large number of labeled samples and can accurately identify various support types such as sheet-like, tree-like, block-like, grid-like, lattice-like, and point-column-like structures.

[0058] Finally, the AI ​​model outputs the support type identification results for each support region, with each support region corresponding to a type label. These labels will serve as key inputs for subsequent process matching.

[0059] For example, taking a 3D-printed metal part as an example, the support structure information to be processed includes three support regions. Region A has the following geometric characteristics: an aspect ratio of 4:1, a thickness of 0.3mm, and is sheet-like, contacting the workpiece surface. Region B has the following geometric characteristics: a main trunk diameter of 2mm, numerous branches, and a tree-like structure, contacting the workpiece at multiple points. Region C has the following geometric characteristics: a mesh density of 5mm × 5mm, a thickness of 0.5mm, and is mesh-like, contacting the workpiece line. Inputting these geometric characteristics into a large AI model, after feature extraction and classification reasoning, the model outputs that Region A is a sheet-like support, Region B is a tree-like support, and Region C is a mesh-like support.

[0060] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation of the present invention.

[0061] Step S200 in the method provided in this application embodiment further includes:

[0062] Construct a support feature rule base, which contains the mapping relationship between different support types and support geometric features;

[0063] Construct a knowledge graph for matching removal processes. In the knowledge graph, nodes represent support types, removal process types, and process parameter types, and edges represent the adaptation relationship between support types and removal processes, and the association relationship between removal processes and process parameters.

[0064] Based on the support type identification result, the geometric features of the current support are determined by traversing the support feature rule base, and then the removal process matching knowledge graph is queried to obtain the corresponding removal process type and initial process parameters.

[0065] In this embodiment, the support feature rule base is a structured database that stores the mapping relationship between different support types and their typical geometric features.

[0066] Removal process matching knowledge graph is a knowledge representation method organized in a graph structure, used to store the association knowledge between support types, removal process types, and process parameter types. Removal process types include at least cutting, milling, grinding, and EDM. Taking grinding as an example, its associated process parameter types include spindle speed, feed rate, contact force, abrasive belt grit size, and oscillation frequency.

[0067] Initial process parameters refer to a set of recommended process parameter values ​​obtained by matching based on the knowledge graph, which serve as the starting point for the subsequent process parameter optimizer, rather than the fixed parameters to be executed at the end.

[0068] In this step, a support feature rule base and a removal process matching knowledge graph are first constructed. During actual operation, based on the support type identification results output by the AI ​​model, the support feature rule base is traversed to obtain the detailed geometric features of the current support, such as thickness and connection method, to supplement the information that the AI ​​model could not refine.

[0069] Then, using the support type and supplementary geometric features as query conditions, the removal process matching knowledge graph is retrieved. The removal process matching knowledge graph stores the optimal removal process type for each support type, as well as the initial process parameters corresponding to that process. By querying the removal process matching knowledge graph, the removal process type and initial process parameters matched for the current support region are output.

[0070] For example, for region A, identified as a sheet-like support, querying the support feature rule base reveals its thickness to be 0.3 mm and its connection method to be surface contact. Retrieving the process matching knowledge graph shows that the sheet-like support is compatible with a cutting process, with initial process parameters of 8000 rpm rotation speed, 1200 mm / min feed rate, and 30 N contact force. For region B, which is a tree-like support, the graph shows it is compatible with a milling process, with initial process parameters of 10000 rpm rotation speed, 800 mm / min feed rate, and 50 N contact force. For region C, which is a mesh-like support, the graph shows it is compatible with a grinding process, with initial process parameters of 6000 rpm rotation speed, 600 mm / min feed rate, and 20 N contact force.

[0071] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0072] Step S200 in the method provided in this application embodiment further includes:

[0073] Collect historical support removal processing data, which includes support type, removal process type, process parameters, and processing results;

[0074] Using support type, removal process type, and process parameter type as nodes, and the adaptation relationship between support type and removal process type, and the association relationship between removal process type and process parameter type as edges, a graph structure of knowledge graph is constructed.

[0075] Based on the removal efficiency, tool wear, and surface damage in historical processing results, the weights of the edges between each node are calculated. The weights are positively correlated with the removal efficiency and negatively correlated with the tool wear and surface damage.

[0076] The constructed graph structure and edge weights are stored in a structured manner to generate a knowledge graph for removing process matching.

[0077] In this embodiment, historical support removal processing data refers to the actual processing records accumulated in previous processing. Each record includes the support type, the removal process used, the actual process parameters used, and the processing result.

[0078] The graph structure of a knowledge graph refers to representing data as nodes and edges. Nodes represent entities, such as support type, process type, and parameter type, while edges represent relationships between entities, such as adaptation relationships and association relationships.

[0079] The edge weight is a numerical value used to quantify the quality of the relationship between nodes. The weight is calculated based on historical processing results: the higher the removal efficiency, the less tool wear, and the less surface damage, the greater the weight; conversely, the less the efficiency, the smaller the weight.

[0080] In this step, a large amount of support removal processing data is first collected from the historical database. Each data record includes the support type, the removal process used, the actual process parameters used, and the processing result.

[0081] Furthermore, a knowledge graph structure is constructed using support type, removal process type, and process parameter type as nodes. Adaptation edges are established between support type nodes and removal process type nodes; association edges are established between removal process type nodes and process parameter type nodes.

[0082] Furthermore, based on the removal efficiency, tool wear, and surface damage levels from historical processing results, the weight of each edge is calculated. The weight calculation formula can use a weighted average method: Weight = α × Normalized removal efficiency + β × (1 - Normalized tool wear) + γ × (1 - Normalized surface damage), where α, β, and γ are preset coefficients to ensure the weight is between 0 and 1. For edges matching support type and removal process, the weight reflects the overall performance of the process on that support type; for edges relating removal process and process parameters, the weight reflects the importance or recommendation level of that parameter for the process.

[0083] Finally, the constructed graph structure and edge weights are stored in a graph database or structured file to generate a process matching removal knowledge graph. This process matching removal knowledge graph can be continuously updated and iterated in subsequent processing. As new data accumulates, the weights can be recalculated, enabling the process matching removal knowledge graph to evolve itself.

[0084] For example, taking a tree-like support structure, historical data records the performance of two processes: milling process with a removal efficiency of 96%, tool wear of 0.03 mm, and surface damage of 0.1 mm; and chiseling process with a removal efficiency of 82%, tool wear of 0.08 mm, and surface damage of 0.3 mm. Weights are assigned based on efficiency first, followed by tool wear, and then surface damage: removal efficiency accounts for 0.5, tool wear for 0.3, and surface damage for 0.2. Then, normalization is performed. For removal efficiency, the maximum value between milling (96%) and chiseling (82%) is 96%. The normalized value for milling is 96% ÷ 96% = 1.00, and the normalized value for chiseling is 82% ÷ 96% = 0.85. Tool wear is converted to a positive score based on the minimum value of 0.03 mm: the score for milling is 1 - (0.03 ÷ 0.08) = 1 - 0.38 = 0.62, and the score for chiseling is 0. Similarly, the normalized value for surface damage can be obtained. Then, the weighted summation is performed: Milling weight = 0.5 × 1.00 + 0.3 × 0.62 + 0.2 × 0.67 = 0.82; Chiseling weight = 0.425. The results show that the milling process has a higher weight, and milling is the preferred method for processing tree-like supports.

[0085] For example, for the grinding process, the knowledge graph stores associated process parameters including contact force, abrasive belt grit size, feed rate, oscillation frequency, and spindle speed. The recommended range for contact force is 5N to 50N, for abrasive belt grit size is P80 to P1000, for feed rate is 300mm / min to 1500mm / min, for oscillation frequency is 1Hz to 5Hz, and for spindle speed is 3000rpm to 12000rpm. The edge weights between each parameter node and the grinding process node are calculated based on historical processing results; higher removal efficiency, lower surface roughness, and less abrasive belt wear result in greater weights.

[0086] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0087] In summary, this step utilizes a large AI model to intelligently identify the geometric features of the support region, accurately distinguishing various complex support types such as sheet-like, tree-like, block-like, grid-like, lattice-like, and point-column-like structures. This addresses the limitations of existing technologies in identifying complex support shapes. By constructing a support feature rule base and a knowledge graph for matching removal processes, the support type is intelligently matched with the removal process and initial process parameters, automating the process from support identification to process decision-making. The knowledge graph is built based on historical processing data and uses weighted quantification of process advantages and disadvantages, ensuring the matching results are data-supported and interpretable. This process forms an identification-matching decision chain, providing precise input for subsequent parameter optimization and path planning, significantly improving the intelligence and adaptability of support removal process planning.

[0088] S300: Input the support type identification result, the initial process parameters and the current tool status to the process parameter optimizer, output the dynamically optimized process parameters, and generate the machining path for the current round based on the dynamically optimized process parameters;

[0089] After identifying the support type and matching the initial process, the technical problem to be solved in this step is how to dynamically optimize the process parameters based on the actual processing conditions and generate a safe and efficient processing path. In existing technologies, process parameters are usually preset and fixed by engineers based on experience, and cannot be adjusted according to real-time conditions such as tool wear, support chatter, and uneven material removal. This leads to low processing efficiency, severe tool wear, and even tool breakage or part damage. Furthermore, in complex scenarios involving multiple support areas and alternating tools, existing technologies lack autonomous process scheduling and path planning capabilities, making it difficult to comprehensively consider factors such as geometric reachability, process sequence constraints, and tool switching costs. They often rely on manual programming, which is inefficient and prone to path interference. Therefore, a method is needed that can dynamically optimize process parameters and autonomously plan processes and paths to achieve efficient, safe, and adaptive support removal processing.

[0090] Step S300 in the method provided in this application embodiment includes:

[0091] A reinforcement learning environment is constructed. The state space of the reinforcement learning environment includes the geometric features of the supporting region, the current tool state, and historical machining results. The action space includes the adjustment amount of spindle speed, feed rate, and contact force. The reward function is positively correlated with the removal efficiency and negatively correlated with the tool wear and surface damage degree.

[0092] Collect historical processing data to initialize the reinforcement learning environment;

[0093] A reinforcement learning algorithm is used to train the process parameter optimization model, with the goal of maximizing the cumulative reward.

[0094] Training stops when the accumulated reward converges to a preset threshold, resulting in a fully trained process parameter optimization model.

[0095] In this embodiment, the reinforcement learning environment is a simulation or modeling environment used to train a process parameter optimization model. The environment defines the state space, action space, and reward function, enabling the model to learn the optimal policy through interaction with the environment.

[0096] The reward function is a feedback signal during reinforcement learning training, used to evaluate the quality of the model's actions. In this embodiment, the reward function is positively correlated with removal efficiency and negatively correlated with tool wear and surface damage.

[0097] Historical processing data refers to data accumulated during past processing, including state information under different working conditions, actions taken, and corresponding reward values. This data is used to initialize the reinforcement learning environment, enabling the model to possess certain prior knowledge at the start of training.

[0098] Reinforcement learning algorithms are algorithms used to train and optimize models. The training objective is to maximize the cumulative reward, that is, to maximize the total reward obtained by the model in long-term interactions.

[0099] In this step, the reinforcement learning environment is first constructed. The state space of the environment is designed as a vector containing the geometric features of the supporting region, the current tool state, and historical machining results; the action space is designed as the speed adjustment, feed rate adjustment, and contact force adjustment; in addition, for the grinding process, the action space also includes the belt grit adjustment, oscillation frequency adjustment, and constant contact force control parameters. The reward function is defined as: Reward = 0.5 × Normalized removal efficiency + 0.3 × (1 - Normalized tool wear) + 0.2 × (1 - Normalized surface damage).

[0100] Furthermore, historical machining data is collected, including hundreds of machining records under different support types and tooling conditions. Each record contains the state, action, and actual reward value. This data is used to initialize the reinforcement learning environment, enabling it to simulate real machining responses.

[0101] Furthermore, a Deep Q-Network (DQN) algorithm is used to train the process parameter optimization model. The model takes the current state as input and outputs the Q-value of each action, selecting the action with the largest Q-value to execute. During training, the model continuously interacts with the environment, collecting data on state, action, reward, and next state, storing it in an experience replay pool, and randomly sampling to update the network parameters.

[0102] When the cumulative reward curve stabilizes and no longer rises significantly, the model is considered to have converged. At this point, training is stopped, the model parameters are saved, and the trained process parameter optimizer is obtained.

[0103] For example, taking a tree-like support structure as an example, a reinforcement learning environment is constructed. The state space vector is [support thickness 2mm, tool wear 0.1mm, previous removal efficiency 92%]; the action space consists of rotational speed adjustment [-200, -100, 0, +100, +200] rpm, feed rate adjustment [-50, 0, +50] mm / min, and contact force adjustment [-2, 0, +2] N, for a total of 5 × 3 × 3 = 45 discrete action combinations. In the reward function, the removal efficiency weight is 0.5, the tool wear weight is 0.3, and the surface damage weight is 0.2. 1000 sets of historical data are collected for environment initialization, and a deep Q-network algorithm is used to train for 50,000 rounds. The cumulative reward increases from the initial -50 to +320 and then converges, resulting in the optimized model after training.

[0104] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0105] Step S300 in the method provided in this application embodiment further includes:

[0106] Based on the dynamically optimized process parameters, the processing sequence of the support area to be removed in the current round is determined;

[0107] A multi-process intelligent scheduling algorithm based on graph neural networks and attention mechanism is adopted to generate a globally optimized process sequence by comprehensively considering geometric reachability constraints, process sequence constraints, tool switching costs and resource occupation conflicts.

[0108] Obtain the kinematic model of the robotic arm, which includes the degree of freedom parameters, range of motion, and mapping relationship between joint angles and end effector pose of each joint of the robotic arm;

[0109] For each position to be processed in the process sequence, the target joint angles of each joint of the robotic arm are obtained by inverse kinematics solution through the kinematic model based on the coordinates and orientation of the position to be processed.

[0110] For two adjacent positions to be processed in the process sequence, with their respective target joint angles as the starting and ending points, a collision-free transition path that satisfies the obstacle avoidance constraint is searched in the configuration space of the robotic arm.

[0111] The collision-free transition paths between the target joint angles corresponding to all positions to be processed and the adjacent positions are spliced ​​together according to the process sequence to generate a complete dynamic processing path for the robotic arm.

[0112] In this embodiment, geometric reachability constraint refers to whether the end effector of the robotic arm can reach the target position under the constraints of the workpiece geometry, including whether it collides with the workpiece or support, and whether it is within the workspace of the robotic arm.

[0113] Attention mechanisms are a technique that focuses on important information by calculating weights. In this algorithm, the attention mechanism calculates the dependency weights between nodes based on their embedding representations; the larger the dependency weight, the more likely the node should be processed.

[0114] Process sequence constraints refer to the sequential requirements between different support areas due to process characteristics, such as removing large external supports first and then processing small internal residues, or rough machining followed by fine machining.

[0115] Tool switching cost refers to the time cost or loss of precision required to change to different tools (such as cutting tools, milling tools, grinding tools).

[0116] A process sequence refers to a sequence formed by arranging the processing tasks of multiple areas to be processed in the optimal order. This sequence determines the order in which the robotic arm performs the tasks.

[0117] The kinematic model of a robotic arm describes the mathematical relationship between the joint angles of the robotic arm and the pose of the end effector. Forward kinematics calculates the end effector pose based on the joint angles, while inverse kinematics calculates the joint angles based on the end effector pose.

[0118] Inverse kinematics refers to determining the joint angles that each joint of a robotic arm should reach, given the target position and orientation of the end effector in space. For robotic arms with redundant degrees of freedom, there may be multiple solutions, and the optimal solution must be selected based on obstacle avoidance, joint constraints, and other conditions.

[0119] Configuration space refers to the mathematical space formed by all possible joint angles of a robotic arm, where each point represents a pose of the robotic arm. Searching for a collision-free path in configuration space allows one to avoid obstacles.

[0120] In this step, the processing sequence of the support areas to be removed in the current round is first determined based on the dynamically optimized process parameters. For example, large supports are removed first, followed by the processing of small residues.

[0121] Furthermore, a multi-process intelligent scheduling algorithm based on Graph Neural Network (GNN) and attention mechanism is employed to generate a globally optimized process sequence. This algorithm constructs a process scheduling graph by treating each support region to be processed as a node and the process sequence constraints and tool switching relationships between regions as edges. Features are extracted from the scheduling graph using the GNN to obtain the embedded representation of each node. The dependency weights between nodes are calculated using the attention mechanism, and the processing priority of each node is determined based on these weights. The process sequence is then output in descending order of priority. This algorithm comprehensively considers geometric reachability, process sequence constraints, tool switching costs, and resource conflicts to achieve intelligent scheduling of processes.

[0122] Furthermore, the kinematic model of the robotic arm is obtained. The robotic arm is usually a 6-axis or 7-axis articulated robot, and the kinematic model includes the degree of freedom parameters of each joint, the range of motion, and the mapping relationship between the joint angles and the pose of the end effector.

[0123] Furthermore, for each position to be processed in the process sequence, inverse kinematics is performed using a kinematic model based on the coordinates and orientation of that position to obtain the target joint angles of each joint of the robotic arm. If multiple solutions exist, the solution closest to the current joint angle is selected to reduce the travel distance.

[0124] Furthermore, for two adjacent processing positions in the process sequence, using their respective target joint angles as the start and end points, a collision-free transition path satisfying obstacle avoidance constraints is searched in the configuration space of the robotic arm. Path search can employ algorithms such as Rapidly-exploring Random Tree (RRT) and Probabilistic Roadmap Method (PRM) to ensure that the robotic arm does not collide with the workpiece, support, or environment during its movement.

[0125] Finally, the collision-free transition paths between the target joint angles corresponding to all the positions to be processed and the adjacent positions are spliced ​​together according to the process sequence to generate a complete dynamic processing path for the robotic arm.

[0126] For example, taking a certain part as an example, after optimizing the process parameters, the processing sequence is determined to be from region A to B and then to C. A graph neural network scheduling algorithm is used to construct a process scheduling graph with 3 nodes. After feature extraction and attention weight calculation, the process sequence A to B and then to C is output. A 6-axis robotic arm kinematic model is obtained, and the inverse kinematics of the target poses in regions A and B is solved to obtain the corresponding target joint angles. The Regression-Regression (RRT) algorithm is used to search for a collision-free transition path from A to B in the configuration space. All target joint angles and transition paths are concatenated to generate a complete processing path.

[0127] In addition, for the polishing process, path planning also needs to consider the following characteristics:

[0128] Firstly, constant force control of contact force: a constant contact pressure should be maintained between the grinding tool and the workpiece surface. During path planning, the tool posture needs to be adjusted in real time according to the normal of the workpiece surface to keep the angle between the tool axis and the normal within a preset range, such as no more than 5 degrees. Secondly, grinding path generation strategy: for large-area support residue areas, a spiral or reciprocating grinding path is used to avoid texture defects caused by grinding in a single direction. For small residue areas, spot grinding or short-stroke reciprocating paths are used. Thirdly, normal matching between the tool and the workpiece surface: through the posture adjustment of the robotic arm, the working surface of the grinding tool is always in contact with the curvature of the workpiece surface.

[0129] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0130] Step S300 in the method provided in this application embodiment further includes:

[0131] Each support area to be processed is treated as a node, and the process sequence constraints and tool switching relationships between support areas are treated as edges to construct a process scheduling graph.

[0132] The process scheduling graph is subjected to feature extraction using a graph neural network to obtain the embedded representation of each node;

[0133] The dependency weights between nodes are calculated using an attention mechanism, and the processing priority of each node is determined based on these dependency weights.

[0134] Output the globally optimized process sequence in descending order of processing priority.

[0135] In this embodiment, the process scheduling diagram is a graph structure data used to represent the constraint relationships between support areas to be processed.

[0136] In this step, each support area to be processed is first abstracted as a node. Edges are then established based on process sequence constraints and tool switching relationships to construct a process scheduling graph. Process sequence constraints can be represented by directed edges, such as A to B, which means A must be processed before B. Tool switching relationships can be represented by undirected edges, with added switching cost weights.

[0137] Furthermore, the process scheduling graph is input into a graph neural network. Through multiple convolutional operations, each node aggregates features from its neighbors to update its own embedding representation. After multiple iterations, the embedding representation of each node integrates its own geometric features, process features, and constraint relationships with other nodes.

[0138] Furthermore, the embeddings of each node are input to the attention mechanism, and the processing priority weights for each node are calculated. The attention mechanism can be a single-layer fully connected network that takes node embeddings as input and outputs a scalar weight, which is then passed through a softmax function (softmax( ) , , The priority score is obtained by normalizing the score of the corresponding labeled node (where n represents the total number of node scores, i represents the current node number, and j represents the score of a node from 1 to n).

[0139] Finally, the nodes are sorted from highest to lowest priority score, and the process sequence is output. This sequence comprehensively considers geometric reachability, process sequence constraints, and tool switching costs, achieving global optimization.

[0140] For example, consider four support regions A, B, C, and D to be processed. Process constraints require A to be processed before B, and C before D. In terms of tool configuration, A and C use milling tools, while B and D use grinding tools; frequent switching is time-consuming. After constructing a process scheduling graph, a graph neural network extracts the features of each node, and an attention mechanism calculates the priority, ultimately generating a process sequence of A, C, B, and D. This sequence first focuses on processing A and C using the same tool, then switches tools to process B and D, avoiding multiple switching and achieving global optimization.

[0141] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0142] In summary, this step constructs a reinforcement learning-based process parameter optimizer, enabling the model to dynamically optimize key parameters such as spindle speed, feed rate, and contact force based on support geometry features, tool status, and historical machining results. This addresses the problem of fixed process parameters and inability to respond to real-time changes in operating conditions in existing technologies. Through a multi-process intelligent scheduling algorithm based on graph neural networks and attention mechanisms, considering factors such as geometric reachability, process sequence constraints, and tool switching costs, autonomous planning and global optimization of the process sequence are achieved. Furthermore, by using a robotic arm kinematics model and configuration space path search, a collision-free complete machining path is generated, overcoming the low efficiency and error-prone nature of traditional manual programming. This step achieves full automation from parameter optimization to path generation, significantly improving the machining efficiency, safety, and adaptability of support removal.

[0143] S400: Perform support removal processing according to the processing path, collect sensing data during the processing, update the workpiece surface information according to the sensing data, recursively remove residual supports, until it is determined that there are no supports to be processed.

[0144] After generating the machining path and starting execution, ensuring the safety and thoroughness of the machining process becomes the technical problem to be solved in this step. In existing technologies, support removal typically employs an open-loop control mode of one-time planning and one-time execution. During machining, it is impossible to perceive tool status and workpiece surface changes in real time. When abnormalities such as tool wear, support chatter, or uneven material removal occur, process parameters cannot be adjusted in time, easily leading to part damage or tool breakage risks. Furthermore, due to the lack of real-time monitoring of machining results, residual supports often remain after each round of machining, requiring manual inspection and reprogramming for secondary machining, resulting in low efficiency and difficulty in ensuring consistent quality. Therefore, a closed-loop control method with real-time sensing, dynamic adjustment, and recursive iteration capabilities is needed to achieve complete support removal.

[0145] Step S400 in the method provided in this application embodiment further includes:

[0146] During the support removal process, visual images, force signals and acoustic emission signals of the processing area are collected in real time as sensing data.

[0147] The surface condition of the processed area is identified based on the visual image to determine whether there is any residual support.

[0148] Based on the force and acoustic emission signals, determine whether the current removal process is in a normal state. When an abnormal signal is detected, pause the process and trigger the re-optimization of process parameters.

[0149] When it is determined from the visual image that there is no remaining support in the current round, the current round of processing ends, and the entire workpiece is scanned in three dimensions. If there are unprocessed support areas, the next round is executed recursively.

[0150] In this embodiment, the sensing data refers to the processing information collected in real time by various sensors, including visual images, force signals, and acoustic emission signals. For the grinding process, visual images can also be used to identify the surface roughness and texture uniformity after grinding, and the surface roughness evaluation value of the grinding area is calculated by image texture analysis algorithm; force signals are used to determine whether the contact force is constant. When the fluctuation amplitude of the force signal exceeds a preset threshold, such as ±3N, the contact force is re-optimized; acoustic emission signals are used to monitor the wear status of the sanding belt. When the characteristic frequency of the acoustic emission signal shifts, it prompts to replace the sanding belt or adjust the contact force.

[0151] Residual support refers to the support material that is not completely removed and remains attached to the surface of the workpiece after a round of processing. It usually manifests as small protrusions, burrs, or residual roots.

[0152] An abnormal signal refers to a force or acoustic emission signal that exceeds a preset threshold range.

[0153] Process parameter re-optimization refers to the process where, after an anomaly is detected, the model inputs the current operating condition into the process parameter optimizer to regenerate optimized parameters that are suitable for the current operating condition, rather than simply pausing processing and waiting for manual intervention.

[0154] Recursive execution refers to the cyclical execution of the complete process of scanning, identification, matching, optimization, processing, and sensing until there are no supports left to be processed on the workpiece surface. Each round dynamically adjusts the strategy based on the processing results of the previous round, achieving the layer-by-layer removal of supports.

[0155] In this step, while the robotic arm performs the support removal process, the model simultaneously activates the perception module. A 3D camera acquires image sequences of the processing area at a fixed frequency to obtain visual information during the support removal process; a force sensor is installed at the end of the robotic arm or the tool gripping area to record the changes in contact force in real time; an acoustic emission sensor is attached to the tool or workpiece surface to capture high-frequency signals generated during processing. These three types of perception data are recorded synchronously with timestamps, forming a complete perception dataset of the processing process.

[0156] Furthermore, the model inputs the acquired visual images into the visual recognition model in real time. After training, the model can identify the surface state of the processed area, determine the presence, location, and extent of residual supports. Simultaneously, the model monitors force and acoustic emission signals in real time, comparing them to preset normal operating thresholds. If the signals are within the threshold range, processing continues; if the signals exceed the threshold, the model immediately pauses the robotic arm's movement, records the time, location, and signal characteristics of the anomaly, uses the current state as new input, calls the process parameter optimizer to recalculate the optimized parameters, generates an adjusted processing path, and then continues execution.

[0157] Finally, after the current round of processing is completed, the model first confirms through visual imaging that there are no residual supports within the processed area, thus ending the current round of processing. Subsequently, the model activates a 3D scanner to scan the entire workpiece, acquiring full-surface 3D data, and re-registers it with the design model to determine if any unprocessed support areas exist. If the registration result shows the existence of new support areas to be processed, the model automatically enters the next round, repeating the entire process from support structure extraction to processing execution until the registration result shows no unprocessed supports, at which point the process ends.

[0158] For example, during the first round of support removal processing on a part, the force signal jumped from the normal 40N to 85N and the acoustic emission signal surged from 0.3V to 1.2V during the processing of the tree-like supports. The model determined this to be abnormal, paused processing, and triggered parameter re-optimization before resuming. After the first round, visual confirmation showed no residue on the surface, but a full 3D scan revealed two untouched tree-like supports within the internal cavity. The model automatically entered a second round of targeted processing. After completion, a second scan confirmed no residue, and the process terminated.

[0159] Furthermore, taking an aero-engine blade as an example, the surface of the blade after printing has sheet-like support residue and printing texture. Through 3D scanning and model registration, 12 areas to be polished were extracted. The AI ​​large model identified the support types as sheet-like support residue and thin-walled texture. Consulting the knowledge graph, sheet-like support matched a coarse polishing process with initial parameters of 15N contact force and P240 abrasive grain size; thin-walled texture matched a fine polishing process with initial parameters of 8N contact force and P400 abrasive grain size. The process parameter optimizer dynamically adjusted the contact force to 18N and 10N respectively based on the abrasive belt wear state. Path planning generated a spiral polishing path with a total length of approximately 8.5 meters. During execution, the force signal remained stable, and the visually recognized surface roughness decreased from Ra6.3μm to Ra0.8μm. After the first round of polishing, two tiny residues were detected. The system recursively executed a second round of fine polishing, ultimately achieving a surface roughness of Ra0.4μm, meeting aerospace standards.

[0160] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0161] In summary, this step constructs a multimodal perception system by acquiring visual images, force signals, and acoustic emission signals, enabling real-time monitoring of the processing. Visual recognition determines the state of residual supports, while force and acoustic emission signals detect process anomalies and trigger parameter re-optimization, forming a closed-loop control of the processing. Through a recursive iterative mechanism, a comprehensive 3D scan is performed after each processing cycle to automatically identify newly exposed support areas and execute the next cycle of processing until all supports are completely removed. This step solves the problems of incomplete support removal due to open-loop control in existing technologies, lack of real-time feedback leading to untimely response to anomalies, and reliance on manual inspection and rework resulting in low efficiency, achieving intelligent closed-loop control for support removal.

[0162] In summary, the embodiments of this application, through the complete process of steps S100 to S400, achieve the following effects: By registering the 3D model with the scanned data, the information of the support structure to be processed is accurately extracted, solving the problem of inaccurate support boundary positioning; by identifying the support type through an AI large model and combining it with knowledge graph matching to remove process and initial parameters, automatic identification and process adaptation of complex supports are achieved; by constructing a process parameter optimizer through reinforcement learning to dynamically adjust parameters such as rotational speed, feed rate, and contact force, and by combining graph neural networks and attention mechanisms to achieve intelligent scheduling of multiple processes and collision-free path planning, the problems of fixed parameters and path dependence on manual intervention are solved; by monitoring the processing process in real time through multimodal perception data and combining iterative mechanisms to remove residual supports round by round, a closed-loop control is formed. Overall, this solves the problems of insufficient support identification capability, single process mode, inability to dynamically adjust parameters, and lack of closed-loop feedback in existing technologies, achieving intelligent adaptive control of the entire process from identification to execution.

[0163] Example 2, as Figure 2 As shown, this embodiment provides a polishing control system based on an AI large model and knowledge graph. The system includes a data acquisition and registration unit, a support identification and process matching unit, a parameter optimization and path planning unit, and an execution and feedback unit. The system includes:

[0164] The data acquisition and registration unit 11 is used to acquire the three-dimensional model data and three-dimensional scanning data of the workpiece to be processed, register the three-dimensional model data and the three-dimensional scanning data, and extract the information of the support structure to be processed on the surface of the workpiece.

[0165] The support identification and process matching unit 12 is used to input the support structure information to be processed into the AI ​​big model, output the support type identification result through the AI ​​big model, and match the corresponding removal process and initial process parameters for various support areas based on the support feature rule base and the removal process matching knowledge graph.

[0166] The parameter optimization and path planning unit 13 is used to input the support type identification result, the initial process parameters and the current tool status to the process parameter optimizer, output the dynamically optimized process parameters, and generate the machining path for the current round based on the dynamically optimized process parameters.

[0167] The execution and feedback unit 14 is used to perform support removal processing according to the processing path, collect sensing data during the processing, update the workpiece surface information according to the sensing data, recursively remove residual supports, until it is determined that there are no supports to be processed.

[0168] The data acquisition and registration unit 11 is specifically used for:

[0169] Obtain the design 3D model data of the workpiece to be processed;

[0170] The actual three-dimensional scanning data of the surface of the workpiece to be processed is acquired using a three-dimensional scanner.

[0171] The three-dimensional model data is registered with the three-dimensional scan data, and the deviation area between the two is calculated.

[0172] The regions within the deviation area that belong to the support structure are taken as support structure information to be processed.

[0173] The support identification and process matching unit 12 is specifically used for:

[0174] The geometric features of each support region in the support structure information to be processed are input into the AI ​​large model. The geometric features include at least the shape features, size features, spatial distribution features and connection method with the workpiece body of the support region.

[0175] The AI ​​model extracts and classifies the geometric features, and outputs the support type identification results for each support region. The support types include at least sheet-like support, tree-like support, block-like support, grid-like support, lattice-like support, and point column support.

[0176] The support identification and process matching unit 12 is also used for:

[0177] Construct a support feature rule base, which contains the mapping relationship between different support types and support geometric features;

[0178] Construct a knowledge graph for matching removal processes. In the knowledge graph, nodes represent support types, removal process types, and process parameter types, and edges represent the adaptation relationship between support types and removal processes, and the association relationship between removal processes and process parameters.

[0179] Based on the support type identification result, the geometric features of the current support are determined by traversing the support feature rule base, and then the removal process matching knowledge graph is queried to obtain the corresponding removal process type and initial process parameters.

[0180] The removal process matching knowledge graph is pre-built through the following steps:

[0181] Collect historical support removal processing data, which includes support type, removal process type, process parameters, and processing results;

[0182] Using support type, removal process type, and process parameter type as nodes, and the adaptation relationship between support type and removal process type, and the association relationship between removal process type and process parameter type as edges, a graph structure of knowledge graph is constructed.

[0183] Based on the removal efficiency, tool wear, and surface damage in historical processing results, the weights of the edges between each node are calculated. The weights are positively correlated with the removal efficiency and negatively correlated with the tool wear and surface damage.

[0184] The constructed graph structure and edge weights are stored in a structured manner to generate a knowledge graph for removing process matching.

[0185] Parameter optimization and path planning unit 13 is specifically used for:

[0186] A reinforcement learning environment is constructed. The state space of the reinforcement learning environment includes the geometric features of the supporting region, the current tool state, and historical machining results. The action space includes the adjustment amount of spindle speed, feed rate, and contact force. The reward function is positively correlated with the removal efficiency and negatively correlated with the tool wear and surface damage degree.

[0187] Collect historical processing data to initialize the reinforcement learning environment;

[0188] A reinforcement learning algorithm is used to train the process parameter optimization model, with the goal of maximizing the cumulative reward.

[0189] Training stops when the accumulated reward converges to a preset threshold, resulting in a fully trained process parameter optimization model.

[0190] The parameter optimization and path planning unit 13 is also used for:

[0191] Based on the dynamically optimized process parameters, the processing sequence of the support area to be removed in the current round is determined;

[0192] A multi-process intelligent scheduling algorithm based on graph neural networks and attention mechanism is adopted to generate a globally optimized process sequence by comprehensively considering geometric reachability constraints, process sequence constraints, tool switching costs and resource occupation conflicts.

[0193] Obtain the kinematic model of the robotic arm, which includes the degree of freedom parameters, range of motion, and mapping relationship between joint angles and end effector pose of each joint of the robotic arm;

[0194] For each position to be processed in the process sequence, the target joint angles of each joint of the robotic arm are obtained by inverse kinematics solution through the kinematic model based on the coordinates and orientation of the position to be processed.

[0195] For two adjacent positions to be processed in the process sequence, with their respective target joint angles as the starting and ending points, a collision-free transition path that satisfies the obstacle avoidance constraint is searched in the configuration space of the robotic arm.

[0196] The collision-free transition paths between the target joint angles corresponding to all positions to be processed and the adjacent positions are spliced ​​together according to the process sequence to generate a complete dynamic processing path for the robotic arm.

[0197] The multi-process intelligent scheduling algorithm based on graph neural networks and attention mechanisms generates a globally optimized process sequence through the following steps:

[0198] Each support area to be processed is treated as a node, and the process sequence constraints and tool switching relationships between support areas are treated as edges to construct a process scheduling graph.

[0199] The process scheduling graph is subjected to feature extraction using a graph neural network to obtain the embedded representation of each node;

[0200] The dependency weights between nodes are calculated using an attention mechanism, and the processing priority of each node is determined based on these dependency weights.

[0201] Output the globally optimized process sequence in descending order of processing priority.

[0202] The execution and feedback unit 14 is specifically used for:

[0203] During the support removal process, visual images, force signals and acoustic emission signals of the processing area are collected in real time as sensing data.

[0204] The surface condition of the processed area is identified based on the visual image to determine whether there is any residual support.

[0205] Based on the force and acoustic emission signals, determine whether the current removal process is in a normal state. When an abnormal signal is detected, pause the process and trigger the re-optimization of process parameters.

[0206] When it is determined from the visual image that there is no remaining support in the current round, the current round of processing ends, and the entire workpiece is scanned in three dimensions. If there are unprocessed support areas, the next round is executed recursively.

[0207] Through the collaborative work of the aforementioned units, the grinding control system based on AI large models and knowledge graphs provided in this application can be applied to metal additive manufacturing scenarios such as complex aerospace components, medical implants, and automotive parts, achieving automated removal of support structures. It can be integrated into robotic support removal workstations or intelligent post-processing production lines, effectively improving the automation level and quality consistency of additive manufacturing post-processing. For the specific workflow and optimization details of this system, please refer to Example 1.

[0208] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0209] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0210] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A refinement control method based on AI large models and knowledge graphs, characterized in that, The method includes: Acquire the three-dimensional model data and three-dimensional scan data of the workpiece to be processed, register the three-dimensional model data and the three-dimensional scan data, and extract the information of the support structure to be processed on the surface of the workpiece. The information of the support structure to be processed is input into the AI ​​big model, and the AI ​​big model outputs the support type identification result. Based on the support feature rule base and the knowledge graph of removal process matching, the corresponding removal process and initial process parameters are matched for each type of support area. The support type identification result, the initial process parameters, and the current tool status are input to the process parameter optimizer, which outputs dynamically optimized process parameters and generates the machining path for the current round based on the dynamically optimized process parameters. The support removal process is performed according to the processing path, and the sensing data during the processing is collected. The workpiece surface information is updated according to the sensing data, and the residual support is recursively removed until it is determined that there is no support to be processed.

2. The polishing control method based on AI large model and knowledge graph according to claim 1, characterized in that, Acquire the 3D model data and 3D scan data of the workpiece to be processed, register the 3D model data and the 3D scan data, and extract the information of the support structure to be processed on the surface of the workpiece, including: Obtain the design 3D model data of the workpiece to be processed; The actual three-dimensional scanning data of the surface of the workpiece to be processed is acquired using a three-dimensional scanner. The three-dimensional model data is registered with the three-dimensional scan data, and the deviation area between the two is calculated. The regions within the deviation area that belong to the support structure are taken as support structure information to be processed.

3. The polishing control method based on AI large model and knowledge graph according to claim 1, characterized in that, The information of the support structure to be processed is input into the AI ​​model, and the AI ​​model outputs the support type identification result, including: The geometric features of each support region in the support structure information to be processed are input into the AI ​​large model. The geometric features include at least the shape features, size features, spatial distribution features and connection method with the workpiece body of the support region. The AI ​​model extracts and classifies the geometric features, and outputs the support type identification results for each support region. The support types include at least sheet-like support, tree-like support, block-like support, grid-like support, lattice-like support, and point column support.

4. The polishing control method based on AI large model and knowledge graph according to claim 1, characterized in that, Based on the supporting feature rule base and the knowledge graph of removal process matching, corresponding removal processes and initial process parameters are matched for various supporting regions, including: Construct a support feature rule base, which contains the mapping relationship between different support types and support geometric features; Construct a knowledge graph for matching removal processes. In the knowledge graph, nodes represent support types, removal process types, and process parameter types, and edges represent the adaptation relationship between support types and removal processes, and the association relationship between removal processes and process parameters. Based on the support type identification result, the geometric features of the current support are determined by traversing the support feature rule base, and then the removal process matching knowledge graph is queried to obtain the corresponding removal process type and initial process parameters.

5. The polishing control method based on AI large model and knowledge graph according to claim 4, characterized in that, The removal process matching knowledge graph is pre-constructed through the following steps: Collect historical support removal processing data, which includes support type, removal process type, process parameters, and processing results; Using support type, removal process type, and process parameter type as nodes, and the adaptation relationship between support type and removal process type, and the association relationship between removal process type and process parameter type as edges, a knowledge graph graph structure is constructed. Based on the removal efficiency, tool wear, and surface damage in historical processing results, the weights of the edges between each node are calculated. The weights are positively correlated with the removal efficiency and negatively correlated with the tool wear and surface damage. The constructed graph structure and edge weights are stored in a structured manner to generate a knowledge graph for removing process matching.

6. The polishing control method based on AI large model and knowledge graph according to claim 1, characterized in that, The process parameter optimizer is pre-built through the following steps: A reinforcement learning environment is constructed. The state space of the reinforcement learning environment includes the geometric features of the supporting region, the current tool state, and historical machining results. The action space includes the adjustment amount of spindle speed, feed rate, and contact force. The reward function is positively correlated with the removal efficiency and negatively correlated with the tool wear and surface damage degree. Collect historical processing data to initialize the reinforcement learning environment; A reinforcement learning algorithm is used to train the process parameter optimization model, with the goal of maximizing the cumulative reward. Training stops when the accumulated reward converges to a preset threshold, resulting in a fully trained process parameter optimization model.

7. The polishing control method based on AI large model and knowledge graph according to claim 1, characterized in that, The processing path for the current round is generated based on the dynamically optimized process parameters, including: Based on the dynamically optimized process parameters, the processing sequence of the support area to be removed in the current round is determined; A multi-process intelligent scheduling algorithm based on graph neural networks and attention mechanisms is adopted to generate a globally optimized process sequence by comprehensively considering geometric reachability constraints, process sequence constraints, tool switching costs and resource occupation conflicts. Obtain the kinematic model of the robotic arm, which includes the degree of freedom parameters, range of motion, and mapping relationship between joint angles and end effector pose of each joint of the robotic arm; For each position to be processed in the process sequence, the target joint angles of each joint of the robotic arm are obtained by inverse kinematics solution through the kinematic model based on the coordinates and orientation of the position to be processed. For two adjacent positions to be processed in the process sequence, with their respective target joint angles as the starting and ending points, a collision-free transition path that satisfies the obstacle avoidance constraint is searched in the configuration space of the robotic arm. The collision-free transition paths between the target joint angles corresponding to all positions to be processed and the adjacent positions are spliced ​​together according to the process sequence to generate a complete dynamic processing path for the robotic arm.

8. The polishing control method based on AI large model and knowledge graph according to claim 7, characterized in that, The multi-process intelligent scheduling algorithm based on graph neural networks and attention mechanisms generates a globally optimized process sequence through the following steps: Each support area to be processed is treated as a node, and the process sequence constraints and tool switching relationships between support areas are treated as edges to construct a process scheduling diagram. The process scheduling graph is subjected to feature extraction using a graph neural network to obtain the embedded representation of each node; The dependency weights between nodes are calculated using an attention mechanism, and the processing priority of each node is determined based on these dependency weights. Output the globally optimized process sequence in descending order of processing priority.

9. The polishing control method based on AI large model and knowledge graph according to claim 1, characterized in that, Collect sensing data during the processing, update the workpiece surface information based on the sensing data, recursively remove residual supports until it is determined that there are no more supports to be processed, including: During the support removal process, visual images, force signals and acoustic emission signals of the processing area are collected in real time as sensing data. The surface condition of the processed area is identified based on the visual image to determine whether there is any residual support. Based on the force and acoustic emission signals, determine whether the current removal process is in a normal state. When an abnormal signal is detected, pause the process and trigger the re-optimization of process parameters. When it is determined from the visual image that there is no remaining support in the current round, the current round of processing ends, and the entire workpiece is scanned in three dimensions. If there are unprocessed support areas, the next round is executed recursively.

10. A polishing control system based on AI large-scale models and knowledge graphs, characterized in that: The system is used to implement the polishing control method based on AI large model and knowledge graph as described in any one of claims 1-9, the system comprising: The data acquisition and registration unit is used to acquire the three-dimensional model data and three-dimensional scanning data of the workpiece to be processed, register the three-dimensional model data and the three-dimensional scanning data, and extract the information of the support structure to be processed on the surface of the workpiece. The support identification and process matching unit is used to input the support structure information to be processed into the AI ​​big model, output the support type identification result through the AI ​​big model, and match the corresponding removal process and initial process parameters for various support regions based on the support feature rule base and the removal process matching knowledge graph. The parameter optimization and path planning unit is used to input the support type identification result, the initial process parameters and the current tool status to the process parameter optimizer, output dynamically optimized process parameters, and generate the machining path for the current round based on the dynamically optimized process parameters. The execution and feedback unit is used to perform support removal processing according to the processing path, collect sensing data during the processing, update the workpiece surface information according to the sensing data, recursively remove residual supports, until it is determined that there are no supports to be processed.

Images