Method for generating a body clamp clamping unit based on multi-agent reinforcement learning
By employing a closed-loop feedback mechanism of multi-agent reinforcement learning and digital twin verification, the problems of static library dependence and manual intervention in the design of existing automotive body fixtures are solved. This enables the automated generation and adaptive optimization of the clamping unit of the body fixture, improving design efficiency and robustness.
Patent Information
- Application Number
- CN202511910439.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-17
- Publication Date
- 2026-06-23
- Estimated Expiration
- 2045-12-17
AI Technical Summary
Existing automotive body fixture designs rely on static libraries and manual intervention, lacking dynamic relationship reasoning and adaptive assembly, resulting in long design cycles and low efficiency.
By employing a multi-agent reinforcement learning approach, standardized input data is generated, and inference is performed using graph neural networks and rule engines. Combined with digital twin verification and self-healing optimization, the automatic generation and iterative optimization of the vehicle body clamping unit are achieved.
It achieves rapid adaptive design of the body clamping unit, shortens the design cycle, improves design efficiency, has self-learning and self-healing capabilities, and is suitable for flexible production of multiple vehicle models.
Smart Images

Figure CN121351641B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent automotive design, and more specifically, to a method for generating body clamping units based on multi-agent reinforcement learning. Background Technology
[0002] Currently, the design of automotive body-in-white (BIW) fixtures mainly relies on manual parametric modeling and empirical rule optimization using CAD software (such as CATIA). Engineers select components (such as locating pins, grippers, and cylinders) from a static parts library based on the 3-2-1 positioning principle, and assemble them into units by manually adjusting their dimensions, layout, and clamping positions.
[0003] A typical workflow includes: drawing 2D sketches and converting them into 3D models; using finite element analysis (FEA) to simulate clamping force distribution and deformation; and using robot simulation software to verify collisions and robot paths. Flexible clamping systems allow manual adjustment of the positioning unit along the T-slot, but require manual intervention. Some advanced technologies integrate sensors and IoT to monitor clamping forces, but still rely heavily on trial and error, with design cycles ranging from hours to days. These solutions are based on static libraries and lack dynamic relationship reasoning and adaptive assembly mechanisms.
[0004] There is currently no effective solution to the above problems. Summary of the Invention
[0005] This invention provides a method for generating vehicle body clamping units based on multi-agent reinforcement learning, which at least solves the technical problem that the design of existing clamping units is based on a static library and lacks dynamic relational reasoning, making adaptive assembly impossible.
[0006] According to one aspect of the present invention, a method for generating a vehicle body clamping unit based on multi-agent reinforcement learning is provided, comprising: generating standardized input data for the vehicle body clamping unit, wherein the standardized input data includes RPS (Reference Point System) grouping prediction results, RPS distance-to-edge parameters, and structural features of vehicle body parts; based on the standardized input data, performing inference on the standardized input data through a graph neural network or a rule-based inference engine to generate a parameter constraint subgraph containing part nodes, constraint relationships, and assembly logic; based on the parameter constraint subgraph, optimizing the parameters of the vehicle body clamping unit through a multi-agent reinforcement learning model to obtain an optimized parameter vector, and generating the vehicle body clamping unit based on the optimized parameter vector.
[0007] According to another aspect of the present invention, a vehicle body clamping unit generation device based on multi-agent reinforcement learning is also provided, comprising: a data acquisition module configured to generate standardized input data for the vehicle body clamping unit, wherein the standardized input data includes RPS group prediction results, RPS distance-k edge distance parameters, and structural features of vehicle body parts; an inference module configured to infer the standardized input data based on the standardized input data using a graph neural network or a rule-based inference engine to generate a parameter constraint subgraph containing part nodes, constraint relationships, and assembly logic; and a generation module configured to optimize the parameters of the vehicle body clamping unit based on the parameter constraint subgraph using a multi-agent reinforcement learning model to obtain an optimized parameter vector, and generate the vehicle body clamping unit based on the optimized parameter vector.
[0008] In this embodiment of the invention, standardized input data for a vehicle body clamping unit is generated. This standardized input data includes RPS grouping prediction results, RPS distance-to-edge parameters, and structural features of the vehicle body parts. Based on this standardized input data, a graph neural network or a rule-based inference engine is used to infer the standardized input data, generating a parameter constraint subgraph containing part nodes, constraint relationships, and assembly logic. Based on this parameter constraint subgraph, a multi-agent reinforcement learning model is used to optimize the parameters of the vehicle body clamping unit, obtaining an optimized parameter vector. The vehicle body clamping unit is then generated based on this optimized parameter vector. This solution addresses the technical problem that existing clamping units are designed based on static libraries and lack dynamic relationship inference, hindering adaptive assembly. Attached Figure Description
[0009] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:
[0010] Figure 1 This is a flowchart of an optional method for generating a vehicle body clamping unit based on multi-agent reinforcement learning according to an embodiment of the present invention;
[0011] Figure 2 This is a flowchart of another optional method for generating a vehicle body clamping unit based on multi-agent reinforcement learning according to an embodiment of the present invention;
[0012] Figure 3 This is a flowchart of an optional method for generating standardized input data according to an embodiment of the present invention;
[0013] Figure 4 This is a flowchart of an optional method for generating a parameter constraint subgraph according to an embodiment of the present invention;
[0014] Figure 5 This is a flowchart of an optional generative multi-agent reinforcement learning method according to an embodiment of the present invention;
[0015] Figure 6 This is a schematic diagram of an optional multi-agent reinforcement learning-based vehicle body clamping unit generation device according to an embodiment of the present invention.
[0016] Figure 7 A schematic diagram of the structure of a computer device suitable for implementing embodiments of the present disclosure is shown. Detailed Implementation
[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.
[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0019] According to an embodiment of the present invention, a method embodiment for generating a vehicle body clamping unit based on multi-agent reinforcement learning is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.
[0020] Figure 1 This is a method for generating vehicle body clamping units based on multi-agent reinforcement learning according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:
[0021] Step S102: Generate standardized input data for the vehicle body clamping unit, wherein the standardized input data includes RPS grouping prediction results, RPS distance-to-edge parameters, and structural features of the vehicle body parts.
[0022] For example, the functional labels output by the preceding RPS model are parsed to obtain the RPS grouping prediction results and the RPS distance margin parameters; a curvature adaptive sampling algorithm is used to dynamically adjust the sampling point density, and the structural features of the vehicle body are extracted from the 3D model of the vehicle body based on the adjusted sampling point density. The preceding RPS model is a pre-trained AI recognition model based on the 3D data of the vehicle body, used to automatically detect and semantically label features (such as holes and surfaces) in the vehicle body structure that may serve as positioning references, and output functional labels containing type, grouping, and constraint relationships for these features.
[0023] Step S104: Based on the standardized input data, reason about the standardized input data through a graph neural network or a rule-based inference engine to generate a parameter constraint subgraph containing part nodes, constraint relationships and assembly logic.
[0024] For example, the standardized input data is parsed to generate query conditions for the knowledge graph, and a query is performed in the knowledge graph based on the query conditions to obtain candidate parts and their associated constraint information; the graph neural network or the rule-based reasoning engine is used to perform embedding calculation and path reasoning on the candidate parts and their associated constraint information to obtain embedding similarity, and the compatibility relationship, mechanical relationship and assembly matching relationship between the candidate parts are determined based on the embedding similarity; based on the compatibility relationship, mechanical relationship and assembly matching relationship, a parameter constraint subgraph containing part nodes, constraint relationships and assembly logic is generated.
[0025] Step S106: Based on the parameter constraint subgraph, the parameters of the vehicle body clamping unit are optimized using a multi-agent reinforcement learning model to obtain an optimized parameter vector, and the vehicle body clamping unit is generated based on the optimized parameter vector.
[0026] For example, based on the parameter constraint subgraph, a part agent model, an assembly agent model, and an optimization agent model are established as the multi-agent reinforcement learning model, and the initial state data for reinforcement learning is obtained using the multi-agent reinforcement learning model. Based on the initial state data, the part agent model optimizes the part parameters of the body fixture clamping unit, the assembly agent optimizes the assembly relationship of the body fixture clamping unit, and the optimization agent optimizes the overall performance of the body fixture clamping unit, thereby obtaining intermediate optimization data reflecting the part parameters, the assembly relationship, and the overall performance. The intermediate optimization data is then fused to generate the optimization parameter vector characterizing the structural configuration and performance indicators of the body fixture clamping unit.
[0027] After obtaining the optimized parameter vector of the vehicle body clamping unit, the method further includes iteratively executing the following until the parameters of the vehicle body clamping unit reach a preset condition:
[0028] First, based on the optimized parameter vector, a digital twin simulation model is used to verify the mechanical stiffness, spatial interference, and assembly feasibility of the vehicle body clamping unit, obtaining verification result data. For example, based on the optimized parameter vector, a digital twin simulation model corresponding to the vehicle body clamping unit is established, and simulation input data to be verified is generated using the digital twin simulation model. Based on the simulation input data, mechanical performance analysis and spatial relationship calculation are performed on the vehicle body clamping unit to obtain the stress, displacement, and interference detection results of the vehicle body clamping unit under stress, which are used as the verification result data.
[0029] Next, based on the verification result data, the parameters of the vehicle body clamping unit that do not meet the preset conditions are self-healed and corrected to generate the corrected optimized parameter vector. Finally, the vehicle body clamping unit is generated based on the optimized parameter vector.
[0030] This application also provides another intelligent generation method for body fixture clamping units. Using the prediction results of the Reference Point System (RPS) grouping and the structural features of the body parts as input, it achieves automated generation and iterative optimization of body fixture clamping units through a closed-loop mechanism of Knowledge Graph (KG), Multi-Agent Reinforcement Learning (MARL), Digital Twin (DT) verification, and self-healing optimization. This application covers a complete closed-loop process from input (RPS labels + geometric features), KG query and reasoning, MARL parameter generation, DT verification feedback, self-healing iteration, to output. Compared with existing technologies, it can quickly generate body fixture clamping units and is suitable for flexible production of multiple vehicle models.
[0031] Specifically, such as Figure 2As shown, the method includes the following steps:
[0032] Step S202, input data.
[0033] The process involves acquiring RPS grouping prediction results, RPS point distance-to-edge parameters, and structural features of the vehicle body (such as local curvature information, normal vectors, surface types, boundary and proximity features, material stiffness, etc.) to generate standardized input data. Specifically, the method for generating standardized input data is as follows: Figure 3 As shown, it includes the following steps:
[0034] Step S2022: Parse the RPS tag.
[0035] Receive the JSON format function tags (e.g., {"type": "double_support_clamp", "confidence": 0.96, "rps_points": [{"coord": [x1,y1,z1], "normal":[nx1,ny1,nz1]}]}) from the preceding RPS model output, and parse out the function type, RPS point coordinates, and normals, which are used as the RPS grouping prediction results and RPS distance margin parameters, respectively.
[0036] Step S2024: Extract geometric features.
[0037] The 3D model of the vehicle body (e.g., CAD model or 3D scan data) is processed to generate a standardized point cloud feature set. This step includes three stages: data preprocessing, curvature calculation and sampling density control, and point cloud feature generation, as detailed below:
[0038] 1) Data preprocessing.
[0039] First, the input 3D model is preprocessed. The preprocessing includes noise removal, voxel downsampling, and normal vector calculation. In the noise removal stage, a statistical outlier filtering algorithm is used to remove outliers, with a noise threshold set to σ=2. In the downsampling stage, voxel filtering is used to control the point cloud density, with the voxel size set to 0.5mm. In the normal vector calculation stage, principal component analysis (PCA) is used to solve for the normal vector of each sampled point based on the distribution of neighboring points.
[0040] 2) Calculate curvature and control sampling density.
[0041] Curvature analysis is performed on the point cloud data, and the sampling density is adaptively adjusted based on the local curvature. The curvature value reflects the geometric complexity of the model surface and is calculated from the rate of change of the normal vector. In this embodiment, the sampling density can be dynamically adjusted using the following formula:
[0042]
[0043] Here, base_density represents the base sampling density, and curvature is the local curvature value. When a high curvature region is detected (such as a hole, edge transition, or corner), the sampling density is increased to twice the base density to enhance the ability to capture geometric details.
[0044] 3) Generate point cloud features.
[0045] After completing curvature adaptive sampling, a standardized set of point cloud features is generated:
[0046]
[0047] in, Represents the i-th point cloud feature. Represents a three-dimensional real space, where N is the total number of sampling points. Each feature point contains attribute information such as three-dimensional coordinates, normal vector, local curvature, and neighborhood density.
[0048] Step S2026, data standardization processing.
[0049] Integrate attribute information such as RPS labels and geometric features into JSON (e.g., {"geometric_features":{"hole_coords":[[x1,y1,z1]],"interference_zones":[[min_x,max_x]]}}) to obtain standardized input data and ensure input consistency.
[0050] Step S204: Generate a parameter constraint subgraph.
[0051] Specifically, the method for generating parameter-constrained subgraphs is as follows: Figure 4 As shown, it includes the following steps:
[0052] Step S2042: Construct a knowledge graph.
[0053] First, structured semantic information is extracted from the historical design dataset to construct a knowledge graph. Specifically, based on over 1000 historical fixture design data points, semantic relationships between parts, interfaces, and process constraints are extracted and organized in the form of RDF triples, for example:<Rest_Unit, composes_with {force_min:1500N, tolerance:0.05mm}, Clamp_Arm> The above triplet indicates that the support unit Rest_Unit can be combined with the clamping arm Clamp_Arm to form a clamping unit. The clamping arm satisfies two constraints: 1) it needs to be able to provide a minimum clamping force force_min of at least 1500 Newtons; 2) the part positioning accuracy is within the tolerance range of 0.05 mm.
[0054] Next, a fine-tuned BERT model (Fine-tuned BERT, F1 > 0.92) is used to identify and classify semantic entities, including part names (e.g., Rest_Unit, Clamp_Arm), interface parameters (coordinates [x, y, z], direction [nx, ny, nz]), and process constraints (e.g., force > 1500N, indicating a clamping force greater than 1500 Newtons). The resulting data is stored in the Neo4j graph database through semantic mapping. In this embodiment, the Neo4j graph database has more than 1000 nodes and more than 5000 relation edges, and supports incremental update operations (update cycle less than 1 hour).
[0055] Step S2044, graph neural network query.
[0056] The standardized input data is parsed, and dynamic path reasoning is performed based on a graph neural network (GNN). This embodiment uses a three-layer GraphSAGE variant model with a node embedding dimension of 128. The semantic embedding vector of the nodes is calculated through neighborhood aggregation and feature propagation mechanisms to identify potential compatibility relationships and constraint matching degrees between parts.
[0057] For example, a semantic search can be performed using the following query path: MATCH(n:Part)-[r:composes_with]->(m:Unit) WHERE n.function='double_support_clamp'. This semantic search means: find all parts whose function is defined as 'double_support_clamp', and identify all units that these parts are connected to. For implicit torque matching relationships between nodes, a matching score is calculated based on cosine similarity.
[0058]
[0059] in, ( ) represents cosine similarity. ( ) represents the feature embedding function. Indicates the clamped unit, This refers to the clamp arm. When the matching score meets the conditions, the two parts are considered compatible in terms of torque or assembly direction.
[0060] Compared to traditional static rule base methods (such as if-then matching), graph neural network query mechanisms can dynamically capture implicit design relationships and semantic similarities, reduce the risk of missing parts, improve design compatibility, effectively avoid assembly interference problems caused by hard-coded rules, and support flexible adaptive queries for multiple vehicle models.
[0061] Step S2046: Output the parametric constraint subgraph.
[0062] After completing GNN inference and path matching, a parameterized constraint subgraph is generated and output in a standardized JSON format. The JSON output includes a parts list, quantities, and corresponding constraint information. For example, {"parts": [{"type": "Rest_Unit", "count":2, "constraints": {"spacing":200mm}}]} means "install two support units. And, the center points of these two support units are precisely arranged 200 mm apart."
[0063] Step S206: Generate multi-agent reinforcement learning.
[0064] Generative Multi-Agent Reinforcement Learning (MARL) Figure 5 As shown, it includes the following steps:
[0065] Step S2062, Input and Initialization.
[0066] First, the parametric constraint subgraph is transformed into the initial state of MARL. Specifically, the part nodes, assembly constraints, and system-level performance indicators in the parametric constraint subgraph are mapped to part agents, assembly agents, and global optimization agents, and the MARL state vector matrix is constructed:
[0067] S = [param_parts; constraints_assembly; metrics_global].
[0068] Here, `param_parts` represents the initial parameters of the parts, including values such as part length, installation height, and normal angle; `constraints_assembly` represents the assembly constraint parameters, including contact surface normals, position tolerances, and alignment requirements; and `metrics_global` represents the global performance metrics, including overall stiffness, stability metrics, and mechanical equilibrium constraints. The total dimension of the state space does not exceed 50.
[0069] To diversify the initialization of the parameter space, a diffusion model was used for random sampling, with a noise step size set to 50, resulting in a set of initial parameters {θ}. k}, where k = 1…10, represents generating 10 sets of initial candidate parameters.
[0070] Step S2064: Design the intelligent agent.
[0071] Based on the initialization, action space, observation space and reward function are designed for the component agent, assembly agent and global optimization agent respectively.
[0072] The component agent is used to adjust the local parameters of a single component. Its action space varies within the range of Δlength ∈ [-5, +5] mm, and its reward function is:
[0073] r_parts = -0.3‖param-constraint‖ + 0.2 × compat_score
[0074] Wherein, param represents the current design parameters of the part, including the current length, installation position or its orientation angle, etc.; constraint represents the process constraints and design range provided by the knowledge graph subgraph, including part tolerances, mounting surface requirements or mechanical constraints, etc.; ‖param-constraint‖ represents the magnitude of the deviation between the current design and the process requirements; compat_score is the part compatibility score obtained based on the embedding similarity of the graph neural network (GNN), with a value range of 0 to 1.
[0075] An assembly agent is used to optimize the fit between multiple parts. Its action space is Δx∈[-10, +10]mm, representing the allowable adjustment range of the assembly coordinates. Its reward function is:
[0076]
[0077] Here, align_3-2-1 represents the assembly alignment calculated based on the 3-2-1 positioning principle. The larger the value, the higher the positioning stability. penalty_dist represents the spatial deviation distance detected during the assembly process. When the deviation exceeds 5mm, a significant penalty will be imposed. The reciprocal of penalty_dist is used to enhance the effect of stronger penalty for larger deviations.
[0078] The optimization agent is used to coordinate the global parameters among the agents so that the local optimization results can meet the global performance goals. The agents exchange information with each other in the form of JSON messages, such as {"part_id":1,"scale_force":1.1}, which means that the locally optimal clamping force of part number 1 needs to be uniformly and proportionally adjusted to 1.1 times the original value, that is, increased by 10%.
[0079] The following code example illustrates the process of generating actions for an intelligent agent:
[0080] for agent in [parts, assembly, optimization]:
[0081] obs = GNN.encode(local_subgraph)
[0082] action = actor_network(obs) # Action generation
[0083] if agent == 'assembly':
[0084] reward=0.4 x alignment_check(action) - 1 / dist_penalty(action)
[0085] Compared with existing single-agent reinforcement learning methods (such as the RL-FD framework that only performs layout optimization), this method introduces a multi-agent cooperation mechanism and combines it with the part relationship embedding provided by GNN to achieve efficient solution of multi-scale non-convex optimization problems (including complex coupling problems such as vibration-force coupling), reducing the number of training convergence steps from the traditional 1000 steps to less than 500 steps.
[0086] Step S2066, Global Reward and Algorithm Update.
[0087] To ensure that the optimization results simultaneously satisfy mechanical performance, assembly non-interference, and cost constraints, a global reward function is constructed:
[0088] R=1.0×(KK target-5.0 × interference vol + 10.0 × cost penalty
[0089] Where K represents the global stiffness of the current clamping unit calculated in the digital twin environment, in N / m; K target This represents the preset target stiffness, which is set to a value in this embodiment. N / m; interference vol The cost represents the spatial interference volume between parts, detected by Boolean operations, and is expressed in cm³. penalty This represents cost-related optimization terms, used to reflect the impact of factors such as the number of parts, material costs, and assembly complexity on the overall cost.
[0090] The training phase employs the CTDE (Centered Training, Distributed Execution) framework and uses an improved Q-learning algorithm for value updates, with the following iterative formula:
[0091] Q(s,a)←Q(s,a)+0.1×[R+0.95×maxQ(s′,a′)-Q(s,a)]
[0092] Here, Q() represents the action-value function, 0.1 is the learning rate, and 0.95 is the discount factor used to balance the weights between the current reward and the future reward. s represents the state, a represents the action, s′ represents the next state, and a′ represents the next action. To improve the model's convergence speed, a Priority Experience Replay (PER) mechanism is used to learn key samples first, and an epsilon-greedy strategy is employed for exploration, with the exploration rate ε gradually decreasing from 1.0 to 0.05 during training.
[0093] Step S2068: Output the optimized parameter vector.
[0094] Output the optimization parameter vector as a JSON string, for example, {"rest_units": [{"pos_x": -100, "height":10}, {"pos_x":100, "height":10}]}, indicating that two support units need to be deployed. The first support unit is installed at X-coordinate -100, with a height of 10 units. The second support unit is installed at X-coordinate 100, also with a height of 10 units.
[0095] Step S208, digital twin verification.
[0096] The optimized parameter vectors generated by MARL are validated to evaluate the degree to which the clamping unit meets the requirements in terms of mechanical performance, spatial interference, and assembly feasibility. The validation process uses the parameter vector JSON output by MARL as input and performs multiphysics solutions in a digital twin simulation model.
[0097] In the mechanical analysis, SymPy is used to solve the static equilibrium equation ∑F = 0 and the vibration modal equations to obtain the structural response under typical loading conditions. The vibration modal displacements need to be controlled below 0.1 mm to ensure structural stability. In stress analysis, the Von Mises equivalent stress is calculated, and its expression is:
[0098] .
[0099] Where, σ vm Von Mises equivalent stress, σ x σ y Let σx and yy be the normal stresses of the structure in the x and y directions, respectively. The omitted terms in the formula represent combinations of shear stress and other principal stresses. Verification requires σy. vm The pressure should not exceed 200 MPa. In spatial interference inspection, Boolean operations are used to find the intersection of the geometric models between parts to obtain the interference volume. If the interference volume is less than 1 cm³, the spatial interference is considered acceptable.
[0100] The output of the digital twin verification is a JSON result containing stress deviation, displacement, and interference volume. For example, {"stress_dev": +20, "displacement": 0.03, "interference_vol": 0.5} indicates that the maximum stress stress_dev is 20% higher than the material's safe allowable stress, the maximum deformation is 0.03 units, a spatial conflict occurred between the fixture and surrounding components, and the estimated interference volume is 0.5 cubic units. Finally, a pass or fail flag is attached.
[0101] Compared to traditional mesh-based finite element simulation methods (such as ANSYS, which typically have high computational complexity), this method employs parametric solution and symbolic computation techniques, significantly reducing overall computational complexity. In typical scenarios, the time required for a single verification run is drastically reduced, and computational efficiency is significantly improved.
[0102] Step S210, self-healing optimization.
[0103] If the digital twin verification results do not meet the preset engineering conditions (such as mechanical stiffness being lower than the target value, spatial interference volume exceeding the threshold, or assembly displacement exceeding the allowable range), the self-healing optimization mechanism will be activated.
[0104] Specifically, the verification result data is first analyzed to extract the deviation vector, including the stiffness deviation ΔK=K. target -K, interference volume V interf and the maximum displacement exceeding the limit Δδ=max(0,δ-δ) limit Subsequently, the bias vector is transformed into a negative reward signal for reinforcement learning and injected into the training loop of the multi-agent reinforcement learning model. Wherein, K... target K is the target stiffness, K is the current stiffness, and δ is the current displacement. limit To limit displacement.
[0105] The multi-agent reinforcement learning model is based on the centralized training-distributed execution (CTDE) framework. It uses the updated reward function to fine-tune the policy networks of the part agent, assembly agent, and optimization agent online. Each agent coordinates to adjust the part parameters, assembly relationships, and system performance configuration based on the reconstructed global and local rewards, and generates a corrected optimized parameter vector.
[0106] This self-healing optimization process constitutes a closed-loop iterative mechanism, executing a maximum of three rounds, with each round taking no more than 30 seconds, until the digital twin verification result meets all preset conditions. This achieves end-to-end autonomous evolution capabilities from design generation, simulation verification, intelligent repair, and re-optimization, significantly improving the system's robustness and engineering applicability.
[0107] Step S212: Output the STEP file.
[0108] After the self-healing cycle is completed and the final optimized parameters are obtained, the parameter vector is converted into a standard STEP file suitable for downstream processes. The system uses the OpenCascade library to generate corresponding 3D geometric models from the geometric parameters in the JSON and outputs a parameterized STEP file, exploded structural diagram, and bill of materials (BOM).
[0109] To ensure manufacturability and assembly accuracy, the system performs a precision check after generating the STEP file. Furthermore, the generated 3D model can be used for augmented reality (AR) overlay verification to enhance visualization.
[0110] Unlike manual export relying on CAD software, this method can automatically generate complete manufacturing files without manually calling intermediate toolchains, shortening the overall design cycle and significantly improving the efficiency of cloud deployment and large-scale collaborative design.
[0111] The embodiments of this application have the following beneficial effects: 1) Automated unit generation. A complete closed loop from RPS grouping prediction to the final STEP file avoids manual intervention. 2) Self-learning and self-healing capabilities. Through closed-loop feedback of MARL and digital twin verification, the system can automatically correct design defects and improve robustness. 3) Combination of knowledge-driven and data-driven approaches. Knowledge graphs ensure constraint compliance, while MARL provides exploratory optimization, taking into account both empirical rules and innovative designs. 4) Significantly improved design efficiency. Unit design, which traditionally takes hours to complete manually, can be completed in minutes with this application's solution. 5) Standardized and parametric output. Directly generating STEP files facilitates integration into existing fixture design processes.
[0112] This application also provides a method for generating a vehicle body clamping unit. The difference from the previous embodiments is that this embodiment uses a rule-based inference engine (e.g., a hybrid rule-embedded dual-channel inference framework) to generate the parameter constraint subgraph. The remaining processing steps are the same as in the previous embodiments and will not be repeated here.
[0113] The parameter constraint subgraph generation method provided in this embodiment will be described in detail below, which includes the following steps:
[0114] 1) Filter candidate parts.
[0115] The candidate part information is input into the rule-based inference engine, which performs hard constraint filtering on the candidate parts through predefined semantic templates and process rules. The semantic templates include functional type matching, interface coordinate range, minimum / maximum mechanical constraints, and installation direction consistency. Any candidate parts that do not meet the hard constraints are removed in subsequent calculations.
[0116] 2) Perform embedded computation and multi-hop path reasoning.
[0117] For the subset of candidate parts output by the rule engine, a local subgraph is extracted in the knowledge graph, and the local subgraph is input into the embedding module (using GraphSAGE) to calculate the node embedding vector.
[0118] To capture implicit relationships and multi-hop dependencies, a multi-hop path weighted scoring function P is introduced. score This is used to evaluate the overall correlation between any two component nodes transmitted via a hop path from 1 to H:
[0119]
[0120] in the formula Represents a node With nodes Multi-hop path weighted scoring Indicates the maximum number of hops. Indicates all lengths of A set of simple paths For the path Similarity or trust score, For decay weights related to path length (e.g.) ,and ), For path. Used as a node identifier for parts.
[0121] 3) The part matching index, i.e., embedding similarity, is obtained by fusing the node embedding vector and the path correlation.
[0122] Based on node embedding vectors Path scoring Define a composite index of embedding similarity and path coupling for part pairs. :
[0123]
[0124] in For the cosine similarity function, the parameters are... The weighted average of similarity and path score can be adjusted online via digital twin feedback; the meanings of the parameters in the formula are as follows: and They are nodes and Embedded vector, Controlling the weights of the two parts, the denominator This is used to normalize the path scores to the (0,1) interval for fusion.
[0125] 4) Calculate the compatibility score for the engineering dimension to obtain the multidimensional composite compatibility score.
[0126] Based on this, a compatibility composite score is defined to reflect engineering constraints from different dimensions. Its score is determined by mechanical matching. Geometric Combination Score Process compliance score The result obtained through adjustable weighted fusion is:
[0127]
[0128] Among them, the mechanical matching score The geometric fit score is calculated from the end moments, stiffness ratio, and safety margin predicted by empirical formulas or pre-trained regression models. The score is given by integrating the part's geometric mapping and the error distribution function of the mating surfaces; process compliance score. The Boolean compliance items output by the rules engine are mapped to numerical scores. For normalized weights and It is used to adjust the design focus in different projects.
[0129] 5) Obtain the set of compatible parts relationships based on composite indicators and multidimensional composite compatibility scores.
[0130] Ultimately and The following formulas are used in combination for criterion decision-making:
[0131]
[0132] in The Sigmoid function is used to map the criterion to the (0,1) interval. This is a proportionality coefficient used to control the relative contributions of S and C. To determine the threshold, if (For example ), then the part is identified , compatible.
[0133] 6) Construct a parameter constraint subgraph based on the set of compatible parts relationships.
[0134] Based on the above criteria, candidate parts are screened in layers according to compatibility, mechanics and assembly matching relationships, generating parameter constraint subgraphs containing information such as part nodes, edges (with edge types labeled as "mechanical constraints", "geometric fit", "process dependence" etc.) and assembly logic (including assembly sequence and interface pairs).
[0135] This application also provides a vehicle body clamping unit generation device based on multi-agent reinforcement learning, such as... Figure 6 As shown, the system includes: a data acquisition module 62, configured to generate standardized input data for the body fixture clamping unit, wherein the standardized input data includes RPS grouping prediction results, RPS distance-edge parameters, and structural features of the body parts; an inference module 64, configured to infer from the standardized input data using a graph neural network or a rule-based inference engine to generate a parameter constraint subgraph containing part nodes, constraint relationships, and assembly logic; and a generation module 66, configured to optimize the parameters of the body fixture clamping unit using a multi-agent reinforcement learning model based on the parameter constraint subgraph to obtain an optimized parameter vector, and generate the body fixture clamping unit based on the optimized parameter vector.
[0136] It should be noted that the vehicle body clamping unit generation device provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the vehicle body clamping unit generation device and the vehicle body clamping unit generation method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.
[0137] Figure 7 A schematic diagram of a computer device suitable for implementing embodiments of the present disclosure is shown. It should be noted that... Figure 7 The computer device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments disclosed herein.
[0138] like Figure 7 As shown, the computer device includes a central processing unit (CPU) 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage section 1008 into a random access memory (RAM) 1003. The RAM 1003 also stores various programs and data required for system operation. The CPU 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.
[0139] The following components are connected to I / O interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to I / O interface 1005 as needed. A removable medium 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 1010 as needed so that computer programs read from it can be installed into storage section 1008 as needed.
[0140] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.
Claims
1. A method for generating clamping units for vehicle body fixtures based on multi-agent reinforcement learning, characterized in that, include: Standardized input data for generating the vehicle body clamping unit is provided, wherein the standardized input data includes RPS grouping prediction results, RPS distance-to-edge parameters, and structural features of the vehicle body parts. Based on the standardized input data, the standardized input data is inferred through a graph neural network or a rule-based inference engine to generate a parameter constraint subgraph containing part nodes, constraint relationships and assembly logic. Based on the parameter constraint subgraph, the parameters of the vehicle body clamping unit are optimized by a multi-agent reinforcement learning model to obtain an optimized parameter vector, and the vehicle body clamping unit is generated based on the optimized parameter vector. The step of inferring from the standardized input data using a graph neural network or a rule-based inference engine to generate a parameter constraint subgraph containing part nodes, constraint relationships, and assembly logic includes: The candidate part information is input into the rule-based inference engine, and the candidate parts are filtered by hard constraints through predefined semantic templates and process rules; Based on the subset of candidate parts output by the inference engine, embedding computation and multi-hop path inference are performed to obtain node embedding vectors and path scores. Calculate the composite index based on the node embedding vector and the path score; A set of compatible parts relationships is obtained based on the composite index and the multidimensional composite compatibility score, and a parameter constraint subgraph is constructed based on the set of compatible parts relationships. The multidimensional composite compatibility score is derived from the mechanical matching score. Geometric Combination Score Process compliance score Obtained through adjustable weighted fusion; Among them, the composite index The following is obtained: in The cosine similarity function has the following parameters. To embed the fusion weights of similarity and path score, and They are nodes and Embedded vector, Rate the path.
2. The method according to claim 1, characterized in that, After obtaining the optimized parameter vector, the method further includes iteratively executing the following until the parameters of the vehicle body clamping unit reach a preset condition: Based on the optimized parameter vector, the mechanical stiffness, spatial interference and assembly feasibility of the vehicle body clamping unit are verified using a digital twin simulation model, and the verification result data is obtained. Based on the verification results, the parameters of the vehicle body clamping unit that do not meet the preset conditions are self-healed and corrected to generate the corrected optimized parameter vector.
3. The method according to claim 2, characterized in that, Standardized input data for generating the vehicle body clamping unit includes: The functional labels output by the preceding RPS model are parsed to obtain the RPS grouping prediction results and the RPS distance margin parameters. A curvature adaptive sampling algorithm is used to dynamically adjust the sampling point density, and the structural features of the vehicle body are extracted from the three-dimensional model of the vehicle body based on the adjusted sampling point density. The structural features of the vehicle body parts are standardized to obtain the standardized input data.
4. The method according to claim 2, characterized in that, Based on the parameter constraint subgraph, the parameters of the vehicle body clamping unit are optimized using a multi-agent reinforcement learning model to obtain an optimized parameter vector, including: Based on the parameter constraint subgraph, a part intelligent agent model, an assembly intelligent agent model, and an optimization intelligent agent model are established as the multi-agent reinforcement learning model, and the initial state data for reinforcement learning is obtained using the multi-agent reinforcement learning model. Based on the initial state data, the part intelligent agent model optimizes the part parameters of the body clamping unit, the assembly intelligent agent optimizes the assembly relationship of the body clamping unit, and the optimization intelligent agent optimizes the overall performance of the body clamping unit, thereby obtaining intermediate optimization data that reflects the part parameters, the assembly relationship, and the overall performance. The intermediate optimization data is fused to generate the optimization parameter vector, which characterizes the structural configuration and performance indicators of the vehicle body clamping unit.
5. The method according to claim 2, characterized in that, Based on the optimized parameter vector, the mechanical stiffness, spatial interference, and assembly feasibility of the vehicle body clamping unit are verified using a digital twin simulation model, and the verification results data are obtained, including: Based on the optimized parameter vector, a digital twin simulation model corresponding to the vehicle body clamping unit is established, and the simulation input data to be verified is generated using the digital twin simulation model. Based on the simulation input data, mechanical performance analysis and spatial relationship calculation are performed on the vehicle body clamping unit to obtain the stress, displacement and interference detection results of the vehicle body clamping unit under stress state, which are used as the verification result data.
6. A vehicle body clamping unit generation device based on multi-agent reinforcement learning, characterized in that, include: The data acquisition module is configured to generate standardized input data for the vehicle body clamping unit, wherein the standardized input data includes RPS grouping prediction results, RPS distance-edge parameters, and structural features of the vehicle body parts; The reasoning module is configured to reason about the standardized input data using a graph neural network or a rule-based reasoning engine, and generate a parameter constraint subgraph containing part nodes, constraint relationships and assembly logic. The generation module is configured to optimize the parameters of the vehicle body clamping unit based on the parameter constraint subgraph using a multi-agent reinforcement learning model to obtain an optimized parameter vector, and generate the vehicle body clamping unit based on the optimized parameter vector. The inference module is further configured as follows: The candidate part information is input into the rule-based inference engine, and the candidate parts are filtered by hard constraints through predefined semantic templates and process rules; Based on the subset of candidate parts output by the inference engine, embedding computation and multi-hop path inference are performed to obtain node embedding vectors and path scores. Calculate the composite index based on the node embedding vector and the path score; A set of compatible parts relationships is obtained based on the composite index and the multidimensional composite compatibility score, and a parameter constraint subgraph is constructed based on the set of compatible parts relationships. The multidimensional composite compatibility score is derived from the mechanical matching score. Geometric Combination Score Process compliance score Obtained through adjustable weighted fusion; Among them, the composite index The following is obtained: in The cosine similarity function has the following parameters. To embed the fusion weights of similarity and path score, and They are nodes and Embedded vector, Rate the path.
7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the method according to any one of claims 1 to 5.
8. A computer device, characterized in that, include: Memory and processor The memory stores computer programs; The processor is configured to execute a computer program stored in the memory, wherein the computer program, when executed, causes the processor to perform the method described in any one of claims 1 to 5.
9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.
Citation Information
Patent Citations
Machining parameter optimization method based on multi-agent evolution reinforcement learning
CN115688563A
Method and device for generating three-dimensional parameterized model based on artificial intelligence
CN120408866A
