Multi-agent cooperative industrial design method and system for complex engineering

Abstract

The invention is suitable for the technical field of industrial design, and provides a multi-agent cooperative industrial design method and system for complex engineering, and the method comprises the following steps: constructing a large model agent set for a complex engineering design task; dividing a design task into sub-tasks which can be executed by an intelligent agent based on a task decomposition algorithm of a knowledge graph and an intention recognition mechanism; a Prompt-to-Code enhancement model is constructed, and task planning targets and constraints of multiple agents are converted into CAD modeling or CAE simulation design instructions capable of being directly executed by an industrial design tool through Prompt embedding; constructing a design version evolution tree, and designing a causal reasoning graph structure based on the interactive influence of target features; and inputting an asynchronous task planning strategy, a design instruction and a version evolution tree feedback result into a large model agent for iterative adaptive optimization. And the design efficiency and the scheme quality of a complex engineering design task can be obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial design, and in particular, to a multi-agent collaborative industrial design method and system for complex engineering. Background Art

[0002] In recent years, with the rapid development of the industrial manufacturing and engineering construction fields, the technical requirements for complex engineering design tasks such as water conservancy projects, bridge construction, high-end equipment manufacturing, and automotive parts manufacturing have been increasing day by day, and the design difficulty and collaborative complexity have increased significantly. Traditional industrial design methods mainly rely on manual experience and semi-automated design tools, and there are problems such as difficult task decomposition, low collaborative efficiency, poor cross-domain communication, and high iteration costs of design schemes. In addition, although current intelligent design tools can improve the automation level of some design links, there are still obvious deficiencies in the overall collaborative design process and complex task adaptive processing.

[0003] The rise of artificial intelligence, especially large language models (LLMs) and multi-agent systems (MASs), has provided new ideas and technical means for solving the above problems. However, in the prior art, the application of large models in the field of industrial design mostly stays in the auxiliary decision-making of single agents or specific functional scenarios, and fails to effectively achieve deep collaboration between different domains and multiple agents. At the same time, the lack of specialized pre-trained models adapted to the characteristics of the industrial design field limits the generalization performance and task execution efficiency of the models. In addition, current task planning methods are mostly static or predefined processes, and it is difficult to flexibly handle the frequently changing design constraints and target requirements in complex engineering design tasks. Therefore, it is necessary to provide a multi-agent collaborative industrial design method and system for complex engineering to solve the above problems. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the purpose of the present invention is to provide a multi-agent collaborative industrial design method and system for complex engineering to solve the problems existing in the above background art.

[0005] The present invention is implemented as follows. A multi-agent collaborative industrial design method for complex engineering, the method includes the following steps: S100, constructing a set of large model agents for complex engineering design tasks, including a structural design agent, a material selection agent, a simulation analysis agent, and a review agent, and establishing a Prompt vocabulary and a task constraint knowledge base through a pre-training method; S200, a task decomposition algorithm based on a knowledge graph and an intention recognition mechanism, dividing the design task into subtasks executable by agents, and realizing the adaptive adjustment of task weights through a graph attention network to form an asynchronous task planning strategy; S300. Build a Prompt-to-Code enhancement model to convert the task planning goals and constraints of multi-agent systems into CAD modeling or CAE simulation design instructions that can be directly executed by industrial design tools through Prompt embedding. S400. Construct an evolutionary tree of design versions. Based on the causal inference graph structure of the interactive influence of design target features, use the attention mechanism to learn the non-linear interactive influence paths between design features, dynamically update the version path of the design scheme, and perform traceability. S500. Input the asynchronous task planning strategy, design instructions, and the feedback results of the version evolution tree into the large model agent for iterative adaptive optimization until the goals and constraint requirements of complex engineering design tasks are met.

[0006] As a further solution of the present invention: The pre-training method includes: S101. Collect resource data for industrial design, including design drawings, simulation data, and technical documents. S102. Construct pre-training tasks, including drawing parsing tasks, simulation data embedding tasks, and document understanding tasks; establish a semantic embedding model to convert design drawings, simulation data, and technical documents into vector representations for calculation; perform self-supervised pre-training. S103. Determine the semantic embedding objective function: The embedding vectors of design drawings, simulation data, and documents are respectively 、 and , and the loss function of the model is: ; Among them, 、 and are weight coefficients for adjusting the losses of each part, 、 and are target embedding vectors, respectively representing the true semantic features of design drawings, simulation data, and documents; through the backpropagation algorithm, optimize the loss function and update the parameters of the embedding model; through the multi-modal fusion strategy of self-supervised learning, learn the correlation between drawings, simulation data, and documents, so that the generated semantic embedding can integrate the information of various data sources. S104. Model evaluation: Evaluate the performance of the trained model in industrial design tasks through the validation set or cross-validation, and evaluate the language understanding ability of the model, the planning accuracy of design tasks, and the integration ability of multi-modal information.

[0007] As a further solution of the present invention: The construction steps of the structural design agent, material selection agent, simulation analysis agent, and review agent include: Structural Design Agent: According to the design requirements, automatically or semi-automatically select the key parameters of the structure, generate or modify the three-dimensional morphological modeling of the design scheme. The structural design agent performs parameter selection and morphological optimization through the objective optimization function, and the objective optimization function is: , where is the design parameter set, represents the objective function of structural design, represents the constraint space of the structural parameters; Material Selection Agent: Identify the material properties that match the design requirements from the material knowledge base, select the appropriate materials according to the performance and constraints, and give a recommended solution for design decision-making. The recommended model for material selection is based on performance index optimization, and the material selection index function is: ; where represents the materials in the material set; represents the material database set; , and represent the strength, weight, and stability performance indicators of the materials respectively; represents the material cost function; , and represent the weight parameters of different performance indicators respectively; Simulation Analysis Agent: Receive the model generated by the structural design agent and the material parameters recommended by the material selection agent, automatically perform simulation analysis based on the CAE simulation tool, verify whether the performance of the design scheme meets the design requirements, and return the simulation results; The simulation analysis process involves finite element analysis: , where represents the structural stiffness matrix; represents the nodal displacement vector; represents the load vector; Review Agent: Receive the output results of the structural design agent, material selection agent, and simulation analysis agent, evaluate the quality and compliance of the design scheme according to the design specifications, standards, and task constraint conditions, and give a judgment on whether the requirements are met and improvement suggestions; The review involves comprehensive scoring or multi-index evaluation, and the comprehensive evaluation index of the design scheme: , where represents the comprehensive quality score of the design scheme; represents the score of the scheme in the th review dimension; represents the weight of the corresponding dimension; represents the total number of review dimensions; Compliance Judgment: When , the design solution passes the review; otherwise, return the optimized agent solution, where represents the minimum requirements for the quality or compliance of the design solution.

[0008] As a further solution of the present invention: the task decomposition algorithm of the intention recognition mechanism realizes the automatic intention classification and subtask mapping of the design task through a dual-channel Transformer intention recognition module that fuses the nodes of the domain knowledge graph and the text features of the design task. The specific steps are as follows: S201, Data preprocessing and feature extraction: Extract text features from the design task text to obtain a text sequence vector representation, and extract task-related node features from the domain knowledge graph to obtain a graph node feature vector representation; Text features: , Graph node features: , where represents the th text embedding vector, represents the th graph node embedding vector; S202, Construct a dual-channel Transformer module: Construct a text channel to capture the dependencies between text sequence features; construct a knowledge graph channel to capture the semantic relationships between graph node features; Encode the text features and graph node features respectively through the multi-head self-attention mechanism. The multi-head attention calculation is: , where , , represent the query, key, and value in the attention mechanism respectively; is the dimension of each attention head; Apply Transformer encoding to the text and graph node features respectively: Text channel encoding, ; Graph channel encoding, ; S203, Dual-channel feature fusion: Fuse the output features of the text channel and the graph channel to generate a fused feature vector: , or further achieve feature fusion through the attention mechanism: , The fusion attention formula can be expressed as: ; where , , represents the learnable parameter in the fusion attention mechanism, is the dimension of the fused feature; S204, Intent Classification and Sub - task Mapping: Based on the fused feature vector, perform intent classification, and according to the classification result, automatically map the corresponding sub - tasks; Intent classification is achieved through a fully - connected layer and the Softmax function: , where is the intent classification result, represents the probability distribution of task types; and are the parameters of the classification layer; Sub - task mapping rule: , where is the mapping rule from task to sub - task, and sub - tasks are automatically selected from the task knowledge base according to the classification result; S205, Output Task Decomposition Result: Output the set of sub - tasks after intent recognition to form the basis for subsequent multi - agent collaborative execution. The final output is: .

[0009] As a further solution of the present invention: The Prompt - to - Code enhancement model includes a cross - modal Prompt encoding mechanism, which improves the domain semantic consistency of design instruction generation and the accuracy of CAD / CAE execution through the collaborative optimization of task text and design - domain - specific DSL embedding. The specific steps are as follows: S301, Design Task Text Feature Extraction: Convert the design task text described in natural language into a text feature embedding vector and process it using a pre - trained language model, , where is the input design task text, is the text embedding feature; S302, Design - Domain - Specific DSL Feature Extraction: Extract feature vectors from design - domain - specific DSL codes to reflect the semantic correspondence between DSL instructions and tasks; S303, Cross - modal Prompt Encoding Mechanism: Adopt a cross - modal attention mechanism to achieve the fusion and collaborative optimization of text embedding and DSL embedding, enabling the two to share the same semantic space, that is: ; where , , is the learnable parameter matrix of cross - modal attention, is the fused cross - modal feature embedding vector, is the dimension of the embedding vector; S304, Prompt - to - Code Generation Module: Use the fused cross - modal feature vector to generate DSL instruction codes executable by design - domain - specific CAD / CAE tools and generate them using a Seq2Seq model: , where is the generated instruction sequence, is the -th DSL instruction element in the instruction sequence, indicating the probability of generating the current instruction element given the fused embedding vector and the previous instruction elements; S305, Domain Semantic Consistency Optimization: Ensure the semantic consistency between the generated instruction sequence and the design task text through the domain semantic consistency loss function, where, is the semantic embedding vector of the generated instruction sequence; is the domain semantic consistency loss, and the optimization objective is to minimize this loss; is the cosine similarity function, measuring the semantic similarity between the generated instruction and the text description; S306, Optimization Objective Function: Train the overall Prompt-to-Code enhancement model, jointly optimize the generation quality and domain semantic consistency, and the comprehensive optimization formula is: where, is the cross-entropy loss of the instruction generation task, is the domain semantic consistency loss, is the hyperparameter for adjusting the weight of the semantic consistency loss; S307, Generation Result Output and Execution: After optimization, output the domain-specific DSL instructions and transfer them to the industrial design tool for automated execution: where, are the finally optimized domain-specific CAD / CAE execution instructions.

[0010] As a further solution of the present invention: In the constructed design version evolution tree, the feature embedding of the design version node is jointly composed of design parameters and performance simulation metrics, and the non-linear interaction relationship between parameter differences and performance changes between different versions is clearly characterized through the graph attention mechanism. The specific steps are as follows: S401, Design Version Node Definition and Feature Construction: Define each design version as a node in the graph structure. The node features include two aspects, namely the design parameter feature and the performance simulation metric feature ; The node feature embedding is: where, is the complete feature embedding of the -th version node, is the feature concatenation operation; S402, Version Evolution Tree Construction: Construct a directed tree structure according to the chronological order or causal relationship of the design versions. The edges represent the evolution or causal dependency relationships between the versions. The initial root node is the initial design version, and each branch represents a different evolution path; Graph structure definition: , where is a set of nodes, representing each design version; is a set of edges, representing the evolutionary relationship between versions; S403, Feature Interaction Analysis of Graph Attention Network Mechanism: Through the graph attention mechanism, learn the non-linear interaction relationship between design parameters and performance changes among version nodes. The graph attention mechanism calculates the update of node feature embeddings, and its graph attention calculation formula: ; where is the node 's attention coefficient to node , reflecting the feature interaction weight between nodes; is the weight matrix of feature linear transformation; is the learning parameter of the attention mechanism; and are the feature embedding vectors of nodes and ; is the concatenation operation; is the set of neighbor nodes of node ; Node feature update formula: , where is the updated node feature vector, is the non-linear activation function; S404, Feature Difference and Performance Change Analysis: Based on the updated node features, clearly depict the influence path of design parameter changes between different versions on performance indicators, and then calculate the feature difference and performance change metric between nodes; The parameter and performance difference of the version change from node to node is expressed as: , where is the Euclidean distance of the feature difference vector between version and version ; Performance Indicator Change Significance Evaluation: , where is the influence degree of performance change between nodes; S405, Dynamic Version Path Update and Traceability Management: According to the node features and difference influence relationship output by GAT, dynamically update the version evolution path, provide version backtracking ability, and facilitate subsequent optimization and fault analysis; According to the influence weight between versions, define the path weight: , where is the weight on the version evolution path, comprehensively considering the attention weight and performance change; Dynamic Path Determination: , where is the optimal version evolution path.

[0011] As a further solution of the present invention: when performing iterative adaptive optimization, a direct preference optimization method and a LoRA parameter efficient adjustment strategy are adopted, and the specific steps are as follows: S501, data preparation and preference annotation: Collect different design schemes automatically generated by the intelligent agent, perform pairwise comparison annotation for different schemes, and clarify the preferences of design experts or the field to obtain a preference dataset. Given two design schemes Annotate the preference relationship; S502, construction of the design field preference fine-tuning model: Use the LoRA technology to adjust the parameters of the large model, reduce the demand for training samples and improve the generalization ability. LoRA decomposes the weight matrix into a low-rank form as: where, is the fixed weight matrix of the original pre-trained model; , are trainable low-rank matrices; S503, implementation of the direct preference optimization method: Use the DPO method to optimize the preference prediction probability distribution of the model by clarifying the preference relationship. For the given scheme the preference optimization loss function is: where, is the preference score given by the model for the scheme ; is the Sigmoid function, is the temperature parameter, which controls the gap sensitivity between preference scores; is the paired-annotated preference dataset; S504, construction of the joint optimization loss function: Construct a joint loss function, comprehensively combine the parameter constraints of LoRA fine-tuning and the preference loss function of DPO, and the comprehensive optimization objective: where, is the loss function for overall training; is the regularization term of the LoRA parameter, which is used to prevent overfitting; is the weight hyperparameter of the LoRA parameter regularization term; S505, iterative optimization and model update: Repeatedly perform the optimization steps of S501 - S504, and the model iteratively updates the parameters according to the latest preference data, gradually improving the generalization ability of the model. The parameter iterative update is: where, is the model parameter at the th iteration; is the learning rate; is the gradient; S506, Adaptive Termination Criterion and Generalization Evaluation: Monitor the generalization performance of the model. When the performance reaches the predetermined requirements or the generalization ability stabilizes, terminate the optimization process.

[0012] Another object of the present invention is to provide a multi-agent collaborative industrial design system for complex engineering, the system comprising: A large model agent set module, configured to construct a large model agent set for complex engineering design tasks, including a structural design agent, a material selection agent, a simulation analysis agent, and a review agent, and establish a Prompt vocabulary and a task constraint knowledge base through a pre-training method; A task decomposition and planning module, configured to divide the design task into subtasks executable by agents through a task decomposition algorithm based on a knowledge graph and an intention recognition mechanism, and adaptively adjust the task weights through a graph attention network to form an asynchronous task planning strategy; A Prompt-to-Code enhanced conversion module, configured to construct a Prompt-to-Code enhanced model, and convert the task planning objectives and constraints of multiple agents into CAD modeling or CAE simulation design instructions directly executable by industrial design tools through Prompt embedding; A design version evolution and management module, configured to construct a design version evolution tree, based on a causal inference graph structure of the interactive influence of design target features, learn the non-linear interactive influence path between design features by means of an attention mechanism, dynamically update the version path of the design scheme, and perform traceability; An iterative adaptive optimization module, configured to input the asynchronous task planning strategy, design instructions, and the feedback results of the version evolution tree into the large model agent for iterative adaptive optimization until the objectives and constraint requirements of the complex engineering design task are met.

[0013] As a further solution of the present invention: the large model intelligent agent set module includes a pre-training sub-module, which constructs a semantic embedding model dedicated to design tasks by using design drawings, simulation data, and technical documents in the field of industrial design, so as to improve the accuracy of intelligent agent language understanding and task planning; the task decomposition and planning module includes an intention recognition sub-module, which uses a dual-channel Transformer model that fuses domain knowledge graph nodes and design task text features to achieve automatic intention classification of design tasks and accurate mapping of sub-tasks; the Prompt-to-Code enhancement conversion module includes a cross-modal Prompt encoding sub-module, which improves the domain semantic consistency of design instruction generation and the accuracy of instruction execution through the collaborative optimization of design task text and design domain-specific DSL embedding; the design version evolution and management module includes a feature embedding sub-module, which constructs feature embeddings using design parameters and performance simulation indicators, and combines graph attention mechanisms to explicitly represent the non-linear interaction relationships of parameter differences and performance changes between design versions.

[0014] As a further solution of the present invention: the complex engineering design task is specifically one or more industrial design tasks in water conservancy projects, bridge construction, high-end equipment manufacturing, and automotive parts manufacturing.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The present invention can perform automatic decomposition of complex engineering design tasks, intelligent task planning, cross-modal design instruction generation, design version evolution management, and adaptive iterative optimization under small sample conditions, effectively improving the automation degree of industrial design tasks, the domain semantic consistency of design schemes, the accuracy of simulation and analysis, and the task generalization ability, reducing the dependence on data scale, and significantly improving the efficiency and quality of complex engineering design. Description of the Drawings

[0016] Figure 1 It is a flowchart of a multi-agent collaborative industrial design method for complex engineering.

[0017] Figure 2 It is a flowchart of a pre-training method in a multi-agent collaborative industrial design method for complex engineering.

[0018] Figure 3 It is a flowchart of a task decomposition algorithm in a multi-agent collaborative industrial design method for complex engineering.

[0019] Figure 4 It is a flowchart of an enhanced model construction in a multi-agent collaborative industrial design method for complex engineering.

[0020] Figure 5It is a flowchart for constructing a version evolution tree in a multi-agent collaborative industrial design method for complex engineering.

[0021] Figure 6 It is a flowchart for adaptive optimization in a multi-agent collaborative industrial design method for complex engineering.

[0022] Figure 7 It is a schematic structural diagram of a multi-agent collaborative industrial design system for complex engineering. Detailed implementation manners

[0023] To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0024] The following describes in detail the specific implementation of the present invention with reference to specific embodiments.

[0025] As Figure 1 shown, an embodiment of the present invention provides a multi-agent collaborative industrial design method for complex engineering, and the method includes the following steps: S100. Construct a domain-specific large model agent set for complex engineering design tasks, specifically including a structural design agent, a material selection agent, a simulation analysis agent and a review agent, and establish a dedicated Prompt vocabulary and a task constraint knowledge base for the design domain through a domain-specific pre-training method; S200. Based on a task decomposition algorithm of a design domain knowledge graph and an intention recognition mechanism, divide complex design tasks into subtasks executable by agents, and adaptively adjust the task weights through a graph attention network to form an asynchronous task planning strategy; S300. Construct a Prompt-to-Code enhanced model for design domain tasks, and convert the task planning objectives and constraints of multi-agents into CAD modeling or CAE simulation design instructions directly executable by industrial design tools through Prompt embedding; S400. Construct a design version evolution tree, based on the causal inference graph structure of the interactive influence of design target features, learn the non-linear interactive influence path between design features with the help of an attention mechanism, and dynamically update the version path of the design scheme and trace it; S500. Input the asynchronous task planning strategy, design instructions and the feedback result of the version evolution tree into the large model agent for iterative adaptive optimization until the objectives and constraint requirements of complex engineering design tasks are met.

[0026] As Figure 2As shown, as a preferred embodiment of the present invention, the pre-training method includes: S101. Collect resource data for industrial design, including design drawings, simulation data, and technical documents. Design drawings: Collect task-related design drawings from the field of industrial design. These drawings usually contain detailed structural designs, dimensions, scales, and other information. Simulation data: Usually includes simulation results of different design schemes under different conditions, such as simulation indicators of stress, temperature, kinematics, etc. Technical documents: Include standards, regulations, design guides, material data manuals, etc. S102. Self-supervised pre-training: Through a learning method that constructs training signals from the input data itself, improve semantic understanding and planning accuracy through self-supervised learning without a large amount of labeled data. That is: First, construct pre-training tasks, including drawing parsing tasks, simulation data embedding tasks, and document understanding tasks. That is, design drawings perform self-supervised learning through parsing tasks, match the structural elements in the drawings with the standard component database, and generate labels to learn the semantic features of the design drawings. Convert the simulation data into vector representations for training the model to learn simulation patterns through the similarity or difference between data. Use technical documents to train the model to understand professional terms and technical requirements in design tasks and establish a semantic embedding model.

[0027] Next, establish a semantic embedding model to convert design drawings, simulation data, and technical documents into vector representations (i.e., "embeddings") that can be used for calculation. Different embedding techniques are used for each data type (drawings, simulation data, documents). For design drawings, common techniques include using convolutional neural networks to extract visual features and then mapping these features to the embedding space. For simulation data, use regression or classification models to convert it into vector representations. For technical documents, use natural language processing techniques such as BERT, GPT, etc. models for text embedding.

[0028] Finally, for the self-supervised pre-training process, use unlabeled design drawings, simulation data, and documents for self-supervised training, and train the embedding model by constructing "pseudo-labels". The goal of the drawing parsing task is to enable the model to infer relevant materials, dimensions, or other design information from the design drawings. The simulation data enhances the prediction accuracy through self-supervised learning algorithms, and backpropagates the parameters predicted by the model to optimize the mapping between the design drawings and the simulation data.

[0029] S103. Optimize the semantic embedding model: Determine the semantic embedding objective function: The embedding vectors of the design drawings, simulation data, and documents are 、 and , and the loss function of the model is: ; Among them, , and are the weight coefficients for adjusting the losses of each part. , and are the target embedding vectors, representing the true semantic features of design drawings, simulation data, and documents respectively; through the backpropagation algorithm, the loss function is optimized, and the parameters of the embedding model are updated; through the multi-modal fusion strategy of self-supervised learning, the correlation between drawings, simulation data, and documents is learned, so that the generated semantic embedding can integrate the information of various data sources and improve the task understanding and planning accuracy. S104, Model evaluation: Evaluate the performance of the trained model in industrial design tasks through the validation set or cross-validation, and evaluate the language understanding ability of the model, the planning accuracy of the design task, and the integration ability of multi-modal information.

[0030] In the embodiments of the present invention, the construction steps of the structure design agent, the material selection agent, the simulation analysis agent, and the review agent include: Structure design agent: Automatically or semi-automatically select the key parameters of the structure (dimensions, shapes, geometric features, etc.) according to the design requirements, generate or modify the three-dimensional shape modeling of the design scheme (automatically generate the model through the CAD software API). The structure design agent performs parameter selection and shape optimization through the target optimization function, and the target optimization function is: Among them, is the design parameter set, represents the target function of the structure design (lightest weight, maximum strength, etc.), represents the constraint space of the structure parameters (manufacturing constraints, dimension constraints, etc.); Material selection agent: Identify the material properties (strength, density, elastic modulus, etc.) that match the design requirements from the material knowledge base, select the appropriate materials according to the performance and constraints, and give a recommended solution for design decision-making. The recommended model for material selection is optimized based on performance indicators, and the material selection index function is: ; Among them, represents the materials in the material set; represents the material database set; , and respectively represent the strength, weight, and stability performance indicators of the materials; represents the material cost function; , and respectively represent the weight parameters of different performance indicators; Simulation analysis agent: Receives the model generated by the structural design agent and the material parameters recommended by the material selection agent, automatically performs simulation analysis based on CAE simulation tools, verifies whether the performance of the design scheme meets the design requirements (such as strength, fatigue, thermal analysis, etc.), and returns simulation results such as safety factor, stress distribution, displacement field, etc.; The simulation analysis process involves finite element analysis: , where represents the structural stiffness matrix; represents the nodal displacement vector; represents the load vector; The performance evaluation formula for simulation results, such as safety factor: , where represents the safety factor; represents the allowable stress of the material; represents the maximum stress obtained in the simulation; Review agent: Receives the output results of the structural design agent, material selection agent, and simulation analysis agent, evaluates the quality and compliance of the design scheme according to design specifications, standards, and task constraints, and gives a judgment on whether it meets the requirements and improvement suggestions; The review involves comprehensive scoring or multi-index evaluation. The comprehensive evaluation index of the design scheme: , where represents the comprehensive quality score of the design scheme; represents the score of the scheme in the th review dimension (such as safety, economy, manufacturability); represents the weight of the corresponding dimension; represents the total number of review dimensions; Compliance judgment: When , the design scheme passes the review; otherwise, return the optimized scheme of the agent, where represents the minimum requirement for the quality or compliance of the design scheme. The above four agents do not work independently, but through shared data interaction and collaboration: Structural design agent → Model → Material selection agent and simulation analysis agent, Material selection agent → Material data → Simulation analysis agent, Simulation analysis agent → Simulation results → Review agent, Review agent → Optimization feedback → Readjust the design → Structural design agent and material selection agent optimize again.

[0031] Such as Figure 3 shown, as a preferred embodiment of the present invention, the task decomposition algorithm of the intention recognition mechanism realizes the automatic intention classification and subtask mapping of the design task through a dual-channel Transformer intention recognition module that fuses domain knowledge graph nodes and design task text features. The specific steps are as follows: S201, Data preprocessing and feature extraction: Extract text features from the design task text (natural language) to obtain the text sequence vector representation, and extract task-related node features (entities, relationships, attributes, etc.) from the domain knowledge graph to obtain the graph node feature vector representation; Text features: , Graph node features: , where represents the th text embedding vector, and represents the th graph node embedding vector; S202, Construct a dual-channel Transformer module: Construct a text channel to capture the dependencies between text sequence features; construct a knowledge graph channel to capture the semantic relationships between graph node features; Encode the text features and graph node features respectively through the multi-head self-attention mechanism. The multi-head attention calculation is: , where , , represent the query, key, and value in the attention mechanism respectively; is the dimension of each attention head; Apply Transformer encoding to the text and graph node features respectively: Text channel encoding, ; Graph channel encoding, ; S203, Dual-channel feature fusion: Fuse the output features of the text channel and the graph channel to generate a fused feature vector: , or further achieve feature fusion through the attention mechanism: . The fusion attention formula can be expressed as: ; where , , represents the learnable parameters in the fusion attention mechanism, and is the dimension of the fused feature; S204, Intent classification and subtask mapping: Based on the fused feature vector, perform intent classification, and according to the classification result, automatically map the corresponding subtasks; Intent classification is achieved through the fully connected layer and the Softmax function: , where is the intent classification result, and represents the probability distribution of the task types; and are the parameters of the classification layer; Subtask mapping rule: , where is the mapping rule from the task to the subtask, and the subtasks are automatically selected from the task knowledge base according to the classification result; S205, Output the task decomposition result: Output the set of subtasks after intention recognition to form the basis for subsequent multi-agent collaborative execution. The final output is: 。

[0032] As Figure 4 shown, as a preferred embodiment of the present invention, the Prompt-to-Code enhancement model includes a cross-modal Prompt encoding mechanism, which improves the domain semantic consistency of design instruction generation and the accuracy of CAD / CAE execution through the collaborative optimization of task text design and domain-specific DSL embedding. The specific steps are as follows: S301, Design task text feature extraction: Convert the design task text described in natural language into a text feature embedding vector and process it using a pre-trained language model, , where is the input design task text, is the text embedding feature; S302, Design domain-specific DSL feature extraction: Extract feature vectors from the domain-specific DSL code of the design field to reflect the semantic correspondence between DSL instructions and tasks; usually use code embedding models such as CodeBERT, CodeT5, etc. for embedding, that is: , where is the domain-specific DSL instruction sequence, is the DSL embedding feature vector; S303, Cross-modal Prompt encoding mechanism (feature fusion): Use a cross-modal attention mechanism to achieve the fusion and collaborative optimization of text embedding and DSL embedding, so that the two share the same semantic space, that is: ; where , , is the learnable parameter matrix of cross-modal attention, is the fused cross-modal feature embedding vector, is the dimension of the embedding vector; S304, Prompt-to-Code generation module (instruction generation): Use the fused cross-modal feature vector to generate DSL instruction codes executable by domain-specific CAD / CAE tools and generate them using a Seq2Seq model: , where is the generated instruction sequence, is the th DSL instruction element in the instruction sequence, represents the probability of generating the current instruction element given the fused embedding vector and the previous instruction elements; S305, Domain Semantic Consistency Optimization: Ensure the semantic consistency between the generated instruction sequence and the design task text through the domain semantic consistency loss function, , where is the semantic embedding vector of the generated instruction sequence; is the domain semantic consistency loss, and the optimization objective is to minimize this loss; is the cosine similarity function, which measures the semantic similarity between the generated instruction and the text description; S306, Optimization Objective Function: Train the overall Prompt-to-Code enhancement model, jointly optimize the generation quality and domain semantic consistency, and the comprehensive optimization formula is: , where is the cross-entropy loss of the instruction generation task, is the domain semantic consistency loss, is the hyperparameter for adjusting the weight of the semantic consistency loss; S307, Generation Result Output and Execution: After optimization, output the domain-specific DSL instructions and transfer them to industrial design tools (such as CAD / CAE software) for automated execution: , where are the finally optimized domain-specific CAD / CAE execution instructions.

[0033] As Figure 5 shown, as a preferred embodiment of the present invention, in the construction of the design version evolution tree, the feature embedding of the design version node is jointly composed of design parameters and performance simulation indicators, and the non-linear interaction relationship between parameter differences and performance changes between different versions is clearly characterized through the graph attention mechanism. The specific steps are as follows: S401, Design Version Node Definition and Feature Construction: Define each design version as a node in the graph structure. The node features include two aspects, namely the design parameter features (used to represent dimensions, material properties, geometric parameters, etc.) and the performance simulation index features (representing CAE simulation results such as strength, weight, stress, vibration, etc.); the node feature embedding is: , where is the complete feature embedding of the th version node, is the feature concatenation operation; S402, Version Evolution Tree Construction: Construct a directed tree structure according to the chronological order or causal relationship of the design versions. The edges represent the evolution or causal dependence relationship between the versions. The initial root node is the initial design version, and each branch represents a different evolution path; the graph structure definition: , where is the node set, representing each design version; is a set of edges, representing the evolutionary relationship between versions; S403, Feature Interaction Analysis of Graph Attention Network Mechanism: Through the graph attention mechanism, learn the non-linear interaction relationship between design parameters and performance changes among version nodes. The graph attention mechanism calculates the update of node feature embeddings, and its graph attention calculation formula: ; Among them, is the node for the node 's attention coefficient, reflecting the feature interaction weight between nodes; is the weight matrix of feature linear transformation; is the learning parameter of the attention mechanism; and are the feature embedding vectors of nodes and ; is the concatenation operation; is the set of neighbor nodes of node ; Node feature update formula: , where is the updated node feature vector, is the non-linear activation function (such as ReLU); S404, Version Feature Difference and Performance Change Analysis: Based on the updated node features, clearly characterize the influence path of design parameter changes between different versions on performance metrics, and then calculate the feature difference and performance change measurement between nodes; The parameter and performance difference of the version change from node to node is expressed as: , where is the version and the version , where is the influence degree of performance change between nodes; S405, Dynamic Version Path Update and Traceability Management: According to the node features and difference influence relationships output by GAT, dynamically update the version evolution path, provide version backtracking ability, and facilitate subsequent optimization and fault analysis; According to the influence weight between versions, define the path weight: , where is the weight on the version evolution path, comprehensively considering the attention weight and performance change; Dynamic path determination: , where is the best (most significantly influential) version evolution path.

[0034] Such asFigure 6 As shown in the figure, as a preferred embodiment of the present invention, when performing iterative adaptive optimization, the direct preference optimization (DPO) method and the LoRA parameter efficient adjustment strategy are adopted. The specific steps are as follows: S501, Data preparation and preference annotation: Collect different design schemes automatically generated by the agent, conduct pairwise comparison annotations for different schemes, and clarify the preferences of design experts or the field to obtain a preference dataset. Given two design schemes Annotate the preference relationship, indicating that the scheme is better than the scheme ; S502, Construction of the design field preference fine-tuning model: Use LoRA (Low-Rank Adaptation) technology to adjust the parameters of the large model, reduce the demand for training samples and improve the generalization ability. LoRA performs low-rank decomposition on the weight matrix as: , where is the fixed weight matrix of the original pre-trained model; , are trainable low-rank matrices , effectively reducing the number of parameters; S503, Implementation of the direct preference optimization method: Use the DPO method to optimize the preference prediction probability distribution of the model by clarifying the preference relationship. The preference optimization loss function for the given scheme is: , where is the preference score given by the model for the scheme ; is the Sigmoid function, is the temperature parameter, which controls the gap sensitivity between preference scores; is the pairwise annotated preference dataset; S504, Construction of the joint optimization loss function: Construct a joint loss function, combine the parameter constraints of LoRA fine-tuning and the preference loss function of DPO, and the comprehensive optimization objective: , where is the loss function for overall training; is the regularization term of LoRA parameters, which is used to prevent overfitting; is the weight hyperparameter of the LoRA parameter regularization term; S505, Iterative optimization and model update: Repeatedly perform the optimization steps of S501 - S504. The model iteratively updates the parameters according to the latest preference data, gradually improving the generalization ability of the model. The parameter iterative update: , where is the model parameter at the th iteration; is the learning rate; is the gradient; S506, Adaptive termination criterion and generalization evaluation: Monitor the generalization performance of the model (such as its performance in the validation set or small-sample tasks). When the performance reaches the predetermined requirements or the generalization ability is stable, terminate the optimization process.

[0035] such as Figure 7 As shown, an embodiment of the present invention also provides a multi-agent collaborative industrial design system for complex engineering. The system includes: The large model agent set module 100 is used to construct a large model agent set for complex engineering design tasks, including a structure design agent, a material selection agent, a simulation analysis agent, and a review agent, and establish a Prompt vocabulary and a task constraint knowledge base through a pre-training method; The task decomposition and planning module 200 is used to divide the design task into subtasks executable by agents through a task decomposition algorithm based on a knowledge graph and an intention recognition mechanism, and adaptively adjust the task weights through a graph attention network to form an asynchronous task planning strategy; The Prompt-to-Code enhanced conversion module 300 is used to construct a Prompt-to-Code enhanced model, and convert the task planning objectives and constraints of multiple agents into CAD modeling or CAE simulation design instructions directly executable by industrial design tools through Prompt embedding; The design version evolution and management module 400 is used to construct a design version evolution tree, based on the causal inference graph structure of the interactive influence of design target features, learn the non-linear interactive influence paths between design features with the help of an attention mechanism, dynamically update the version path of the design scheme and perform traceability; The iterative adaptive optimization module 500 is used to input the asynchronous task planning strategy, design instructions, and the feedback results of the version evolution tree into the large model agent for iterative adaptive optimization until the objectives and constraint requirements of complex engineering design tasks are met.

[0036] In the embodiments of the present invention, the large model intelligent agent set module 100 includes a pre-training sub-module, which constructs a semantic embedding model dedicated to design tasks by using design drawings, simulation data, and technical documents in the field of industrial design, so as to improve the accuracy of intelligent agent language understanding and task planning; the task decomposition and planning module 200 includes an intention recognition sub-module, which uses a dual-channel Transformer model that fuses domain knowledge graph nodes and design task text features to achieve automatic intention classification of design tasks and accurate mapping of sub-tasks; the Prompt-to-Code enhancement conversion module 300 includes a cross-modal Prompt encoding sub-module, which improves the domain semantic consistency of design instruction generation and the accuracy of instruction execution through the collaborative optimization of design task text and design domain-specific DSL embedding; the design version evolution and management module 400 includes a feature embedding sub-module, which constructs feature embeddings by using design parameters and performance simulation indicators, and combines graph attention mechanisms to explicitly represent the non-linear interaction relationships of parameter differences and performance changes between design versions.

[0037] In the embodiments of the present invention, the complex engineering design task is specifically one or more industrial design tasks in water conservancy projects, bridge construction, high-end equipment manufacturing, and automotive parts manufacturing.

[0038] The above only describes the preferred embodiments of the present invention in detail, and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

[0039] It should be understood that although the steps in the flowcharts of the embodiments of the present invention are shown in sequence according to the arrows, these steps do not necessarily have to be executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps does not have a strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages do not necessarily have to be executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages does not necessarily have to be sequential, but can be executed alternately or alternately with at least a part of other steps or sub-steps or stages of other steps.

[0040] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0041] After considering the specification and the disclosure of the embodiments, those skilled in the art will readily conceive of other embodiments of the present disclosure. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include known common knowledge or conventional technical means in the technical field not disclosed in the present disclosure. The specification and the embodiments are only regarded as exemplary, and the true scope and spirit of the present disclosure are pointed out by the claims.

Claims

1. A multi-agent collaborative industrial design method for complex engineering, characterized in that, The method includes the following steps: S100. Construct a set of large model agents for complex engineering design tasks, including a structural design agent, a material selection agent, a simulation analysis agent, and a review agent. Establish a Prompt vocabulary and a task constraint knowledge base through a pre-training method; S200. Based on a task decomposition algorithm of a knowledge graph and an intention recognition mechanism, divide the design task into subtasks executable by agents, and adaptively adjust the task weights through a graph attention network to form an asynchronous task planning strategy; S300. Construct a Prompt-to-Code enhancement model, and convert the task planning objectives and constraints of multiple agents into CAD modeling or CAE simulation design instructions directly executable by industrial design tools through Prompt embedding; S400. Construct a design version evolution tree. Based on the causal inference graph structure of the interactive influence of design target features, use the attention mechanism to learn the non-linear interactive influence path between design features, dynamically update the version path of the design scheme, and perform traceability; S500. Input the asynchronous task planning strategy, design instructions, and the feedback results of the version evolution tree into the large model agent for iterative adaptive optimization until the objectives and constraint requirements of complex engineering design tasks are met.

2. The multi-agent collaborative industrial design method for complex engineering according to claim 1, characterized in that, The pre-training method includes: S101. Collect resource data for industrial design, including design drawings, simulation data, and technical documents; S102. Construct pre-training tasks, including drawing parsing tasks, simulation data embedding tasks, and document understanding tasks; establish a semantic embedding model to convert design drawings, simulation data, and technical documents into vector representations for calculation; perform self-supervised pre-training; S103. Determine the semantic embedding objective function: The embedding vectors of the design drawings, simulation data, and documents are respectively , and , and the loss function of the model is as follows: ； Among them, , and are the weight coefficients for adjusting the losses of each part. , and are the target embedding vectors, representing the true semantic features of design drawings, simulation data, and documents respectively. Through the backpropagation algorithm, the loss function is optimized and the parameters of the embedding model are updated. Through the multi-modal fusion strategy of self-supervised learning, the correlations between drawings, simulation data, and documents are learned, so that the generated semantic embeddings can integrate the information of various data sources. S104. Model evaluation: Evaluate the performance of the trained model in industrial design tasks through a validation set or cross-validation, and evaluate the model's language understanding ability, the planning accuracy of design tasks, and the integration ability of multi-modal information.

3. The multi-agent collaborative industrial design method for complex engineering according to claim 1, characterized in that The construction steps of the structural design agent, the material selection agent, the simulation analysis agent, and the review agent include: Structural Design Agent: According to the design requirements, automatically or semi-automatically select the key parameters of the structure, generate or modify the three-dimensional morphological modeling of the design scheme. The Structural Design Agent performs parameter selection and morphological optimization through the objective optimization function, and the objective optimization function is: , where is the design parameter set, represents the objective function of the structural design, represents the constraint space of the structural parameters; Material selection agent: Identify material properties that match the design requirements from the material knowledge base, select appropriate materials according to performance and constraints, and give recommended solutions for design decisions. The recommended model for material selection is optimized based on performance indicators. The material selection index function is: ； Among them, represents the material in the material set; represents the material database set; 、 and respectively represent the strength, weight and stability performance indexes of the material; represents the material cost function; 、 and respectively represent the weight parameters of different performance indexes; Simulation analysis agent: Receives the model generated by the structural design agent and the material parameters recommended by the material selection agent, automatically performs simulation analysis based on the CAE simulation tool, verifies whether the performance of the design scheme meets the design requirements, and returns the simulation results; The simulation analysis process involves finite element analysis: , where represents the structural stiffness matrix; represents the nodal displacement vector; represents the load vector; Review Agent: Receives the output results of the Structural Design Agent, Material Selection Agent, and Simulation Analysis Agent, evaluates the quality and compliance of the design solution according to design specifications, standards, and task constraints, and gives a judgment on whether the requirements are met and improvement suggestions; The review involves comprehensive scoring or multi-index evaluation. The comprehensive evaluation index of the design solution is: , where represents the comprehensive quality score of the design solution; represents the score of the solution in the th review dimension; represents the weight of the corresponding dimension; represents the total number of review dimensions; Compliance judgment: When , the design plan passes the review; otherwise, return the agent optimization plan, where represents the minimum requirement for the quality or compliance of the design plan.

4. The multi-agent collaborative industrial design method for complex engineering according to claim 1, characterized in that, The task decomposition algorithm of the intention recognition mechanism realizes the automatic intention classification and subtask mapping of design tasks through a dual-channel Transformer intention recognition module that fuses domain knowledge graph nodes and design task text features. The specific steps are as follows: S201. Data preprocessing and feature extraction: Extract text features from the design task text to obtain a text sequence vector representation, and extract task-related node features from the domain knowledge graph to obtain a graph node feature vector representation; Text features: , graph node features: , where represents the th text embedding vector, represents the th graph node embedding vector; S202, Construct a dual-channel Transformer module: Construct a text channel to capture the dependencies between text sequence features; construct a knowledge graph channel to capture the semantic relationships between graph node features; encode the text features and graph node features respectively through the multi-head self-attention mechanism. The multi-head attention calculation is: , where , , represent the query, key, and value in the attention mechanism respectively; is the dimension of each attention head; apply Transformer encoding to the text and graph node features respectively: Text channel encoding, ; Graph channel encoding, ; S203, Dual-channel Feature Fusion: Fuse the output features of the text channel and the atlas channel to generate a fused feature vector: , or further achieve feature fusion through an attention mechanism: , and the fusion attention formula can be expressed as: ； Among them, , , represent the learnable parameters in the fusion attention mechanism, is the dimension of the fused features; S204, Intent Classification and Sub - task Mapping: Based on the fused feature vector, perform intent classification, and according to the classification result, automatically map the corresponding sub - tasks; Intent classification is implemented through a fully - connected layer and the Softmax function: , where, is the intent classification result, represents the probability distribution of task types; and are the parameters of the classification layer; Sub - task mapping rule: , where, is the mapping rule from tasks to sub - tasks, and sub - tasks are automatically selected from the task knowledge base according to the classification result; S205, Output the task decomposition result: Output the set of subtasks after intention recognition to form the basis for subsequent multi-agent collaborative execution, and finally output: .

5. The multi-agent collaborative industrial design method for complex engineering according to claim 1, wherein The Prompt-to-Code enhancement model includes a cross-modal Prompt encoding mechanism, which improves the domain semantic consistency of design instruction generation and the accuracy of CAD / CAE execution through the collaborative optimization of design task text and domain-specific DSL embedding in the design field. The specific steps are as follows: S301, Design task text feature extraction: Convert the design task text described in natural language into a text feature embedding vector, and process it using a pre-trained language model, , where, is the input design task text, is the text embedding feature; S302, Design Domain-Specific DSL Feature Extraction: Extract feature vectors from design domain-specific DSL code to reflect the semantic correspondence between DSL instructions and tasks; S303, Cross-Modal Prompt Encoding Mechanism: Adopt a cross-modal attention mechanism to achieve the fusion and collaborative optimization of text embedding and DSL embedding, enabling them to share the same semantic space, that is: ； Among them, , , is a learnable parameter matrix for cross-modal attention, is the fused cross-modal feature embedding vector, is the dimension of the embedding vector; S304, Prompt-to-Code Generation Module: Using the fused cross-modal feature vectors, generate DSL instruction codes executable by domain-specific CAD / CAE tools, and use the Seq2Seq model for generation: , where is the generated instruction sequence, is the th DSL instruction element in the instruction sequence, represents the probability of generating the current instruction element given the fused embedding vector and the previous instruction elements; S305, Domain Semantic Consistency Optimization: Ensure the semantic consistency between the generated instruction sequence and the design task text through the domain semantic consistency loss function, , where is the semantic embedding vector of the generated instruction sequence; is the domain semantic consistency loss, and the optimization goal is to minimize this loss; is the cosine similarity function, which measures the semantic similarity between the generated instruction and the text description; S306, Optimize the objective function: Train the overall Prompt-to-Code enhancement model, jointly optimize the generation quality and domain semantic consistency, and the comprehensive optimization formula is: , where is the cross-entropy loss for the instruction generation task, is the domain semantic consistency loss, is the hyperparameter for adjusting the weight of the semantic consistency loss; S307, Generation result output and execution: After optimization is completed, output domain-specific DSL instructions and transfer them to the industrial design tool for automated execution: , where The finally optimized domain-specific CAD / CAE execution instructions.

6. The multi-agent collaborative industrial design method for complex engineering according to claim 1, wherein In the construction of the design version evolution tree, the feature embedding of the design version node is jointly composed of design parameters and performance simulation indicators. Through the graph attention mechanism, the non-linear interaction relationship between parameter differences and performance changes among different versions is clearly characterized. The specific steps are as follows: S401, Design Version Node Definition and Feature Construction: Define each design version as a node in a graph structure. The node features include two aspects, namely design parameter features and performance simulation index features ; The node feature embedding is as follows: , where is the complete feature embedding of the -th version node, is the feature concatenation operation; S402, Version Evolution Tree Construction: Construct a directed tree structure according to the chronological order or causal relationship of design versions. The edges represent the evolution or causal dependencies between versions. The initial root node is the initial design version, and each branch represents a different evolution path; Graph Structure Definition: , where is the set of nodes, representing each design version; is the set of edges, representing the evolution relationship between versions; S403, Feature Interaction Analysis of Graph Attention Network Mechanism: Through the graph attention mechanism, learn the non-linear interaction relationship between design parameters and performance changes among version nodes. The graph attention mechanism calculates the update of node feature embedding, and its graph attention calculation formula: ； Among them, is the node is the attention coefficient for the node reflecting the feature interaction weight between nodes; is the weight matrix of the feature linear transformation; is the learning parameter of the attention mechanism; and are the feature embedding vectors of nodes and ; is the concatenation operation; is the set of neighbor nodes of node ; The node feature update formula: , where is the updated node feature vector, is the non-linear activation function; S404, Version Feature Difference and Performance Variation Analysis: Based on the updated node features, clearly characterize the impact path of design parameter changes between different versions on performance metrics, and then calculate the feature differences and performance variation metrics between nodes; nodes to node The parameter and performance differences of version changes are expressed as: , where is the Euclidean distance of the feature difference vector between version and version ; Significance evaluation of performance metric changes: , where is the degree of impact of performance changes between nodes. S405, Dynamic Version Path Update and Traceability Management: Dynamically update the version evolution path based on the node features and differential impact relationships output by GAT, provide version backtracking capabilities for subsequent optimization and fault analysis; Define path weights according to the impact weights between versions: , where is the weight on the version evolution path, comprehensively considering the attention weight and performance changes; Dynamic path determination: , where is the optimal version evolution path.

7. The multi-agent collaborative industrial design method for complex engineering according to claim 1, characterized in that When performing iterative adaptive optimization, adopt the direct preference optimization method and the LoRA parameter efficient adjustment strategy. The specific steps are as follows: S501, Data Preparation and Preference Annotation: Collect different design solutions automatically generated by the agent, conduct pairwise comparison annotations for different solutions, and clarify the preferences of design experts or the field to obtain a preference dataset. Given two design solutions Annotate the preference relationship; S502, Design Domain Preference Fine-Tuning Model Construction: Use LoRA technology to adjust the parameters of the large model, reduce the training sample requirements and improve the generalization ability. LoRA performs low-rank decomposition on the weight matrix as follows: , where is the fixed weight matrix of the original pre-trained model; , are trainable low-rank matrices; S503, Implementation of the direct preference optimization method: Using the DPO method, by clarifying the preference relationship, optimize the preference prediction probability distribution of the model for a given solution The preference optimization loss function is as follows: , where is the preference score given by the model for the solution ; is the Sigmoid function is the temperature parameter, controlling the gap sensitivity between preference scores; is the preference dataset with pairwise annotations S504, Construction of Joint Optimization Loss Function: Construct a joint loss function that combines the parameter constraints of LoRA fine-tuning and the preference loss function of DPO, and the comprehensive optimization objective: , where is the loss function for overall training; is the regularization term of LoRA parameters, which is used to prevent overfitting; is the weight hyperparameter of the LoRA parameter regularization term S505, Iterative Optimization and Model Update: Repeatedly perform the optimization steps of S501 - S504. The model iteratively updates its parameters based on the latest preference data, gradually enhancing the model's generalization ability. Parameter Iterative Update: , where is the model parameter for the th iteration; is the learning rate; is the gradient; S506, Adaptive Termination Criterion and Generalization Evaluation: Monitor the generalization performance of the model. When the performance reaches the predetermined requirements or the generalization ability is stable, terminate the optimization process.

8. A multi-agent collaborative industrial design system for complex engineering, characterized in that, The system includes: A large model intelligent agent set module, which is used to construct a large model intelligent agent set for complex engineering design tasks, including a structural design intelligent agent, a material selection intelligent agent, a simulation analysis intelligent agent, and a review intelligent agent. Establish a Prompt vocabulary and a task constraint knowledge base through pre-training methods; A task decomposition and planning module, which is used for a task decomposition algorithm based on a knowledge graph and an intention recognition mechanism to divide the design task into subtasks that can be executed by intelligent agents, and realize the adaptive adjustment of task weights through a graph attention network to form an asynchronous task planning strategy; A Prompt-to-Code Enhanced Conversion Module, which is used to construct a Prompt-to-Code enhanced model to convert the task planning goals and constraints of multiple intelligent agents into CAD modeling or CAE simulation design instructions that can be directly executed by industrial design tools through Prompt embedding; A design version evolution and management module, which is used to construct a design version evolution tree, based on the causal inference graph structure of the interactive influence of design target features, learn the non-linear interactive influence path between design features through the attention mechanism, dynamically update the version path of the design scheme and perform traceability; An iterative adaptive optimization module, which is used to input the asynchronous task planning strategy, design instructions, and the feedback results of the version evolution tree into the large model intelligent agent for iterative adaptive optimization until the goals and constraint requirements of complex engineering design tasks are met.

9. The multi-agent collaborative industrial design system for complex engineering according to claim 8, characterized in that, The large model intelligent agent set module includes a pre-training sub-module. The pre-training sub-module uses design drawings, simulation data, and technical documents in the industrial design field to construct a semantic embedding model dedicated to design tasks, improving the language understanding and task planning accuracy of intelligent agents; The task decomposition and planning module includes an intention recognition sub-module, which uses a dual-channel Transformer model that fuses the nodes of the domain knowledge graph and the text features of the design task to achieve automatic intention classification of the design task and precise mapping of sub-tasks; the Prompt-to-Code enhancement conversion module includes a cross-modal Prompt encoding sub-module, which improves the domain semantic consistency of the design instruction generation and the accuracy of instruction execution through the collaborative optimization of the design task text and the embedding of the design domain-specific DSL; the design version evolution and management module includes a feature embedding sub-module, which constructs feature embeddings using design parameters and performance simulation metrics, and combines graph attention mechanisms to explicitly represent the non-linear interaction relationships of parameter differences and performance changes between design versions.

10. The multi-agent collaborative industrial design system for complex engineering according to claim 8, wherein, The complex engineering design tasks are specifically one or more industrial design tasks in water conservancy projects, bridge construction, high-end equipment manufacturing, and automotive parts manufacturing.

Images