A multi-agent collaborative industrial design method and system for complex engineering

Abstract

This invention, applicable to the field of industrial design technology, provides a multi-agent collaborative industrial design method and system for complex engineering projects. The method includes the following steps: constructing a large-scale intelligent agent collection for complex engineering design tasks; using a task decomposition algorithm based on a knowledge graph and intent recognition mechanism to divide the design task into agent-executable subtasks; constructing a Prompt-to-Code enhancement model to convert the multi-agent task planning goals and constraints into CAD modeling or CAE simulation design instructions directly executable by industrial design tools through prompt embedding; constructing a design version evolution tree based on a causal reasoning graph structure that reflects the interactive influence of design goal features; and inputting asynchronous task planning strategies, design instructions, and version evolution tree feedback results into the large-scale intelligent agent for iterative adaptive optimization. This method can significantly improve the design efficiency and solution quality of complex engineering design tasks.

Description

Technical Field

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

[0002] In recent years, with the rapid development of industrial manufacturing and engineering construction, 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, significantly increasing the difficulty of design and the complexity of collaboration. Traditional industrial design methods, which rely primarily on manual experience and semi-automated design tools, suffer from problems such as difficulty in task decomposition, low collaboration efficiency, poor cross-domain communication, and high costs of iterative design solutions. Furthermore, while current intelligent design tools can improve the automation level of some design links, they still have significant shortcomings in the overall collaborative design process and the adaptive processing of complex tasks.

[0003] The rise of artificial intelligence, especially large language models (LLMs) and multi-agent systems (MASs), has provided new ideas and technical means to solve the above problems. However, in existing technologies, the application of large models in the field of industrial design is mostly limited to assisting decision-making in a single agent or specific functional scenarios, and has failed to effectively achieve deep collaboration across domains and multiple agents. At the same time, the lack of specialized pre-trained models adapted to the characteristics of the industrial design field has limited the generalization performance of the model and the efficiency of task execution. In addition, current task planning methods are mostly static or predefined processes, which make it 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] In view of the deficiencies in the prior art, the purpose of the present invention is to provide a multi-agent collaborative industrial design method and system for complex engineering projects to solve the problems existing in the above-mentioned background technology.

[0005] The present invention is implemented as follows: a multi-agent collaborative industrial design method for complex engineering projects, the method comprising the following steps:

[0006] S100 builds a large-scale intelligent agent collection for complex engineering design tasks, including structural design agents, material selection agents, simulation analysis agents, and review agents. It establishes a prompt vocabulary and task constraint knowledge base through pre-training methods.

[0007] S200, a task decomposition algorithm based on knowledge graphs and intent recognition mechanisms, divides design tasks into agent-executable subtasks and adaptively adjusts task weights through a graph attention network to form an asynchronous task planning strategy;

[0008] S300 builds a Prompt-to-Code enhancement model that converts multi-agent task planning goals and constraints into CAD modeling or CAE simulation design instructions that can be directly executed by industrial design tools through prompt embedding;

[0009] S400 builds a design version evolution tree based on a causal reasoning graph structure of the interaction between design target features. It uses an attention mechanism to learn the nonlinear interaction paths between design features, dynamically updates the version path of the design solution, and performs traceability.

[0010] S500 inputs the asynchronous task planning strategy, design instructions, and version evolution tree feedback results into the large model agent for iterative adaptive optimization until the objectives and constraints of the complex engineering design task are met.

[0011] As a further solution of the present invention: the pre-training method includes:

[0012] S101, collect resource data for industrial design, including design drawings, simulation data and technical documents;

[0013] S102: Construct pre-training tasks, including drawing parsing, simulation data embedding, and document understanding; establish a semantic embedding model to convert design drawings, simulation data, and technical documents into vector representations for computation; and perform self-supervised pre-training.

[0014] S103, determine the semantic embedding objective function: the embedding vectors of design drawings, simulation data and documents are 、 and , the loss function of the model for:

[0015] ;

[0016] in, 、 and is the weight coefficient for adjusting the loss of each part, 、 and is the target embedding vector, representing the true semantic features of the design drawing, simulation data, and document respectively. Through the back-propagation algorithm, the loss function is optimized and the parameters of the embedding model are updated. Through the multimodal fusion strategy of self-supervised learning, the correlation between the drawings, simulation data, and documents is learned, so that the generated semantic embedding can integrate information from various data sources.

[0017] S104, Model Evaluation: Evaluate the performance of the trained model in industrial design tasks through validation sets or cross-validation, and evaluate the model's language comprehension ability, planning accuracy of design tasks, and ability to integrate multimodal information.

[0018] As a further solution of the present invention: the steps of constructing the structure design agent, the material selection agent, the simulation analysis agent, and the review agent include:

[0019] Structural design agent: Based on design requirements, it automatically or semi-automatically selects key parameters of the structure and generates or modifies the 3D morphological modeling of the design scheme. The structural design agent performs parameter selection and morphological optimization through the target optimization function, which is: ,in, is the design parameter set, represents the objective function of structural design, represents the constraint space of structural parameters;

[0020] Material selection agent: Identifies material properties that match design requirements from the material knowledge base, selects appropriate materials based on performance and constraints, and provides recommendations for design decisions. The material selection recommendation model is based on performance index optimization. The material selection index function is:

[0021] ;

[0022] in, Represents a material in a material collection; Represents a collection of material databases; 、 and Respectively represent the strength, weight and stability performance indicators of the material; represents the material cost function; 、 and Respectively represent the weight parameters of different performance indicators;

[0023] Simulation Analysis Agent: This agent receives the model generated by the Structural Design Agent and the material parameters recommended by the Material Selection Agent. It automatically performs simulation analysis based on CAE simulation tools to verify whether the performance of the design solution meets the design requirements and returns the simulation results. The simulation analysis process involves finite element analysis: ,in, represents the structural stiffness matrix; represents the node displacement vector; represents the load vector;

[0024] Review Agent: Receives the output of the structural design agent, material selection agent, and simulation analysis agent. Based on the design specifications, standards, and task constraints, it evaluates the quality and compliance of the design solution, and makes judgments on whether the requirements are met and provides suggestions for improvement. The review involves comprehensive scoring or multi-index evaluation. The comprehensive evaluation indicators of the design solution are: ,in, Indicates the comprehensive quality score of the design scheme; Indicates that the scheme The scores on each evaluation dimension; Indicates the weight of the corresponding dimension; Indicates the total number of dimensions reviewed;

[0025] Compliance judgment: When , the design plan passes the review; otherwise, the agent optimization plan is returned, where Indicates the minimum requirements for design quality or compliance.

[0026] As a further solution of the present invention, the task decomposition algorithm of the intent recognition mechanism realizes automatic intent classification and subtask mapping of design tasks through a dual-channel Transformer intent recognition module that integrates domain knowledge graph nodes and design task text features. The specific steps are as follows:

[0027] S201, data preprocessing and feature extraction: extract text features from the design task text to obtain a text sequence vector representation, extract task-related node features from the domain knowledge graph to obtain a graph node feature vector representation; text features: , graph node features: ,in, Indicates the text embedding vectors, Indicates the graph node embedding vectors;

[0028] S202, build a dual-channel Transformer module: build a text channel to capture the dependency between text sequence features; build a knowledge graph channel to capture the semantic relationship between graph node features; encode text features and graph node features respectively through a multi-head self-attention mechanism. The multi-head attention calculation is: ,in, 、 、 Represent the query, key, and value in the attention mechanism respectively; is the dimension of each attention head; Transformer encoding is applied to text and graph node features respectively: text channel encoding, ; Spectral channel coding, ;

[0029] 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 implement feature fusion through the attention mechanism: , the fusion attention formula can be expressed as:

[0030] ;

[0031] in, , , represents the learnable parameters in the fusion attention mechanism, is the dimension of fusion features;

[0032] S204, Intent Classification and Subtask Mapping: Intent classification is performed based on the fused feature vector, and the corresponding subtasks are automatically mapped according to the classification results. Intent classification is achieved through the fully connected layer and the Softmax function: ,in, is the intent classification result, Represents the probability distribution of task types; and is the classification layer parameter; subtask mapping rule: ,in, It is a task to subtask mapping rule that automatically selects subtasks from the task knowledge base based on the classification results;

[0033] S205, output task decomposition results: Output the subtask set after intention recognition, forming the basis for subsequent multi-agent collaborative execution, and finally output: .

[0034] As a further solution of the present invention, the Prompt-to-Code enhancement model includes a cross-modal Prompt encoding mechanism. By co-optimizing the embedding of design task text and design domain-specific DSL, the domain semantic consistency of design instruction generation and the accuracy of CAD / CAE execution are improved. The specific steps are as follows:

[0035] 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. ,in, is the text feature, Embed features for text;

[0036] S302, design domain specific DSL feature extraction: extracting feature vectors from the design domain specific DSL code to reflect the semantic correspondence between DSL instructions and tasks;

[0037] S303, Cross-modal Prompt Encoding Mechanism: A cross-modal attention mechanism is used to achieve the fusion and collaborative optimization of text embedding and DSL embedding, so that the two share the same semantic space, namely:

[0038] ;

[0039] in, , , is the learnable parameter matrix of cross-modal attention, is the fused cross-modal feature embedding vector, is the embedding vector dimension;

[0040] S304, Prompt-to-Code Generation Module: Utilizes the fused cross-modal feature vectors to generate DSL instruction code executable by domain-specific CAD / CAE tools. This is done using a Seq2Seq model. ,in, is the generated instruction sequence, The first instruction in the sequence DSL directive elements, represents the probability of generating the current instruction element given the fused embedding vector and the previous instruction element;

[0041] S305, domain semantic consistency optimization: ensure the semantic consistency of the generated instruction sequence and the design task text through the domain semantic consistency loss function, ,in, To generate semantic embedding vectors for instruction sequences; 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 instructions and the text description;

[0042] S306, optimization objective function: train the overall Prompt-to-Code enhancement model, jointly optimize generation quality and domain semantic consistency, and the comprehensive optimization formula is: ,in, The cross entropy loss for the instruction generation task, is the domain semantic consistency loss, To adjust the hyperparameters of semantic consistency loss weight;

[0043] S307, generate result output and execute: After the optimization is completed, the domain-specific DSL instructions are output and passed to the industrial design tool for automated execution: ,in, The final optimized domain-specific CAD / CAE execution instructions.

[0044] As a further solution of the present invention: in the construction of the design version evolution tree, the feature embedding of the design version node is composed of design parameters and performance simulation indicators. The nonlinear interactive relationship between parameter differences and performance changes between different versions is clearly characterized through the graph attention mechanism. The specific steps are as follows:

[0045] S401, design version node definition and feature construction: Each design version is defined as a node in the graph structure. The node features include two aspects: design parameter features and performance simulation index characteristics ; Node feature embedding is: ,in, For the The complete feature embedding of the version node, It is the feature splicing operation;

[0046] S402, version evolution tree construction: Construct a directed tree structure based on the chronological order or causal relationship of the design versions. The edges represent the evolution or causal dependency between versions. The initial root node is the initial design version, and each branch represents a different evolution path. Graph structure definition: ,in, is a set of nodes, representing each design version; is a set of edges, representing the evolutionary relationship between versions;

[0047] S403, feature interaction analysis of the graph attention network mechanism: Through the graph attention mechanism, the nonlinear interaction relationship between design parameters and performance changes between learning version nodes is learned. The graph attention mechanism calculates the node feature embedding update. Its graph attention calculation formula is:

[0048] ;

[0049] in, For nodes For Node The attention coefficient reflects the feature interaction weight between nodes; is the weight matrix of feature linear transformation; Learning parameters for the attention mechanism; and For nodes and The feature embedding vector of For splicing operation; For nodes The neighbor node set of ; node feature update formula: ,in, is the updated node feature vector, is a nonlinear activation function;

[0050] S404, version feature difference and performance change analysis: Based on the updated node features, clearly characterize the impact path of design parameter changes between different versions on performance indicators, and then calculate the feature difference and performance change metrics between nodes; To Node The parameter and performance differences of the version changes are expressed as follows: ,in, For version With version Euclidean distance between feature difference vectors; performance index change significance evaluation: ,in, is the impact degree of performance changes between nodes;

[0051] S405, Dynamic Version Path Update and Traceability Management: Based on the node characteristics and difference impact relationships output by GAT, the version evolution path is dynamically updated, providing version backtracking capabilities to facilitate subsequent optimization and fault analysis; path weights are defined based on the impact weights between versions: ,in, The weight of the version evolution path is determined by comprehensively considering the attention weight and performance changes; dynamic path determination: ,in, The best version evolution path.

[0052] 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:

[0053] S501, Data Preparation and Preference Annotation: Collect different design solutions automatically generated by the agent, perform pairwise comparison annotations on different solutions, clarify the preferences of design experts or fields, and obtain a preference dataset. Given two design solutions Label preference relations;

[0054] S502, design domain preference fine-tuning model construction: use LoRA technology to adjust the parameters of large models, reduce the demand for training samples and improve generalization ability, LoRA weight matrix Perform low-rank decomposition as: ,in, is the fixed weight matrix of the original pre-trained model; 、 is a trainable low-rank matrix;

[0055] S503, direct preference optimization method implementation: using the DPO method, by clarifying the preference relationship, optimizing the model preference prediction probability distribution, for a given solution The preference optimization loss function is: ,in, Model pair solution the preference score given; is the Sigmoid function, is the temperature parameter, controlling the sensitivity of the gap between preference scores; It is a pairwise labeled preference dataset;

[0056] S504, Joint Optimization Loss Function Construction: Construct a joint loss function that integrates the parameter constraints of LoRA fine-tuning and the preference loss function of DPO, and optimizes the overall goal: ,in, is the loss function of the overall training; is the regularization term of the LoRA parameter, used to prevent overfitting; is the LoRA parameter regularization term weight hyperparameter;

[0057] S505, iterative optimization and model update: Repeat the optimization steps S501-S504. The model iteratively updates parameters based on the latest preference data, gradually improving the model's generalization ability. Parameters are iteratively updated: ,in, For the Model parameters for the iteration; is the learning rate; is the gradient;

[0058] S506, Adaptive Termination Criteria and Generalization Evaluation: Monitor the generalization performance of the model and terminate the optimization process when the performance reaches the predetermined requirements or the generalization ability is stable.

[0059] Another object of the present invention is to provide a multi-agent collaborative industrial design system for complex engineering projects, the system comprising:

[0060] The large-model agent collection module is used to build a large-model agent collection for complex engineering design tasks, including structural design agents, material selection agents, simulation analysis agents, and review agents. The prompt vocabulary and task constraint knowledge base are established through pre-training methods.

[0061] The task decomposition and planning module is used to implement a task decomposition algorithm based on knowledge graphs and intent recognition mechanisms, dividing design tasks into subtasks that can be executed by agents. It also uses a graph attention network to achieve adaptive adjustment of task weights and form an asynchronous task planning strategy.

[0062] The Prompt-to-Code enhanced conversion module is used to build a Prompt-to-Code enhanced model, which converts the multi-agent task planning goals and constraints into CAD modeling or CAE simulation design instructions that can be directly executed by industrial design tools through prompt embedding;

[0063] The design version evolution and management module is used to build a design version evolution tree. Based on the causal reasoning graph structure of the interaction between design target features, it uses the attention mechanism to learn the nonlinear interaction path between design features, dynamically update the version path of the design solution, and conduct traceability.

[0064] The iterative adaptive optimization module is used to input asynchronous task planning strategies, design instructions, and version evolution tree feedback results into the large model agent for iterative adaptive optimization until the objectives and constraints of complex engineering design tasks are met.

[0065] As a further solution of the present invention: the large-model intelligent agent collection module includes a pre-training sub-module, which uses design drawings, simulation data and technical documents in the field of industrial design to construct a semantic embedding model dedicated to design tasks, thereby improving 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 integrates domain knowledge graph nodes and design task text features to achieve automatic intention classification of design tasks and precise mapping of subtasks; the Prompt-to-Code enhanced 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 uses design parameters and performance simulation indicators to construct feature embedding, and combines the graph attention mechanism to clearly characterize the nonlinear interactive relationship between parameter differences and performance changes between design versions.

[0066] 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 automobile parts manufacturing.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] The present invention can perform automated 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 degree of automation of industrial design tasks, the domain semantic consistency of design solutions, the accuracy of simulation and analysis, and the task generalization capability, reducing dependence on data scale, and significantly improving the efficiency and quality of complex engineering design. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a flowchart of a multi-agent collaborative industrial design method for complex engineering projects.

[0070] Figure 2 This is a flowchart of the pre-training method in a multi-agent collaborative industrial design method for complex engineering.

[0071] Figure 3 This is a flowchart of the task decomposition algorithm in a multi-agent collaborative industrial design method for complex engineering.

[0072] Figure 4 Flowchart for enhancing model building in a multi-agent collaborative industrial design approach for complex engineering projects.

[0073] Figure 5 A flowchart for constructing a version evolution tree in a multi-agent collaborative industrial design method for complex engineering projects.

[0074] Figure 6 This is a flowchart of adaptive optimization in a multi-agent collaborative industrial design method for complex engineering projects.

[0075] Figure 7 This is a structural diagram of a multi-agent collaborative industrial design system for complex engineering projects. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention is 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 intended to limit the present invention.

[0077] The specific implementation of the present invention is described in detail below with reference to specific embodiments.

[0078] like Figure 1 As shown, an embodiment of the present invention provides a multi-agent collaborative industrial design method for complex engineering, the method comprising the following steps:

[0079] S100 builds a domain-specific large-scale agent collection for complex engineering design tasks, including a structural design agent, a material selection agent, a simulation analysis agent, and a review agent. It uses domain-specific pre-training methods to establish a prompt vocabulary and task constraint knowledge base dedicated to the design domain.

[0080] S200, a task decomposition algorithm based on the design domain knowledge graph and intent recognition mechanism, divides complex design tasks into agent-executable subtasks and adaptively adjusts task weights through a graph attention network to form an asynchronous task planning strategy;

[0081] S300 builds a prompt-to-code enhancement model for design-related tasks. It converts the multi-agent task planning goals and constraints into CAD modeling or CAE simulation design instructions that can be directly executed by industrial design tools through prompt embedding.

[0082] S400 builds a design version evolution tree based on a causal reasoning graph structure of the interaction between design target features. It uses an attention mechanism to learn the nonlinear interaction paths between design features, dynamically updates the version path of the design solution, and performs traceability.

[0083] S500 inputs the asynchronous task planning strategy, design instructions, and version evolution tree feedback results into the large model agent for iterative adaptive optimization until the objectives and constraints of the complex engineering design task are met.

[0084] like Figure 2 As shown, as a preferred embodiment of the present invention, the pre-training method includes:

[0085] S101, collects resource data for industrial design, including design drawings, simulation data, and technical documentation; Design drawings: collect task-related design drawings from the industrial design field, which usually contain detailed structural design, dimensions, proportions, and other information. Simulation data: usually includes simulation results of different design schemes under different conditions, such as stress, temperature, kinematics, and other simulation indicators. Technical documentation: includes standards, regulations, design guidelines, material data manuals, etc.

[0086] S102, self-supervised pre-training: Through a learning method that constructs training signals from the input data itself, self-supervised learning is used to improve semantic understanding and planning accuracy in the absence of a large amount of labeled data, namely: first, a pre-training task is constructed, including a drawing parsing task, a simulation data embedding task, and a document understanding task; that is, the design drawings are self-supervised through the parsing task, the structural elements in the drawings are matched with the standard component database, and labels are generated to learn the semantic features of the design drawings; the simulation data is converted into a vector representation for training the model to learn the simulation pattern through the similarity or difference between the data; the technical document training model is used to understand the professional terms and technical requirements in the design task, and a semantic embedding model is established.

[0087] Next, a semantic embedding model is established to convert design drawings, simulation data, and technical documents into vector representations (i.e., "embeddings") that can be used for calculations. Different embedding techniques are used for each data type (drawings, simulation data, documents). For design drawings, commonly used techniques include convolutional neural networks to extract visual features and then mapping these features to the embedding space. For simulation data, regression or classification models are used to convert it into vector representations. For technical documents, natural language processing technologies such as BERT, GPT, and other models are used for text embedding.

[0088] Finally, the self-supervised pre-training process uses unlabeled design drawings, simulation data, and documents for self-supervised training, and trains 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 uses a self-supervised learning algorithm to enhance prediction accuracy, backpropagate the model's predicted parameters, and optimize the mapping between design drawings and simulation data.

[0089] 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 , the loss function of the model for:

[0090] ;

[0091] in, 、 and is the weight coefficient for adjusting the loss of each part, 、 and is the target embedding vector, representing the true semantic features of the design drawings, simulation data, and documents respectively. Through the back-propagation algorithm, the loss function is optimized and the parameters of the embedding model are updated. Through the multimodal fusion strategy of self-supervised learning, the correlation between drawings, simulation data, and documents is learned, so that the generated semantic embedding can integrate information from various data sources, improving task understanding and planning accuracy.

[0092] S104, Model Evaluation: Evaluate the performance of the trained model in industrial design tasks through validation sets or cross-validation, and evaluate the model's language comprehension ability, planning accuracy of design tasks, and ability to integrate multimodal information.

[0093] In an embodiment of the present invention, the steps of constructing the structural design agent, the material selection agent, the simulation analysis agent, and the review agent include:

[0094] Structural design agent: Based on design requirements, it automatically or semi-automatically selects key parameters of the structure (size, shape, geometric features, etc.), generates or modifies the 3D morphological model of the design scheme (automatically generates the model through the CAD software API). The structural design agent performs parameter selection and morphological optimization through the target optimization function, which is: ,in, is the design parameter set, represents the objective function of structural design (lightest weight, maximum strength, etc.), Represents the constraint space of structural parameters (manufacturing constraints, size constraints, etc.);

[0095] Material selection agent: Identifies material properties (strength, density, elastic modulus, etc.) that match design requirements from the material knowledge base, selects appropriate materials based on performance and constraints, and provides recommendations for design decisions. The material selection recommendation model is based on performance index optimization. The material selection index function is:

[0096] ;

[0097] in, Represents a material in a material collection; Represents a collection of material databases; 、 and Respectively represent the strength, weight and stability performance indicators of the material; represents the material cost function; 、 and Represent the weight parameters of different performance indicators respectively;

[0098] 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 design performance 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: ,in, represents the structural stiffness matrix; represents the node displacement vector; Represents the load vector; simulation result performance evaluation formula, such as safety factor: ,in, represents the safety factor; Indicates the allowable stress of the material; represents the maximum stress obtained in the simulation;

[0099] Review Agent: Receives the output of the structural design agent, material selection agent, and simulation analysis agent. Based on the design specifications, standards, and task constraints, it evaluates the quality and compliance of the design solution, and makes judgments on whether the requirements are met and provides suggestions for improvement. The review involves comprehensive scoring or multi-index evaluation. The comprehensive evaluation indicators of the design solution are: ,in, Indicates the comprehensive quality score of the design scheme; Indicates that the scheme Scores on each evaluation dimension (such as safety, economy, and manufacturability); Indicates the weight of the corresponding dimension; Indicates the total number of dimensions reviewed;

[0100] Compliance judgment: When , the design plan passes the review; otherwise, the agent optimization plan is returned, where Represents the minimum requirements for design quality or compliance. The four agents described above do not work independently, but rather interact and collaborate through shared data: 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 → Re-adjustment of Design → Further Optimization by the Structural Design Agent and Material Selection Agent.

[0101] like Figure 3 As shown in FIG, as a preferred embodiment of the present invention, the task decomposition algorithm of the intent recognition mechanism realizes the automatic intent classification and subtask mapping of the design task through a dual-channel Transformer intent recognition module that integrates the domain knowledge graph nodes and the design task text features. The specific steps are as follows:

[0102] S201, Data Preprocessing and Feature Extraction: Extract text features from the design task text (natural language) to obtain a text sequence vector representation, extract task-related node features (entities, relationships, attributes, etc.) from the domain knowledge graph, and obtain a graph node feature vector representation; Text features: , graph node features: ,in, Indicates the text embedding vectors, Indicates the graph node embedding vectors;

[0103] S202, build a dual-channel Transformer module: build a text channel to capture the dependency between text sequence features; build a knowledge graph channel to capture the semantic relationship between graph node features; encode text features and graph node features respectively through a multi-head self-attention mechanism. The multi-head attention calculation is: ,in, 、 、 Represent the query, key, and value in the attention mechanism respectively; is the dimension of each attention head; Transformer encoding is applied to text and graph node features respectively: text channel encoding, ; Spectral channel coding, ;

[0104] 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 implement feature fusion through the attention mechanism: , the fusion attention formula can be expressed as:

[0105] ;

[0106] in, , , represents the learnable parameters in the fusion attention mechanism, is the dimension of fusion features;

[0107] S204, Intent Classification and Subtask Mapping: Intent classification is performed based on the fused feature vector, and the corresponding subtasks are automatically mapped according to the classification results. Intent classification is achieved through the fully connected layer and the Softmax function: ,in, is the intent classification result, Represents the probability distribution of task types; and is the classification layer parameter; subtask mapping rule: ,in, It is a task to subtask mapping rule that automatically selects subtasks from the task knowledge base based on the classification results;

[0108] S205, output task decomposition results: Output the subtask set after intention recognition, forming the basis for subsequent multi-agent collaborative execution, and finally output: .

[0109] like Figure 4As shown in FIG, as a preferred embodiment of the present invention, the Prompt-to-Code enhancement model includes a cross-modal Prompt encoding mechanism. By co-optimizing the embedding of design task text and design domain-specific DSL, the domain semantic consistency of design instruction generation and the accuracy of CAD / CAE execution are improved. The specific steps are as follows:

[0110] 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. ,in, is the text feature, Embed features for text;

[0111] S302, Design Domain Specific DSL Feature Extraction: Extract feature vectors from the design domain specific DSL code to reflect the semantic correspondence between DSL instructions and tasks. This is usually done using code embedding models such as CodeBERT and CodeT5, namely: ,in, is a domain-specific DSL instruction sequence, Embed feature vector for DSL;

[0112] S303, Cross-modal Prompt Encoding Mechanism (Feature Fusion): A cross-modal attention mechanism is used to achieve the fusion and collaborative optimization of text embedding and DSL embedding, so that the two share the same semantic space, namely:

[0113] ;

[0114] in, , , is the learnable parameter matrix of cross-modal attention, is the fused cross-modal feature embedding vector, is the embedding vector dimension;

[0115] S304, Prompt-to-Code Generation Module (Instruction Generation): Utilizes the fused cross-modal feature vectors to generate DSL instruction codes executable by domain-specific CAD / CAE tools. This is done using a Seq2Seq model. ,in, is the generated instruction sequence, The first instruction in the sequence DSL directive elements, represents the probability of generating the current instruction element given the fused embedding vector and the previous instruction element;

[0116] S305, domain semantic consistency optimization: ensure the semantic consistency of the generated instruction sequence and the design task text through the domain semantic consistency loss function, ,in, To generate semantic embedding vectors for instruction sequences; 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 instructions and the text description;

[0117] S306, optimization objective function: train the overall Prompt-to-Code enhancement model, jointly optimize generation quality and domain semantic consistency, and the comprehensive optimization formula is: ,in, The cross entropy loss for the instruction generation task, is the domain semantic consistency loss, To adjust the hyperparameters of semantic consistency loss weight;

[0118] S307, Generate Result Output and Execute: After the optimization is completed, the domain-specific DSL instructions are output and passed to industrial design tools (such as CAD / CAE software) for automated execution: ,in, The final optimized domain-specific CAD / CAE execution instructions.

[0119] like Figure 5 As shown in FIG. 1 , 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 composed of design parameters and performance simulation indicators. The nonlinear interactive relationship between parameter differences and performance changes between different versions is clearly characterized through the graph attention mechanism. The specific steps are as follows:

[0120] S401, design version node definition and feature construction: Each design version is defined as a node in the graph structure. The node features include two aspects: design parameter features (used to represent dimensions, material properties, geometric parameters, etc.) and performance simulation index characteristics (representing CAE simulation results such as strength, weight, stress, and vibration); node feature embedding is: ,in, For the The complete feature embedding of the version node, It is the feature splicing operation;

[0121] S402, version evolution tree construction: Construct a directed tree structure based on the chronological order or causal relationship of the design versions. The edges represent the evolution or causal dependency between versions. The initial root node is the initial design version, and each branch represents a different evolution path. Graph structure definition: ,in, is a set of nodes, representing each design version; is a set of edges, representing the evolutionary relationship between versions;

[0122] S403, feature interaction analysis of the graph attention network mechanism: Through the graph attention mechanism, the nonlinear interaction relationship between design parameters and performance changes between learning version nodes is learned. The graph attention mechanism calculates the node feature embedding update. Its graph attention calculation formula is:

[0123] ;

[0124] in, For nodes For Node The attention coefficient reflects the feature interaction weight between nodes; is the weight matrix of feature linear transformation; Learning parameters for the attention mechanism; and For nodes and The feature embedding vector of For splicing operation; For nodes The neighbor node set of ; node feature update formula: ,in, is the updated node feature vector, is a nonlinear activation function (such as ReLU);

[0125] S404, version feature difference and performance change analysis: Based on the updated node features, clearly characterize the impact path of design parameter changes between different versions on performance indicators, and then calculate the feature difference and performance change metrics between nodes; To Node The parameter and performance differences of the version changes are expressed as: ,in, For version With version Euclidean distance between feature difference vectors; performance index change significance evaluation: ,in, is the impact degree of performance changes between nodes;

[0126] S405, Dynamic Version Path Update and Traceability Management: Based on the node characteristics and difference impact relationships output by GAT, the version evolution path is dynamically updated, providing version backtracking capabilities to facilitate subsequent optimization and fault analysis; path weights are defined based on the impact weights between versions: ,in, The weight of the version evolution path is determined by comprehensively considering the attention weight and performance changes; dynamic path determination: ,in, is the best (most significant impact) version evolution path.

[0127] like Figure 6 As shown, 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, and the specific steps are as follows:

[0128] S501, Data Preparation and Preference Annotation: Collect different design solutions automatically generated by the agent, perform pairwise comparison annotations on different solutions, clarify the preferences of design experts or fields, and obtain a preference dataset. Given two design solutions Label preference relations, Representation scheme Better than the solution ;

[0129] S502, Design Domain Preference Fine-tuning Model Construction: Use LoRA (Low Rank Adaptation) technology to adjust the parameters of large models, reduce the demand for training samples and improve generalization ability. LoRA adjusts the weight matrix Perform low-rank decomposition as: ,in, is the fixed weight matrix of the original pre-trained model; 、 is a trainable low-rank matrix , effectively reducing the number of parameters;

[0130] S503, direct preference optimization method implementation: using the DPO method, by clarifying the preference relationship, optimizing the model preference prediction probability distribution, for a given solution The preference optimization loss function is: ,in, Model pair solution the preference score given; is the Sigmoid function, is the temperature parameter, controlling the sensitivity of the gap between preference scores; It is a pairwise labeled preference dataset;

[0131] S504, Joint Optimization Loss Function Construction: Construct a joint loss function that integrates the parameter constraints of LoRA fine-tuning and the preference loss function of DPO, and optimizes the overall goal: ,in, is the loss function of the overall training; is the regularization term of the LoRA parameter, used to prevent overfitting; is the LoRA parameter regularization term weight hyperparameter;

[0132] S505, iterative optimization and model update: Repeat the optimization steps S501-S504. The model iteratively updates parameters based on the latest preference data, gradually improving the model's generalization ability. Parameters are iteratively updated: ,in, For the Model parameters for the iteration; is the learning rate; is the gradient;

[0133] S506, Adaptive Termination Criteria and Generalization Evaluation: Monitor the generalization performance of the model (such as performance in the validation set or small sample tasks), and terminate the optimization process when the performance meets the predetermined requirements or the generalization ability is stable.

[0134] like Figure 7 As shown, an embodiment of the present invention further provides a multi-agent collaborative industrial design system for complex engineering projects, the system comprising:

[0135] The large-model agent collection module 100 is used to construct a large-model agent collection for complex engineering design tasks, including a structural design agent, a material selection agent, a simulation analysis agent, and a review agent. The prompt vocabulary and task constraint knowledge base are established through pre-training methods.

[0136] The task decomposition and planning module 200 is used to divide the design task into subtasks that can be executed by the agent based on the task decomposition algorithm based on the knowledge graph and the intention recognition mechanism, and to achieve adaptive adjustment of task weights through the graph attention network to form an asynchronous task planning strategy;

[0137] The Prompt-to-Code enhanced conversion module 300 is used to build a Prompt-to-Code enhanced model, which converts the multi-agent task planning goals and constraints into CAD modeling or CAE simulation design instructions that can be directly executed by industrial design tools through prompt embedding;

[0138] The design version evolution and management module 400 is used to construct a design version evolution tree. Based on the causal reasoning graph structure of the interaction between design target features, it uses the attention mechanism to learn the nonlinear interaction path between design features, dynamically update the version path of the design solution, and conduct traceability.

[0139] The iterative adaptive optimization module 500 is used to input the asynchronous task planning strategy, design instructions and version evolution tree feedback results into the large model intelligent agent for iterative adaptive optimization until the objectives and constraints of the complex engineering design task are met.

[0140] In an embodiment of the present invention, the large-model intelligent agent collection module 100 includes a pre-training submodule, which uses design drawings, simulation data and technical documents in the field of industrial design to construct a semantic embedding model dedicated to the design task, thereby improving the agent's language understanding and task planning accuracy; the task decomposition and planning module 200 includes an intent recognition submodule, which uses a dual-channel Transformer model that integrates domain knowledge graph nodes and design task text features to achieve automatic intent classification of design tasks and precise mapping of subtasks; the Prompt-to-Code enhanced conversion module 300 includes a cross-modal Prompt encoding submodule, 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 submodule, which uses design parameters and performance simulation indicators to construct feature embedding, and combines the graph attention mechanism to clearly characterize the nonlinear interactive relationship between parameter differences and performance changes between design versions.

[0141] In an embodiment 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 automobile parts manufacturing.

[0142] The above is only a detailed description of the preferred embodiments of the present invention, which is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0143] It should be understood that, although the various steps in the flow chart of each embodiment of the present invention are shown in sequence according to the indication of the arrows, these steps are not necessarily performed in sequence according to the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in order, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0144] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When executed, the program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may 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 (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0145] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the disclosure in the specification and examples. 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 common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered merely as exemplary, and the true scope and spirit of the present disclosure are indicated by the claims.

Claims

1. A multi-agent collaborative industrial design method for complex engineering projects, characterized by: The method comprises the following steps: S100 builds a large-scale intelligent agent collection for complex engineering design tasks, including structural design agents, material selection agents, simulation analysis agents, and review agents. It establishes a prompt vocabulary and task constraint knowledge base through pre-training methods. S200, a task decomposition algorithm based on knowledge graphs and intent recognition mechanisms, divides design tasks into agent-executable subtasks and adaptively adjusts task weights through a graph attention network to form an asynchronous task planning strategy; S300 builds a Prompt-to-Code enhancement model that converts multi-agent task planning goals and constraints into CAD modeling or CAE simulation design instructions that can be directly executed by industrial design tools through prompt embedding; S400 builds a design version evolution tree based on a causal reasoning graph structure of the interaction between design target features. It uses an attention mechanism to learn the nonlinear interaction paths between design features, dynamically updates the version path of the design solution, and performs traceability. S500 inputs the asynchronous task planning strategy, design instructions, and version evolution tree feedback results into the large model agent for iterative adaptive optimization until the objectives and constraints of the complex engineering design task are met.

2. The multi-agent collaborative industrial design method for complex engineering according to claim 1 is characterized in that: The pre-training method comprises: S101, collect resource data for industrial design, including design drawings, simulation data and technical documents; S102: Construct pre-training tasks, including drawing parsing, simulation data embedding, and document understanding; establish a semantic embedding model to convert design drawings, simulation data, and technical documents into vector representations for computation; and perform self-supervised pre-training. S103, determine the semantic embedding objective function: the embedding vectors of design drawings, simulation data and documents are 、 and , the loss function of the model for: ； in, 、 and is the weight coefficient for adjusting the loss of each part, 、 and is the target embedding vector, representing the true semantic features of the design drawing, simulation data, and document respectively. Through the back-propagation algorithm, the loss function is optimized and the parameters of the embedding model are updated. Through the multimodal fusion strategy of self-supervised learning, the correlation between the drawings, simulation data, and documents is learned, so that the generated semantic embedding can integrate information from various data sources. S104, Model Evaluation: Evaluate the performance of the trained model in industrial design tasks through validation sets or cross-validation, and evaluate the model's language comprehension ability, planning accuracy of design tasks, and ability to integrate multimodal information.

3. The multi-agent collaborative industrial design method for complex engineering according to claim 1 is characterized in that: The steps of constructing the structural design agent, material selection agent, simulation analysis agent and review agent include: Structural design agent: Based on design requirements, it automatically or semi-automatically selects key parameters of the structure and generates or modifies the 3D morphological modeling of the design scheme. The structural design agent performs parameter selection and morphological optimization through the target optimization function, which is: ,in, is the design parameter set, represents the objective function of structural design, represents the constraint space of structural parameters; Material selection agent: Identifies material properties that match design requirements from the material knowledge base, selects appropriate materials based on performance and constraints, and provides recommendations for design decisions. The material selection recommendation model is based on performance index optimization. The material selection index function is: ； in, Represents a material in a material collection; Represents a collection of material databases; 、 and Respectively represent the strength, weight and stability performance indicators of the material; represents the material cost function; 、 and Respectively represent the weight parameters of different performance indicators; Simulation Analysis Agent: This agent receives the model generated by the Structural Design Agent and the material parameters recommended by the Material Selection Agent. It automatically performs simulation analysis based on CAE simulation tools to verify whether the performance of the design solution meets the design requirements and returns the simulation results. The simulation analysis process involves finite element analysis: ,in, represents the structural stiffness matrix; represents the node displacement vector; represents the load vector; Review Agent: Receives the output of the structural design agent, material selection agent, and simulation analysis agent. Based on the design specifications, standards, and task constraints, it evaluates the quality and compliance of the design solution, and makes judgments on whether the requirements are met and provides suggestions for improvement. The review involves comprehensive scoring or multi-index evaluation. The comprehensive evaluation indicators of the design solution are: ,in, Indicates the comprehensive quality score of the design scheme; Indicates that the scheme The scores on each evaluation dimension; Indicates the weight of the corresponding dimension; Indicates the total number of dimensions reviewed; Compliance judgment: When , the design plan passes the review; otherwise, the agent optimization plan is returned, where Indicates the minimum requirements for design quality or compliance.

4. The multi-agent collaborative industrial design method for complex engineering according to claim 1, characterized in that: The task decomposition algorithm of the intent recognition mechanism achieves automatic intent classification and subtask mapping of design tasks through a dual-channel Transformer intent recognition module that integrates 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, extract task-related node features from the domain knowledge graph to obtain a graph node feature vector representation; Text features: , graph node features: ,in, Indicates the text embedding vectors, Indicates the graph node embedding vectors; S202, build a dual-channel Transformer module: build a text channel to capture the dependency between text sequence features; build a knowledge graph channel to capture the semantic relationship between graph node features; encode text features and graph node features respectively through a multi-head self-attention mechanism. The multi-head attention calculation is: ,in, 、 、 Represent the query, key, and value in the attention mechanism respectively; is the dimension of each attention head; Transformer encoding is applied to text and graph node features respectively: text channel encoding, ; Spectral channel coding, ; 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 implement feature fusion through the attention mechanism: , the fusion attention formula can be expressed as: ； in, , , represents the learnable parameters in the fusion attention mechanism, is the dimension of fusion features; S204, Intent Classification and Subtask Mapping: Intent classification is performed based on the fused feature vector, and the corresponding subtasks are automatically mapped according to the classification results. Intent classification is achieved through the fully connected layer and the Softmax function: ,in, is the intent classification result, Represents the probability distribution of task types; and is the classification layer parameter; subtask mapping rule: ,in, It is a task to subtask mapping rule that automatically selects subtasks from the task knowledge base based on the classification results; S205, output task decomposition results: Output the subtask set after intention recognition, forming 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 4 is characterized in that: The proposed Prompt-to-Code enhancement model includes a cross-modal prompt encoding mechanism that improves the domain semantic consistency of design instruction generation and the accuracy of CAD / CAE execution by co-optimizing the embedding of design task text and design domain-specific DSL. 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. ,in, Embed features for text; S302, design domain specific DSL feature extraction: extracting feature vectors from the design domain specific DSL code to reflect the semantic correspondence between DSL instructions and tasks; S303, Cross-modal Prompt Encoding Mechanism: A cross-modal attention mechanism is used to achieve the fusion and collaborative optimization of text embedding and DSL embedding, so that the two share the same semantic space, namely: ； in, , , is the learnable parameter matrix of cross-modal attention, is the fused cross-modal feature embedding vector, is the embedding vector dimension; S304, Prompt-to-Code Generation Module: Utilizes the fused cross-modal feature vectors to generate DSL instruction code executable by domain-specific CAD / CAE tools. This is done using a Seq2Seq model. ,in, is the generated instruction sequence, The first instruction in the sequence DSL directive elements, represents the probability of generating the current instruction element given the fused embedding vector and the previous instruction element; S305, domain semantic consistency optimization: ensure the semantic consistency of the generated instruction sequence and the design task text through the domain semantic consistency loss function, ,in, To generate semantic embedding vectors for instruction sequences; 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 instructions and the text description; S306, optimization objective function: train the overall Prompt-to-Code enhancement model, jointly optimize generation quality and domain semantic consistency, and the comprehensive optimization formula is: ,in, The cross entropy loss for the instruction generation task, is the domain semantic consistency loss, To adjust the hyperparameters of semantic consistency loss weight; S307, generate result output and execute: After the optimization is completed, the domain-specific DSL instructions are output and passed to the industrial design tool for automated execution: ,in, The final optimized domain-specific CAD / CAE execution instructions.

6. The multi-agent collaborative industrial design method for complex engineering according to claim 1, characterized in that: In the design version evolution tree, the feature embedding of the design version node is composed of design parameters and performance simulation indicators. The nonlinear interactive 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: Each design version is defined as a node in the graph structure. The node features include two aspects: design parameter features and performance simulation index characteristics ; The node feature embedding is: ,in, For the The complete feature embedding of the version node, It is the feature splicing operation; S402, version evolution tree construction: Construct a directed tree structure based on the chronological order or causal relationship of the design versions. The edges represent the evolution or causal dependency between versions. The initial root node is the initial design version, and each branch represents a different evolution path. Graph structure definition: ,in, 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 the graph attention network mechanism: Through the graph attention mechanism, the nonlinear interaction relationship between design parameters and performance changes between learning version nodes is learned. The graph attention mechanism calculates the node feature embedding update. Its graph attention calculation formula is: ； in, For nodes For Node The attention coefficient reflects the feature interaction weight between nodes; is the weight matrix of feature linear transformation; Learning parameters for the attention mechanism; and For nodes and The feature embedding vector of For splicing operation; For nodes The neighbor node set of ; node feature update formula: ,in, is the updated node feature vector, is a nonlinear activation function; S404, version feature difference and performance change analysis: Based on the updated node features, clearly characterize the impact path of design parameter changes between different versions on performance indicators, and then calculate the feature difference and performance change metrics between nodes; To Node The parameter and performance differences of the version changes are expressed as: ,in, For version With version Euclidean distance between feature difference vectors; performance index change significance evaluation: ,in, is the impact degree of performance changes between nodes; S405, Dynamic Version Path Update and Traceability Management: Based on the node characteristics and difference impact relationships output by GAT, the version evolution path is dynamically updated, providing version backtracking capabilities to facilitate subsequent optimization and fault analysis; path weights are defined based on the impact weights between versions: ,in, The weight of the version evolution path is determined by comprehensively considering the attention weight and performance changes; dynamic path determination: ,in, The best 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, a direct preference optimization method and a LoRA parameter efficient adjustment strategy are used. The specific steps are as follows: S501, Data Preparation and Preference Annotation: Collect different design solutions automatically generated by the agent, perform pairwise comparison annotations on different solutions, clarify the preferences of design experts or fields, and obtain a preference dataset. Given two design solutions Label preference relations; S502, design domain preference fine-tuning model construction: use LoRA technology to adjust the parameters of large models, reduce the demand for training samples and improve generalization ability, LoRA weight matrix Perform low-rank decomposition as: ,in, is the fixed weight matrix of the original pre-trained model; 、 is a trainable low-rank matrix; S503, direct preference optimization method implementation: using the DPO method, by clarifying the preference relationship, optimizing the model preference prediction probability distribution, for a given solution The preference optimization loss function is: ,in, Model pair solution the preference score given; is the Sigmoid function, is the temperature parameter, controlling the sensitivity of the gap between preference scores; It is a pairwise labeled preference dataset; S504, Joint Optimization Loss Function Construction: Construct a joint loss function that integrates the parameter constraints of LoRA fine-tuning and the preference loss function of DPO, and optimizes the overall goal: ,in, is the loss function of the overall training; is the regularization term of the LoRA parameter, used to prevent overfitting; is the LoRA parameter regularization term weight hyperparameter; S505, iterative optimization and model update: Repeat the optimization steps S501-S504. The model iteratively updates parameters based on the latest preference data, gradually improving the model's generalization ability. Parameters are iteratively updated: ,in, For the Model parameters for the iteration; is the learning rate; is the gradient; S506, Adaptive Termination Criteria and Generalization Evaluation: Monitor the generalization performance of the model and terminate the optimization process when the performance reaches the predetermined requirements or the generalization ability is stable.

8. A multi-agent collaborative industrial design system for complex engineering projects, characterized by: The system comprises: The large-model agent collection module is used to build a large-model agent collection for complex engineering design tasks, including structural design agents, material selection agents, simulation analysis agents, and review agents. The prompt vocabulary and task constraint knowledge base are established through pre-training methods. The task decomposition and planning module is used to implement a task decomposition algorithm based on knowledge graphs and intent recognition mechanisms, dividing design tasks into subtasks that can be executed by agents. It also uses a graph attention network to achieve adaptive adjustment of task weights and form an asynchronous task planning strategy. The Prompt-to-Code enhanced conversion module is used to build a Prompt-to-Code enhanced model, which converts the multi-agent task planning goals and constraints into CAD modeling or CAE simulation design instructions that can be directly executed by industrial design tools through prompt embedding; The design version evolution and management module is used to build a design version evolution tree. Based on the causal reasoning graph structure of the interaction between design target features, it uses the attention mechanism to learn the nonlinear interaction path between design features, dynamically update the version path of the design solution, and conduct traceability. The iterative adaptive optimization module is used to input asynchronous task planning strategies, design instructions, and version evolution tree feedback results into the large model agent for iterative adaptive optimization until the objectives and constraints 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 agent collection module includes a pre-training sub-module, which uses design drawings, simulation data and technical documents in the field of industrial design to build a semantic embedding model dedicated to design tasks, thereby improving the agent's language understanding and task planning accuracy. The task decomposition and planning module includes an intent recognition submodule, which uses a dual-channel Transformer model that integrates domain knowledge graph nodes and design task text features to achieve automatic intent classification of design tasks and precise mapping of subtasks; the Prompt-to-Code enhanced conversion module includes a cross-modal Prompt encoding submodule, 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 submodule, which uses design parameters and performance simulation indicators to construct feature embedding, and combines the graph attention mechanism to clearly characterize the nonlinear interactive relationship between parameter differences and performance changes between design versions.

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

Images