A model cluster-based intelligent design method and system for complex systems

Abstract

The application belongs to the field of system engineering, and particularly discloses a complex system intelligent design method and system based on model clusters, which comprises the following steps: determining design features of a system to be designed, wherein the design features comprise architecture features, component features and parameter features; searching for the most similar information in a system design historical knowledge base based on the design features by using a search enhancement generation method; further generating generalization rules of system design by a large language model based on the most similar information; generating a model cluster based on the generalization rules, wherein the model cluster comprises multiple design models; modeling the optimal design model in the model cluster as a Markov decision process, taking the design model simulation performance index as a reward function, and obtaining the optimal design model through reinforcement learning. The application can realize more efficient, comprehensive and accurate complex system scheme design.

Description

Technical Field

[0001] This invention belongs to the field of systems engineering, and more specifically, relates to an intelligent design method and system for complex systems based on model clusters. Background Technology

[0002] In the design of complex systems such as aircraft, rockets, and automobiles, design verification and evaluation models currently face urgent transformation needs and challenges: 1) In terms of scheme design, the design model, which heavily relies on the designer's experience, produces very limited feasible solutions, making it difficult to cover a more comprehensive design exploration space, resulting in schemes only achieving optimality within a small local design space. 2) In terms of scheme verification, the verification model, which is mainly based on physical experiments, has a long cycle, high cost, and is difficult to cover extreme operating conditions, while large-sample simulation experiments face challenges such as large modeling workload and long overall simulation time, making it difficult to achieve a balance between comprehensiveness and efficiency in scheme verification. 3) In terms of scheme evaluation, due to the large number and variety of evaluation indicators, a single evaluation algorithm is difficult to apply to the high-precision and high-efficiency evaluation of all indicators, and manually selecting evaluation algorithms makes it difficult for scheme evaluation work to achieve comprehensive and accurate confirmation of design indicators.

[0003] Therefore, in response to the changing needs of complex system development models and the challenges faced by its design verification technology, there is an urgent need for a more efficient, high-quality, and comprehensive method for designing, verifying, and evaluating complex systems. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method and system for intelligent design of complex systems based on model clusters, the purpose of which is to achieve more efficient, comprehensive and accurate design of complex system solutions.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for intelligent design of complex systems based on model clusters is proposed, comprising the following steps:

[0006] Determine the design characteristics of the system to be designed, including architectural characteristics, component characteristics, and parameter characteristics;

[0007] Based on design features, a retrieval-enhanced generation method is used to retrieve the most similar information from the system design history knowledge base; then, based on the most similar information, a large language model is used to generate generalization rules for the system design.

[0008] A model cluster is generated based on generalization rules, which includes multiple design models;

[0009] The optimal design model will be selected from the model cluster and modeled as a Markov decision process. The simulation performance index of the design model will be used as the reward function, and the optimal design model will be obtained through reinforcement learning.

[0010] As a further preferred embodiment, the architectural features include the system's architectural composition, connection relationships, and interface types; the component features include the specific models of each component in the architecture; and the parameter features include the key design parameters and performance indicators of the components.

[0011] As a further optimization, based on design features, a retrieval-enhanced generation method is used to retrieve the most similar information from the system design history knowledge base, including:

[0012] A system design history knowledge base is formed by preprocessing and vectorizing historical system design data in advance.

[0013] The design features are vectorized, and based on the vectorized design features, the most similar information is retrieved from the system design history knowledge base using a retrieval enhancement generation method.

[0014] As a further preferred option, a cluster of models generated based on generalization rules includes:

[0015] The generalization rules for system design include architecture generalization generation rules, component generalization generation rules, and parameter generalization generation rules;

[0016] An architecture-level solution model is generated based on the architecture generalization generation rules; a component-level solution model is generated based on the architecture-level solution model and the component generalization generation rules; and a parameter-level solution model is generated based on the component-level solution model and the parameter generalization generation rules, thereby forming a design model.

[0017] As a further preferred approach, an architecture-level solution model is generated based on the architecture generalization generation rules, including:

[0018] The architecture topology is automatically generated based on the architecture generalization generation rules. This architecture topology is expressed using SysML. Then, the architecture topology expressed in SysML is mapped and converted into the architecture topology expressed in Modelica to obtain the architecture-level solution model.

[0019] As a further preferred approach, the A3C network is adopted as the decision network for the Markov decision process, i.e., the global network. During the reinforcement learning process, each design model scheme in the model cluster corresponds to a simulation task to obtain simulation performance indicators. A multi-threaded parallel computing method is adopted to evenly partition and distribute the simulation tasks corresponding to the model cluster to each working thread. Each working thread copies the global network parameters as a local model copy. Each working thread independently performs reinforcement learning and updates the global network parameters, which are then distributed to other working threads.

[0020] As a further preferred method, the state space determination method for a Markov decision process is as follows:

[0021] By using word embedding technology and the pre-trained language model BERT, the semantic information of natural language text in the model cluster is analyzed. The numerical parameter features in the model cluster are processed by neural networks to realize the multi-dimensional representation of the structure, semantics and parameters of the designed model, forming a state space.

[0022] As a further preferred option, the following steps are also included:

[0023] Construct an initial evaluation indicator tree;

[0024] Based on the initial evaluation index tree, a retrieval-enhanced generation method is used to search the historical model evaluation index vector knowledge base to obtain the most similar information. This most similar information and rules are then combined to generate prompt words. A large language model is used to generate index aggregation rules, automatically aggregating some evaluation indices in the evaluation index tree to form several index subtrees. The simulation results of the optimal design model are then associated with the indices in the evaluation index tree. The optimal design model is evaluated based on the evaluation index tree to obtain the evaluation results.

[0025] As a further optimization, the indicators aggregated in each indicator subtree adopt the same evaluation algorithm, and the evaluation algorithm is selected based on the large model.

[0026] According to another aspect of the present invention, a model cluster-based intelligent design system for complex systems is provided, including a processor, the processor being used to execute the above-described model cluster-based intelligent design method for complex systems.

[0027] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages:

[0028] 1. This invention integrates Model-Based Systems Engineering (MBSE) with the field of artificial intelligence. Through model cluster-driven, highly generalized design, it expands the exploration space in the solution design stage, thereby providing more innovative possibilities. Furthermore, in the solution verification stage, simulation verification enables efficient operation of the model cluster and rapid closed-loop verification of the solution, shortening the closed-loop iteration cycle. This invention can promote the exploration of more high-quality design solutions for complex systems in the early stages of development, achieving more efficient, comprehensive, and accurate integrated design verification.

[0029] 2. When acquiring model clusters based on artificial intelligence, RAG technology was introduced. By retrieving relevant information from external knowledge bases, it is ensured that the AI's answers are based on the latest and most accurate factual evidence, which significantly reduces the "illusion" phenomenon, that is, the problem of AI fabricating false information.

[0030] 3. In reinforcement learning, simulation performance metrics are used as the reward function, and a multi-threaded parallel computing method is adopted to achieve low-redundancy parallel simulation, thereby systematically improving simulation efficiency. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of a complex system intelligent design method based on model clusters, as described in an embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram illustrating the definition of highly generalizable design features in an embodiment of the present invention.

[0033] Figure 3 This is a schematic diagram illustrating the training of high generalization design rules in an embodiment of the present invention.

[0034] Figure 4 This is a schematic diagram of the automatic generalization generation method for model clusters in an embodiment of the present invention.

[0035] Figure 5 This is a schematic diagram of the dynamic low-redundancy simulation method according to an embodiment of the present invention.

[0036] Figure 6 This is a schematic diagram of the parallel simulation method for model clusters in an embodiment of the present invention.

[0037] Figure 7 This is a schematic diagram of the low-redundancy parallel fusion simulation method according to an embodiment of the present invention.

[0038] Figure 8 This is a schematic diagram of the multi-dimensional fusion method for evaluation indicators in an embodiment of the present invention.

[0039] Figure 9 This is a schematic diagram illustrating the adaptive switching of the evaluation algorithm in an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0041] This invention provides an intelligent design method for complex systems based on model clusters, such as... Figure 1 As shown, it includes the following steps:

[0042] S1. Model Cluster-Driven High Generalization Design: High generalization design feature definition to form system high generalization design features; high generalization design rule training based on machine learning to form system high generalization design rules; automatic generalization generation of model clusters based on the master model to output system design scheme model clusters.

[0043] S11. Definition of High Generalization Design Features: To achieve high generalization design in a solution, including architectural generalization, component generalization, and parameter generalization, it is first necessary to define the high generalization design features of the complex system. Define architectural features, component features, and parameter features according to different dimensions of architectural design, component selection, and parameter design. For example... Figure 2 As shown, the details are as follows:

[0044] Architecture characteristics definition: The core elements are the architectural composition, connection relationships, and interface types of a complex system. Taking the architecture characteristics of a solid rocket propulsion subsystem of an aircraft as an example, the aircraft architecture includes first-level, second-level, and third-level solid rocket propulsion subsystems. Each level of solid rocket propulsion subsystem can be composed of components such as a solid engine casing, ignition device, combustion chamber, nozzle, pressure sensor, and thrust termination device. Solid rocket propulsion subsystem architectures exhibit significant differences in fuel supply, thrust control, and structural layout.

[0045] Component feature definition: The core element is the specific model of each component in the architecture, which is represented by a library of replaceable components and combinations of replaceable components. Taking the component features of the solid rocket propulsion subsystem of an aircraft as an example, its combustion chamber and nozzle have a variety of models and specifications of components that can be used as replaceable components, and the use of different component models has a significant impact on the design results.

[0046] Parameter characteristic definition: The core elements are the key design parameters and performance indicators of the selected components, which are represented by key parameters and key parameter boundaries. Taking the parameter characteristics of the solid rocket motor subsystem of an aircraft as an example, parameters such as the characteristic velocity, standard burning rate, propellant density, and pressure index of the combustion chamber, the nozzle exit diameter and throat diameter, and the shell wall thickness, radius, and shell length directly affect the overall design performance of the aircraft.

[0047] S12. Training of High-Generalization Design Rules: Based on a clear understanding of high-generalization design characteristics, to achieve model cluster generation based on the master model, it is necessary to have architecture generalization design rules, component generalization selection rules, and parameter generalization definition rules to achieve automated derivation and ensure effective coverage of the design space. Retrieval-Augmented Generation (RAG) technology is used to train the high-generalization design rules, including steps such as knowledge source preprocessing, knowledge vectorization, feature vectorization, similarity retrieval, and generalization rule generation. For example... Figure 3 As shown, the details are as follows:

[0048] (1) Pre-build a system design history knowledge base, including:

[0049] Knowledge source preprocessing: Preprocessing of knowledge sources such as historical model architecture topology, shelf product component models, and industry standard parameter ranges, including text extraction and optical character recognition.

[0050] Knowledge vectorization: The preprocessed knowledge is vectorized to form a vector knowledge base of system architecture, components, and parameters, i.e., a system design history knowledge base.

[0051] (2) Feature vectorization: Feature vectorization of highly generalizable design features, including architectural features, component features, and parameter features.

[0052] (3) Similarity retrieval: By combining the system's high generalization design feature vector with the system's historical design knowledge base, similarity retrieval is performed to obtain the most similar information.

[0053] (4) Generalization rule generation: Based on the rule-generated prompt words, combined with the most similar information, generalization rules are generated through a large language model, including architecture generalization generation rules, component generalization generation rules and parameter generalization generation rules.

[0054] S13. Automatic Generalization Generation of Model Clusters: A model cluster is a series of generalized solution model representations, including multiple design models of a complex system. Automatic generalization generation of model clusters is achieved through model architecture generalization, model component generalization, and model parameter generalization. The generation of system model clusters represents the automatic generalization generation of a series of system solutions. For example... Figure 4 As shown, the details are as follows:

[0055] Architecture-level solution model generation: Based on architecture generalization generation rules, m architecture topologies are automatically generated, and their architectural components, connection relationships, and interface types are expressed using SysML models. To better correspond with product components, connections, and interfaces, the SysML-expressed architecture models are mapped and converted into Modelica-expressed architecture models. Based on SysML-Modelica mapping conversion technology, the m SysML architecture models A1-Am are mapped and converted into m Modelica architecture models A1-Am.

[0056] Component-level solution model generation: Taking one of the architecture-level solution models Ai as the baseline, based on the component generalization generation rules, n component-level solution models Ai.C1-Ai.Cn corresponding to the architecture design Ai are automatically generated through component selection, thus generating mn component-level solution models A1.C1-Am.Cn.

[0057] Parameter-level solution model generation: Taking one of the component-level solution models Ai.Cj as the baseline, based on the parameter generalization generation rule, s parameter-level solution models Ai.Cj.P1-Ai.Cj.Ps corresponding to the component selection Ai.Cj are automatically generated through parameter transformation, thus generating mns parameter-level solution models A1.C1.P1-Am.Cn.Ps.

[0058] S2. Combining reinforcement learning-based dynamic low-redundancy simulation and model cluster parallel simulation mechanism, the optimal design model is selected from the model cluster.

[0059] Dynamic low-redundancy simulation, such as Figure 5 As shown, a Markov decision process is used to select models. Based on reinforcement learning algorithms, during the simulation process, models that do not meet the design requirements are dynamically eliminated based on the feedback from the simulation results. This reduces redundant simulation time for ineffective models and transforms the solution of the entire set of model clusters into the solution of effective model clusters, improving overall simulation efficiency. For example, reinforcement learning can judge the effectiveness based on the simulation results. If the results are good, models with similar parameter values ​​are selected for further simulation; if the results are bad, models with significantly different parameter values ​​are selected for further simulation. Each simulation can find a locally optimal solution and eliminate locally suboptimal solutions. Details are as follows:

[0060] Feature extraction is performed on the model cluster to form a state space: word embedding technology and the pre-trained language model BERT (Bidirectional Encoder Representations from Transformers) are used to parse the semantic information of natural language text such as component names and parameter descriptions; neural networks are used to process numerical parameter features to achieve multi-dimensional representation of the model's structure, semantics, and parameters.

[0061] Global A3C Network Establishment: Establish a global A3C (Asynchronous Advantage Actor-Critic) network architecture, including setting the Actor network output scheme selection strategy, the Critic network evaluating the scheme state value, and establishing a global parameter sharing space.

[0062] Initialize the DQN surrogate model: Design a DQN (Deep Q-Network) surrogate model space exploration strategy, including a value network for real-time decision-making and a target value network for stable training; initialize network parameters and experience replay memory pool, and define the solver for model selection as the environment for reinforcement learning interaction; use the simulation performance index of each designed model as the reward function, and use Markov decision process to select models to achieve end-to-end training from features to decisions.

[0063] The proxy model explores in the state space: the state is the model feature vector, the current action is selected by the maximum Q value, and the memory unit is generated by combining the reward function value and stored in the replay memory pool.

[0064] Batch data is sampled from memory data: the error between the current value network's prediction and the target value network is calculated, network parameters are optimized, and the optimal selection strategy is iteratively learned in the complete set of efficient search models. The curse of dimensionality is mitigated through experience replay and target network techniques, avoiding local optima traps while ensuring exploration efficiency.

[0065] A closed-loop feedback mechanism for simulation results and optimization strategies is established: through reinforcement learning reward feedback and two-layer screening constraint feedback, a closed loop is formed encompassing feature extraction, strategy learning, screening optimization, and performance feedback, achieving low-redundancy screening in a large-scale model cluster. This outputs a complete set of effective model solutions and the optimal model solution within the current model cluster.

[0066] Furthermore, for parallel simulation of model clusters: a multi-threaded parallel computing method is adopted, and through optimized allocation of computing resources and real-time monitoring of abnormal states, a systematic improvement in simulation efficiency is achieved. For example... Figure 6 As shown, the details are as follows:

[0067] Requesting a simulation task: A simulation task is generated for each model scheme in the entire model cluster. The simulation task checks the current idle thread queue for thread resources in turn. If there are no idle threads, the simulation task is stored in the waiting queue. If there are available thread resources, the simulation task is assigned to the thread for simulation and the thread is removed from the idle thread queue.

[0068] Simulation variable parameter monitoring: Monitor all simulation process data of the selected variable at fixed time intervals; if the monitored variable parameter exceeds the limit, the simulation is terminated; if the monitored variable does not exceed the limit, monitoring continues until the simulation is completed.

[0069] The model cluster performs simulations in a loop: when a simulation task in a certain thread terminates, the thread is added back to the idle thread queue; the above steps are repeated until the entire model cluster has completed its simulation.

[0070] Furthermore, based on dynamic low-redundancy simulation and parallel simulation, a low-redundancy parallel fusion simulation is designed: multiple worker thread replicas are deployed through a master-slave architecture, with each worker thread executing strategy exploration in parallel within an independent simulation environment. This generates actions such as adjusting model component architecture and optimizing parameter configurations, and collects multi-dimensional reward signals containing performance metrics, computational efficiency, and feasibility constraints in real time for global synchronization. Figure 7 As shown, the details are as follows:

[0071] Worker thread creation: Create and start N worker threads, evenly partition and distribute the model clusters among the worker threads, and copy the global network parameters as a local model copy for each worker thread.

[0072] Worker thread execution: Each worker thread independently executes the reinforcement learning process. Based on the characteristics of the model scheme and the historical simulation results of the model scheme, it executes reinforcement learning actions, selects the next model scheme for simulation, and runs continuously in a loop.

[0073] Global parameter update: After each state run within a worker thread, the value function parameters are updated to the global network after gradient calculation. The global network then updates its parameters through gradient calculation and aggregation, and distributes the updates to each worker thread.

[0074] Main thread completion: The main thread monitors the operation of each worker thread. When the termination condition is met, such as the total number of simulations reaching a threshold or the selected reinforcement learning model scheme remaining unchanged after multiple iterations, the simulation operation is terminated, and the effective model cluster and the preferred model scheme are output.

[0075] S3. Adaptive Algorithm Switching Evaluation and Analysis: Define the scheme evaluation index tree to form the initial evaluation index tree; aggregate the evaluation indexes from multiple dimensions to form the aggregated indexes for system scheme evaluation; adaptively switch the evaluation algorithm to form a comprehensive and reasonable scheme evaluation and analysis result to guide the optimal selection of schemes.

[0076] S31. Definition of the Solution Evaluation Index Tree: Through hierarchical index items, multi-data source mapping mechanisms, integration of subjective and objective weighting, and multi-algorithm collaboration, the tree achieves accurate transformation from abstract evaluation requirements to specific simulation data, thereby enabling comprehensive system verification of complex systems. The construction of the evaluation index tree is based on the following principles:

[0077] Hierarchical structural design principle: Hierarchical structural design reflects the overall characteristics of the system and the interaction between its subsystems. Taking a solid rocket motor in an aircraft as an example, the design is decomposed from the top level down into basic indicators including propellant grain mass, mass ratio, diameter, and length, and performance indicators including maximum combustion chamber pressure, specific impulse, average thrust, and maximum thrust. This hierarchical division clearly presents the coupling relationship between structure and performance.

[0078] The principle of linking indicators with simulation data: all indicator items are linked to specific simulation data and quantification methods. Taking a solid rocket motor for aircraft as an example, the propellant grain mass is calculated through geometric modeling and density simulation, the combustion chamber pressure is solved using an internal ballistic model, and the specific impulse is derived based on simulation data of thrust and mass flow rate, ensuring that the indicators are measurable and dynamically updated, providing data support for performance evaluation.

[0079] The principle of avoiding redundancy at the same level: Indicators at the same level are mutually exclusive to avoid information redundancy between them. Taking solid rocket motors for aircraft as an example, basic indicators are stratified by mass, geometry, and efficiency, while performance indicators are analyzed by pressure, energy conversion, and thrust characteristics, with no information overlap.

[0080] The principle of balancing universality and domain specificity: The indicator system needs to balance universality and domain specificity. Taking solid rocket motors for aircraft as an example, general indicators are retained while highlighting the characteristics of solid rocket motors, so as to meet the common evaluation needs of engines and accurately reflect the domain technical characteristics of solid rocket motors.

[0081] S32. Multi-dimensional integration of evaluation indicators: Utilizing retrieval-enhanced generation technology, historical model evaluation indicator experience is incorporated to automatically aggregate evaluation indicators, forming indicator subtrees, thereby updating the evaluation indicator tree. For example... Figure 8 As shown, the details are as follows:

[0082] (1) Pre-construct a knowledge base of historical model evaluation index vectors, including:

[0083] Knowledge source preprocessing: The knowledge source of historical model evaluation indicators is preprocessed, mainly by structured text conversion;

[0084] Knowledge vectorization: The preprocessed knowledge is vectorized to form a vector knowledge base of historical model evaluation indicators.

[0085] (2) Indicator vectorization: Vectorize the indicator names in the initial evaluation indicator tree.

[0086] (3) Similarity retrieval: By combining the indicator vector with the knowledge base of historical model evaluation indicator vectors, similarity retrieval is performed to obtain the most similar information.

[0087] (4) Rule training and generation: Based on the rules, prompt words are generated, and index aggregation rules and matching evaluation algorithms are generated through a large language model for recommendation.

[0088] (5) Evaluation index aggregation: According to the index aggregation rules, the evaluation indexes are classified and aggregated into multiple index subtrees T[1:n]. The indexes aggregated in each index subtree are subject to the same evaluation algorithm.

[0089] S33. Adaptive Algorithm Switching Evaluation: For different types of scheme metrics, an adaptive switching rule training method based on a large model is used to dynamically select different adaptation evaluation algorithms. For example... Figure 9 As shown, the details are as follows:

[0090] Scheme simulation result variable acquisition: Through low-redundancy parallel fusion simulation, the mns model clusters of the A1.C1.P1-Am.Cn.Ps scheme are simulated and run to obtain effective scheme simulation result variables.

[0091] Simulation variables and index mapping association: Combining the evaluation index tree and subtrees T1-Tn formed by multi-dimensional aggregation, the simulation variables of the scheme are implemented as indicators and associated with the corresponding indicators, serving as the basis for scheme evaluation.

[0092] Adaptive switching rule training: Learn the characteristics of different evaluation algorithms in the evaluation algorithm library, match the characteristics of the indicator subtrees, and output a dynamic selection mechanism and algorithm.

[0093] Adaptive dynamic evaluation analysis: Based on simulation variables associated with the indicators, the indicators of different dimensions are dynamically evaluated. During the evaluation process, the appropriate evaluation algorithm is intelligently selected for the indicator subtrees of different dimensions to obtain the evaluation results of each indicator subtree.

[0094] This invention addresses the need for transformation in the development model of complex systems and the challenges faced by its design verification technology. It proposes a design, verification, and evaluation method based on the intersection of Model-Based Systems Engineering (MBSE) and Artificial Intelligence (AI). Through model cluster-driven, highly generalized design, it expands the design space and improves the quality and efficiency of design solutions. Dynamic, low-redundancy parallel simulation verification enables efficient operation of model clusters and rapid closed-loop verification of solutions. Adaptive algorithm switching evaluation analysis achieves comprehensive and accurate evaluation and selection of solution indicators. This promotes the exploration of more high-quality design solutions in the early stages of development for complex systems, as well as more efficient, comprehensive, and accurate solution verification and evaluation. The research results can provide a technical and methodological foundation for the transformation of development models for complex systems such as aircraft, rockets, and automobiles.

[0095] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and 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 within the scope of protection of the present invention.

Claims

1. A model cluster based intelligent design method for complex systems, characterized in that, The method comprises the following steps: determining the design features of the system to be designed, the design features including architecture features, component features and parameter features; the architecture features including the architecture composition, connection relationship and interface type of the system, the component features including the specific model of each component in the architecture, and the parameter features including the key design parameters and performance indicators of the components; based on the design features, using a retrieval enhancement generation method to retrieve the most similar information in the system design historical knowledge base; and then based on the most similar information, generating generalization rules for system design through a large language model; generating a model cluster based on the generalization rules, the model cluster including multiple design models; generating a model cluster based on the generalization rules, comprising: The generalization rules for system design include architecture generalization generation rules, component generalization generation rules and parameter generalization generation rules; generating an architecture level scheme model according to the architecture generalization generation rules; based on the architecture level scheme model, generating a component level scheme model according to the component generalization generation rules; based on the component level scheme model, generating a parameter level scheme model according to the parameter generalization generation rules, thereby forming a design model; Screening the optimal design model in the model cluster as a Markov decision process, taking the design model simulation performance indicator as the reward function, and obtaining the optimal design model through reinforcement learning.

2. The model cluster based intelligent design method of complex systems as claimed in claim 1, wherein, Based on the design features, using a retrieval enhancement generation method to retrieve the most similar information in the system design historical knowledge base, comprising: forming a system design historical knowledge base by pre-processing and vectorizing the system design historical data in advance; vectorizing the design features, and based on the vectorized design features, using a retrieval enhancement generation method to retrieve the most similar information in the system design historical knowledge base.

3. The model cluster based intelligent design method of complex systems as claimed in claim 1, wherein, Generating an architecture level scheme model according to the architecture generalization generation rules, comprising: automatically generating an architecture topology according to the architecture generalization generation rules, the architecture topology being expressed by SysML; then mapping and converting the architecture topology expressed by SysML into an architecture topology expressed by Modelica, to obtain an architecture level scheme model.

4. The model cluster based intelligent design method of complex system as claimed in claim 1, wherein, Using an A3C network as the decision network of the Markov decision process, i.e. a global network; In the reinforcement learning process, each design model scheme in the model cluster corresponds to a simulation task to obtain a simulation performance indicator; using a multi-thread parallel computing method, the simulation tasks corresponding to the model cluster are evenly partitioned and distributed to each work thread, each work thread copies the global network parameters as a local model copy, each work thread independently performs reinforcement learning and updates the global network parameters, which are then issued to other work threads.

5. The model cluster based intelligent design method of complex systems as claimed in claim 1, wherein, The state space determination method of the Markov decision process is: using word embedding technology and a pre-trained language model BERT to analyze the semantic information of natural language text in the model cluster, using a neural network processing model to process numerical parameter features in the model cluster, realizing multi-dimensional representation of the structure, semantics and parameters of the design model, and forming a state space.

6. The model cluster based intelligent design method of complex system according to any one of claims 1-5, characterized in that, Further comprising the following steps: building an initial evaluation index tree; Based on the initial evaluation index tree, a retrieval enhancement generation method is used to retrieve in the historical model evaluation index vector knowledge base to obtain the most similar information; the most similar information and the rule generation prompt word are combined to generate an index aggregation rule through a large language model to automatically aggregate part of the evaluation indexes in the evaluation index tree to form several index sub-trees; the simulation result of the optimal design model is associated with the indexes in the evaluation index tree, and the optimal design model is evaluated according to the evaluation index tree to obtain an evaluation result.

7. The model cluster based intelligent design method of complex systems as claimed in claim 6, wherein, The aggregated indexes in each index sub-tree use the same evaluation algorithm, and the evaluation algorithm is selected based on a large model.

8. A model cluster based intelligent design system for complex systems, characterized in that, The processor is configured to perform the model cluster-based intelligent design method for a complex system according to any one of claims 1-7.

Images