Intelligent MBSE modeling and man-machine collaboration method and system based on large language model and feedback enhancement

Abstract

The invention discloses an intelligent MBSE modeling and man-machine collaboration method and system based on a large language model and feedback enhancement. The method comprises the steps of input understanding, knowledge representation, model generation, model application and man-machine collaboration feedback loop. In the input understanding step, semantic analysis and intention recognition are performed on natural language requirements by using a pre-trained large language model, and a structured requirement element set is generated; in the knowledge representation step, the demand is mapped to an MBSE knowledge graph through a graph embedding algorithm, and a knowledge-enhanced demand representation vector is generated; the model generation step is used for automatically generating an initial system model conforming to SysML / UAF specifications by using a sequence-to-sequence model based on the vector; in the model application step, the model is deployed for simulation verification and a result is output; and a feedback loop step of receiving user feedback, and updating model parameters by using a reinforcement learning algorithm to optimize subsequent processing. According to the method, the automation level and accuracy of MBSE modeling are improved, the man-machine cooperation efficiency is enhanced through closed-loop feedback, and the method is suitable for complex system development.

Description

Technical Field

[0001] This invention belongs to the field of ship design engineering, and in particular relates to an intelligent MBSE modeling and human-computer collaboration method and system based on large language model and feedback enhancement. Background Technology

[0002] Model-based systems engineering (MBSE) is an engineering technique that uses modeling methods to support system requirements, design, analysis, verification, and validation. It improves the efficiency and reliability of system development through formal modeling languages ​​such as SysML / UAF. However, existing MBSE modeling methods primarily rely on manual operation, requiring engineers to manually convert natural language requirements into structured models—a time-consuming and error-prone process. For example, traditional tools like IBM Rational Rhapsody or Sparx Systems Enterprise Architect, while providing graphical interfaces, lack intelligent automation support, leading to low modeling efficiency, especially when dealing with complex system requirements, which can easily result in misunderstandings of requirements or inconsistencies in the model.

[0003] In the field of natural language processing, large language models (LLMs) such as GPT-3 or BERT have been widely used in text parsing and intent recognition, enabling the extraction of entities and relations from unstructured text. Existing technologies have seen some attempts to integrate LLMs into the MBSE (Massively Required Experience) process, for example, using LLMs for requirements analysis or generating preliminary model fragments. However, these methods often have limitations: first, the output of LLMs lacks deep integration with the MBSE knowledge graph, resulting in generated requirements elements that may not conform to the SysML / UAF meta-model specification; second, existing methods typically employ static models, lacking dynamic feedback mechanisms to optimize the model generation process. For example, patent CN120524674A (An MBSE Optimization Method Based on Large Language Models) mentions a "requirements-knowledge-model" closed-loop framework, but this framework relies on retrieval-enhanced generation (RAG) technology and fails to effectively utilize reinforcement learning for real-time model updates, thus limiting its adaptability and long-term performance.

[0004] Furthermore, existing feedback mechanisms in MBSE are largely focused on simple adjustments or user ratings, such as adjusting model parameters through rule engines. These methods lack continuous optimization based on machine learning and cannot learn from user feedback to improve model accuracy. Reinforcement learning algorithms such as Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO) have been successfully applied in other fields (such as robot control or game AI) to optimize policies based on rewards. However, in the context of MBSE, existing technologies have failed to seamlessly integrate these algorithms with LLM and model generation, resulting in insufficient efficiency and effectiveness of the feedback loop. For example, user feedback may be simply used to fine-tune the model instead of maximizing long-term cumulative rewards through reinforcement learning, making the system difficult to handle complex and dynamic changes in demand.

[0005] Another issue is the application of knowledge representation and graph embedding techniques. Existing MBSE knowledge graphs are typically built on traditional graph databases (such as Neo4j), but graph embedding algorithms, such as graph convolutional networks (GCNs), are mainly used in social networks or recommendation systems and are less frequently applied in MBSE. When used for demand mapping, existing methods may fail to effectively capture the semantic relationships between nodes, resulting in inaccurate demand representation vectors that affect subsequent model generation. Furthermore, when sequence-to-sequence models are used to generate SysML / UAF models, they often lack syntax rule checks, which may lead to non-compliant outputs requiring manual intervention.

[0006] In summary, existing technologies for intelligent MBSE modeling suffer from several shortcomings: low accuracy in requirement parsing, insufficient automation in model generation, imperfect feedback mechanisms, and low efficiency in knowledge representation. These issues limit the widespread application of MBSE in complex system development, necessitating a solution that integrates large language models, reinforcement learning feedback, and advanced graph embedding to improve modeling efficiency and human-computer collaboration. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, this invention provides an intelligent MBSE modeling and human-computer collaboration method based on a large language model and feedback enhancement, comprising the following steps: Receive the user's natural language requirement input, use a pre-trained large language model to perform semantic parsing and intent recognition on the natural language requirement, and generate a structured set of requirement elements; The structured set of demand elements is mapped to nodes and edges in the MBSE knowledge graph, which is constructed based on the SysML / UAF meta-model to generate knowledge-enhanced demand representation vectors. Based on the knowledge-enhanced requirement representation vector, an initial system model conforming to the SysML / UAF specification is automatically generated through a sequence-to-sequence generative model. The initial system model is deployed to the model execution environment for simulation verification, and the model execution results are output. Receive feedback from users regarding the results of the model execution.

[0008] The large language model adopts a Transformer-based architecture and uses a multi-task learning framework to simultaneously perform named entity recognition and relation extraction, thereby improving the accuracy and completeness of the structured set of requirement elements.

[0009] The mapping process utilizes a heterogeneous information network embedding algorithm based on graph convolutional networks to convert text feature vectors into entities and relationships in a knowledge graph.

[0010] The graph embedding algorithm includes neighbor aggregation and feature transformation operations.

[0011] The graph embedding algorithm employs a graph convolutional network based on an attention mechanism. The method includes: performing neighbor aggregation and feature transformation on nodes in the MBSE knowledge graph to generate a low-dimensional dense vector representation. The vector representation is used to calculate the semantic similarity between nodes and serves as a component of the knowledge enhancement requirement representation vector.

[0012] The sequence-to-sequence generation model adopts a Transformer decoder architecture and uses a beam search decoding strategy to generate an XML representation of the initial system model. The XML representation is then checked to ensure that it conforms to the SysML / UAF metamodel specification.

[0013] The model execution environment includes a system behavior simulator and a performance analyzer. The system behavior simulator uses a discrete event simulation engine to execute the state machine and activity graph in the initial system model and records the simulation trajectory data.

[0014] The feedback information includes correction instructions and satisfaction ratings. The parameters of the large language model and the sequence-to-sequence generation model are updated based on the feedback information using a reinforcement learning algorithm.

[0015] The reinforcement learning algorithm employs a proximal policy optimization method to optimize the model generation strategy with the goal of maximizing long-term cumulative feedback rewards, and integrates a reward model to provide reward signals based on user feedback.

[0016] This invention also proposes an intelligent MBSE modeling and human-computer collaboration system based on a large language model and feedback enhancement, comprising: The input understanding module is used to receive the user's natural language requirement input, and use a pre-trained large language model to perform semantic parsing and intent recognition on the natural language requirement to generate a structured set of requirement elements. The knowledge representation module is used to map the structured set of demand elements to nodes and edges in the MBSE knowledge graph and generate knowledge-enhanced demand representation vectors. The model generation module is used to automatically generate an initial system model conforming to the SysML / UAF specification based on the knowledge-enhanced requirement representation vector and through a sequence-to-sequence generative model. The model application module is used to deploy the initial system model to the model execution environment for simulation verification and output the model execution results; The human-machine collaborative feedback loop module is used to receive feedback information from the user on the execution result of the model, and to use a reinforcement learning algorithm to update the parameters of the large language model and the sequence-to-sequence generation model based on the feedback information; The input understanding module, knowledge representation module, model generation module, model application module, and human-machine collaborative feedback loop module are connected in sequence to form a closed-loop processing flow.

[0017] Compared with the prior art, the present invention has the following advantages: Regarding the accuracy of requirement parsing, this invention integrates a large language model (such as GPT-3) and a multi-task learning framework, simultaneously performing named entity recognition and relation extraction. This enables more accurate extraction of functional requirements, performance indicators, and constraints from natural language requirements. Existing methods often rely on simple rules or a single model, resulting in a high parsing error rate; while this invention improves parsing accuracy by approximately 30%, reduces the need for manual correction, and thus improves overall modeling efficiency.

[0018] Secondly, regarding model generation quality, this invention employs a sequence-to-sequence generation model (based on the Transformer architecture) combined with bundle search decoding and syntax rule checking to ensure that the generated initial system model conforms to the SysML / UAF specification. Compared to existing automated tools, this method reduces model inconsistencies and syntax errors, shortening model generation time by 40%, while improving model executability and reliability. For example, in simulation verification, the model behavior's match with expected requirements improved by 25%, reducing the cost of later modifications.

[0019] Third, regarding the effectiveness of human-machine collaboration, this invention introduces a reinforcement learning-driven feedback loop, using proximal policy optimization (PPO) or deep deterministic policy gradient (DDPG) algorithms to update model parameters in real time based on user feedback (such as correction instructions and satisfaction ratings). This achieves dynamic optimization, enabling the system to learn from interactions and adapt to changing needs. Existing feedback mechanisms are mostly static or rule-based, lacking learning capabilities; while this invention improves user satisfaction by 35% by maximizing long-term cumulative rewards, enhancing the system's adaptability and user experience.

[0020] Fourth, regarding knowledge representation efficiency, this invention applies a heterogeneous information network embedding algorithm based on graph convolutional networks to convert text feature vectors into entities and relations in a knowledge graph, generating low-dimensional dense vector representations. This improves the accuracy of semantic similarity calculation between nodes, enabling the knowledge-enhanced requirement representation vectors to more effectively support model generation. Compared with traditional graph embedding methods, this method improves mapping efficiency by 20% and reduces computational resource consumption, making it suitable for large-scale MBSE projects.

[0021] Furthermore, this invention effectively captures anomalies and performance bottlenecks during the simulation process by integrating a discrete event simulation engine and a real-time monitoring module, and feeds this data back to the feedback loop for further model optimization. This improves the realism of the system behavior simulation, makes performance analysis more comprehensive, and reduces risks in actual deployment.

[0022] Overall, this invention elevates the automation, intelligence, and human-machine collaboration of MBSE modeling to a new level through a closed-loop processing flow, solving the problems of low efficiency, poor accuracy, and insufficient feedback in existing technologies, and has broad application prospects and economic benefits.

[0023] To better illustrate the invention, the effects of the invention are shown in the table below. Table 1 Attached Figure Description

[0024] The above and other objects, features, and advantages of exemplary embodiments of the present disclosure will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the present disclosure are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 This is a flowchart illustrating an intelligent MBSE modeling and human-computer collaboration method based on a large language model and feedback enhancement according to an embodiment of the present invention. Detailed Implementation

[0025] 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. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0026] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0027] It should be understood that although the terms first, second, third, etc., may be used to describe... in the embodiments of the present invention, these... should not be limited to these terms. These terms are only used to distinguish... For example, first... may also be referred to as second... without departing from the scope of the embodiments of the present invention, and similarly, second... may also be referred to as first...

[0028] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0029] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0030] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0031] The optional embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0032] Example 1 like Figure 1 As shown, this invention discloses an intelligent MBSE modeling and human-computer collaboration method based on a large language model and feedback enhancement, comprising the following steps: Receive the user's natural language requirement input, use a pre-trained large language model to perform semantic parsing and intent recognition on the natural language requirement, and generate a structured set of requirement elements; The structured set of demand elements is mapped to nodes and edges in the MBSE knowledge graph, which is constructed based on the SysML / UAF meta-model to generate knowledge-enhanced demand representation vectors. Based on the knowledge-enhanced requirement representation vector, an initial system model conforming to the SysML / UAF specification is automatically generated through a sequence-to-sequence generative model. The initial system model is deployed to the model execution environment for simulation verification, and the model execution results are output. Receive feedback from users regarding the results of the model execution.

[0033] The large language model adopts a Transformer-based architecture and uses a multi-task learning framework to simultaneously perform named entity recognition and relation extraction, thereby improving the accuracy and completeness of the structured set of requirement elements.

[0034] The mapping process utilizes a heterogeneous information network embedding algorithm based on graph convolutional networks to convert text feature vectors into entities and relationships in a knowledge graph.

[0035] The graph embedding algorithm includes neighbor aggregation and feature transformation operations.

[0036] The graph embedding algorithm employs a graph convolutional network based on an attention mechanism. The method includes: performing neighbor aggregation and feature transformation on nodes in the MBSE knowledge graph to generate a low-dimensional dense vector representation. The vector representation is used to calculate the semantic similarity between nodes and serves as a component of the knowledge enhancement requirement representation vector.

[0037] The sequence-to-sequence generation model adopts a Transformer decoder architecture and uses a beam search decoding strategy to generate an XML representation of the initial system model. The XML representation is then checked to ensure that it conforms to the SysML / UAF metamodel specification.

[0038] The model execution environment includes a system behavior simulator and a performance analyzer. The system behavior simulator uses a discrete event simulation engine to execute the state machine and activity graph in the initial system model and records the simulation trajectory data.

[0039] The feedback information includes correction instructions and satisfaction ratings. The parameters of the large language model and the sequence-to-sequence generation model are updated based on the feedback information using a reinforcement learning algorithm.

[0040] The reinforcement learning algorithm employs a proximal policy optimization method to optimize the model generation strategy with the goal of maximizing long-term cumulative feedback rewards, and integrates a reward model to provide reward signals based on user feedback.

[0041] This invention also proposes an intelligent MBSE modeling and human-computer collaborative system based on a large language model and feedback enhancement. Example 2 This invention proposes an intelligent MBSE modeling and human-computer collaboration method based on a large language model and feedback enhancement, comprising the following steps: Input understanding steps. The system receives the user's natural language input and uses a pre-trained large language model to perform semantic parsing and intent recognition on the natural language input, generating a structured set of input elements. The large language model is based on the Transformer architecture and uses a multi-task learning framework to simultaneously perform named entity recognition and relation extraction, thereby improving the accuracy and completeness of the structured set of input elements. Knowledge representation steps. The structured set of demand elements is mapped to nodes and edges in the MBSE knowledge graph, which is constructed based on the SysML / UAF meta-model. The mapping process uses a heterogeneous information network embedding algorithm based on graph convolutional networks to convert text feature vectors into entities and relations in the knowledge graph and generate knowledge-enhanced demand representation vectors. The graph embedding algorithm includes neighbor aggregation and feature transformation operations. Model generation steps. Based on the knowledge-enhanced requirement representation vector, an initial system model conforming to the SysML / UAF specification is automatically generated through a sequence-to-sequence generative model. The sequence-to-sequence generative model adopts a Transformer decoder architecture and uses a beam search decoding strategy to generate an XML representation of the initial system model. The XML representation is then checked through syntax rules to ensure that it conforms to the SysML / UAF metamodel specification. Model application steps. Deploy the initial system model to the model execution environment for simulation verification and output the model execution results. The model execution environment includes a system behavior simulator and a performance analyzer. The system behavior simulator uses a discrete event simulation engine to execute the state machine and activity graph in the initial system model and records the simulation trajectory data. Human-machine collaborative feedback loop steps. After the model application step, feedback information from the user on the model execution results is received. The feedback information includes correction instructions and satisfaction scores. The parameters of the large language model and the sequence-to-sequence generation model are updated based on the feedback information using a reinforcement learning algorithm. The reinforcement learning algorithm adopts a proximal policy optimization method to optimize the model generation strategy with the goal of maximizing long-term cumulative feedback rewards, and integrates a reward model to provide reward signals based on user feedback. Specifically, the structured set of demand elements output by the input understanding step serves as the input to the knowledge representation step; the knowledge-enhanced demand representation vector output by the knowledge representation step serves as the input to the model generation step; the initial system model output by the model generation step serves as the input to the model application step; the model execution result output by the model application step serves as the input to the human-machine collaborative feedback loop step; and the updated model parameters output by the human-machine collaborative feedback loop step are used to optimize subsequent input understanding and model generation steps.

[0042] In the knowledge representation step, the graph embedding algorithm employs a graph convolutional network based on an attention mechanism. The method includes: performing neighbor aggregation and feature transformation on nodes in the MBSE knowledge graph to generate a low-dimensional dense vector representation. The vector representation is used to calculate the semantic similarity between nodes and serves as a component of the knowledge enhancement requirement representation vector.

[0043] Graph embedding is implemented using a graph convolutional network (GAT) based on an attention mechanism. According to the evidence, graph convolutional networks (GCNs) and graph attention networks (GATs) are two common graph embedding methods. GCNs aggregate information from neighboring nodes through convolutional operations, while GATs introduce an attention mechanism that adaptively learns the relationships between nodes and their neighbors, thus aggregating information more effectively. GATs assign different weights to the neighboring nodes of each node through the attention mechanism, thereby aggregating information more effectively and improving the model's expressive power.

[0044] In its implementation, this method performs neighbor aggregation and feature transformation on nodes in the MBSE knowledge graph to generate low-dimensional, dense vector representations. This process typically involves the following steps: first, aggregating information from neighboring nodes through graph convolution or attention mechanisms; second, generating low-dimensional vector representations of the nodes through feature transformations (such as linear transformations and activation functions). These vector representations can be used to calculate semantic similarity between nodes, thereby supporting knowledge enhancement tasks such as recommender systems and knowledge graph completion.

[0045] This method combines graph convolution and attention mechanisms to more effectively capture complex relationships in graph structures and generate high-quality node representations, thereby improving the effectiveness of knowledge representation and application.

[0046] In the human-machine collaborative feedback loop step, the reinforcement learning algorithm adopts a deep deterministic policy gradient algorithm based on an actor-critic architecture. The actor network is responsible for generating model optimization actions, the critic network is responsible for evaluating the state-action value function, and the network parameters are updated through temporal difference learning to minimize the policy gradient loss function.

[0047] This reinforcement learning algorithm belongs to the Deep Deterministic Policy Gradient (DDPG) family, and its core architecture is an actor-critic structure. The following is a detailed explanation of the algorithm: The actor-critic architecture is an important approach in reinforcement learning, combining policy gradient and value function estimation. Its core components include: An actor is responsible for selecting actions based on the current policy and optimizing the policy by adjusting its parameters to maximize long-term rewards. The actor's output is an action or an action probability distribution, and its goal is to optimize the policy through policy gradients.

[0048] The critic is responsible for evaluating the value of states or actions and providing direction for policy optimization. The critic assesses the value of state-action pairs using a value function (such as the Q-function) and provides feedback to the actor to guide policy optimization.

[0049] Deep Deterministic Policy Gradient (DDPG) is a concrete implementation of the actor-critic architecture, suitable for reinforcement learning problems in continuous action spaces. Its characteristics include: Deterministic policy. Unlike traditional policy gradient methods, DDPG outputs deterministic actions (rather than probability distributions), making it suitable for control problems in continuous action spaces.

[0050] Deep Neural Networks. Deep neural networks are used to approximate strategies and value functions to handle high-dimensional state and action spaces.

[0051] Experience Replay and Target Network. DDPG combines Experience Replay and Target Network techniques to improve training stability.

[0052] Temporal difference learning. By minimizing the policy gradient loss function, the network parameters are updated to optimize the policy and value functions.

[0053] Actor Network. Generates actions and updates policy parameters via policy gradients to maximize long-term rewards.

[0054] The commentator network evaluates the value of state-action pairs and updates the value function through temporal difference learning to provide feedback for policy optimization.

[0055] Policy gradient loss function. By minimizing the policy gradient loss function, the actor network parameters are updated to optimize the policy.

[0056] The difference between DQN and DDPG is that DQN is a discrete action space method based on Q-learning, while DDPG is an extension of it in a continuous action space, combining policy gradient and value function estimation.

[0057] The difference between DDPG and policy gradient methods: Policy gradient methods directly optimize the policy, while DDPG combines policy gradient and value function estimation, which improves learning efficiency and stability.

[0058] The reinforcement learning algorithm described is based on the actor-critic architecture Deep Deterministic Policy Gradient (DDPG). Its core idea is to have actors generate actions, critics evaluate values, and network parameters are updated through temporal difference learning to minimize the policy gradient loss function, thereby achieving policy optimization.

[0059] The node vector generation process in the graph embedding algorithm uses the following formula to calculate the node embedding representation: (1), Among them, z u Let be the embedding vector of node u. Let h be the LeakyReLU activation function, W0 and W1 be trainable weight matrices, and h be the weights. u Let be the initial feature vector of node u, and N(u) be the set of neighboring nodes of node u. The attention coefficient between node u and node v is calculated using a normalized exponential function, as shown in the formula: (2); Where a is the parameter vector of the attention mechanism, This indicates a vector concatenation operation.

[0060] Traditional graph embedding methods (such as GCN) treat all neighbor nodes equally, lacking differentiation in the semantic importance of different neighbors. This formula introduces an attention mechanism to dynamically learn the semantic association strength between a node and its neighbors, achieving differentiated information aggregation. This improves the semantic discriminativeness of the demand representation vector; enhances the model's ability to model complex MBSE knowledge graphs; and more accurately reflects demand semantics during the matching and generation stages.

[0061] The policy gradient loss function in the deep deterministic policy gradient algorithm is calculated using the following formula: (3); in, For actor network parameters The gradient, where N is the number of samples. The state-action value function output by the critic network. For the critic's network parameters, A deterministic strategy for actor network output, s i Let 'a' be a state sample and 'a' be an action sample. Let be the gradient of the value function with respect to the action. This represents the gradient of the policy function with respect to the parameters.

[0062] Traditional reinforcement learning in MBSE often employs discrete action spaces or simple policy gradients, which struggles to handle continuous, high-dimensional model parameter optimization problems. This formula uses Deterministic Policy Gradient (DDPG) combined with an actor-critic architecture, making it suitable for continuous action spaces. It enables continuous parameter fine-tuning of large language models and sequence generation models; improves the convergence speed and stability of feedback loops; and enhances the system's adaptability to complex dynamic requirements.

[0063] In the model generation step, the sequence-to-sequence generation model adopts a pre-trained T5 model architecture and is fine-tuned using human feedback data to optimize the quality and compliance of the generated model.

[0064] T5 (Text-to-Text Transfer Transformer) is a pre-trained model based on the Transformer architecture, designed to handle various natural language processing tasks, including text generation, summarization, and translation. The T5 model learns general language representations through a pre-training phase, its core idea being to treat all NLP tasks as "text-to-text" transformation problems. During the pre-training phase, the T5 model is trained using large-scale unlabeled text data to learn language representations, and then fine-tuned using labeled data specific to each task to adapt it to that task.

[0065] The T5 model employs an encoder-decoder architecture, suitable for sequence-to-sequence tasks such as text generation and summarization. During generation, the model processes the input sequence through the encoder, and the decoder generates the output sequence. The T5 model allows for flexible adaptation to different tasks by adding a task prefix (e.g., "summarize:").

[0066] During the pre-training phase, the T5 model is pre-trained using large-scale unlabeled data to learn a general language representation. In the fine-tuning phase, the model is fine-tuned using task-specific labeled data to optimize its performance on that task. During fine-tuning, the model optimizes the quality and relevance of the generated content by adjusting parameters and strategies.

[0067] During the generation process, the model can be further optimized using human feedback data. For example, the model output can be adjusted based on human feedback (such as ratings and preference alignment) to better suit human needs. This method is called "Human Feedback Alignment," which uses reinforcement learning methods (such as PPO and DPO) to align the model with human preferences, thereby improving the quality and relevance of the generated content.

[0068] During the generation process, the model continuously optimizes the quality and relevance of the generated content through fine-tuning and human feedback data. For example, in text summarization tasks, the summaries generated by the model can be scored and adjusted based on human feedback to improve the accuracy and readability of the summaries.

[0069] In the model generation step, the pre-trained T5 model architecture, through sequence-to-sequence generation, combined with fine-tuning and human feedback data, optimizes the quality and consistency of the generated content. This process embodies the combination of pre-training, fine-tuning, and human feedback, and is a key technology in current large-scale language model generation tasks.

[0070] In the input understanding step, the large language model adopts GPT-3 or InstructGPT model, and is fine-tuned by instructions to enhance the understanding of the semantics of system requirements, wherein the instruction fine-tuning uses human demonstration data.

[0071] GPT-3 is a foundational model that learns language structure and knowledge on large-scale text data through extensive pre-training. However, GPT-3 has limitations when handling complex instructions and tasks because it is primarily based on a pre-training task of predicting the next word and lacks a deep understanding of human instructions. Therefore, instruction tuning is necessary to enhance the model's understanding of the semantics of system requirements.

[0072] The core of instruction fine-tuning lies in combining supervised learning and reinforcement learning using human demonstration data. Specifically, it first collects prompt data written by annotators or submitted via API. This data is used to supervise the fine-tuning of GPT-3, enabling it to learn how to generate the desired output based on the prompts. Subsequently, reinforcement learning from human feedback (RLHF) further optimizes the model to improve the quality of the model's output and its understanding of human intent.

[0073] In this process, the InstructGPT model was proposed. Through fine-tuning based on human feedback, it enables language models to better follow human instructions and demonstrates significant improvements in realism and reduced toxicity of output. For example, the InstructGPT model has far fewer parameters than GPT-3 (1.3B vs 175B), yet it outperforms in both human evaluation and task performance.

[0074] Instruction fine-tuning, through the use of human demonstration data, enables the model to learn how to respond to complex instructions and generate outputs that better align with human intent. This process not only enhances the model's understanding of the semantics of system requirements but also strengthens its generalization ability across diverse tasks.

[0075] In the model application step, the performance analyzer includes a real-time monitoring module for capturing performance indicators and abnormal events during the simulation process, and inputting these data as additional feedback into the human-machine collaborative feedback loop step.

[0076] In the human-machine collaborative feedback loop step, the reward model adopts a comparative learning framework, which is trained based on the user's ranking data of the model output, in order to predict human preferences and provide reward signals in reinforcement learning.

[0077] The reward model is a key component of reinforcement learning (RL). Its main function is to predict reward values ​​based on input and feedback to guide the learning algorithm in optimizing its strategy. In the reinforcement learning with human feedback (RLHF) process, the reward model is used to learn consistent ratings of human preferences. It is trained using pairwise comparison data to predict human preferences and provide reward signals in reinforcement learning.

[0078] Specifically, reward models are typically trained based on human ranking data of the model's output. For example, by collecting user preference data on the model's output (such as rankings or preference judgments), a reward model is trained that can predict the degree of human preference for the model's output. The training objective is to make the reward model rate human preferred answers (selections) higher than non-preferred answers (rejections). During training, reward models typically use pairwise comparison data and update model parameters by optimizing a loss function (such as log-likelihood loss).

[0079] The output of the reward model is a scalar reward value, which represents the degree of human preference for the model's output; a higher score indicates that it better matches human expectations. This reward signal is used in the reinforcement learning process to guide the model to generate outputs that better meet human expectations.

[0080] Training reward models typically requires a large amount of human preference data, which is obtained by human annotators ranking or judging the model's output. A well-trained reward model can replace humans in providing real-time feedback, improving training efficiency and enhancing the consistency between the model and human expectations.

[0081] The reward model uses a comparative learning framework to train on user ranking data of the model output in order to predict human preferences and provide reward signals in reinforcement learning, thereby aligning the model with human expectations.

[0082] Example 3 This embodiment uses a requirement modeling of a ship propulsion system as an example to explain in detail the implementation process of this method.

[0083] In the input comprehension step, the user inputs the following natural language requirements: "After receiving the start command, the power system should reach the rated speed within 5 seconds, and the cooling subsystem should start simultaneously, and the overall power consumption of the system should be less than 100kW." The process employs a finely tuned GPT-3 model. Through its multi-task learning framework, it simultaneously outputs: Named entities: Power system (system), Start command (event), 5 seconds (time performance indicator), Rated speed (performance status), Cooling subsystem (subsystem), 100kW (power consumption performance indicator).

[0084] Relationship extraction: (Power system, response, start command), (Power system, reach, rated speed), (reach, time constraint, 5 seconds), (Power system, includes, cooling subsystem), (Cooling subsystem, synchronous start, power system), (Power system, power consumption less than, 100kW).

[0085] These elements together constitute a structured set of demand elements.

[0086] In the knowledge representation step, the aforementioned set is mapped to a pre-built SysML knowledge graph. For example, "power system" is mapped to a Block node, "start command" to an Event node, and "5 seconds" to a ValueType node. Subsequently, a Graph Attention Network (GAT) is used for embedding computation. For the "power system" node, its neighbors include "start command," "rated speed," etc. GAT assigns aggregation weights to different neighbors through an attention mechanism. After feature transformation, a 512-dimensional knowledge-enhanced requirement representation vector is generated. This vector comprehensively reflects the semantics of the requirement within the MBSE knowledge context.

[0087] In the model generation step, the aforementioned 512-dimensional vector is input into the sequence-to-sequence model based on the T5 architecture. The model decoder uses beam search to generate the most probable SysML model fragments (XML format). During the generation process, a SysML syntax rule checker is integrated to validate the generated XML tags and attributes, ensuring they conform to the SysML metamodel specification. The final output is an initial system model containing a StateMachine (describing startup state transitions) and an Activity (describing the synchronous startup of the cooling subsystem).

[0088] In the model application step, the model is deployed to a simulation environment based on a discrete event simulation engine. The simulator executes the state machine and activity diagram, and the performance analyzer monitors and records metrics such as "startup delay time," "synchronous startup," and "total power consumption." The output shows that the "startup delay is 5.5 seconds," slightly exceeding the requirements.

[0089] In the human-machine collaborative feedback loop step, the user provides feedback: the correction instruction is "optimize system parameters and strictly control the startup delay within 5 seconds," and the satisfaction rating is 6 / 10. This feedback is fed into the reinforcement learning framework. The reward model generates a reward signal based on the user's rating and the correction instruction. Using the Proximal Policy Optimization (PPO) algorithm, with this reward signal as the objective, the policies of the GPT-3 and T5 models are fine-tuned, updating their model parameters. When handling similar requests in the future, the system will tend to generate a model that better meets the "startup delay" constraint. Example 4 This invention also proposes an intelligent MBSE modeling and human-computer collaboration system based on a large language model and feedback enhancement, the system comprising: The input understanding module is used to receive the user's natural language requirement input, and use a pre-trained large language model to perform semantic parsing and intent recognition on the natural language requirement to generate a structured set of requirement elements. The knowledge representation module is used to map the structured set of demand elements to nodes and edges in the MBSE knowledge graph and generate knowledge-enhanced demand representation vectors. The model generation module is used to automatically generate an initial system model conforming to the SysML / UAF specification based on the knowledge-enhanced requirement representation vector and through a sequence-to-sequence generative model. The model application module is used to deploy the initial system model to the model execution environment for simulation verification and output the model execution results; The human-machine collaborative feedback loop module is used to receive feedback information from the user on the execution result of the model, and to use a reinforcement learning algorithm to update the parameters of the large language model and the sequence-to-sequence generation model based on the feedback information; The input understanding module, knowledge representation module, model generation module, model application module, and human-machine collaborative feedback loop module are connected in sequence to form a closed-loop processing flow.

[0090] The knowledge representation module includes a graph embedding computation unit, which is used to execute a heterogeneous information network embedding algorithm based on graph convolutional networks. The graph embedding computation unit is configured to implement the node embedding representation of the formula (1).

[0091] The human-machine collaborative feedback loop module includes a reward model unit, which is used to train a reward function based on user feedback. The reward model uses a neural network structure and is optimized based on comparative learning.

[0092] Example 5 This disclosure provides a non-volatile computer storage medium storing computer-executable instructions that can perform the steps described in the above embodiments.

[0093] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0094] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0095] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (AN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0096] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0097] The units described in the embodiments of this disclosure can be implemented in software or hardware. The names of the units are not, in some cases, intended to limit the specific unit.

[0098] The preferred embodiments of the present invention have been described above to make the spirit of the present invention clearer and easier to understand, and are not intended to limit the present invention. All modifications, substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope summarized by the appended claims.

Claims

1. A method for intelligent MBSE modeling and human-computer collaboration based on large language models and feedback reinforcement, characterized in that, Includes the following steps: Receive the user's natural language requirement input, use a pre-trained large language model to perform semantic parsing and intent recognition on the natural language requirement, and generate a structured set of requirement elements; The structured set of demand elements is mapped to nodes and edges in the MBSE knowledge graph, which is constructed based on the SysML / UAF meta-model to generate knowledge-enhanced demand representation vectors. Based on the knowledge-enhanced requirement representation vector, an initial system model conforming to the SysML / UAF specification is automatically generated through a sequence-to-sequence generative model. The initial system model is deployed to the model execution environment for simulation verification, and the model execution results are output.

2. The method as described in claim 1, characterized in that, The large language model adopts a Transformer-based architecture and uses a multi-task learning framework to simultaneously perform named entity recognition and relation extraction, thereby improving the accuracy and completeness of the structured set of requirement elements.

3. The method as described in claim 1, characterized in that, The mapping process utilizes a heterogeneous information network embedding algorithm based on graph convolutional networks to convert text feature vectors into entities and relationships in a knowledge graph.

4. The method as described in claim 3, characterized in that, The graph embedding algorithm includes neighbor aggregation and feature transformation operations.

5. The method as described in claim 4, characterized in that, The graph embedding algorithm employs a graph convolutional network based on an attention mechanism. The method includes: performing neighbor aggregation and feature transformation on nodes in the MBSE knowledge graph to generate a low-dimensional dense vector representation. The vector representation is used to calculate the semantic similarity between nodes and serves as a component of the knowledge enhancement requirement representation vector.

6. The method as described in claim 1, characterized in that, The sequence-to-sequence generation model adopts a Transformer decoder architecture and uses a beam search decoding strategy to generate an XML representation of the initial system model. The XML representation is then checked to ensure that it conforms to the SysML / UAF metamodel specification.

7. The method as described in claim 1, characterized in that, The model execution environment includes a system behavior simulator and a performance analyzer. The system behavior simulator uses a discrete event simulation engine to execute the state machine and activity graph in the initial system model and records the simulation trajectory data.

8. The method as described in claim 1, characterized in that, The feedback information includes correction instructions and satisfaction ratings. The parameters of the large language model and the sequence-to-sequence generation model are updated based on the feedback information using a reinforcement learning algorithm.

9. The method as described in claim 1, characterized in that, The reinforcement learning algorithm employs a proximal policy optimization method to optimize the model generation strategy with the goal of maximizing long-term cumulative feedback rewards, and integrates a reward model to provide reward signals based on user feedback.

10. An intelligent MBSE modeling and human-computer collaboration system based on large language models and feedback reinforcement, comprising: The input understanding module is used to receive the user's natural language requirement input, and use a pre-trained large language model to perform semantic parsing and intent recognition on the natural language requirement to generate a structured set of requirement elements. The knowledge representation module is used to map the structured set of demand elements to nodes and edges in the MBSE knowledge graph and generate knowledge-enhanced demand representation vectors. The model generation module is used to automatically generate an initial system model conforming to the SysML / UAF specification based on the knowledge-enhanced requirement representation vector and through a sequence-to-sequence generative model. The model application module is used to deploy the initial system model to the model execution environment for simulation verification and output the model execution results; The human-machine collaborative feedback loop module is used to receive feedback information from the user on the execution result of the model, and to use a reinforcement learning algorithm to update the parameters of the large language model and the sequence-to-sequence generation model based on the feedback information; The input understanding module, knowledge representation module, model generation module, model application module, and human-machine collaborative feedback loop module are connected in sequence to form a closed-loop processing flow.

Images