Space docking mechanism state reconstruction and optimization method based on twin agent

Abstract

The application discloses a space docking mechanism state reconstruction and optimization method based on a twin intelligent agent, and belongs to the technical field of ground tests of space docking mechanisms. The method comprises the following steps: constructing a twin intelligent agent and a space docking field knowledge graph; receiving a natural language test task; the twin intelligent agent analyzes the natural language test task into structured working condition parameters by combining the knowledge graph with an engineering constraint through a knowledge retrieval mechanism; calling a pre-constructed multi-fidelity structure response proxy model to predict the full-field displacement of the space docking mechanism according to the structured working condition parameters; extracting a maximum deformation value based on the predicted full-field displacement, and identifying a risk if the maximum deformation value exceeds an allowable deformation threshold; generating a candidate working condition adjustment scheme based on the constraint of the knowledge graph and calling the proxy model again for verification until a safety condition is met. The multi-fidelity proxy model realizes rapid reconstruction of the full-field displacement under sparse monitoring conditions, and improves the state monitoring efficiency and optimization decision-making capability of the ground test of the space docking mechanism.

Description

Technical Field

[0001] This application belongs to the field of ground test technology for aerospace docking mechanisms, specifically involving a method for state reconstruction and optimization of aerospace docking mechanisms based on twin intelligent agents. Background Technology

[0002] The docking mechanism is a key mechanical component for spacecraft to achieve capture, buffering, correction, and rigid connection, and its structural response characteristics directly affect the safety of the docking mission. Due to the high-risk and irreversible nature of on-orbit docking missions, the docking mechanism usually needs to obtain its structural response characteristics in advance through buffering tests and numerical simulations during the ground phase to support risk assessment and operational design.

[0003] In existing technologies, one type of method mainly relies on ground-based physical experiments and sensor monitoring to obtain key response data of the docking mechanism. However, due to limitations in the number of sensors, installation locations, and wiring conditions, only a small number of measurement points can typically be obtained, making it difficult to acquire information on the overall displacement and local high-risk areas of complex docking mechanisms. Another type of method uses finite element analysis for modeling and calculation. However, aerospace docking mechanisms often involve contact nonlinearity, material parameter deviations, assembly errors, and uncertain boundary conditions during buffer tests, making it difficult for finite element models to fully characterize the actual test response. Furthermore, high-fidelity finite element calculations are time-consuming, making it difficult to meet the needs of real-time state reconstruction and immediate risk identification during the test.

[0004] In recent years, although digital twin and surrogate model technologies have been introduced into the field of structural response prediction, existing systems mostly rely on offline simulation data, which is insufficient for online adaptation to unknown working conditions and measured deviations. Moreover, most of them remain at the level of state display and lack the ability to make intelligent decisions and close-loop optimization by integrating domain knowledge. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method for reconstructing and optimizing the state of aerospace docking mechanisms based on twin agents. This method combines knowledge graph constraints with large language model reasoning and achieves rapid reconstruction of full-field displacement under sparse monitoring conditions through a multi-fidelity agent model.

[0006] The present invention provides a method for state reconstruction and optimization of aerospace docking mechanisms based on twin intelligent agents, comprising the following steps: Construct a twin intelligent agent and a knowledge graph for the aerospace docking field, wherein the knowledge graph contains entities and relationships of structural components, load conditions, deformation thresholds and engineering rules; Upon receiving a natural language experiment task, the twin agent uses a knowledge retrieval mechanism combined with engineering constraints in the knowledge graph to parse the natural language experiment task into structured engineering condition parameters. The pre-built multi-fidelity structural response proxy model is invoked to predict the full-field displacement of the aerospace docking mechanism based on the structural chemical condition parameters. The maximum deformation value is extracted based on the predicted full-field displacement. If the maximum deformation value exceeds the allowable deformation threshold, a risk is identified. Candidate working condition adjustment schemes are generated based on the constraints of the knowledge graph, and the multi-fidelity structural response proxy model is called again to verify the candidate working condition adjustment schemes until the safety conditions are met. The safety conditions include at least the maximum deformation value not exceeding the allowable deformation threshold.

[0007] By calling a multi-fidelity structural response proxy model, the full-field displacement can be predicted quickly under sparse monitoring conditions. This overcomes the shortcomings of traditional finite element simulation, which is time-consuming and cannot meet real-time requirements. It can quickly obtain the full-field deformation distribution of the docking mechanism during the test, providing complete data support for condition assessment.

[0008] As a further limitation of the technical solution of this invention, the steps for constructing a knowledge graph in the field of aerospace docking include: Based on the engineering semantics of aerospace docking mechanisms, a domain ontology is established, defining the hierarchical relationships and attributes of the structural components, load conditions, deformation thresholds, and engineering rules. Cleaning, semantic slicing, and overlapping window division were performed on engineering documents related to aerospace docking mechanisms. The large language model is called to extract entities, relations and attributes from the sliced ​​text, the extraction results are organized into triples, and the extracted entities are used as candidate entities. An Embedding semantic alignment mechanism is introduced to vectorize candidate entities and calculate the semantic similarity between them. Candidate entities with semantic similarity exceeding a preset threshold are merged into the same entity, eliminating redundant entities and their corresponding redundant triples. The processed triples are written into a graph database to form a knowledge graph of the aerospace docking domain containing nodes and edges.

[0009] Based on the engineering semantics of aerospace docking mechanisms, a domain ontology is established, defining the hierarchical relationships and attributes of structural components, load conditions, deformation thresholds, and engineering rules, so that the knowledge graph can accurately represent the intrinsic relationship between the structural composition and test parameters of aerospace docking mechanisms.

[0010] By calling a large language model to extract entities, relations, and attributes from engineering documents and organize them into triples, the automated knowledge extraction and structured representation of engineering knowledge in the aerospace docking field is realized. Compared with the method of manually building a knowledge base, the construction cost is greatly reduced and the construction efficiency of knowledge graph is improved.

[0011] An embedding semantic alignment mechanism is introduced to vectorize candidate entities and merge entities based on semantic similarity. This resolves semantic conflicts and redundancy issues caused by inconsistencies in synonyms, abbreviations, and naming in engineering documents, improving the data quality and retrieval accuracy of the knowledge graph. The processed triples are written into a graph database to form a knowledge graph containing nodes and edges, supporting subsequent knowledge retrieval through entity links, Cypher queries, and multi-hop path expansion. This enables agents to quickly obtain engineering constraints such as structural composition, load boundaries, material thresholds, and failure modes when receiving experimental tasks.

[0012] As a further limitation of the technical solution of the present invention, the step of constructing the twin agent includes: A modular workflow is constructed based on a large language model. The workflow includes at least a user input receiving node, a task parsing node, a knowledge graph retrieval node, a model invocation node, a risk assessment node, and a strategy generation node. The workflow enables data integration between the natural language interaction interface and the aerospace docking domain knowledge graph.

[0013] The system achieves modular orchestration of the twin agent's functions, with parameters passed between nodes through a unified data field. This ensures smooth data flow and logical consistency between the agent and other stages, including task parsing, knowledge retrieval, state prediction, risk assessment, and strategy generation. The agent can complete the entire process from natural language understanding to engineering constraint retrieval and structured output within a unified interaction framework, ensuring that the agent's task parsing process is always constrained by domain knowledge.

[0014] As a further limitation of the technical solution of the present invention, the steps of the twin agent combining the engineering constraints in the knowledge graph through the knowledge retrieval mechanism include: A graph retrieval-enhanced generation mechanism is adopted, which performs entity linking, Cypher query, and multi-hop path expansion after receiving the natural language experiment task; The constraint subgraphs related to the current task are extracted from the knowledge graph, and the corresponding original engineering text fragments are recalled to serve as constraints for generating structural engineering condition parameters.

[0015] By extracting constraint subgraphs relevant to the current task from the knowledge graph, the problem of large language models lacking specialized knowledge of aerospace docking in general domains is overcome, making the parsing results and strategy suggestions generated by the agent traceable in engineering. Recalling the corresponding original engineering text fragments and using them together with the constraint subgraph as constraints for generating structured engineering condition parameters achieves dual enhancement of structured knowledge retrieval and unstructured text retrieval, ensuring the comprehensiveness of knowledge acquisition and providing semantic basis for the agent's reasoning, thus avoiding the problem of insufficient generalization ability caused by purely data-driven approaches.

[0016] As a further limitation of the technical solution of the present invention, the step of parsing the natural language experimental task into structured chemical condition parameters includes: A large language model is invoked to perform intent recognition and key information extraction for a natural language experimental task, thereby obtaining key information. Through the knowledge retrieval mechanism, engineering constraint subgraphs and original text fragments that match the key information are retrieved from the knowledge graph. The natural language experiment task, the engineering constraint subgraph, and the original text fragment are input into the large language model. The large language model is used for logical reasoning and format conversion to generate structural engineering condition parameters that conform to the preset data structure. The structural engineering condition parameters include at least: load direction vector and load magnitude, boundary conditions and constraint states of the docking mechanism, position identifiers of target monitoring nodes, and corresponding allowable deformation thresholds.

[0017] Experimenters no longer need to memorize complex parameter formats and instruction specifications, lowering the operational threshold. By inputting the natural language experimental task, the engineering constraint subgraph obtained from the knowledge graph retrieval, and the original engineering text fragments into the large language model for logical reasoning and format conversion, structured parameter generation under knowledge graph constraints is achieved. This provides a standardized input interface for subsequent invocation of the multi-fidelity structured response proxy model, ensuring seamless integration between the twin agent's natural language understanding module and the proxy model's prediction module.

[0018] As a further limitation of the technical solution of the present invention, the step of calling a pre-built multi-fidelity structural response surrogate model to predict the full-field displacement of the aerospace docking mechanism includes: The structural chemical condition parameters and sparse monitoring response are used as input features and input into the multi-fidelity structural response proxy model. The multi-fidelity structural response proxy model, based on the input features, outputs the displacement increment of each discrete node of the aerospace docking mechanism through the fusion operation of physical prior calculation and residual correction. The initial coordinates of each discrete node are obtained, and the displacement increment is superimposed on the corresponding initial coordinates to generate the updated coordinates of each discrete node, thereby obtaining the overall displacement prediction result.

[0019] This approach overcomes the problem of insufficient generalization ability of purely data-driven models under unknown conditions, while also compensating for the deviation between finite element models and actual responses caused by simplification assumptions and preset parameters. It achieves the reconstruction of the entire field displacement distribution from sparse measurement point data, solving the technical challenge of obtaining complete structural state information due to the limited number of sensors and sparse monitoring data in ground tests of aerospace docking mechanisms.

[0020] As a further limitation of the technical solution of the present invention, the multi-fidelity structural response proxy model includes an RBF physical prior branch, a gMLP residual correction branch, a spatial gating unit, and a gating fusion layer. The RBF physical prior branch is used to learn the basic displacement field mapping based on finite element simulation data; The spatial gating unit is used to extract the spatial correlation between different nodes; The gMLP residual correction branch is used to learn nonlinear residuals based on measured data and to correct them in combination with the spatial correlation extracted by the spatial gating unit. The gated fusion layer is used to nonlinearly fuse the physical prediction results output by the RBF physical prior branch with the high-dimensional feature responses output by the gMLP residual correction branch to obtain the full-field displacement prediction results.

[0021] By extracting spatial correlations between different nodes using spatial gating elements, the surrogate model can capture the spatial continuity and transmission patterns of structural deformation, improving the spatial consistency of the overall displacement prediction results. Furthermore, by using the gMLP residual correction branch to learn nonlinear residuals based on measured data and combining this with the spatial correlations extracted by the spatial gating elements, the surrogate model can effectively learn the nonlinear deviations between finite element simulation results and actual experimental responses, improving the model's adaptability to real experimental conditions.

[0022] As a further limitation of the technical solution of the present invention, the pre-construction process of the multi-fidelity structural response proxy model includes the following training stages: Finite element physical prior training stage: Using finite element simulation data as samples, the RBF physical prior branch is trained to enable the model to obtain the basic displacement field prediction capability based on finite element analysis. Measured residual correction stage: Freeze the parameters of the RBF physical prior branch, train the gMLP residual correction branch and the spatial gating unit using historical test measurement data, and learn the nonlinear residual between the finite element prediction results and the measured response; Online update phase: Structural response data is continuously collected through key monitoring nodes deployed on the aerospace docking mechanism, the parameters of the RBF physical prior branch are frozen, a local residual loss function is constructed using the real-time monitoring data as a sample, and based on the local residual loss function, only the weight matrix in the gated fusion layer is updated with a small step gradient.

[0023] During the online update phase, real-time monitoring data from key monitoring nodes is used to perform small-step gradient updates on the weight matrix of the gated fusion layer. This enables the surrogate model to continuously adapt to current experimental boundary disturbances and unknown operating conditions. Meanwhile, by freezing the underlying RBF physical parameters, the physical stability of the model is ensured, preventing overall model drift due to overfitting to real-time noise.

[0024] As a further limitation of the technical solution of the present invention, the step of extracting the maximum deformation value based on the predicted full-field displacement, and determining that a risk has been identified if the maximum deformation value exceeds the allowable deformation threshold, and generating candidate working condition adjustment schemes based on the constraints of the knowledge graph includes: The three-dimensional displacement vectors of each discrete node in the full-field displacement prediction results are traversed, and the absolute deformation of each node is calculated. Extract the maximum value from the absolute deformation amounts as the maximum deformation value, and obtain the spatial coordinate identifier of the corresponding node; The maximum deformation value is compared with a preset allowable deformation threshold. If the maximum deformation value is greater than the allowable deformation threshold, a risk is identified, and the spatial coordinate identifier and the corresponding load parameter are marked as a risk source. Extract the sensitive load parameters or boundary conditions that cause the maximum deformation value to exceed the allowable deformation threshold; The knowledge retrieval mechanism is used to recall mechanical consistency rules and material safety thresholds related to the sensitive load parameters from the knowledge graph. The sensitive load parameters, the mechanical consistency rules, and the material safety threshold are input into the large language model, which then generates at least one candidate working condition adjustment scheme within a preset engineering feasible domain.

[0025] By extracting the maximum deformation value and obtaining the spatial coordinates of the corresponding nodes, high-risk areas were accurately located, allowing test personnel to intuitively understand the most dangerous parts of the structure. The spatial coordinates and corresponding load parameters were marked as risk sources, and sensitive load parameters or boundary conditions that caused the maximum deformation value to exceed the allowable deformation threshold were extracted, enabling risk source analysis. A knowledge retrieval mechanism was used to recall mechanical consistency rules and material safety thresholds related to sensitive load parameters from a knowledge graph. These sensitive load parameters, mechanical consistency rules, and material safety thresholds were then input into a large-scale language model to generate candidate working condition adjustment schemes within a preset engineering feasible domain. This ensured that the generated adjustment schemes were subject to the dual constraints of mechanical consistency rules and material safety thresholds, avoiding the problem of schemes generated by a general large-scale language model exceeding the engineering safety boundary.

[0026] As a further limitation of the technical solution of the present invention, the multi-fidelity structural response proxy model is invoked again to verify the candidate operating condition adjustment scheme until the safety conditions are met, specifically including: The candidate working condition adjustment scheme is used as a new structural working condition parameter and input into the multi-fidelity structural response proxy model to obtain the corresponding verification full-field displacement. Extract the maximum deformation value from the verified full-field displacement, and determine whether the maximum deformation value does not exceed the allowable deformation threshold. If the value does not exceed the limit, the safety condition is deemed met, and the candidate working condition adjustment scheme is output as the final optimization result. If the limit is exceeded, the twin agent is triggered to regenerate the next candidate working condition adjustment scheme and repeat the verification process until the preset maximum number of iterations is reached or the safety conditions are met.

[0027] It can automatically search for operating condition solutions that meet safety requirements without human intervention. When the maximum number of iterations is reached and the safety conditions still cannot be met, the iteration stops and prompts for manual review, ensuring the system's safety and engineering practicality. It effectively solves the shortcomings of existing digital twin systems, which mostly stay at the level of status display and offline analysis and lack the ability to generate strategies and perform closed-loop verification. This upgrades the system from a simple status monitoring tool to an intelligent test assistance system with autonomous decision-making and continuous optimization capabilities.

[0028] As can be seen from the above technical solutions, this application has the following advantages: It integrates domain knowledge into the agent's reasoning process in a structured form, enabling the agent to understand the engineering semantics of the aerospace docking mechanism. This provides a traceable engineering constraint basis for subsequent task analysis, risk identification, and strategy generation, avoiding the generation of strategies that do not meet the engineering feasible domain due to the lack of domain constraints in general-purpose large language models. The twin agent receives natural language experimental tasks and automatically parses them into structured chemical condition parameters. Experiment personnel do not need professional simulation modeling knowledge to complete the condition configuration through natural language interaction, reducing the operational threshold of manual analysis in ground experiments and improving the efficiency and user-friendliness of experimental task configuration.

[0029] The system employs automatic risk identification based on maximum deformation value and allowable deformation threshold, along with closed-loop optimization through knowledge graph-based candidate adjustment scheme generation and re-verification of multi-fidelity proxy models. When structural deformation exceeds limits, the system can automatically generate adjustment schemes that meet engineering constraints and verify their effectiveness, reducing manual trial and error and repeated analysis, thus improving experimental efficiency and safety. Attached Figure Description

[0030] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart illustrating the method provided by the present invention.

[0032] Figure 2 This is a diagram illustrating a specific implementation process of the present invention. Detailed Implementation

[0033] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0035] This invention proposes a method for state reconstruction and optimization of aerospace docking mechanisms based on a twin agent, comprising a physical simulation layer, a cognitive interaction layer, and a strategy guidance layer. The physical simulation layer integrates finite element simulation data, ground test sensor data, and real-time monitoring data, and achieves rapid full-field displacement prediction through a multi-fidelity structural response proxy model. The cognitive interaction layer transforms natural language test tasks into structural working conditions and uses GraphRAG to call upon aerospace docking domain knowledge graphs, providing the agent with constraints such as structural composition, load boundaries, material thresholds, and failure rules. The strategy guidance layer, upon detecting excessive structural deformation or local risks, calls a large language model to generate working condition adjustment schemes and verifies them again through a proxy model, thus forming a closed loop of test task analysis, model invocation, state reconstruction, risk identification, and strategy optimization. Figure 1 and Figure 2 As shown, it includes the following steps: S1. Construct a twin intelligent agent and a knowledge graph in the field of aerospace docking, wherein the knowledge graph contains entities and relationships of structural components, load conditions, deformation thresholds and engineering rules; A space docking mechanism was selected as the physical test object, and a corresponding ground-based buffer test platform was built. The space docking mechanism includes at least a docking ring, buffer components, a connection mechanism, and a load-bearing path structure. According to the test mission requirements, sensors were deployed at key locations on the docking ring, buffer components, connection points, load-bearing path, and areas prone to concentrated local deformation to collect structural response data of the docking mechanism under different load directions, load magnitudes, and boundary conditions. The data collected by the sensors includes at least one of strain, displacement, and force.

[0036] Simultaneously, a three-dimensional geometric model was established based on the actual structural dimensions, component connection methods, and assembly relationships of the aerospace docking mechanism, and a corresponding finite element analysis model was built in finite element software. Material parameters, contact relationships, boundary conditions, and load application methods were set in the finite element model to obtain full-field displacement response data under different typical working conditions. A set of typical working conditions covering the parameter space was generated through Latin hypercube sampling. The full-field displacement field under each working condition was calculated using the finite element model, and response data from measurement points under the same or similar working conditions were collected through a ground-based buffer test platform. These data collectively form the data foundation for subsequent digital twin mapping and RgMN surrogate model training. Specifically, finite element simulation data serves as low-fidelity full-field response data, ground test measurement point data serves as high-fidelity measured data, and online monitoring data serves as the data source for subsequent model correction and state reconstruction.

[0037] After completing the preparation of the test object, finite element model, and initial test data, a digital test twin intelligent agent for the aerospace docking mechanism was constructed. This digital test twin intelligent agent includes a 3D visualization model, a finite element analysis model, test measurement point information, a structural response database, node mapping relationships, and backend prediction services.

[0038] The 3D geometric model of the aerospace docking mechanism is imported into the digital experimental twin intelligent agent visualization environment to form an interactive 3D display object. Then, the node coordinates, element information, load conditions, and displacement response results in the finite element model are exported and their correspondences are established with the mesh nodes in the target model of the digital experimental twin intelligent agent. For the test sensor measuring points, their installation positions on the physical structure are recorded and mapped to the corresponding positions in the finite element model and the digital twin model.

[0039] Since the finite element source mesh and the target mesh of the digital twin are not completely identical, a unified node mapping mechanism is established using the KNN-IDW interpolation method: first, several neighboring nodes near the target node are searched in the finite element source mesh; then, inverse distance weights are set according to the distance, and the displacement value on the target node is calculated. This mechanism is used to map the finite element full-field displacement results or the node displacement results output by the surrogate model onto the digital twin 3D model. After the mapping is completed, the digital twin system updates the node coordinates of the 3D model according to the predicted displacement and displays the deformation magnitude of different regions through color mapping.

[0040] After completing the construction of the digital experimental twin agent, the finite element simulation data, historical test measurement point data, and online monitoring data were processed in a unified manner. The data from different sources were converted into different formats, and the fields of working condition parameters, measurement point numbers, node coordinates, displacement response, and time information were standardized so that they could be used by the subsequent agent model and agent workflow.

[0041] For finite element full-field displacement data, the coordinates of the finite element source mesh and the digital twin target mesh are normalized to eliminate the influence caused by inconsistencies in coordinate scales and origins between different models. For historical experimental measurement point data, a mapping relationship between measurement points and target mesh nodes is established based on the sensor placement. If the measurement point location does not completely coincide with the target mesh node, a neighboring node is selected or the corresponding location is determined through interpolation. For online monitoring data, a real-time data receiving format is set so that it can be used as model input or online correction basis during system operation.

[0042] A multi-fidelity structural response proxy model (RgMN) is constructed. This model serves as a structural response prediction tool invoked by a twin agent. Its inputs are operating parameters and sparse monitoring responses, and its output is the full-field displacement prediction results on the target mesh of the aerospace docking mechanism. The RgMN model includes an RBF physical prior branch, a gMLP residual correction branch, space-gated cells, and a gated fusion layer.

[0043] The RBF physical prior branch is used to learn the mapping relationship between load cases and foundation displacement fields in finite element simulation samples. K-means clustering is performed on the joint input features of load cases and space, which consist of normalized load case parameters and target mesh node coordinates. The cluster centers are selected as the initial centers of the RBF physical prior branch, ensuring that the RBF centers cover the main structural response modes. The expression for the RBF physical prior branch is: (1) in, , This is a vector of structural chemical condition parameters; For the first Normalized coordinates of each target grid node; This is a combined input feature of working conditions and space; The number of radial basis functions. Output weights for three-dimensional displacement. For the first One RBF center; For kernel width; For the first The physical prior displacement of each node.

[0044] The gMLP residual correction branch is used to learn the nonlinear deviation between the finite element simulation results and the actual experimental response. Spatial gated elements are used to extract the spatial correlation between different nodes. The core calculation process of the gMLP residual correction branch is as follows: (2) in, The input feature matrix for gMLP; and These are the input projection weights and biases, respectively; For activation function, These are the hidden features after splitting; This is the spatial gating weight matrix; The bias term of the space gating unit; SGU is the space gating unit; For normalization operators; This represents the Hadamard product. and To output projection parameters; This is the high-fidelity response matrix output by the residual correction branch of gMLP.

[0045] The gated fusion layer is used to nonlinearly fuse the physical prediction results output by the RBF physical prior branch with the high-dimensional feature responses output by the gMLP residual correction branch to obtain the full-field displacement prediction results: (3) in, This represents the final predicted total displacement. The output of the dynamic gain function is used to adjust the contribution intensity of the RBF physical prior branch at each node and in each displacement direction according to the current working condition. This is the physical prior displacement output by the RBF branch. ; This is a bias compensation term used to correct systematic errors.

[0046] After obtaining the predicted nodal displacements, a nodal offset reconstruction method is further used to recover the complete displacement field distribution. A finite element method with nodal nodes is provided. The initial coordinates are Then its updated coordinates can be expressed as: (4) in, For the first Predicted displacement of each node; These are the updated node coordinates.

[0047] The training of the RgMN multifidelity structural response surrogate model is carried out in three stages.

[0048] The first stage is the finite element physical prior training stage. Finite element simulation data is used as the main training data, with typical working condition parameters as input and full-field displacement results mapped to the target mesh under the corresponding working conditions as output. By training the RBF physical prior branch, the model obtains the basic structural response characteristics of the docking mechanism under different load conditions, possessing a basic predictive capability close to the finite element analysis results.

[0049] The second stage is the measured residual correction stage. The RBF physical prior branch parameters obtained from the first stage training are loaded, and the main parameters of this branch are frozen. Then, the gMLP residual correction branch and spatial gating element are trained using historical experimental measurement data. Real experiments involve factors such as material parameter deviations, assembly errors, contact disturbances, and uncertainties in boundary conditions, often resulting in discrepancies between the finite element prediction results and the measured response. Through this stage of training, the model learns the nonlinear residuals between the finite element results and the actual measurement point responses, improving its adaptability to real experimental conditions.

[0050] The third phase is the online update phase. To improve the model's adaptability to real-world engineering environments, structural response data is continuously collected through key monitoring nodes deployed on the aerospace docking mechanism. And construct the local residual loss function: (5) in, This refers to the online update time; The set of monitoring nodes participating in online updates; The number of monitoring nodes; This represents the total displacement predicted by the current model. Select a matrix for the nodes.

[0051] During the online update process, the underlying RBF physical parameters are frozen, and only the weight matrix in the SGU is updated. The small-step gradient update has the following iterative form: (6) in, η For learning rate, For the first The weight matrix during each iteration. The updated weight matrix, The loss function with respect to the weight matrix The gradient is calculated. After model training, RgMN is deployed as a callable backend prediction service. This service receives structural condition parameters and real-time measurement point data, and outputs the full-field displacement prediction results on the target mesh. Deployment can be done via a local service interface or an HTTP API interface. The preset input format for the model includes load direction, load magnitude, boundary conditions, key measurement point responses, and target output mesh information; the preset output format includes target mesh node displacements, maximum deformation values, key region responses, and node offset data. The RgMN proxy model becomes a structural response prediction tool that can be called by the twin agent.

[0052] A knowledge graph for the aerospace docking domain is constructed. This knowledge graph is used to provide domain knowledge and engineering constraints for aerospace docking mechanisms for twin agents.

[0053] A domain ontology is established based on the structural composition, testing procedures, and risk analysis requirements of the aerospace docking mechanism. Entity types within the domain ontology include structural components, material parameters, sensing points, load conditions, boundary conditions, deformation thresholds, failure modes, and engineering rules; relationship types include component connection relationships, load application relationships, measuring point association relationships, response association relationships, threshold constraint relationships, and failure induction relationship relationships.

[0054] The process involves organizing structural specification documents, experimental plans, finite element analysis reports, test records, and engineering experience documents. These documents are then cleaned, semantically sliced, and have overlapping windows defined. A large language model is used to extract entities, relations, and attributes from the sliced ​​text, and the extraction results are formatted as triples, with the extracted entities serving as candidate entities. To address the issues of synonyms, abbreviations, and inconsistent naming in the engineering documents, an embedding semantic alignment mechanism is introduced. Candidate entities are vectorized, and the semantic similarity between them is calculated. Candidate entities with semantic similarity exceeding a preset threshold are merged into a single entity, eliminating redundant entities and their corresponding redundant triples.

[0055] The processed triples are written into the Neo4j graph database to form a knowledge graph for the aerospace docking domain, containing nodes and edges. Entity and relation indexes are built on the graph database, and knowledge retrieval interfaces for entity linking, Cypher queries, and multi-hop path expansion are configured. This graph stores information such as structural composition, load boundaries, material thresholds, measurement point layout, failure modes, and historical control rules, providing a knowledge foundation for subsequent GraphRAG retrieval and agent reasoning.

[0056] After completing the knowledge graph construction, configure the GraphRAG retrieval enhancement module. This module is used to perform entity linking, Cypher queries, and multi-hop path expansion in the Neo4j knowledge graph after the agent receives the experimental task, based on the structural components, load directions, boundary conditions, monitoring objects, risk descriptions, and optimization objectives in the task.

[0057] The agent first extracts structured keywords from the user task, including target components, load direction, load magnitude, boundary states, and key metrics. The GraphRAG module links entities based on these keywords to determine their corresponding entities in the knowledge graph; then, it uses Cypher queries to obtain relevant entities and relationships; for constraint information with indirect relationships, it expands through multi-hop paths to obtain load boundaries, material thresholds, failure modes, and engineering rules. In addition to graph retrieval, the GraphRAG module also recalls original engineering text fragments relevant to the current task from the original engineering corpus based on semantic similarity. The graph subgraph provides structured constraints, while the text fragments provide semantic basis. Both types of information are input into the large language model, ensuring that the generated structured instructions and optimization suggestions have engineering basis and traceability.

[0058] The GraphRAG retrieval and generation process is represented as follows: (7) in, This is the current interactive command. For structured semantic representation, For conversation history, For semantic parsing operators, A structured knowledge graph built on Neo4j; This is a collection of original engineering corpora. This represents the graph retrieval process based on entity links, Cypher queries, and multi-hop path expansion. This represents the text context recall process based on semantic similarity. This represents the large language model generation process under the joint constraints of graph evidence and textual context. Adjust the original candidate work condition vectors generated for the large language model.

[0059] After completing the construction of the RgMN agent model, knowledge graph, GraphRAG retrieval module and digital twin system, a twin agent is built.

[0060] A modular workflow is constructed based on a large language model. This workflow includes at least a user input receiving node, a task parsing node, a knowledge graph retrieval node, a model invocation node, a risk assessment node, and a strategy generation node. This workflow facilitates data integration between the natural language interaction interface and the aerospace docking domain knowledge graph.

[0061] The twin agent combines the engineering constraints in the knowledge graph with a knowledge retrieval mechanism. Specifically, it adopts a graph retrieval-enhanced generation mechanism. After receiving the natural language test task, it performs entity linking, Cypher query, and multi-hop path expansion. It extracts the constraint subgraph related to the current task from the knowledge graph and recalls the corresponding original engineering text fragments, which together serve as constraints for generating structured engineering condition parameters.

[0062] S2. Receive natural language test task. The twin agent uses a knowledge retrieval mechanism combined with the engineering constraints in the knowledge graph to parse the natural language test task into structured engineering condition parameters. The twin agent receives experimental tasks input by the user through an interactive interface using natural language. These tasks include load direction, load magnitude, boundary conditions, monitoring locations, deformation thresholds, or optimization objectives. For example, the user can request the system to analyze the overall deformation of the docking mechanism under a specific load direction, or request the system to provide a working condition adjustment plan while ensuring the maximum deformation does not exceed a set threshold.

[0063] After receiving user input, the twin agent parses the natural language by combining conversation history and task parsing prompts. Specifically, the twin agent invokes a large language model to perform intent recognition and key information extraction for the natural language experimental task, obtaining key information; it then retrieves engineering constraint subgraphs and original text fragments matching the key information from the knowledge graph through a knowledge retrieval mechanism; finally, it inputs the natural language experimental task, engineering constraint subgraphs, and original text fragments into the large language model, using the model for logical reasoning and format conversion to generate structural engineering condition parameters conforming to a preset data structure. These structural engineering condition parameters include at least: load direction vector and load magnitude, boundary conditions and constraint states of the docking mechanism, location identifiers of target monitoring nodes, and corresponding allowable deformation thresholds.

[0064] After generating the structural engineering condition parameters, the twin agent performs engineering logic verification on these parameters to determine if they meet the basic mechanics consistency rules in the knowledge graph. If the verification fails, a message indicating missing parameters or logical conflicts is generated and fed back to the twin agent, triggering it to initiate follow-up questions to the user to supplement and improve the structural engineering condition parameters. The twin agent generates follow-up questions based on the specific type of failure, merges these questions with the user's supplementary input, and performs another engineering logic verification until the verification passes or the preset number of follow-up questions is reached.

[0065] S3. Call the pre-built multi-fidelity structural response proxy model to predict the full-field displacement of the aerospace docking mechanism based on the structural chemical condition parameters; The agent invokes the RgMN proxy model based on structured execution instructions. Specifically, structural chemical condition parameters and sparse monitoring responses are used as input features and fed into the multi-fidelity structural response proxy model. Based on the input features, the multi-fidelity structural response proxy model outputs the displacement increments of each discrete node of the aerospace docking mechanism through a fusion operation of physical prior calculations and residual corrections. The initial coordinates of each discrete node are obtained, and the displacement increments are superimposed on the corresponding initial coordinates to generate the updated coordinates of each discrete node, thereby obtaining the overall displacement prediction result.

[0066] The prediction results include the displacement values ​​of each target grid node, the response of key areas, the maximum deformation value, and node offset information for digital twin display. The backend prediction service sends the total node displacement results to the digital twin frontend. The digital twin system updates the node coordinates of the 3D model based on the node displacements and displays the deformation magnitude at different locations using color mapping. For areas with large local deformation, the system uses highlighted colors, risk markers, or text prompts to allow experimenters to visually observe the overall deformation trend and local high-risk areas.

[0067] Meanwhile, the agent extracts the maximum deformation value, key area deformation value, local deformation concentration location and key measurement point response from the full field displacement results, and uses these results as input for risk identification and safety determination.

[0068] S4. Extract the maximum deformation value based on the predicted full-field displacement. If the maximum deformation value exceeds the allowable deformation threshold, a risk is identified. Generate candidate working condition adjustment schemes based on the constraints of the knowledge graph, and call the multi-fidelity structural response proxy model again to verify the candidate working condition adjustment schemes until the safety conditions are met. The safety conditions include at least the maximum deformation value not exceeding the allowable deformation threshold.

[0069] After obtaining the full-field state reconstruction results, the twin agent compares the maximum deformation value, key area response, and local deformation concentration with the safety threshold and failure rules in the knowledge graph. Specifically, it traverses the three-dimensional displacement vectors of each discrete node in the full-field displacement prediction results and calculates the absolute deformation of each node; it extracts the maximum value of the absolute deformation as the maximum deformation value and obtains the spatial coordinate identifier of the corresponding node; it compares the maximum deformation value with the preset allowable deformation threshold. If the maximum deformation value is greater than the allowable deformation threshold, a risk is identified, and the spatial coordinate identifier and the corresponding load parameters are marked as risk sources.

[0070] Simultaneously, the localized high-deformation areas are matched and compared with failure modes in the knowledge graph. If the spatial distribution characteristics of the localized high-deformation areas match the failure modes recorded in the knowledge graph, the current operating condition is determined to be risky.

[0071] If the current prediction result does not exceed the allowable threshold and no local deformation concentration related to the failure mode occurs, the agent determines that the current operating condition meets the test requirements and outputs the current state reconstruction result, key response description, and test analysis conclusion. This result is also displayed in the digital twin system for test personnel to view.

[0072] If the prediction result exceeds the safety threshold, or if a locally highly deformable region matches a failure mode in the knowledge graph, the agent determines that the current operating condition is risky and enters the operating condition optimization phase. At this time, the system records the current operating condition parameters, risk location, exceedance indicators, and knowledge graph retrieval basis as input for subsequent strategy generation.

[0073] When the system identifies excessive structural deformation or local risks, the agent invokes a large language model to generate candidate working condition adjustment schemes. Specifically, it extracts sensitive load parameters or boundary conditions that cause the maximum deformation value to exceed the allowable deformation threshold; it retrieves mechanical consistency rules and material safety thresholds related to the sensitive load parameters from the knowledge graph through a knowledge retrieval mechanism; and it inputs the sensitive load parameters, mechanical consistency rules, and material safety thresholds into the large language model, which then generates at least one candidate working condition adjustment scheme within a preset engineering feasible domain.

[0074] To avoid large language models generating load adjustment schemes that violate engineering boundaries, a feasible region is defined, consisting of upper and lower load limits, mechanical consistency constraints, material thresholds, and knowledge graph rules. (8) in, For the first Current structural chemical condition parameters of the wheel; Adjustments for candidate operating conditions; and These are the load boundaries; This represents a physical constraint function consisting of mechanical consistency, stability, and safety thresholds. The current physical state; This represents the engineering rule constraint function obtained from the knowledge graph retrieval; This includes constraint subgraphs, relationship paths, material thresholds, and engineering rule evidence retrieved from the knowledge graph. Candidate solutions are generated based on the current risk type, overall displacement distribution, knowledge graph constraints, and engineering text, including reducing load magnitude, adjusting load direction, modifying boundary conditions, changing test parameters, or prompting manual review.

[0075] To ensure that candidate solutions meet engineering requirements, the system performs feasibility domain verification. If a candidate solution exceeds the feasibility domain, it is first corrected or projected to meet the engineering allowable range before being input as a new operating condition parameter into the subsequent verification process.

[0076] Subsequently, the agent inputs the new load condition parameters back into the RgMN surrogate model for rapid prediction. Specifically, the candidate load condition adjustment scheme is input as the new structural load condition parameter into the multi-fidelity structural response surrogate model to obtain the corresponding verification full-field displacement; the verification maximum deformation value is extracted from the verification full-field displacement, and it is determined whether the verification maximum deformation value does not exceed the allowable deformation threshold. If it does not exceed the threshold, the safety condition is deemed met, and the candidate load condition adjustment scheme is output as the final optimization result.

[0077] If the verification result still does not meet the safety requirements, the twin agent is triggered to regenerate the next candidate operating condition adjustment scheme, and the above verification process is repeated. The closed-loop reasoning and execution process can be represented as: (9) in, The engineering feasible region projection operator is used to restrict candidate solutions to the th... Wheel feasible domain Inside; The adjusted full-field predicted displacement matrix; For multi-fidelity structural response proxy model; For the first Three-dimensional predicted displacement of each node; To allow for deformation threshold, For safety assessment results; when When, it means the optimized state meets the safety requirements; when At that point, the system proceeds to the next round of strategy correction.

[0078] To ensure the stability of the closed-loop process, the system sets a maximum number of iterations. When a working condition that meets the safety requirements is obtained within the maximum number of iterations, the system outputs this solution as a recommended result. If the safety requirements are still not met after reaching the maximum number of iterations, the system outputs the final prediction result and generates structural-level engineering suggestions based on the knowledge graph, prompting the test personnel to reduce the load, adjust the boundary conditions, or conduct manual verification.

[0079] The above-described specific embodiments fully implement the method flow of this invention. Through the digital twin system, test personnel can observe the predicted distribution of the full-field displacement of the docking mechanism in real time within a three-dimensional visualization environment, avoiding the incomplete state information problem caused by relying solely on limited measurement point data in traditional methods. Through the closed-loop optimization mechanism of the intelligent agent, the system can automatically generate and verify working condition adjustment schemes after risk identification, reducing the number of manual trial and error attempts and repeated analyses, and improving test efficiency. The introduction of knowledge graph constraints ensures that the strategy generated by the intelligent agent always remains within the engineering safety boundary, avoiding the problem of non-compliance with the physical feasible region that may arise from general large language models. Through the physical prior and data-driven fusion architecture of the RgMN multi-fidelity agent model, rapid reconstruction of the full-field displacement under sparse measurement point conditions is achieved, providing complete and real-time state data support for risk identification and strategy optimization.

[0080] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for state reconstruction and optimization of aerospace docking mechanisms based on twin intelligent agents, characterized in that, Includes the following steps: Construct a twin intelligent agent and a knowledge graph for the aerospace docking field, wherein the knowledge graph contains entities and relationships of structural components, load conditions, deformation thresholds and engineering rules; Upon receiving a natural language experiment task, the twin agent uses a knowledge retrieval mechanism combined with engineering constraints in the knowledge graph to parse the natural language experiment task into structured engineering condition parameters. The pre-built multi-fidelity structural response proxy model is invoked to predict the full-field displacement of the aerospace docking mechanism based on the structural chemical condition parameters. The maximum deformation value is extracted based on the predicted full-field displacement. If the maximum deformation value exceeds the allowable deformation threshold, a risk is identified. Candidate working condition adjustment schemes are generated based on the constraints of the knowledge graph, and the multi-fidelity structural response proxy model is called again to verify the candidate working condition adjustment schemes until the safety conditions are met. The safety conditions include at least the maximum deformation value not exceeding the allowable deformation threshold.

2. The method for state reconstruction and optimization of aerospace docking mechanisms based on twin agents according to claim 1, characterized in that, The steps involved in constructing a knowledge graph for the aerospace docking field include: Based on the engineering semantics of aerospace docking mechanisms, a domain ontology is established, defining the hierarchical relationships and attributes of the structural components, load conditions, deformation thresholds, and engineering rules. Cleaning, semantic slicing, and overlapping window division were performed on engineering documents related to aerospace docking mechanisms. The large language model is called to extract entities, relations and attributes from the sliced ​​text, the extraction results are organized into triples, and the extracted entities are used as candidate entities. An Embedding semantic alignment mechanism is introduced to vectorize candidate entities and calculate the semantic similarity between them. Candidate entities with semantic similarity exceeding a preset threshold are merged into the same entity, eliminating redundant entities and their corresponding redundant triples. The processed triples are written into a graph database to form a knowledge graph of the aerospace docking domain containing nodes and edges.

3. The method for state reconstruction and optimization of aerospace docking mechanisms based on twin agents according to claim 1, characterized in that, The steps to construct a twin agent include: A modular workflow is constructed based on a large language model. The workflow includes at least a user input receiving node, a task parsing node, a knowledge graph retrieval node, a model invocation node, a risk assessment node, and a strategy generation node. The workflow enables data integration between the natural language interaction interface and the aerospace docking domain knowledge graph.

4. The method for state reconstruction and optimization of aerospace docking mechanisms based on twin agents according to claim 3, characterized in that, The steps taken by the twin agent to combine engineering constraints in the knowledge graph through a knowledge retrieval mechanism include: A graph retrieval-enhanced generation mechanism is adopted, which performs entity linking, Cypher query, and multi-hop path expansion after receiving the natural language experiment task; The constraint subgraphs related to the current task are extracted from the knowledge graph, and the corresponding original engineering text fragments are recalled to serve as constraints for generating structural engineering condition parameters.

5. The method for state reconstruction and optimization of aerospace docking mechanisms based on twin agents according to claim 1, characterized in that, The steps for parsing the natural language test task into structured chemical condition parameters include: A large language model is invoked to perform intent recognition and key information extraction for a natural language experimental task, thereby obtaining key information; Through the knowledge retrieval mechanism, engineering constraint subgraphs and original text fragments that match the key information are retrieved from the knowledge graph. The natural language experiment task, the engineering constraint subgraph, and the original text fragment are input into the large language model. The large language model is used for logical reasoning and format conversion to generate structural engineering condition parameters that conform to the preset data structure. The structural engineering condition parameters include at least: load direction vector and load magnitude, boundary conditions and constraint states of the docking mechanism, position identifiers of target monitoring nodes, and corresponding allowable deformation thresholds.

6. The method for state reconstruction and optimization of aerospace docking mechanisms based on twin agents according to claim 1, characterized in that, The steps for predicting the full-field displacement of a space docking mechanism by invoking a pre-built multi-fidelity structural response surrogate model include: The structural chemical condition parameters and sparse monitoring response are used as input features and input into the multi-fidelity structural response proxy model. The multi-fidelity structural response proxy model, based on the input features, outputs the displacement increment of each discrete node of the aerospace docking mechanism through the fusion operation of physical prior calculation and residual correction. The initial coordinates of each discrete node are obtained, and the displacement increment is superimposed on the corresponding initial coordinates to generate the updated coordinates of each discrete node, thereby obtaining the overall displacement prediction result.

7. The method for state reconstruction and optimization of aerospace docking mechanisms based on twin agents according to claim 1, characterized in that, The multifidelity structural response surrogate model includes the RBF physical prior branch, the gMLP residual correction branch, the spatial gating unit, and the gating fusion layer; The RBF physical prior branch is used to learn the basic displacement field mapping based on finite element simulation data; The spatial gating unit is used to extract the spatial correlation between different nodes; The gMLP residual correction branch is used to learn nonlinear residuals based on measured data and to correct them in combination with the spatial correlation extracted by the spatial gating unit. The gated fusion layer is used to nonlinearly fuse the physical prediction results output by the RBF physical prior branch with the high-dimensional feature responses output by the gMLP residual correction branch to obtain the full-field displacement prediction results.

8. The method for state reconstruction and optimization of aerospace docking mechanisms based on twin agents according to claim 7, characterized in that, The pre-construction process of the multi-fidelity structural response surrogate model includes the following training phases: Finite element physical prior training stage: Using finite element simulation data as samples, the RBF physical prior branch is trained to enable the model to obtain the basic displacement field prediction capability based on finite element analysis. Measured residual correction stage: Freeze the parameters of the RBF physical prior branch, train the gMLP residual correction branch and the spatial gating unit using historical test measurement data, and learn the nonlinear residual between the finite element prediction results and the measured response; Online update phase: Structural response data is continuously collected through key monitoring nodes deployed on the aerospace docking mechanism, the parameters of the RBF physical prior branch are frozen, a local residual loss function is constructed using real-time monitoring data as samples, and based on the local residual loss function, only the weight matrix in the gated fusion layer is updated with a small step gradient.

9. The method for state reconstruction and optimization of aerospace docking mechanisms based on twin agents according to claim 1, characterized in that, The steps of extracting the maximum deformation value based on the predicted full-field displacement, and determining that a risk has been identified if the maximum deformation value exceeds the allowable deformation threshold, and generating candidate working condition adjustment schemes based on the constraints of the knowledge graph, include: Traverse the three-dimensional displacement vectors of each discrete node in the full-field displacement prediction results and calculate the absolute deformation of each node. Extract the maximum value from the absolute deformation amounts as the maximum deformation value, and obtain the spatial coordinate identifier of the corresponding node; The maximum deformation value is compared with a preset allowable deformation threshold. If the maximum deformation value is greater than the allowable deformation threshold, a risk is identified, and the spatial coordinate identifier and the corresponding load parameter are marked as a risk source. Extract the sensitive load parameters or boundary conditions that cause the maximum deformation value to exceed the allowable deformation threshold; The knowledge retrieval mechanism is used to recall mechanical consistency rules and material safety thresholds related to the sensitive load parameters from the knowledge graph. The sensitive load parameters, the mechanical consistency rules, and the material safety threshold are input into the large language model, which then generates at least one candidate working condition adjustment scheme within a preset engineering feasible domain.

10. The method for state reconstruction and optimization of aerospace docking mechanisms based on twin agents according to claim 1, characterized in that, The multi-fidelity structural response proxy model is invoked again to verify the candidate operating condition adjustment scheme until the safety conditions are met, specifically including: The candidate working condition adjustment scheme is used as a new structural working condition parameter and input into the multi-fidelity structural response proxy model to obtain the corresponding verification full-field displacement. Extract the maximum deformation value from the verified full-field displacement, and determine whether the maximum deformation value does not exceed the allowable deformation threshold. If the value does not exceed the limit, the safety condition is deemed met, and the candidate working condition adjustment scheme is output as the final optimization result. If the limit is exceeded, the twin agent is triggered to regenerate the next candidate working condition adjustment scheme and repeat the verification process until the preset maximum number of iterations is reached or the safety conditions are met.

Images