PDDL space mission modeling method and system based on enhanced constraint large model

Abstract

The invention provides a PDDL space mission modeling method and system based on an enhanced constraint large model, and relates to the technical field of space mission planning. According to the modeling method, the time sequence dependence, the resource upper limit and the probabilistic constraint are uniformly expressed through a semantic-constraint combined state extension diagram, so that the expression ability of the model to a complex space task is effectively improved; through a local increment updating mechanism, local modification can be performed on the premise of keeping an original structure when a task changes, so that the dynamic adaptability of task modeling is improved; through a predicate extension module and a feedback learning mechanism, the system has the capability of continuously absorbing new task types, platform differences and spatial-temporal characteristics, and has evolvability; through direct visualization of a graph structure, large model behaviors are transparent; by introducing a form, semantics and simulation triple verification mechanism, problems can be detected in advance, more errors can be found and automatically repaired before planning, and the reliability of the model is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of space mission planning, and particularly relates to a PDDL space mission modeling method and system based on an enhanced constraint large model. BACKGROUND

[0002] Space mission planning is a core technology to ensure the efficient and reliable operation of spacecraft. In the early space activities, mission planning was mainly focused on single satellites or a small number of satellites, and usually adopted a serial working mode of "demand-annotation-execution-feedback-processing". The task objectives were relatively independent, and the response rhythm was relatively fixed. However, with the significant increase in the number of on-orbit spacecraft and the diversification of mission types, the focus of space mission planning has gradually expanded from a single platform to the global coordination of multiple satellites and multiple tasks. Not only is it necessary to integrate the scheduling of on-orbit services such as earth observation, communication relay, and navigation enhancement, but it is also necessary to coordinate new types of tasks such as deep space exploration, on-orbit maintenance, and space experiments, to achieve system-level resource optimization and task coordination. Under this background, the mission planning system needs to face multiple challenges: on the one hand, it needs to coordinate heterogeneous spacecraft platforms, multiple types of payloads, and limited satellite-ground resources to meet the strong constraint conditions brought by multiple time windows and task conversion sequences of observation, data transmission, and other tasks; on the other hand, it also needs to adapt to uncertain factors such as changes in the space environment, fluctuations in the state of the load, and the insertion of high-priority tasks, to achieve dynamic response and rapid re-planning. This makes space mission planning gradually evolve into a complex decision-making problem with high dynamics and strong constraints.

[0003] As a standardized planning problem modeling language, Planning Domain Definition Language (PDDL) can separate domain knowledge from specific problems and is also applicable to space mission planning, which is a complex problem. However, its rigorous syntax also constitutes a usage threshold, resulting in the dependence of model construction and updating on professional programmers, which is inefficient in high-dynamic scenarios and requires repeated iterations in strong constraint scenarios. The emergence of large models provides a new path to solve this bottleneck: it can replace the human brain to directly understand natural language task instructions containing priorities, space-time constraints, and automatically convert them into structured PDDL problem definition files.

[0004] The existing PDDL domain modeling method based on large models still has problems such as insufficient constraint expression ability, poor model dynamic adaptability, lack of uncertainty modeling and process explainability guarantee in the context of space mission planning, which are specifically manifested in: Limited constraint expression ability: the original technology mainly generates basic action predicates, and the constraint form is limited to single-step logical relationships, which cannot describe multi-dimensional time sequence dependencies and resource competition relationships, as well as probabilistic constraints such as cloud layers, weather, and communication interference.

[0005] Weak dynamic adaptability of the model: the original technology is static generation of PDDL model, when the task environment or parameter changes, the whole task description must be input again, and the PDDL model must be generated again.

[0006] Task subject change leads to dynamic change of constraint structure: in space task planning, different spacecrafts, different platforms, different orbit types and different ground station configurations will cause significant changes in task constraints. Since the traditional PDDL domain model is constructed by using static predicates and fixed action templates, its constraint structure cannot be automatically adapted according to different task subjects.

[0007] The generation process lacks explainability: the large model output process of the original technology is completely black-boxed, and the generated PDDL model is difficult to trace the basis for generation.

[0008] The model consistency and verification ability is insufficient: the original technology relies on manual or single round verification tools (such as VAL) to detect syntax errors, but cannot detect logical consistency and constraint conflict.

[0009] These problems make it difficult for the generated PDDL model to accurately describe the complex timing and resource relationship among multiple tasks, and it is also difficult to quickly adjust and update when the task environment or parameters change, thereby limiting the depth of application of the large model in space task planning. SUMMARY

[0010] In view of the shortcomings of the prior art, the present application provides a PDDL space task modeling method and system based on an enhanced constraint large model, which solves the problem that the PDDL space task modeling method based on a large model cannot accurately describe the complex timing and resource relationship among multiple tasks, and cannot quickly adjust and update when the task environment or parameters change.

[0011] To achieve the above purpose, the present application is realized by the following technical scheme: A PDDL space task modeling method based on an enhanced constraint large model, the modeling method comprising: S1. Space knowledge base retrieval Combine knowledge graph and vector retrieval enhancement mechanism to automatically complete the prompt words of the large model and improve its domain understanding; at the same time, provide real-time state information of task related entities to ensure the parameter accuracy of the subsequent generated PDDL domain model; S2. Multi-modal task input Change the multi-modal input into a unified task package interface standard, all modules only need to parse the standard JSON schema, which can process different input sources; S3. Constraint enhancement The actions and related constraints involved in the task are expressed using a graph structure, which makes the generation process both interpretable and able to automatically adapt to the dynamic task environment. S4.PDDL Domain Model Generation Based on the enhanced sequence constraint intermediate representation, the automatic transformation from natural language task semantics to PDDL planning models is achieved through semantic alignment, predicate normalization, and controlled predicate generation mechanisms; and dedicated domain models are dynamically generated for each task based on the general domain model. S5. Verification and Closed-Loop Correction By combining formal verification, simulation verification, and semantic verification, the system automatically detects logical errors, resource conflicts, and semantic deviations in the planning model. It also achieves automatic model repair through a correction strategy generator based on a large language model. Furthermore, it constructs a closed-loop feedback system across modules to send the verification results back to the upstream module for updating the knowledge base data and prompts for each module, thereby enabling continuous optimization and adaptive evolution of the model.

[0012] Preferably, in S1, the aerospace knowledge base retrieval module includes: a graph database, a vector database, and a state database; The graph database is used to integrate structured knowledge in the fields of aerospace telemetry and control, mission planning, orbital mechanics, and resource constraints, and organizes and stores it in the form of a knowledge graph; where nodes in the knowledge graph represent entities in the aerospace field; and edges represent relationships between entities. The vector database is used to store unstructured semantic information related to aerospace mission planning. Semantic text fragments are converted into vector representations by an embedding model and then stored in the vector database. The status database is used to summarize and maintain real-time or near-real-time status information of spacecraft and telemetry, tracking and command (TT&C) related entities. The data required for maintenance is provided directly by other departments or obtained from previous input updates.

[0013] Preferably, step S1 specifically includes: extracting entity names from the module input, searching for corresponding nodes in the graph database, and obtaining structured knowledge of the neighborhood. ; Embed the module input as a vector The module calculates the cosine similarity between the input vector and the vector in the vector database, and determines whether the semantic text similarity is greater than a set threshold. semantic text The similarity calculation formula is: The retrieved structured knowledge and semantic text are fused into a context-enhanced text for direct reading by the large model. The fusion formula is as follows: in, For the first structured knowledge; For the first semantic text; For the first weight coefficient / confidence of structured knowledge; For the first weight coefficient / confidence of semantic text; If there is a description about the state of the entity in the module input, it will be recognized and extracted by the large model for updating the state database in the space knowledge base retrieval module.

[0014] Preferably, the S2 specifically comprises: S2.1. Natural language input and retrieval enhancement Input the natural language or structured task in the module into the space knowledge base retrieval module to obtain context-enhanced text; If there is a description about the state of the entity in the module input, it will be recognized and extracted by the large model for updating the state database in the space knowledge base retrieval module; S2.2. Other modal input processing and unification Detect the input type through the input interface. If it is an external data stream input, use the input to update the state database in the space knowledge base retrieval module. If it is other non-natural language input, automatically trigger the multi-modal large model to generate a natural language description of the data; S2.3. JSON script structured generation The natural language description obtained in S2.1 and S2.2 is combined with the context-enhanced text obtained in S2.1 to perform process-based processing, such as entity recognition, verb extraction, and time standardization; Output the task requirements as structured JSON scripts, including: task id, task timestamp, task natural language description, task source, and priority, etc.

[0015] Preferably, the S3 specifically comprises: Use the large model to decompose the JSON script into multiple executable sub-tasks to generate a high-level task tree; Model the constraints involved in the current task with the help of constraint templates, including: temporal constraints, causal constraints, resource constraints, and uncertainty constraints; Temporal constraints are used to represent the sequence of actions: ; represents the action After the action is completed ; Causal constraints are used to indicate that the result of a previous action is a prerequisite for the next action: ; Resource constraints represent the limitations on a spacecraft's resource capabilities: ; indicates at any time ,resource The amount used shall not exceed the remaining amount; among which Indicates at time Resources The set of all actions that generate consumption; Uncertainty constraints represent the probability of a successful action under uncertainties such as weather, link quality, and cloud cover. ; The constraints involved in the mission come not only from mission instructions, but also from the aerospace knowledge base retrieval module and the reasoning supplements and constraint modeling rules set by the large model. The high-level task tree containing action semantics is merged with the modeled constraints into a state extension graph structure. The state extension graph structure is defined as follows: in, It is a set of action nodes, each node encapsulates the action subject, object, resource requirements, and the estimated time window derived from the low-precision recursive model; These are semantic logical dependency edges generated by the high-level task tree; These are temporal constraint edges related to the sequence; This is a set of non-temporal constraints, specifically characterized by introducing node state variables: Resource constraints: defined as the state variable "current resource value", requiring that it never fall below a preset threshold throughout the entire path; Uncertainty constraint: Defined as the state variable "path success rate", its value is the product of the success probabilities of each node in the path, and the final result is required to be no less than the reliability threshold; The state extension graph structure It supports external visualization and manual modification, and the semantic information and source of each node and constraint (obtained from large model inference / directly provided by the task / obtained by knowledge base retrieval) are also marked on the graph; The state extension graph structure It supports dynamic updates. When the upper-level module detects a task change, the system does not regenerate the entire graph, but instead adopts an incremental update strategy that combines local topological sorting with shortest path adjustment. in, Indicates a new node; Indicates the new constraint edge relationship; This indicates the shortest path update strategy; Indicates a new node Sorting a set; Indicates the new graph structure local Perform topology operations.

[0016] Preferably, S4 specifically includes: During predicate generation, the state extension graph structure As an intermediate representation for constraint enhancement, it is first necessary to perform pruning operations on it to retain only the action nodes and corresponding edges that satisfy the temporal, resource and uncertainty constraints in the graph. Expand the state graph structure after constraint pruning The predicates and related attributes (action subject, object type, resource requirements and time window, etc.) involved are extracted, and the preconditions and action effects of their implementation are obtained from the constraint edges, and their semantic descriptions are read from the semantic edges. Extracted predicates and existing predicate lists Semantic alignment, if the extracted predicate and the predicates in the list have a cosine similarity At that time, the two are considered equivalent, and the existing predicates are used directly (for predicates that are highly bound to the space platform due to heterogeneity, time window and mission object, the relevant attributes of the predicates need to be updated to adapt to mission requirements); if the cosine similarity between the extracted predicates and the predicates in the list is... Then, further interpretive comparisons are performed using a large model; if the extracted predicates and the predicates in the list have a cosine similarity... When this happens, the predicate is listed as a potential new candidate predicate, among which and A similarity threshold set by the user; Perform multi-condition filtering on candidate predicates: Predicates are high-frequency words in dialogues, appearing frequently. ,in The set frequency threshold; The predicate has high semantic value, and the relevant corpus attention mechanism scores involved in the reasoning process are as follows: in, For predicates (query vectors); For the first part involved in the reasoning process One relevant corpus (key vector); Indicates the scaling factor; This represents the normalized attention function; The set score threshold; Predicates are independent; through large-scale model reasoning, it is determined that they cannot be represented by combinations of existing predicates. If all three conditions are met, a new predicate is created using the large model based on the fixed syntax, and the new predicate definition and its semantic interpretation are updated in the predicate list; otherwise, the existing predicate combination is used. The following prompts and context generated by the aerospace knowledge base retrieval module guide the generation of mission-specific PDDL question files and domain files for the large model: A detailed description of the PDDL generation task, outlining the components of subsequent inputs and expected outputs; Generate one or two examples from other fields to illustrate the input and output formats; The JSON file describing the current task; The predicate phrases related to this task, extracted from the previous step; Preferably, S5 specifically includes: For PDDL model and state extension graph Static formal verification includes: Predicate closure detection ensures that predicates are defined and correctly referenced in all PDDL models through syntactic scanning. Parameter consistency verification: Type inference ensures that all parameter types in the PDDL model match correctly. Constraint acyclicity detection is performed by topological sorting to ensure the constraint graph is acyclic. The temporal constraints in the graph form a directed acyclic graph; Large-scale semantic verification of the PDDL model is performed by having the large model reread the PDDL domain and question file, restore them to natural language descriptions, and compare them with the original task semantics using cosine similarity to determine if there are any deviations. When a conflict or deviation is detected, the large model-driven correction strategy generator intervenes, adopts different correction logic according to the error type, and calls the large model to generate a readable error report. After confirming that the PDDL static test is successful, the PDDL model is solved using a symbolic solver, and the model solution is verified through simulation to confirm its feasibility in reality. During simulation, the orbital data from the aerospace knowledge base retrieval module and the ground telemetry and control window are used for simulation execution: a high-precision orbital recursion simulator is introduced; satellite visibility window, attitude constraints, and frequency band usage are calculated; whether the mission execution sequence meets all real-world constraints is checked, and if not, corrections are made, and a readable error report is generated by calling the large model. The validation results not only correct the current model, but also feed back to the preceding modules. Furthermore, the domain knowledge and the Prompt templates and rule templates of each module are continuously updated through gradient descent of the loss function, enabling the system to continuously accumulate task modeling experience during long-term operation and possess dynamic learning and knowledge accumulation capabilities.

[0017] A PDDL space mission modeling system based on an enhanced constraint large model, the modeling system includes: a space knowledge base retrieval module, a multimodal mission input module, a constraint enhancement module, a PDDL domain model generation module, and a verification and closed-loop correction module; The aerospace knowledge base retrieval module combines knowledge graphs and vector retrieval enhancement mechanisms to automatically complete the prompts for large models; it also provides real-time status information for mission-related entities. The multimodal task input module is used to transform multimodal input into a unified task package interface standard. All subsequent modules only need to parse the standard JSON schema to handle different input sources. The constraint enhancement module is used to express the actions and related constraints involved in the task using a graph structure, so that the generation process is both interpretable and can automatically adapt to the dynamic task environment. The PDDL domain model generation module is based on the enhanced sequence constraint intermediate representation. Through semantic alignment, predicate normalization, and controlled predicate generation mechanisms, it realizes the automatic transformation from natural language task semantics to PDDL planning models. On the basis of general domain models, it dynamically generates special domain models for each task. The closed-loop correction module combines formal verification, simulation verification, and semantic verification to automatically detect logical errors, resource conflicts, and semantic deviations in the planning model. It also uses a correction strategy generator based on a large language model to automatically repair the model. The module further constructs a cross-module closed-loop feedback system to send the verification results back to the upstream module for updating the knowledge base data and prompts for each module, thereby achieving continuous optimization and adaptive evolution of the model.

[0018] This invention provides a PDDL-based aerospace mission modeling method and system based on an enhanced constrained large model. Compared with existing technologies, it has the following advantages: In this invention, a semantic-constraint joint state extension graph is used to uniformly express temporal dependencies, resource limits, and probabilistic constraints, expanding the constraint types to more than four categories, effectively improving the model's ability to represent complex aerospace missions. Through a local incremental update mechanism, the entire model does not need to be reconstructed when missions change, allowing for local modifications while maintaining the original structure, reducing update latency, improving model stability, and enhancing the dynamic adaptability of mission modeling. A predicate extension module and feedback learning mechanism enable the system to continuously absorb new mission types, platform differences, and spatiotemporal features. As usage increases, the system's predicate and constraint libraries will be continuously optimized, exhibiting evolvability. Direct visualization of the graph structure allows for locating the source of each constraint and action node, enabling engineers to manually review or modify the generated results, making the behavior of the large model transparent. By introducing a triple verification mechanism of formal, semantic, and simulation verification, issues such as undefined predicates, temporal loops, action semantic deviations, and unexecutable planned paths can be detected in advance, allowing more errors to be discovered and automatically corrected before planning, significantly improving the model's reliability. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the modeling system in an embodiment of the present invention.

[0021] Figure 2 This is a flowchart illustrating the workflow of the input and retrieval module in an embodiment of the present invention.

[0022] Figure 3 This is a flowchart illustrating the workflow of the multimodal task input module in an embodiment of the present invention.

[0023] Figure 4 This is a flowchart illustrating the operation of the constraint enhancement module in an embodiment of the present invention.

[0024] Figure 5 This is a flowchart of the PDDL domain model generation module in an embodiment of the present invention.

[0025] Figure 6 This is a logic table for correcting various errors when modifying the PDDL model in a large model embodiment of the present invention.

[0026] Figure 7 This is a flowchart illustrating the process of the verification and closed-loop correction module in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] This application provides a PDDL aerospace mission modeling method and system based on an enhanced constraint large model, which solves the problems that the PDDL aerospace mission modeling method based on a large model is difficult to accurately describe the complex temporal and resource relationships between multiple missions, and is also difficult to quickly adjust and update when the mission environment or parameters change.

[0029] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0030] Example: like Figures 1-7 As shown, this invention provides a PDDL space mission modeling method based on an enhanced constraint large model, the modeling method comprising: S1. Aerospace Knowledge Base Retrieval A retrieval module for the aerospace knowledge base is constructed, including: a graph database, a vector database, and a state database; The graph database is used to integrate structured knowledge from fields such as aerospace telemetry and control, mission planning, orbital mechanics, and resource constraints, and organizes and stores it in the form of a knowledge graph. Nodes in the knowledge graph represent entities in the aerospace field, including but not limited to spacecraft, payloads, ground stations, orbit types, resource types, mission actions, and constraint objects. Edges represent relationships between entities, including but not limited to execution relationships, dependency relationships, constraint relationships, and resource occupancy relationships. This database provides existing engineering rules, physical constraints, and prior knowledge of the field. The vector database is used to store unstructured semantic information related to space mission planning. The semantic information includes at least text content such as historical mission dialogues, mission summary reports, command audit logs, and manually annotated descriptions. The semantic text fragments are converted into vector representations by the embedding model and stored in the vector database to support retrieval operations based on semantic similarity, thereby providing historical experience, implicit rules, and semantic supplementary information for the large model reasoning process. The state database is used to summarize and maintain real-time or near-real-time state information related to spacecraft and telemetry, tracking, and command (TT&C) entities. The state information includes at least data such as spacecraft orbital parameters, remaining resources, ground station usage, visibility windows, and global weather and communication link status. The data required for maintenance is provided directly by other departments or obtained from previous input updates. This state database is used to provide dynamic parameter support to ensure that the generated PDDL domain model can reflect the actual operating conditions under the current mission cycle. In the PDDL modeling process described in this invention, the task modeling process sequentially passes through multiple functional modules. Each module is executed by a large language model under predefined prompt constraints, performing parsing, reasoning, and generation operations. To improve the accuracy and consistency of the large model's generation in the strongly constrained scenario of aerospace missions, the system introduces an aerospace knowledge base retrieval module as an external knowledge enhancement unit during the execution of each module. The aerospace knowledge base retrieval module dynamically determines the retrieval target based on the input content of the current module and queries the aerospace knowledge base for domain knowledge related to the current modeling stage. The retrieval target includes at least the current task semantics, action description, candidate predicates, constraints, or state parameters. After the retrieval results are filtered and fused, they are organized into an enhanced context and input into the large model along with the prompts of the corresponding modules to guide its generation process. Aerospace knowledge base retrieval includes: Extract entity names from module input, search for corresponding nodes in the graph database, and obtain structured knowledge of the neighborhood. ; Embed the module input as a vector The module calculates the cosine similarity between the input vector and the vector in the vector database, and determines whether the semantic text similarity is greater than a set threshold. semantic text The similarity calculation formula is: The retrieved structured knowledge and semantic text are fused into a context-enhanced text for direct reading by the large model. The fusion formula is as follows: in, For the first Structured knowledge; For the first Semantic text; For the first The weighting coefficients / confidence levels of structured knowledge are related to the source of the knowledge; For the first The weight coefficient / confidence of a semantic text is related to the source of the knowledge; If the module input contains descriptions of entity states, these will be identified and extracted by the large model to update the state database in the aerospace knowledge base retrieval module. S2. Multimodal task input S2.1. Enhanced Natural Language Input and Retrieval Input the natural language or structured task from the module into the aerospace knowledge base retrieval module to obtain context-enhanced text; If the module input contains descriptions of entity states, these will be identified and extracted by the large model to update the state database in the aerospace knowledge base retrieval module. S2.2. Other Modal Input Processing and Unification The input type is detected through the input interface. If it is an external data stream input, the input is used to update the status database in the aerospace knowledge base retrieval module. If it is other non-natural language input, the natural language description of the multimodal large model generation data is automatically triggered. S2.3. JSON Script Structure Generation The natural language descriptions obtained from S2.1 and S2.2 are combined with the context-enhanced text obtained from S2.1 for process-oriented processing to complete process-oriented tasks such as entity recognition, verb extraction, and time standardization. Output the task requirements as a structured JSON script, including: task ID, task timestamp, task natural language description, task source and priority, etc. S3. Constraint Enhancement The large model is used to decompose the JSON script into multiple executable subtasks, generating a high-level task tree to characterize the semantic hierarchy and execution order between tasks; Constraint templates are used to model the constraints involved in the current task. These constraint templates include: temporal constraints, causal constraints, resource constraints, and uncertainty constraints. Timing constraints are used to represent the order in which actions occur: ; indicates an action In action Finish Execute after the specified time; Causal constraints are used to indicate that the result of a previous action is a prerequisite for the next action: ; Resource constraints represent the limitations on a spacecraft's resource capabilities: ; indicates at any time ,resource The amount used shall not exceed the remaining amount; among which Indicates at time Resources The set of all actions that generate consumption; Uncertainty constraints represent the probability of a successful action under uncertainties such as weather, link quality, and cloud cover. ; The constraints involved in the mission come not only from mission instructions, but also from the aerospace knowledge base retrieval module and the reasoning supplements and constraint modeling rules set by the large model. The high-level task tree containing action semantics is merged with the modeled constraints into a state extension graph structure. The state extension graph structure is defined as follows: This structure serves as an enhanced intermediate representation for subsequent model generation. in, It is a set of action nodes, each node encapsulates the action subject, object, resource requirements, and the estimated time window derived from the low-precision recursive model; These are semantic logical dependency edges generated by the high-level task tree; These are temporal constraint edges related to the sequence; This is a set of non-temporal constraints, specifically characterized by introducing node state variables: Resource constraints: defined as the state variable "current resource value", requiring that it never fall below a preset threshold throughout the entire path; Uncertainty constraint: Defined as the state variable "path success rate", its value is the product of the success probabilities of each node in the path, and the final result is required to be no less than the reliability threshold; The state extension graph structure It supports external visualization and manual modification, and the semantic information and source of each node and constraint (obtained from large model inference / directly provided by the task / obtained by knowledge base retrieval) are also marked on the graph; The state extension graph structure It supports dynamic updates. When the upper-level module detects a task change, the system does not regenerate the entire graph, but instead adopts an incremental update strategy that combines local topological sorting with shortest path adjustment. in, Indicates a new node; Indicates the new constraint edge relationship; This indicates the shortest path update strategy; Indicates a new node Sorting a set; Indicates the new graph structure local Perform topology operations.

[0031] S4.PDDL Domain Model Generation During predicate generation, the state extension graph structure As an intermediate representation for constraint enhancement, it is first necessary to perform pruning operations on it to retain only the action nodes and corresponding edges that satisfy the temporal, resource and uncertainty constraints in the graph, thereby avoiding the generation of inconsistent or unexecutable predicate structures in the unconstrained semantic space of the large model. Expand the state graph structure after constraint pruning The predicates and related attributes (action subject, object type, resource requirements and time window, etc.) involved are extracted, and the preconditions and action effects of their implementation are obtained from the constraint edges, and their semantic descriptions are read from the semantic edges. Extracted predicates and existing predicate lists Semantic alignment, if the extracted predicate and the predicates in the list have a cosine similarity At that time, the two are considered equivalent, and the existing predicates are used directly (for predicates that are highly bound to the space platform due to heterogeneity, time window and mission object, the relevant attributes of the predicates need to be updated to adapt to mission requirements); if the cosine similarity between the extracted predicates and the predicates in the list is... Then, further interpretive comparisons are performed using a large model; if the extracted predicates and the predicates in the list have a cosine similarity... When this happens, the predicate is listed as a potential new candidate predicate; Perform multi-condition filtering on candidate predicates: Predicates are high-frequency words in dialogues, appearing frequently. ,in The set frequency threshold; The predicate has high semantic value, and the relevant corpus attention mechanism scores involved in the reasoning process are as follows: in, For predicates (query vectors); For the first part involved in the reasoning process One relevant corpus (key vector); Indicates the scaling factor; This represents the normalized attention function; The set score threshold; Predicates are independent; through large-scale model reasoning, it is determined that they cannot be represented by combinations of existing predicates. If all three conditions are met, a new predicate is created using the large model based on the fixed syntax, and the new predicate definition and its semantic interpretation are updated in the predicate list; otherwise, the existing predicate combination is used. The following prompts and context generated by the aerospace knowledge base retrieval module guide the generation of mission-specific PDDL question files and domain files for the large model: A detailed description of the PDDL generation task, outlining the components of subsequent inputs and expected outputs; Generate one or two examples from other fields to illustrate the input and output formats; The JSON file describing the current task; The predicate phrases related to this task, extracted from the previous step; S5. Verification and Closed-Loop Correction For PDDL model and state extension graph Static formal verification includes: Predicate closure detection ensures that predicates are defined and correctly referenced in all PDDL models through syntactic scanning. Parameter consistency verification: Type inference ensures that all parameter types in the PDDL model match correctly. Constraint acyclicity detection is performed by topological sorting to ensure the constraint graph is acyclic. The temporal constraints in the graph form a directed acyclic graph; Large-scale semantic verification of the PDDL model is performed by having the large model reread the PDDL domain and question file, restore them to natural language descriptions, and compare them with the original task semantics using cosine similarity to determine if there are any deviations. When a conflict or deviation is detected, the large model-driven correction strategy generator intervenes, adopts different correction logic according to the error type, and calls the large model to generate a readable error report. After confirming that the PDDL static test is successful, the PDDL model is solved using a symbolic solver, and the model solution is verified through simulation to confirm its feasibility in reality. During the simulation, the orbital data in the aerospace knowledge base retrieval module and the ground telemetry and control window need to be combined for simulation execution: a high-precision orbit recursive simulator is introduced; the satellite visibility window, attitude constraints, and frequency band usage are calculated; whether the mission execution sequence meets all real-world constraints is checked, and if not, it is corrected, and a readable error report is generated by calling the large model.

[0032] The validation results not only correct the current model, but also feed back to the preceding modules. Furthermore, the domain knowledge and the Prompt templates and rule templates of each module are continuously updated through gradient descent of the loss function, enabling the system to continuously accumulate task modeling experience during long-term operation and possess dynamic learning and knowledge accumulation capabilities.

[0033] This invention provides a PDDL space mission modeling system based on an enhanced constraint large model. The modeling system includes: a space knowledge base retrieval module, a multimodal mission input module, a constraint enhancement module, a PDDL domain model generation module, and a verification and closed-loop correction module. The aerospace knowledge base retrieval module combines knowledge graphs and vector retrieval enhancement mechanisms to automatically complete the prompts for the large model, improving its domain understanding. It also provides real-time status information for mission-related entities, ensuring the accuracy of parameters in the subsequently generated PDDL domain model. This module serves as the knowledge support layer of the entire system, responsible for providing the large model with external memory of "domain facts" and "mission patterns." Its input is the prompt instructions provided to the large model, and its output is the prompt instructions supplemented with enhanced context. Specifically, this module includes three types of knowledge: a standardized knowledge graph (graph database) integrating knowledge from aerospace telemetry and control, mission planning, orbital mechanics, and resource constraints; semantic text fragments (vector database) composed of historical mission dialogues, mission summary reports, and command audit logs; and a status database summarizing spacecraft orbital parameters, ground station usage, and global meteorological data to display the real-time status of the spacecraft and related telemetry and control entities. The multimodal task input module is used to transform multimodal inputs into a unified task package interface standard. All modules only need to parse the standard JSON schema to handle different input sources, reducing the interface alignment cost of each module. This module is the input layer and data standardization center of the entire system. It is responsible for collecting raw information related to aerospace missions from multiple sources and converting it into a unified structured task package that the system can parse for use by subsequent modules. The inputs include multimodal information such as natural language, satellite remote sensing images, and external data streams represented by telemetry status, orbit prediction, meteorological and cloud data. The output is a structured JSON script. The constraint enhancement module is used to represent the actions and related constraints involved in the task using a graph structure, making the generation process both interpretable and automatically adaptable to the dynamic task environment, taking into account both the semantic generalization ability of the large model and the logical rigor of rule modeling. This module is an intermediate layer for domain model generation, using the semantic reasoning of the large model and the constraint rules in the aerospace knowledge base, and leveraging the reasoning ability of the large model to generate a unified constraint graph structure. The input is the JSON script output by the multimodal task input module, and the output is a graph structure representing temporal constraints, resource constraints, and uncertain constraints. The PDDL domain model generation module, based on enhanced sequence constraint intermediate representation, achieves automatic conversion from natural language task semantics to PDDL planning models through semantic alignment, predicate normalization, and controlled predicate generation mechanisms. Building upon a general domain model (composed of a general predicate list), it dynamically generates specialized domain models for each task, overcoming the constraint specificity challenges caused by dynamic time windows and heterogeneous platforms (such as multiple types of satellites and ground facilities) in aerospace mission planning. This not only ensures the formal consistency of the generated models but also endows the system with self-learning and knowledge evolution capabilities. This module serves as the implementation layer for domain model generation. Based on enhanced sequence constraint intermediate representation, it achieves automatic generation from natural language tasks to PDDL domain models through automatic predicate extraction, action template generation, and constraint mapping. The input is the state extension graph structure generated by the large model semantic-constraint enhancement module. The output is a dynamically updated list of predicates and a domain file and issue file for PDDL; The closed-loop correction module combines formal verification, simulation verification, and semantic verification to automatically detect logical errors, resource conflicts, and semantic deviations in the planning model. It also uses a correction strategy generator based on a large language model to automatically repair the model. Furthermore, the module constructs a cross-module closed-loop feedback system, sending verification results back to the upstream module to update the knowledge base data and prompts for each module, enabling continuous optimization and adaptive evolution of the model. This module is responsible for verifying the generated results and the model's self-evolution. After the PDDL model is generated, it automatically detects logical errors, constraint conflicts, temporal inconsistencies, and uncertainties. Through formal verification and semantic correction, the closed loop enables the model to have self-detection, self-interpretation, and self-correction capabilities. The inputs are the PDDL file generated by S4 and the constraint graph generated by S3; the outputs are a readable error report and the corrected PDDL model.

[0034] In summary, compared with the prior art, the present invention has the following beneficial effects: In this embodiment of the invention, a semantic-constraint joint state extension graph is used to uniformly express temporal dependencies, resource limits, and probabilistic constraints, expanding the constraint types to more than four categories, effectively improving the model's ability to express complex aerospace missions. Through a local incremental update mechanism, the entire model does not need to be reconstructed when the mission changes, allowing for local modifications while maintaining the original structure, reducing update latency, improving model stability, and enhancing the dynamic adaptability of mission modeling. The predicate extension module and feedback learning mechanism enable the system to continuously absorb new mission types, platform differences, and spatiotemporal features. As usage increases, the system's predicate and constraint libraries will be continuously optimized, exhibiting evolvability. Direct visualization of the graph structure allows for locating the source of each constraint and action node, enabling engineers to manually review or modify the generated results, making the behavior of the large model transparent. By introducing a triple verification mechanism of formal, semantic, and simulation verification, issues such as undefined predicates, temporal loops, action semantic deviations, and unexecutable planned paths can be detected in advance, allowing more errors to be discovered and automatically corrected before planning, significantly improving the model's reliability.

[0035] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0036] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A PDDL-based aerospace mission modeling method based on an enhanced constraint large model, characterized in that, The modeling method includes: S1. Aerospace Knowledge Base Retrieval By combining knowledge graphs and vector retrieval enhancement mechanisms, the system automatically completes the prompts for large models, improving their domain understanding; at the same time, it provides real-time status information of task-related entities to ensure the accuracy of parameters in the subsequently generated PDDL domain models. S2. Multimodal task input By transforming multimodal inputs into a unified task package interface standard, all modules only need to parse the standard JSON schema to handle different input sources; S3. Constraint Enhancement The actions and related constraints involved in the task are expressed using a graph structure, which makes the generation process both interpretable and able to automatically adapt to the dynamic task environment. S4.PDDL Domain Model Generation Based on the enhanced sequence constraint intermediate representation, the automatic transformation from natural language task semantics to PDDL planning models is achieved through semantic alignment, predicate normalization, and controlled predicate generation mechanisms; and dedicated domain models are dynamically generated for each task based on the general domain model. S5. Verification and Closed-Loop Correction By combining formal verification, simulation verification, and semantic verification, the system automatically detects logical errors, resource conflicts, and semantic deviations in the planning model. It also achieves automatic model repair through a correction strategy generator based on a large language model. Furthermore, it constructs a closed-loop feedback system across modules to send the verification results back to the upstream module for updating the knowledge base data and prompts for each module, thereby achieving continuous optimization and adaptive evolution of the model.

2. The PDDL space mission modeling method based on an enhanced constraint large model as described in claim 1, characterized in that, In S1, the aerospace knowledge base retrieval module includes: a graph database, a vector database, and a state database; The graph database is used to integrate structured knowledge in the fields of aerospace telemetry and control, mission planning, orbital mechanics, and resource constraints, and organizes and stores it in the form of a knowledge graph; where nodes in the knowledge graph represent entities in the aerospace field; and edges represent relationships between entities. The vector database is used to store unstructured semantic information related to aerospace mission planning. Semantic text fragments are converted into vector representations by an embedding model and then stored in the vector database. The status database is used to summarize and maintain real-time or near-real-time status information of spacecraft and telemetry, tracking and command (TT&C) related entities. The data required for maintenance is provided directly by other departments or obtained from previous input updates.

3. The PDDL space mission modeling method based on an enhanced constraint large model as described in claim 2, characterized in that, S1 specifically includes: extracting entity names from the module input, searching for corresponding nodes in the graph database, and obtaining structured knowledge of the neighborhood. ; Embed the module input as a vector The module calculates the cosine similarity between the input vector and the vector in the vector database, and determines whether the semantic text similarity is greater than a set threshold. semantic text The similarity calculation formula is: ； The retrieved structured knowledge and semantic text are fused into a context-enhanced text for direct reading by the large model. The fusion formula is as follows: ； in, For the first Structured knowledge; For the first Semantic text; For the first Weight coefficients / confidence levels of structured knowledge; For the first Weight coefficients / confidence levels of semantic text; If the module input contains descriptions of entity states, these will be identified and extracted by the large model to update the state database in the aerospace knowledge base retrieval module.

4. The PDDL space mission modeling method based on an enhanced constraint large model as described in claim 1, characterized in that, S2 specifically includes: S2.

1. Enhanced Natural Language Input and Retrieval Input the natural language or structured task from the module into the aerospace knowledge base retrieval module to obtain context-enhanced text; If the module input contains descriptions of entity states, these will be identified and extracted by the large model to update the state database in the aerospace knowledge base retrieval module. S2.

2. Other Modal Input Processing and Unification The input type is detected through the input interface. If it is an external data stream input, the input is used to update the status database in the aerospace knowledge base retrieval module. If it is other non-natural language input, the natural language description of the multimodal large model generation data is automatically triggered. S2.

3. JSON Script Structure Generation The natural language descriptions obtained from S2.1 and S2.2 are combined with the context-enhanced text obtained from S2.1 for process-oriented processing to complete process-oriented tasks such as entity recognition, verb extraction, and time standardization. Output the task requirements as a structured JSON script, including: task ID, task timestamp, task natural language description, task source and priority, etc.

5. The PDDL space mission modeling method based on an enhanced constraint large model as described in claim 4, characterized in that, S3 specifically includes: The JSON script is broken down into multiple executable subtasks using a large model, generating a high-level task tree; Constraint templates are used to model the constraints involved in the current task. These constraint templates include: temporal constraints, causal constraints, resource constraints, and uncertainty constraints. Timing constraints are used to represent the order in which actions occur: ; indicates an action In action Finish Execute after the specified time; Causal constraints are used to indicate that the result of a previous action is a prerequisite for the next action: ; Resource constraints represent the limitations on a spacecraft's resource capabilities: ; indicates at any time ,resource The amount used shall not exceed the remaining amount; among which Indicates at time Resources The set of all actions that generate consumption; Uncertainty constraints represent the probability of a successful action under uncertainties such as weather, link quality, and cloud cover. ; The constraints involved in the mission come not only from mission instructions, but also from the aerospace knowledge base retrieval module and the reasoning supplements and constraint modeling rules set by the large model. The high-level task tree containing action semantics is merged with the modeled constraints into a state extension graph structure. The state extension graph structure is defined as follows: ； in, It is a set of action nodes, each node encapsulates the action subject, object, resource requirements, and the estimated time window derived from the low-precision recursive model; These are semantic logical dependency edges generated by the high-level task tree; These are temporal constraint edges related to the sequence; This is a set of non-temporal constraints, specifically characterized by introducing node state variables: Resource constraints: defined as the state variable "current resource value", which requires that the value not fall below a preset threshold throughout the entire path; Uncertainty constraint: Defined as the state variable "path success rate", its value is the product of the success probabilities of each node in the path, and the final result is required to be no less than the reliability threshold; The state extension graph structure It supports external visualization and manual modification, and the semantic information and source of each node and constraint (obtained from large model inference / directly provided by the task / obtained by knowledge base retrieval) are also marked on the graph; The state extension graph structure It supports dynamic updates. When the upper-level module detects a task change, the system does not regenerate the entire graph, but instead adopts an incremental update strategy that combines local topological sorting with shortest path adjustment. ； in, Indicates a new node; Indicates the new constraint edge relationship; This indicates the shortest path update strategy; Indicates a new node Sorting a set; Indicates the new graph structure local Perform topology operations.

6. The PDDL space mission modeling method based on an enhanced constraint large model as described in claim 5, characterized in that, S4 specifically includes: During predicate generation, the state extension graph structure As an intermediate representation for constraint enhancement, it is first necessary to perform pruning operations on it to retain only the action nodes and corresponding edges that satisfy the temporal, resource and uncertainty constraints in the graph. Expand the state graph structure after constraint pruning The predicates and related attributes (action subject, object type, resource requirements and time window, etc.) involved are extracted, and the preconditions and action effects of their implementation are obtained from the constraint edges, and their semantic descriptions are read from the semantic edges. Extracted predicates and existing predicate lists Semantic alignment, if the extracted predicate and the predicates in the list have a cosine similarity At that time, the two are considered equivalent, and the existing predicates are used directly (for predicates that are highly bound to the space platform due to heterogeneity, time window and mission object, the relevant attributes of the predicates need to be updated to adapt to mission requirements); if the cosine similarity between the extracted predicates and the predicates in the list is... Then, further interpretive comparisons are performed using a large model; if the extracted predicates and the predicates in the list have a cosine similarity... When this happens, the predicate is listed as a potential new candidate predicate, among which and A similarity threshold set by the user; Perform multi-condition filtering on candidate predicates: Predicates are high-frequency words in dialogues, appearing frequently. ,in The set frequency threshold; The predicate has high semantic value, and the relevant corpus attention mechanism scores involved in the reasoning process are as follows: ； in, For predicates (query vectors); For the first part involved in the reasoning process One relevant corpus (key vector); Indicates the scaling factor; This represents the normalized attention function; The set score threshold; Predicates are independent; through large-scale model reasoning, it is determined that they cannot be represented by combinations of existing predicates. If all three conditions are met, a new predicate is created using the large model based on the fixed syntax, and the new predicate definition and its semantic interpretation are updated in the predicate list; otherwise, the existing predicate combination is used. The following prompts and contextual guidance generated by the aerospace knowledge base retrieval module guide the generation of mission-specific PDDL question files and domain files for the large model: A detailed description of the PDDL generation task, outlining the components of subsequent inputs and expected outputs; Generate one or two examples from other fields to illustrate the input and output formats; The JSON file describing the current task; The predicate phrases related to this task, obtained from the previous step.

7. The PDDL space mission modeling method based on an enhanced constraint large model as described in claim 6, characterized in that, S5 specifically includes: For PDDL model and state extension graph Static formal verification includes: Predicate closure detection ensures that predicates are defined and correctly referenced in all PDDL models through syntactic scanning. Parameter consistency verification: Type inference ensures that all parameter types in the PDDL model match correctly. Constraint acyclicity detection is performed by topological sorting to ensure the constraint graph is acyclic. The temporal constraints in the graph form a directed acyclic graph; Large-scale semantic verification of the PDDL model is performed by having the large model reread the PDDL domain and question file, restore them to natural language descriptions, and compare them with the original task semantics using cosine similarity to determine if there are any deviations. When a conflict or deviation is detected, the large model-driven correction strategy generator intervenes, adopts different correction logic according to the error type, and calls the large model to generate a readable error report. After confirming that the PDDL static test is successful, the PDDL model is solved using a symbolic solver, and the model solution is verified through simulation to confirm its feasibility in reality. During simulation, the orbital data from the aerospace knowledge base retrieval module and the ground telemetry and control window are used for simulation execution: a high-precision orbital recursion simulator is introduced; satellite visibility window, attitude constraints, and frequency band usage are calculated; whether the mission execution sequence meets all real-world constraints is checked, and if not, corrections are made, and a readable error report is generated by calling the large model. The validation results not only correct the current model, but also feed back to the preceding modules. Furthermore, the domain knowledge and the Prompt templates and rule templates of each module are continuously updated through gradient descent of the loss function, enabling the system to continuously accumulate task modeling experience during long-term operation and possess dynamic learning and knowledge accumulation capabilities.

8. A PDDL space mission modeling system based on an enhanced constraint large model, characterized in that, The modeling system includes: an aerospace knowledge base retrieval module, a multimodal mission input module, a constraint enhancement module, a PDDL domain model generation module, and a verification and closed-loop correction module; The aerospace knowledge base retrieval module combines knowledge graphs and vector retrieval enhancement mechanisms to automatically complete the prompts for large models; it also provides real-time status information for mission-related entities. The multimodal task input module is used to transform multimodal input into a unified task package interface standard. All subsequent modules only need to parse the standard JSON schema to handle different input sources. The constraint enhancement module is used to express the actions and related constraints involved in the task using a graph structure, so that the generation process is both interpretable and can automatically adapt to the dynamic task environment. The PDDL domain model generation module is based on the enhanced sequence constraint intermediate representation. Through semantic alignment, predicate normalization, and controlled predicate generation mechanisms, it realizes the automatic transformation from natural language task semantics to PDDL planning models. Based on the general domain model, it dynamically generates special domain models for each task. The closed-loop correction module combines formal verification, simulation verification, and semantic verification to automatically detect logical errors, resource conflicts, and semantic deviations in the planning model. It also uses a correction strategy generator based on a large language model to automatically repair the model. The module further constructs a cross-module closed-loop feedback system to send the verification results back to the upstream module for updating the knowledge base data and prompts for each module, thereby achieving continuous optimization and adaptive evolution of the model.

Images