Intelligent inference scenario generation and decision making method and system

Abstract

The present application relates to a kind of intelligent deduction scenario generation and decision-making method and system, according to the type of combat field and task scene type, professional field data is arranged to form standard knowledge block, and standard knowledge block is fine-tuned training of domain large model;Extract the scenario elements in user natural language instruction, and retrieve standard knowledge block in combination with initial situation information, generate scenario draft;According to the scenario draft modification opinion of user feedback, carry out scenario parameter adjustment and simulation deduction entity configuration, realize deduction;According to the real-time battlefield situation information obtained in the process of deduction and feedback to user, realize auxiliary decision-making.The present application is through the " decision-making intention-algorithm calling " intelligent mapping and encapsulation technology based on large model, " the person in loop " guiding question and answer iteration and dynamic deduction closed loop, the modularization intelligent decision-making process chain design for typical task, forms closed loop intelligent decision-making process, significantly improves the intelligent level and efficiency of deduction preparation and decision analysis.

Description

Technical Field

[0001] This invention relates to the field of military simulation technology, and in particular to an intelligent simulation scenario generation and decision-making method and system. Background Technology

[0002] In complex system simulation fields such as system-level training simulation and combat simulation, the construction of scenario is the prerequisite and foundation for conducting simulation and analysis. Traditional scenario construction relies heavily on domain experts manually writing configuration files, a cumbersome, time-consuming, and labor-intensive process that struggles to quickly respond to dynamically changing situations and diverse scenario requirements. While some template- or rule-based auxiliary generation methods exist, their generalization ability is weak, their flexibility is poor, they struggle to understand and process high-level, ambiguous natural language descriptions, and they cannot adapt to complex and ever-changing application scenarios.

[0003] In recent years, Large Language Models (LLMs) have made groundbreaking progress in natural language understanding and generation, providing new possibilities for intelligent scenario generation. However, directly applying general-purpose large models to specialized fields faces serious challenges: First, general-purpose models lack deep domain knowledge, easily generating "illusionary" content that does not conform to facts or professional norms; second, there is a huge gap between the text descriptions generated by the model and the structured data formats (such as JSON and XML configuration files) that can be executed by large-scale inference systems, requiring a lot of manual intervention for conversion; third, the lack of effective natural language interaction mechanisms makes it impossible for users to guide and fine-tune the generation process in real time, making it difficult to meet the high requirements for accuracy and reliability in practical applications.

[0004] Currently common scenario generation systems that combine large language models, such as the operational scenario random generation system and application provided by Chinese invention patent CN115358477A, ​​read historical operational scenario files from a scenario database, and automatically divide the scenarios and actions according to the user-input scenario generation parameters and task types using natural language processing algorithms. They then extract the equipment, personnel, and action models for each stage, automatically map them to the corresponding model entities in the model database, and automatically generate system-recognizable simulation scenarios according to the user-specified stage divisions using a random number algorithm. However, relying on historical operational scenario files, the generated simulation scenarios lack innovation and diversity, have limited coverage of fields and scenarios, and are difficult to adapt to complex and highly variable battlefield environments.

[0005] Therefore, existing technologies suffer from problems such as low automation and intelligence in concept construction, difficulty in integrating professional domain knowledge, uncontrollable generation results, and low efficiency in human-computer interaction. There is an urgent need for an intelligent solution that can deeply integrate domain knowledge, support natural language interaction, and achieve full-process intelligent execution from intent understanding to automatic execution. Summary of the Invention

[0006] To address the technical problems existing in the prior art, the present invention aims to provide an intelligent simulation scenario generation and decision-making method and system, which can realize the understanding of the user's high-level intentions during military simulation, shifting from "free generation" to "constrained construction", assisting in the generation of high-quality simulation scenarios and the formation of executable decision-making schemes, and significantly improving the efficiency and scientific nature of simulation preparation and decision-making.

[0007] To achieve the above-mentioned objectives, this invention provides an intelligent scenario generation and decision-making method, comprising the following steps:

[0008] Step S1: Organize the data in the professional field to construct a knowledge corpus and a knowledge retrieval module; the knowledge retrieval module is used to retrieve the knowledge corpus based on the provided search keywords and return several knowledge fragments;

[0009] Step S2: Fine-tune and train the domain-wide model using the knowledge corpus; the domain-wide model is used to understand the intent of the input natural language to generate search keywords, and to generate draft scenarios, auxiliary decision-making schemes and standard configuration files based on the obtained knowledge fragments;

[0010] Step S3: Obtain the scenario instructions from the user, and generate a scenario draft using the domain big model and knowledge retrieval module; the scenario instructions are natural language instructions.

[0011] Step S4: Based on the user-confirmed draft scenario, use the domain model and the knowledge retrieval module to perform a deduction of the physical configuration of the large system based on the user-confirmed draft scenario, and generate a standard configuration file.

[0012] Step S5: Send the configuration file to the simulation system for loading and simulation.

[0013] Step S6: Obtain real-time situational data from the simulation system, extract the real-time situational data, obtain real-time situational information, and feed it back to the user.

[0014] Step S7: Obtain decision instructions from the user; based on the real-time situation information and the decision instructions, generate an auxiliary decision-making scheme using the domain big model and the knowledge retrieval module; and perform a deduction system entity configuration on the auxiliary decision-making scheme; the decision instructions are natural language instructions.

[0015] Step S8: The auxiliary decision-making scheme, after being configured with physical entity, is sent to the simulation system for loading and simulation, and the simulation results are fed back to the user.

[0016] According to one technical solution of the present invention, the knowledge corpus includes a conceptual element knowledge corpus, a category knowledge corpus, an instruction template knowledge corpus, and a parameter value range and initialization knowledge corpus;

[0017] The hypothetical element knowledge corpus is built based on the interface of the large-scale simulation system and industry standards, including interface specification standard knowledge blocks and simulation rule knowledge blocks;

[0018] The category knowledge corpus is built on a proprietary intelligence database and includes a dedicated dictionary for entity recognition and a correlation graph of "equipment-capabilities-weaknesses".

[0019] The instruction template knowledge base is built based on the equipment instruction interface specification and includes a mapping table of "natural language intent -> instruction code -> parameter list";

[0020] The parameter value range and the initial knowledge corpus are built based on the equipment manual, including a set of equipment parameter verification rules and "environment-parameter" association rules.

[0021] According to a technical solution of the present invention, in step S2, the model fine-tuning training based on a multiple constraint mechanism specifically includes:

[0022] a. It is stipulated that when generating draft scenarios and auxiliary decision-making schemes, the domain big model must and can only reference the Top-K fragments retrieved by the knowledge retrieval module; if the fragments retrieved by the knowledge retrieval module are empty, the domain big model shall reply "unknown";

[0023] b. It is stipulated that when generating the standard configuration file, each character generated by the domain big model shall be verified in real time to determine whether the content currently generated by the domain big model conforms to the predefined configuration file structure;

[0024] c. During the training process of the domain-wide model, the output of the domain-wide model is trained with high weights for "domain terminology" and "equipment parameters" using low-rank adaptation technology.

[0025] According to one technical solution of the present invention, step S3 specifically includes:

[0026] Step S31: Obtain the intended instruction from the user, and use the domain big model to perform intent recognition on the intended instruction to obtain the intended draft search keywords;

[0027] Step S32: Using the knowledge retrieval module, search the category knowledge corpus, the instruction template knowledge base, and the parameter value range and initialization knowledge corpus according to the scenario draft search keywords, and return several knowledge fragments;

[0028] Step S33: Using the domain-wide model, generate a draft scenario based on the knowledge fragments returned by the knowledge retrieval module, and then provide it to the user.

[0029] Step S34: Obtain the draft proposal modification opinions from users, use the domain big model to perform intent recognition on the draft proposal modification opinions, and obtain the draft proposal modification search keywords;

[0030] Step S35: Directly modify the draft scenario by modifying the search keywords based on the draft scenario using the domain-wide model; or, after modifying the search keywords based on the draft scenario using the knowledge retrieval module and returning several knowledge fragments, modify the draft scenario using the domain-wide model based on the knowledge fragments returned by the knowledge retrieval module.

[0031] According to a technical solution of the present invention, in steps S4 and S7, the physical configuration of the simulation deduction specifically includes:

[0032] The domain-wide model is used to identify the intent of the draft scenario or auxiliary decision-making scheme and obtain entity-specific configuration search keywords.

[0033] Using the knowledge retrieval module, based on the entity configuration retrieval keywords, the system retrieves the desired element knowledge corpus and the parameter value range and initialization knowledge corpus, and returns several knowledge fragments.

[0034] Using the aforementioned domain-wide model, a standard configuration file is generated based on several knowledge fragments returned by the knowledge retrieval module.

[0035] According to one technical solution of the present invention, step S7 specifically includes:

[0036] Step S71: Obtain real-time situational information and decision instructions from the user, use the domain big model to perform intent recognition, and generate auxiliary decision-making scheme retrieval keywords;

[0037] Step S72: Using the knowledge retrieval module, search the category knowledge corpus, the instruction template knowledge base, and the parameter value range and initialization knowledge corpus according to the retrieval keywords of the auxiliary decision scheme, and return several knowledge fragments;

[0038] Step S73: Using the domain-wide model, generate an auxiliary decision-making scheme based on the knowledge fragments returned by the knowledge retrieval module, and provide feedback to the user;

[0039] Step S74: Obtain modification opinions on the auxiliary decision-making scheme from users, use the domain big model to perform intent recognition on the modification opinions of the auxiliary decision-making scheme, and obtain the search keywords for modification of the auxiliary decision-making scheme;

[0040] Step S75: Directly modify the auxiliary decision-making scheme by modifying the search keywords according to the auxiliary decision-making scheme using the domain big model; or, after modifying the search keywords according to the auxiliary decision-making scheme using the knowledge retrieval module and returning several knowledge fragments, modify the auxiliary decision-making scheme using the domain big model based on the knowledge fragments returned by the knowledge retrieval module.

[0041] According to one aspect of the present invention, an intelligent scenario generation and decision-making system is provided for implementing the above-described intelligent scenario generation and decision-making method, characterized in that the system comprises a data and situation layer, a training and simulation integration layer, and a decision-making and execution layer:

[0042] The data and situation layer includes:

[0043] The document organization module is used to extract document content, convert formats, and semantically slice professional domain data according to combat domains and scenarios, forming standard knowledge blocks and building a knowledge corpus.

[0044] The knowledge retrieval module is used to vectorize and embed the standard knowledge blocks and build an index, and to retrieve knowledge fragments based on the index according to the search keywords.

[0045] The situation reading module is used to acquire real-time or historical situation data, and perform protocol adaptation and format standardization to form situation information;

[0046] The integrated training and promotion layer includes:

[0047] The training framework module is used to fine-tune the training of the domain-wide model based on the data and knowledge fragments produced by the situation layer.

[0048] The inference management module is used for deploying large-scale domain models;

[0049] The decision-making and execution layer includes:

[0050] The scenario construction module is used to call the domain big model, perform intent understanding on the input scenario instructions to generate search keywords, and generate scenario drafts and standard configuration files based on the obtained knowledge fragments;

[0051] The decision support module is used to call the domain-wide model, perform intent understanding on the input decision instructions and real-time situation information to generate search keywords, and generate decision support schemes and their standard configuration files based on the acquired knowledge fragments.

[0052] The user interaction module is used to achieve human-loop interaction with users and obtain natural language commands from users.

[0053] The large system simulation module is used to connect with the simulation large system and realize data exchange with the simulation large system.

[0054] According to one technical solution of the present invention, the document organization module includes a content extraction submodule, a data slicing submodule, and a format conversion submodule.

[0055] According to one technical solution of the present invention, the training framework module includes a sample generation submodule, a feature extraction submodule, a model training submodule, a performance evaluation submodule, and a Kanban display submodule.

[0056] According to one technical solution of the present invention, the scene construction module includes:

[0057] The scenario-assisted generation submodule is used to call the domain-wide model, perform intent understanding on the input scenario instructions to generate search keywords, and generate and adjust the scenario draft based on the obtained knowledge fragments;

[0058] The entity configuration generation submodule is used to call the domain-wide model, perform simulation and deduction of the user-confirmed scenario draft, and generate a standardized configuration file.

[0059] Compared with the prior art, the present invention has the following advantages:

[0060] 1) High level of intelligence: From user natural language input to complete scenario generation, no manual coding or configuration files are required;

[0061] 2) Solving the "semantic gap" problem: By constructing knowledge blocks based on interface specifications, it achieves automatic and accurate conversion from natural language to executable code of the simulation inference system, rather than simple text filling;

[0062] 3) Solving the "illusion" problem: By introducing category constraints and parameter boundaries based on proprietary intelligence and equipment manuals, the model is forced to be generated within a closed "real dataset", ensuring the military rationality of the scenario;

[0063] 4) Human-in-the-loop collaboration: Through the command template knowledge base, rapid mapping from commander's intentions to equipment actions is achieved;

[0064] 5) High system integration: It perfectly integrates large-scale model technology with existing mature simulation and deduction systems, giving full play to their respective advantages;

[0065] 6) Significant military application value: It meets the urgent needs of modern warfare for rapid, accurate, and intelligent decision support. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0067] Figure 1 This schematic diagram illustrates the structure of an intelligent scenario generation and decision-making system according to an embodiment of the present invention.

[0068] Figure 2 This diagram illustrates the workflow of an intelligent scenario generation and decision-making system according to an embodiment of the present invention.

[0069] Figure 3 The flowchart illustrates an intelligent scenario generation and decision-making method provided in one embodiment of the present invention. Detailed Implementation

[0070] The description of the embodiments in this specification should be taken in conjunction with the accompanying drawings, which should form part of the complete specification. In the drawings, the shape or thickness of the embodiments may be exaggerated and may be indicated in a simplified or convenient manner. Furthermore, parts of the various structures in the drawings will be described separately; it is worth noting that elements not shown in the figures or not described in words are in a form known to those skilled in the art.

[0071] The descriptions of the embodiments herein, including any references to directions and orientations, are for ease of description only and should not be construed as limiting the scope of the invention. The following description of preferred embodiments involves combinations of features, which may exist independently or in combination; the invention is not particularly limited to the preferred embodiments. The scope of the invention is defined by the claims.

[0072] like Figure 1 As shown, the present invention provides an intelligent scenario generation and decision-making method, which includes the following steps:

[0073] Step S1: Organize the data in the professional field, construct a knowledge corpus, and form standard knowledge blocks; the knowledge corpus includes a conceptual element knowledge corpus, a category knowledge corpus, an instruction template knowledge corpus, and a parameter value range and initialization knowledge corpus.

[0074] In this invention, the organization of professional domain data is based on the type of combat domain and mission scenario. The types of combat domain include: space situational awareness, space control, positioning, navigation and timing, intelligence surveillance and reconnaissance, satellite communication, environmental monitoring, and missile early warning.

[0075] The types of mission scenarios include: perception scenarios, offensive and defensive scenarios, scenarios that enable joint operations, and strategic strike / defense scenarios.

[0076] The data in the relevant professional fields includes data from military regulations, historical cases, equipment databases, and expert knowledge bases, which retrieve relevant equipment performance and tactical regulations.

[0077] The standard knowledge blocks are stored in the form of question-and-answer pairs.

[0078] The scenario element knowledge corpus includes interface specification standard knowledge blocks and simulation rule knowledge blocks. The text generated by general large-scale models is often natural language description, while large-scale inference systems require strictly structured data (such as JSON / XML). Using the scenario element knowledge corpus, the underlying interface definitions (in .IDL or .proto file format) and scenario file schemas of the large-scale inference system are parsed to construct syntax-level knowledge blocks that satisfy the requirements of the large-scale inference system.

[0079] The hypothetical element knowledge corpus is built based on the interface of the inference system and industry standards, and specifically includes:

[0080] Step S111, Interface Specification Corpus: Parse the configuration file of the large-scale deduction system, extract the mandatory field constraints of entities, attributes, and behaviors in the large-scale deduction system, and form interface specification standard knowledge blocks;

[0081] The configuration files for the large-scale simulation system include the API interface files or configuration file schemas of the large-scale simulation system.

[0082] The above methods can limit the form of hypothetical elements such as equipment in the simulation system, and prevent the mandatory fields of parameters such as entities, attributes, and behaviors in the generated configuration files or the format of hypothetical elements from not meeting the requirements of the simulation system.

[0083] Step S112, Standard Specification Embedding: Parse and simulate the standard documents of the large system to generate simulation rule knowledge blocks;

[0084] The standard documentation for the large-scale simulation system includes scenario writing specifications and model framework specifications used by the system. Rule-based knowledge blocks record the mandatory elements and formatting requirements for simulation tasks as defined by the system. For example, it specifies that a "perception task" must include mandatory elements such as "task identifier," "spatiotemporal region," "troop composition," "perception equipment performance parameters," and "target object configuration parameters," along with their parameter formats. This method ensures that the text descriptions output by the model, when converted into configuration files, do not require manual correction of field names and formats and can directly pass the system's syntax validation.

[0085] The category knowledge corpus includes a dedicated entity recognition dictionary and a "equipment-capability-weakness" association graph. Built upon its proprietary intelligence database, the category knowledge corpus specifically includes:

[0086] Step S121, Entity Category Determination: Extract entity categories from the proprietary intelligence database and construct a dedicated dictionary for entity recognition;

[0087] Extract equipment models (such as a certain type of fighter jet or a certain missile system), unit numbers, and geographical names from the proprietary intelligence database to construct a dedicated dictionary for Named Entity Recognition (NER). This facilitates the forced retrieval of this category library when generating scenario drafts, ensuring that the generated entities are real entities or pre-set hypothetical targets in the current simulation context.

[0088] Step S122, Relationship Graph Construction: Based on the proprietary intelligence database and entity recognition dictionary, construct a relationship graph of "equipment-capabilities-weaknesses";

[0089] The "equipment-capability-weakness" relationship map describes the relationships between "equipment-capability-weakness". For example, "a certain type of anti-ship missile" and "a certain type of shipborne air defense system" have a specific "adversarial / being-adversarial" relationship. Through the "equipment-capability-weakness" relationship map, the realism of the scenario can be constrained, avoiding low-level "cross-over" errors.

[0090] The instruction template knowledge base includes a mapping table of "natural language intent -> instruction code -> parameter list". During the decision support phase, the user's natural language instructions need to be mapped to specific equipment control instructions. A "translator" from natural language to machine instructions is constructed using the instruction template knowledge base.

[0091] The instruction template knowledge base is built based on the equipment instruction interface specification, and specifically includes:

[0092] Step S13: Analyze and deduce the underlying instruction set of the large system, and establish a mapping table of "natural language intent -> instruction code -> parameter list".

[0093] The instruction template knowledge base can be used for knowledge retrieval during configuration file generation, facilitating the generation of binary instruction streams or structured messages that can be directly executed by the simulation engine from a large-scale domain model. For example, when a user inputs "turn on electronic interference," the large-scale domain model identifies the intent. Based on the intent recognition result, the matching instruction code is found to be CMD_JAMMING. Parameters such as frequency range and interference power are extracted from the matched instruction code. This mapping table achieves zero-loss conversion from natural language to machine-executable code, thereby improving the efficiency of human-computer interaction.

[0094] The parameter value range and initialization knowledge corpus includes a set of equipment parameter verification rules and "environment-parameter" association rules. Through this corpus, the rationality issues of scenario initialization and parameter adjustment can be resolved, and a physical boundary constraint library for standard configuration files input into the large-scale simulation system can be constructed.

[0095] The parameter value range and initialization knowledge corpus are built based on the equipment manual and specifically include:

[0096] Step S141, Parameter Boundary Constraints: Extract the limit parameters from the equipment manual and construct a parameter verification rule set.

[0097] Step S142: Initialize the logic chain: Obtain the relationship between the environment and limit parameters in the assembly manual, and establish the "environment-parameter" association rules. For example, automatically constrain the value range of "carrier-based aircraft take-off and landing parameters" based on "sea state level". Ensure that the generated scenario is physically and engineeringally predictable, and avoid the large-scale system from crashing due to parameter overflow.

[0098] The parameter value range and the initial knowledge corpus are used to ensure the rationality of the data in the generated scenario draft. When generating the scenario draft, the knowledge retrieval module will automatically call this knowledge corpus to filter the numerical rationality.

[0099] Step S2: Fine-tune the domain-wide model using the standard knowledge blocks;

[0100] To address the "illusion" problem in the reasoning process of large domain models, this invention introduces multiple constraint mechanisms during the model training and inference phases to ensure that large domain models "strictly select knowledge base data." In step S2, model fine-tuning training based on the multiple constraint mechanisms specifically includes:

[0101] a. It is stipulated that when generating draft scenarios and auxiliary decision-making schemes, the domain big model must and can only reference the Top-K fragments retrieved by the knowledge retrieval module; if the fragments retrieved by the knowledge retrieval module are empty, the domain big model shall reply "unknown";

[0102] When generating hypothetical elements (such as troop formation), the domain-wide model can be forced to reference only the Top-K fragments returned by the knowledge retrieval module. If the data is not in the knowledge base, the domain-wide model should respond with "unknown" instead of fabricating it. Through the strong coupling constraints enhanced by the above retrieval, the presence of non-real data in the output results of the domain-wide model can be avoided.

[0103] b. It is stipulated that when generating the standard configuration file, each character generated by the domain big model shall be verified in real time to determine whether the content currently generated by the domain big model conforms to the predefined configuration file structure;

[0104] During the configuration file generation phase, JSON Schema constraint decoding technology is employed. Each character generated by the model is validated in real-time to ensure it conforms to the predefined schema structure. This schema constraint decoding effectively guarantees that the generated configuration file conforms to the definition of the large-scale model configuration file.

[0105] c. During the training process of the domain-wide model, the output of the domain-wide model is trained with high weights for "domain terminology" and "equipment parameters" using low-rank adaptation technology.

[0106] In fine-tuning training, low-rank adaptation techniques are used to specifically train "domain terminology" and "specific equipment parameters" with high weights, suppressing interference from general corpora. This fundamentally eliminates the probability of the model producing "common sense errors," achieving high-fidelity cleaning of the fine-tuning data and ensuring the determinism and authority of the output.

[0107] Step S3: Obtain the scenario instructions from the user, and generate a scenario draft using the domain big model and knowledge retrieval module; the scenario instructions are natural language instructions.

[0108] Step S3 specifically includes:

[0109] Step S31: Obtain the intended instructions from the user, use the domain big model to perform intent recognition on the intended instructions, and obtain the intended draft search keywords;

[0110] The search keywords for the draft scenario should include at least one operational domain element and / or one scenario element. Operational domain elements include: domain capability name, core function, core domain elements, and operational intent / target direction. Scenario elements include: battlefield environment (time, location), organization, equipment, deployment, and basic operations.

[0111] Among them, the core elements of the domain are the key to the identification of the domain big model. The category knowledge corpus stores several core elements of the domain corresponding to the domain capability names. After the domain big model is fine-tuned and trained using the category knowledge corpus, it can determine the domain capabilities to be generated based on the identified core elements of the domain. It can also identify the core capabilities and combat intentions / target directions by calling the "equipment-capability-weakness" relationship graph through the domain capabilities, thereby generating complete scenario draft search keywords.

[0112] Step S32: Using the knowledge retrieval module, search the category knowledge corpus, instruction template knowledge corpus, and parameter value range and initialization knowledge corpus based on the scenario draft keywords to retrieve the scenario elements and return several knowledge fragments;

[0113] Step S33: Using the domain big model, generate a draft scenario based on the knowledge fragments returned by the knowledge retrieval module and provide feedback to the user;

[0114] Step S34: Obtain the draft proposal modification opinions from users, use the neighborhood large model to identify the intent of the draft proposal modification opinions, and obtain the draft proposal modification search keywords;

[0115] Step S35: Directly modify the draft scenario by modifying the search keywords based on the draft scenario using the domain big model; or, after modifying the search keywords based on the draft scenario using the knowledge retrieval module and returning several knowledge fragments, modify the draft scenario using the neighborhood big model based on the knowledge fragments returned by the knowledge retrieval module.

[0116] Step S4: Based on the user-confirmed draft scenario, use the domain model and the knowledge retrieval module to perform a deduction of the physical configuration of the large system based on the user-confirmed draft scenario, and generate a standard configuration file.

[0117] The standard knowledge block is retrieved according to the predefined generation instructions;

[0118] Step S5: Obtain initial situation information. Based on the scenario generation instructions, knowledge fragment retrieval results, and situation information, generate a scenario draft through the domain big model and provide feedback to the user.

[0119] The draft scenario includes: the force composition and deployment recommendations of both sides, the main mission phases, and an overview of the command relationship.

[0120] Step S6: Obtain modification opinions from users on the draft scenario, adjust the scenario parameters according to the modification opinions, and perform entity configuration for simulation based on the adjusted draft scenario to generate a standard configuration file.

[0121] Step S5: Send the configuration file to the simulation system for loading and simulation.

[0122] In step S5, obtaining real-time situational information specifically includes:

[0123] The real-time situational data is acquired, and key events are identified and interpreted from the real-time situational data to generate the real-time situational information. The real-time situational information includes basic spatiotemporal information, target status information, environmental status information, capability operation information, key event information, decision command information, etc.

[0124] Step S6: Obtain real-time situational data from the simulation system, extract the real-time situational data, obtain real-time situational information, and feed it back to the user.

[0125] Step S7: Obtain decision instructions from the user; based on the real-time situation information and the decision instructions, generate an auxiliary decision-making scheme using the domain big model and the knowledge retrieval module; and perform a deduction system entity configuration on the auxiliary decision-making scheme; the decision instructions are natural language instructions.

[0126] Step S7 specifically includes:

[0127] Step S71: Obtain real-time situational information and decision instructions from users, use a domain big model to perform intent recognition, and generate auxiliary decision-making scheme retrieval keywords;

[0128] Step S72: Using the knowledge retrieval module, search the category knowledge corpus, instruction template knowledge base, and parameter value range and initialization knowledge corpus based on the auxiliary decision-making scheme search keywords, and return several knowledge fragments;

[0129] Step S73: Using the domain-wide model, generate an auxiliary decision-making scheme based on the knowledge fragments returned by the knowledge retrieval module, and provide feedback to the user;

[0130] Step S74: Obtain modification opinions on the auxiliary decision-making scheme from users, use the domain big model to identify the intent of the modification opinions, and obtain the search keywords for modification of the auxiliary decision-making scheme;

[0131] Step S75: Directly modify the decision support scheme by modifying the search keywords according to the decision support scheme using the domain big model; or, after modifying the search keywords according to the decision support scheme using the knowledge retrieval module and returning several knowledge fragments, modify the decision support scheme according to the knowledge fragments returned by the knowledge retrieval module using the domain big model.

[0132] Step S8: The auxiliary decision-making scheme, after being configured with physical entity, is sent to the simulation system for loading and simulation, and the simulation results are fed back to the user.

[0133] like Figure 2 and 3 As shown, the intelligent scenario generation and decision-making system provided by the present invention includes a data and situation layer, a training and simulation integration layer, and a decision-making and execution layer.

[0134] The data and situation layer, serving as the system's data foundation, is responsible for the access, processing, standardization, and retrieval of multi-source data. The data and situation layer includes a document organization module, a knowledge retrieval module, and a situation reading module.

[0135] The document processing module is used to extract content, convert formats, and semantically segment professional domain data from documents based on operational domains and scenarios, forming standard knowledge blocks and constructing a knowledge corpus. The module extracts content, converts formats, and semantically segments input documents containing professional domain data (such as Word and PDF reports), generating standardized "knowledge blocks" suitable for large-scale domain model processing. This module comprises three sub-modules: content extraction, format conversion, and data segmentation, ensuring that the raw data is preprocessed efficiently and in a standardized manner.

[0136] The knowledge retrieval module is used to vectorize and embed the standard knowledge blocks and build an index. It generates instructions based on situational information and scenarios, and performs retrieval based on search keywords and the index, returning knowledge fragments. By providing semantic similarity-based retrieval capabilities (RAG), the knowledge retrieval module provides accurate and relevant knowledge sources for upper-layer applications.

[0137] The situation reading module is used to acquire real-time or historical situation data, and perform protocol adaptation and format standardization to form situation information. The situation reading module can read real-time or historical situation data from external simulation systems (i.e., large systems), perform protocol adaptation and format standardization, and provide the latest contextual information for decision-making.

[0138] The integrated training and inference layer serves as the system's AI capability engine, responsible for cultivating, managing, and providing domain-specific AI model services; the integrated training and inference layer includes a training framework module and an inference management module.

[0139] The training framework module is used for fine-tuning the large-scale domain model based on data and knowledge blocks generated from the situational layer. Utilizing standard knowledge blocks produced by the document processing module, the training framework module automates the fine-tuning of the large-scale domain model. It includes sub-modules for feature extraction, sample generation, model training, performance evaluation, and dashboard visualization. The feature extraction sub-module automatically generates training samples based on standard knowledge blocks. The sample generation sub-module then cleans and formats the data. The model training sub-module uses efficient fine-tuning techniques such as LoRA to fine-tune the parameters of the large-scale domain model. Finally, the performance evaluation sub-module performs multi-dimensional capability assessment, completing the training of the large-scale domain model. During the training process, the dashboard visualization sub-module provides visual monitoring, thus forming a closed loop of capability from data to model.

[0140] The inference management module is used for version management and deployment of the domain-specific large-scale models trained by the training framework module. It can save parameters of different training batches of the domain-specific large-scale models and manage the versions of the trained models. The inference management module utilizes high-performance inference frameworks such as vLLM for deployment, providing highly available and low-latency model inference services to the upper layers.

[0141] The decision-making and execution layer, serving as the top layer for system-user interaction, implements the core business logic. This layer includes a scenario construction module, a decision support module, a user interaction module, and a large-scale system simulation module.

[0142] The scenario construction module invokes the domain-wide model to understand the intent of the input scenario instructions, generate search keywords, and generate scenario drafts and standard configuration files based on the acquired knowledge fragments. The core function of the scenario construction module is intelligent scenario generation within a "human-in-the-loop" context. The module includes an auxiliary scenario generation submodule and an entity configuration generation submodule. The auxiliary scenario generation submodule combines the user's natural language requirements and retrieved knowledge to generate a scenario draft; the entity configuration generation submodule precisely parses the draft and converts it into a standardized configuration file (such as JSON) that the inference system can load.

[0143] The decision support module retrieves the standard knowledge blocks based on decision instructions and real-time situational data, and generates decision support schemes based on the retrieval results using a domain-wide model. Built on intelligent agent frameworks such as Qwen-Agent, the module understands user natural language instructions, automatically plans steps, and intelligently invokes planning algorithms encapsulated in the larger system via Model Context Protocol (MCP) to achieve decision support.

[0144] The user interaction module is used to achieve human-in-the-loop interaction with users and obtain natural language commands from users. The user interaction module can provide a graphical interface as a carrier of human-in-the-loop interaction, supporting user input, result display, and manual verification and correction of the intended configuration.

[0145] The large-scale system simulation module is used to connect with the large-scale simulation system and realize data exchange with it. The large-scale system simulation module can also act as an adapter between the system and the external simulation environment, realizing protocol conversion and data mapping between internal commands and external simulation system APIs.

[0146] In this invention, the intelligent simulation scenario generation and decision-making system adopts a layered and modular design concept, dividing the system into a three-layer architecture: data and situation layer, simulation and training integration layer, and decision and execution layer. Each layer only undertakes the core work related to its own business, which greatly reduces development and maintenance costs. At the same time, it decouples the simulation business process, facilitates independent expansion of each layer, and adapts to rapid changes in business needs.

[0147] The core workflow of the system is as follows: Figure 2 It clearly illustrates the logical relationships and data flow between the modules at each layer, and its workflow is as follows:

[0148] In the data and situation layer, the situation reading module is used to read the situation data source of the pre-training task domain and typical task scenarios, and automatically complete the data protocol adaptation.

[0149] In the integrated training and inference layer, the training framework module and the inference management module enable fine-tuning training of large language models for various domains and task scenarios, version management of final model parameters, and final model deployment, providing inference services for large language models for the decision-making and execution layers.

[0150] In the decision-making and execution layer, through the scenario construction module, decision support module, user interaction module and large system simulation module, a human-in-the-loop is formed with the user to assist the user in realizing the scenario construction, matching selection and calling of the underlying task planning model required for task simulation;

[0151] In the decision support process, the user interaction module establishes a human-in-the-loop relationship with the user. Based on the user's needs, it uses the knowledge retrieval module and the situation reading module to provide the data support required by the task planning module, and provides reference opinions such as priority ranking for the user to select planning methods and planning objects.

[0152] The present invention will be further described in detail below with reference to specific embodiments.

[0153] This implementation method takes the generation of a "reconnaissance mission scenario" as an example to explain the various stages of the intelligent scenario generation and decision-making method and system.

[0154] 1) User input stage:

[0155] Users input natural language commands through the user interaction module: "Construct a war scenario with countries A and B as the opposing sides."

[0156] After receiving the instruction, the auxiliary scenario generation submodule of the scenario construction module uses the inference management module to call the fine-tuned large language model for intent understanding and identify key elements such as: "Attacker: Country A", "Defender: Country B", "Task type (combat simulation)", "Drill type (training)", etc. The system automatically supplements the default parameters such as: "Drill duration (72 hours)", "Geographical area (a certain sea)", "Scenario complexity (medium)".

[0157] 2) Knowledge retrieval and scenario draft generation stage:

[0158] First, the knowledge retrieval module is invoked, using keywords such as "counter-attack mission," "country A's attack equipment," "country B's defense equipment," and "strike and counter-attack" to retrieve relevant knowledge fragments on equipment performance and tactical regulations from military doctrines, historical cases, and equipment databases.

[0159] Subsequently, it combines user commands and retrieved knowledge into prompt words, calls the domain big model deployed by the reasoning management module, and obtains the current battlefield situation from the big system (including combat drone parameters, missile parameters, airport / ground station distribution, countermeasure weapon parameters, etc.), reads the existing combat equipment database information of country A and country B, and generates a detailed scenario draft, which includes the force composition of the opposing sides, deployment suggestions, main mission phases and command relationship overview.

[0160] 3) Human-in-the-loop verification and configuration phase:

[0161] The draft scenario is presented to users through a user interaction module. After reviewing it, users can provide suggestions for modification, such as "increasing the number of surveillance equipment in Country B to Type 4," "adjusting the simulation area to a certain sea area," and "adding factors affecting the attack effect based on weather conditions."

[0162] When revising the draft scenario, modifications can be made based on different suggestions from users. These modifications include "simple parameter modifications" and "modifications requiring search support," as detailed below:

[0163] Simple parameter modification: Extract parameter information from the modification suggestions and modify or replace the corresponding parameters of the scenario operation. Simple parameter modification can directly execute the draft modification operation without calling the knowledge retrieval module. It is suitable for scenarios where the modification suggestions only involve adjusting the parameters already existing in the scenario, without adding unknown data or involving the supplementation of external environment / rules.

[0164] Modifications requiring external data support: Identify the modification elements in the suggested modifications, call the knowledge retrieval module to perform a search based on the identified modification elements, and then execute the modification operation. Modifications requiring retrieval support are applicable to scenarios where the suggested modifications involve adding unknown data, adjusting the core scenario (such as the simulation area), or supplementing new influencing factors (such as meteorological conditions), and specifically include the following steps:

[0165] ① By calling the large domain model through the reasoning management module, the core elements in the modification suggestions are first identified (such as "a certain sea area" corresponding to "geographical / hydrological data", and "meteorological conditions" corresponding to "meteorological parameters + association rules").

[0166] ② Call the knowledge retrieval module to retrieve and supplement relevant data (such as environmental data of a certain sea area, and rules on the impact of meteorological conditions on the attack effect);

[0167] ③ After the retrieval is completed, the entity configuration generation sub-module is linked through the reasoning management module (such as updating the troop deployment and environmental parameters related to the simulation area) and the simulation scenario draft is modified synchronously.

[0168] After user confirmation, the entity configuration generation submodule is launched. It again leverages the capabilities of the large model to parse the final draft scenario into a structured list of entities (such as entity type, coordinates, and quantity) and a list of relationships (command chain). The parsed JSON structure is then returned to the user for final confirmation.

[0169] After the user confirms that everything is correct on the verification interface, the submodule generates the final standard configuration file.

[0170] 4) Loading and simulation phase:

[0171] The large system simulation module receives the configuration file and sends it to the simulation system. The system completes the creation and loading of the scenario file, initializes the forces and environmental parameters of both sides, and establishes a real-time data connection channel to receive simulation situation data, either manually or automatically.

[0172] 5) Decision Support Stage:

[0173] During the simulation, the auxiliary decision-making module identifies key events and generates situational interpretation information based on the current confrontation situation, such as "Country A's equipment is about to enter Country B's interference zone." Users can issue commands, such as "Country A adjusts its equipment maneuver plan to avoid interference" or "Country B activates its jamming station to suppress reconnaissance." The intelligent agent in the auxiliary decision-making module automatically generates call commands for algorithms such as equipment maneuver planning and weapon allocation planning. These commands are then used by the large system simulation module to call the corresponding functions of the large system and the results are fed back to the user, forming a closed-loop decision-making process.

[0174] Through the above specific implementation methods, the practicality and advancement of the present invention in the design of intelligent decision-making processes for typical tasks have been fully verified, providing an innovative solution for command and decision-making in the context of intelligent warfare.

[0175] Finally, it should be noted that the above description represents a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A method for intelligent scenario generation and decision-making, characterized in that, Includes the following steps: Step S1: Organize the data in the professional field and build a knowledge corpus and knowledge retrieval module; The knowledge retrieval module is used to retrieve the knowledge corpus based on the provided search keywords and return several knowledge fragments; Step S2: Use the knowledge corpus to fine-tune and train the domain-wide model; The domain-wide model is used to understand the intent of the input natural language, generate search keywords, and generate draft scenarios, decision support schemes, and standard configuration files based on the obtained knowledge fragments. Step S3: Obtain the scenario instructions from the user, and generate a scenario draft using the domain big model and knowledge retrieval module; the scenario instructions are natural language instructions. Step S4: Based on the user-confirmed draft scenario, use the domain model and the knowledge retrieval module to perform a deduction of the physical configuration of the large system based on the user-confirmed draft scenario, and generate a standard configuration file. Step S5: Send the configuration file to the simulation system for loading and simulation. Step S6: Obtain real-time situational data from the simulation system, extract the real-time situational data, obtain real-time situational information, and feed it back to the user. Step S7: Obtain decision instructions from the user; based on the real-time situation information and the decision instructions, generate an auxiliary decision-making scheme using the domain big model and the knowledge retrieval module; and perform a deduction system entity configuration on the auxiliary decision-making scheme; the decision instructions are natural language instructions. Step S8: The auxiliary decision-making scheme with completed physical configuration is sent to the simulation system for loading and simulation, and the simulation results are fed back to the user.

2. The intelligent scenario generation and decision-making method according to claim 1, characterized in that, The knowledge corpus includes a conceptual element knowledge corpus, a category knowledge corpus, an instruction template knowledge corpus, and a parameter value range and initialization knowledge corpus. The scenario element knowledge corpus is constructed based on the interface of the large simulation system and industry standards. The scenario element knowledge corpus includes interface specification standard knowledge blocks and simulation rule knowledge blocks. The category knowledge corpus is built based on a proprietary intelligence database. The category knowledge corpus includes a dedicated dictionary for entity recognition and a correlation graph of "equipment-capabilities-weaknesses". The instruction template knowledge base is built based on the equipment instruction interface specification, and the instruction template knowledge base includes a mapping table of "natural language intent -> instruction code -> parameter list"; The parameter value range and initialization knowledge corpus are constructed based on the equipment manual. The parameter value range and initialization knowledge corpus include a set of equipment parameter verification rules and "environment-parameter" association rules.

3. The intelligent scenario generation and decision-making method according to claim 2, characterized in that, In step S2, the model fine-tuning training based on a multiple constraint mechanism specifically includes: a. It is stipulated that when generating draft scenarios and auxiliary decision-making schemes, the domain big model must and can only reference the Top-K fragments retrieved by the knowledge retrieval module; if the fragments retrieved by the knowledge retrieval module are empty, the domain big model shall reply "unknown"; b. It is stipulated that when generating the standard configuration file, each character generated by the domain big model shall be verified in real time to determine whether the content currently generated by the domain big model conforms to the predefined configuration file structure; c. During the training process of the domain-wide model, the output of the domain-wide model is trained with high weights for "domain terminology" and "equipment parameters" using low-rank adaptation technology.

4. The intelligent scenario generation and decision-making method according to claim 2, characterized in that, Step S3 specifically includes: Step S31: Obtain the intended instruction from the user, and use the domain big model to perform intent recognition on the intended instruction to obtain the intended draft search keywords; Step S32: Using the knowledge retrieval module, search the category knowledge corpus, the instruction template knowledge base, and the parameter value range and initialization knowledge corpus according to the scenario draft search keywords, and return several knowledge fragments; Step S33: Using the domain-wide model, generate a draft scenario based on the knowledge fragments returned by the knowledge retrieval module, and then provide it to the user. Step S34: Obtain the draft proposal modification opinions from users, use the domain big model to perform intent recognition on the draft proposal modification opinions, and obtain the draft proposal modification search keywords; Step S35: Directly modify the draft scenario by modifying the search keywords based on the draft scenario using the domain-wide model; or, after modifying the search keywords based on the draft scenario using the knowledge retrieval module and returning several knowledge fragments, modify the draft scenario using the domain-wide model based on the knowledge fragments returned by the knowledge retrieval module.

5. The intelligent scenario generation and decision-making method according to claim 2, characterized in that, In steps S4 and S7, the physical configuration specifically includes: The domain-wide model is used to identify the intent of the draft scenario or auxiliary decision-making scheme and obtain entity configuration search keywords. Using the knowledge retrieval module, based on the entity configuration retrieval keywords, the system retrieves the desired element knowledge corpus and the parameter value range and initialization knowledge corpus, and returns several knowledge fragments. Using the aforementioned domain-wide model, a standard configuration file is generated based on several knowledge fragments returned by the knowledge retrieval module.

6. The intelligent scenario generation and decision-making method according to claim 2, characterized in that, Step S7 specifically includes: Step S71: Obtain real-time situational information and decision instructions from the user, use the domain big model to perform intent recognition, and generate auxiliary decision-making scheme retrieval keywords; Step S72: Using the knowledge retrieval module, search the category knowledge corpus, the instruction template knowledge base, and the parameter value range and initialization knowledge corpus according to the retrieval keywords of the auxiliary decision scheme, and return several knowledge fragments; Step S73: Using the domain-wide model, generate an auxiliary decision-making scheme based on the knowledge fragments returned by the knowledge retrieval module, and provide feedback to the user; Step S74: Obtain modification opinions on the auxiliary decision-making scheme from users, use the domain big model to perform intent recognition on the modification opinions of the auxiliary decision-making scheme, and obtain the search keywords for modification of the auxiliary decision-making scheme; Step S75: Directly modify the auxiliary decision-making scheme by modifying the search keywords according to the auxiliary decision-making scheme using the domain big model; or, after modifying the search keywords according to the auxiliary decision-making scheme using the knowledge retrieval module and returning several knowledge fragments, modify the auxiliary decision-making scheme using the domain big model based on the knowledge fragments returned by the knowledge retrieval module.

7. An intelligent scenario generation and decision-making system, used to implement the intelligent scenario generation and decision-making method as described in any one of claims 1 to 6, characterized in that, The system comprises a data and situational awareness layer, a training and propagation integration layer, and a decision-making and execution layer. The data and situation layer includes: The document organization module is used to extract document content, convert formats, and semantically slice professional domain data according to combat domains and scenarios, forming standard knowledge blocks and building a knowledge corpus. The knowledge retrieval module is used to vectorize and embed the standard knowledge blocks and build an index, and to retrieve knowledge fragments based on the index according to the search keywords. The situation reading module is used to acquire real-time or historical situation data, and perform protocol adaptation and format standardization to form situation information; The integrated training and promotion layer includes: The training framework module is used to fine-tune the training of the domain-wide model based on the data and knowledge fragments produced by the situation layer. The inference management module is used for deploying large-scale domain models; The decision-making and execution layer includes: The scenario construction module is used to call the domain big model, perform intent understanding on the input scenario instructions to generate search keywords, and generate scenario drafts and standard configuration files based on the obtained knowledge fragments; The decision support module is used to call the domain-wide model, perform intent understanding on the input decision instructions and real-time situation information to generate search keywords, and generate decision support schemes and their standard configuration files based on the acquired knowledge fragments. The user interaction module is used to achieve human-loop interaction with users and obtain natural language commands from users. The large system simulation module is used to connect with the simulation large system and realize data exchange with the simulation large system.

8. The intelligent scenario generation and decision-making system according to claim 7, characterized in that, The document organization module includes a content extraction submodule, a data slicing submodule, and a format conversion submodule.

9. The intelligent scenario generation and decision-making system according to claim 7, characterized in that, The training framework module includes a sample generation submodule, a feature extraction submodule, a model training submodule, a performance evaluation submodule, and a Kanban display submodule.

10. The intelligent scenario generation and decision-making system according to claim 7, characterized in that, The scene construction module includes: The scenario-assisted generation submodule is used to call the domain-wide model, perform intent understanding on the input scenario instructions to generate search keywords, and generate and adjust the scenario draft based on the obtained knowledge fragments; The entity configuration generation submodule is used to call the domain-wide model, perform simulation and deduction of the user-confirmed scenario draft, and generate a standardized configuration file.

Images