Large-model-driven dynamic and static target correlation modeling, value analysis and intention deduction method and system

Abstract

The invention relates to a large-model-driven dynamic and static target correlation modeling, value analysis and intention deduction method and a large-model-driven dynamic and static target correlation modeling, value analysis and intention deduction system. The method comprises the following steps: acquiring dynamic target data and static environment data; performing association modeling on the dynamic target data and the static environment data by using the large model to form a time-space coupled knowledge graph; the knowledge graph is utilized to drive the large model to carry out double-path analysis, one path drives the large model to execute incremental graph updating on newly added data, the other path takes dynamic and static association in the knowledge graph as constraint conditions to be embedded into a reasoning link of the large model, and the large model is utilized to carry out system value evaluation and intention deduction; and carrying out fusion presentation on a system value evaluation result and an intention deduction result on a visual platform. According to the method, deep collaboration of the knowledge graph and the large model can be realized, the intelligent analysis capability on the value and behavior intention of the target system is enhanced on the basis of associated modeling, and the accuracy and systematicness of situation cognition and decision support in a complex environment are comprehensively improved.

Description

Technical Field

[0001] This invention specifically relates to the field of target analysis and intent deduction, and more specifically to a large-model-driven method and system for dynamic and static target association modeling, value analysis, and intent deduction based on spatiotemporally coupled knowledge graphs. Addressing the need for fusion and analysis of multi-source heterogeneous data in complex environments, this invention utilizes spatiotemporally coupled knowledge graph technology to achieve deep association modeling between dynamic targets (such as mobile units and unmanned vehicles) and static targets (such as facilities and landmarks). Combined with value analysis and intent deduction algorithms, it deduces the systemic value, potential behavioral patterns, and intentions of targets in real time. This method and system can operate in a domestically developed and controllable environment, providing efficient and secure technical support for situational awareness, threat assessment, and decision support. Background Technology

[0002] Currently, the fields of target analysis, intent deduction, and system value assessment mainly rely on three types of technical means. The first type is knowledge-based reasoning methods, including expert systems, rule matching, intuitionistic fuzzy reasoning, and multi-attribute decision analysis. The core of these methods lies in using predefined expert knowledge bases and logical rules to match and infer target behavioral characteristics. In system value analysis, this is typically manifested in evaluating the target's functional role and static importance within the system through predefined rules. The second type is probabilistic models based on statistical learning, typically represented by Bayesian networks, evidence theory, and hidden Markov models. These models infer target intent by establishing probabilistic dependencies between states or events and fusing multi-source uncertainty information. This type of method can further be used to quantify the influence or criticality probability of target nodes in dynamic interactive networks, providing a dynamic and uncertain perspective for system value analysis. The third type is temporal modeling techniques based on deep learning, mainly including recurrent neural networks, long short-term memory networks, and Transformer models. These techniques can automatically learn complex temporal features and potential patterns in the target's historical behavioral sequences to predict future intent or behavioral trends. Current research frontiers are exploring the use of such technologies to automatically mine functional patterns and synergistic laws of targets from group interaction data, thereby enabling emergent assessment of system capabilities.

[0003] This invention mainly relates to two known technologies, including: 1) Knowledge Graph Technology. A knowledge graph is a semantic network with an "entity-relationship-attribute" core structure. It organizes triples using graph databases (such as Neo4j) to transform scattered information into a structured knowledge network. Its construction relies on natural language processing technologies, including entity recognition, relation extraction, and knowledge completion. In applications, knowledge graphs act as "semantic navigation," analyzing logical links between entities (such as the "dependency" and "call" relationships between system modules), significantly improving the retrieval accuracy of multi-hop reasoning problems. Compared to traditional text matching, knowledge graphs offer advantages in dynamic adaptability and interpretability; however, their construction requires high-quality data cleaning, entity disambiguation (merging synonymous entities), and community clustering optimization; otherwise, they are susceptible to noise interference.

[0004] 2) Large-scale model technology. Large-scale models (such as GPT-4.1 and DeepSeek-R1) are based on the Transformer architecture and acquire general semantic understanding capabilities through pre-training on massive amounts of text data. Technological advancements in 2025 will focus on two aspects: first, lightweight architecture and efficient inference; second, domain specialization and multimodal fusion. Current challenges include: illusion suppression for complex tasks and computational limitations for edge deployments.

[0005] Currently, technological developments in the field of target analysis and intent deduction mainly revolve around three types of methods, including: 1) Traditional knowledge-based reasoning methods, such as rule matching, intuitionistic fuzzy reasoning, and multi-attribute decision-making, heavily rely on rule bases built from expert experience for logical inference. Their advantage lies in the clear and interpretable decision-making process. However, these methods have static limitations. For example, they struggle to adapt to changes in UAV behavior in dynamic airspace environments, and rule updates rely on manual intervention, leading to low timeliness. Furthermore, as the rule base expands, traversal computation significantly increases time overhead, potentially causing analysis delays in scenarios with high real-time requirements. More importantly, these methods are limited to representing static hierarchical relationships, making it difficult to quantify and assess the impact value between dynamic and static targets, such as the impact of specific ground-based equipment (e.g., navigation base stations) on the overall communication efficiency of a UAV swarm.

[0006] 2) Probabilistic models based on statistical learning, such as Bayesian networks and evidence theory, infer intent by fusing probabilities from multi-source data (e.g., flight trajectories, state signals), and can handle uncertain information relatively well. However, their bottleneck lies in their weak ability to model temporal dynamics, making it difficult to effectively capture the collaborative intent implicit in continuous action sequences. Furthermore, these models have high requirements for prior probabilities and data quality; in practical applications, missing data or noise interference can easily lead to inference bias. At the level of system value analysis, this method struggles to characterize the complex dynamic dependency network formed between static and dynamic objectives, resulting in its inadequacy in assessing the impact of critical node failures on system resilience.

[0007] 3) Deep learning time series models, such as LSTM and Transformer, can automatically learn complex patterns from historical behavioral sequence data (such as speed, heading, and formation changes), demonstrating high accuracy in cluster target intent prediction tasks. However, their applications have significant limitations: First, model training requires massive amounts of high-quality labeled data, while effective samples in real-world scenarios are often scarce; second, complex models have high computational resource requirements, making real-time deployment at the edge difficult; most importantly, the "black box" nature of the model makes its decision-making logic difficult to interpret, which not only affects the credibility of intent judgment results but also prevents the model from clearly explaining its assessment basis for the relative importance of various facilities in the system, limiting its application in critical decision-making. Summary of the Invention

[0008] This invention proposes a large-model-driven method and system for dynamic and static target association modeling, system value analysis, and intent inference based on spatiotemporally coupled knowledge graphs. Its core lies in constructing a unified spatiotemporal knowledge graph framework, deeply integrating and associating real-time behavioral data of dynamic targets with static environmental elements to form a semantic network that fully reflects the constraints and interactions of the physical world. Within this framework, the knowledge graph and the large language model form a bidirectional, reinforcing synergistic mechanism: the knowledge graph, as a structured foundation with spatiotemporal semantics, provides the large language model with real-world logical constraints and relational topologies, effectively suppressing the risk of illusions in model generation and significantly improving the credibility and interpretability of the intent inference process; simultaneously, it fully utilizes the powerful natural language understanding and semantic extraction capabilities of the large language model to automatically identify target entities, extract interaction events, and construct relational links from multi-source heterogeneous text data, achieving low-cost, high-efficiency construction and dynamic updating of the knowledge graph, greatly reducing reliance on manual annotation and expert verification.

[0009] Based on the aforementioned collaborative mechanism, this invention not only models the relationships between dynamic and static targets but also further supports in-depth deduction from behavior to intent and multi-dimensional evaluation of the target system's value. This invention can infer a target's potential behavioral intent based on its location, connectivity strength, and historical behavioral patterns in the spatiotemporal graph; and can quantitatively evaluate the functional role and strategic value of a specific target within the overall system based on graph indicators such as network centrality and node influence. All analysis results are integrated and visualized through a Geographic Information System (GIS), intuitively presenting the comprehensive situation of target trajectory, relationships, intent probability, and value weight. This invention adopts a modular architecture and a universal API design, supporting flexible access and deployment of large models of varying capabilities, meeting the needs of diverse reasoning tasks while ensuring the traceability and reproducibility of the reasoning process in specific deployment environments. By constructing a semantic understanding foundation that couples dynamic and static elements, achieving deep collaboration between knowledge graphs and large models, and strengthening the intelligent analysis capabilities of target system value and behavioral intent based on association modeling, this invention comprehensively improves the accuracy and systematic nature of situational awareness and decision support in complex environments.

[0010] The technical solution adopted in this invention is as follows: A large-model-driven method for modeling, value analysis, and intent deduction of dynamic and static target associations includes the following steps: Acquire dynamic target data and static environment data; Large models are used to model the correlation between dynamic target data and static environmental data, forming a spatiotemporally coupled knowledge graph. The knowledge graph drives the large model to perform dual-path parsing. One path drives the large model to perform incremental graph updates on new data, while the other path embeds the dynamic and static relationships in the knowledge graph as constraints into the reasoning link of the large model, and uses the large model to perform system value assessment and intent inference. The results of system value assessment and intent deduction are integrated and presented on a visualization platform.

[0011] Furthermore, the method of using a large model to model the correlation between dynamic target data and static environmental data, forming a spatiotemporally coupled knowledge graph, includes: Ontology modeling is performed, defining three key elements: categories, relationships between categories, and attributes, to establish structured constraints for the knowledge graph; Based on the category and relation constraints defined in ontology modeling, a structured extraction instruction template is generated. This template dynamically binds entity type enumeration values ​​and relation type enumeration values, requiring the large model to accurately extract entity attributes and relationships between entities from the input text and output them in standard JSON format. The JSON format data output by the large model enters the post-processing pipeline, where quality is ensured through multi-level verification, including performing basic integrity checks first and then logical rule verification. Data that passes verification is processed separately by the transformation engine: dynamic and static target data enters the structured database, while entity relationships are imported into the graph database to generate a relational network. Non-standard relationships are automatically mapped to ontology standard relationships and transformation logs are recorded. When the output of the large model is missing a key attribute, the default value filling strategy is activated. If an anomaly is detected, the manual verification interface is triggered and the writing of related data is frozen. The knowledge graph that is finally constructed is traced back to the original text through data lineage marking.

[0012] Furthermore, the three elements of definition—category, relationship between categories, and attribute—are defined as follows: For static target categories, the defined attributes include: unique identifier, text description, scope of influence, precise geographic coordinates, creation time, last update time, and running status; For dynamic target categories, the defined attributes include: unique identifier, trajectory sequence, real-time velocity, heading angle, and status; For event categories, the defined attributes include: start time and end time, location of occurrence, event description, and event classification.

[0013] Furthermore, a human-machine collaborative correction mechanism is adopted to verify the identification errors or missing data in the automated extraction results of the large model. Two verification modes are supported: one is the direct database operation mode, where users can directly edit entity attributes, delete redundant relationships, or complete missing events in the graph list view; the other is the visual annotation mode, which supports marking entity boundaries or connecting lines to build relationships on the original text.

[0014] Furthermore, the step of embedding dynamic and static relationships in the knowledge graph as constraints into the reasoning chain of the large model, and using the large model for system value assessment and intent deduction, includes: Users can select a specific geographical area on the map interface, retrieve all dynamic and static targets within that area, perform spatial queries using the graph database, and extract sub-maps. Based on the entity identifiers in the sub-graph, relevant attributes and event details are supplemented from the business database and integrated into structured JSON data; The JSON data, along with the user-specified analysis objectives and time window, is injected into two specialized analysis engines: the intent deduction engine and the value analysis engine. The intent inference engine uses prompt templates, requiring large models to infer the future behavior, motivations, and potential risks of specific dynamic targets based on the relationships and temporal behaviors of subgraphs; The value analysis engine uses prompt templates to require large models to evaluate the system value indicators of certain static facilities to dynamic objectives in a specific task context, based on the subgraph network structure, including functional support and dependence strength. The outputs of the intent deduction engine and the value analysis engine are filtered out by the rule validator to remove conclusions that violate physical laws or logical contradictions. The results of the validation are then input into the GIS visualization engine for multi-layer dynamic rendering and visualization.

[0015] Furthermore, the multi-layer dynamic rendering includes: Spatial Relationship Layer: Maps knowledge graph nodes to a map, with node size dynamically scaling based on threat value or value weight; Event heatmap: A semi-transparent heatmap of event clustering is generated using a kernel density estimation algorithm; Trajectory prediction layer: used to replay historical trajectories and mark the predicted path for the next few hours with green dashed lines; Causal chain layer: The key evidence chain supporting the deduction conclusion or value assessment is clearly displayed in the decision dashboard using a tree-like topology diagram.

[0016] Furthermore, the integration and presentation of the system value assessment and intent deduction results on the visualization platform includes: simultaneously outputting the dynamic target trajectory, intent deduction conclusions, and system value relationship diagram generated based on question and answer to the decision dashboard, forming a closed-loop decision support system of "data fusion - knowledge construction - interactive analysis and deduction - situation presentation".

[0017] A large-model-driven system for dynamic and static target association modeling, value analysis, and intent inference using the above-mentioned method includes: The ontology modeling module is used to abstractly model the conceptual layer of the knowledge graph by configuring attribute constraints and to define the system value assessment indicators. The graph management module supports manual editing and visual preview, as well as direct manipulation of graph elements to maintain an association network that conforms to ontology constraints; The knowledge generation and verification module is used to call large models to automatically process text input, extract entity relationships and generate candidate maps, and provide a visual annotation interface for manual review and correction. The value analysis and intent inference module is used to integrate GIS geographic information systems and provide joint analysis capabilities for dynamic and static targets, including map-based network display, spatiotemporal retrieval of regional bounding boxes, value calculation engine and intent inference engine, to realize the quantification of the importance of target nodes, the assessment of system resilience and the generation and visualization of behavioral intent, and to map abstract relationships and values ​​to physical space. The system management module is used to coordinate the operation of the ontology modeling module, graph management module, knowledge generation and verification module, and value analysis and intent inference module.

[0018] The beneficial effects of this invention stem from the synergistic innovation of two core technological mechanisms, systematically enhancing the depth and reliability of target system value analysis and intent deduction. Addressing the limitations of traditional rule-based methods in adaptability to dynamic environments, this invention achieves dynamic knowledge evolution and intent insight through a dual-role large-scale model collaborative mechanism: the knowledge graph construction engine utilizes the semantic parsing and relation extraction capabilities of the large-scale model to transform multi-source text data into a graph relation network rich in spatiotemporal semantics in real time, replacing the traditional manual rule writing process. Its technical advantage lies in the large-scale model's ability to adaptively learn new entity relationships and behavioral patterns, continuously optimizing the situational expression in the graph through an incremental update mechanism, thereby supporting the dynamic evaluation of the target's functional role and dependencies within the system. Simultaneously, the domain reasoning engine embeds the dynamic and static associations already constructed in the knowledge graph as structured constraints into the large-scale model's reasoning chain, effectively suppressing logical deviations in the generation process and providing a reliable cognitive foundation for system value assessment and intent deduction.

[0019] To overcome the shortcomings of probabilistic models in temporal modeling and data quality dependence, this invention constructs a spatiotemporally coupled knowledge graph foundation that achieves multi-dimensional dynamic modeling and value quantification capabilities. This foundation introduces a spatiotemporal evolution mechanism into traditional knowledge graphs by establishing cross-dimensional semantic mappings of dynamic and static targets, enabling the system to simultaneously perceive the implicit relationships between environmental constraints and behavioral sequences. Based on this, this invention can calculate value indicators such as the centrality and influence of target nodes in the network, and provide continuous and traceable evidence of behavioral evolution for intent deduction, thereby overcoming the limitations of probabilistic models in characterizing complex dependencies and system value transmission paths.

[0020] To address the data dependency and black-box decision-making problems of deep learning models, this invention constructs an interpretable inference loop through GIS visualization and intent analysis engines. Its core breakthrough lies in mapping abstract knowledge relationships to concrete physical space: when a user selects an area of ​​interest on an electronic map, a regional sub-map is generated in real time through spatial indexing and injected as context into the large model for intent inference and value analysis. This mechanism possesses dual interpretability: on the input side, spatial boundaries clearly define the inference scope and environmental elements; on the output side, the large model is required to return an inference evidence chain containing behavioral basis and associated paths. Utilizing dynamic rendering technology, the four-dimensional decision-making logic of "goal-environment-behavior-value" is ultimately presented in a visual form, significantly reducing dependence on labeled data and breaking the black box of deep learning model decision-making through spatial interpretability.

[0021] Furthermore, this invention supports dynamic adjustment of the ontology structure based on domain characteristics and continuous optimization of the Prompt template, forming a feedback-driven inference capability evolution mechanism. Under limited computing power, complex analysis tasks can be decomposed into parallel components, enabling distributed execution of sub-tasks such as entity recognition, relation extraction, value calculation, and intent inference. For specialized scenarios, the system supports calling domain-fine-tuned small models specializing in high-frequency basic operations, forming a collaborative architecture with the general-purpose large model where "the large model coordinates complex reasoning, and the small model ensures efficient processing," thus improving the overall efficiency and practicality of the system while ensuring inference depth. Attached Figure Description

[0022] Figure 1 This is a diagram illustrating the overall architecture of the method of this invention.

[0023] Figure 2 This is a flowchart of the steps of the method of the present invention.

[0024] Figure 3 This is a task prompt template.

[0025] Figure 4 This is a flowchart of the deduction process.

[0026] Figure 5 This is a template for prompting a simulation task.

[0027] Figure 6 This is a task prompt template for calculating the system value. Detailed Implementation

[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0029] The main contents of this invention include: 1. The dual-role collaboration mechanism of the large model.

[0030] In this invention, the large model is given a dual role to achieve a collaborative mechanism for domain enhancement. First, as an automated knowledge graph construction tool, the large model, with its instruction parsing and semantic understanding capabilities, automatically identifies and extracts entities, attributes, events, and their relationships from multi-source texts such as reports and news, significantly reducing reliance on manual annotation and supporting low-cost construction and continuous evolution of the graph. Second, as a domain-adaptive inference engine, the large model relies on the static target system value weights, dynamic behavioral sequences, and environmental constraints inherent in the constructed graph to perform interpretable intent deduction, outputting structured conclusions including trajectory prediction, behavioral motivation analysis, and impact scope assessment. The core innovation of this mechanism lies in constructing an integrated "construction-reasoning" dual-engine framework, which not only solves the problems of low efficiency and lagging updates in traditional manual knowledge graph construction but also effectively suppresses the knowledge gaps and logical illusions that exist in pure large models in professional domains. This invention supports flexible scheduling of different types of models through a general API layer. For example, a dedicated structure prediction model can be used to perform entity and relation extraction, while a generative large model can be selected to handle complex intent inference tasks. When the inference accuracy is insufficient in a specific domain (such as unmanned system behavior analysis), domain-labeled data can be introduced to fine-tune the inference model, improving scenario adaptability without reconstructing the system architecture. This design ensures that the inference process is interpretable and the results are traceable, while realizing closed-loop decision support from multi-source information to system value judgments and intent conclusions.

[0031] 2. The knowledge base foundation of knowledge graphs.

[0032] This invention constructs a unified framework for the association of dynamic and static targets by introducing a spatiotemporally coupled knowledge graph as the core knowledge foundation. Its innovation lies in two key aspects: First, leveraging the automated extraction capabilities (entity / relationship recognition) of large models significantly reduces the reliance on manual intervention in knowledge graph construction, supporting efficient generation and updates under manual verification and overcoming the bottleneck of cumbersome and inefficient traditional graph construction. Second, it utilizes the structured and interpretable network advantages of the knowledge graph, embedding its real-world relationships as constraints into the reasoning chain of large models, effectively suppressing model illusions. The key breakthrough lies in establishing a cross-dimensional mapping mechanism for dynamic and static targets: extracting attributes such as coordinates, type, and creation time from static targets (e.g., landmarks, facilities) to construct spatial relationship networks such as "landmark-dependent-facility" and "organization-deployment-region"; extracting attributes such as location, speed, and time series from dynamic targets (e.g., unmanned equipment, vehicles) to construct affiliation and configuration relationships such as "control center-scheduling-drone swarm" and "drone-equipped-multispectral sensor"; and specifically establishing dynamic and static interaction relationships (e.g., "drone-inspection-power facility") to form a spatiotemporally coupled knowledge graph. In traditional knowledge graph structures, event attributes are introduced as a crucial basis for spatiotemporal analysis. Each target should be associated with a network of related events, serving as important evidence of historical evolution. This graph carries the spatiotemporal attributes and event imprints of all targets, providing traceable foundational data for intent deduction. When analyzing events related to targets in a specific region, by querying sub-graphs and integrating domain hints, system value and predictive conclusions based on real-world interconnected networks are generated, enabling accurate situational deduction under environmental constraints.

[0033] This invention designs a dedicated system adapted to the aforementioned methods, achieving engineering implementation from knowledge modeling to value assessment and intent deduction through a modular architecture. The core innovation lies in the technical integration of five collaborative modules: 1) Ontology Modeling Module: Driven by business experts, it abstracts and models the conceptual layer of the knowledge graph by configuring attribute constraints (such as type definition and relation inheritance rules), and supports the definition of system value assessment indicators (such as node centrality and functional dependency strength), thus establishing a semantic foundation for subsequent quantitative analysis; 2) Graph Management Module: Supports manual editing and visual preview, allowing direct manipulation of graph elements such as events, entities, relationships, and attributes, and maintaining an association network that conforms to ontology constraints; 3) Knowledge generation and verification module: It calls the large model to automatically process text input (such as domain documents), performs entity relation extraction and generates candidate graphs, and provides a visual annotation interface for manual review and correction, so as to achieve efficient collaboration of "machine initial screening - manual fine-tuning". 4) Value Analysis and Intent Deduction Module: Deeply integrated with GIS geographic information system, providing joint analysis capabilities for dynamic and static targets: In addition to map-based network display and spatiotemporal retrieval of regional selection, the key feature is the integration of value calculation engine and intent deduction engine to realize the quantification of the importance of target nodes, system resilience assessment, and generation and visualization of behavioral intent, mapping abstract relationships and values ​​to physical space; 5) System Management Module: Maintains basic data and coordinates the operation of various components.

[0034] The innovation of this system lies in the seamless integration of knowledge modeling, automated construction, human-machine collaborative verification, system value analysis, and intent inference engine to form a closed-loop workflow, enabling the dynamic and static target correlation modeling, system value quantification, and intent inference methods to have the ability to be applied on a large scale.

[0035] like Figure 1 The diagram illustrates an embodiment of the large-model-driven dynamic and static target association modeling, system value analysis, and intent inference method of this invention, employing a layered collaborative architecture. First, dynamic target data (coordinates / velocities / time series) and static environmental data (landmark coordinates / facility attributes) are received through the data input layer, and both are simultaneously input into the spatiotemporal coupled knowledge graph module. In this module, the core task is to perform entity association modeling, generate a dynamic and static relationship network, and attach computable weight attributes (such as facility function type, communication range, etc.) to entities and relationships, thereby constructing a knowledge graph rich in semantic and weighted features as a structured foundation. This knowledge graph drives the large language model engine to perform dual-path parsing: one path drives the automatic entity relationship extraction tool to perform incremental graph updates on new data; the other path serves as the core of analysis and reasoning, automatically extracting the spatiotemporal subgraph of a specific area after the user selects it on the GIS visualization platform and submitting it to the large model. Based on this context (spatiotemporal subgraph), the large model performs two types of in-depth analysis: first, it infers the target's intent and predicts its behavioral logic within the region; second, it responds to specific user queries and performs system value analysis on specific nodes, such as quantifying the degree of support and influence weight provided by node A (e.g., a navigation base station) to node B (e.g., a drone) in completing its task within a selected region. Finally, all inference and analysis results are integrated and presented on a GIS visualization platform. Dynamic target trajectories, intent inference conclusions, and system value relationship diagrams generated based on question-and-answer sessions are simultaneously output to the decision dashboard, forming a closed-loop decision support system of "data fusion - knowledge construction - interactive analysis and inference - situation presentation".

[0036] In one embodiment of the large-model-driven dynamic and static target association modeling, value analysis, and intent inference system of this invention, a Java+Python dual-stack microservice architecture is adopted. The business service layer is implemented through Spring Cloud (Java), and the algorithm service layer is built using Flask (Python). Combined with the domestically developed basic software environment, Galaxy Kylin, an independent and controllable technology system is formed. The front-end provides interactive interfaces such as a business expert console and graph editing through the Vue.js framework. The back-end services are divided into: 1. Java Business Microservice Group: Ontology Building Service (Constraint Configuration Engine + Ontology Verification) is developed using Spring Boot and connects to the Renmin University Kingbase database via JDBC to store ontology constraints; Graph Building Service (Visual Editor + Storage Interface) is based on Neo4j Java to drive operations on the graph database and supports Cypher queries and real-time preview.

[0037] 2. Python Algorithm Microservices Group: Knowledge Generation and Verification Service (LLM Scheduling API + Automatic Extraction Engine) calls large models through standard API interfaces and combines MinIO to manage document data; Analysis Tool Service integrates GIS spatial computing and calls GeoTools to process geographic data.

[0038] 3. System Management Module: Implements unified OAuth2 authentication based on Spring Security, and coordinates resource scheduling through the Nacos service registry.

[0039] 4. Modules collaborate via standardized protocols: The front-end is proxied by Nginx, using the G6 framework for rendering visualization maps and leveraging RESTful / WebSocket for efficient communication. The data layer employs a heterogeneous storage strategy: Kingbase manages ontology metadata, Neo4j stores associated networks, and MinIO carries documents and GIS spatial files, meeting performance and security requirements under high-concurrency scenarios.

[0040] The following describes the specific implementation method of each step in the large-model-driven dynamic and static target association modeling, value analysis, and intent inference method of this invention. The step flow is as follows: Figure 2 As shown.

[0041] 1) Knowledge system modeling Before constructing a knowledge graph, this invention first performs ontology modeling as the core foundation of the knowledge system. As an abstract conceptual framework, the ontology establishes structured constraints for the knowledge graph by defining three elements: categories, relationships between categories, and attributes. This method has domain-wide applicability. In the specific implementation of the association modeling and intent inference domains involved in this invention, key ontology categories constructed include entity types such as geographical regions, mobile vehicles, control centers, organizations, and fixed sensors. These categories form networked connections through semantic relationships. For example, a "located" relationship is established between mobile vehicles and geographical regions, a "monitoring" relationship is formed between fixed sensors and mobile vehicles, and a "belonging to" jurisdictional relationship is established between mobile vehicles and control centers. Furthermore, it can be extended to typical dynamic and static target association patterns such as "dependence," "control," and "deployment" based on specific data conditions.

[0042] At the attribute definition level, a strict standardized design is adopted: For static target categories, core attributes are configured as follows: unique identifier (string type), text description (string), influence range (radius value, unit: kilometers), precise geographic coordinates (WGS84 latitude and longitude), creation time (timestamp), last update time (timestamp), and operational status (enumerated values: such as normal / damaged / maintenance, etc.). For dynamic target categories, dynamic attributes are defined as follows: unique identifier (string), trajectory sequence (a set of latitude and longitude coordinates marked with timestamps), real-time speed (floating-point number, unit: km / h), heading angle (0-360 degrees floating-point number), and status (enumerated values: such as forward / retreat / stayed, etc.). For event categories, start and end times (timestamps), occurrence location (associated target entity), event description (text), and event category (enumerated values: such as economic / cooperation / public opinion / other, etc.) are set.

[0043] After completing ontology modeling, a complete knowledge constraint system is formed, and all subsequent analysis tasks are executed in accordance with the rules of this ontology. The model possesses the characteristic of "build once, reuse long-term," supporting dynamic expansion through two maintenance mechanisms: first, a batch import / export function based on Excel templates, employing predefined field validation rules to ensure data compliance; second, a built-in visual modeling interface providing drag-and-drop category management, relationship configuration, and attribute editing components. The final storage layer adopts a heterogeneous architecture—category metadata and attribute definitions are stored in the Renmin University Kingbase relational database (table structure: ontology_class, ontology_property, relation_type, event_data), while the entity relationship network is persisted to the Neo4j graph database (node ​​labels correspond to ontology categories, edge type mapping relationship definitions). Dual-database collaboration is achieved through the JDBC / Bolt protocol, providing standardized data access interfaces for modules such as knowledge extraction and intent inference.

[0044] 2) Acquisition of basic data The basic data acquisition stage of this invention uses a simulated dataset conforming to ontology constraints as the input source to ensure data integrity and format standardization. Input data is limited to plain text format (e.g., TXT / Doc). This data source can be open-source data or legally purchased data, etc. The required data depends on the specific ontology construction, and data integrity must be guaranteed. Incomplete data may cause deviations in subsequent analysis, but will not affect the overall system and method's usability.

[0045] 3) Knowledge Graph Generation This invention achieves knowledge graph construction through a large model-driven automated pipeline. First, based on the category and relation constraints defined in the ontology modeling stage, it generates structured extraction instruction templates, such as... Figure 3 As shown, this template dynamically binds entity type enumeration values ​​(geographic region, mobile vehicle, control center, etc.) and relationship type enumeration values ​​(auxiliary / deployment / dependency, etc.). It requires the large model to accurately extract entity attributes (static target coordinates, dynamic target trajectory sequences) and relationships between entities from the input text and output them in standard JSON format. The extraction task can be broken down into multiple parts, and the content of the prompt template can be dynamically adjusted according to the actual domain.

[0046] The JSON data output from the large model enters the post-processing pipeline, where multi-level validation ensures quality: first, a basic integrity check is performed, followed by logical rule verification. Data that passes validation is then processed by the transformation engine: dynamic and static target data are entered into a structured database, while entity relationships are batch-imported into Neo4j using Cypher statements to generate a relational network. For non-standard relationships, the system automatically maps them to standard ontology relationships and records the transformation log.

[0047] A fault-tolerance mechanism is designed throughout the entire process: when key attributes are missing from the large model output, a default value filling strategy is activated (e.g., taking the median of similar values ​​when base defense values ​​are not labeled); if anomalies such as coordinate land-sea conflicts are detected, a manual verification interface is triggered and the writing of related data is frozen. The final constructed knowledge graph traces back to the original text through data lineage markers. Here, the data lineage marker is a unique traceability identifier assigned to each piece of input original text data (e.g., a non-repeatable hash value or data source ID).

[0048] 4) Manual data verification This invention provides a human-machine collaborative correction mechanism for the data verification process. For identification errors or missing data in the automated extraction results, a dedicated verification platform enables efficient correction. The system first displays the knowledge triples (entity, relation, attribute) output by the large model alongside the original text, supporting two core verification modes: one is a direct database operation mode, where users can directly edit entity attributes, delete redundant relations, or complete missing events in the graph list view; the other is a visual annotation mode, which supports marking entity boundaries or constructing relations by highlighting words on the original text. All modifications are synchronized to the underlying database in real time. Attribute changes are updated via JDBC to update the Renmin University of China (RUC) JinCang database, and relation adjustments are rewritten using the Bolt protocol to rewrite the Neo4j graph structure.

[0049] 5) Value analysis, intent deduction, and GIS visualization The specific implementation process of this invention in the GIS visualization and intelligent analysis stage is as follows: Figure 4 As shown, the core steps are as follows: When a user selects a specific geographical area on the map interface (e.g., a rectangular area defined by the East China Sea from 125°E to 130°E and 28°N to 32°N), the system first retrieves all dynamic and static targets within that area using Leaflet.js, and then performs spatial queries using the Neo4j graph database (e.g., MATCH (n)-[r]-(m) WHERE point.distance(n.coord, $selected_point)<$radius RETURN...) to extract associated sub-graphs. The system then supplements relevant attributes and event details from business databases such as KingbaseES based on the entity identifiers in the sub-graphs. Finally, the search results are integrated into structured JSON data, serving as the basis for subsequent analysis.

[0050] Implementation of value analysis and intent deduction: The JSON data, along with user-specified analysis objectives, time windows, and other parameters, is injected into two specialized analysis engines.

[0051] Intent deduction engine: uses dedicated prompt templates (structure such as...) Figure 5 As shown in the figure, the large model is required to infer the future behavior, motivation and potential risks of a specific dynamic target based on the correlation and temporal behavior of the sub-graph.

[0052] Value Analysis Engine: Uses another dedicated prompt template (structure as follows) Figure 6 As shown in the figure, the large model is required to evaluate the system value indicators such as the degree of functional support and dependence of certain static facilities (such as navigation stations and airports) on dynamic targets (such as drones and vehicles) under specific task backgrounds based on the network structure of the subgraph.

[0053] Functional support refers to the potential support a static facility can provide to a dynamic target in completing a specific task, based on its functional attributes (such as communication range, navigation accuracy, and energy supply capacity) and its location in the network topology. This indicator combines node centrality metrics (such as degree centrality and betweenness centrality) in graph networks with the inherent attributes of the facility. For example, a navigation station located at the intersection of multiple critical paths (i.e., possessing high betweenness centrality) typically has high functional support.

[0054] The strength of dependency refers to the degree to which a dynamic target depends on a certain type of static facility during the task completion process. This indicator can be quantified by the strength and uniqueness of the relationship between the dynamic target and the static facility, as well as the impact of the facility's failure on task completion. Calculated, it reflects the reverse influence of dynamic target nodes on static facility nodes. For example, if a drone swarm can only rely on a specific airport for resupply, the swarm's dependency on that airport is strong; conversely, if multiple backup airports exist, the dependency weakens. The calculation basis for this part is: the system pre-calculates various network centrality indicators (such as degree centrality, proximity centrality, betweenness centrality, eigenvector centrality, etc.) of nodes in the subgraph as basic quantitative data through algorithms built into graph databases (such as Neo4j) or custom queries.

[0055] The value analysis engine injects the aforementioned objective network metrics, entity attribute data (such as facility type and status), and user-defined task context into the large model. Based on the domain knowledge and reasoning logic in the prompt template, the large model weights, fuses, and semantically interprets this data, ultimately outputting easily understandable assessments of "functional support" and "dependency strength." Furthermore, by integrating the built-in algorithms of the graph database, the system automatically performs a series of graph computational analyses after generating the spatiotemporal subgraph, directly quantifying key topological metrics of nodes in the network, including centrality and influence. After extracting the relational subgraph of a specific region from a graph database (such as Neo4j), the system calls its built-in graph algorithm library to perform predefined network analyses on that subgraph. This typically includes calculating node degree centrality (a measure of the number of direct connections), betweenness centrality (a measure of the frequency with which a node appears as an "intermediate node" on the shortest path), proximity centrality (a measure of the ease with which a node reaches all other nodes in the network), and eigenvector centrality (a measure of the importance of a node's connection to high-influence nodes). These calculations produce objective, network-structure-based numerical results that collectively constitute a quantitative description of node "centrality" and "influence." These raw graph metrics are then encapsulated along with entity attribute data and input as core evidence into a large-scale model-driven value analysis engine. Based on this, the large model performs semantic interpretation and comprehensive weighting in conjunction with the task context, ultimately generating high-level value assessment conclusions such as domain-specific "functional support" and "dependency strength."

[0056] Rule Validation and Visualization: The outputs from both engines pass through a rule validator to filter out conclusions that violate physical laws or logical contradictions. The validated results are then input into the GIS visualization engine for multi-layer dynamic rendering, including: (1) Spatial Relationship Layer: Using the G6 graph visualization library, knowledge graph nodes are mapped to the map, and the node size can be dynamically scaled according to the threat value or value weight.

[0057] (2) Event heat map: The Turf.js kernel density estimation algorithm is used to generate a semi-transparent heat map of event clustering.

[0058] (3) Trajectory prediction layer: integrates the Cesium Timeline component, replays historical trajectories, and marks the predicted path for the next 12 hours with a green dashed line.

[0059] (4) Causal chain layer: In the decision dashboard, the key evidence chain supporting the deduction conclusion or value assessment is clearly displayed in a tree topology diagram.

[0060] Users can perform in-depth analysis through interactive dashboards, such as dragging the timeline to trigger trajectory playback, or adjusting the selection area to update the spatiotemporal statistics panel in real time (e.g., comparing the number of dynamic and static targets). All visualization elements are linked to the original knowledge graph nodes through data lineage tags, ensuring full traceability of the entire analysis and decision-making process.

[0061] It should be understood that the methods and systems described and implemented in the above embodiments of the present invention can be achieved in other ways. For example, the above module division can be implemented in other ways, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. Each step and module in the present invention can be implemented as a software functional unit and can be stored in a computer-readable storage medium, including several instructions to cause a computer device to execute some or all of the steps of the method described in the present invention. For example, one embodiment of the present invention provides a computer device (computer, server, smartphone, etc.) including a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for executing each step of the method of the present invention. For example, another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disk, etc.) storing a computer program, which, when executed by a computer, implements each step of the method of the present invention. For example, another embodiment of the present invention provides a computer program product including a computer program, which, when executed by a computer, implements the steps of the method of the present invention.

[0062] The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and to implement it accordingly. Those skilled in the art will understand that various substitutions, changes, and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the content disclosed in the embodiments of this specification; the scope of protection of the present invention is defined by the claims.

Claims

1. A large model driven dynamic and static target association modeling, value analysis and intention deduction method, characterized in that, The method comprises the following steps: acquiring dynamic target data and static environment data; using a large model to perform associated modeling on the dynamic target data and the static environment data, and forming a spatio-temporal coupled knowledge graph; using the knowledge graph to drive the large model to perform double-path analysis, one path of which drives the large model to perform incremental graph updating on new data, and the other path embeds the dynamic and static association in the knowledge graph as a constraint condition into an inference link of the large model, and uses the large model to perform system value evaluation and intention deduction; fusing and presenting the results of the system value evaluation and intention deduction on a visualization platform.

2. The method of claim 1, wherein, The use of a large model to perform associated modeling on dynamic target data and static environment data to form a spatio-temporal coupled knowledge graph comprises: performing ontology modeling to define three elements of categories, inter-category relationships and attributes, and establishing structured constraints for the knowledge graph; based on the categories and relationship constraints defined by the ontology modeling, generating a structured extraction instruction template, which dynamically binds entity type enumeration values and relationship type enumeration values, and requires the large model to accurately extract entity attributes and entity relationships from the input text and output in standard JSON format; the JSON format data output by the large model enters a post-processing pipeline, which ensures quality through multi-level checking, including first performing basic integrity checking and then performing logical rule verification, and the data that passes the verification is processed by a conversion engine: dynamic and static target data enters a structured database, and entity relationships are imported into a graph database to generate an association network, and for irregular relationships, automatically map them to ontology standard relationships and record conversion logs; when the large model output is missing key attributes, a default value filling strategy is started, and if an anomaly is found, an artificial review interface is triggered and the associated data writing is frozen, and the finally constructed knowledge graph is traced back to the original text through data bloodline marking.

3. The method of claim 2, wherein, The definition of the three elements of categories, inter-category relationships and attributes is as follows: for static target categories, the defined attributes include: unique identifier, text description, impact range, precise geographic coordinates, establishment time, last update time and running status; for dynamic target categories, the defined attributes include: unique identifier, trajectory sequence, real-time speed, heading angle and state; for event categories, the defined attributes include: start time and end time, occurrence location, event description and event classification.

4. The method of claim 2, wherein, A man-machine cooperative correction mechanism is adopted to correct the recognition errors or missing data in the automatic extraction results of the large model, supporting two correction modes: one is a database direct operation mode, in which a user directly edits entity attributes, deletes redundant relationships or supplements missing events in a graph list view; the other is a visual annotation mode, which supports word marking on the original text to mark entity boundaries or connection to build relationships.

5. The method of claim 1, wherein, The use of the knowledge graph to drive the large model to perform double-path analysis, one path of which drives the large model to perform incremental graph updating on new data, and the other path embeds the dynamic and static association in the knowledge graph as a constraint condition into an inference link of the large model, and uses the large model to perform system value evaluation and intention deduction, comprises: a user frames a specific geographic area on a map interface, retrieves all dynamic and static targets in the area, performs spatial query using a graph database, and extracts a sub-graph. According to the entity identification in the sub-graph, the relevant attributes and event details are supplemented from the business database, and are integrated into structured JSON data; The JSON data and the user-specified analysis target and time window are respectively injected into two special analysis engines: an intention deduction engine and a value analysis engine; The intention deduction engine uses a prompt template to require the large model to deduce the future behavior, motivation and potential risks of a specific dynamic target based on the association and time sequence behavior of the sub-graph; The value analysis engine uses a prompt template to require the large model to evaluate the system value indicators of certain static facilities to the dynamic target in a specific task context, including function support and dependency strength, based on the network structure of the sub-graph; The output results of the intention deduction engine and the value analysis engine are filtered by a rule checker to filter out conclusions that violate physical laws or logical contradictions, and the results that pass the verification are input into a GIS visualization engine for multi-layer dynamic rendering and visual presentation.

6. The method of claim 5, wherein, The multi-layer dynamic rendering includes: Spatial relationship layer: mapping knowledge graph nodes to a map, with node size dynamically scaled according to threat value or value weight; Event heat layer: using kernel density estimation algorithm to generate a semi-transparent heat map of event clustering degree; Trajectory prediction layer: used to replay historical trajectory and mark the predicted path for several hours in the future with green dashed line; Causal chain layer: clearly showing the key evidence chain supporting the deduction conclusion or value assessment in a tree topology graph in the decision board.

7. The method of claim 1, wherein, The fusion and presentation of the system value assessment and intention deduction results on the visualization platform include: synchronously outputting the dynamic target trajectory, intention deduction conclusion and system value relationship graph generated based on question and answer to the decision board, forming a closed-loop decision support system of "data fusion-knowledge construction-interactive analysis and deduction-situation presentation".

8. A large model driven dynamic and static target correlation modeling, value analysis and intention deduction system using the method of any one of claims 1-7. Including: Ontology modeling module, used for abstract modeling of the concept layer of the knowledge graph by configuring attribute constraints, and defining system value assessment indicators; Graph management module, supporting manual editing and visual preview, as well as direct operation of graph elements, to maintain an associated network that meets the ontology constraints; Knowledge generation and checking module, used for calling a large model to automatically process text input, performing entity relationship extraction and generating candidate graphs, while providing a visual annotation interface for manual review and correction; Value analysis and intention deduction module, used for integrating a GIS geographic information system to provide joint analysis capability of dynamic and static targets, including map-based associated network display, regional frame selection spatio-temporal retrieval, value calculation engine and intention deduction engine, to realize target node importance quantification, system resilience assessment and behavior intention generation and visualization, and to map abstract relationships and values to physical space; System management module, used for unified coordination of the ontology modeling module, graph management module, knowledge generation and checking module, and value analysis and intention deduction module.

9. A computer device, comprising: Including a memory and a processor, the memory stores a computer program configured to be executed by the processor, the computer program includes instructions for executing the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by a computer to implement the method in any one of claims 1-7.

Images