Target recognition system and method based on large language model and knowledge graph

Abstract

The invention discloses a target recognition system and method based on a large language model and a knowledge graph, and the system comprises a data collection module, a data preprocessing and feature extraction module, a data fusion module, a knowledge base module, an LLM reasoning module, and a result output and presentation module, and relates to the technical field of target recognition and tracking. According to the target recognition system and method based on the large language model and the knowledge graph, an intelligent system capable of effectively integrating multi-source heterogeneous intelligence data and automatically carrying out high-precision identity and intention recognition can be constructed, so that the efficiency and depth of intelligence analysis are improved, and decision making is better supported.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target identification and tracking, in particular to a target identification system and method based on a large language model and a knowledge graph. BACKGROUND

[0002] Existing technologies in the field of reconnaissance target identification usually rely on the combination of multiple professional systems and manual analysis. For example, radar systems are responsible for target detection and tracking, providing physical parameters (position, speed, height, etc.) of the target; signal intelligence (SIGINT) systems are responsible for intercepting and analyzing electronic signals (such as radar emission signals, communication signals) of the target, extracting electronic features; electro-optical / infrared (EO / IR) systems provide image or video data of the target; open source intelligence (OSINT) systems collect information from public channels.

[0003] In terms of target identification (especially identity recognition), existing systems mainly use rule-based matching or limited feature comparison methods. For example, according to the specific pulse characteristics of radar signals, the radar model is identified, and then the aircraft model equipped with the radar is inferred; or the radar cross section (RCS) of the target is compared with the known target library. In terms of target intent recognition, more reliance is placed on experienced analysts to make judgments on manually integrated data and target behavior patterns. Some advanced systems use machine learning-based classifiers to identify specific behavior patterns or match based on a pre-set scenario library. Data fusion technology is also used to associate data from different sensors to the same target, but usually only stays at the data level of association, rather than deep semantic understanding and logical reasoning.

[0004] Although there are some attempts to automate part of the identification process based on machine learning, such as using deep learning for image recognition or signal classification, these methods often focus on pattern recognition of single modal data, or require a large number of labeled samples, and it is difficult to effectively integrate and utilize the complex associations between multi-source heterogeneous data and the prior knowledge of human experts. The existing technology still highly depends on manual complex logical reasoning and knowledge integration to promote low-level sensor features to high-level semantic understanding (identity, intent) level.

[0005] The existing technology has the following main problems: a. Difficulty in data heterogeneity and fusion: the format and modality of raw data from different sensors are vastly different, and it is still challenging to effectively and timely associate radar tracks, electronic signals, communication content, images, public information, etc. to the same target and perform unified representation, especially to overcome the inconsistencies in time, space and semantics. Traditional methods often require complex customized interfaces and algorithms when fusing different types of data, and it is difficult to adapt to new data sources or data formats.

[0006] b. Reliance on manual analysis and lack of real-time performance: Existing advanced intelligence analysis, especially intent recognition, relies heavily on experienced human analysts who need to spend a lot of time on the fused fragmented information for judgment and logical reasoning. This leads to low efficiency of the analysis process, making it difficult to meet the demand for near real-time intelligence in modern high-dynamic environments. In rapidly changing battlefield situations or emergencies, traditional analysis processes may not be able to provide timely key warnings.

[0007] c. Knowledge management and application are scattered and unstructured: The business knowledge of experts (e.g., common tactics of a particular unit, typical flight profiles of a certain type of aircraft in different missions, the correlation between electronic warfare features and operational intent) often exists in unstructured form in documents, databases, and even analysts' experience, making it difficult to systematize, structure, and effectively utilize in automated systems. Although rule-based expert systems attempt to encode knowledge, the rule base construction and maintenance cost is high, and it is difficult to cover all complex and variable scenarios, with poor robustness.

[0008] d. Limited recognition accuracy and robustness: Recognition methods based on a single feature or simple rules are vulnerable to target camouflage, electronic interference, data noise, or unknown behavior patterns, leading to false positives, false negatives, or failure to recognize. Existing methods have limited recognition accuracy and robustness in handling incomplete, uncertain, or deceptive behavior information.

[0009] e. Difficulty in deep intent inference: Existing technologies lack the ability to infer the deep intent of targets (e.g., distinguishing between training, demonstration, reconnaissance, attack preparation, etc.), especially when complex logical reasoning is required to combine target continuous behavior sequences, communication content, external environment, historical patterns, and other multi-dimensional information. Traditional methods are difficult to handle, while this is crucial for threat level judgment and countermeasure strategy formulation. In summary, existing technologies have significant technical defects in handling massive multi-source heterogeneous intelligence data, automating high-precision identity and intent recognition, effectively managing and utilizing expert knowledge, and meeting real-time requirements.

[0010] In summary, existing technologies have significant technical defects in handling massive multi-source heterogeneous intelligence data, automating high-precision identity and intent recognition, effectively managing and utilizing expert knowledge, and meeting real-time requirements. SUMMARY

[0011] To overcome the shortcomings of existing technologies, the present application provides a target recognition system and method based on large language models and knowledge graphs, which solves the problem of significant technical defects in handling massive multi-source heterogeneous intelligence data, automating high-precision identity and intent recognition, effectively managing and utilizing expert knowledge, and meeting real-time requirements.

[0012] To achieve the above object, the present application is realized by the following technical solutions: a target recognition system based on a large language model and a knowledge graph, comprising a data acquisition module, a data preprocessing and feature extraction module, a data fusion module, a knowledge base module, an LLM inference module, and a result output and presentation module. Data acquisition module: responsible for acquiring raw data streams in real time or quasi-real time from various reconnaissance sensors and open data sources. These data include but are not limited to radar echo or track data, electronic signal data, intercepted communication data, image / video data, public report information, etc.

[0013] Data preprocessing and feature extraction module: clean, denoise, and standardize the collected raw data.

[0014] Use domain-specific algorithms and pre-set business knowledge rules to extract structured features from different types of data. For example: Extract from radar data: target ID, accurate spatio-temporal coordinates, speed, height, heading, acceleration, maneuver type, radar cross-section (RCS) feature, track type.

[0015] Extract from SIGINT data: signal source ID, signal type, working mode, pulse feature, frequency feature, modulation feature, communication protocol, recognized call sign, keyword, phrase.

[0016] Extract from OSINT data: information such as geographic location, time, event, organization, equipment model associated with the target.

[0017] Extract from image / video data: recognized aircraft model, paint, serial number, and mounted type.

[0018] Output the extracted features in a unified format and associate them with the target ID.

[0019] Data fusion module: receives feature data about the same target at different times or generated by different sensors from the data preprocessing and feature extraction module.

[0020] Use multi-sensor tracking and association algorithms to associate these scattered feature data to a unique logical target ID.

[0021] Integrate all the associated features into a comprehensive feature package about the current state and behavior of the target.

[0022] Knowledge base module: build and maintain a knowledge graph containing domain expert business knowledge.

[0023] KG's nodes include, but are not limited to: aircraft models, radar models, unit of troops, military bases, mission types, flight profile types, electronic signal signatures, geographical areas, specific call signs, known tactical patterns, etc.

[0024] KG's edges represent various associations between entities, such as: equipped with (Unit_X, Aircraft_Type_A), emitting signals (Radar_Type_B, Signal_Pattern_C), performing missions (Aircraft_Type_A, Mission_Type_D), , commonly used call signs (Unit_X, Callsign_W), , cooperating with (Mission_Type_D, Aircraft_Type_F), etc.

[0025] Expert operational knowledge is encoded into the KG by defining entities, relationships, and attributes, and even ontology rules. The KG supports query and reasoning operations for retrieving background knowledge related to input features.

[0026] LLM reasoning module: This is the core reasoning engine of the invention. It receives the target integrated feature package from the data fusion module.

[0027] Knowledge enhancement mechanism: According to the input target features, this module automatically queries the knowledge graph to retrieve the most relevant entities, relationships, and facts related to these features. For example, query "signal feature B" is associated with which radar and aircraft models; query "call sign Cobra" is associated with which troops and aircraft models; query "high altitude circling" is a typical profile of which tasks.

[0028] The integrated target feature data after fusion and the relevant background knowledge retrieved from the KG are organized into a structured prompt and input into the large language model. The prompt design includes clear task instructions, input data format specifications, reasoning step guidance, and output format requirements.

[0029] After receiving the prompt, the large language model uses its powerful natural language understanding ability, pattern recognition ability, and logical reasoning ability to analyze the input features and background knowledge. The LLM can find associations, recognize patterns, and perform multi-step logical reasoning in these information to infer the identity and intent of the target.

[0030] Identification logic: The LLM compares the target's features with the aircraft model, radar model, unit, call sign, and other information in the KG. For example, if the features match the F-16's performance envelope, the radar signature is AN / APG-68, and the call sign belongs to a unit known to be equipped with F-16s, the LLM logically infers that the target is an F-16 and belongs to that unit.

[0031] Intent Recognition Logic: LLM analyzes the target's behavioral sequences, electronic activity patterns, and communication content, and combines the KG associations between mission types and typical behavioral patterns with information about the current external environment to infer the target's mission intent. For example, if a target suddenly climbs and activates its target tracking radar after flying at low altitude and high speed in a sensitive area, and the communication content mentions "lock," LLM logically infers that its intention may be a ground attack. LLM can also identify logical conflicts between data. For example, if the radar track indicates eastward flight, but the communication content indicates westward flight, LLM can infer that there is a deceptive intent.

[0032] LLM generates structured analysis results, including the identified target identity, inferred target intent, key evidence supporting these conclusions, and any anomalies or inconsistencies detected.

[0033] Result output and presentation module: receives the structured analysis results output by the LLM reasoning module.

[0034] The results are formatted and pushed to intelligence analyst workstations, command and control systems, or other downstream information systems.

[0035] The target situation is visually presented on the user interface, including target location, track, identified identity information, current intention, and confidence level, and can be traced back to the specific feature data and knowledge graph information supporting these conclusions, enhancing the interpretability of the results.

[0036] Trigger alerts based on confidence or intent type.

[0037] The present invention also discloses a method for a target recognition system based on a large language model and a knowledge graph, and the specific deployment and use process is as follows: After the system is started, each module enters operation. The data acquisition module continuously receives external data. The data preprocessing and feature extraction module and the data fusion module process this data in real time, generating a comprehensive feature package for the target and triggering an LLM inference request. Based on the request, the LLM inference module retrieves relevant knowledge from the knowledge base module (KG) and feeds it, along with the feature package, into the LLM for inference. The inference results are returned to the result output and presentation module for presentation and distribution. Analysts can monitor the target's situation and review analysis results through the user interface. If necessary, they can manually intervene or provide feedback to the knowledge base to optimize the system.

[0038] Alternative components or other similar technical solutions: Knowledge base representation: In addition to the knowledge graph, some simple rules or associations can also be directly encoded in the prompt template of the LLM. However, when dealing with a large amount of complex knowledge that requires multi-hop reasoning, the knowledge graph has a significant advantage. This invention prefers to use a knowledge graph, but is compatible with some knowledge injected through PromptTemplates.

[0039] LLM model: Different architectures or sizes of large language models can be used, and selection is made according to reasoning ability, speed, resource requirements and security. Additional fine-tuning may be required for military, intelligence and other fields.

[0040] Feature extraction algorithm: Different signal processing, image recognition, and text analysis algorithms can be used according to the specific sensor type and data characteristics.

[0041] Data fusion algorithm: Various multi-sensor fusion algorithms can be used, such as Bayesian networks, DS evidence theory, and extensions of Kalman filters.

[0042] The invention provides a target recognition system and method based on large language models and knowledge graphs. Compared with the prior art, it has the following beneficial effects: 1. The target recognition system and method based on large language models and knowledge graphs significantly improve the accuracy and robustness of recognition: By integrating heterogeneous data from radar, SIGINT, OSINT and other sources, combining the rich expert business knowledge in KG and the powerful cross-modal / cross-domain logical reasoning ability of LLM, the target information can be cross-verified from multiple angles, effectively reducing the misjudgment caused by incomplete, uncertain or noisy single-source data, significantly improving the accuracy of identity and intent recognition, especially when facing unknown or deceptive behavior. For example, traditional methods may only identify the model based on radar signal characteristics, while this invention can combine call signs, communication content, known unit information, and even image features for multi-dimensional confirmation, greatly improving the confidence of recognition.

[0043] 2. The target recognition system and method based on large language models and knowledge graphs realizes automation and near real-time analysis: The complex feature integration, knowledge association and logical reasoning process are automated, greatly shortening the intelligence analysis cycle, enabling the system to respond to dynamic target situation in near real time, meeting the rapid decision-making needs of modern warfare or emergencies. Traditionally, analysts take hours or even days to complete tasks, while the system can provide preliminary conclusions within minutes.

[0044] 3、The target recognition system and method based on large language model and knowledge graph effectively manages and utilizes expert knowledge: structured knowledge graph form is adopted to manage expert business knowledge, so that the knowledge is clearer, easier to maintain and expand, and can be efficiently utilized by the reasoning system. The complexity explosion and maintenance difficulty problems faced by traditional rule system are avoided. The RAG mechanism ensures that the LLM can obtain the most relevant background knowledge during reasoning, improving the relevance and accuracy of reasoning.

[0045] 4、The target recognition system and method based on large language model and knowledge graph deepens the intention recognition level: LLM can perform deeper and more detailed intention inference by virtue of its understanding ability of text (communication content, OSINT) and structured information (features, KG), distinguish different potential motives under similar behavior patterns (such as distinguishing routine training from attack rehearsal), and even recognize and analyze complex deception behaviors, providing deep intelligence insights that traditional methods cannot obtain.

[0046] 5、The target recognition system and method based on large language model and knowledge graph improves system flexibility and adaptability: when new aircraft, radar, tactics or new intelligence sources appear, the system can adapt by mainly updating the knowledge graph or incrementally training / adjusting the prompt of the LLM, without the need to redesign and implement complex hard-coded logic or rule sets.

[0047] 6、The target recognition system and method based on large language model and knowledge graph reduces the workload of manual work: automates tedious data integration and preliminary analysis work, liberates analysts from the ocean of massive information, and enables them to focus on higher-level verification, strategic research and decision support of system output results. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The figure is a schematic diagram of the system of the present application.

[0049] In the figure: 1, data acquisition module; 2, data preprocessing and feature extraction module; 3, data fusion module; 4, knowledge base module; 5, LLM reasoning module; 6, result output and presentation module. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application are described clearly and completely. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0051] Please refer to Figure 1The embodiment of the application provides a technical scheme: a target recognition system based on a large language model and a knowledge graph, comprising a data acquisition module 1, a data preprocessing and feature extraction module 2, a data fusion module 3, a knowledge base module 4, an LLM inference module 5 and a result output and presentation module 6.

[0052] The specific implementation of each module of the system is as follows: Data acquisition module 1: A dedicated data interface (for example, a network socket, an API, a file system monitor) is used to establish a connection with external data sources such as radar stations, SIGINT listening stations, OSINT crawlers, image recognition systems and the like.

[0053] The functions of data reception, preliminary timestamp synchronization and format conversion are realized, and the original data is distributed to downstream modules. For example, real-time radar track streams, intercepted electronic signal packets, and structured information crawled from the network are sent to a message queue or a temporary database.

[0054] Data preprocessing and feature extraction module 2: Special processing units are developed or integrated for different types of data.

[0055] Radar processing unit: realize track filtering and smoothing, calculate based on speed, height, acceleration, turn rate calculator, and make preliminary target type tendency judgment based on RCS feature comparison (need to preset typical RCS feature library). Output a JSON object containing target ID, time, position, speed vector, height, maneuver type code, and estimated RCS level.

[0056] SIGINT processing unit: integrate signal sorter, identifier (identify signal type / radar model based on deep learning or spectral analysis algorithm), modulator / demodulator, keyword extractor (preset military-related keyword list). Output a JSON object containing signal source ID, time, signal type code, working mode code, identified call sign, and intercepted text fragments.

[0057] OSINT processing unit: integrate network crawler, text entity identifier (identify organization, location, equipment model, event), geocoder. Output a JSON object containing associated target ID (if can be associated), time, information summary, involved entity, and geographic coordinates.

[0058] Image processing unit: integrate aircraft model identification model (based on convolutional neural network, etc.), paint recognition, and mounting recognition model. Output a JSON object containing target ID (if can be associated), time, identified model, paint feature, and mounting type.

[0059] Operational knowledge is embodied in this case as parameter settings, threshold judgments, and preliminary classification rules used to guide feature extraction, such as defining high-speed high-G turns as a threshold for fighter aircraft characteristics.

[0060] Data fusion module 3: Maintain a list of target IDs using a multi-sensor tracker.

[0061] For newly received feature data packets, determine whether they belong to an existing target ID or whether a new target ID needs to be created based on temporal, spatial location, motion trend, and the relevance of the feature itself (e.g., whether the SIGINT signal source location matches the radar track location).

[0062] Integrate all feature packets associated with the same target ID in chronological order to form a comprehensive snapshot of the target's current state or a sequence of behavior records over a period of time. For example, a target object includes its latest radar state, recent SIGINT activity records, related communication content, and background information associated from OSINT.

[0063] Knowledge base module 4 (based on knowledge graph): Use a graph database (such as Neo4j, ArangoDB) to store the knowledge graph.

[0064] Design ontology to define entity types and relationship types.

[0065] Import structured or semi-structured knowledge sorted by experts through data entry tools or ETL processes to fill the graph. For example, extract information from aircraft performance manuals, military organizational structure diagrams, electronic reconnaissance manuals, and historical battle analysis reports and convert them into graph format.

[0066] Provide API interfaces based on graph query languages such as Cypher or Gremlin for LLM reasoning modules to perform knowledge retrieval.

[0067] LLM reasoning module 5: Deploy one or more LLM model instances (such as Llama, Mistral, or other secure models in a private environment based on the Transformer architecture). You can fine-tune the model as needed to better understand military terminology and concepts.

[0068] Implement RAG manager: When receiving a target comprehensive feature package, the RAG manager analyzes the features, identifies key entities and attributes (such as "signal type = feature B", "call sign = Cobra", "flight profile = high-altitude circling").

[0069] The RAG manager calls the knowledge base API to retrieve relevant nodes and edges in the KG with these entities and attributes as query conditions.

[0070] For example, query the KG for entities and relationships that are directly or indirectly associated with "Feature B", "Cobra", and "High altitude circling" (e.g., "What radar does Feature B belong to?", "Which aircraft uses that radar?", "What unit does the Cobra call sign belong to?", "What aircraft does that unit equip?", "What is the typical profile of High altitude circling for a mission?").

[0071] The retrieved KG fragments (usually converted into concise textual descriptions) are organized into a structured Prompt along with the target composite feature and sent to the LLM. The Prompt design takes into account the context window limit, providing only the most relevant knowledge.

[0072] After the LLM receives the Prompt, it performs reasoning according to the reasoning steps specified in the Prompt (e.g., first evaluate identity evidence, then evaluate intent evidence, and finally make a comprehensive judgment).

[0073] For example, the Prompt can guide the LLM to: 1. Analyze the feature package and list the key observation facts; 2. Analyze the knowledge graph fragments and list the relevant background knowledge; 3. Combine facts and background knowledge to infer the target identity, explain the basis and confidence; 4. Combine facts and background knowledge to infer the target intent, explain the basis and confidence; 5. Point out any contradictions between facts and background knowledge.

[0074] The LLM outputs the reasoning results in a preset JSON format.

[0075] Result output and presentation module 6: Implement a post-processing unit to verify the format of the LLM output JSON and parse the content.

[0076] Write the results to the intelligence database and update the target situation record.

[0077] Develop a user interface based on Web or client / server architecture to visualize the target's location on the electronic map, track, identity label, and intent label. Users can click on the target to view detailed feature data, LLM reasoning conclusions, confidence, and specific evidence chains supporting the conclusions (highlighting relevant features and knowledge graph facts).

[0078] Implement an alarm rule that triggers a sound / visual alarm and pushes a notification to relevant personnel when the LLM infers a high-confidence specific intent (such as attack preparation) or detects an anomaly.

[0079] A feedback mechanism is provided to allow analysts to label or correct the identification results of the LLM and use these feedbacks to optimize the KG or LLM model.

[0080] System structure and module relationship: The whole system is in a pipeline architecture. The data acquisition module 1 is the input end, which sends the original data to the preprocessing layer. The data preprocessing and feature extraction module 2 processes different data sources in parallel, and outputs standardized features. The data fusion module 3 associates the features to the target. The knowledge base module 4 is an independent knowledge storage and query center, which provides background knowledge for the LLM reasoning module 5. The LLM reasoning module 5 is the core processing unit, which combines the fusion features and KG knowledge for reasoning. The result output module in the result output and presentation module 6 is the output end, which presents the results to the user or downstream system. Each module interacts through standardized data interfaces such as message queues, databases, and APIs. The RAG mechanism between the knowledge base module 4 and the LLM reasoning module 5 is the key connection for effective use of knowledge.

[0081] The embodiment of the present application provides a technical scheme: a method of target identification system based on large language model and knowledge graph, comprising the following steps: Step S1: real-time acquisition of heterogeneous original data of radar, signal intelligence (SIGINT), open source intelligence (OSINT), electro-optical / infrared (EO / IR) through the data acquisition module 1; Step S2: cleaning, denoising and standardization processing of the original data through the data preprocessing and feature extraction module 2, and extraction of structured features: Extract target ID, space-time coordinates, speed, height, heading, acceleration, maneuver type, radar cross section (RCS) features from radar data; Extract signal source ID, signal type, working mode, pulse feature, communication protocol, call sign and keyword from SIGINT data; Extract associated geographic location, event, equipment model from OSINT data; Extract aircraft model, painting, serial number, and mounting type from EO / IR data; Step S3: associate the features of the same target to a unique logical target ID through the multi-sensor tracking algorithm in the data fusion module 3, and generate a comprehensive feature package; Step S4: the knowledge base module 4 retrieves the background knowledge related to the target features based on the knowledge graph (KG). The KG nodes include aircraft model, radar model, unit, task type, tactical mode, and the edge relationship includes equipment relationship, deployment relationship, and tactical association; Step S5: organize the comprehensive feature package and KG retrieval results into a structured Prompt through the LLM reasoning module 5, and input the large language model (LLM) for reasoning: Identity recognition: Compare target features with the association of aircraft models, military units, call signs in KG; Intention recognition: Combine behavior sequences, electronic activity patterns, communication content, and KG task associations to infer deep intentions; Step S6: Output structured results including identity, intention, confidence, supporting evidence, and anomaly alerts through the result output and presentation module 6.

[0082] Notes for operation: Data synchronization and time calibration: Ensure that data from different sensors has accurate timestamps and strict time synchronization during fusion and reasoning.

[0083] Quality and update of knowledge graph: The accuracy and completeness of KG directly affect the reasoning results. A continuous knowledge update and verification mechanism needs to be established to ensure that KG reflects the latest intelligence and expert consensus.

[0084] Security and privacy of LLM: Deploy in isolated environment to avoid data leakage risk. Strictly review and filter input Prompt and output results.

[0085] Reasoning delay and resource optimization: LLM reasoning, especially complex Prompt and large models, may require significant computing resources and time. Performance optimization such as using more efficient models, hardware acceleration, optimizing Prompt design, batch reasoning, etc. is needed to meet real-time requirements.

[0086] Explainability and confidence evaluation: Although LLM is a black box, guiding its output reasoning process (such as Chain-of-Thought) and providing supporting evidence through Prompt engineering can improve the credibility of results. It is necessary to develop or utilize technology to evaluate the confidence of LLM output and clearly present it to analysts.

[0087] Adversarial learning: The system should have the ability to identify the deceptive behavior of the enemy (such as launching fake signals, deliberately executing misleading flight profiles) to mislead the system. This may require encoding known deception patterns in KG or training LLM to identify contradictions and anomalies in data.

[0088] Human-machine collaboration: This system is a tool to enhance the ability of human analysts, not a complete replacement. The system should be designed to facilitate human analysts to intervene, verify, correct and provide feedback, forming a closed-loop optimization.

[0089] Through the above specific embodiments, the present application can construct an intelligent system that can effectively integrate multi-source heterogeneous intelligence data and automatically perform high-precision identity and intention recognition, thereby improving the efficiency and depth of intelligence analysis and better supporting decision-making.

[0090] It has to be noted that, in the present document, the terms "first", "second", etc. merely serve to identify different entities or actions from each other, without necessarily requiring or implying any actual relationship or order between these entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0091] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application as defined by the appended claims and their equivalents.

Claims

1. A system for target recognition based on a large language model and knowledge graph, characterized in that: It includes a data acquisition module (1), a data preprocessing and feature extraction module (2), a data fusion module (3), a knowledge base module (4), an LLM reasoning module (5) and a result output and presentation module (6).

2. The object recognition system based on a large language model and knowledge graph according to claim 1 is characterized in that: The data acquisition module (1) is responsible for acquiring raw data streams from various reconnaissance sensors and open data sources in real time or quasi-real time.

3. The object recognition system and method based on a large language model and knowledge graph according to claim 1 is characterized by: The data preprocessing and feature extraction module (2) cleans, denoises, and standardizes the collected raw data, and extracts structured features from different types of data.

4. The object recognition system and method based on a large language model and knowledge graph according to claim 1, characterized in that: The data fusion module (3) receives feature data about the same target generated at different times or by different sensors from the data preprocessing and feature extraction module (2).

5. The object recognition system and method based on a large language model and knowledge graph according to claim 1 is characterized by: The knowledge base module (4) constructs and maintains a knowledge graph containing the business knowledge of domain experts.

6. The object recognition system and method based on a large language model and knowledge graph according to claim 1, characterized in that: The result output and presentation module (6) receives the structured analysis result output by the LLM reasoning module (5).

7. The object recognition method based on a large language model and a knowledge graph according to any one of claims 1 to 6, characterized in that: The following steps are involved: Step S1: Real-time acquisition of heterogeneous raw data of radar, signal intelligence (SIGINT), open source intelligence (OSINT), and electro-optical / infrared (EO / IR) through the data acquisition module (1); Step S2: The raw data is cleaned, denoised and normalized through the data preprocessing and feature extraction module (2), and structured features are extracted: Extract target ID, time and space coordinates, speed, altitude, heading, acceleration, maneuver type, and radar cross-section (RCS) characteristics from radar data; Extract signal source ID, signal type, operating mode, pulse characteristics, communication protocol, call sign, and keywords from SIGINT data; Extract relevant geographic locations, events, and equipment models from OSINT data; Extract aircraft model, paint scheme, aircraft serial number, and mount type from EO / IR data; Step S3: The multi-sensor tracking algorithm in the data fusion module (3) associates the features of the same target to a unique logical target ID to generate a comprehensive feature package; Step S4: The knowledge base module (4) retrieves background knowledge related to target features based on the knowledge graph KG. KG nodes include aircraft models, radar models, troop units, mission types, and tactical modes. Edge relationships include equipment relationships, deployment relationships, and tactical associations. Step S5: The comprehensive feature package and KG search results are organized into a structured prompt through the LLM reasoning module (5), and input into the large language model LLM for reasoning: Identity recognition: Compare the target features with the aircraft model, military unit, and call sign in the KG; Intent recognition: Inferring deep intent by combining behavioral sequences, electronic activity patterns, communication content, and KG task associations; Step S6: Output structured results through the result output and presentation module (6), including identity, intention, confidence, supporting evidence and abnormal alarm.

8. The object recognition method based on a large language model and a knowledge graph according to claim 7, characterized in that: The KG construction in step S4 specifically includes: A graph database is used to store entities and relationships, and entity types and relationship types are defined through ontology rules.

9. The object recognition method based on a large language model and a knowledge graph according to claim 7, characterized in that: The LLM reasoning in step S5 adopts a retrieval enhancement generation mechanism: Parse the key entities in the target features and query the KG to obtain multi-hop related knowledge.

10. The object recognition method based on a large language model and a knowledge graph according to claim 7, characterized in that: The result output of step S6 includes: Visualize target tracks, identity tags, and intention tags on electronic maps; Highlight the feature data and KG evidence chain that support the conclusion; An alarm is triggered when the intent confidence level is greater than 90% or a logical conflict is detected.

Images