Radiation source recognition reasoning question-answering system and method based on graph retrieval enhanced generation

Abstract

The invention discloses a radiation source recognition reasoning question-answering system and method based on graph retrieval enhancement generation, and belongs to the technical field of artificial intelligence and signal processing. The system comprises a data processing module, a knowledge graph construction module, a retrieval and reasoning module, a graph retrieval enhancement generation module and an output module. The method comprises the following steps: preprocessing multivariate data through a three-layer data architecture to generate a structured triple; constructing a dynamic knowledge graph; top-to-bottom accurate retrieval and bottom-to-top context optimization are realized by adopting a U-retrieval strategy; generating a traceable recognition analysis and processing suggestion based on the retrieval sub-graph in combination with a large language model; and outputting a final result with reference and term interpretation. According to the method, the problems that knowledge updating is lagged and reasoning cannot be explained in a traditional radiation source identification system are solved, the identification accuracy, the real-time performance and the result interpretability are remarkably improved, and the method is suitable for radiation source intelligent identification and reasoning tasks in a complex electromagnetic environment.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and signal processing technology, and in particular to a radiation source signal identification reasoning question answering system and method based on graph retrieval augmented generation (GraphRAG) technology. Technical Background

[0002] With the rapid development of wireless communication, radar, navigation, and various electronic devices, the electromagnetic spectrum environment is characterized by multi-source heterogeneity and dynamic changes. Traditional radiation source identification and question-answering systems based on static knowledge bases often face problems such as lagging knowledge updates, insufficient reasoning depth, and lack of interpretability.

[0003] GraphRAG technology combines the structured reasoning capabilities of knowledge graphs with the generative capabilities of large language models. However, its direct application in the highly specialized field of signal processing, radiation source identification, still faces the following challenges: insufficient accuracy in domain knowledge retrieval; complex relationships between signal features and semantics; and high requirements for interpretability and traceability of identification and reasoning results. Therefore, a system is needed that can integrate multi-source signal features, achieve knowledge traceability, and possess efficient reasoning capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide a radiation source identification reasoning question-answering system and method based on GraphRAG, so as to solve the problems of lagging knowledge updates and unexplainable reasoning in existing systems.

[0005] This invention achieves efficient retrieval and interpretable reasoning for multi-source electromagnetic signals by designing a three-layer data architecture, a dynamic knowledge graph construction mechanism, and a U-retrieval and GraphRAG reasoning process, thereby improving the accuracy and real-time performance of radiation source identification.

[0006] Compared with existing technologies, this invention has the following advantages: This method provides a structured triplet knowledge representation based on [entity, source, definition], ensuring the traceability of knowledge points; it adopts a U-retrieval strategy to achieve precise top-down positioning and bottom-up context optimization, which improves retrieval efficiency and ensures the integrity of the global context; the GraphRAG inference module, combined with a large language model, can generate verifiable and interpretable radiation source identification and signal analysis results; the system output includes citation sources and terminology explanations, enhancing the credibility and traceability of the results.

[0007] The overall process of the system of this invention includes the following steps:

[0008] S1. Based on the traditional "entity-relationship" knowledge graph, the system integrates three key data layers: RAG data layer: containing user-owned radiation source signal feature data; resource layer: accessing authoritative radiation source identification papers and technical reports; definition layer: embedding standard definitions of signal processing terminology. Through a large language model, the system intelligently parses and maps these three layers of data, automatically generating structured triples of [entity, source, definition].

[0009] S2. Construction of Dynamic Knowledge Graph: The system utilizes a large language model to perform deep semantic understanding on preprocessed data, automatically identifying and extracting key entities (such as signal feature parameters, equipment models, recognition algorithms, etc.). Through semantic similarity matching algorithms, entities are associated with authoritative materials in the resource layer, while simultaneously mapping to standard terms in the definition layer, forming complete knowledge units. This process ensures the traceability of knowledge points and supports dynamic incremental updates.

[0010] S3. Retrieval and Inference: The retrieval and inference module adopts a U-retrieval strategy, achieving efficient and comprehensive information retrieval through hierarchical retrieval and iterative optimization.

[0011] Top-down precise retrieval: drilling down layer by layer from high-level signal type concepts to specific radiation source instances;

[0012] Bottom-up response optimization: By backtracking through context, specific nodes are associated with the overall recognition system to ensure the integrity of the global context.

[0013] This method effectively focuses the search scope by precisely retrieving from high-level signal classification concepts down to specific radiation source instances through top-down retrieval; then, it links specific nodes to the overall signal identification system through bottom-up contextual backtracking, ensuring that the global context is not lost.

[0014] S4. GraphRAG Inference Generates Radiation Source Identification, Analysis, and Processing Suggestions. Utilizing GraphRAG (Graph Retrieval Enhanced Generation) technology, based on the retrieved radiation source entity-relationship subgraph, executable processing suggestions are generated. Combined with the reasoning capabilities of LLM (Large Language Model), relevant identification information is retrieved from the radiation source knowledge graph, and traceable signal feature analysis and processing suggestions are generated.

[0015] S5. Output the final answer with citations and terminology explanations. The final output includes: traceable citation sources; explanations of identified terms; and structured identification and reasoning results.

[0016] Compared to traditional multimodal reasoning tasks for radiation source identification, this system achieves significant breakthroughs and improvements in methodology. The system's innovation is mainly reflected in three dimensions: knowledge representation, retrieval mechanism, and reasoning process.

[0017] Knowledge representation level: Triples are used to achieve full-link tracing of signal processing knowledge;

[0018] At the retrieval mechanism level: U-retrieval strategy collaborative optimization, balancing retrieval efficiency and global integrity;

[0019] At the reasoning process level: an end-to-end reasoning chain is realized from the original signal features to the high-level identification conclusions, and the results are transparent and verifiable.

[0020] At the knowledge representation level, by employing structured triples of [entity, source, definition], the system achieves full-dimensional tracing of radiation source knowledge, effectively solving the knowledge silo problem existing in traditional identification systems. Regarding the retrieval mechanism, this system proposes a dedicated labeling system for radiation sources (including signal priority, modulation type, frequency band range, etc.) and achieves synergistic optimization with the U-retrieval strategy. This innovation achieves accurate retrieval with millisecond-level response speed, significantly improving retrieval efficiency compared to traditional global retrieval, while ensuring the integrity of the global context through a bottom-up contextual backtracking mechanism. In the reasoning process, the system innovatively constructs an end-to-end interpretable reasoning chain from original signal features to complete radiation source identification. The generated identification analysis results not only include specific technical solutions but also automatically link authoritative papers, technical manuals, and other sources, embedding standardized terminology explanations, making the decision-making basis of the entire identification process completely transparent and verifiable. Attached Figure Description

[0021] Figure 1 This is a flowchart of radiation source signal acquisition and preprocessing;

[0022] Figure 2 This is a flowchart demonstrating the use of the "U-retrieval" strategy;

[0023] Figure 3 This is a schematic diagram of question-answering reasoning implemented using GraphRAG. Detailed Implementation

[0024] The present invention will now be described in detail:

[0025] S1. Preprocessing of Multivariate Data: Based on the traditional "entity-relationship" knowledge graph, the system integrates three key data layers: the RAG data layer contains user-owned private radiation source signal feature datasets, such as measured signal parameters and signal acquisition environment records; the resource layer accesses authoritative signal processing academic papers and technical reports to ensure the accuracy of technical parameters and identification theories; and the definition layer embeds standard definitions of communication engineering terminology, such as standardized explanations of professional terms like modulation methods and signal bandwidth. Through intelligent parsing and association mapping of these three layers of data using a large language model, the system automatically generates structured triples of [entity, source, definition]. For example, it associates specific radiation source signal feature entities with theoretical explanations in the "Modern Signal Processing" technical report and standard definitions in "Communication Engineering Terminology." This three-layer data architecture not only achieves full-link traceability of knowledge points but also provides a highly reliable knowledge foundation for subsequent signal identification reasoning through cross-validation of authoritative sources and standard definitions. During preprocessing, the system also emphasizes unified encoding of multimodal data, associating unstructured data such as signal waveforms and spectrum diagrams with text descriptions for storage, ensuring that retrieval results include both professional explanations and intuitive displays of relevant signal features.

[0026] S2. Construction of Dynamic Knowledge Graph: This module first uses a large language model to perform deep semantic understanding on preprocessed multi-source radiation source data, automatically identifying and extracting key entities, including equipment models, signal characteristics, and recognition rules. Based on entity extraction, the system employs a semantic similarity-based matching algorithm to accurately associate each entity with authoritative technical documents, standards, and specifications in the resource layer. Simultaneously, the module automatically associates standardized terminology definitions in the definition layer, forming complete [entity, source, definition] triplet knowledge units. This construction method not only achieves full-dimensional tracing of knowledge points but also establishes a multi-level relationship network between entities, such as "signal association" and "feature derivation," through the semantic understanding capabilities of LLM. This provides a knowledge base with both professional depth and timeliness for subsequent signal recognition reasoning based on the knowledge graph. In particular, the system dynamically evaluates the consistency between newly acquired signal data and existing knowledge. When an unrecorded entity is detected, the incremental update mechanism of the knowledge graph is automatically triggered to ensure the system continuously maintains optimal recognition capabilities.

[0027] S3. The retrieval and reasoning module innovatively realizes intelligent retrieval and identification analysis of radiation source signals. The system adopts a U-retrieval strategy, first using a top-down precise retrieval approach, starting from the signal classification system. For example, starting with high-level concepts such as "modulated signal," it gradually drills down to the "phase modulation" signal type, ultimately locking onto specific modulation methods such as "QPSK modulation." This hierarchical retrieval method effectively filters out irrelevant data. After obtaining the precise target, the system initiates a bottom-up contextual backtracking, associating specific signal nodes upwards with their respective signal classification systems, and horizontally expanding to associated technical nodes such as "spectral characteristics" and "demodulation methods," ultimately forming a complete signal identification and analysis chain. This method effectively focuses the retrieval scope by drilling down layer by layer from high-level signal concepts to specific instances through top-down precise retrieval; then, it associates specific nodes with the overall identification system through bottom-up contextual backtracking, ensuring that the global context is not lost. A schematic diagram of the specific method is shown below. Figure 2 As shown.

[0028] When a user enters question Q, query tags are first generated using Q, and then a large language model is used. Generate tactical tags : Search downwards from the highest level to find the most matching subgraph. Let the set of labels for the i-th layer be . The j-th label is The retrieval formula is:

[0029] ,in Cosine similarity is calculated in the target subgraph. In the process, based on the similarity between the query and entity content, extract... Related entities , here

[0030] This refers to the maximum number of entities that can be retrieved.

[0031]

[0032] in This represents a text encoding function that maps text to a high-dimensional vector. Indicate the problem Vector representation of; It is a physical entity Content description; To expand the hop count of neighbors in a knowledge graph, if Then retrieve the entity All triplet neighbors within 2 hops. Expand its Jumping Triple Neighbor To form a retrieval subgraph , refers to entities of All triples within the jump, including those with Entities and relationships that are directly or indirectly connected; From The relevant entity set is extracted. After locating and querying the most relevant subgraph step by step from high level to low level, the initial response is optimized by utilizing the global context of the high-level labels.

[0033] S4. GraphRAG Inference Generates Radiation Source Identification, Analysis, and Processing Suggestions. Utilizing GraphRAG (Graph Retrieval Enhanced Generation) technology, based on the retrieved radiation source entity-relationship subgraph, executable processing suggestions are generated. Combined with the inference capabilities of LLM (Large Language Model), relevant identification information is retrieved from the radiation source knowledge graph, and traceable signal feature analysis and processing suggestions are generated. Specific operations are as follows: Figure 3 As shown.

[0034] Input question Q + retrieval subgraph Using large language models Generate preliminary tactical analysis: ; Integrate higher-level labels (such as "modulation identification theory" and "spectral feature analysis") layer by layer to optimize the response: ,

[0035] For the first The optimized response content is then reviewed. Further source tracing and credibility verification ensure that each suggestion originates from reliable data, supporting decision verification, and a retrieval subgraph is calculated. average confidence level :

[0036]

[0037] in This represents a text encoding function that maps text to a high-dimensional vector. Indicate the problem Vector representation of; It is a physical entity The content description. If the confidence level is less than the threshold (0.7), the message "Manual review required" will be displayed.

[0038] S5. Generating Explanatory Answers. Tests were conducted in various experimental environments (Neo4j graph database, localized large language model deployment, OneAPI deployment, etc.) for different types of questions, including global search, local search, and comprehensive queries. Results show that the GraphRAG-based graph reasoning radiation source signal recognition question-answering system significantly improves the accuracy, depth, and reasoning ability for complex logical relationships compared to traditional LLM question answering and traditional KGQA when handling knowledge-intensive tasks in specialized domains. It can accurately identify the core themes in documents, integrate multi-source knowledge, and provide answers with professional depth and practical case support. Compared to traditional large models, this method provides source references and more professional identification and recommendation analysis.

Claims

1. A radiation source identification reasoning question-answering system based on graph retrieval enhancement, characterized in that, The system includes: a data processing module for preprocessing multivariate data to construct a three-layer data architecture comprising a RAG data layer, a resource layer, and a definition layer, and generating structured triples using a large language model; a knowledge graph construction module for constructing a dynamic knowledge graph based on the structured triples to achieve entity recognition, association, and mapping; a retrieval and reasoning module for employing a U-retrieval strategy to achieve top-down precise retrieval and bottom-up context optimization, retrieving subgraphs related to the query; a graph retrieval enhancement generation module for generating traceable radiation source identification analysis and processing suggestions based on the retrieved subgraphs and combined with a large language model; and an output module for outputting recognition results with citations and terminology explanations.

2. The radiation source identification reasoning question-answering system based on graph retrieval enhancement generation according to claim 1, characterized in that, The RAG data layer includes user-owned private radiation source signal feature data; the resource layer includes authoritative radiation source identification papers and technical reports; and the definition layer includes standard definitions of signal processing terminology.

3. The radiation source identification reasoning question-answering system based on graph retrieval enhancement generation according to claim 1 or 2, characterized in that, The structured triple takes the form of [entity, source, definition].

4. The radiation source identification reasoning question-answering system based on graph retrieval enhancement generation according to claim 1, characterized in that, The U-retrieval strategy includes: a top-down retrieval phase, which drills down from high-level signal classification concepts to specific radiation source instances; and a bottom-up optimization phase, which uses context backtracking to associate specific nodes with the overall identification system.

5. The radiation source identification reasoning question-answering system based on graph retrieval enhancement generation according to claim 4, characterized in that, The top-down retrieval phase includes a retrieval process based on a semantic similarity matching algorithm, and the retrieval formula is as follows: ,in Cosine similarity calculation; query tags are generated using Q-factor, and a large language model is employed. Generate tactical tags : Search downwards from the highest level to find the most matching subgraph. The label set of the i-th layer is The j-th label is 6. The radiation source identification reasoning question-answering system based on graph retrieval enhancement generation according to claim 1, characterized in that, The graph retrieval enhancement generation module is also used to verify the credibility of the generated suggestions. The confidence calculation formula is as follows: ,in This represents a text encoding function that maps text to a high-dimensional vector. Indicate the problem Vector representation of; It is a physical entity Content description; To expand the hop count of neighbors in the knowledge graph; if the confidence level is less than the threshold, prompt "Manual review required".

7. The radiation source identification reasoning question-answering system based on graph retrieval enhancement generation according to claim 1, characterized in that, The knowledge graph construction module is also used to realize the dynamic incremental update of the knowledge graph.

8. A radiation source identification reasoning question-answering method based on graph retrieval enhancement, characterized in that, The method includes: preprocessing multivariate data to construct a three-layer data architecture comprising a RAG data layer, a resource layer, and a definition layer, and generating structured triples using a large language model; constructing a dynamic knowledge graph based on the structured triples; employing a U-retrieval strategy to achieve top-down precise retrieval and bottom-up context optimization, retrieving and querying related subgraphs; generating traceable radiation source identification analysis and processing suggestions based on the retrieved subgraphs and the large language model; and outputting identification results with citations and terminology explanations.

9. The radiation source identification reasoning question-answering method based on graph retrieval enhancement generation according to claim 8, characterized in that, The U-retrieval strategy includes: a top-down retrieval phase, which drills down from high-level signal classification concepts to specific radiation source instances; and a bottom-up optimization phase, which uses context backtracking to associate specific nodes with the overall identification system.

10. The radiation source identification reasoning question-answering method based on graph retrieval enhancement generation according to claim 8 or 9, characterized in that, The method also includes verifying the credibility of the generated suggestions. If the confidence level is lower than a set threshold, a manual review is required.

Images