Earthquake Disaster Emergency Simulation System Based on Multi-Agent and Knowledge Graph

The earthquake disaster emergency simulation system based on multi-agent and knowledge graphs solves the problems of weak multi-hop logical reasoning ability and entity confusion in earthquake emergency response, and achieves efficient and reliable emergency decision support, ensuring the system's compliance and second-level response capability.

CN123117370APending Publication Date: 2026-11-17GUANGDONG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611275681.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies have weak multi-hop logical reasoning capabilities in earthquake emergency response, semantic confusion caused by entities with the same name but different meanings, and insufficient system robustness, making it impossible to achieve a highly available, second-level industrial-grade disaster prevention and mitigation control closed loop.

Method used

An earthquake disaster emergency simulation system based on multi-agent and knowledge graph is adopted. By constructing a five-level hierarchical topology task tree, joint entity disambiguation unit and dynamic fault-tolerant replanning unit, multi-agent collaborative reasoning is realized. Combined with graph database and vector database cluster, multi-hop logical reasoning and entity disambiguation are performed to ensure the compliance and traceability of decision-making.

Benefits of technology

It significantly improves multi-hop logical reasoning capabilities, eliminates entity confusion, enhances the robustness and reliability of the system, and can provide structured emergency decision support within a second-level response time.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117370A_ABST
    Figure CN123117370A_ABST
Patent Text Reader

Abstract

This invention discloses an earthquake disaster emergency simulation system based on multi-agent and knowledge graph, relating to the fields of computer applications and disaster prevention and mitigation automation. The system includes a data access and preprocessing module, a knowledge graph storage module, a multi-agent collaborative reasoning module, and a business output module. Its core adopts a multi-agent closed-loop architecture of "planning-execution-verification-replanning." This invention achieves multi-hop logical reasoning of the knowledge graph by constructing a five-level hierarchical topology task tree, solving the problem of easily broken links in traditional vector retrieval reasoning; it constructs ontology boundary protection through a three-stage joint entity disambiguation mechanism, eliminating decision illusions caused by homonymous entities; and it improves system engineering robustness through dynamic replanning, degradation fault tolerance, and token budget control. This invention can achieve fully traceable, traceable earthquake emergency task adaptive assignment, supporting a highly reliable, second-level industrial-grade disaster prevention and mitigation control closed loop.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of computer applications and automated disaster prevention and mitigation technologies, specifically to an earthquake disaster emergency simulation system based on multi-agent systems and knowledge graphs. Background Technology

[0002] With the development of Large Language Model (LLM) and Retrieval-Augmented Generation (RAG) technologies, vector retrieval-based intelligent question answering and knowledge augmentation generation methods have been widely applied in the field of natural disaster emergency response.

[0003] Existing technologies typically employ the following processing flow: (1) Chunking is performed on long texts of emergency response plans; (2) Use the text embedding model to map text blocks into vector representations and store them in a vector database; (3) After receiving the disaster information input, it is encoded into a query vector and similarity retrieval is performed in the vector space based on cosine similarity; (4) After recalling the relevant text fragments, input them into the large language model for unified generation, thereby outputting emergency response suggestions or disposal plans.

[0004] The aforementioned vector-based RAG method has some effectiveness in general question answering and knowledge enhancement tasks, but in complex decision-making scenarios such as earthquake emergency response, which have strong constraints, multi-level hierarchical responsibility structures, and strict legal constraints, it still has the following limitations: 1. Weak multi-hop logical reasoning ability, unable to handle cross-document, multi-level emergency responsibility network simulations. Emergency response after an earthquake has a strict legal hierarchy. Traditional vector RAGs only perform local searches based on literal or semantic similarity, lacking an understanding of the macro-level topological structure of documents. When it is necessary to deduce long-link, cross-document, complex multi-hop logic such as "disaster trigger response level, activation of responsible departments, and execution of standard task codes" based on specific earthquake magnitude and geographical conditions, chain-like reasoning is prone to breakage, failing to automatically form a structured flow scheme for a complete chain of evidence.

[0005] 2. Entities with the same name but different meanings can lead to semantic confusion, which can easily cause misjudgments or erroneous responses to secondary disasters. In real earthquake disaster reduction collaboration systems, there are a large number of heterogeneous entities with highly similar place names and organization names but completely different rights and responsibilities. Traditional vector retrieval lacks the ontology boundary protection of knowledge graphs. In cross-regional and cross-departmental collaborative computing, the proximity of semantic feature vector spaces can easily create the illusion of regional or responsibility misalignment, leading to the generation of incorrect emergency resource allocation decisions.

[0006] 3. Linear instruction flows lack dynamic self-healing and compliance verification mechanisms, resulting in insufficient system robustness. Traditional question-and-answer or instruction issuance systems employ "one-way chain" calls. When the large language model outputs incorrect formats or external retrieval returns high-noise information, the system will directly crash or output non-compliant content. The system lacks multi-agent collaboration and adaptive replanning capabilities, making it unable to achieve dynamic fault tolerance and compliance self-verification in long-link inference, and thus unable to support highly available, second-level response industrial-grade disaster prevention and mitigation control closed loops. Summary of the Invention

[0007] To address the shortcomings mentioned in the background art, this invention provides an earthquake disaster emergency simulation system based on multi-agent and knowledge graph.

[0008] To solve the above-mentioned technical problems, the technical solution provided by the present invention is: an earthquake disaster emergency simulation system based on multi-agent and knowledge graph, including a data access and preprocessing module, a knowledge graph storage module, a multi-agent collaborative reasoning module, and a business output module; The data access and preprocessing module is used to access multi-source heterogeneous earthquake disaster information and complete unified formatting processing; The knowledge graph storage module includes a graph database and a vector database cluster, which are used to store an earthquake emergency knowledge graph composed of a five-level hierarchical topology task tree, an entity vector index, a standard operating procedure rule base, and a historical experience case base. The multi-agent collaborative reasoning module includes a master agent, an execution agent network, a verification agent, and a replanning agent. It has a built-in joint entity disambiguation unit and a dynamic fault-tolerant replanning unit. It completes disaster simulation and task assignment decision-making through a closed-loop process of "planning-execution-verification-replanning". The business output module is used to integrate the full chain of evidence simulation results and generate a structured emergency simulation report and a standardized resource assignment work order.

[0009] Furthermore, the five-level hierarchical task tree, from top to bottom, consists of the emergency process root node, task domain node, task category node, standard task code node, and responsible department / professional resource node; each node contains a unique node identifier, node name, node type, and legal basis clause number attribute.

[0010] Furthermore, the joint entity disambiguation unit adopts a three-stage mechanism of "string recall + vector rearrangement + NIL detection": The first phase involves full matching, prefix matching, suffix matching, and pinyin matching based on the standard entity list to recall the candidate entity set. The second stage calculates the cosine similarity between the query vector and the candidate entity vector using a text embedding model to complete the semantic ranking of candidate entities. The third stage performs boundary detection through dual similarity thresholds. When the highest similarity is lower than the lower threshold, the electron (NIL) signal is triggered, the system suppresses fuzzy generalization generation and turns to external information verification.

[0011] Furthermore, the master control agent is used to extract key projection parameters such as magnitude, epicenter node, and population density from standardized disaster information, decompose the projection task into a sequence of sub-tasks that can be executed in parallel or serially, and encapsulate them into structured planning instructions in JSON format and send them to the execution agent network.

[0012] Furthermore, the execution agent network is used to assemble graph database query statements according to subtask types and perform multi-hop path retrieval along the five-level hierarchical topology task tree; the retrieval transmission logic is: specific magnitude and geographical conditions → trigger → response level → activation → responsible department → execution → standard task code → association → professional rescue resources and technical support nodes; the retrieval process records complete intermediate node information and encapsulates it into an evidence chain structure containing path nodes, path edges and complete path descriptions.

[0013] Furthermore, the verification agent is used to perform double verification on the inference results: Sufficiency verification involves checking whether the four decision-making elements—response level determination, list of responsible departments, standard task codes, and evidence chain traceability information—are complete. Compliance verification compares key decisions such as response level and task assignment with standard operating procedure rules in the knowledge graph to verify their legal compliance and the authenticity of the provisions.

[0014] Furthermore, the replanning agent is used to trigger dynamic replanning when verification fails. The strategy adjustment includes reducing the generation temperature of the large language model, expanding the range of hop counts for graph queries, and correcting external search keywords. The system sets a maximum retry threshold. When the threshold is reached, a degradation fault tolerance mechanism is triggered, a list of uncertain factors to be verified is output, and the decision-making power is transferred to humans.

[0015] Furthermore, the multi-agent collaborative reasoning module also includes a token budget control unit, which calculates the context token consumption in real time through a word segmenter that matches the target large model; it employs text-chunkBM25 retrieval and embedding a joint scoring machine to prune low-information-density content, performs entity-level triplet summarization pruning on multi-turn dialogues, and controls the context window size and token consumption.

[0016] Furthermore, when the information within the graph is insufficient, the execution intelligent agent network calls a compliant external retrieval engine to obtain supplementary information, which is then fused into the evidence chain after webpage cleaning and structured parsing. In cross-regional and cross-departmental collaborative scenarios, the joint entity disambiguation unit is called to perform secondary verification on key nodes of the retrieval path.

[0017] The advantages of this invention compared to existing technologies are: 1. Significantly improved multi-hop logical reasoning ability. This invention organizes emergency knowledge into an explicit entity-relationship triplet network, and replaces discrete vector similarity retrieval with multi-hop traversal of topological paths. Each step of the deduction retains a complete chain of evidence, naturally adapting to the hierarchical transmission logic of emergency responsibilities, fundamentally solving the problem of easily broken cross-document and multi-level reasoning links.

[0018] 2. Breakthrough improvements in entity disambiguation and illusion suppression. The three-stage joint disambiguation mechanism replaces traditional unbounded vector space retrieval by using closed candidate set matching and ontology boundary detection. It constructs an entity boundary protection layer from the bottom layer, completely eliminating confusion of entities with the same name and decision illusions in cross-regional and cross-departmental collaborations.

[0019] 3. The system's robustness and closed-loop control capabilities are substantially enhanced. The multi-agent closed-loop architecture realizes a complete control flow of verification-feedback-replanning, and together with retry circuit breaking and degradation fault tolerance mechanisms, it ensures that the system latency is controllable. Dynamic memory pruning and token budget control avoid context overflow, greatly improving the engineering stability and concurrent throughput of long-link inference, and can support industrial-grade disaster mitigation control closed loops with second-level response. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the physical hardware deployment topology of the system of the present invention; Figure 2 This is a diagram of the multi-agent collaborative architecture of the present invention; Figure 3 This is a flowchart of the overall control program of the system of the present invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings.

[0022] Combined with appendix Figure 1-3 An earthquake disaster emergency simulation system based on multi-agent and knowledge graph includes a data access and preprocessing module, a knowledge graph storage module, a multi-agent collaborative reasoning module, and a business output module.

[0023] Data access and preprocessing module: Receives disaster data from earthquake monitoring networks, local government disaster reporting systems, and publicly available internet information platforms, and completes unified formatting processing.

[0024] The knowledge graph storage module includes a graph database and a vector database cluster, storing knowledge graph data, entity vector indexes, standard operating procedure rule bases, and historical experience case bases. It provides graph query and vector retrieval services. The knowledge graph adopts a five-level hierarchical topology task tree structure, from top to bottom: emergency process root node, task domain node, task category node, standard task code node, and responsible department / professional resource node. Each node carries attributes such as a unique identifier, name, type, and legal basis clause number.

[0025] The multi-agent collaborative reasoning module includes a master agent, an execution agent network, a verification agent, and a replanning agent. It incorporates a joint entity disambiguation unit and a dynamic fault-tolerant replanning unit. The overall process adopts a closed-loop collaborative flow of Plan-Execute-Validate-Report. The master agent is responsible for extracting disaster parameters, decomposing sub-tasks, and issuing planning instructions. The execution agent network is responsible for multi-hop graph retrieval, external supplementary retrieval, and secondary entity disambiguation verification. The verification agent is responsible for dual verification of sufficiency and compliance. The replanning agent is responsible for strategy adjustment and degradation fault tolerance when verification fails. The joint entity disambiguation unit adopts a three-stage mechanism of "string recall + vector rearrangement + NIL detection" to complete entity matching and boundary detection in a closed candidate set limited by the knowledge graph ontology boundary, thus eliminating decision illusions caused by homonymous entities.

[0026] Business output module: Displays simulation results, generates structured emergency reports, sends resource scheduling work orders to external emergency management systems, realizes closed-loop execution of disaster response, and has a built-in Token budget control unit. It accurately calculates token consumption through a matching word segmenter, combines semantic scoring to prune redundant context, and performs dynamic memory pruning to avoid context window overflow.

[0027] This embodiment provides an earthquake disaster emergency simulation system based on multi-agent and knowledge graph. The system is deployed in a distributed computer hardware environment, and the hardware nodes include: a data access and preprocessing server, a graph database and vector database cluster, a multi-agent collaborative reasoning computing node, and a service output and command issuance terminal. The overall system workflow is divided into an offline construction phase and an online simulation phase, and the specific implementation process is as follows: I. Offline Build Phase (Steps S1-S3)

[0028] Step S1: Preprocessing and entity relation extraction of unstructured emergency response plan text; S1-1 Text Acquisition and Cleaning; The system reads unstructured texts such as the "National Earthquake Emergency Response Plan" and related regional emergency response frameworks. It then performs format standardization processing on the collected text, including removing irrelevant headers and footers, converting tables to text descriptions, and standardizing paragraph merging and splitting.

[0029] S1-2 Entity and Relation Extraction Based on Large Language Model; The cleaned text is segmented into several semantically complete text segments according to chapter structure. For each text segment, the following structured extraction prompt words are constructed and input into the large language model. The large language model outputs a set of structured entity relation triples E={(h,r,t)|h∈H,r∈R,t∈T} based on the above prompt words, where H is the head entity set, R is the relation type set, and T is the tail entity set.

[0030] Construction of the S1-3 five-level hierarchical topology task tree; The system organizes the extracted triple set hierarchically, constructing a topological task tree with the following five-level structure. Each node in the task tree contains attribute information: node unique identifier (node_id), node name (name), node type (node_type), legal basis clause number (legal_basis), etc.

[0031] Table 1. Topology Task Tree of Five-Level Structure

[0032] Step S2: Construction and storage of the knowledge graph; S2-1 graph database modeling; An attribute graph data model is used to map the entities extracted in step S1 to nodes in the graph database, and the relationships to edges. The definitions of nodes and edges are as follows.

[0033] Node: Each entity corresponds to one node, and the node contains a set of attributes (such as name, type, description, legal basis, etc.). Edge: Each relation corresponds to a directed edge, and the edge contains attributes (such as relation type, confidence level, etc.).

[0034] S2-2 spectral data imported into the database; Write the above nodes and edges into the graph database, and perform the following operations during the database insertion process: Generate a globally unique identifier for each node; Create indexes between nodes (especially composite indexes for node names and types to accelerate subsequent retrieval); The five-level hierarchical topology is explicitly stored in the form of paths, which facilitates quick location during multi-hop retrieval.

[0035] The graphical organization of the three types of core data in S2-3; The graph should contain at least the following three types of core data subgraphs: SOP rule base subgraph: Stores the triggering relationships between disaster level, emergency response level, and response measures; The executor network subgraph stores the various departments and organizations, their responsibilities, and collaborative relationships. Historical experience database sub-graph: Stores historical disaster cases and standard handling templates.

[0036] Step S3: Construction of the entity disambiguation benchmark library; The establishment of the S3-1 standard entity directory; Extract all place name nodes and organization name nodes from the knowledge graph to construct a standard entity directory library. Each standard entity records its standardized name, alias set, administrative division, unique code, and other information.

[0037] Pre-calculation of entity vectors in S3-2; For each entity in the standard entity directory, a pre-trained text embedding model is used, here BGE is chosen to calculate its semantic vector, which is stored in the vector database to form an entity vector index, which is used for similarity calculation in the subsequent vector reordering stage.

[0038] II. Online simulation phase (steps S4-S13) Step S4: Collection and input of multi-source unstructured disaster information in the earthquake zone; S4-1 Multi-Source Data Access; The system receives real-time reports of multi-source unstructured disasters in the earthquake zone from the following channels: Quantitative data (structured) generated by the earthquake monitoring network, including magnitude, epicenter location, and time of occurrence. Text descriptions (unstructured) sent back by on-site rescue teams, such as "abnormal ground fracturing and unusual noises occurred in a certain area"; Disaster situation reports submitted by local governments (semi-structured); Unstructured disaster descriptions emerging from public platforms.

[0039] S4-2 Unified formatting of input information; The system converts the aforementioned multi-source heterogeneous data into a standardized text input format Q={location_description,magnitude,time,damage_description,...}, where location_description is the location information text in the disaster description, magnitude is the magnitude value, and damage_description is the disaster damage description text.

[0040] Step S5: Joint entity disambiguation (one of the core innovations); This step aims to accurately map the place names and organization names extracted in the disaster description in step S4 to the corresponding standard entity nodes in the knowledge graph, thereby solving the "misattribution" computational illusion caused by homonymous entities in cross-regional and cross-departmental collaborations in traditional vector retrieval.

[0041] S5-1 Phase 1 – String Recall; For the location description in the input text Q, perform the following operations: String matching based on standard entity directories (including full matching, prefix matching, suffix matching, and pinyin matching); Recall all candidate standard entities C_candidate={c1,c2,...,c k}, where k≥1.

[0042] S5-2 Second Stage – Vector Rearrangement; For the candidate entity set C_candidate recalled in the first phase, perform semantic vector similarity ranking: The location description portion of the input text Q is converted into a query vector using an embedding model. V q =Embed(Q_location); Calculate V q With each candidate entity vector V i Cosine similarity between them: ; Candidate entities are sorted from high to low similarity, and the Top-N candidates are selected (N is preferably 3).

[0043] S5-3 Phase 3 – NIL (Not in the Vocabulary) Detection Perform boundary detection on candidate entities after vector rearrangement: If the highest similarity is greater than or equal to the preset threshold θ1, then the entity is directly selected as the disambiguation result; if the highest similarity is less than the preset threshold... If θ1 is greater than or equal to the preset threshold θ2, then ambiguity is determined and context disambiguation is triggered: extract the context information in the input text and make a joint judgment with the associated attributes of the candidate entity; If the highest similarity is less than the preset threshold θ2, the NIL signal is triggered, determining that the current entity is not within the ontology boundary of the knowledge graph. The system automatically suppresses fuzzy generalization generation and turns to an external secure network retrieval engine to conduct cross-source information verification.

[0044] Mapping of spectral nodes in S5-4 disambiguation results; After the above three-stage disambiguation process, the locations and organization names in the disaster description are accurately mapped to the corresponding standard entity nodes in the knowledge graph, resulting in a mapping result set M={(mention_i,node_id_i)}.

[0045] Step S6: Task decomposition and planning of the master control agent (Orchestrator); Standardized extraction of S6-1 disaster parameters; The master control agent receives disaster information after entity disambiguation and extracts a set of key inference parameters P={magnitude, epicenter (epicenter node ID), population_density, building_type,...}.

[0046] S6-2 Task Breakdown Based on the inference parameters P, the master control agent decomposes the complex disaster simulation task into several sub-task sequences T={t1,t2,...,t...} that can be executed in parallel or sequentially. m}, including but not limited to: Subtask t1: Disaster severity assessment, determining which level of emergency response corresponds to; Subtask t2: Identify responsible departments, identifying which departments should be activated; Subtask t3: Standard task code matching, specifying which SOP task codes each responsible department should execute; Subtask t4: Professional resource node matching, matching corresponding rescue resources and technical support nodes.

[0047] Generation of S6-3 planning instructions; The master agent encapsulates the above subtask sequence into structured planning instructions and issues them to the executor agent network. The planning instructions are in JSON object format, containing fields such as: task ID, task type, input parameters, and expected output format.

[0048] Step S7: Perform graph multi-hop retrieval using the ExecutorNetwork. Assembly of S7-1 map query statements; The executing agent receives subtasks from the master agent and assembles the corresponding graph database query statements according to the subtask type.

[0049] S7-2 multi-hop topology path retrieval; The executing agent submits the assembled Cypher statement to the graph database execution engine, performing multi-hop searches along the five-level hierarchical topology task tree. The search path follows the following propagation logic: Specific magnitude and geographical conditions → Triggering → Response level → Activation → Responsible departments → Execution → Standard mission code → Association → Professional rescue resources and technical support nodes Each jump records the complete information of the intermediate nodes, forming a structured retrieval result with a complete chain of evidence.

[0050] S7-3 evidence chain encapsulation of search results; The path results returned by multi-hop retrieval are encapsulated into an evidence chain structure: Evidence={path_nodes:[node1, node2,...],path_edges:[edge1, edge2,...],full_trace: complete path description}, ensuring that each step of the deduction has a legal basis and is traceable.

[0051] Step S8: Secondary verification of entity disambiguation; After the intelligent agent completes the graph query, if the retrieval path involves cross-regional and cross-departmental collaborative scenarios, the system calls the entity disambiguation module (step S5) again to perform secondary verification on the key nodes in the path to ensure that the node path returned by the retrieval is consistent with the administrative division and institutional affiliation of the disaster-stricken area, and to prevent path mismatch caused by homonyms.

[0052] Step S9: External data supplementation retrieval (conditional trigger); When the information within the graph is insufficient to support a complete deduction, the agent network invokes an external retrieval engine for supplementary searching. S9-1 Generation of search keywords; Based on the currently acquired map node information, the names of core entities (such as epicenter location names and response level codes) are extracted as search keywords.

[0053] S9-2 External Data Retrieval; By initiating retrieval requests to external data sources through domestically compliant APIs, supplementary information can be captured from compliant data sources such as WeChat official accounts and government platforms.

[0054] In-depth analysis of S9-3 webpage content; The messy web page text returned by the search is cleaned by a web page deep parser and converted into structured Markdown text. Core factual information such as time, place, event description, and data source is extracted.

[0055] S9-4 Supplementary Information: Graph Fusion; The core facts extracted from external searches are associated with existing graph nodes and added to the evidence chain.

[0056] Step S10: Verify the compliance of the validator; S10-1 Sufficiency verification; Verify that the agent checks whether the inference results returned by the agent network contain complete decision elements: 1. Does it include a clear response level determination? 2. Does it include a complete list of responsible departments? 3. Does it include the standard task codes corresponding to each department? 4. Does it contain complete evidence chain tracing information? If any element is missing, it is determined to be "insufficient" and proceeds to step S11.

[0057] S10-2 Compliance Verification; The verification agent compares and verifies key decisions (response level, task assignment, etc.) in the inference results with the SOP rules stored in the knowledge graph: 1. Check whether the determined response level matches parameters such as magnitude and damage. 2. Check whether the task assignment complies with the mandatory provisions of the "National Earthquake Emergency Response Plan" and related regulations; 3. Check whether the legal basis clause numbers in the chain of evidence actually exist in the source document.

[0058] 4. If the verification passes, proceed to step S12; if the verification fails, proceed to step S11.

[0059] Step S11: Replanning and Dynamic Self-Healing Mechanism; S11-1 Encapsulation of error messages; When the verification in step S10 fails, the verification agent encapsulates the missing or non-compliant information into an error feedback object.

[0060] S11-2 replanning trigger; Error feedback is sent back to the master agent (Orchestrator), triggering a replanning process. The master agent, based on the current error tracking, dynamically adjusts its strategy and re-executes the simulation. The strategy adjustments during replanning include: The generation temperature of the large language model is dynamically reduced from 0.3 in the first round to 0.1 or 0.0, which forcibly suppresses the model's divergence. Adjust the hop count range for graph queries (e.g., expand from 1-5 hops to 1-7 hops). Correct the search keywords and re-retrieve external data.

[0061] S11-3 Retry Count Control and Degradation Fault Tolerance; The system sets a maximum retry count threshold N_max (N_max is preferably 3). At each retry, the retry counter retry_count increments by 1. If retry_count<N_max, a new round of inference after replanning is executed (return to step S6). If retry_count≥N_max, a degraded fault tolerance mechanism is triggered: the verification agent forcibly terminates the planning, marks the current information degree as "locally sufficient", clearly outputs the "list of uncertain elements to be verified" in the form of a highlighted box in the final report, and safely transfers the decision-making power to the human commander, ensuring that the system does not fall into an infinite dead loop.

[0062] S11-4 Dynamic memory pruning; During multiple retries, the system performs dynamic memory pruning in real time. The original initial user query and the core objective of the master agent are retained. The failed original web crawling content of up to ten thousand words from the previous round is forcibly removed, and only the structured missing information assertions produced by the verification agent are retained. Entity-level triple summary extraction is performed on the multi-round dialogue history to compress redundant context.

[0063] Through the above pruning mechanism, it is ensured that the context window sent to the large language model at each retry is not overwhelmed by invalid information, reducing Token consumption.

[0064] Step S12: Token budget control and dynamic context management; S12-1 Accurate calculation of Token budget; Before each large language model call is executed, the system invokes a Tokenizer (a tokenizer matching the target large model) to accurately calculate the Token consumption of the current context.

[0065] S12-2 Dynamic context trimming; If the total Token amount of the current context exceeds Available_Tokens, the system triggers dynamic context trimming, preferentially removing content with low information density, such as HTML tags of web pages, irrelevant navigation text, etc. Key paragraph extraction based on semantic paragraph similarity is performed on long texts, and only paragraphs with high relevance are retained. The specific technology used in this patent is the joint scoring of Text-ChunkBM25+Embedding. The redundant content is replaced with the refined knowledge graph triple summary.

[0066] Through the above trimming, the Token consumption of a single request can be reduced by more than 40%, completely eliminating system crashes caused by context overflow.

[0067] Step S13: Structured emergency inference report generation and resource assignment; S13-1 Assembly of structured report; The system integrates the results of the full chain of evidence simulation to generate a structured emergency simulation report containing the following: Basic information and assessment conclusions regarding the disaster: magnitude, epicenter location, affected area, and response level determination; Legal basis and evidence chain tracing: the provision number and original citation of each step of the deduction; List of departments to be activated: Names of departments to be activated, descriptions of their responsibilities, and collaborative relationships; Standard Task Code Assignment Table: Standard task codes and specific operating instructions to be executed by each department; Professional resource node assignment scheme: a list of corresponding professional earthquake disaster relief resources and technical support nodes.

[0068] S13-2 Generation and issuance of emergency support resource assignment work orders; The system automatically generates standardized resource allocation work orders based on the simulation results, including fields such as work order number, issuance time, executing department, task code, operation instructions, and completion deadline. The work orders are distributed to the corresponding responsible nodes (such as the monitoring and forecasting department, the publicity and education center, and the geophysical institute) through the system interface, thus establishing a complete technical loop from "reporting of multi-source unstructured disasters in the earthquake zone" to "precise resource allocation".

[0069] (iv) Logical relationships and data flow between each step.

[0070] The logical relationship between steps S1 to S13 above is as follows: Steps S1-S3 (offline construction) form the system's data infrastructure layer, which only needs to be executed once or updated periodically during the entire system's operating cycle, providing a knowledge graph and disambiguation benchmark library for steps S4-S13 (online inference). Steps S4-S5 are the input and preprocessing layers for disaster information, responsible for standardizing multi-source heterogeneous data and completing entity disambiguation; Steps S6-S9 constitute the multi-agent collaborative inference layer, which is the core decision engine of the system; Steps S10-S11 constitute the quality assurance and fault tolerance layer, ensuring the compliance of the simulation results and the engineering robustness of the system; Step S12 is the resource control layer, which runs through the entire online simulation process to ensure the efficient use of system resources; Step S13 is the business output layer, completing the full business loop from decision-making to execution.

[0071] The layers communicate with each other through standardized data interfaces, achieving functional decoupling and modular reuse.

[0072] Advantages of this patented technical solution: (a) Significant improvement in multi-step logical reasoning ability Traditional vector retrieval (RAG) schemes calculate vector similarity by physically segmenting unstructured text into blocks. Essentially, this is local semantic matching, lacking the ability to perceive the macroscopic topological structure of documents. When faced with long-chain deductions spanning multiple documents and involving five levels of hierarchical topological relationships—such as "specific magnitude and geographical conditions → triggering response level → activating responsible departments → matching standard task codes"—vector retrieval can only perform isolated similarity calculations between individual text blocks, failing to establish cross-block and cross-document association paths, making chain-like reasoning prone to breakage.

[0073] This invention organizes emergency knowledge into an explicit entity-relationship-entity triplet network by constructing a five-level hierarchical topology task tree and storing it in a graph database. The executing agent performs multi-hop graph traversal along the topological path, recording complete intermediate node and edge information at each step of the deduction, forming a traceable chain of full evidence. The structured nature of the graph allows the system to naturally understand the causal transmission logic of "trigger → activation →", rather than relying on statistical semantic similarity guesses.

[0074] (II) Breakthrough improvements in entity disambiguation and hallucination suppression One of the core flaws of traditional vector retrieval (RAG) schemes lies in the lack of ontology boundary protection. In the embedding vector space, entities with similar semantics but completely different responsibilities may be confused due to the similarity of their text feature vector spaces, leading to misattribution or misalignment of responsibilities in the search results. This computational illusion can cause serious secondary disasters and misjudgments in emergency decision-making scenarios.

[0075] This invention designs a three-stage joint disambiguation mechanism of "string recall + vector rearrangement + NIL detection", which fundamentally constructs an entity ontology boundary protection layer, wherein: During the string retrieval phase, deterministic matching is performed based on the standard entity list to ensure that no candidate entities are missed; In the vector rearrangement stage, a refined semantic ranking is performed within the candidate set, rather than an unbounded search is performed across the entire vector space; During the NIL detection phase, when the similarity of all candidate entities is lower than the preset threshold, the entity is explicitly determined to be "not in the knowledge graph". The system actively suppresses fuzzy generalization generation and turns to external verification instead of forcibly giving an incorrect match.

[0076] The essential difference of this mechanism is that traditional vector retrieval performs similarity matching in an infinitely open vector space, while this invention performs deterministic selection and boundary detection in a closed candidate set limited by the knowledge graph ontology boundary, thus eliminating entity confusion in cross-regional and cross-departmental collaboration from the bottom up.

[0077] (III) Substantial Enhancement of System Robustness and Closed-Loop Control Capability Traditional vector retrieval (RAG) schemes employ a linear, unidirectional instruction flow: user query → vector retrieval → LLM generation → output. If any step in this process fails (e.g., the retrieval returns a large amount of noise, or the LLM output format is incorrect), the entire system will crash or output non-compliant content, lacking dynamic self-healing and fault tolerance mechanisms.

[0078] This invention constructs a four-stage multi-agent closed-loop control architecture: Plan-Execute-Validate-Report. Through the decoupling design of the master agent (Orchestrator), the execution agent network (Executor), and the validator (Validator), the following core robustness guarantees are achieved: Sufficiency verification and closed-loop feedback: The verification agent performs double verification (sufficiency + compliance) on the execution results. If the conditions are not met, it triggers replanning instead of unidirectional linear output. Retry count circuit breaking and degradation fault tolerance: Set a maximum retry threshold (preferably 3 times), and after the threshold is reached, the system is safely handed over to the human commander to ensure that the upper limit of system latency is controllable and to prevent infinite loops. Dynamic memory pruning: In multiple rounds of retries, redundant context is removed, and only core entities and missing information assertions are retained to prevent context window explosion; Precise control of the token budget: Real-time calculation of token consumption, reserved output space, and prevention of long reports being truncated.

[0079] Other alternatives to this patent: (1) Replacement of knowledge graph storage methods In addition to being stored in a graph database, the five-level hierarchical topology task tree can also be stored in a relational database (multi-hop queries are achieved through entity tables and relational tables in conjunction with recursive common table expressions) or an in-memory graph structure (all data is loaded into the in-memory adjacency table when the system starts, and high-speed traversal is achieved through in-memory computation). All of the above methods can realize the storage and multi-hop retrieval of structured knowledge.

[0080] (2) Alternatives to entity disambiguation methods The three-stage disambiguation mechanism of "string recall + vector rearrangement + NIL detection" can be replaced by a pure rule disambiguation engine based on a rule dictionary, or by a context disambiguation method based on a large language model (inputting the entity reference and its context into the large language model and allowing it to autonomously match and select from the candidate list). Both methods can eliminate the confusion of entities with the same name but different meanings.

[0081] (3) Replacement of multi-agent cooperative architecture The Plan-Execute-Validate-Report four-stage architecture can be replaced by a hierarchical decomposition architecture (the global scheduling agent splits the task by domain, and the domain expert agents execute it in parallel) or a decentralized collaboration architecture (each agent collaborates autonomously based on a publish-subscribe model and reaches consensus through a voting mechanism), both of which can complete the decomposition and execution of disaster simulation tasks.

[0082] (4) Alternatives to verification and reprogramming mechanisms The verification agent function can be replaced by a rule verification module (encoding compliance rules into logical expressions and judging whether the deduction result satisfies all constraints one by one). The replanning mechanism can be replaced by a backoff strategy based on heuristic rules or a decision correction mechanism based on historical cases, both of which can achieve self-healing recovery when the deduction fails.

[0083] (5) Alternatives to dynamic memory pruning The pruning method based on triplet summary extraction can be replaced by a sliding window pruning algorithm based on information freshness and relevance scoring, or a hierarchical summary compression method based on large language models. Both methods can achieve context window protection and token loss reduction in multi-round retry scenarios.

[0084] (6) Replacement of system deployment and output methods The system can be deployed in a standalone computing environment, a cloud computing platform, or an edge computing node. Functional modules can be deployed in a centralized or distributed manner as needed. In addition to structured reports, the system output can also be replaced with voice broadcasting, visual large-screen displays, or standardized API data interface output.

[0085] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An earthquake disaster emergency simulation system based on multi-agent and knowledge graph, characterized in that, It includes a data access and preprocessing module, a knowledge graph storage module, a multi-agent collaborative reasoning module, and a business output module; The data access and preprocessing module is used to access multi-source heterogeneous earthquake disaster information and complete unified formatting processing; The knowledge graph storage module includes a graph database and a vector database cluster, which are used to store an earthquake emergency knowledge graph composed of a five-level hierarchical topology task tree, an entity vector index, a standard operating procedure rule base, and a historical experience case base. The multi-agent collaborative reasoning module includes a master agent, an execution agent network, a verification agent, and a replanning agent. It has a built-in joint entity disambiguation unit and a dynamic fault-tolerant replanning unit. It completes disaster simulation and task assignment decision-making through a closed-loop process of "planning-execution-verification-replanning". The business output module is used to integrate the full chain of evidence simulation results and generate a structured emergency simulation report and a standardized resource assignment work order.

2. The earthquake disaster emergency simulation system based on multi-agent and knowledge graph as described in claim 1, characterized in that, The five-level hierarchical task tree consists of, from top to bottom, the emergency process root node, the task domain node, the task category node, the standard task code node, and the responsible department / professional resource node; each node contains a unique node identifier, node name, node type, and legal basis clause number attributes.

3. The earthquake disaster emergency simulation system based on multi-agent and knowledge graph as described in claim 1, characterized in that, The joint entity disambiguation unit adopts a three-stage mechanism of "string recall + vector rearrangement + NIL detection": The first phase involves full matching, prefix matching, suffix matching, and pinyin matching based on the standard entity list to recall the candidate entity set. The second stage calculates the cosine similarity between the query vector and the candidate entity vector using a text embedding model to complete the semantic ranking of candidate entities. The third stage performs boundary detection using dual similarity thresholds. When the highest similarity is lower than the lower threshold, the NIL signal is triggered, the system suppresses fuzzy generalization generation, and switches to external information verification.

4. The earthquake disaster emergency simulation system based on multi-agent and knowledge graph as described in claim 1, characterized in that, The master control agent is used to extract key projection parameters such as magnitude, epicenter, and population density from standardized disaster information, decompose the projection task into a sequence of sub-tasks that can be executed in parallel or serially, and encapsulate them into structured planning instructions in JSON format and send them to the execution agent network.

5. The earthquake disaster emergency simulation system based on multi-agent and knowledge graph as described in claim 4, characterized in that, The execution agent network is used to assemble graph database query statements according to subtask types and perform multi-hop path retrieval along the five-level hierarchical topology task tree; the retrieval transmission logic is: specific magnitude and geographical conditions → trigger → response level → activation → responsible department → execution → standard task code → association → professional rescue resources and technical support nodes; the retrieval process records complete intermediate node information and encapsulates it into an evidence chain structure containing path nodes, path edges and complete path descriptions.

6. The earthquake disaster emergency simulation system based on multi-agent and knowledge graph as described in claim 1, characterized in that, The verification agent is used to perform double verification on the inference results: Sufficiency verification involves checking whether the four decision-making elements—response level determination, list of responsible departments, standard task codes, and evidence chain traceability information—are complete. Compliance verification compares key decisions such as response level and task assignment with standard operating procedure rules in the knowledge graph to verify their legal compliance and the authenticity of the provisions.

7. The earthquake disaster emergency simulation system based on multi-agent and knowledge graph as described in claim 6, characterized in that, The replanning agent is used to trigger dynamic replanning when verification fails. The strategy adjustment includes reducing the generation temperature of the large language model, expanding the range of hop counts for graph queries, and correcting external search keywords. The system sets a maximum retry threshold. When the threshold is reached, a degradation fault tolerance mechanism is triggered, a list of uncertain factors to be verified is output, and the decision-making power is transferred to humans.

8. The earthquake disaster emergency simulation system based on multi-agent and knowledge graph as described in claim 1, characterized in that, The multi-agent collaborative reasoning module also includes a token budget control unit, which calculates the context token consumption in real time through a word segmenter matched with the target large model; it uses Text-ChunkBM25 and an embedded joint scoring machine to prune low information density content, performs entity-level triplet summarization pruning on multi-turn dialogues, and controls the context window size and token consumption.

9. The earthquake disaster emergency simulation system based on multi-agent and knowledge graph as described in claim 1, characterized in that, When the information within the graph is insufficient, the execution agent network calls a compliant external retrieval engine to obtain supplementary information, which is then fused into the evidence chain after webpage cleaning and structured parsing. In cross-regional and cross-departmental collaborative scenarios, the joint entity disambiguation unit is called to perform secondary verification on key nodes of the retrieval path.