CN-121600667-B - Disaster early warning system and method based on large language model multi-agent

CN121600667BCN 121600667 BCN121600667 BCN 121600667BCN-121600667-B

Abstract

The invention discloses a disaster early warning system and method based on a large language model multi-agent, wherein the system comprises a situation awareness agent, an execution agent, a domain expert agent group, a system sharing state module and a tool library module, the situation awareness agent is used for monitoring data and triggering early warning workflow signals, the execution agent is used for decomposing and routing tasks, the domain expert agent group comprises an emergency management agent and a plurality of expert agents in different domains, the domain expert agent group is used for executing professional tasks in each domain and writing task analysis results in each domain back to a unified state, the situation awareness agent, the execution agent and the domain expert agent group call corresponding tools in the tool library module and perform information interaction through the system sharing state module, and early warning information is automatically generated by the emergency management agent according to a preset template. The invention can improve the accuracy and consistency of early warning and judgment and shorten the time from risk discovery to forming an executable action scheme.

Inventors

WANG NAIYU
WANG JIE
BIAN XUECHENG
LU WEIMING
SHEN YONGLIANG
WANG YINGJUN

Assignees

浙江大学

Dates

Publication Date: 20260508
Application Date: 20260122

Claims (10)

1. A disaster early warning system based on a large language model multi-agent is characterized by comprising: The situation awareness intelligent agent is at least used for continuously monitoring multisource observation and forecast data of a target area, updating and anomaly identification are carried out on the environment state, and an early warning workflow signal is generated when a preset trigger condition is met; Executing an agent, at least for receiving a situation awareness agent or an external task request, decomposing and routing a task, and performing cross-agent discussion and negotiation based on a task result to form a consistent conclusion; The domain expert intelligent agent group comprises an emergency management intelligent agent and a plurality of expert intelligent agents in different domains, and is at least used for receiving and executing the tasks allocated by the execution intelligent agents and writing back the task analysis results in each domain to a unified state; The system sharing state module is used for providing a system sharing state, a unified information interaction hub and a semantic space for each intelligent agent; The tool library module is used for providing required tools for each intelligent body; The intelligent tool library module is used for calling corresponding tools in the tool library module by the situation awareness intelligent agent, the execution intelligent agent and the domain expert intelligent agent group, carrying out information interaction through the system sharing state module, and finally automatically generating early warning information by the emergency management intelligent agent according to a preset template.
2. The disaster early warning system based on large language model multi-agent according to claim 1, wherein expert agents in different fields at least include any one or more of the following agents: meteorological agent, living building agent, electric power agent, natural resource agent, water conservancy agent, maritime agent, traffic agent, marine economy agent, and agricultural agent.
3. The disaster early warning system based on large language model multi-agent according to claim 1, wherein situation awareness agent, execution agent and domain expert agent group are constructed based on the same large language model, and the inference framework MRKL-ReAct is uniformly adopted, and the inference process comprises knowledge activation stage, tool selection stage and result disbelief stage.
4. The disaster early warning system based on large language model multi-agent according to claim 1, wherein information flow between situation awareness agent, execution agent and domain expert agent group is realized by the following routing modes: sequentially routing, namely automatically triggering analysis of expert intelligent agents in the downstream field by taking the output of the expert intelligent agents in the former field as a triggering condition; the conditional routing is to determine whether to automatically activate the corresponding workflow according to whether the risk index exceeds a threshold value; and (3) parallel routing, namely carrying out parallel analysis on different disaster dimensions by the domain expert intelligent agent group, and synchronizing analysis results in a system sharing state module.
5. The disaster early warning system based on large language model multi-agent according to claim 1, wherein the tool library module used for constructing situation awareness agent, execution agent and domain expert agent group comprises the following tools: the universal tool and the data access tool at least comprise data cleaning, quality control, anomaly detection, threshold monitoring, spatial interpolation and statistical analysis tools; The task scheduling and conflict resolution tool at least comprises a task resolution strategy, a routing and matching algorithm and a multi-round negotiation and conflict resolution mechanism tool, and is used for converting a natural language task into an executable subtask sequence; And the domain-specific analysis tool is used for configuring a corresponding professional model and API, and is called by a domain expert intelligent agent group through a unified tool interface.
6. The disaster early warning system based on the large language model multi-agent according to claim 1, wherein a plurality of domain expert agents contained in the domain expert agent group and tools used by the domain expert agents are configured in a modularized mode, a system sharing state adopts a version graph structure, a current value is saved, a history track is recorded, and the agents are accessed and updated in a permission range to support collaborative reasoning on the premise of keeping role division.
7. The disaster early warning method based on the large language model multi-agent is characterized by being applied to the disaster early warning system based on the large language model multi-agent as claimed in any one of claims 1 to 6, and comprises the following steps: The environmental state is updated and abnormal identification is carried out through multisource observation and forecast data of a situation awareness agent continuously monitoring target area, and whether preset early warning triggering conditions are met or not is judged; When the early warning triggering condition is met or an external task request is received, the execution agent analyzes the task and decomposes the task into a plurality of subtasks; according to semantic matching between subtask descriptions and capability descriptions of expert agents in each field, the execution agents automatically distribute subtasks to the matched expert agents for execution; Based on role prompt and information of a system sharing state module, the expert intelligent agent in each field distributed to the task calls corresponding tools in the tool library module to execute subtasks in parallel, and then writes back task analysis results in each field to a unified state; the executing agent calls expert agents in all fields to conduct cross-agent discussion and negotiation on preset task results to form a consistent conclusion, and based on the consistent conclusion, early warning information is automatically generated by the emergency management agent according to a preset template.
8. The disaster early warning method based on large language model multi-agent according to claim 7, wherein the tasks to be completed by the disaster early warning system based on large language model multi-agent at least comprise evaluating the warning level, analyzing the disaster risk, generating disaster report, and making action plan.
9. The disaster early warning method based on large language model multi-agent according to claim 8, wherein executing agent calls each domain expert agent to perform cross agent discussion and negotiation on preset task results to form a consistent conclusion, specifically comprising: Executing an agent to call expert agents in each field to perform cross-agent discussion and negotiation on preset task results, wherein the preset task results at least comprise generation of disaster reports and establishment of action plans, and if inconsistent description or logic conflict is found, reinitiating targeted subtasks or request supplementary explanation to form a multi-round negotiation mechanism of 'generation-check-correction', so that final conclusions are consistent in semantic level, namely consistent conclusions are formed, and a sharing state is written back.
10. The disaster early warning method based on large language model multi-agent according to any one of claims 7 to 9, further comprising the steps of, after automatically generating early warning information by emergency management agent according to a preset template based on a coincidence conclusion: And submitting the generated early warning information to an emergency command center for manual auditing and formal release, and updating related decision records in a system sharing state module according to auditing feedback.

Description

Disaster early warning system and method based on large language model multi-agent Technical Field The invention relates to the technical field of artificial intelligence and natural disaster emergency management, in particular to a disaster early warning system and method based on a large language model multi-agent. Background At present, in early warning and emergency management of natural disasters, especially typhoons, storm and other meteorological disasters, a multisource information integration and threshold warning technology based on a numerical forecasting model, a monitoring station network and a geographic information system is commonly adopted. The method is characterized in that departments of weather, water conservancy, natural resources and the like respectively establish information systems of the industry, rainfall, wind field, water level, landslide risk and the like in a future period are predicted through a numerical weather forecast model, a flood evolution model, a geological disaster susceptibility model and the like, then corresponding early warning grades are triggered according to a preset index system and grading standards (such as rainfall, wind power and water level exceed a certain threshold), and early warning information is pushed through a business system in the departments or public-oriented release channels. In order to improve the cooperative efficiency, comprehensive disaster early warning and emergency command platforms are built in some areas, and multi-department data such as weather, water conservancy, emergency, natural resources and the like are tried to be gathered on the same platform, so that centralized display of early warning information, consultation and judgment and joint consultation of consultation are realized. The platform integrates data of business systems of departments through interfaces to provide unified map display, plan management, task circulation and consultation record functions, and part of the systems also introduce a rule engine or a workflow engine to carry out flow management on key links such as early warning release, consultation start, plan start and the like. In addition, with the development of artificial intelligence technology, the prior art has also begun to explore the use of machine learning models or knowledge graph technologies to evaluate disaster risks and intelligently interpret early warning information, such as training a loss prediction model based on historical disaster conditions and meteorological data, or constructing a disaster scene knowledge base to assist in research and judgment. There are also disclosed techniques that propose using conversational artificial intelligence to provide emergency personnel with auxiliary functions such as policy inquiry, plan inquiry, and manuscript writing. However, most of these technologies still use a single model or a single task as a core, and mainly solve the intelligentized problem of a certain link, and a complete flow collaborative framework penetrating through 'monitoring-predicting-evaluating-deciding-acting' is not formed yet. In objective terms, the above prior art has the following problems and disadvantages in practical application: (1) The departments are split and the semantics are not uniform, namely, although the comprehensive platform integrates the data of multiple departments on an interface, the bottom layer is still the simple splicing of a plurality of independent systems, the index system, the professional terms and the data structure used by each department are large in difference, and the unified semantic representation and the sharing state model are lacking, so that automatic reasoning and collaborative analysis are difficult to be carried out in the unified semantic space among the cross departments, and the manual consultation is still seriously relied on. (2) The existing early warning process is usually circulated according to a fixed sequence, and multi-stage experts and management staff are required to check and issue step by step. Even if part of the system introduces a workflow engine, the existing manual flow is mainly electronic and formalized, and the analysis path and task division are difficult to flexibly adjust according to the disaster development situation and the temporary decision requirement, so that the time from risk identification to forming an executable action scheme is long, and the gold early warning time is not easy to grasp. (3) The prior art mainly adopts a grading early warning method taking a threshold value as a core, and the early warning grade is judged mainly based on single or few physical indexes such as rainfall, wind speed, water level and the like, so that the amplification or weakening effects of disaster bearing body exposure, vulnerability and infrastructure difference on disaster influence in different areas are difficult to fully consider, and the composite and cascading disaster scenes such as typhoons,