A grassroots emergency command plan intelligent generation method based on a large language model and multi-source knowledge fusion
By constructing a knowledge foundation of rule graphs and experience graphs, and combining it with a large language model, the problems of static template rigidity and low response efficiency in the grassroots emergency command plan system are solved. This enables dynamic adjustment and rapid response of emergency command plans, and generates emergency plans with clear legal basis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 浪潮智慧城市科技有限公司
- Filing Date
- 2026-01-13
- Publication Date
- 2026-06-05
AI Technical Summary
The existing grassroots emergency command and control plan system suffers from problems such as rigid static templates, difficulty in passing on experience, defects in the application of large models, and low response efficiency. It is unable to dynamically adjust, quickly retrieve historical cases, and generate emergency plans with clear legal basis.
By constructing a knowledge foundation driven by both rule graphs and experience graphs, and by integrating large language models with multi-source knowledge, a causal association index and a three-dimensional vector deconstruction method are established to generate traceable, assessable, and evolvable emergency command plans.
It enables dynamic adjustment of emergency command plans, rapid response, and clear legal basis, thereby improving emergency response efficiency and ensuring the flexibility and interpretability of the generated plans.
Smart Images
Figure CN122152784A_ABST
Abstract
Description
Technical Field
[0001] This invention discloses an intelligent generation method for grassroots emergency command plans based on the fusion of large language models and multi-source knowledge, which relates to the field of artificial intelligence technology. Background Technology
[0002] Currently, there are some main problems in the grassroots emergency command and control plan system: Static templates are rigid: Contingency plans rely on manually compiled electronic documents and cannot be dynamically adjusted according to real-time disaster situations. They lack flexible response mechanisms to complex disasters such as typhoons, urban flooding, and secondary accidents.
[0003] Difficulties in passing on experience: Historical case studies are stored in scattered text reports, failing to form structured knowledge. New commanders struggle to quickly retrieve similar cases, highlighting the reliance on personal experience.
[0004] Large model application defects: The general large language model has serious illusion problems, the generated plans lack legal basis and interpretability, and cannot be linked to local resource ledgers and historical data, resulting in insufficient practicality.
[0005] Response efficiency bottleneck: Manually reviewing contingency plans, comparing cases, and coordinating resources takes an average of more than 2 hours, which is difficult to meet the requirements of emergency response within the golden 4 hours. Summary of the Invention
[0006] This invention addresses the problems of existing technologies by providing a method for intelligent generation of grassroots emergency command plans based on the fusion of large language models and multi-source knowledge. By constructing a knowledge base driven by both rule graphs and experience graphs, the plan generation process can be made traceable, assessable, and evolvable.
[0007] The specific solution proposed in this invention is as follows: This invention provides a method for intelligent generation of grassroots emergency command plans based on the fusion of large language models and multi-source knowledge, including: Step 1: Build a data access layer to convert real-time data into a unified format through standardized interfaces: Real-time data is collected from multiple heterogeneous data streams and converted into a unified event description format. This unified event description format includes event code, event space, event credibility, event impact, event level, and event details. Historical grassroots emergency reports, response records, and expert commentary documents were collected, and after data cleaning and anonymization, they were tagged according to time, location, and disaster type to establish a case material library. Step 2: Construct a knowledge fusion layer and establish a causal relationship index: Establish an emergency response plan rules knowledge graph: construct a five-tuple of disaster type, risk level, response action, responsible entity, and resource requirements, and store static knowledge, including legal provisions, resource ledgers, and job responsibilities; A three-dimensional vector deconstruction method is used to establish a historical case library: key information from the cases is extracted, event feature vectors are extracted based on the key information, decision-making action chain vectors are formulated, and the effectiveness of the decision-making actions is evaluated using an effect evaluation matrix. Establish a causal relationship index, including an event index, an action chain index, and an effect index; Step 3: Build the LLM inference engine layer and fine-tune it using a large model to adapt it to the basic emergency response domain: An emergency intent parser is established to parse natural language commands into structured event frames. A dual-path parallel RAG engine was built, which simultaneously performs retrieval and outputs answers using both the emergency response plan rule knowledge graph and the case material library. A causal graph constraint decoder is established, using the causal association index as constraints to convert the search results into a directed causal graph. Citation tags are then inserted during the LLM decoding stage. The system generates a self-assessment of quality based on rule coverage and resource matching. It compares the effects of the newly generated solutions with those of the old cases, calculates the confidence score, and then proceeds to the next step based on the confidence score. Step 4: Build an intelligent interaction layer that provides natural language interfaces and visualized rendering of contingency plans. The results of manual handling are fed back into the knowledge fusion layer to achieve incremental updates of the case library and graph, forming a positive cycle of use-feedback-optimization.
[0008] Furthermore, in step 2 of the aforementioned intelligent generation method for grassroots emergency command plans based on the fusion of large language models and multi-source knowledge, a three-dimensional vector deconstruction method is used to establish a historical case library, including: The UIE model is used to uniformly process the basic event extraction task, extract event features, and identify entities including time, location, and disaster type. The sequence labeling model based on BIO annotation identifies subject-action-object triples, extracts decision actions, and sorts them by time. Effect evaluation extraction: Numerical indicators predicted by the training regression model, including casualties and property losses, are extracted for effect evaluation and the evaluation results are quantified. Vector encoding: One-Hot+Embedding encoding is used for event features to obtain event feature vectors, Transformer encoding is used for decision action chains to obtain decision action chain vectors, and normalized vectors are used for effect evaluation to form an effect evaluation matrix.
[0009] Furthermore, in step 2 of the aforementioned intelligent generation method for grassroots emergency command plans based on large language models and multi-source knowledge fusion, establishing a causal association index includes: index establishment and online recall. Create an event index: Use the inner product similarity search index structure to create an event index, with a default recall of 50 records; Create an action chain index: Use subject and operation keywords to create an action chain index, with a preset recall of 50 results; Create an effect index: Filter by range according to the direction of minimum expected loss to create an effect index. Merge sorting: Take the union of candidates in the three-way index, re-sort and output the Top 3 for reference by the LLM inference layer; Generate case causal subgraphs: Mine the causal relationships between decision actions and effects in the case material library, construct case causal subgraphs for similar case retrieval and decision path reuse.
[0010] Furthermore, in step 3 of the method for intelligent generation of grassroots emergency command plans based on the fusion of a large language model and multi-source knowledge, when simultaneously performing retrieval and outputting answers using the emergency plan rule knowledge graph and the case material library, the standard response process is obtained from the emergency plan rule knowledge graph, and the top 3 similar cases are obtained from the case material library. These are combined into a mixed context and sent to the large model for the large model to output the answer.
[0011] Furthermore, step 4 of the aforementioned method for intelligently generating grassroots emergency command plans based on the fusion of large language models and multi-source knowledge specifically includes: Through a natural language interface, it supports both spoken language and text input, automatically recognizes dialects, and provides digital spoken language expression capabilities. The contingency plan is visualized and rendered in the form of a flowchart, intuitively presenting the generated plan to the command personnel. Different colors are used to indicate the confidence level at each node. It provides knowledge tracing; when commanders click on a contingency plan task, they can drill down to the original entries in the emergency plan rules knowledge graph or case material library to view the original text of the regulations or details of the handling of similar historical events. Provide a manual review node, and after the contingency plan is output, set up the confirmation, adjustment, and rejection process.
[0012] This invention also provides an intelligent generation system for grassroots emergency command plans based on large language models and multi-source knowledge fusion, including a data access module, a knowledge fusion module, a reasoning module, and an interaction module. The data access module builds a data access layer, converting real-time data into a unified format through standardized interfaces. Real-time data is collected from multiple heterogeneous data streams and converted into a unified event description format. This unified event description format includes event code, event space, event credibility, event impact, event level, and event details. Historical grassroots emergency reports, response records, and expert commentary documents were collected, and after data cleaning and anonymization, they were tagged according to time, location, and disaster type to establish a case material library. The knowledge fusion module constructs a knowledge fusion layer and establishes a causal relationship index: Establish an emergency response plan rules knowledge graph: construct a five-tuple of disaster type, risk level, response action, responsible entity, and resource requirements, and store static knowledge, including legal provisions, resource ledgers, and job responsibilities; A three-dimensional vector deconstruction method is used to establish a historical case library: key information from the cases is extracted, event feature vectors are extracted based on the key information, decision-making action chain vectors are formulated, and the effectiveness of the decision-making actions is evaluated using an effect evaluation matrix. Establish a causal relationship index, including an event index, an action chain index, and an effect index; The inference module builds an LLM inference engine layer, using a large model for fine-tuning and adaptation to grassroots emergency response scenarios. An emergency intent parser is established to parse natural language commands into structured event frames. A dual-path parallel RAG engine was built, which simultaneously performs retrieval and outputs answers using both the emergency response plan rule knowledge graph and the case material library. A causal graph constraint decoder is established, using the causal association index as constraints to convert the search results into a directed causal graph. Citation tags are then inserted during the LLM decoding stage. The system generates a self-assessment of quality based on rule coverage and resource matching. It compares the effects of the newly generated solutions with those of the old cases, calculates the confidence score, and then proceeds to the next step based on the confidence score. The interactive module builds an intelligent interaction layer that provides natural language interfaces and visualized rendering of contingency plans. The results of manual handling are fed back into the knowledge fusion layer to achieve incremental updates of the case library and graph, forming a positive cycle of use-feedback-optimization.
[0013] Furthermore, the knowledge fusion module of the intelligent generation system for grassroots emergency command plans based on large language models and multi-source knowledge fusion adopts a three-dimensional vector deconstruction method to establish a historical case library, including: The UIE model is used to uniformly process the basic event extraction task, extract event features, and identify entities including time, location, and disaster type. The sequence labeling model based on BIO annotation identifies subject-action-object triples, extracts decision actions, and sorts them by time. Effect evaluation extraction: Numerical indicators predicted by the training regression model, including casualties and property losses, are extracted for effect evaluation and the evaluation results are quantified. Vector encoding: One-Hot+Embedding encoding is used for event features to obtain event feature vectors, Transformer encoding is used for decision action chains to obtain decision action chain vectors, and normalized vectors are used for effect evaluation to form an effect evaluation matrix.
[0014] Furthermore, the knowledge fusion module of the intelligent generation system for grassroots emergency command plans based on large language models and multi-source knowledge fusion establishes a causal association index, including: index establishment and online retrieval. Create an event index: Use the inner product similarity search index structure to create an event index, with a default recall of 50 records; Create an action chain index: Use subject and operation keywords to create an action chain index, with a preset recall of 50 results; Create an effect index: Filter by range according to the direction of minimum expected loss to create an effect index. Merge sorting: Take the union of candidates in the three-way index, re-sort and output the Top 3 for reference by the LLM inference layer; Generate case causal subgraphs: Mine the causal relationships between decision actions and effects in the case material library, construct case causal subgraphs for similar case retrieval and decision path reuse.
[0015] Furthermore, in the reasoning module of the grassroots emergency command plan intelligent generation system based on a large language model and multi-source knowledge fusion, when simultaneously performing retrieval and outputting answers using the emergency plan rule knowledge graph and the case material library, the system retrieves the standard response process from the emergency plan rule knowledge graph and the top 3 similar cases from the case material library, combining them into a mixed context and sending it to the large model for the large model to output the answer.
[0016] Furthermore, the interactive module of the grassroots emergency command plan intelligent generation system based on large language models and multi-source knowledge fusion supports language input and text input through a natural language interface, automatically recognizes dialects, and provides digital spoken language expression capabilities. The contingency plan is visualized and rendered in the form of a flowchart, intuitively presenting the generated plan to the command personnel. Different colors are used to indicate the confidence level at each node. It provides knowledge tracing; when commanders click on a contingency plan task, they can drill down to the original entries in the emergency plan rules knowledge graph or case material library to view the original text of the regulations or details of the handling of similar historical events. Provide a manual review node, and after the contingency plan is output, set up the confirmation, adjustment, and rejection process.
[0017] The advantages of this invention are: Based on the fusion of large language models and multi-source knowledge, this method generates grassroots emergency command plans, primarily targeting the grassroots emergency command plan system. It transforms historical cases into structured documents, links them with local resource ledgers and historical data, and combines the reasoning ability of large language models with the causal learning ability of historical cases to dynamically generate and adjust emergency command plans according to real-time disaster situations, ensuring that the generated plans have legal basis and flexible response mechanisms. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the emergency reasoning process of the method of the present invention.
[0019] Figure 2 This is a schematic diagram of the application architecture of the method of the present invention. Detailed Implementation
[0020] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention. Example
[0021] This invention provides a method for intelligent generation of grassroots emergency command plans based on the fusion of large language models and multi-source knowledge, including: Step 1: Build a data access layer to convert real-time data into a unified format through standardized interfaces: Step 11: Collect real-time data through multi-source heterogeneous data streams and convert the real-time data into a unified event description format. The unified event description format includes event code, event space, event credibility, event impact, event level, and event details, as shown in Table 1.
[0022] Table 1
[0023] Step 12: Collect 30 years of historical grassroots emergency reports, response records, and expert commentary documents. After data cleaning and anonymization, tagged them according to time, location, and disaster type to establish a case material library. Step 2: Construct a knowledge fusion layer and establish a causal relationship index: Step 21: Establish an emergency response plan rule knowledge graph: Construct a five-tuple of disaster type, risk level, response action, responsible entity, and resource requirements, and store static knowledge, including legal provisions, resource ledgers, and job responsibilities; Step 22: Employ a three-dimensional vector deconstruction method to establish a historical case library: extract key information from cases, extract event feature vectors based on the key information, formulate decision-making action chain vectors, and simultaneously use an effect evaluation matrix to assess the effectiveness of decision-making actions. Specifically, the event feature vector focuses on describing event characteristics, recording the event's occurrence time, location, disaster type, intensity level, and scope of impact; the decision-making action chain vector focuses on describing event handling actions, recording event actions, responsible parties, operational content, handling time, and prerequisite dependencies; and the effect evaluation matrix focuses on describing the event's post-event impact, recording casualties, property damage, public satisfaction, and whether secondary disasters occurred.
[0024] Three-dimensional vector deconstruction methods may specifically include: The UIE model is used to uniformly process the basic event extraction task, extract event features, and identify entities including time, location, and disaster type. The sequence labeling model based on BIO annotation identifies subject-action-object triples, extracts decision actions, and sorts them by time. Effect evaluation extraction: Numerical indicators predicted by the training regression model, including casualties and property losses, are extracted for effect evaluation and the evaluation results are quantified. Vector encoding: One-Hot+Embedding encoding is used for event features to obtain event feature vectors, Transformer encoding is used for decision action chains to obtain decision action chain vectors, and normalized vectors are used for effect evaluation to form an effect evaluation matrix.
[0025] Step 23: Establish a causal relationship index, including an event index, an action chain index, and an effect index.
[0026] The establishment of a causal association index includes: index creation and online recall. Create an event index: Use the inner product similarity search index structure to create an event index, with a default recall of 50 records; Create an action chain index: Use subject and operation keywords to create an action chain index, with a preset recall of 50 results; Create an effect index: Filter by range according to the direction of minimum expected loss to create an effect index. Merge sorting: Take the union of candidates in the three-way index, re-sort and output the Top 3 for reference by the LLM inference layer; Generate case causal subgraphs: Mine the causal relationships between decision actions and effects in the case material library, construct case causal subgraphs for similar case retrieval and decision path reuse.
[0027] Step 3: Build the LLM inference engine layer and fine-tune it using a large model to adapt it to the basic emergency response domain: Step 31: Establish an emergency intent parser to parse natural language commands into structured event frames. Step 32: Build a dual-path parallel RAG engine, and simultaneously perform retrieval and output of answers using the emergency plan rule knowledge graph and the case material library respectively. The retrieval obtains the standard response process from the emergency plan rule knowledge graph, and the retrieval obtains the top 3 similar cases from the case material library. These are combined into a mixed context and sent to the large model for the large model to output the answer.
[0028] Step 33: Establish a causal graph constraint decoder, using the causal association index as a constraint condition, and convert the search results into a directed causal graph such as nodes = actions / resources / risk points, edges = facilitating / inhibiting relationships. Insert citation tags in the LLM decoding stage to improve the consistency between the generated content and historical experience and rules, and ensure that each step of generation can be traced back to a specific rule or case source.
[0029] Step 34: Generate a quality self-assessment. Based on rule coverage and resource matching, compare the effects of the newly generated solution with those of the old case, calculate the confidence level, and execute the next action according to the confidence level score. A confidence level score range can be set. If the confidence level exceeds a certain value, it is marked as highly reliable and enters the intelligent interaction layer. If it is below a certain value, it triggers regeneration.
[0030] Step 4: Build an intelligent interaction layer that provides natural language interfaces and visualized rendering of contingency plans. The results of manual handling are fed back into the knowledge fusion layer to achieve incremental updates of the case library and graph, forming a positive cycle of use-feedback-optimization.
[0031] Specifically, it can include: Through a natural language interface, it supports both spoken language and text input, automatically recognizes dialects, and provides digital spoken language expression capabilities. The contingency plan is visualized and rendered in the form of a flowchart, intuitively presenting the generated plan to the command personnel. Different colors are used to indicate the confidence level at each node. It provides knowledge tracing; when commanders click on a contingency plan task, they can drill down to the original entries in the emergency plan rules knowledge graph or case material library to view the original text of the regulations or details of the handling of similar historical events. It provides a manual review node, and after the contingency plan is output, it sets up a process for confirmation, adjustment, and rejection. If the command personnel adjust the plan, the change is recorded and sent back to the case library, realizing incremental updates of the case library and the map, forming a positive cycle of "use-feedback-optimization".
[0032] The method of this invention integrates multi-source heterogeneous data by constructing a data access layer, converts it into a unified format through a standardized interface, builds a knowledge fusion layer to construct an emergency plan knowledge graph and a historical case knowledge base, establishes a causal association index, constructs an LLM inference engine layer to fine-tune and adapt to the grassroots emergency field, and finally provides a natural language interface and plan visualization rendering by building an intelligent interaction layer, sets up manual review nodes, and inputs the results of manual handling back into the knowledge fusion layer to update and improve the case knowledge base.
[0033] Example 2 This invention also provides an intelligent generation system for grassroots emergency command plans based on large language models and multi-source knowledge fusion, including a data access module, a knowledge fusion module, a reasoning module, and an interaction module. The data access module builds a data access layer, converting real-time data into a unified format through standardized interfaces. Real-time data is collected from multiple heterogeneous data streams and converted into a unified event description format. This unified event description format includes event code, event space, event credibility, event impact, event level, and event details. Historical grassroots emergency reports, response records, and expert commentary documents were collected, and after data cleaning and anonymization, they were tagged according to time, location, and disaster type to establish a case material library. The knowledge fusion module constructs a knowledge fusion layer and establishes a causal relationship index: Establish an emergency response plan rules knowledge graph: construct a five-tuple of disaster type, risk level, response action, responsible entity, and resource requirements, and store static knowledge, including legal provisions, resource ledgers, and job responsibilities; A three-dimensional vector deconstruction method is used to establish a historical case library: key information from the cases is extracted, event feature vectors are extracted based on the key information, decision-making action chain vectors are formulated, and the effectiveness of the decision-making actions is evaluated using an effect evaluation matrix. Establish a causal relationship index, including an event index, an action chain index, and an effect index; The inference module builds an LLM inference engine layer, using a large model for fine-tuning and adaptation to grassroots emergency response scenarios. An emergency intent parser is established to parse natural language commands into structured event frames. A dual-path parallel RAG engine was built, which simultaneously performs retrieval and outputs answers using both the emergency response plan rule knowledge graph and the case material library. A causal graph constraint decoder is established, using the causal association index as constraints to convert the search results into a directed causal graph. Citation tags are then inserted during the LLM decoding stage. The system generates a self-assessment of quality based on rule coverage and resource matching. It compares the effects of the newly generated solutions with those of the old cases, calculates the confidence score, and then proceeds to the next step based on the confidence score. The interactive module builds an intelligent interaction layer that provides natural language interfaces and visualized rendering of contingency plans. The results of manual handling are fed back into the knowledge fusion layer to achieve incremental updates of the case library and graph, forming a positive cycle of use-feedback-optimization.
[0034] The information interaction and execution process between the modules in the above system are based on the same concept as the method embodiment of the present invention, and the specific details can be found in the description in the method embodiment of the present invention, and will not be repeated here.
[0035] Similarly, the system of this invention generates grassroots emergency command plans based on the fusion of large language models and multi-source knowledge. It is mainly aimed at the grassroots emergency command plan system. It forms structured documents from historical cases, links them with local resource ledgers, historical data, etc., and combines the reasoning ability of large language models with the causal learning ability of historical cases to dynamically generate and adjust emergency command plans according to the real-time disaster situation, ensuring that the generated plans have legal basis and flexible response mechanisms.
[0036] It should be noted that not all steps and modules in the above processes and system structures are mandatory; some steps or modules can be omitted as needed. The execution order of each step is not fixed and can be adjusted as required. The system structures described in the above embodiments can be physical or logical structures. That is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or they may be jointly implemented by certain components in multiple independent devices.
[0037] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.
Claims
1. A method for intelligent generation of grassroots emergency command plans based on the fusion of large language models and multi-source knowledge, characterized by: include: Step 1: Build a data access layer to convert real-time data into a unified format through standardized interfaces: Real-time data is collected from multiple heterogeneous data streams and converted into a unified event description format. This unified event description format includes event code, event space, event credibility, event impact, event level, and event details. Historical grassroots emergency reports, response records, and expert commentary documents were collected, and after data cleaning and anonymization, they were tagged according to time, location, and disaster type to establish a case material library. Step 2: Construct a knowledge fusion layer and establish a causal relationship index: Establish an emergency response plan rules knowledge graph: construct a five-tuple of disaster type, risk level, response action, responsible entity, and resource requirements, and store static knowledge, including legal provisions, resource ledgers, and job responsibilities; A three-dimensional vector deconstruction method is used to establish a historical case library: key information from the cases is extracted, event feature vectors are extracted based on the key information, decision-making action chain vectors are formulated, and the effectiveness of the decision-making actions is evaluated using an effect evaluation matrix. Establish a causal relationship index, including an event index, an action chain index, and an effect index; Step 3: Build the LLM inference engine layer and fine-tune it using a large model to adapt it to the basic emergency response domain: An emergency intent parser is established to parse natural language commands into structured event frames. A dual-path parallel RAG engine was built, which simultaneously performs retrieval and outputs answers using both the emergency response plan rule knowledge graph and the case material library. A causal graph constraint decoder is established, using the causal association index as constraints to convert the search results into a directed causal graph. Citation tags are then inserted during the LLM decoding stage. The system generates a self-assessment of quality based on rule coverage and resource matching. It compares the effects of the newly generated solutions with those of the old cases, calculates the confidence score, and then proceeds to the next step based on the confidence score. Step 4: Build an intelligent interaction layer that provides natural language interfaces and visualized rendering of contingency plans. The results of manual handling are fed back into the knowledge fusion layer to achieve incremental updates of the case library and graph, forming a positive cycle of use-feedback-optimization.
2. The method for intelligent generation of grassroots emergency command plans based on the fusion of large language models and multi-source knowledge as described in claim 1. Its characteristic is that step 2 employs a three-dimensional vector deconstruction method to establish a historical case library, including: The UIE model is used to uniformly process the grassroots event extraction task, extract event features, and identify entities including time, location, and disaster type. The sequence labeling model based on BIO annotation identifies subject-action-object triples, extracts decision actions, and sorts them by time. Effect evaluation extraction: Numerical indicators predicted by the training regression model, including casualties and property losses, are extracted for effect evaluation and the evaluation results are quantified. Vector encoding: One-Hot+Embedding encoding is used for event features to obtain event feature vectors, Transformer encoding is used for decision action chains to obtain decision action chain vectors, and normalized vectors are used for effect evaluation to form an effect evaluation matrix.
3. A method for intelligent generation of grassroots emergency command plans based on the fusion of large language models and multi-source knowledge, as described in claim 1 or 2, characterized in that establishing a causal association index in step 2 includes: Index building and online recall: Create an event index: Use the inner product similarity search index structure to create an event index, with a default recall of 50 records; Create an action chain index: Use subject and operation keywords to create an action chain index, with a preset recall of 50 results; Create an effect index: Filter by range according to the direction of minimum expected loss to create an effect index. Merge sorting: Take the union of candidates in the three-way index, re-sort and output the Top 3 for reference by the LLM inference layer; Generate case causal subgraphs: Mine the causal relationships between decision actions and effects in the case material library, construct case causal subgraphs for similar case retrieval and decision path reuse.
4. The intelligent generation method for grassroots emergency command plans based on the fusion of large language models and multi-source knowledge as described in claim 1, characterized in that: In step 3, when the emergency response plan rule knowledge graph and the case material library are used to simultaneously perform the retrieval and output the answer, the standard response process is obtained from the emergency response plan rule knowledge graph, and the top 3 similar cases are obtained from the case material library. These are combined into a mixed context and sent to the large model for the large model to output the answer.
5. The intelligent generation method for grassroots emergency command plans based on the fusion of large language models and multi-source knowledge as described in claim 1, characterized in that: Step 4 specifically includes: Through a natural language interface, it supports both spoken language and text input, automatically recognizes dialects, and provides digital spoken language expression capabilities. The contingency plan is visualized and rendered in the form of a flowchart, intuitively presenting the generated plan to the command personnel. Different colors are used to indicate the confidence level at each node. It provides knowledge tracing; when commanders click on a contingency plan task, they can drill down to the original entries in the emergency plan rules knowledge graph or case material library to view the original text of the regulations or details of the handling of similar historical events. Provide a manual review node, and after the contingency plan is output, set up the confirmation, adjustment, and rejection process.
6. A grassroots emergency command plan intelligent generation system based on the fusion of large language models and multi-source knowledge, characterized by: It includes a data access module, a knowledge fusion module, a reasoning module, and an interaction module. The data access module builds a data access layer, converting real-time data into a unified format through standardized interfaces. Real-time data is collected from multiple heterogeneous data streams and converted into a unified event description format. This unified event description format includes event code, event space, event credibility, event impact, event level, and event details. Historical grassroots emergency reports, response records, and expert commentary documents were collected, and after data cleaning and anonymization, they were tagged according to time, location, and disaster type to establish a case material library. The knowledge fusion module constructs a knowledge fusion layer and establishes a causal relationship index: Establish an emergency response plan rules knowledge graph: construct a five-tuple of disaster type, risk level, response action, responsible entity, and resource requirements, and store static knowledge, including legal provisions, resource ledgers, and job responsibilities; A three-dimensional vector deconstruction method is used to establish a historical case library: key information from the cases is extracted, event feature vectors are extracted based on the key information, decision-making action chain vectors are formulated, and the effectiveness of the decision-making actions is evaluated using an effect evaluation matrix. Establish a causal relationship index, including an event index, an action chain index, and an effect index; The inference module builds an LLM inference engine layer, using a large model for fine-tuning and adaptation to grassroots emergency response scenarios. An emergency intent parser is established to parse natural language commands into structured event frames. A dual-path parallel RAG engine was built, which simultaneously performs retrieval and outputs answers using both the emergency response plan rule knowledge graph and the case material library. A causal graph constraint decoder is established, using the causal association index as constraints to convert the search results into a directed causal graph. Citation tags are then inserted during the LLM decoding stage. The system generates a self-assessment of quality based on rule coverage and resource matching. It compares the effects of the newly generated solutions with those of the old cases, calculates the confidence score, and then proceeds to the next step based on the confidence score. The interactive module builds an intelligent interaction layer that provides natural language interfaces and visualized rendering of contingency plans. The results of manual handling are fed back into the knowledge fusion layer to achieve incremental updates of the case library and graph, forming a positive cycle of use-feedback-optimization.
7. The intelligent generation system for grassroots emergency command plans based on the fusion of large language models and multi-source knowledge as described in claim 6, characterized in that knowledge... The fusion module employs a three-dimensional vector deconstruction method to establish a historical case library, including: The UIE model is used to uniformly process the grassroots event extraction task, extract event features, and identify entities including time, location, and disaster type. The sequence labeling model based on BIO annotation identifies subject-action-object triples, extracts decision actions, and sorts them by time. Effect evaluation extraction: Numerical indicators predicted by the training regression model, including casualties and property losses, are extracted for effect evaluation and the evaluation results are quantified. Vector encoding: One-Hot+Embedding encoding is used for event features to obtain event feature vectors, Transformer encoding is used for decision action chains to obtain decision action chain vectors, and normalized vectors are used for effect evaluation to form an effect evaluation matrix.
8. The intelligent generation system for grassroots emergency command plans based on the fusion of large language models and multi-source knowledge as described in claim 6, characterized in that: The knowledge fusion module establishes a causal association index, including: index creation and online recall. Create an event index: Use the inner product similarity search index structure to create an event index, with a default recall of 50 records; Create an action chain index: Use subject and operation keywords to create an action chain index, with a preset recall of 50 results; Create an effect index: Filter by range according to the direction of minimum expected loss to create an effect index. Merge sorting: Take the union of candidates in the three-way index, re-sort and output the Top 3 for reference by the LLM inference layer; Generate case causal subgraphs: Mine the causal relationships between decision actions and effects in the case material library, construct case causal subgraphs for similar case retrieval and decision path reuse.
9. The intelligent generation system for grassroots emergency command plans based on the fusion of large language models and multi-source knowledge as described in claim 6, characterized in that: When the reasoning module simultaneously performs retrieval and outputs the answer using the emergency plan rule knowledge graph and the case material library, the retrieval obtains the standard response process from the emergency plan rule knowledge graph and the retrieval obtains the top 3 similar cases from the case material library, combining them into a mixed context and sending it to the large model for the large model to output the answer.
10. The intelligent generation system for grassroots emergency command plans based on the fusion of large language models and multi-source knowledge as described in claim 1, characterized in that: The interaction module supports both spoken and text input via a natural language interface, automatically recognizes dialects, and provides digital spoken language expression capabilities. The contingency plan is visualized and rendered in the form of a flowchart, intuitively presenting the generated plan to the command personnel. Different colors are used to indicate the confidence level at each node. It provides knowledge tracing; when commanders click on a contingency plan task, they can drill down to the original entries in the emergency plan rules knowledge graph or case material library to view the original text of the regulations or details of the handling of similar historical events. Provide a manual review node, and after the contingency plan is output, set up the confirmation, adjustment, and rejection process.