Method and device for generating coping scheme description information, storage medium and electronic equipment

Abstract

The embodiment of the invention provides a coping scheme description information generation method and device, a storage medium and electronic equipment, and the method comprises the steps: obtaining real-time fire behavior data and a coping scheme query request, carrying out the fusion retrieval according to the real-time fire behavior data and the coping scheme query request, obtaining a plurality of target knowledge nodes, and storing the target knowledge nodes in a database; and determining a knowledge node text associated with the plurality of target knowledge nodes according to a pre-constructed knowledge graph and the plurality of target knowledge nodes, and generating coping scheme description information according to the knowledge node text and the real-time fire data. According to the forest fire emergency decision-making method and device, the response scheme is dynamically generated in combination with the real-time fire data and the enhanced retrieval generation technology, the problems of limited decision-making intelligence level and poor dynamic adaptability in the forest fire scene in the related technology are solved, and the real-time performance and flexibility of the forest fire emergency decision-making are improved.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and emergency decision-making, and more specifically, to a method, apparatus, storage medium, and electronic device for generating response plan description information. Background Technology

[0002] In current forest fire emergency command systems, emergency plans are often stored in a scattered, document-based manner, lacking structured processing and semantic association. For example, although the Yantai City Forest Fire Prevention Monitoring and Command System integrates multi-source data such as satellite remote sensing, UAV thermal imaging, ground sensors, and manual reporting, its plan library is still mainly composed of unstructured text, making it difficult to quickly and accurately match and retrieve relevant plan content in emergency situations.

[0003] In terms of decision support, the following methods are mainly adopted: 1. Rule-based decision engines: These systems rely on fixed rules and processes formulated by experts, triggering corresponding decisions through preset conditions. For example, when a fire point is detected, the system automatically matches a preset fire-fighting plan based on the size and location of the fire point. 2. Case-Based Reasoning (CBR) technology: Retrieving similar cases from a historical case database for decision reference. For example, the Daxinganling region has built a "Cloud-based Forest Fire Prevention Visual Management System," achieving communication support capabilities of "full coverage, all-time availability, and full controllability," but its case database has low retrieval efficiency and is difficult to dynamically correlate with real-time fire data. 3. Dynamic spread models: Based on the location of the fire point and real-time meteorological data, predicting the future spread direction, area, speed, and intensity of the fire. For example, the FARSITE model simulates fire spread using the Rothermel fire line propagation rate equation and Huygens wave theory, but its integration with the decision support system is not high, lacking real-time data-driven intelligent decision-making capabilities.

[0004] Therefore, the current forest fire emergency command system still has the following problems: 1. The knowledge base of emergency plans is fragmented, lacking structured processing and semantic association, making it difficult to quickly and accurately match and call relevant plan content in emergency situations. 2. The level of decision-making intelligence is limited, relying on human experience or simple rule engines, resulting in limited accuracy and difficulty in dynamically associating with real-time fire data. 3. Poor dynamic adaptability, making it difficult to dynamically adjust decision-making strategies based on factors such as fire intensity changes, weather conditions, and terrain features, failing to meet the real-time and flexibility requirements of forest fire emergency decision-making.

[0005] There is no good solution to the above problems in the relevant technologies. Summary of the Invention

[0006] This application provides a method, apparatus, storage medium, and electronic device for generating response scheme description information, in order to at least solve the problems of limited intelligence level and poor dynamic adaptability in forest fire scenarios in related technologies.

[0007] According to one aspect of the embodiments of this application, a method for generating response plan description information is provided, comprising: acquiring real-time fire data and a response plan query request, wherein the real-time fire data is used to represent event information of an ongoing fire event, and the response plan query request is used to request a response plan for the fire event; performing a fusion retrieval based on the real-time fire data and the response plan query request to obtain multiple target knowledge nodes; determining knowledge node text associated with the multiple target knowledge nodes based on a pre-constructed knowledge graph and the multiple target knowledge nodes; and generating response plan description information based on the knowledge node text and the real-time fire data, wherein the response plan description information is used to represent a response plan for the fire event.

[0008] According to another aspect of the embodiments of this application, an apparatus for generating response plan description information is also provided, comprising: an acquisition module, configured to acquire real-time fire data and a response plan query request, wherein the real-time fire data is used to represent event information of an ongoing fire event, and the response plan query request is used to request a response plan for the fire event; a fusion retrieval module, configured to perform a fusion retrieval based on the real-time fire data and the emergency decision query request to obtain multiple target knowledge nodes; a reasoning module, configured to determine knowledge node text associated with the multiple target knowledge nodes based on a pre-constructed knowledge graph and the multiple target knowledge nodes; and a generation module, configured to generate response plan description information based on the knowledge node text and the real-time fire data, wherein the response plan description information is used to represent a response plan for the fire event.

[0009] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer program, wherein the computer program is configured to perform the steps in any of the above method embodiments when executed by a processor.

[0010] According to another aspect of the embodiments of this application, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform the steps in any of the method embodiments described above.

[0011] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to perform the steps of any of the above method embodiments through the computer program.

[0012] This application obtains real-time fire data and response plan query requests, performs a fusion retrieval based on the real-time fire data and response plan query requests to obtain multiple target knowledge nodes, determines the knowledge node text associated with the multiple target knowledge nodes based on a pre-constructed knowledge graph and the multiple target knowledge nodes, and generates response plan description information based on the knowledge node text and real-time fire data. This application can dynamically generate response plans by combining real-time fire data and enhanced retrieval generation technology, solving the problems of limited intelligence and poor dynamic adaptability in forest fire scenarios in related technologies, and improving the real-time nature and flexibility of forest fire emergency decision-making. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the structure of the forest fire emergency command system in the embodiments of this application;

[0014] Figure 2 This is a flowchart illustrating a method for generating solution description information according to an embodiment of this application;

[0015] Figure 3 This is a schematic diagram of the knowledge graph construction process in one embodiment of this application;

[0016] Figure 4 This is a schematic diagram of the knowledge base construction process in an exemplary embodiment of this application;

[0017] Figure 5 This is a schematic diagram of the fusion retrieval process in one embodiment of this application;

[0018] Figure 6 This is a schematic diagram of the fusion retrieval process in an exemplary embodiment of this application;

[0019] Figure 7 This is a schematic diagram of the decision generation process in one embodiment of this application;

[0020] Figure 8 This is a schematic diagram of the overall workflow of a forest fire emergency command system in an exemplary embodiment of this application;

[0021] Figure 9 This is a schematic diagram of the system interaction flow in one embodiment of this application;

[0022] Figure 10 This is a structural block diagram of a device for generating solution description information in one embodiment of this application. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] The embodiments in this application can be applied to the fields of artificial intelligence and emergency decision-making, especially in forest fire emergency decision-making scenarios.

[0026] The emergency decision-making technology in related technologies mainly suffers from the following problems:

[0027] 1. Fragmented Emergency Response Plan Knowledge Base: In existing systems, forest fire emergency response plans are mostly stored in document form, lacking structured processing and semantic association. This makes it difficult to quickly and accurately match and retrieve relevant plan content in emergency situations. For example, although some cities' forest fire emergency response plans include seven parts such as general principles, organizational command system, early warning and monitoring, and emergency response, there is a lack of effective semantic association and retrieval mechanisms between these parts.

[0028] 2. Limited level of intelligent decision-making: Traditional decision support systems often rely on human experience or simple rule engines, failing to effectively integrate multi-source data (such as terrain, weather, historical cases, etc.) for comprehensive analysis and intelligent decision-making. For example, while case-based reasoning systems can retrieve similar cases from historical case databases, the retrieval process is overly reliant on human experience, resulting in limited accuracy and difficulty in dynamically correlating with real-time fire data.

[0029] 3. Poor dynamic adaptability: Traditional systems are unable to dynamically adjust decision-making strategies based on factors such as changes in fire intensity, weather conditions, and terrain features, and cannot meet the real-time and flexibility requirements of forest fire emergency decision-making.

[0030] 4. Poor decision interpretability: The decision recommendations generated by the RAG system based on pure vector retrieval lack logical chains and evidence traceability, making it difficult for commanders to understand the reasoning process behind the decision, which affects the credibility and execution efficiency of the decision.

[0031] 5. Limited search technology: Existing systems mostly use a single search technology (such as pure vector or pure keyword), which cannot meet the needs of semantic understanding and precise matching.

[0032] This application aims to solve the aforementioned technical problems by integrating Retrieval-Augmented Generation (RAG) technology with knowledge graph technology. It achieves a solution particularly suitable for forest fire emergency decision-making scenarios, combining historical decision-making information with real-time monitoring data. Through a hybrid retrieval technique using "vector + keyword," it provides more accurate and interpretable forest fire emergency decision support, addressing the challenges of real-time response, resource allocation, and scientific decision support in forest fire emergency decision-making. By constructing an intelligent forest fire command and decision-making system, it enables rapid response to forest fires.

[0033] Figure 1 This is a schematic diagram of the forest fire emergency command system in the embodiments of this application, as shown below. Figure 1 As shown, the system includes the following structure:

[0034] The computing device 10 is used to execute the steps in the various method embodiments of this application and has user interaction and data processing capabilities. The computing device 10 may include mobile terminals (such as smartphones, tablets, etc.), fixed terminals (such as computer terminals, command center workstations), servers, etc.

[0035] Memory 20 is used for long-term storage of structured historical decision documents. Memory 20 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. Memory 20 may be internal or external memory of the computing device 10, or it may be a memory remotely located equivalent to the computing device 20.

[0036] The fire monitoring equipment 30 is used to collect detailed data on forest fires in real time. The fire monitoring equipment 30 includes, but is not limited to, ground sensors, unmanned aerial vehicle (UAV) detectors, and satellite remote sensing equipment. For example, an integrated "space-air-ground" monitoring system can be adopted, using various technologies such as satellite remote sensing, high-altitude observation equipment, thermal imaging dual-spectrum gimbals, and intelligent checkpoint monitoring to construct a monitoring network.

[0037] In this embodiment, the computing device 10 can obtain phase light data from the memory 20 and the fire monitoring device 30 based on user instructions, and generate a real-time response plan for a forest fire scenario by running the response plan description information generation method in this application, so as to facilitate the deployment of fire fighting and rescue work and the scheduling of related resources.

[0038] According to one aspect of the embodiments of this application, a method for generating solution description information is provided, which can be executed by a computing device 10. Figure 2 This is a flowchart illustrating a method for generating solution description information according to an embodiment of this application, such as... Figure 2 As shown, the process of this method may include the following steps:

[0039] Step S202: Obtain real-time fire data and response plan query request;

[0040] Step S204: Perform a fusion retrieval based on the real-time fire data and the response plan query request to obtain multiple target knowledge nodes;

[0041] Step S206: Based on the pre-constructed knowledge graph, determine the knowledge node text associated with the multiple target knowledge nodes;

[0042] Step S208: Generate response plan description information based on the knowledge node text and the real-time fire data.

[0043] In this embodiment, the real-time fire data is used to represent event information of an ongoing fire event, and the response solution query request is used to request a response solution for the fire event. It should be noted that the query here is actually to generate a new response solution using enhanced retrieval generation technology; the response solution description information is used to represent the response solution for the fire event, that is, the emergency decision-making solution in the forest fire scenario.

[0044] The entities performing the above steps can be mobile terminals, computer terminals, servers, or other computing devices, but are not limited to these.

[0045] The method for generating response plan description information in this embodiment can be applied to the fields of artificial intelligence and emergency decision-making, especially in forest fire emergency decision-making scenarios. Through the above steps, response plans can be dynamically generated by combining real-time fire data and enhanced retrieval generation technology. This solves the problems of limited intelligence and poor dynamic adaptability in forest fire scenarios in related technologies, improving the real-time nature and flexibility of forest fire emergency decision-making.

[0046] In this embodiment, knowledge graphs can make the generated response plans (i.e., forest fire emergency decision-making plans) more interpretable, allowing decision-makers to clearly understand the formation process and basis of the decisions. By combining vector retrieval and keyword retrieval, a fusion retrieval technology is designed that can capture both "keywords" and understand "context," improving the accuracy and comprehensiveness of the retrieval. This provides more reliable information for generating high-quality decision-making plans for large models, thereby improving the accuracy and interpretability of forest fire emergency decisions.

[0047] In some embodiments, step S204, which involves performing a fusion retrieval based on the real-time fire data and the response plan query request to obtain multiple target knowledge nodes, may include the following steps:

[0048] Step S2042: Perform a splicing operation on the real-time fire data and the response plan query request to obtain spliced ​​text;

[0049] Step S2044: Perform vector retrieval in a pre-built vector database based on the concatenated text to obtain vector retrieval results;

[0050] Step S2046: Based on the concatenated text, perform keyword retrieval in a pre-built keyword index library to obtain keyword retrieval results;

[0051] Step S2048: Determine the multiple target knowledge nodes based on the vector retrieval results and the keyword retrieval results.

[0052] In some embodiments, step S2044, which involves performing vector retrieval in a pre-constructed vector database based on the concatenated text to obtain vector retrieval results, may include: vectorizing the concatenated text to obtain a target vector; performing vector retrieval on the target vector in the vector database to obtain a set of vectors matching the target vector; and obtaining a set of similarities corresponding one-to-one with the set of vectors and a first set of candidate knowledge nodes corresponding one-to-one with the set of vectors, wherein the similarity between each vector in the set of vectors and the target vector satisfies a preset similarity condition; wherein the vector retrieval results include the first set of candidate knowledge nodes and the set of similarities.

[0053] In one exemplary embodiment, vectorization processing includes vectorization encoding via an embedding model such as BGE-m3 or text2vec.

[0054] In this embodiment, a vector, also known as an embedding vector or a semantic vector, is a high-dimensional semantic vector generated from an entire concatenated text or a knowledge block / information block. The vectors in the vector database are generated during the construction of the knowledge graph. Therefore, each vector can correspond to a knowledge node, which is used to describe the semantic information of the node. Exemplarily, similarity can be calculated by methods such as cosine similarity and Euclidean distance. The similarity condition can be set as a similarity threshold. If it is higher than this similarity threshold, it can be considered that the semantics of two vectors are similar. For example, "blazing flames" and "flying fire" are semantically similar.

[0055] In some embodiments, the keyword retrieval in the pre-constructed keyword index library according to the concatenated text in step S2046 to obtain a keyword retrieval result may include: performing word segmentation on the concatenated text to obtain a set of target words; performing keyword retrieval on the set of target words in the keyword index library to obtain a set of keywords that match the set of target word units, and obtaining a second set of candidate knowledge nodes indexed by the set of keywords and a set of overlap degrees corresponding to the second set of candidate knowledge nodes one by one, where each keyword in the set of keywords is the same as or contains one word in the set of target words, and each overlap degree in the set of overlap degrees is determined based on a keyword set and a target word set related to the corresponding candidate knowledge node in the second set of candidate knowledge nodes. Each keyword in the keyword set indexes to the corresponding candidate knowledge node, and the target word set includes the words in the set of target words that are the same as or contained in each keyword in the keyword set; where the keyword retrieval result includes the second set of candidate knowledge nodes and the set of overlap degrees.

[0056] In this embodiment, the keyword index library is generated during the construction of the knowledge graph and is used to store the index information of each keyword, including the frequency and position of the keyword in the corresponding text. At the same time, each keyword can index to the corresponding knowledge node. Multiple keywords (also known as a keyword set) can be included in the attributes of each knowledge node.

[0057] In this embodiment, overlap means that two words have an inclusion relationship, or some characters are the same or completely the same. Precise matching of information can be achieved through keyword retrieval, and only exactly the same or partially the same words appear in the retrieval result. Exemplarily, the keyword index library can be constructed based on Whoosh or Elasticsearch, and keyword retrieval can be implemented using the BM25 algorithm, but the present application is not limited thereto.

[0058] In some embodiments, step S2048, determining the plurality of target knowledge nodes based on the vector retrieval results and the keyword retrieval results, may include: when the vector retrieval results include a first group of candidate knowledge nodes with a one-to-one correspondence and a set of similarities, and the keyword retrieval results include a second group of candidate knowledge nodes with a one-to-one correspondence and a set of overlaps, performing a deduplication operation on the first group of candidate knowledge nodes and the second group of candidate knowledge nodes to obtain a third group of candidate knowledge nodes; determining the similarity and overlap corresponding to each candidate knowledge node in the third group of candidate knowledge nodes based on the set of similarities and the set of overlaps; determining the matching degree corresponding to each candidate knowledge node in the third group of candidate knowledge nodes based on the similarity and overlap corresponding to each candidate knowledge node in the third group of candidate knowledge nodes; determining K candidate knowledge nodes in the third group of candidate knowledge nodes as the plurality of target knowledge nodes, where K is a preset positive integer greater than 1, and the matching degree corresponding to the K candidate knowledge nodes is greater than the matching degree corresponding to the candidate knowledge nodes in the third group of candidate knowledge nodes other than the K candidate knowledge nodes.

[0059] In this embodiment, the first set of candidate knowledge nodes is a set of candidate knowledge nodes that correspond one-to-one with a set of vectors. The set of vectors includes vectors that match the target vector obtained by vector retrieval of the target vector in the vector database. The target vector is a vector obtained by vectorizing the concatenated text. The set of similarities corresponds one-to-one with the set of vectors. The similarity between each vector in the set of vectors and the target vector satisfies a preset similarity condition.

[0060] In this embodiment, the second set of candidate knowledge nodes includes candidate knowledge nodes indexed by a set of keywords. The set of keywords includes keywords that match the set of target words obtained by keyword retrieval of a set of target words in the keyword index library. The set of target words includes words obtained by word segmentation of the concatenated text. Each keyword in the set of keywords is the same as or contains a word in the set of target words. Each overlap in the set of overlap is determined based on the set of keywords and the set of target words related to the candidate knowledge nodes in the second set of candidate knowledge nodes. Each keyword in the set of keywords is indexed to the corresponding candidate knowledge node. The set of target words includes words in the set of target words that are the same as or contained in each keyword in the set of keywords.

[0061] In some embodiments, similarity may also be referred to as vector retrieval score, vector similarity, semantic similarity, etc., to represent the score of vector retrieval results; overlap may also be referred to as keyword retrieval score, concept overlap, keyword hit score, overlap, etc., to represent the score of keyword retrieval results. Matching degree may also be referred to as comprehensive retrieval score, etc., to represent the score of comprehensive retrieval results.

[0062] In some embodiments, determining the similarity and overlap corresponding to each candidate knowledge node in the third set of candidate knowledge nodes based on the set of similarities and the set of overlaps may include at least one of the following:

[0063] (1) When the first part of the candidate knowledge nodes in the third group of candidate knowledge nodes have a corresponding similarity in the same group of similarities and a corresponding overlap in the same group of overlaps, the similarity and overlap of each candidate knowledge node in the first part of the candidate knowledge nodes are respectively set to the similarity in the same group of similarities and the overlap in the same group of overlaps.

[0064] (2) If the second part of the candidate knowledge nodes in the third group of candidate knowledge nodes has a corresponding similarity in the same group of similarities but no corresponding overlap in the same group of overlaps, the similarity and overlap of each candidate knowledge node in the second part of the candidate knowledge nodes shall be set to the similarity and 0 respectively in the same group of similarities.

[0065] (3) If the third part of the candidate knowledge nodes in the third group of candidate knowledge nodes does not have a corresponding similarity in the same group of similarities, but has a corresponding overlap in the same group of overlaps, the similarity and overlap of each candidate knowledge node in the third part of the candidate knowledge nodes shall be set to 0 and the overlap corresponding to the same group of overlaps, respectively.

[0066] In some embodiments, determining the matching degree of each candidate knowledge node in the third group of candidate knowledge nodes based on the similarity and overlap of each candidate knowledge node in the third group of candidate knowledge nodes may include: normalizing the similarity and overlap of each candidate knowledge node in the third group of candidate knowledge nodes to obtain a set of normalized similarity and a set of normalized overlap; and performing a weighted operation on the set of normalized similarity and the set of normalized overlap to obtain the matching degree of each candidate knowledge node in the third group of candidate knowledge nodes.

[0067] For example, the weighted operation can be implemented using the following formula: Matching degree = Similarity degree α+ overlap (1-α); where the preset weight is α, that is, the weight of the vector retrieval score is set to α, and the weight of the keyword retrieval score is set to 1-α. This weight can be optimized through actual testing or machine learning.

[0068] In some embodiments, step S206, which involves determining the knowledge node text associated with the plurality of target knowledge nodes based on a pre-constructed knowledge graph, may include:

[0069] Step S2062: In the knowledge graph, starting from each of the plurality of target knowledge nodes, perform an N-hop search along edges with edge weights greater than a preset threshold to obtain at least one extended knowledge node associated with each target knowledge node, where N is a preset positive integer. The knowledge graph includes a set of knowledge nodes and a set of edges. Each edge in the set of edges connects two knowledge nodes in the set of knowledge nodes. Each edge has a predetermined edge weight, which is determined based on the overlap and similarity between the two knowledge nodes connected by the edge.

[0070] Step S2064: Concatenate the text corresponding to each of the multiple target knowledge nodes and the text corresponding to at least one extended knowledge node associated with each target knowledge node to obtain the knowledge node text.

[0071] In this embodiment, N-hop refers to all nodes / edges reachable by taking N steps from the starting point. N-hop search refers to explicitly limiting the search scope on the knowledge graph to subgraphs whose distance from the starting node is ≤ N hops, and then returning the results to downstream tasks. This embodiment expands the fusion retrieval results through N-hop search, improving the content richness of the knowledge node text. In the process of constructing the knowledge graph, the edge weight of each edge is obtained by weighted fusion based on the vector similarity and keyword overlap between the texts corresponding to the two connected knowledge nodes, which can comprehensively reflect the semantic similarity between two adjacent nodes and whether there are overlapping tags / concepts / keywords.

[0072] In one exemplary embodiment, the concatenation can be achieved by first concatenating the text of each target knowledge node and the text of its associated extended knowledge nodes sequentially, and then concatenating the concatenated texts related to each target knowledge node sequentially. However, this application is not limited to this. For example, a simple concatenation strategy can be set as P={p1,…,pk}, where P is the concatenated knowledge node text, p1,…,pk are the texts of each knowledge node, and preset separators such as spaces and line breaks can be set between the texts.

[0073] In an exemplary embodiment, since the text of adjacent nodes may have overlapping areas, deduplication can be performed on adjacent text during the splicing process. For example, the final length of the knowledge node text can be limited, and some of the extended node text can be deleted based on the length limit during the splicing process to improve the efficiency of subsequent large model processing.

[0074] In some embodiments, step S208, which involves generating response plan description information based on the knowledge node text and the real-time fire data, may include the following steps:

[0075] Step S2082: Generate target prompt words based on the knowledge node text, the real-time fire data, and the preset prompt word template, wherein the prompt word template is used to instruct the large language model to perform at least one preset response solution generation task;

[0076] Step S2084: Input the target prompt word into the large language model to obtain the coping solution description information output by the large language model, wherein the coping solution description information includes the generation result of each coping solution generation task in the at least one coping solution generation task.

[0077] In some embodiments, the at least one response plan generation task includes at least one of the following: determining the fire extinguishing path, fire extinguishing method, quantity of fire extinguishing resources, and allocation plan of fire extinguishing resources based on the fire terrain, meteorological conditions, and fire intensity; and determining the rescue path, quantity of rescue resources, and allocation plan of rescue resources based on personnel casualties and fire location.

[0078] In an exemplary embodiment, the prompt word template can be set as follows: "Based on real-time fire data and knowledge node text, perform the following preset tasks: In terms of fire extinguishing strategy formulation, generate a scientific fire extinguishing path based on the fire terrain, meteorological conditions, and fire intensity, including determining where to start fire extinguishing, what fire extinguishing method to use, etc., while making reasonable resource allocation, clarifying the required number of fire extinguishing equipment (such as fire trucks, water pumps, fire extinguishing bombs, etc.), the number of firefighters, and their deployment direction; In terms of rescue force deployment, generate the optimal dispatch path for rescue teams based on casualties and fire location, plan the order and route of rescue teams to various disaster points, and determine the required resources for rescue, such as the quantity and deployment plan of medical equipment, stretchers, first aid medicines, etc. The model is required to generate answers only based on the given context to prevent the model from fabricating information." Users can set different prompt word templates according to experience or actual scenario needs, and this application does not impose any restrictions on this.

[0079] In some embodiments, the real-time fire data includes at least one of the following: fire terrain, meteorological conditions, fire intensity, casualties, and fire location. For example, real-time fire data can be monitored through one or more methods such as sensors, drones, and remote sensing satellites; this application is not limited to these methods.

[0080] In some embodiments, the response solution query request can be generated based on user interaction. For example, users can generate it through text output, selection of query conditions, etc., in the system interface. For instance, the response solution query request could be "What forest fire control measures should be taken now?", but this application does not limit this.

[0081] In some embodiments, before obtaining real-time fire data and response plan query requests in step S202, the method may further include: obtaining historical response plan description information, wherein the historical response plan description information is used to represent historical response plans corresponding to at least one fire event that has occurred in a predetermined forest area; performing data cleaning and structuring processing on the historical response plan description information to obtain structured information; and constructing the knowledge graph, vector database, and keyword index library based on the structured information.

[0082] In some embodiments, the step of cleaning and structuring the historical response plan description information to obtain structured information may include: cleaning the historical response plan description information to remove invalid and duplicate data, and obtaining cleaned information; and extracting the value corresponding to each field in the tuple template from the cleaned information according to a preset tuple template to obtain the structured information.

[0083] In one exemplary embodiment, duplicate data may include identical disposal methods, document paragraphs, or statements, while invalid data may include missing fields, garbled characters, or meaningless strings.

[0084] In some embodiments, the preset tuple template includes at least two of the following fields: fire type field, fire cause field, time and season of occurrence field, fire location and terrain field, command and rescue division of labor field, fire extinguishing strategy field, and resource requirement field.

[0085] For example, the pre-defined plural template can adopt the "5W2H" paradigm, including (What, Why, When, Where, Who, How, How much), with the following specific meanings: What, fire type (surface fire, crown fire, etc.); Why, cause of fire (lightning strike, human error, etc.); When, time of fire occurrence and seasonal characteristics; Where, location of the fire and surrounding terrain; Who, division of labor between command and rescue teams; How, firefighting strategy (e.g., "establishing firebreaks should avoid steep slopes"); How much, resource requirements and time requirements. However, this application is not limited to this.

[0086] In some embodiments, constructing the knowledge graph, vector database, and keyword index based on the structured information may include the following steps:

[0087] (1) The structured information is divided into a group of information blocks, wherein the total number of characters in each information block in the group of information blocks is less than or equal to a first preset value, and the number of overlapping characters between two adjacent information blocks in the group of information blocks is equal to a second preset value; for example, the first preset value can be set to 500 and the second preset value can be set to 50, but this application does not limit this, and the values ​​can be adjusted according to user needs.

[0088] (2) Construct the knowledge graph based on the set of information blocks, wherein the knowledge graph includes a set of knowledge nodes and a set of edges, each knowledge node in the set of knowledge nodes is used to represent a corresponding information block in the set of information blocks, each edge in the set of edges connects two knowledge nodes in the set of knowledge nodes, and each edge has a predetermined edge weight; wherein the node attributes of each knowledge node include the corresponding information block, the vector corresponding to the corresponding information block, and the concept tag extracted from the information block, the concept tag includes multiple keywords and / or entities, and the edge weight is determined based on the overlap of the keywords corresponding to the two knowledge nodes connected by the corresponding edge and the similarity of the vectors corresponding to the two knowledge nodes;

[0089] (3) Perform vectorization processing on each information block in the set of information blocks to obtain the vector corresponding to each information block, wherein the vector database includes a set of vectors, and the set of vectors includes the vector corresponding to each information block in the set of information blocks; wherein each vector in the set of vectors corresponds to a knowledge node in the knowledge graph, and the vector is a multidimensional semantic vector;

[0090] (4) Perform word segmentation on each information block in the set of information blocks to obtain a set of words corresponding to each information block, determine each word in the different words in the set of words as a keyword, and obtain multiple keywords; index each keyword in the multiple keywords to a knowledge node in the knowledge graph; and determine the frequency or number of times each keyword appears in the information blocks that correspond one-to-one with each knowledge node in the knowledge graph, wherein the keyword index library includes the multiple keywords, the knowledge node indexed by each keyword in the multiple keywords, and the frequency or number of times each keyword in the multiple information blocks that correspond one-to-one with each knowledge node.

[0091] This application does not impose strict restrictions on the construction order of the aforementioned knowledge graph, vector database, and keyword index. For example, vectorization processing and word segmentation processing can be executed in parallel and independently, and the complete construction process of the knowledge graph can also be carried out synchronously with vectorization processing and word segmentation processing. However, the node attributes and edge connections in the knowledge graph depend on the results of vectorization processing and word segmentation processing.

[0092] In this embodiment, each information block corresponds to a knowledge node in the knowledge graph. When constructing the knowledge graph, the node framework of the knowledge graph can be built first based on the information blocks. Then, each information block is processed by word segmentation and vectorization to generate corresponding vectors, keywords, and other information. These vectors and keywords are then filled into the attributes of the knowledge nodes. Next, the similarity between the vectors corresponding to each knowledge node is calculated, the overlap between the keywords corresponding to each knowledge node is calculated, and the edge weight is determined by combining the similarity and overlap between each pair of nodes (for example, a weighted average of similarity and overlap can be used to obtain the edge weight value). Finally, edge connections can be established between nodes with edge weights greater than a certain value (e.g., 0.6), thereby completing the construction of the entire knowledge graph database. Edge connections visually represent the degree of association between two nodes and also help expand the search results in subsequent retrieval processes, facilitating the visualization of the reasoning process.

[0093] In some embodiments, after generating response plan description information based on the knowledge node text and the real-time fire data in step S208, the method further includes: generating inference information related to the response plan description information based on the knowledge graph; wherein the knowledge graph includes a set of knowledge nodes and a set of edges, each edge in the set of edges connects two knowledge nodes in the set of knowledge nodes, each edge has a predetermined edge weight, and the set of knowledge nodes includes the plurality of target knowledge nodes; wherein the inference information includes the plurality of target knowledge nodes, the node attributes of each of the plurality of target knowledge nodes, the edges connected to each of the plurality of target knowledge nodes, and the edge weights of the edges.

[0094] In this embodiment, the reasoning information is used to demonstrate the reasoning logic of the solution, which helps to increase the user's trust in the final result.

[0095] In this embodiment, a knowledge graph is constructed to achieve structured management of historical cases, enhancing semantic connections between information and improving retrieval efficiency. During a fire, relevant cases can be dynamically retrieved based on real-time fire data, and decision-making strategies can be dynamically adjusted based on factors such as fire intensity changes, weather conditions, and terrain features, improving the real-time nature and flexibility of forest fire emergency decision-making. By employing a "vector + keyword" fusion retrieval method, the needs of semantic understanding and precise matching can be balanced, improving the quality of retrieval results. This application also combines the knowledge graph to achieve a visual representation of the reasoning path, improving the credibility and execution efficiency of decision results.

[0096] This application involves two core functions: knowledge base construction and fusion retrieval. The specific implementation methods of these two core functions will be described below with reference to the attached figures.

[0097] The knowledge base in this application focuses on emergency decision-making for forest fires, storing historical decision-making information covering detailed data such as the time and location of the fire, its size, weather conditions, the measures taken, and their effects. The knowledge base includes a knowledge graph, a vector database, and keyword index data blocks. By constructing this knowledge base, historical decision-making information (such as documents and case studies) can be transformed from text into knowledge nodes, thereby establishing a knowledge graph and achieving structured information management.

[0098] In one embodiment of this application, in order to improve the interpretability of decisions, a knowledge graph can be constructed using graph retrieval-augmented generation (Graph RAG) technology.

[0099] Figure 3This is a schematic diagram of the knowledge graph construction process in one embodiment of this application, such as... Figure 3 As shown, this process mainly consists of two parts: converting text into knowledge nodes, and constructing a knowledge graph. Specifically, this process may include the following steps:

[0100] Step S301, Data Acquisition and Preprocessing: Collect data such as the text of the city's forest fire emergency plan, historical cases, and the distribution of emergency resources, transform them into structured data, and annotate them using the "5W2H" paradigm. That is, annotate based on a preset seven-tuple template.

[0101] The "5W2H" paradigm includes (What, Why, When, Where, Who, How, How much), with the following specific meanings: What, fire type (surface fire, crown fire, etc.); Why, cause of fire (lightning strike, human error, etc.); When, time of fire occurrence and seasonal characteristics; Where, location of the fire and surrounding terrain; Who, division of labor between command and rescue teams; How, firefighting strategy (e.g., "firebreaks should avoid steep slopes"); How much, resource requirements and time constraints.

[0102] In one exemplary embodiment, the following structured data can be obtained based on the "5W2H" paradigm:

[0103] {

[0104] What: "Surface fire"

[0105] Why: Lightning strike

[0106] "When": "Summer"

[0107] "Where": "ridge area",

[0108] "Who": "Forest fire brigade, medical team",

[0109] How: "Establishing firebreaks and aerial firefighting",

[0110] "How much": "within 2 hours"

[0111] }

[0112] Step S302, Document Slicing: Since historical decision-making and handling documents related to forest fires are often quite long, direct processing can easily lead to memory overflow in large models. Therefore, the documents need to be sliced ​​appropriately using a sliding window method.

[0113] For example, the slicing formula can be set as: for i in range(0, len(text), chunk_size -overlap);

[0114] Here, `chunk_size` represents the size of each knowledge block, and `overlap` represents the overlapping portion of adjacent knowledge blocks to avoid knowledge gaps. For example, a complete fire response report can be divided into chunks of 500 words each, with a 50-word overlap to ensure the continuity of knowledge.

[0115] Step S303, Concept Extraction: Analyze each knowledge block using a large model (such as Llama, GPT) and automatically extract 5 to 10 keywords or entities as node labels.

[0116] For example, taking the description of a forest fire response measure as "using water pumps in relay to pump water to suppress the fire, while organizing personnel to open up firebreaks", after processing by a large model, core concepts such as "water pump relay pumping", "firebreaks" and "fire suppression" can be extracted, providing a basis for subsequent node association.

[0117] Step S304, Vectorization: For each knowledge block, generate a corresponding high-dimensional vector (also called an embedding vector or semantic vector) using an embedding model (such as BGE-m3, text2vec). Batch processing is used to improve efficiency. Use a vector database (such as Milvus) to store the semantic vectors.

[0118] Step S305, Keyword Indexing: First, each knowledge block is segmented using a word segmentation tool suitable for the Chinese forest fire domain, breaking the knowledge block text into independent word units. An index is built based on the segmentation results, recording the frequency and location of each word in each knowledge block, forming a keyword index library that can be quickly queried.

[0119] For example, to optimize retrieval speed, BM25 retrieval can integrate with professional search engine tools such as Whoosh and Elasticsearch: when using Whoosh, its built-in index structure and query optimization mechanism can support efficient keyword matching and score calculation; when choosing Elasticsearch, its distributed architecture and inverted index technology can achieve millisecond-level response in large-scale forest fire knowledge bases, which is especially suitable for scenarios that need to process massive amounts of historical decision data.

[0120] Step S306: Construct the knowledge graph. Define nodes and establish edge connections between nodes according to the edge connection strategy.

[0121] The nodes are defined as follows: Each knowledge block is a node, and the node attributes include text content, concept tags, and corresponding embedding vectors. For example, a knowledge block node about the meteorological conditions of a fire has text content that records detailed data such as wind speed, temperature, and humidity at the time, concept tags that include "wind speed," "temperature," and "humidity," and also has corresponding embedding vectors.

[0122] Edge connection strategy: If two nodes share a common concept, a connection is established. The edge weight (edge_weight) can be calculated by weighted averaging of concept overlap and vector similarity. For example, when edge_weight > 0.6, an edge connection can be established. For instance, the "artificial rainmaking" response node and the "meteorological conditions suitable for artificial rainmaking" meteorological node are connected because they share the common concept of "artificial rainmaking" and the calculated edge weight meets the condition, thus making the implicit connections between knowledge blocks explicit.

[0123] For example, the weighted average formula for edge weights is as follows:

[0124] edge_weight=0.7 similarity +0.3 len(shared_concepts) / min(len(node_i.concepts), len(node_j.concepts)));

[0125] Where edge_weight is the edge weight, similarity is the vector similarity, shared_concepts are the concepts shared by the two nodes, node_i.concepts are all the concepts of node i, node_j.concepts are all the concepts of node j, and len() is used to count the length of the concepts.

[0126] Through the above steps, a vector database, a keyword index library, and a knowledge graph can be constructed, enabling the structured processing of historical decision-making and response documents for forest fires. This greatly facilitates the retrieval and generation of emergency plans in subsequent forest fire scenarios.

[0127] Figure 4 This is a schematic diagram of the knowledge base construction process in an exemplary embodiment of this application, such as... Figure 4 As shown, through processes such as document slicing, concept extraction, vectorization, and keyword indexing, a vector database and a keyword index can be constructed. Then, based on the definition of nodes and edges, a "node-edge structure" is established, ultimately forming a knowledge graph for forest fire emergency decision-making.

[0128] In this embodiment, Milvus can be used as the vector database. The keyword index library can be integrated with professional search engine tools such as Elasticsearch to improve subsequent retrieval efficiency and the response speed of the solution generation method. Other vector databases or search engines can also be used, and this application is not limited to them.

[0129] In this embodiment, the node attributes of each node may include the text content of the corresponding knowledge block, concept tags, and embedding vectors. For example, the edge connection strategy may be set to establish an edge connection between two nodes that share a common concept and have an edge weight > 0.6. However, this application is not limited to this.

[0130] In one embodiment of this application, fusion retrieval further optimizes the retrieval and answer generation process based on Graph RAG. Specifically, fusion retrieval uses both vector retrieval and keyword retrieval methods in parallel. Based on preset weight values, it calculates a comprehensive score for the two retrieval results, and finally sorts and removes duplicates to obtain the final retrieval results. The score for vector retrieval results is also referred to as vector similarity, semantic similarity, or similarity, while the score for keyword retrieval results is also referred to as concept overlap, keyword hit score, or overlap.

[0131] Figure 5 This is a schematic diagram of the fusion retrieval process in one embodiment of this application, as shown below. Figure 5 As shown, the process includes the following steps:

[0132] Step S501, Vector Retrieval: Combine real-time fire data with query statements (e.g., "What forest fire control measures should be taken under current high temperature and high wind speed conditions"), and vectorize them using an embedding model (e.g., BGE-m3, text2vec). In the constructed knowledge graph vector database, calculate the semantic similarity with each node (knowledge block) and filter out nodes with similar semantics. For example, semantic similarity is commonly calculated using methods such as cosine similarity and Euclidean distance.

[0133] Step S502, Keyword retrieval: Combine real-time fire data with query statements, perform word segmentation, extract keywords such as "high temperature", "high wind speed" and "response measures", and use keyword retrieval algorithms (such as the BM25 algorithm) to search for nodes containing these keywords in the keyword index database to achieve accurate positioning of key information.

[0134] In this embodiment, steps S501 and S502 can be run in parallel to improve the retrieval efficiency of fusion retrieval.

[0135] Step S503, Score Normalization and Weighted Fusion: The semantic similarity score obtained from vector retrieval and the keyword hit score obtained from BM25 keyword retrieval are normalized to the range [0,1]. Weighted fusion is performed according to a preset weight α (e.g., the weight of the vector retrieval score is α, and the weight of the BM25 keyword retrieval score is 1-α; this weight can be optimized through actual testing or machine learning) to obtain the comprehensive score for each node.

[0136] For example, the BM25 score can be scaled to the [0,1] interval using the Min-Max normalization method; and the cosine similarity calculated with vector retrieval is normalized using Z-Score to eliminate scale differences.

[0137] For example, when α approaches 0.5, vector retrieval and keyword retrieval results each account for half, which is suitable for most scenarios; when α approaches 1, the retrieval is more biased towards semantic understanding; when α approaches 0, the retrieval is more biased towards keyword matching. This application does not impose restrictions on the value of α.

[0138] Step S504, Comprehensive Sorting and Top-K Selection: Sort all nodes in descending order based on the comprehensive score, and select Top-K, which are the K nodes with the highest comprehensive scores. The value of K is set according to actual needs, such as K=10.

[0139] In some embodiments, in addition to Top-K nodes, the graph traversal mechanism of the knowledge graph can be combined (as step S505) to avoid duplicate or conflicting information. Starting from these nodes, multi-hop reasoning is performed along paths with high edge weights using breadth-first search combined with a priority queue to mine more indirectly related knowledge nodes and expand the retrieval scope.

[0140] Figure 6 This is a schematic diagram of the fusion retrieval process in an exemplary embodiment of this application, such as... Figure 6 As shown, the system input is a combination of user queries and real-time fire data, such as "firefighting measures under high temperature and strong winds". Based on the above fusion retrieval process, specific examples of vector retrieval and keyword retrieval are as follows.

[0141] Step S501, the vector retrieval branch, may include the following steps: vectorization of the input (query statement + real-time fire data) (e.g., using an embedding model); querying a vector database (e.g., Miluvs); calculating semantic similarity (e.g., cosine similarity or Euclidean distance); and determining the vector retrieval score.

[0142] Step S502, the keyword retrieval branch, can be implemented using the BM25 algorithm. This branch may include the following steps: word segmentation processing of the input (query statement + real-time fire data) (e.g., using a domain-specific word segmentation tool); keyword extraction; keyword index query (e.g., Whoosh, Elasticsearch); BM25 score calculation; and determination of the keyword retrieval score.

[0143] In this embodiment, after obtaining the vector retrieval score and the keyword retrieval score respectively, they can be weighted and fused (such as by weighted average operation) to obtain a comprehensive score. The comprehensive score is then sorted, and the K nodes with the highest comprehensive scores (Top-K) are selected. Then, a Graph RAG is traversed based on the Top-K nodes to expand related knowledge nodes. Finally, a knowledge block of the retrieval results is generated based on the Top-K nodes and the expanded knowledge nodes.

[0144] In this embodiment of the application, after the fusion retrieval is completed, the decision generation stage can begin. Decision generation can be implemented based on a large model.

[0145] Figure 7 This is a schematic diagram of the decision generation process in one embodiment of this application, as shown below. Figure 7 As shown, the process includes the following steps: The decision generation process is as follows:

[0146] Step S701: The selected node text (including the initial Top-K nodes and the nodes expanded by graph traversal) can be concatenated to form a complete context.

[0147] Specifically, real-time fire data, including real-time fire terrain data (such as slope and vegetation cover), meteorological data (such as wind speed, wind direction, temperature, and humidity), fire intensity data (such as fire line length and spread rate), as well as information on casualties and fire location, are fused with the retrieved context. To ensure the effectiveness of the large model input, the length of the context needs to be controlled to avoid exceeding the model's limit in terms of the number of tokens.

[0148] Step S702: Input the context into the large model (such as GPT-4, Wenxin Yiyan, etc.), and set system prompt words to explicitly require the large language model to execute the preset decision task based on the context after the fusion of data, and finally output the decision solution.

[0149] For example, the pre-defined decision-making tasks may include at least one of the following: In terms of firefighting strategy formulation, based on the fire terrain, meteorological conditions, and fire intensity, generate a scientific firefighting path, including determining where to begin firefighting, what firefighting methods to use, etc., while simultaneously allocating resources reasonably, specifying the required number of firefighting equipment (such as fire trucks, water pumps, fire extinguishing bombs, etc.), the number of firefighters, and their deployment direction; In terms of rescue force deployment, based on casualties and fire location, generate the optimal dispatch path for rescue teams, plan the order and route of rescue teams to various disaster sites, and simultaneously determine the required resources for rescue, such as the quantity and deployment plan of medical equipment, stretchers, and first-aid medicines. The model is required to generate answers only based on the given context to prevent the model from fabricating information. The decision-making tasks in this application are only examples and can be adjusted according to user experience and actual scenario requirements. For example, if other types of firefighting or rescue resources are involved, the prompts can be appropriately adjusted or corresponding decision-making tasks can be added.

[0150] Step S703 involves tracing the knowledge graph back to specific knowledge nodes to demonstrate the decision-making basis and reasoning path. This step is optional and can increase the credibility of the decision-making plan, facilitating its rapid implementation.

[0151] Through the above steps, Graph RAG can be implemented, which generates decision-making schemes based on the fusion retrieval results of knowledge graphs.

[0152] In one embodiment of this application, a forest fire emergency command system (system structure as follows) Figure 1 The overall workflow (as shown) is divided into three main stages: knowledge preparation, real-time response, and result output.

[0153] Figure 8 This is a schematic diagram of the overall workflow of a forest fire emergency command system in an exemplary embodiment of this application, as shown below. Figure 8 As shown, the overall workflow is as follows:

[0154] S1, Knowledge Preparation Phase (Offline Processing). The input for this phase is historical decision-making documents, such as fire reports, response plans, and rescue records. The outputs of this phase are a knowledge graph, a vector database, and a keyword index.

[0155] Specifically, the knowledge preparation phase includes the following steps:

[0156] S1-1, Historical Data Collection and Preprocessing: Collect various document data related to historical decision-making and handling of forest fires, including fire reports, handling plans, rescue records, etc. Clean the data, remove duplicate and invalid information, and standardize the format.

[0157] S1-2, Document Slicing and Feature Extraction: The preprocessed document is sliced ​​using a sliding window method to generate several knowledge chunks; for each knowledge chunk, the core concept labels are extracted using a large model, and the corresponding embedding vectors are generated using a pre-trained language model.

[0158] S1-3, Knowledge Graph Construction: Using knowledge blocks as nodes, establish edge connections based on common concepts and semantic similarity between nodes, calculate edge weights, construct a knowledge graph containing knowledge of forest fire emergency decision-making, and store it in a graph database; at the same time, perform word segmentation on knowledge blocks, and build a keyword index library based on search engines such as Whoosh or Elasticsearch to support subsequent retrieval.

[0159] S2, Real-time Response Phase. The inputs for this phase are real-time fire data and user queries.

[0160] Specifically, the real-time response phase includes the following steps:

[0161] S2-1, Real-time Fire Data Acquisition and Query Reception: Collects real-time fire data through sensors, drones, on-site reporting, etc., including fire terrain, meteorological conditions, fire intensity, casualties, fire location, etc.; and receives emergency decision query requests input by users.

[0162] S2-2, Fusion retrieval execution: The user query request is vectorized and segmented. The semantically similar nodes are found in the knowledge graph through vector retrieval, and the keyword matching nodes are found in the index through keyword retrieval algorithm (such as BM25). The scores of the two retrieval results are normalized and weighted and fused. After comprehensive ranking, the Top-K nodes are selected, and multi-hop reasoning is performed in combination with graph traversal mechanism to expand relevant knowledge nodes.

[0163] S2-3, Context Fusion and Decision Generation: The retrieved knowledge node texts are concatenated to form a basic context, which is then integrated with real-time fire data and input into a large language model. The large language model then generates decision-making schemes for fire extinguishing strategy formulation and rescue force allocation according to preset task requirements.

[0164] S3, Result Output Stage.

[0165] Specifically, this stage involves the output of decision-making schemes and reasoning paths: outputting emergency decision-making schemes generated by a large language model, and displaying the reasoning paths based on the decision-making through a knowledge graph, including the knowledge nodes involved and the connections between nodes, to achieve the interpretability of the decision.

[0166] The forest fire emergency command system in this application involves multiple layers of system interaction, including: user layer, interaction interface layer, data layer, core processing layer, and execution layer, as detailed below:

[0167] User layer: Users are forest fire emergency command personnel who perform operations such as query input, data viewing, and command issuance through the system's interactive interface. They are the main operators and decision receivers of the system.

[0168] Interactive Interface Layer: The system interactive interface is the core interaction carrier between users and the system. It has functions such as data display, request input, command forwarding, and result presentation, and supports the visualization of real-time fire data, decision-making plans, and reasoning paths.

[0169] Data Layer: This includes real-time data acquisition devices and an offline knowledge base. Real-time data acquisition devices include sensors, drones, and on-site reporting terminals, responsible for collecting dynamic data such as fire terrain, weather, and fire intensity. The offline knowledge base contains structured historical cases, emergency plans, and other data, stored in the form of knowledge graphs and keyword indexes.

[0170] The core processing layer consists of a data processing module, a fusion retrieval module, and a decision generation module. The data processing module performs preprocessing such as cleaning and format conversion on the real-time collected data; the fusion retrieval module performs "vector + keyword" fusion retrieval and combines knowledge graphs for multi-hop reasoning; the decision generation module merges the retrieval results with real-time data and inputs them into a large model to generate decision schemes.

[0171] Execution layer: The execution terminal includes mobile terminals carried by the rescue team, fire equipment control systems, etc. It receives and executes command instructions issued by users through the interactive interface, and at the same time feeds back the execution status to the system interactive interface, forming a closed-loop interaction.

[0172] Figure 9 This is a schematic diagram of the system interaction flow in one embodiment of this application, such as... Figure 9 As shown, the process includes the following steps:

[0173] In step S901, emergency command personnel input a decision query request (such as "the handling plan for surface fires in mountain ridge areas under high temperature and strong wind weather") through the system interactive interface, and can also view real-time fire data transmitted by real-time data acquisition equipment and preprocessed by the data processing module.

[0174] In step S902, the system interface forwards the query request and real-time data to the fusion retrieval module. The fusion retrieval module retrieves relevant knowledge from the offline knowledge base, executes the fusion retrieval process, obtains the Top-K nodes, and expands the relevant knowledge through graph traversal.

[0175] In step S903, the decision generation module receives the retrieval results, integrates them with real-time fire data, and inputs them into the large model to generate decision content including fire extinguishing strategies and resource allocation plans, while extracting reasoning paths from the knowledge graph.

[0176] In step S904, the system's interactive interface visualizes the decision-making plan and reasoning path for emergency command personnel, who then issue command instructions to the execution terminal based on the plan through the interface.

[0177] Step S905: Execute the terminal execution command and send the execution progress, on-site feedback and other status information back to the system interaction interface to realize the interactive closed loop of the entire emergency decision-making process.

[0178] Through the above steps, in forest fire emergency decision-making scenarios, commanders only need to output a query request, and the system can automatically obtain real-time fire data and generate the most suitable fire-fighting strategy and resource allocation plan through a series of processing (fusion retrieval, large model processing, etc.), and display the corresponding reasoning path. This effectively improves the decision-making speed of commanders and the response speed of fire events, and can effectively reduce the spread of forest fires and avoid further losses.

[0179] The following key technical points are involved in the embodiments of this application:

[0180] 1. Application of RAG technology in forest fire decision-making: This is the first time that Retrieval Enhanced Generation (RAG) technology has been applied to the field of forest fire emergency command. By integrating historical decision-making information with real-time fire monitoring data, emergency decision support is achieved. Specifically, this includes a knowledge graph and RAG technology combination architecture, as well as a dynamic correlation mechanism between historical and real-time data, which solves the problem of inefficient decision-making in traditional systems that rely on human experience.

[0181] 2. Structured Emergency Response Plan Knowledge Base Based on the "5W2H" Paradigm: The "5W2H" paradigm (What, Why, When, Where, Who, Which, How, How much) is used to structure and annotate forest fire emergency response plan texts, historical cases, and emergency resource data, forming multi-dimensional semantic relationships. The specific fields of "5W2H" are defined as follows: What corresponds to the fire type, Why corresponds to the cause of the fire, When corresponds to the time and season of occurrence, Where corresponds to the fire location and terrain, Who corresponds to the division of command and rescue responsibilities, How corresponds to the firefighting strategy, and How much corresponds to the resource requirements and time constraints.

[0182] 3. Knowledge graph construction technology based on historical decision-making and handling information: By using technologies such as named entity recognition, relation extraction and semantic role labeling, structured knowledge is extracted from unstructured historical handling information to construct a professional knowledge graph in the field of forest fire decision-making.

[0183] 4. "Vector + Keyword" Fusion Retrieval Mechanism: Combining the advantages of vector retrieval (semantic understanding) and BM25 keyword retrieval (exact matching), this mechanism improves the accuracy and comprehensiveness of retrieval results through dynamic weight allocation and score fusion.

[0184] 5. Methods to enhance the interpretability of decisions: By using path queries and visualizations of knowledge graphs, decision-making schemes generated by large models are associated with specific knowledge nodes in the knowledge graph, displaying the connection relationships and reasoning logic chains between nodes, clarifying the association paths between decision-making schemes and historical decision information and real-time data, thereby providing the logical chain and basis for decision generation, and improving the credibility and interpretability of decisions.

[0185] By employing the aforementioned technical approaches, we can effectively address the problems existing in related technologies, such as fragmented emergency plan knowledge bases, limited level of decision-making intelligence, poor dynamic adaptability, poor decision interpretability, and limited retrieval technology. This leads to the construction of an intelligent forest fire command and decision-making system, enabling rapid response to forest fires.

[0186] The technical solutions in the embodiments of this application have the following beneficial effects.

[0187] 1. Enhanced Decision Interpretability: By constructing a knowledge graph within the knowledge base, the causal and relational relationships between historical decision-making information are clearly represented. When generating decision-making solutions, the knowledge graph allows tracing the basis and logical derivation process of the decision, enabling decision-makers to clearly understand why a certain measure was taken and its relationship with other factors (such as weather conditions and rescue resources), thereby improving the interpretability and credibility of the decision. This is achieved through the technical means of constructing a knowledge graph, directly solving the problem of insufficient decision interpretability in existing technologies.

[0188] 2. Improved accuracy and comprehensiveness of retrieval: The fusion retrieval technology combines the advantages of vector retrieval and keyword retrieval. Vector retrieval can understand the semantics of the query and find semantically similar information, avoiding the inability of keyword retrieval to handle semantic drift; BM25 keyword retrieval can accurately hit keywords, compensating for the potential omission of key matches by vector retrieval. The fusion of the two ensures that the retrieval results include semantically relevant information without omitting key keyword matches, improving the accuracy and comprehensiveness of the retrieval and providing richer and more reliable information support for generating high-quality decision-making solutions for large models. This is achieved by designing a "vector + keyword" fusion retrieval process, effectively solving the shortcomings of existing single retrieval technologies and improving the performance of the entire RAG system.

[0189] 3. Improved quality and efficiency of emergency decision-making: Based on more interpretable decisions and more accurate and comprehensive retrieval information, the large model can generate emergency decision-making solutions that are more closely aligned with real-world scenarios. Decision-makers can understand and accept decisions more quickly, reducing decision-making time and costs, and improving the efficiency and effectiveness of forest fire emergency response.

[0190] The results generated by the technical solution in this application will be compared and evaluated with those generated by the traditional RAG retrieval solution from four core dimensions: relevance of retrieval information, adaptability of decision-making scheme, interpretability, and rationality of resource allocation. The specific judgment criteria are as follows:

[0191] 1. Relevance of Retrieved Information: Traditional RAG retrieval schemes (pure vector or pure keyword) have significant limitations: pure vector retrieval is prone to irrelevant information being mixed in due to semantic generalization; for example, searching for "mountain surface fire extinguishing measures" may lead to information related to grassland surface fire control. Pure keyword retrieval struggles to handle semantically similar but different keywords; for example, searching for "steep slope fire line control" cannot effectively match related content such as "high slope fire extinguishing strategies." This application uses a fusion retrieval of "vector + keyword," leveraging vector retrieval to capture semantic connections while precisely targeting core information through keyword retrieval. In all test scenarios, experts judged that the historical cases and contingency plans retrieved by this scheme had a higher degree of matching with the current fire situation, contained no obviously irrelevant information, and did not omit key control points, demonstrating significantly better relevance than traditional schemes.

[0192] 2. Decision-Making Applicability Dimension: Traditional RAG retrieval solutions generate decision-making solutions that are highly general but lack specificity. Traditional solutions only provide routine firefighting measures, while this application considers the priority of evacuating residents, thus improving the feasibility of the solution. This application generates decision-making solutions that are more closely aligned with the actual needs of the scenario due to its more accurate retrieval information and the integration of real-time fire data and multi-hop reasoning from knowledge graphs. For example, the generated solution clearly specifies the operational sequence of "first organizing the evacuation of residents to a safe area, then using firebreaks to block the spread of the fire, and coordinating priority support from nearby fire-fighting forces." Experts judge that the scenario adaptability of this decision-making solution far exceeds that of traditional solutions.

[0193] 3. Explainability Dimension: Traditional RAG retrieval schemes generate decision-making solutions that lack clear logical traceability, leaving experts unable to ascertain the source and basis of specific measures within the scheme. This application utilizes knowledge graphs to achieve visualized traceability of decision-making basis, clearly demonstrating the connection paths between decision-making measures and historical cases and emergency plans in all scenarios.

[0194] 4. Resource Allocation Rationality Dimension: Traditional RAG retrieval schemes are prone to configuration imbalances in resource allocation. This application combines emergency resource distribution data from the knowledge graph with resource consumption patterns from historical cases to generate a more rational resource allocation scheme. Experts judged that its resource allocation scheme better meets actual operational needs and can effectively improve emergency response efficiency.

[0195] Based on the evaluation results of the above four dimensions, the customized RAG technology method for forest fire emergency decision-making scenarios proposed in this application is superior to traditional RAG retrieval methods in terms of retrieval information quality, decision-making relevance, interpretability, and resource allocation rationality. It can better meet the professional needs of forest fire emergency decision-making and has significant technical advantages.

[0196] To facilitate understanding of the technical solutions in this application, the specific explanations of the technical terms used in this application are as follows:

[0197] RAG Technology Principle: RAG (Retrieval Augmented Generation) is an AI technology that combines retrieval and generation capabilities, integrating retrieval and generation models. When generating text, the RAG model first uses a retrieval model to retrieve information relevant to the input from a large amount of data, and then uses this retrieved information to generate the answer or text.

[0198] Application of the “5W1H” paradigm in the field of emergency response: The “5W1H” paradigm (What, Why, When, Where, Who, How) has been applied in the emergency plans of power grid companies, but this application extends it to the forest fire scenario and adds the “How much” field (such as resource demand and time requirements), thus forming the “5W2H” paradigm, and further technically structured emergency plan library.

[0199] Application of Graph RAG technology in knowledge organization: Graph RAG is an enhanced solution that integrates knowledge graph technology with traditional RAG. Its core lies in transforming unstructured text into a structured "node-edge" knowledge network. In this application, the application of Graph RAG is manifested as follows: using knowledge blocks related to forest fire emergency decision-making as nodes, edges are established through common concepts and semantic similarity to form a logically related knowledge graph. This structure not only supports accurate knowledge retrieval but also mines indirect connections between nodes through multi-hop reasoning. For example, from the "high temperature weather" node, an edge connecting to the "fire spread acceleration" node leads to the "firebreak setting" node, providing more comprehensive knowledge support for decision generation. Compared to traditional RAG, which relies solely on vector matching, Graph RAG significantly improves the interpretability of decisions and the accuracy of knowledge retrieval through the explicit relationships of the knowledge graph.

[0200] Adaptation Method for Word Segmentation Tools in the Forest Fire Domain: Given the specialized nature of Chinese forest fire emergency decision-making scenarios, the word segmentation tool used in this application must meet the precise segmentation requirements of domain-specific terms. Specifically, this includes: expanding the custom thesaurus for specialized fire-related terms and firefighting measures such as "surface fire," "crown fire," and "relay pumping"; optimizing the word segmentation algorithm to identify geographically relevant compound words such as "ridge area" and "steep slope terrain"; and correcting the word segmentation results based on contextual semantics to avoid splitting "establishing firebreaks" into irrelevant single characters or words. Through this adaptation, the keyword index library is constructed to better align with the expression habits in the forest fire domain, improving the accuracy of BM25 keyword retrieval.

[0201] Collaborative Storage Architecture of Vector and Graph Databases: This application adopts a collaborative storage model using vector databases (such as Milvus) and graph databases (such as Neo4j). The vector database stores the high-dimensional semantic vectors of knowledge blocks, supports fast vector similarity calculation, and meets semantic retrieval requirements. The graph database stores the node-edge structure and attribute information of the knowledge graph (such as node concept labels and edge weights), supporting efficient graph traversal and relation queries. The two are linked through unique identifiers for knowledge blocks. When performing fusion retrieval, the vector database returns semantically similar node IDs, and the graph database queries the corresponding node attributes and associated edge information based on these IDs, realizing collaborative calling of the two databases and balancing retrieval efficiency and knowledge association mining capabilities.

[0202] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0203] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / random access memory (RAM), magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0204] According to another aspect of the embodiments of this application, an apparatus for generating solution description information is also provided. This apparatus can be used to implement the solution description information generation method provided in the above embodiments, and will not be repeated hereafter. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0205] Figure 10 This is a structural block diagram of a device for generating solution description information according to an embodiment of this application, such as... Figure 10 As shown, the apparatus for generating the solution description information includes the following structure:

[0206] The acquisition module 102 is used to acquire real-time fire data and response plan query requests, wherein the real-time fire data is used to represent event information of an ongoing fire event, and the response plan query request is used to request a query for a response plan for the fire event.

[0207] The fusion retrieval module 104 is used to perform fusion retrieval based on the real-time fire data and the response plan query request to obtain multiple target knowledge nodes;

[0208] Reasoning module 106 is used to determine the knowledge node text associated with the multiple target knowledge nodes based on the pre-constructed knowledge graph and the multiple target knowledge nodes;

[0209] The generation module 108 is used to generate response plan description information based on the knowledge node text and the real-time fire data, wherein the response plan description information is used to represent the response plan for the fire event.

[0210] It should be noted that the acquisition module 102 in this embodiment can be used to execute the above step S202, the fusion retrieval module 104 in this embodiment can be used to execute the above step S204, the reasoning module 106 in this embodiment can be used to execute the above step S206, and the generation module 108 in this embodiment can be used to execute the above step S208.

[0211] The device for generating solution description information in this application embodiment can also perform the steps in the above method embodiments. For specific embodiments, please refer to the detailed description in the preceding method embodiments, which will not be repeated here.

[0212] It should be noted that the above modules can be implemented by software or hardware. For the latter, they can be implemented in the following ways, but are not limited to: all the above modules are located in the same processor; or, the above modules are located in different processors in any combination.

[0213] According to another aspect of the embodiments of this application, a computer-readable storage medium is provided, the computer-readable storage medium including a stored program, wherein the program executes the steps in any of the above method embodiments when it is run.

[0214] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, ROMs, RAMs, portable hard drives, magnetic disks, or optical disks.

[0215] According to another aspect of the embodiments of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor is configured to perform the steps of any of the method embodiments described above via the computer program. In an exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0216] Specific examples in this embodiment can be found in the examples described in the above embodiments and exemplary implementations, and will not be repeated here.

[0217] According to another aspect of the embodiments of this application, a computer program product is also provided, the computer program product including a computer program / instructions containing program code for performing the method shown in the flowchart.

[0218] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0219] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.

Claims

1. A method for generating solution description information, characterized in that, include: Obtain real-time fire data and response plan query request, wherein the real-time fire data is used to represent event information of an ongoing fire event, and the response plan query request is used to request a query for the response plan of the fire event; Based on the real-time fire data and the response plan query request, multiple target knowledge nodes are obtained through fusion retrieval. Based on the pre-constructed knowledge graph and the multiple target knowledge nodes, determine the knowledge node text associated with the multiple target knowledge nodes; Based on the knowledge node text and the real-time fire data, a response plan description is generated, wherein the response plan description is used to represent the response plan for the fire event.

2. The method according to claim 1, characterized in that, The process of fusing and retrieving data based on real-time fire data and response plan query requests yields multiple target knowledge nodes, including: Perform a concatenation operation on the real-time fire data and the response plan query request to obtain the concatenated text; Vector retrieval results are obtained by performing vector retrieval on the concatenated text in a pre-built vector database. Based on the concatenated text, a keyword search is performed in a pre-built keyword index library to obtain keyword search results; The multiple target knowledge nodes are determined based on the vector retrieval results and the keyword retrieval results.

3. The method according to claim 2, characterized in that, The step of performing vector retrieval in a pre-constructed vector database based on the concatenated text to obtain vector retrieval results includes: The concatenated text is vectorized to obtain the target vector; Vector retrieval is performed on the target vector in the vector database to obtain a set of vectors that match the target vector, and a set of similarities corresponding one-to-one with the set of vectors and a first set of candidate knowledge nodes corresponding one-to-one with the set of vectors are obtained, wherein the similarity between each vector in the set of vectors and the target vector satisfies a preset similarity condition. The vector retrieval results include the first set of candidate knowledge nodes and the set of similarities.

4. The method according to claim 2, characterized in that, The step of performing keyword retrieval in a pre-built keyword index based on the concatenated text to obtain keyword retrieval results includes: The concatenated text is segmented to obtain a set of target words; In the keyword index library, a keyword retrieval is performed on the set of target words to obtain a set of keywords that match the set of target word units. A second set of candidate knowledge nodes indexed by the set of keywords and a set of overlap degrees corresponding one-to-one with the second set of candidate knowledge nodes are also obtained. Each keyword in the set of keywords is the same as or contains a word in the set of target words. Each overlap degree in the set of overlap degrees is determined based on a set of keywords and a set of target words related to the candidate knowledge nodes corresponding to the second set of candidate knowledge nodes. Each keyword in the set of keywords is indexed to the corresponding candidate knowledge node. The set of target words includes words in the set of target words that are the same as or contained in each keyword in the set of keywords. The keyword retrieval results include the second group of candidate knowledge nodes and the first group of overlap.

5. The method according to claim 2, characterized in that, The step of determining the multiple target knowledge nodes based on the vector retrieval results and the keyword retrieval results includes: When the vector retrieval results include a first group of candidate knowledge nodes with a one-to-one correspondence and a set of similarities, and the keyword retrieval results include a second group of candidate knowledge nodes with a one-to-one correspondence and a set of overlaps, a deduplication operation is performed on the first group of candidate knowledge nodes and the second group of candidate knowledge nodes to obtain a third group of candidate knowledge nodes. Based on the set of similarities and the set of overlaps, the similarity and overlap of each candidate knowledge node in the third group of candidate knowledge nodes are determined. Based on the similarity and overlap of each candidate knowledge node in the third group of candidate knowledge nodes, the matching degree of each candidate knowledge node in the third group of candidate knowledge nodes is determined. K candidate knowledge nodes from the third group of candidate knowledge nodes are determined as the plurality of target knowledge nodes, where K is a preset positive integer greater than 1, and the matching degree corresponding to the K candidate knowledge nodes is greater than the matching degree corresponding to the candidate knowledge nodes other than the K candidate knowledge nodes in the third group of candidate knowledge nodes.

6. The method according to claim 5, characterized in that, The step of determining the similarity and overlap of each candidate knowledge node in the third set of candidate knowledge nodes based on the set of similarities and the set of overlaps includes: If the first part of the candidate knowledge nodes in the third group of candidate knowledge nodes have a corresponding similarity in the same group of similarities and a corresponding overlap in the same group of overlaps, then the similarity and overlap of each candidate knowledge node in the first part of the candidate knowledge nodes are respectively set to the similarity in the same group of similarities and the overlap in the same group of overlaps. If the second part of the candidate knowledge nodes in the third group of candidate knowledge nodes has a corresponding similarity in the same group of similarities but no corresponding overlap in the same group of overlaps, then the similarity and overlap of each candidate knowledge node in the second part of the candidate knowledge nodes are set to the similarity and 0 respectively in the same group of similarities. If the third part of the candidate knowledge nodes in the third group of candidate knowledge nodes does not have a corresponding similarity in the same group of similarities, but has a corresponding overlap in the same group of overlaps, then the similarity and overlap of each candidate knowledge node in the third part of the candidate knowledge nodes are set to 0 and the overlap corresponding to the same group of overlaps, respectively.

7. The method according to claim 5, characterized in that, The step of determining the matching degree of each candidate knowledge node in the third group of candidate knowledge nodes based on the similarity and overlap of each candidate knowledge node includes: The similarity and overlap of each candidate knowledge node in the third group of candidate knowledge nodes are normalized to obtain a set of normalized similarity and a set of normalized overlap. A weighted operation is performed on the set of normalized similarities and the set of normalized overlaps to obtain the matching degree corresponding to each candidate knowledge node in the third set of candidate knowledge nodes.

8. The method according to any one of claims 1 to 7, characterized in that, The step of determining the knowledge node text associated with the multiple target knowledge nodes based on the pre-constructed knowledge graph and the multiple target knowledge nodes includes: In the knowledge graph, starting from each of the plurality of target knowledge nodes, an N-hop search is performed along the edges whose weights are greater than a preset threshold to obtain at least one extended knowledge node associated with each target knowledge node, where N is a preset positive integer. The knowledge graph includes a set of knowledge nodes and a set of edges, where each edge in the set of edges connects two knowledge nodes in the set of knowledge nodes, and each edge has a predetermined edge weight. The text corresponding to each of the multiple target knowledge nodes is concatenated with the text corresponding to at least one extended knowledge node associated with each target knowledge node to obtain the knowledge node text.

9. The method according to any one of claims 1 to 7, characterized in that, The step of generating response plan description information based on the knowledge node text and the real-time fire data includes: The target prompt word is generated based on the knowledge node text, the real-time fire data, and the preset prompt word template, wherein the prompt word template is used to instruct the large language model to perform at least one preset response solution generation task; The target prompt word is input into the large language model to obtain the coping solution description information output by the large language model, wherein the coping solution description information includes the generation result of each coping solution generation task in the at least one coping solution generation task.

10. The method according to claim 9, characterized in that, The at least one response generation task includes at least one of the following: Based on the fire site topography, meteorological conditions, and fire intensity, determine the fire extinguishing route, fire extinguishing method, quantity of fire extinguishing resources, and allocation plan for fire extinguishing resources; Based on the casualties and the location of the fire, determine the rescue route, the quantity of rescue resources, and the allocation plan for the rescue resources.

11. The method according to any one of claims 1 to 7, characterized in that, The real-time fire data includes at least one of the following: fire terrain, meteorological conditions, fire intensity, casualties, and fire location.

12. The method according to any one of claims 1 to 7, characterized in that, Before obtaining real-time fire data and response plan query requests, the method further includes: Obtain historical response plan description information, wherein the historical response plan description information is used to represent historical response plans corresponding to at least one fire event that has occurred in a predetermined forest area; The historical response plan description information is cleaned and structured to obtain structured information; The knowledge graph, vector database, and keyword index are constructed based on the structured information.

13. The method according to claim 12, characterized in that, The process of cleaning and structuring the historical response plan description information to obtain structured information includes: The historical response plan description information is cleaned to remove invalid and duplicate data, resulting in cleaned information. The structured information is obtained by extracting the value corresponding to each field in the tuple template from the cleaned data according to the preset tuple template.

14. The method according to claim 13, characterized in that, The preset multi-group template includes at least two of the following fields: fire type, fire cause, time and season of occurrence, fire location and terrain, division of command and rescue responsibilities, firefighting strategy, and resource requirement.

15. The method according to claim 12, characterized in that, The construction of the knowledge graph, vector database, and keyword index based on the structured information includes: The structured information is divided into a group of information blocks, wherein the total number of characters in each information block in the group of information blocks is less than or equal to a first preset value, and the number of overlapping characters between two adjacent information blocks in the group of information blocks is equal to a second preset value. The knowledge graph is constructed based on the set of information blocks, wherein the knowledge graph includes a set of knowledge nodes and a set of edges. Each knowledge node in the set of knowledge nodes is used to represent a corresponding information block in the set of information blocks. Each edge in the set of edges connects two knowledge nodes in the set of knowledge nodes. Each edge has a predetermined edge weight. Each information block in the set of information blocks is vectorized to obtain a vector corresponding to each information block. The vector database includes a set of vectors, which includes a vector corresponding to each information block in the set of information blocks. Each vector in the set of vectors corresponds to a knowledge node in the knowledge graph. Each information block in the set of information blocks is segmented to obtain a set of words corresponding to each information block. Each word in the set of words is identified as a keyword, resulting in multiple keywords. Each keyword is indexed to a knowledge node in the knowledge graph. The frequency or number of times each keyword appears in the information blocks corresponding to each knowledge node in the knowledge graph is determined. The keyword index library includes the multiple keywords, the knowledge nodes indexed to each keyword, and the frequency or number of times each keyword appears in the multiple information blocks corresponding to each knowledge node.

16. The method according to any one of claims 1 to 7, characterized in that, After generating response plan description information based on the knowledge node text and the real-time fire data, the method further includes: Based on the knowledge graph, generate inference information related to the description information of the response plan; The knowledge graph includes a set of knowledge nodes and a set of edges. Each edge in the set of edges connects two knowledge nodes in the set of knowledge nodes. Each edge has a predetermined edge weight. The set of knowledge nodes includes the plurality of target knowledge nodes. The reasoning information includes the plurality of target knowledge nodes, the node attributes of each of the plurality of target knowledge nodes, the edges connected to each of the plurality of target knowledge nodes, and the edge weights of the edges.

17. An apparatus for generating solution description information, characterized in that, include: The acquisition module is used to acquire real-time fire data and response plan query requests, wherein the real-time fire data is used to represent event information of an ongoing fire event, and the response plan query requests are used to request responses to the fire event. The fusion retrieval module is used to perform a fusion retrieval based on the real-time fire data and the response plan query request to obtain multiple target knowledge nodes; The reasoning module is used to determine the knowledge node text associated with the multiple target knowledge nodes based on the pre-constructed knowledge graph and the multiple target knowledge nodes; The generation module is used to generate response plan description information based on the knowledge node text and the real-time fire data. The response plan description information is used to represent the response plan for the fire event.

18. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 16.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of the method according to any one of claims 1 to 16.

20. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 16.

Images