A method, equipment, and medium for extracting forest fire operational elements based on natural language understanding technology.

Abstract

This invention discloses a method, device, and medium for extracting forest fire operational elements based on natural language understanding technology. The method includes: constructing a forest fire operational knowledge graph; extracting key features from multimodal data; fusing the key features of the multimodal data; extracting forest fire operational elements based on a trained semantic understanding model; and converting the extracted operational elements into structured data for system integration and application. This invention significantly improves the efficiency and accuracy of forest fire operational element extraction, realizing the function of automatically and accurately extracting key operational elements from multi-source heterogeneous data. Through a dynamic weight adjustment mechanism, the system can automatically adjust the contribution of each modality feature according to the forest fire development status, making the element extraction more aligned with actual operational needs. This effectively solves the problems of one-sided and static element extraction in existing technologies, providing high-quality data support for forest fire monitoring and decision-making.

Description

Technical Field

[0001] This invention relates to the field of forest fire business element extraction technology, and in particular to a method, equipment and medium for extracting forest fire business elements based on natural language understanding technology. Background Technology

[0002] Current forest fire monitoring systems primarily rely on video surveillance and smoke / fire recognition technologies, using AI image algorithms to detect smoke and flames in forests. However, these technologies only provide visual evidence of fires and cannot provide a comprehensive understanding of forest fire operations. In practice, forestry management departments need to extract key operational elements from multi-source heterogeneous data (including field report texts, meteorological data, monitoring records, satellite imagery, etc.), such as the time, location, scale, fire development, and meteorological conditions of the fire, to support real-time early warning and decision-making. However, manually extracting these elements is not only inefficient but also prone to information omissions or errors due to subjective factors, making it difficult to meet the high demands of modern forestry fire prevention management for timeliness and accuracy.

[0003] While existing Natural Language Processing (NLP) technologies have made significant progress in general domains, they face three major challenges in the vertical scenario of forest fire prevention: First, high domain knowledge barriers. The cross-use of forestry terminology (such as "humus layer thickness" and "canopy closure") and meteorological terminology (such as "daily relative humidity range") leads to semantic drift in general models, for example, misclassifying "dry understory" as "soil drought." Second, difficulty in multimodal data fusion. Text data is often associated with heterogeneous data such as meteorological values, geographic coordinates, and remote sensing images, but existing methods lack cross-modal semantic alignment mechanisms, making it difficult to locate specific risk areas from text such as "increased wind on the north slope." Third, weak dynamic temporal reasoning. Fire risk development is affected by real-time changes in meteorological conditions; for example, "gusts of wind up to level 7 in the next 3 hours" needs to be dynamically extrapolated in conjunction with the current fire situation, but traditional NLP models cannot handle such temporal dependencies, leading to a disconnect between warning results and actual risks.

[0004] In recent years, although some studies have attempted to apply natural language processing technology to specific vertical fields, a systematic solution is still lacking in the special scenario of forest fire prevention and control. Existing technologies are mostly limited to the simple recognition of general entities such as people, locations, and times, failing to deeply integrate with the forest fire business knowledge system, and also lacking the ability to verify the logical consistency of the extracted results. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a method, device, and medium for extracting forest fire business elements based on natural language understanding technology, thereby enabling efficient and accurate processing and analysis of forest fire-related text information and effectively solving many problems of existing technologies in forest fire prevention and control scenarios.

[0006] This invention provides a method, device, and medium for extracting forest fire business elements based on natural language understanding technology.

[0007] First aspect: A method for extracting forest fire operational elements based on natural language understanding technology, including:

[0008] S1. Construct a knowledge graph for forest fire operations;

[0009] S2. Receive multimodal data, perform preprocessing, and extract key features from the multimodal data;

[0010] S3. Fuse key features of multimodal data to form a unified multimodal feature representation;

[0011] S4. Extracting forest fire business elements based on the trained semantic understanding model;

[0012] S5. Convert the extracted business elements into structured data for system integration and application.

[0013] In one embodiment of the present invention, step S1 includes the following steps:

[0014] S11. Collect historical forest fire reports, meteorological data, and multi-source operational data from geographic information systems;

[0015] S12. Extract key business elements from the data to form an entity set;

[0016] S13. Define the semantic relationships between entities and establish logical connections;

[0017] S14. Organize entities and logical relationships into a knowledge graph structure, represented as G=(E, R, T), where: E is the set of entities, R is the set of relations, and T is the set of triples.

[0018] In one embodiment of the present invention, the multimodal data in S2 includes text data, image data, and geographic information data, wherein:

[0019] Text data preprocessing includes:

[0020] We employ a word segmentation method specific to the forest fire business domain to perform word segmentation processing on the text data;

[0021] A forest fire business terminology dictionary was used to perform entity recognition on word segmentation to identify business elements.

[0022] Use sequence labeling technology to label business elements;

[0023] The labeled business elements are vectorized to obtain text feature vectors;

[0024] Image data preprocessing includes:

[0025] Fire area detection is performed on video footage to identify the range and intensity of the fire, and each image feature is vectorized to obtain the image feature vector.

[0026] The geographic information preprocessing process includes:

[0027] Extract the geographical location coordinates and topographic information features of the fire site, and vectorize each geographical feature to obtain the geographical feature vector;

[0028] In one embodiment of the present invention, step S3, which fuses key features of multimodal data to form a unified multimodal feature representation, includes:

[0029] S31. Normalize the text feature vector, image feature vector, and geographic feature vector respectively to generate the corresponding standardized feature vector, expressed by the formula:

[0030]

[0031] Where X is the original feature value, μ is the feature mean, and σ is the feature standard deviation;

[0032] S32. A dynamic weight adjustment mechanism is adopted to dynamically adjust the weights of each modal feature according to the forest fire development status. The weight allocation formula is as follows:

[0033]

[0034] Where α is the text feature weight coefficient, β is the image feature weight coefficient, γ is the geographic feature weight coefficient, and s is the forest fire development stage;

[0035] S33. A weighted fusion method is used to fuse multimodal features. The fusion formula is expressed as:

[0036]

[0037] Where M is the fused multimodal feature vector, T' is the standardized text feature vector, I' is the standardized image feature vector, and G' is the standardized geographic feature vector.

[0038] In one embodiment of the present invention, the semantic understanding model is built based on the semantic understanding model of Transformer, and the model architecture is as follows:

[0039] A multi-layer Transformer encoder is used, with each layer containing a multi-head self-attention mechanism and a feedforward neural network;

[0040] A multi-task learning framework is adopted to extract multiple business elements simultaneously;

[0041] In one embodiment of the present invention, the semantic understanding model of the Transformer adopts a dynamic context-aware mechanism, introducing temporal location encoding into the multi-head self-attention mechanism to represent the stages of forest fire development, expressed by the formula:

[0042]

[0043] in, Encode the time location vector. This is the stage of forest fire development. For dimensional indexing, For model dimensions.

[0044] In one embodiment of the present invention, the multi-task learning framework for the semantic understanding model of the Transformer includes:

[0045] An independent output header is designed for each business element. A multi-task learning approach is adopted, and a loss function is used to simultaneously optimize the extraction of multiple elements. The loss function is expressed as:

[0046]

[0047] in, This is the total loss function; Extracting loss for time elements; Extracting losses for location elements; Losses are extracted for scale factors; Extracting losses from factors contributing to the development of the fire; Losses were extracted for meteorological conditions. These are the weighting coefficients for each element's task.

[0048] In one embodiment of the present invention, step S5, converting the extracted business elements into structured data, includes:

[0049] The confidence score for each feature is calculated using the following formula:

[0050]

[0051] in, This is the confidence vector. This is the score vector output by the model. The function is a normalization function to ensure that the confidence level is in the range [0, 1].

[0052] Structured data is generated according to a predefined JSON format, which contains all key business elements and their confidence levels.

[0053] Second aspect: An electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, performs the steps of the method provided in the first aspect.

[0054] Third aspect: A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect.

[0055] The beneficial effects of this invention are:

[0056] 1. The method of this invention significantly improves the efficiency and accuracy of forest fire business element extraction by constructing a knowledge graph specifically for forest fire operations and designing a dynamic context-aware semantic understanding model. Compared with existing technologies, this invention solves the problem of insufficient semantic understanding of forest fire operations by general natural language processing models, and realizes the function of automatically and accurately extracting key business elements from multi-source heterogeneous data. Through a dynamic weight adjustment mechanism, the system can automatically adjust the contribution of each modality feature according to the forest fire development status (initial stage, development stage, and extinction stage), making element extraction more in line with actual business needs, effectively solving the problems of one-sided and static element extraction in existing technologies, and providing high-quality data support for forest fire monitoring and decision-making.

[0057] 2. The method of this invention demonstrates significant technical advantages in practical applications. Testing and verification show that the element extraction accuracy of this method can reach over 95%, a significant improvement compared to existing technologies. The element extraction efficiency is reduced from 2-3 hours for manual processing to 5-10 minutes for automatic processing, an efficiency increase of 12-36 times. Simultaneously, the structured data generated by this invention can be directly integrated into forest fire early warning systems and decision support platforms without additional data conversion and processing, improving system integration efficiency. In actual forest fire emergency response, the method of this invention can provide timely and accurate fire information, offering crucial support to decision-makers, significantly improving emergency response speed and effectiveness, and effectively reducing losses caused by forest fires.

[0058] 3. This invention deeply integrates forest fire operational knowledge with natural language understanding technology, constructing a dedicated semantic analysis framework for forest fire operational scenarios. Through a dynamic context-aware mechanism, the system can automatically adjust the focus of element extraction according to the stage of forest fire evolution, achieving accurate understanding of the forest fire situation. Simultaneously, the multimodal feature fusion method designed in this invention effectively integrates multi-source data such as text, images, and geographic information, solving the data silo problem and making the system's perception of the forest fire situation more comprehensive and accurate. This technological breakthrough not only fills the technical gap in intelligent extraction of forestry operational elements but also promotes the deep application of natural language processing technology in vertical industry scenarios, providing crucial technical support for the construction of smart forestry.

[0059] 4. The implementation and application of the method of this invention significantly improves the level of intelligence in forestry fire prevention and management. In actual deployment, the system can extract key elements of forest fires in real time and accurately, providing high-precision data support for fire situation assessment, resource allocation, and response plan formulation, increasing the accuracy of forest fire early warning by more than 35% and shortening emergency response time by more than 40%. At the same time, the method of this invention reduces the reliance of forestry management departments on manual element extraction, alleviating the burden on staff and improving work efficiency. Furthermore, the structured output format seamlessly integrates with existing forest fire monitoring systems without requiring additional interface development, reducing system integration costs and demonstrating significant economic benefits and broad application prospects. Attached Figure Description

[0060] Figure 1 This is a schematic flowchart of the method of the present invention;

[0061] Figure 2 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0062] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0063] The current extraction of forest fire business elements faces the following problems: high domain knowledge barriers, lack of cross-modal semantic alignment mechanisms, and weak dynamic temporal reasoning, resulting in a disconnect between early warning results and actual risks.

[0064] To address the aforementioned issues, this invention provides a method for extracting forest fire operational elements based on natural language understanding technology. This method automatically and accurately extracts forest fire operational elements from multi-source heterogeneous data, solving problems such as incomplete and inaccurate element extraction and the inability to dynamically adapt to forest fire development in existing technologies. It provides efficient and reliable intelligent technical support for forestry fire prevention management. By constructing a dedicated knowledge graph for forest fire operations, designing a dynamic weight adjustment mechanism, achieving dynamic context-aware semantic understanding, and generating directly integrateable structured data—containing complete operational elements and confidence information—this method can be directly integrated into forest fire monitoring and decision-making systems without additional processing, significantly improving data utilization efficiency.

[0065] Example 1:

[0066] like Figure 1 As shown in the figure, this embodiment discloses a method for extracting forest fire business elements based on natural language understanding technology, including the following steps:

[0067] S1. Construct a knowledge graph for forest fire operations.

[0068] A dedicated knowledge graph for forest fire operations is constructed as the foundational framework for semantic understanding. This knowledge graph contains three core entities:

[0069] Fire event entity (including attributes such as time, location, scale, fire development, and type of fire source);

[0070] Meteorological conditions (including attributes such as wind force, temperature, humidity, and precipitation);

[0071] Geographic information entities (including attributes such as altitude, vegetation type, and terrain slope) and the semantic relationships between them (such as influence, association, and determination).

[0072] The specific construction method includes: collecting multi-source business data such as historical forest fire reports, meteorological data, and geographic information system data; extracting key business elements from the data to form an entity set; then, defining semantic relationships between entities, establishing logical connections, organizing entities and logical relationships into a knowledge graph structure, and forming a unified semantic representation. The forest fire business knowledge graph is represented as G=(E, R, T).

[0073] Where: E is the entity set, containing all entities related to forest fire operations; R is the relation set, containing semantic associations between entities; and T is the set of triples, where each triple is (E... i R j E k ), representing entity E i With entity E k Through logical relationship R j Connected.

[0074] The innovation of the knowledge graph lies in its customization for the characteristics of forest fire operations. Compared with general knowledge graphs, it contains more accurate semantic relationships and professional terms related to forest fire operations, providing targeted support for subsequent semantic understanding.

[0075] S2. Receive multimodal data and preprocess it to extract key features of the multimodal data; the multimodal data includes text data, image data and geographic information data.

[0076] The text data preprocessing includes: first, using a forest fire business domain-specific word segmentation method to segment the text; then, applying a forest fire business terminology dictionary to perform entity recognition of the segmented text, identifying business elements such as "time," "location," and "scale"; next, using sequence labeling technology to tag the segmented text, labeling the business element types, etc.; and finally, vectorizing the text to obtain the text feature vector.

[0077] Image data preprocessing includes: detecting fire areas in video footage, identifying the fire's range and intensity characteristics, and vectorizing each image feature to obtain image feature vectors.

[0078] Geographic information preprocessing includes: extracting the geographical coordinates (longitude and latitude) of the fire site, obtaining the topographic information (elevation, slope, vegetation type, etc.) of the fire area, and vectorizing the geographic features to obtain geographic feature vectors.

[0079] In the process of extracting key features from multimodal data, based on the forest fire business knowledge graph, semantic association and fusion are performed on identified business elements such as "time," "location," and "scale," as well as the fire's extent and intensity, geographical coordinates, and topographic information of the fire area. Various predefined relationships and rules within the forest fire business knowledge graph are used to perform correlation analysis on different types of features and business elements. For example, the "time" identified in the text is matched with the time features of the fire's occurrence in the image to determine the accurate time of the forest fire. "Location" is combined with the fire's geographical coordinates and topographic information to construct a complete geographical scene of the forest fire's occurrence. For the business element of "scale," a comprehensive evaluation is performed, combining the fire's location characteristics and regional features, to determine the actual size of the forest fire.

[0080] Simultaneously, the reasoning capabilities of knowledge graphs are leveraged to uncover the potential information hidden behind these features and elements. For example, based on the terrain information and key features of the fire, the possible direction and speed of forest fire spread can be inferred. Through semantic association and fusion of multimodal data features, a more comprehensive and accurate set of forest fire operational element information is formed, providing a solid data foundation for subsequent forest fire operational element extraction and analysis.

[0081] S3. Fuse key features of multimodal data to form a unified multimodal feature representation.

[0082] First, the text feature vector, image feature vector, and geographic feature vector are normalized to ensure that the magnitudes of each feature are consistent, generating corresponding standardized feature vectors, expressed by the formula:

[0083]

[0084] Where X is the original feature value, μ is the feature mean, and σ is the feature standard deviation;

[0085] Then, a dynamic weight adjustment mechanism is adopted to dynamically adjust the weights of each modal feature according to the forest fire development status (initial stage, development stage, and extinction stage). The weight allocation formula is as follows:

[0086]

[0087] Where α is the text feature weight coefficient, β is the image feature weight coefficient, γ is the geographic feature weight coefficient, and s is the forest fire development stage (s=1 represents the initial stage, s=2 represents the development stage, and s=3 represents the extinguishing stage).

[0088] Finally, a weighted fusion method is used to fuse the multimodal features. The fusion formula is expressed as follows:

[0089]

[0090] Where M is the fused multimodal feature vector, T' is the standardized text feature vector, I' is the standardized image feature vector, and G' is the standardized geographic feature vector.

[0091] S4. Extract forest fire business elements based on the trained semantic understanding model.

[0092] This embodiment constructs a semantic understanding model based on Transformer to extract business elements related to forest fires. The model architecture includes: a multi-layer Transformer encoder with multi-modal feature vectors as input. Each Transformer layer contains a multi-head self-attention mechanism and a feedforward neural network. The model is designed as a multi-task learning framework to simultaneously complete multiple business element extraction tasks.

[0093] The dynamic context-aware mechanism includes introducing time-location encoding into the self-attention mechanism to represent the stages of wildfire development. The time-location encoding formula is as follows:

[0094]

[0095] in, Encode the time location vector. The stages of forest fire development are t=1 for the initial stage, t=2 for the development stage, and t=3 for the extinguishing stage. For dimensional indexing, For model dimensions.

[0096] The Transformer semantic understanding model's multi-task learning framework includes: designing an independent output head for each business element, employing a multi-task learning approach, and utilizing a loss function to simultaneously optimize the extraction of multiple elements. The loss function is expressed as:

[0097]

[0098] in, This is the total loss function; Extracting loss for time elements; Extracting losses for location elements; Losses are extracted for scale factors; Extracting losses from factors contributing to the development of the fire; Losses were extracted for meteorological conditions. These are the weighting coefficients for each element's task.

[0099] This method integrates the forest fire development status into the model's context-aware mechanism, enabling the model to dynamically adjust the semantic understanding focus according to the forest fire evolution stage, and achieve accurate extraction of key elements at different stages.

[0100] S5. Convert the extracted business elements into structured data for system integration and application.

[0101] This invention converts extracted business elements into structured data, facilitating system integration and use, including element extraction and confidence calculation.

[0102] First, confidence scores are calculated. The model outputs the extraction results for each business element, and a confidence score is calculated for each element, representing the reliability of the extraction results. The confidence score calculation formula is as follows:

[0103]

[0104] in, This is the confidence vector. This is the score vector output by the model. The function is a normalization function to ensure that the confidence level is in the range [0, 1].

[0105] Then, structured data is generated according to a predefined JSON format, which contains all key business elements and their confidence levels.

[0106] Finally, the system is integrated and applied. The structured data is output through a standard API interface. The forest fire early warning system and decision support platform automatically parse and use the structured data, which is directly used for fire situation assessment, resource allocation, and response plan formulation.

[0107] The structured data generated using this method contains complete business elements and confidence information, and can be directly integrated into the forest fire monitoring and decision-making system without additional processing, significantly improving data utilization efficiency.

[0108] Through the above steps, the method in this embodiment achieves the function of automatically and accurately extracting forest fire business elements from multi-source heterogeneous data, solving the problems of incomplete and inaccurate element extraction and inability to dynamically adapt to the development status of forest fires in existing technologies, and providing efficient and reliable intelligent technical support for forestry fire prevention and management. The innovation of this invention lies in constructing a dedicated knowledge graph for forest fire business, designing a dynamic weight adjustment mechanism, realizing dynamic context-aware semantic understanding, and generating structured data that can be directly integrated.

[0109] Application examples:

[0110] This application example uses a forest fire incident in a mountainous area of ​​XX City, XX Province on October 10, 2025 as a case study to illustrate the application process of the method of this invention in a real-world scenario. The forest fire occurred at 14:30 on that day, located at 29.5°N, 110.3°E, at an altitude of 500 meters, covered by coniferous forest vegetation, with a slope of 20 degrees, covering an area of ​​50 hectares, and spreading northeastward. The meteorological conditions were wind force 3-4 and a temperature of 25°C.

[0111] First, a forest fire business knowledge graph can be constructed. This knowledge graph can be built based on historical data from the forestry department and includes three core entities:

[0112] The physical elements of a fire incident (including time, location, scale, fire development, and fire source type attributes).

[0113] Meteorological conditions (including wind force, temperature, humidity, and precipitation attributes);

[0114] Geographic information entities (including altitude, vegetation type, and slope attributes).

[0115] Furthermore, key semantic relationships between entities have been established, such as the association rules between "fire spread direction" and "meteorological conditions," and the association rules between "fire area" and "vegetation type." The knowledge graph has covered all business elements, providing accurate basis for subsequent semantic understanding.

[0116] Then, multimodal data preprocessing is performed. The system receives three types of raw data:

[0117] Text report: "At 14:30 on October 10, a forest fire broke out in XX District, XX City, with a fire area of ​​50 hectares. The fire is spreading northeastward, with wind force of 3-4, temperature of 25℃, in a coniferous forest area, at an altitude of 500 meters and a slope of 20 degrees."

[0118] Video surveillance footage: showing smoke and flame characteristics in the fire area;

[0119] Geographic Information System (GIS) data: provides information on fire location coordinates, elevation, slope, and vegetation type;

[0120] After processing: text data is used to identify business elements such as "time", "scale" and "direction of spread"; image data is used to extract the fire range and intensity features; and geographic data is used to obtain information such as coordinates, altitude and vegetation type to generate standardized feature vectors for subsequent fusion.

[0121] Then, multimodal feature fusion is performed: the system determines that the forest fire is in the development stage (stage identifier s=2), and calculates dynamic weights:

[0122] Text weight α = 1 - (2 / 3) = 1 / 3

[0123] Image weight β = 2 / 6 = 1 / 3

[0124] Geographic weight γ = 2 / 6 = 1 / 3

[0125] After normalizing the three types of features, the formula is as follows: The data is then fused to generate a unified feature vector. The fusion result ensures that text, images, and geographic information contribute appropriately according to the fire's development stage.

[0126] Then, context-aware semantic understanding is performed dynamically.

[0127] The fused features are input into the Transformer model, which perceives the "development phase" state (t=2) through temporal location encoding, achieving dynamic contextual understanding. The model simultaneously extracts all business elements, and the output is as follows:

[0128] Fire time: 14:30 on October 10, 2023 (98% confidence level)

[0129] Location information: XX District, XX City, XX Province, altitude 500 meters, vegetation type is coniferous forest (confidence level 95%).

[0130] Fire size: 50 hectares (92% confidence level)

[0131] Fire development: Spreading northeastward, currently in the development stage (96% confidence level)

[0132] Meteorological conditions: Wind force 3-4, temperature 25℃ (94% confidence level)

[0133] The model accurately identifies the correlation between "fire spreading northeast" and "wind force 3-4", avoiding misjudgment of technical terms by general models.

[0134] Finally, structured business elements are generated.

[0135] The system converts the extracted results into structured data, presenting complete business information in natural language. This structured data is transmitted in real time to the forest fire early warning system via a standard interface. The system automatically parses the data and generates a fire situation map, allowing decision-makers to obtain complete fire information within 5 minutes. Compared to traditional manual processing (which takes 2-3 hours), efficiency is improved by 36 times; the accuracy rate of element extraction is 95%, 30 percentage points higher than the industry average of 65%. In this incident, the precise fire spread direction and meteorological conditions data provided by the system enabled emergency teams to deploy to the northeast direction 15 minutes in advance, effectively controlling the spread of the fire and reducing losses by approximately 30%.

[0136] This application example has been fully verified. The technical solution achieves an accuracy rate of over 95% in extracting elements in real forest fire scenarios, dynamically adapts to the development of forest fires, and the structured data can be directly used for decision support, significantly improving the efficiency of forestry fire prevention.

[0137] The present invention also provides an electronic device, Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 2 As shown, the electronic device may include a processor, a communications interface, memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory, for example, to execute the following method:

[0138] S1. Construct a knowledge graph for forest fire operations;

[0139] S2. Receive multimodal data, perform preprocessing, and extract key features from the multimodal data;

[0140] S3. Fuse key features of multimodal data to form a unified multimodal feature representation;

[0141] S4. Extracting forest fire business elements based on the trained semantic understanding model;

[0142] S5. Convert the extracted business elements into structured data for system integration and application.

[0143] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0144] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments, including, for example:

[0145] S1. Construct a knowledge graph for forest fire operations;

[0146] S2. Receive multimodal data, perform preprocessing, and extract key features from the multimodal data;

[0147] S3. Fuse key features of multimodal data to form a unified multimodal feature representation;

[0148] S4. Extracting forest fire business elements based on the trained semantic understanding model;

[0149] S5. Convert the extracted business elements into structured data for system integration and application.

[0150] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0151] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, 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 can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A forest fire business element extraction method based on natural language understanding technology, characterized by, include: S1. Construct a knowledge graph for forest fire operations; S2. Receive multimodal data and preprocess it to extract key features of the multimodal data and generate text feature vectors, image feature vectors and geographic feature vectors respectively. The multimodal data includes text data, image data and geographic information data. S3. Fuse key features of multimodal data to form a unified multimodal feature representation, including: S31. Normalize the text feature vector, image feature vector, and geographic feature vector respectively to generate the corresponding standardized feature vector, expressed by the formula: Where X is the original feature value, μ is the feature mean, and σ is the feature standard deviation; S32. A dynamic weight adjustment mechanism is adopted to dynamically adjust the weights of each modal feature according to the forest fire development status. The weight allocation formula is as follows: Where α is the text feature weight coefficient, β is the image feature weight coefficient, γ is the geographic feature weight coefficient, and s is the forest fire development stage, with s=1 representing the initial stage, s=2 representing the development stage, and s=3 representing the extinguishing stage. S33. A weighted fusion method is used to fuse multimodal features. The fusion formula is expressed as: Where M is the fused multimodal feature vector, T' is the standardized text feature vector, I' is the standardized image feature vector, and G' is the standardized geographic feature vector; S4. Extracting forest fire business elements based on the trained semantic understanding model; The semantic understanding model is built upon the Transformer semantic understanding model, and its architecture is as follows: A multi-layer Transformer encoder is used, with multimodal feature vectors as input. Each layer contains a multi-head self-attention mechanism and a feedforward neural network. A multi-task learning framework is adopted to extract multiple business elements simultaneously; The semantic understanding model of the Transformer adopts a dynamic context-aware mechanism, introducing temporal and positional encoding into the multi-head self-attention mechanism to represent the stages of wildfire development, as expressed by the formula: in, This is a time-location encoded vector, where t represents the stage of wildfire development: t=1 indicates the initial stage, t=2 indicates the development stage, and t=3 indicates the extinguishing stage. For dimensional indexing, For model dimensions; S5. Convert the extracted business elements into structured data for system integration and application.

2. The forest fire business element extraction method based on natural language understanding technology according to claim 1, characterized in that, S1 includes the following steps: S11. Collect historical forest fire reports, meteorological data, and multi-source operational data from geographic information systems; S12. Extract key business elements from the data to form an entity set; S13. Define the semantic relationships between entities and establish logical connections; S14. Organize entities and logical relationships into a knowledge graph structure, represented as G=(E, R, T), where: E is the set of entities, R is the set of relations, and T is the set of triples. 3.The forest fire business element extraction method based on natural language understanding technology according to claim 1, wherein, In S2: Text data preprocessing includes: We employ a word segmentation method specific to the forest fire business domain to perform word segmentation processing on the text data; A forest fire business terminology dictionary was used to perform entity recognition on word segmentation to identify business elements. Use sequence labeling technology to label business elements; The labeled business elements are vectorized to obtain text feature vectors; Image data preprocessing includes: Fire area detection is performed on video footage to identify the range and intensity of the fire, and each image feature is vectorized to obtain the image feature vector. The geographic information preprocessing process includes: Extract the geographical coordinate features and topographic information features of the fire site, and vectorize each geographical feature to obtain the geographical feature vector.

4. The forest fire business element extraction method based on natural language understanding technology according to claim 1, characterized in that, The semantic understanding model multi-task learning framework of the Transformer includes: An independent output header is designed for each business element. A multi-task learning approach is adopted, and a loss function is used to simultaneously optimize the extraction of multiple elements. The loss function is expressed as: in, This is the total loss function; Extracting loss for time elements; Extracting losses for location elements; Losses are extracted for scale factors; Extracting losses from factors contributing to the development of the fire; Losses were extracted for meteorological conditions. These are the weighting coefficients for each element's task.

5. The method for extracting forest fire operational elements based on natural language understanding technology according to claim 1, characterized in that, The S5 step involves converting the extracted business elements into structured data, including: The confidence score for each feature is calculated using the following formula: in, This is the confidence vector. This is the score vector output by the model. The function is a normalization function to ensure that the confidence level is in the range [0, 1]. Structured data is generated according to a predefined JSON format, which contains all key business elements and their confidence levels. Structured data is output through a standard API interface. The forest fire early warning system and decision support platform automatically parse and use the structured data, which is then directly used for fire situation assessment, resource allocation, and response plan formulation.

6. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the forest fire business element extraction method based on natural language understanding technology as described in any one of claims 1 to 5.

7. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the forest fire business element extraction method based on natural language understanding technology as described in any one of claims 1 to 5.

Images