Intelligent disaster emergency scheme generation method and system

Abstract

The invention discloses an intelligent disaster emergency scheme generation method, which comprises the steps of configuring a multi-channel data interface to access multi-source data, identifying named entities through a deep learning model, constructing a unified disaster feature space, and outputting a standardized disaster feature vector; inputting the vector into an emergency domain knowledge graph to perform double-engine reasoning and rule reasoning to generate a compliance necessary action set, and performing case reasoning to generate a supplementary action set based on similarity retrieval; after the action sets are combined, space-time constraint verification and conflict resolution are carried out, and after serialization, a large model output scheme is input through PromptEngine. According to the method, unified convergence and standardization of multi-source data are realized, dynamic disaster situations and static knowledge are linked, regulation rigid constraint and disposal flexible adaptation are balanced, and scientific, compliant, feasible and efficient decision support is provided for emergency command.

Description

Technical Field

[0001] This invention belongs to the field of computer application technology, specifically a method and system for generating intelligent disaster emergency response plans. Background Technology

[0002] In the field of natural disaster emergency response, scientific, compliant, and executable action plans are the central nervous system of disaster relief command. Currently, various levels of government have established relatively complete emergency response plan systems, but these plans are mainly stored in static, unstructured text formats (such as PDF documents and paper compilations). The generation of plans after a disaster mainly relies on manual expert review or simple keyword searches (such as "magnitude 6 earthquake"), directly retrieving entire historical documents as the basis for command. However, this "static document management" model exposes deep-seated response mechanism failures when facing sudden disasters with highly time-varying and non-linear characteristics, mainly in the following two dimensions.

[0003] Existing contingency plans are based on fixed procedures pre-set in idealized scenarios. However, the evolution of real disasters is a dynamic process, influenced in real time by multiple variables such as weather, geography, and population distribution. Because current technology lacks a mechanism to deconstruct unstructured documents into the smallest indivisible execution units, the system cannot fine-grainedly reorganize the plan content based on real-time sensor data (such as road closure information and rainfall changes). This results in generated plans often containing a large number of redundant or even invalid instructions (e.g., recommending vehicle transport in areas where roads are already blocked), failing to achieve accurate mapping between situational awareness and action instructions.

[0004] Existing decision support systems typically employ only a single reasoning model: either rigid matching based on a rule-based system or similarity retrieval based on historical cases. In complex disaster scenarios, pure rule-based reasoning struggles to cover all emergencies, while pure case-based reasoning is prone to overlooking current legal and regulatory constraints by simply applying historical experience. The lack of a hybrid reasoning architecture that simultaneously integrates rigid legal constraints with the flexible reuse of historical experience results in solutions that are either too rigid to be implemented or, while flexible, contain compliance risks, failing to meet the dual standards of scientific rigor and legality required for command and decision-making. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for generating intelligent disaster emergency response plans, so as to solve the following technical problems mentioned in the background art: In existing technologies, there are challenges in the effective aggregation and standardized representation of dynamic, multi-source, heterogeneous disaster data, as well as insufficient precise adaptation between static emergency knowledge and dynamic disaster characteristics. This makes it difficult to balance the rigid constraints of regulations with the flexible needs of handling complex disasters.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for generating intelligent disaster emergency response plans includes the following steps: S1 is configured with a multi-channel data interface to access sensor time-series data, weather forecasts, and on-site voice streams; it uses a deep learning model to perform named entity recognition, extracting entity triples related to disaster type, disaster intensity, and environmental constraints; it constructs a unified disaster feature space, mapping discrete information into high-dimensional numerical vectors, and outputs standardized disaster feature vectors. S2, input the disaster feature vector into the pre-constructed emergency domain knowledge graph and perform dual-engine reasoning. The emergency domain knowledge graph is constructed using the smallest granularity action unit of the deconstructed traditional plan as the basic unit. Dual-engine reasoning includes: Through a rule-based reasoning engine, preset legal provisions are transformed into IF-THEN production rules and a rule base is established. The rule base is traversed to determine whether the disaster feature vector satisfies the rule antecedents, and a set of mandatory actions that satisfy the rule antecedents is generated. By using a case-based reasoning engine, effective handling measures are retrieved from the emergency domain knowledge graph based on the similarity between disaster feature vectors and historical cases, and a supplementary action set is generated. S3 merges the mandatory action set and the supplementary action set to obtain the initial action set, performs spatiotemporal constraint verification on the initial action set and resolves conflicting instructions; serializes the verified action list, inputs it into the large model through PromptEngineering and outputs the disaster emergency response plan.

[0007] Furthermore, the calculation of the disaster feature vector is as follows: Among them, for disaster types ( (), using one-hot encoding for identification; Regarding the intensity of the disaster ( ) and environmental parameters ( ), and then normalize it.

[0008] Furthermore, the rule-based reasoning engine employs the Drools rule engine and / or the SWRL language to describe hard regulatory provisions.

[0009] Furthermore, the triplet structure of the emergency response domain knowledge graph is defined as <action, relationship, constraint>.

[0010] Furthermore, the formula for calculating the similarity between the disaster feature vector and historical cases is as follows: in: Disaster vector , Historical case vectors ; Preset weights for different feature dimensions; Top-K extraction: Select Extract the effective measures from the first K historical cases and add them as supplementary action atoms to the supplementary action set.

[0011] Furthermore, spatiotemporal constraint verification includes: The GIS operator is invoked to verify the implementation of the action command. The Intersect operation is executed, and the judgment logic is as follows: If the return value is Null, it is determined to be a space constraint conflict, and the system executes the conflict resolution strategy.

[0012] Furthermore, generating disaster emergency response plans specifically includes: The verified action atomic sequence is input into the natural language generation model to generate a command and dispatch summary that includes disaster assessment, response level, organizational structure, resource allocation list and action roadmap.

[0013] A smart disaster emergency response plan generation system, used to implement the method described in any of the preceding aspects, comprising: Feature processing module: Used to configure multi-channel data interfaces, access sensor time-series data, weather forecasts and on-site voice streams; use deep learning models to perform named entity recognition, extract entity triples related to disaster type, disaster intensity and environmental constraints; construct a unified disaster feature space, map discrete information into high-dimensional numerical vectors, and output standardized disaster feature vectors; Dual-engine reasoning module: used to input disaster feature vectors into a pre-built emergency domain knowledge graph for dual-engine reasoning; Solution assembly module: It is used to merge the mandatory action set and the supplementary action set to obtain the initial action set, perform spatiotemporal constraint verification on the initial action set and resolve conflicting instructions; serialize the verified action list, input it into the large model through PromptEngineering and output the disaster emergency plan.

[0014] Furthermore, the feature processing module supports multi-channel data interfaces, adapting to the access and processing of numerical, textual, and unstructured data.

[0015] Furthermore, in the dual-engine reasoning module, the rule reasoning unit has a built-in IF-THEN production rule base, and the case reasoning unit has a built-in historical disaster case vector database.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This method enables the unified aggregation and standardized transformation of multi-source heterogeneous disaster data, generating complete disaster feature vectors. It effectively connects the original dynamic disaster data with the subsequent static knowledge reasoning process, allowing static knowledge and rules to accurately adapt to dynamic disaster characteristics and avoiding the one-sided interpretation of dynamic disasters by static texts.

[0017] This method is supported by an emergency response knowledge graph and generates action sets that are both compliant and adaptable through dual-engine reasoning. The rule reasoning stage adheres to the core requirements of regulations, solidifying the rigid compliance bottom line of the solution, while the case reasoning stage provides diversified disposal options, giving the solution flexibility to adapt to complex disasters. Subsequently, through the optimization, verification and conflict resolution of the action sets, the method further coordinates the regulatory requirements and actual disposal needs. The final natural language emergency solution not only ensures the standardization of regulatory expressions but also fits the actual disposal scenario, achieving a precise balance between the rigid constraints of regulations and the flexible adaptation of disposal. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the technical route of the feature processing module of the present invention; Figure 2 This is a schematic diagram of the technical route of the dual-engine inference module of the present invention; Figure 3 This is a schematic diagram of the technical route for assembling the module of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: S1, as Figure 1 As shown, a multi-channel data interface is configured to access sensor time-series data, weather forecasts, and on-site voice streams; a deep learning model is used for named entity recognition to extract entity triples related to disaster type, disaster intensity, and environmental constraints; a unified disaster feature space is constructed to map discrete information into high-dimensional numerical vectors and output standardized disaster feature vectors.

[0021] This step is mainly used to achieve unified aggregation and standardized processing of sensor time-series data, weather forecast data, and on-site voice streams. At the same time, it completes the abstract transformation from raw data to disaster characteristics, providing standardized input with strong adaptability and complete information for subsequent reasoning.

[0022] Specifically, firstly, a multi-channel heterogeneous data access interface is established. Addressing the inconsistency in time granularity between second-level sensor data and hourly weather forecast data, a strategy of time axis alignment and sliding window aggregation can be adopted. This involves segmenting and statistically summarizing high-frequency sensor data according to the time granularity of weather forecasts, while preserving the temporal trend characteristics of the sensor data. To address the low accuracy of on-site speech-to-text translation in noisy environments, a speech enhancement preprocessing module can be introduced, combined with noise suppression algorithms and echo cancellation technology. Then, a speech-to-text and multi-model fusion named entity recognition (NER) scheme is employed, integrating BERT and a domain-specific fine-tuning model to improve entity extraction accuracy.

[0023] During the named entity extraction stage, based on the different degrees of disaster, key dimensions such as disaster scope, population density, and resource reserves can be supplemented to form a multi-dimensional entity triplet consisting of disaster type, disaster intensity, environmental constraints, disaster scope, and resource status. To address the problem that high-dimensional vectors are prone to losing key contextual information, a context-aware feature embedding method is adopted to fuse the entity triplet with the temporal and semantic features of the original data. A unified disaster feature space is constructed through the GraphSAGE model to generate a high-dimensional numerical vector that combines global context and local key information.

[0024] S2, as Figure 2 As shown, the disaster feature vector is input into the pre-constructed emergency domain knowledge graph for dual-engine reasoning. The emergency domain knowledge graph is constructed using the smallest granularity action unit of the deconstructed traditional plan as the basic unit. Dual-engine reasoning includes: Through a rule-based reasoning engine, preset legal provisions are transformed into IF-THEN production rules and a rule base is established. The rule base is traversed to determine whether the disaster feature vector satisfies the rule antecedents, and a set of mandatory actions that satisfy the rule antecedents is generated. By using a case-based reasoning engine, effective handling measures are retrieved from the emergency domain knowledge graph based on the similarity between disaster feature vectors and historical cases, and a supplementary action set is generated.

[0025] This step is mainly used to generate a set of actions that are both compliant and adaptable, based on standardized disaster feature vectors and combined with emergency response knowledge graphs.

[0026] Specifically, the first step is to construct a knowledge graph for the emergency response field. This involves a strategy of hierarchical decomposition, domain expert annotation, and dynamic calibration. Traditional emergency plans are decomposed into three levels: plan level, response stage, and specific action. This forms a unified annotation system for the smallest unit of action. Furthermore, a granularity adjustment mechanism is established to dynamically adapt the granularity based on the complexity of the disaster.

[0027] In the rule and regulation transformation stage, legal provisions are converted into a combination of fuzzy rules and probability thresholds, constructing a hierarchical rule base containing mandatory and optional rules. In the dual-engine reasoning stage, the rule-based reasoning engine uses a fuzzy matching and threshold judgment mode to traverse the rule base and generate a set of mandatory actions, ensuring the legal compliance of the solution. The case-based reasoning engine uses the "feature vector cosine similarity + domain weight correction" algorithm to retrieve effective handling measures for similar historical cases and generate a supplementary action set. Action dependency edges are introduced into the knowledge graph, and the Petri net model is used to visualize and verify the parallel, mutual exclusion, and conditional triggering relationships between actions, avoiding potential dependency conflicts in advance.

[0028] In this step, the similarity matching of case reasoning in the dual-engine reasoning can capture disaster details not covered by regulations, and improve the adaptability of reasoning results to actual scenarios; the fuzzy matching mode of the rule engine not only adheres to the core requirements of regulations, but also gives the reasoning process adaptability flexibility through probability threshold settings; the supplementary action set generated by case reasoning enriches the disposal options and realizes flexible disposal on the basis of compliance; and the action dependency modeling further ensures the coordination and feasibility of the action set.

[0029] S3, as Figure 3 As shown, the mandatory action set and the supplementary action set are merged to obtain the initial action set. The initial action set is then subjected to spatiotemporal constraint verification and conflict instructions are resolved. The verified action list is serialized, input into the large model through PromptEngineering, and the disaster emergency response plan is output.

[0030] This step optimizes and validates the initial action set generated by the dual-engine inference, resolves conflicts, and adapts to spatiotemporal and resource constraints. Finally, a large model transforms the structured action list into a natural language emergency response plan. Specifically, the mandatory action set and the supplementary action set are first merged and deduplicated. A two-dimensional verification model with spatiotemporal constraints and resource constraints is introduced. A linear programming algorithm is used to coordinate and schedule the execution time, spatial scope, and resource requirements of the actions. A priority rule of legal priority > disaster urgency > case success rate is adopted to resolve resource competition conflicts. In the solution generation stage, a situational awareness feedback interface is embedded to connect to the multi-source data interface of S1 in real time. A threshold for changes in disaster situation is set. When the dynamic disaster exceeds the threshold, the action set is re-inferred and the solution is updated. In the PromptEngineering stage, a Prompt template with a structured action list, disaster context, and legal constraint description is used. The optimized action list is input into the large model to generate a clear and standardized natural language emergency solution, while retaining the structured action list as the basic data for solution adjustment.

[0031] The collaborative verification of resource constraints and spatiotemporal constraints in this step enables the action set to better adapt to the actual resource distribution and spatiotemporal conditions, improving the feasibility of action implementation. The natural language scheme generated by the large model not only ensures the standardization of legal expressions but also enhances its relevance to actual response scenarios by incorporating the disaster context.

[0032] This method enables the unified aggregation and standardized transformation of multi-source heterogeneous disaster data, generating complete disaster feature vectors. It effectively connects the original dynamic disaster data with the subsequent static knowledge reasoning process, allowing static knowledge and rules to accurately adapt to dynamic disaster characteristics and avoiding the one-sided interpretation of dynamic disasters by static texts.

[0033] This method is supported by an emergency response knowledge graph and generates action sets that are both compliant and adaptable through dual-engine reasoning. The rule reasoning stage adheres to the core requirements of regulations, solidifying the rigid compliance bottom line of the solution, while the case reasoning stage provides diversified disposal options, giving the solution flexibility to adapt to complex disasters. Subsequently, through the optimization, verification and conflict resolution of the action sets, the method further coordinates the regulatory requirements and actual disposal needs. The final natural language emergency solution not only ensures the standardization of regulatory expressions but also fits the actual disposal scenario, achieving a precise balance between the rigid constraints of regulations and the flexible adaptation of disposal.

[0034] Example 2: The specific calculation of the disaster feature vector is as follows: Among them, for disaster types ( (), using one-hot encoding for identification; Regarding the intensity of the disaster ( ) and environmental parameters ( ), and then normalize it.

[0035] For discrete disaster types, a one-hot encoding method is used to map different disaster types into non-overlapping binary vectors for identification; for continuous disaster intensity and environmental parameters, normalization is used to scale their values ​​to a uniform range (usually 0-1); finally, the components after the above processing are integrated to form a complete disaster feature vector.

[0036] This disaster feature vector calculation method is used to transform heterogeneous disaster information into a standardized numerical vector form, realizing a unified representation of discrete, continuous, and time-dimensional disaster data. It provides a consistent and computable disaster information carrier for subsequent emergency reasoning, plan generation, and other stages, solving the problem that different types of disaster data are difficult to directly integrate and calculate.

[0037] One-hot encoding ensures clear distinction of disaster types and model identifiability. Normalization eliminates the dimensional differences in disaster intensity and environmental parameters, avoiding interference from numerical differences in different dimensions of data in subsequent calculations. A unified feature vector format allows various types of disaster information to participate efficiently in the subsequent reasoning process, which not only improves the standardization of disaster information representation but also enhances its adaptability to subsequent algorithms and models, laying the foundation for the accurate use of disaster information.

[0038] Example 3: The rule-based reasoning engine uses the Drools rule engine and / or SWRL language to describe hard regulatory clauses.

[0039] Specifically, for mandatory regulatory provisions, if the "Drools rule engine and SWRL language" combination mode is adopted, it is necessary to first combine the knowledge graph ontology structure of the emergency domain, use SWRL language to describe the semantic association of "applicable scenario-compliant action" in the regulations to generate basic rules, and then convert these semantic rules into "IF (antecedent: disaster features match the applicable conditions of the regulations) - THEN (consequence: execute the corresponding compliant action)" production rules supported by Drools and enter them into the rule base; if only one method is used, it is only necessary to complete the writing and loading of the corresponding rules (Drools directly writes production rules, and SWRL writes semantic rules based on the knowledge graph); during the reasoning process, the engine will traverse the rule base, determine whether the current disaster feature vector satisfies the antecedent of the rule, trigger and generate a set of mandatory actions that meet the requirements of the regulations.

[0040] Among them, the Drools rule engine has efficient rule matching and batch execution capabilities, which can quickly complete the automated verification of large-scale legal provisions; the SWRL language has strong semantic compatibility with emergency domain knowledge graphs, which can accurately capture the semantic logic of legal provisions and ensure the accurate association between rules and disaster information.

[0041] Example 4: The triple structure of the emergency response domain knowledge graph is defined as <action, relation, constraint>. Specifically, firstly, the core emergency response action types (such as personnel evacuation, material allocation, facility reinforcement, etc.) are identified, and the constraints corresponding to each action (including time constraints, resource constraints, scenario constraints, etc.) are also identified. Next, the semantic relationships between actions and constraints are defined (such as "limited by", "must satisfy", "depends on", etc.). Finally, specific actions, relations, and constraints are constructed into triple instances in the format of <action, relation, constraint> (e.g., <"mountain road clearing action", "limited by", "continuous rainfall does not exceed moderate rain level">), and these triples are entered into the ontology and instance library of the emergency response domain knowledge graph to complete the structured storage of knowledge. This triple structure makes the association logic between emergency actions and constraints clearer and the semantics more explicit, enabling the inference engine to directly retrieve the corresponding constraints for verification when matching actions, effectively improving the practical feasibility of generating action sets. At the same time, the structured triple form facilitates the maintenance and updating of the knowledge graph. When constraints (such as adjustments to action time limits due to regulatory updates) change, only the corresponding triple instance needs to be modified, reducing the maintenance cost of the knowledge graph and ensuring the timeliness and accuracy of knowledge.

[0042] Example 5: The formula for calculating the similarity between disaster feature vectors and historical cases is as follows: in: Disaster vector , Historical case vectors ; Preset weights for different feature dimensions; Top-K extraction: Select Extract the effective measures from the first K historical cases and add them as supplementary action atoms to the supplementary action set.

[0043] Specifically, first determine the current disaster feature vector A (i.e., V). input ) and historical case feature vector B (i.e., V) history For each feature dimension of the two, calculate the cosine similarity Cos(A) under that dimension. i B j ), and compare it with the preset weight ω of the corresponding dimension. iMultiply the results and sum the results of all dimensions to obtain the comprehensive similarity Sim(A,B) between the current disaster situation and the historical case. Then, filter out the historical cases with similarity Sim>0, select the top K cases with the highest similarity, extract the effective measures in these cases and add them as supplementary action atoms to the supplementary action set.

[0044] By setting preset weights for different feature dimensions, the impact of key disaster features on matching results can be highlighted, improving the accuracy of matching similar cases. The application of cosine similarity can effectively measure the semantic and numerical correlation between vectors, ensuring the rationality of similarity calculation. The "Sim>0" filtering condition combined with Top-K extraction not only filters out cases that are irrelevant to the current disaster situation, but also retains the historical handling experience with the highest adaptability, making the supplementary action set both in line with the current disaster situation and practically effective.

[0045] Example 6: Spatiotemporal constraint verification includes: The GIS operator is invoked to verify the implementation of the action command. The Intersect operation is executed, and the judgment logic is as follows: If the return value is Null, it is determined to be a space constraint conflict, and the system executes the conflict resolution strategy.

[0046] The core function of this spatiotemporal constraint verification method is to use GIS spatial analysis capabilities to verify the compatibility between emergency action instructions and actual geographic space, accurately identify spatial constraint conflicts caused by mismatch in geographic scope, provide clear judgment criteria for subsequent conflict resolution, avoid generating invalid action instructions that cannot be implemented in actual geographic space, and ensure the spatial feasibility of emergency plans.

[0047] Specifically, the system first calls the GIS operator to perform geospatial verification of the proposed emergency action instructions. The core operation is the Intersect operation, which calculates the intersection between the geographic execution range of the action instructions and the actual usable spatial area (such as the rescue reachable area, the resource coverage area, etc.). The judgment logic is set as follows: if the Intersect operation returns null, it means that the execution range of the action instructions and the actual usable space have no intersection, and it is determined that there is a spatial constraint conflict. The system then triggers the preset conflict resolution strategy.

[0048] This method relies on the professional spatial analysis capabilities of GIS operators. The Intersect operation can accurately depict the relationship between action commands and geospatial data, ensuring the accuracy of spatial constraint conflict determination and avoiding subjective bias in human judgment of geographic adaptability. The clear discrimination logic of "returning a null value means conflict" standardizes and automates the conflict identification process, improving verification efficiency.

[0049] Example 7: The specific steps involved in generating a disaster emergency response plan include: The verified action atomic sequence is input into the natural language generation model to generate a command and dispatch summary that includes disaster assessment, response level, organizational structure, resource allocation list and action roadmap.

[0050] Specifically, the input is first defined as the sequence of action atoms after being verified by spatiotemporal constraints and conflict resolution. This sequence is then fed into the natural language generation model as the core input data. Based on the preset emergency plan text structure, the model automatically extracts key information from the action atoms and logically integrates them to generate a complete command and dispatch summary that includes disaster assessment, response level, organizational structure, resource allocation list, and action roadmap, ensuring that each information module is complete and logically coherent.

[0051] By integrating multi-dimensional core information, commanders can quickly and comprehensively grasp the disaster situation and response requirements, avoiding decision-making biases caused by fragmented information.

[0052] Example 8: The method for generating intelligent disaster emergency response plans, and its specific implementation process, are as follows: The system first receives real-time multimodal raw data streams from satellite remote sensing, drone aerial photography, ground sensor networks, social media, emergency hotline voice, and meteorological and hydrological monitoring systems via multi-channel data interfaces. These streams include sensor time-series data, weather forecasts, and on-site voice streams. All data undergoes a rigorous standardized preprocessing procedure before entering the processing core. In the spatiotemporal alignment stage, linear interpolation is used to align data with inconsistent time granularities, such as the weather forecast time point T. f Nearby sensor data i (time t) i ) and s i+1 (time t) i+1 The aligned numerical calculation formula is as follows: .

[0053] During the noise filtering stage, the numerical sensor data stream X={x1,x2,...,x...} is processed. N Perform Z-score anomaly detection based on a sliding window, calculate the mean μ and standard deviation σ within the window, and standardize the values. If |z t If |>3|zt|>3, it is considered an outlier and is removed.

[0054] During the format normalization stage, for continuous disaster intensity v level and environmental parameter v env Perform max-min normalization: ; Scale it to the [0,1] interval; for discrete disaster types ν type Then, one-hot encoding is used to convert it into a binary vector.

[0055] Finally, all processed features are concatenated and mapped into a high-dimensional numerical vector to construct a standardized disaster feature vector V. input .

[0056] The system will V input The input is fed into a pre-built emergency domain knowledge graph for dual-engine collaborative reasoning. In the case-based reasoning engine, the current disaster vector A (i.e., V) needs to be calculated. input The similarity between vector A and historical case B is calculated using the weighted cosine similarity formula: ; in, It is the preset weight of the i-th feature dimension, satisfying d represents the total number of feature dimensions. The system selects the top K historical cases with a similarity Sim(A,B) exceeding a preset threshold θ (e.g., θ=0.75), extracts their effective measures, and generates a supplementary action set.

[0057] The system optimizes, verifies, and resolves conflicts in the generated initial action set. In the resource constraint verification phase, the problem is modeled as a multi-commodity flow network optimization model with capacity constraints, and its mixed-integer linear programming (MILP) model is as follows: Objective function: ; For each node i and resource k, the flow conservation constraint is: ; For each edge (i,j), the capacity constraint is: ; Nonnegativity constraint: ; Where K is the set of resource categories, E is the set of edges (transportation paths), and c ijk x is the unit cost of transporting resource k from node i to j. ijk It refers to the quantity of goods transported, s ik and d ik These represent the supply and demand of node i for resource k, respectively, u k U is the unit volume or weight coefficient of resource k. ij It is the upper limit of the transportation capacity of path (i,j).

[0058] In the spatiotemporal constraint verification, the system calls GIS spatial analysis operators, for example, by calculating the polygon P representing the execution range of the action command. action With the danger zone polygon P risk intersecting area Make a judgment. If If so, then a spatial conflict is determined to exist.

[0059] Finally, the system inputs the validated and optimized action sequence into the large language model via PromptEngineering to generate the final solution. The system also includes a dynamic update mechanism that automatically triggers incremental inference when new data causes significant changes in the feature vector. The criterion for a significant change is: for any dimension i of the feature vector, the amount of change... Exceeding the preset threshold ϵ of this dimension i (For example, the water depth dimension ϵ=0.5). By organically integrating the above calculation formula into each key stage of the method implementation, the technical solution of this invention achieves full-process quantification, calculability, and verifiability from multi-source data perception to intelligent decision output, significantly improving the objectivity, accuracy, and reliability of emergency response plan generation.

[0060] Example 9: The specific implementation process of the intelligent disaster emergency response plan generation method is as follows: The system first integrates multimodal raw data streams in real time from satellite remote sensing, UAV aerial photography, ground sensor networks, social media, emergency hotline voice, and meteorological and hydrological monitoring systems via multi-channel data interfaces. These streams include sensor time-series data, weather forecasts, and on-site voice streams. This data access architecture is compatible with multi-source heterogeneous data, covers a wide range of data sources, and closely matches the actual data needs of disaster emergency scenarios. This provides a sufficient data foundation for solution generation and demonstrates significant feasibility.

[0061] Before entering the core processing layer, all data undergoes a rigorous standardized preprocessing procedure, including spatiotemporal alignment, noise filtering, format normalization, and semantic disambiguation. During the spatiotemporal alignment stage, the system uses a unified WGS84 geographic coordinate system and UTC+8 time base to resample and register all input data, ensuring strict synchronization of timestamps and geographic coordinates across different sources. To address the inconsistency in time granularity between second-level sensor data and hourly weather forecast data, a time axis alignment and sliding window aggregation strategy is employed. For example, using the weather forecast time point as a reference, linear interpolation or segmented aggregation (such as taking the average or maximum value) is used to align the sensor data.

[0062] Furthermore, a geographic registration error correction model can be introduced into the spatiotemporal alignment stage to improve spatial alignment accuracy through ground control point calibration. A temporal consistency verification mechanism can be added to the temporal alignment stage to verify the rationality of the interpolation results. For missing sensor data, a temporal data completion model based on an attention mechanism (such as Transformer-XL) can be used to fill in missing values. At the same time, the outlier detection mechanism can be optimized, and dynamic thresholds can be set in combination with domain expert experience to avoid misjudgments caused by a single Z-score standard. For image data, based on the reconstruction error detection of convolutional autoencoders, auxiliary judgment features such as illumination and weather can be added to improve the accuracy of invalid area identification.

[0063] The specific data processing strategy is as follows: For numerical sensor data, a Z-score normalization and outlier detection mechanism based on a sliding window (e.g., a 60-minute window with a 5-minute sliding window) is adopted. The Z-score normalization formula is: In the formula, x is the data point to be standardized, μ is the mean of the data within the sliding window, and σ is the standard deviation of the data within the sliding window. The outlier judgment criterion is that when |z|>3, the data point is regarded as an invalid reading and is discarded or marked. For image data, reconstruction error detection based on convolutional autoencoder is used to identify invalid regions caused by cloud cover or equipment failure. For text and speech data, after improving the signal-to-noise ratio through speech enhancement preprocessing (such as spectral subtraction and Wiener filtering) and combining it with noise suppression algorithms, named entity recognition technology based on BERT and other models fine-tuned on emergency domain corpora is used to extract key entities such as disaster type, disaster intensity, environmental constraints, disaster range, location of trapped personnel, secondary risk points and resource needs from the text and translated speech to form structured entity triples.

[0064] To address the issues of inaccurate dialect and technical terminology recognition and incomplete named entity extraction in text and speech data processing, an expanded corpus covering dialects from multiple regions and technical terms in the emergency response field was constructed, and the BERT model was fine-tuned. An entity relationship extraction auxiliary model was introduced to supplement missing key entities through contextual semantic association analysis. At the same time, a manual review interface was added to automatically retain high-confidence entities and prompt manual verification for low-confidence entities.

[0065] For image and video data, visual features such as road damage levels (classified into four levels: intact, minor, moderate, and completely disrupted), building collapse probability (0-1 continuous values), and water depth distribution are extracted using convolutional neural networks (e.g., ResNet-50 backbone network combined with a task-specific head). All extracted structured information is mapped to a predefined emergency domain ontology model, and a unified disaster feature space is constructed using graph neural network models such as GraphSAGE. Discrete and heterogeneous disaster information and their correlations are embedded into high-dimensional numerical vectors, ultimately outputting a standardized disaster feature vector V. input This vector is usually represented as V input =[ν type ,ν level ,ν env ,ν time ,...], where disaster type ν type One-hot encoding is used (e.g., earthquake [1,0,0,...], flood [0,1,0,...]), and continuous parameters such as disaster intensity ν are used. level The environmental parameter νenv is subjected to max-min normalization to the [0,1] interval, thereby achieving deep fusion and standardized representation of multi-source heterogeneous data. This fusion scheme is based on mature deep learning and ontology modeling technologies, with a clear technical path and good feasibility.

[0066] Subsequently, the system inputs the generated standardized disaster feature vectors into a pre-constructed emergency domain knowledge graph for dual-engine collaborative reasoning. This knowledge graph is composed of a regulatory knowledge graph and a historical experience graph, and uses the smallest granularity action unit obtained from deconstructing traditional contingency plans as its basic building block. Its construction approach aligns with the core requirements of compliance and experience in emergency decision-making, and is highly feasible.

[0067] Furthermore, a dynamic update mechanism for the knowledge graph can be established. The legal knowledge graph is connected to the official information release interfaces of national and local emergency management departments. Through text change detection algorithms, it can capture updated legal content in real time, automatically trigger knowledge node updates, and synchronize them to the rule base. The historical experience graph adds a regional feature dimension, dividing the experience sub-database according to disaster-prone areas, and optimizing the regional weight of case retrieval by combining factors such as regional economic level and infrastructure distribution. In the legal text parsing stage, a multi-model fusion parsing strategy is introduced, combining dependency parsing, semantic role labeling, and domain expert rule bases to perform multi-dimensional verification of ambiguous texts, ensuring the accuracy of the parsing results.

[0068] The regulatory knowledge graph uses current national and local emergency management laws, regulations, and technical standards as data sources. It extracts compliance constraints such as subject obligations, behavioral prohibitions, and resource quota limits through a legal text parser (combined with dependency parsing and rule templates). These constraints are then modeled as semantic nodes with subject-verb-object triple structures. The nodes are connected by relational edges such as applicable conditions and trigger thresholds to form a machine-readable and reasonable compliance rule network. The Drools rule engine can be used to deploy IF-THEN production rules, and / or the SWRL language can be used for semantic rule description and reasoning. Historical experience graphs use historical disaster event handling records and exercise evaluation reports as data sources. Through event pattern mining algorithms (such as gSpan frequent subgraph mining), they identify high-frequency disaster characteristics-response action mapping relationships and deconstruct contingency plan documents into atomic minimum execution units. Each unit is defined as the smallest instruction block with a unique identifier, input preconditions (represented by first-order logic formulas), output effect declaration, required resource list, execution timing window, geographical scope (GeoJSON format), and associated legal provision number. Its triple structure can be defined as <action, relationship, constraint>.

[0069] The dual-engine reasoning process specifically includes: The rule-based reasoning engine first transforms pre-defined legal provisions into IF-THEN production rules and establishes a rule base. It then iterates through the rule base to determine whether the current disaster feature vector satisfies the rule antecedents, generating a set of mandatory actions that meet the rigid requirements of the regulations. Simultaneously, the case-based reasoning engine is activated, performing a search based on the similarity between the disaster feature vector and historical case feature vectors. The similarity calculation uses a weighted cosine similarity formula, the core expression of which is: ; in Let be the cosine similarity of the i-th feature. Let ||Ai|| and ||Bi|| represent the dot product of the i-th dimension of feature vectors A and B, respectively; A is the current disaster vector Vinput, and B is the historical case vector Vhistory. These are pre-set weights for different characteristic dimensions (such as disaster type, intensity, environment, etc.) that satisfy the constraints. The weights can be dynamically adjusted based on expert experience or historical feedback.

[0070] Furthermore, an analytic hierarchy process (AHP) combined with a dynamic weight adjustment mechanism based on historical response feedback can be introduced. Entropy weighting can be used to correct subjective expert weights, enhancing the scientific rigor of weight settings. An adaptive similarity threshold mechanism can be adopted, dynamically adjusting the threshold θ based on the current disaster type and urgency (e.g., appropriately lowering the threshold to 0.7 in major disaster scenarios and raising it to 0.8 in general disaster scenarios). Simultaneously, the diversity constraints of case retrieval can be increased to avoid homogenization of search results. A dual-engine reasoning conflict adjudication mechanism can be established, constructing an adjudication rule base based on the principle that legal priority is higher than experience priority and emergency scenarios are higher than routine scenarios, to quantitatively evaluate and decide on conflict outcomes.

[0071] The system selects the top K historical cases (e.g., K=50) with similarity exceeding a preset threshold θ (e.g., θ=0.75), extracts their verified effective response measures, and generates a supplementary action set. The dual engines achieve semantic alignment through a shared disaster ontology model, ensuring consistency in reasoning. After generating the mandatory action set and the supplementary action set, the system merges and deduplicates them to obtain the initial action set.

[0072] Next, the system optimizes, verifies, and resolves conflicts in the initial action set. This process involves triple verification of spatiotemporal constraints, resource constraints, and logical consistency, and this verification system comprehensively covers the core constraints for implementing the emergency response plan.

[0073] In the spatiotemporal constraint verification, real-time traffic and meteorological dynamic data are introduced to update the GIS spatial analysis operator, and a path risk assessment model is added to conduct multi-dimensional safety verification of the rescue route. In the resource constraint verification model, emergency resource priority weights are added (e.g., medical resources have higher priority than material transportation), and actual constraints such as vehicle load limits and emergency lane priority are considered. The objective function is optimized to be a multi-objective optimization model with the lowest total transportation cost and the best delivery time of high-priority resources. An improved genetic algorithm is used to improve the solution efficiency. In the logical consistency verification, an emergency action sequence logic rule base is constructed. Combining the Petri net model and causal relationship graph, the system comprehensively detects the problems of missing causal dependencies, time sequence inversion, and effect conflicts in the action sequence. At the same time, logical exception rules for extreme scenarios are added.

[0074] The specific verification and optimization strategies are as follows: Spatiotemporal constraint verification involves calling GIS spatial analysis operators to perform spatial relationship analysis on the geographical execution range of action commands (such as Intersect, Buffer, NetworkAnalysis). For example, the Intersect operation is executed to determine whether the rescue route intersects with a dangerous area (such as a landslide area). If the intersecting area exceeds a safety threshold or a key road segment is blocked, it is determined to be a spatial constraint conflict, and the system will execute a preset conflict resolution strategy, such as replanning the route or adjusting the action sequence.

[0075] Resource constraint verification is performed by constructing a multi-commodity flow network with capacity constraints and optimizing it using a mixed-integer linear programming model. The model is specifically described as follows: The objective function is... Where K is the set of resource categories and E is the set of edges (transportation paths), Let k be the unit cost of transporting resource k from node i to j. For the quantity of goods transported. Constraints include: Flow conservation constraint: For each node and resources ,satisfy , where N is the set of nodes, sik is the supply of resource k at node i, and dik is the demand for resource k at node i; Capacity constraint: For each path ,satisfy , where uk is the unit volume or weight coefficient of resource k, and Uij is the upper limit of the transport capacity of path (i,j); Nonnegativity constraint: For all The model is established. It takes the resource requirements of atomic execution units as input, combines a real-time material inventory database and a transportation capacity model, and uses a solver (such as Gurobi) to find the resource optimization allocation scheme with the lowest total transportation cost. The logical consistency verification module checks for missing causal dependencies, inverted timing, or conflicting effects between action sequences, for example, by constructing a temporal logic rule base or a Petri net model. All verification and conflict resolution processes are coordinated by a central coordinating controller, which maintains the global task state machine and makes decisions based on the rule of priority > disaster urgency > case success rate.

[0076] Finally, the system serializes the action atomic sequences after multiple verifications and optimizations, and inputs them, along with the disaster context and regulatory constraints, into a large language model (such as GPT-4 or ERNIE) using a PromptEngineering template (which includes sections such as "task background," "structured input," and "output format requirements"). Based on these structured inputs, the large model automatically generates the "Emergency Response Guidance Plan" natural language document.

[0077] Example 10: This embodiment uses a practical example to illustrate the method of the present invention in detail, as follows: Scene background: County A in southern China experienced a historically rare and prolonged torrential downpour. Located in a typical mountainous region, it comprises 8 townships and 126 administrative villages, some of which are situated along rivers and have weak infrastructure. The 24-hour rainfall reached 250mm, triggering flash floods that trapped three riverside villages (Villages B, C, and D). Two county-level highways (X01 and X02) were completely cut off by landslides, and five village roads suffered moderate damage. Approximately 800 people were stranded, and power and communication were disrupted in some areas. The meteorological department forecasts moderate rain for the next 24 hours, with a high risk of secondary disasters. The County A Emergency Command Center activated the intelligent disaster emergency response plan generation system of this invention to quickly generate a precise emergency plan.

[0078] The system synchronously accesses the following six types of core data through multi-channel data interfaces: Ground sensor data: Water level data from 12 hydrological monitoring stations within the county, with the water level in the B village section exceeding the warning level by 1.2m; Real-time rainfall data from 3 rain gauge stations, with an increase of 50mm in the past hour; Data from 8 soil moisture sensors, with an average humidity of 92%.

[0079] Weather forecast data: The provincial meteorological department updates the forecast every 6 hours. It is expected to rain moderately in the next 24 hours, with a cumulative rainfall of 30-50 mm and gusts of up to level 6.

[0080] Satellite remote sensing data: 10-meter resolution imagery from the Gaofen-6 satellite identified two landslide areas around Village B, covering an area of ​​approximately 0.3 km².

[0081] Drone aerial data: The Emergency Management Bureau deployed three drones to take aerial photos of the trapped villages and obtain images of road damage. The X01 highway from K12 to K15 is completely cut off, and there is water accumulation of 0.8m in many places on the road from C village to D village.

[0082] Emergency hotline and social media data: The emergency hotline received 46 calls for help within one hour, and 23 requests for help were captured from online social media platforms. The core requests were "people trapped", "lack of drinking water" and "sudden illness of the elderly".

[0083] Basic geographic and resource data: County GIS map; Village B has one primary school shelter with a capacity of 500 people; Village C has a village committee shelter with a capacity of 300 people; the county warehouse has 8 tons of drinking water, 5 tons of food, and 20 boxes of first aid medicines; the county fire and rescue brigade has 20 people, the armed police squadron has 30 people, and there are 6 village-level emergency teams with a total of 80 people.

[0084] Using UTC+8 time as the reference, the second-level sensor data is aggregated according to the 6-hour granularity of weather forecasts (taking the maximum value every 6 hours), and satellite remote sensing and UAV imagery are registered with GIS map coordinates to mark the latitude and longitude of landslide areas and road damage points.

[0085] Z-score detection with a 60-minute sliding window was used to detect water level sensor data, and one outlier caused by flood impact (z=3.2>3) was removed; convolutional autoencoder reconstruction error detection was used to detect UAV imagery, and two invalid images caused by cloud cover were removed.

[0086] After denoising the hotline voice using spectral subtraction, it was converted to text. The BERT model was then fine-tuned using a dialect-expanded corpus to extract entity triples: <“Village B”, “Number of people trapped”, “350 people”>, <“Village C”, “Emergency needs”, “First aid medicine”>, <“X01 Highway”, “Damage level”, “Complete disruption”>.

[0087] ResNet-50 was used to extract road damage levels from UAV imagery, identifying 2 areas of complete road closure and 3 areas of moderate damage; water depth distribution was also analyzed; one-hot encoding [0,1,0,0] was applied to disaster types (corresponding to earthquake, flash flood, typhoon, and debris flow), along with disaster intensity (0.85, normalized based on rainfall and water level exceeding warning levels) and environmental parameters (e.g., 0.92 for 250mm rainfall and 0.92 for 92% soil moisture), ultimately generating a standardized disaster feature vector. Vinput=[0,1,0,0, 0.85,0.92, 0.92, 18:00] (dimensions: disaster type, disaster intensity, rainfall, soil moisture, current time).

[0088] Integrate the "Emergency Response Law", "Flood Control Regulations", and "Emergency Plan for Mountain Flood Disasters in County A" and convert them into IF-THEN rules: If 24-hour rainfall is ≥200mm and people are trapped - THEN, activate Level II emergency response, prioritizing the evacuation of trapped personnel; If the road is completely blocked and the number of people trapped is greater than or equal to 300, then helicopter rescue will be activated. If there are seriously injured persons, then emergency medical supplies will be delivered within 4 hours.

[0089] By analyzing similar flash flood cases in County A and neighboring counties in recent years, we extract the smallest action unit and construct a ternary set: <"Helicopter rescue", "limited by", "visibility ≥ 5km">, <"Reservoir discharge", "must meet", "downstream personnel evacuation completed">, <"Supply delivery", "dependent on", "drone range ≥ 50km">.

[0090] Through the interface of the county emergency management bureau's official website, the latest flood control instructions (such as "add 3 temporary shelters") are synchronized, and the shelter distribution nodes in the knowledge graph are automatically updated.

[0091] The Drools rule engine traverses the regulatory rule base. Given Vinput that meets the criteria of "rainfall ≥ 200mm, people trapped, road completely blocked," it generates mandatory actions. A Level II emergency response was activated, and a command center was established.

[0092] People trapped in villages B, C, and D were transferred to shelters.

[0093] The A County Reservoir has been activated for flood discharge.

[0094] Five tons of drinking water and three tons of food were allocated from the county's warehouse.

[0095] The city's emergency management bureau was coordinated to dispatch two helicopters.

[0096] Calculate the weighted cosine similarity between Vinput and historical cases (disaster type weight 0.3, disaster intensity 0.2, rainfall 0.3, soil moisture 0.2), select the top-3 similar cases (similarity 0.82, 0.78, 0.76), and extract supplementary actions: Village-level emergency teams were activated to pre-position and evacuate the elderly, infirm, and disabled.

[0097] Drones delivered emergency medical supplies to Village C.

[0098] Goods were transported via roads through neighboring County Y.

[0099] Reinforce the river embankment in Village B.

[0100] Power restoration efforts will prioritize ensuring power supply to shelters.

[0101] The GIS operator was invoked to perform an Intersect operation on the ground transportation of materials. It was found that the planned route via X01 highway intersected with the landslide area by 1200m2, which exceeded the safety threshold, and a spatial conflict was determined.

[0102] Conflict resolution: The route was replanned, starting from the county seat, via the Y County Highway to the temporary handover point in D Village; helicopters were simultaneously deployed to transport emergency medical supplies.

[0103] Establish a mixed-integer linear programming model: supply point (county warehouse), demand point (350 people in village B, 280 people in village C, and 170 people in village D), resources (5 tons of drinking water, 3 tons of food, and 20 boxes of first-aid medicines), and transportation route capacity (Y County highway load limit of 10 tons / trip). Solution results: The shipment was transported in two batches. The first batch departed at 12:00, transporting 3 tons of drinking water and 2 tons of food to the transfer point in Village D, from where it was distributed to Villages B and C by drone. The second batch departed at 15:00, transporting 2 tons of drinking water, 1 ton of food, and 20 boxes of first-aid medicine. Simultaneously, a helicopter transported 5 boxes of first-aid medicine, arriving at Village C at 13:30.

[0104] PromptEngineering Template Design: Role: Decision Advisor, Emergency Command Center of County A Background: County A received a total of 250 mm of rainfall from June 15 to 18, triggering flash floods that trapped 800 people in villages B, C, and D. Two county-level roads were completely cut off, and moderate rain is expected to continue for the next 24 hours.

[0105] Core Action List: 1. Response level: Level II emergency response, command center established.

[0106] 2. Personnel transfer: From 10:00 to 14:00, 800 people from villages B, C, and D were transferred to the shelter. The village-level emergency teams assisted in the evacuation of the elderly, weak, sick, and disabled.

[0107] 3. Rescue Support: Two helicopters were coordinated to arrive at 13:00 to transport emergency medical supplies to Village C and transfer seriously injured personnel to the county hospital.

[0108] 4. Material allocation: Transport 5 tons of drinking water and 3 tons of food in two batches, with the first batch of materials to be delivered before 14:00.

[0109] 5. Engineering measures: At 14:30, the A County reservoir will start releasing floodwaters, and at 16:00, the river embankment reinforcement in B Village will begin.

[0110] 6. Road repair: Organize a 50-person repair team to prioritize clearing the road to the D village connection point and restore traffic within 24 hours.

[0111] Output requirements: Generate a structured command and dispatch overview, including disaster assessment, organizational structure, action sequence, resource list, and risk warnings.

[0112] The resulting Level II emergency response command plan for flash flood disaster in County A is as follows: Disaster assessment: Our county has been hit by a once-in-a-century torrential rain, with 250 mm of rainfall in 24 hours, triggering flash floods. 800 people in 3 villages are trapped, 2 county-level roads are completely cut off, and power and communication are disrupted in some areas. Moderate rain is expected to continue in the next 24 hours, with a high risk of secondary landslides and riverbank breaches.

[0113] Organizational Structure: Commander-in-Chief: XXX Deputy Commanders: XXX, XXX It has four working groups: Rescue team: Fire and Rescue Brigade and Armed Police Squadron, responsible for personnel transfer and search and rescue; The materials team, comprised of the County Development and Reform Bureau and the Civil Affairs Bureau, is responsible for the allocation and transportation of materials. Medical team: County Health Commission and County Hospital, responsible for medical treatment and drug supply; Support team: County Transportation Bureau and Power Company, responsible for road repair and power restoration.

[0114] Key action sequence: 10:00-14:00: Complete the transfer of people from villages B, C, and D, prioritizing the evacuation of the elderly, weak, sick, and disabled, and set up temporary medical points in the shelters.

[0115] 13:00: The helicopter arrived at the county's general aviation airport, transporting 5 boxes of emergency medical supplies to Village C and picking up 2 seriously injured people.

[0116] 14:00: The first batch of supplies (3 tons of drinking water and 2 tons of food) arrived at the D Village handover point and were distributed to various villages by drone.

[0117] 14:30: The reservoir in County A started releasing floodwater, strictly controlling the flow rate to ≤50m³ / s, and the downstream monitoring point reported the water level every 15 minutes.

[0118] 16:00: Commence the riverbank reinforcement project in Village B, deploying 2,000 sandbags and 3 excavators.

[0119] 20:00: The second batch of supplies was delivered. The road repair team entered the site and the D Village connection point was opened within 24 hours.

[0120] Resource allocation list: Rescue forces: 20 firefighters, 30 armed police officers, 80 village-level emergency team members, and 2 helicopters.

[0121] Emergency supplies: 5 tons of drinking water, 3 tons of food, 20 boxes of first aid medicines, 2,000 sandbags, and 3 excavators.

[0122] Refuge sites: B Village Primary School (500 people), C Village Committee (300 people), and 3 newly added temporary shelters.

[0123] Example 11: A smart disaster emergency response plan generation system, used to implement the method described in any of the preceding aspects, comprising: Feature processing module: Used to configure multi-channel data interfaces, access sensor time-series data, weather forecasts and on-site voice streams; use deep learning models to perform named entity recognition, extract entity triples related to disaster type, disaster intensity and environmental constraints; construct a unified disaster feature space, map discrete information into high-dimensional numerical vectors, and output standardized disaster feature vectors; Dual-engine reasoning module: used to input disaster feature vectors into a pre-built emergency domain knowledge graph for dual-engine reasoning; Solution assembly module: It is used to merge the mandatory action set and the supplementary action set to obtain the initial action set, perform spatiotemporal constraint verification on the initial action set and resolve conflicting instructions; serialize the verified action list, input it into the large model through PromptEngineering and output the disaster emergency plan.

[0124] Furthermore, the feature processing module supports multi-channel data interfaces, adapting to the access and processing of numerical, textual, and unstructured data.

[0125] Furthermore, in the dual-engine reasoning module, the rule reasoning unit has a built-in IF-THEN production rule base, and the case reasoning unit has a built-in historical disaster case vector database.

[0126] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for generating intelligent disaster emergency response plans, characterized in that, Including the following methods: S1 is configured with a multi-channel data interface to access sensor time-series data, weather forecasts, and on-site voice streams; it uses a deep learning model to perform named entity recognition, extracting entity triples related to disaster type, disaster intensity, and environmental constraints; it constructs a unified disaster feature space, mapping discrete information into high-dimensional numerical vectors, and outputs standardized disaster feature vectors. S2, input the disaster feature vector into the pre-constructed emergency domain knowledge graph and perform dual-engine reasoning. The emergency domain knowledge graph is constructed using the smallest granularity action unit of the deconstructed traditional plan as the basic unit. Dual-engine reasoning includes: Through a rule-based reasoning engine, preset legal provisions are transformed into IF-THEN production rules and a rule base is established. The rule base is traversed to determine whether the disaster feature vector satisfies the rule antecedents, and a set of mandatory actions that satisfy the rule antecedents is generated. By using a case-based reasoning engine, effective handling measures are retrieved from the emergency domain knowledge graph based on the similarity between disaster feature vectors and historical cases, and a supplementary action set is generated. S3 merges the mandatory action set and the supplementary action set to obtain the initial action set, performs spatiotemporal constraint verification on the initial action set and resolves conflicting instructions; serializes the verified action list, inputs it into the large model through PromptEngineering and outputs the disaster emergency response plan.

2. The method for generating intelligent disaster emergency response plans according to claim 1, characterized in that, The specific calculation of the disaster feature vector is as follows: Among them, for disaster types ( (), using one-hot encoding for identification; Regarding the intensity of the disaster ( ) and environmental parameters ( ), and then normalize it.

3. The method for generating an intelligent disaster emergency response plan according to claim 1, characterized in that: The rule-based reasoning engine uses the Drools rule engine and / or SWRL language to describe hard regulatory clauses.

4. The method for generating an intelligent disaster emergency response plan according to claim 1, characterized in that: The triplet structure of knowledge graphs in the emergency response domain is defined as <action, relationship, constraint>.

5. The method for generating an intelligent disaster emergency response plan according to claim 1, characterized in that, The formula for calculating the similarity between disaster feature vectors and historical cases is as follows: in: Disaster vector , Historical case vectors ; Preset weights for different feature dimensions; Top-K extraction: Select Extract the effective measures from the first K historical cases and add them as supplementary action atoms to the supplementary action set.

6. The method for generating an intelligent disaster emergency response plan according to claim 1, characterized in that, Spatiotemporal constraint verification includes: The GIS operator is invoked to verify the implementation of the action command. The Intersect operation is executed, and the judgment logic is as follows: If the return value is Null, it is determined to be a space constraint conflict, and the system executes the conflict resolution strategy.

7. The method for generating an intelligent disaster emergency response plan according to claim 1, characterized in that, The specific steps involved in generating a disaster emergency response plan include: The verified action atomic sequence is input into the natural language generation model to generate a command and dispatch summary that includes disaster assessment, response level, organizational structure, resource allocation list and action roadmap.

8. An intelligent disaster emergency response plan generation system, used to implement the method described in any one of claims 1-7, characterized in that, include: Feature processing module: Used to configure multi-channel data interfaces, access sensor time-series data, weather forecasts and on-site voice streams; use deep learning models to perform named entity recognition, extract entity triples related to disaster type, disaster intensity and environmental constraints; construct a unified disaster feature space, map discrete information into high-dimensional numerical vectors, and output standardized disaster feature vectors; Dual-engine reasoning module: used to input disaster feature vectors into a pre-built emergency domain knowledge graph for dual-engine reasoning; Solution assembly module: It is used to merge the mandatory action set and the supplementary action set to obtain the initial action set, perform spatiotemporal constraint verification on the initial action set and resolve conflicting instructions; serialize the verified action list, input it into the large model through PromptEngineering and output the disaster emergency plan.

9. The intelligent disaster emergency response plan generation system according to claim 8, characterized in that, The feature processing module supports multi-channel data interfaces, adapting to the access and processing of numerical, textual, and unstructured data.

10. The intelligent disaster emergency response plan generation system according to claim 8, characterized in that, In the dual-engine reasoning module, the rule reasoning unit has a built-in IF-THEN production rule base, and the case reasoning unit has a built-in historical disaster case vector database.

Images