A multi-source data emergency rescue knowledge reasoning method based on a large model

Abstract

This invention relates to the fields of emergency rescue and artificial intelligence technology, and discloses a multi-source data emergency rescue knowledge reasoning method based on a large model. The method acquires multi-source heterogeneous emergency rescue data from emergency rescue scenarios, preprocesses this data to obtain multi-source data with a deep unified representation, and inputs it into an emergency rescue knowledge model to obtain enhanced representation data, structured knowledge data, and causal relationship data. A fusion reasoning algorithm is then used to perform interpretable knowledge reasoning on the enhanced representation data, structured knowledge data, and causal relationship data to generate reasoning results. The reasoning results are evaluated to output emergency rescue decision-making recommendations. Through a three-step process of weighted fusion, confidence assessment, and conflict resolution, the method ensures that the decision-making recommendations have high confidence, are logically consistent, and conform to standards, making them suitable for on-site command and improving the reliability, accuracy, practicality, and safety of decision-making.

Description

Technical Field

[0001] This invention relates to the fields of emergency rescue and artificial intelligence technology, specifically to a multi-source data emergency rescue knowledge reasoning method based on a large model. Background Technology

[0002] Currently, existing emergency rescue knowledge reasoning methods mostly rely on a single data source or simple data splicing. However, multi-source emergency data is heterogeneous and diverse, covering structured, semi-structured, and unstructured data. Traditional data processing methods struggle to achieve efficient parsing and standardization of various types of data. Furthermore, existing methods are disconnected from the integration and reasoning of multi-source data. Data integration is merely superficial splicing without in-depth mining based on professional emergency knowledge. The reasoning models have poor scenario adaptability, resulting in low efficiency and poor reliability in emergency rescue decision-making. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the purpose of this invention is to provide a multi-source data-based emergency rescue knowledge reasoning method based on a large model.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A method for emergency rescue knowledge reasoning based on multi-source data using a large model includes: Acquire multi-source heterogeneous emergency rescue data in emergency rescue scenarios; Based on spatiotemporal graph neural networks and large models, the multi-source heterogeneous emergency rescue data is preprocessed to obtain multi-source data with deep unified representation. The multi-source data is input into the emergency rescue knowledge model to obtain enhanced representation data, structured knowledge data, and causal relationship data for the multi-source data. A fusion reasoning algorithm is used to perform interpretable knowledge reasoning on the enhanced representation data, structured knowledge data, and causal relationship data to generate reasoning results; The reasoning results are evaluated and processed to output emergency rescue decision-making suggestions.

[0005] In this invention, preferably, the multi-source heterogeneous emergency rescue data includes static data and dynamic data. The static data includes at least one of rescue case database, emergency standard documents, rescue plans, and geographic information basic data. The dynamic data includes at least one of environmental monitoring data, fire equipment operation data, on-site feedback data, and on-site video or image data.

[0006] In this invention, preferably, the preprocessing of the multi-source heterogeneous emergency rescue data based on spatiotemporal graph neural networks and large models to obtain multi-source data with a deep unified representation includes: The first emergency rescue data is obtained by performing format normalization, cleaning and quality control on the multi-source heterogeneous emergency rescue data; The first emergency rescue data is mapped to a unified spatiotemporal unit, and various heterogeneous information is encoded and spliced ​​into a disaster vector of a unified dimension by combining a large model. A spatiotemporal correlation model is constructed using spatiotemporal units as nodes, disaster vectors as node features, and spatial adjacency and temporal dependency relationships as edges. Based on the spatiotemporal correlation model, deep unified representation learning is performed on the first emergency rescue data to output multi-source data with both semantic consistency and spatiotemporal correlation.

[0007] In this invention, preferably, the method further includes enhancing emergency domain knowledge based on the multi-source data to construct the emergency rescue knowledge model: A pre-training dataset is constructed based on the multi-source data; The large language model is pre-trained based on the pre-trained dataset; The pre-trained large language model is then jointly trained. Construct an emergency rescue sub-scenario Prompt template and input it into the jointly trained large language model to generate an emergency rescue knowledge model.

[0008] In this invention, preferably, the construction of the pre-training dataset based on the multi-source data includes: A corpus is constructed based on the rescue and emergency response professional corpus from the multi-source data, and the rescue and emergency response professional corpus is associated and mapped with the representation vector to construct the first aligned dataset. Based on the emergency rescue-specific knowledge graph in the multi-source data, the entities and relationships in the graph are vector-encoded and associated with the representation vectors to construct a second aligned dataset. Virtual augmented samples are generated based on the small sample representation vectors in the multi-source data to construct a small sample dataset.

[0009] In this invention, preferably, the pre-training of the large language model based on the pre-trained dataset includes: Based on the first aligned dataset, the large language model is pre-trained using corpus masking. Based on the second alignment dataset, knowledge graph alignment pre-training is performed on the large language model; Based on the aforementioned small sample dataset, the large language model is pre-trained with small sample enhancement.

[0010] In this invention, preferably, the step of employing a fusion reasoning algorithm on the enhanced representation data, structured knowledge data, and causal correlation data to perform interpretable knowledge reasoning and generate reasoning results includes: Causal reasoning obtains the initial probability of the reasoning target based on causal correlation data processing; Rule-based reasoning involves matching structured knowledge data with domain rule expressions, obtaining reasoning conclusions based on the matched rules, and correcting the initial probability of the reasoning target based on the rules. Case-based reasoning, based on a comprehensive knowledge system and a historical case database for emergency rescue, extracts feature vectors from historical case scenarios, calculates the similarity between the feature vectors and the current scenario, selects the handling plan corresponding to the feature vectors of the historical case scenarios with high similarity, and supplements and optimizes the reasoning conclusions obtained based on rules. The reasoning results obtained from causal reasoning, rule-based reasoning, and case-based reasoning are fused and optimized using counterfactual intervention algorithms and confidence propagation algorithms.

[0011] In this invention, preferably, the reasoning results obtained above are fused and optimized using the counterfactual intervention algorithm, including: Identify the targets and scenarios for counterfactual intervention; Based on the aforementioned counterfactual intervention targets and scenarios, counterfactual hypotheses are constructed and intervention samples are generated. The intervention sample is input into the emergency rescue knowledge model for counterfactual reasoning, and the probability change is calculated; Key causal elements are selected based on the probability change, and the causal relationship data is corrected based on the key causal elements. Output counterfactual intervention results.

[0012] In this invention, preferably, the inference results obtained above are fused and optimized using the confidence propagation algorithm, including: Obtain the reasoning results and corresponding initial confidence levels of the causal reasoning, rule-based reasoning, and case-based reasoning; With the reasoning objective as the core node and the enhanced representation data, structured knowledge data, and causal relationship data mentioned in the reasoning as child nodes, a confidence propagation network is constructed. Based on the propagation path, the initial confidence of each sub-node is multiplied by the corresponding propagation coefficient to obtain the contribution confidence of each sub-node to the core node, and the sum of the contribution confidence is calculated to obtain the initial total confidence after fusion. The initial total confidence score is corrected by introducing a confidence score correction factor to obtain the final confidence score; By integrating the causal correlation data corrected by counterfactual intervention, the final confidence level obtained from confidence propagation, and the specific conclusions of the three-level inference, a complete inference result is output.

[0013] In this invention, preferably, the step of evaluating and processing the reasoning result to output emergency rescue decision-making suggestion data includes: Construct a set of reasoning results based on the reasoning results; A weighted fusion algorithm is used to fuse the inference results based on the reliability weights of the inference results; The fused inference results are evaluated and corrected. Conflict resolution is performed on the fused inference results based on resolution rules; Once the conflict is resolved, emergency response decision-making recommendations will be generated.

[0014] Compared with the prior art, the beneficial effects of the present invention are: The method of this invention, by constructing a spatiotemporal correlation model, can uniformly encode and spatiotemporally align multi-source heterogeneous emergency rescue data, improving the utilization rate and representation accuracy of emergency rescue data. It also jointly trains emergency rescue text, knowledge graphs, small-sample augmentation, and a large model to construct a three-layer knowledge system of enhanced representation, structured rules, and causal relationships, enabling the large model to possess professional rescue semantics and logical judgment capabilities. Through a three-layer fusion mechanism of causal reasoning, rule-based reasoning, and case-based reasoning, combined with counterfactual intervention algorithms and confidence propagation algorithms, the reasoning process is based on sufficient evidence and is logically rigorous, avoiding the one-sidedness and inexplicability problems caused by a single reasoning method. Through a three-step process of weighted result fusion, confidence assessment, and conflict resolution, it ensures that decision recommendations have high confidence, are free of logical contradictions, and comply with standards, making them usable for on-site command and improving the reliability, accuracy, practicality, and safety of decision-making. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a multi-source data-based emergency rescue knowledge reasoning method based on a large model, as described in this invention.

[0016] Figure 2 This is a schematic diagram of the method flow for step S2 of the present invention.

[0017] Figure 3 This is a schematic diagram of the spatiotemporal correlation model described in this invention.

[0018] Figure 4 This is a schematic diagram of the method flow for step S3 of the present invention.

[0019] Figure 5 This is a schematic diagram of the method flow for step S4 of the present invention.

[0020] Figure 6 This is a schematic diagram of the method flow for step S5 of the present invention. Detailed Implementation

[0021] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0023] Please see Figure 1 A preferred embodiment of the present invention provides a multi-source data emergency rescue knowledge reasoning method based on a large model, applicable to various emergency rescue scenarios such as fires, building collapses, and floods. It enables efficient processing, deep fusion, and accurate knowledge reasoning of multi-source heterogeneous data, providing reliable knowledge support for emergency rescue decision-making. The method includes: S1. Acquire multi-source heterogeneous emergency rescue data in emergency rescue scenarios; S2. Based on the spatiotemporal graph neural network and large model, the multi-source heterogeneous emergency rescue data is preprocessed to obtain multi-source data with deep unified representation. S3. Input the multi-source data into the emergency rescue knowledge model to obtain enhanced representation data, structured knowledge data, and causal relationship data for the multi-source data; S4. The enhanced representation data, structured knowledge data, and causal relationship data are subjected to a fusion reasoning algorithm to perform interpretable knowledge reasoning and generate reasoning results. S5. Evaluate and process the reasoning results, and output emergency rescue decision-making suggestions data.

[0024] In this embodiment, in step S1, the multi-source heterogeneous emergency rescue data includes static data and dynamic data. The static data includes at least one of the following: rescue case library, emergency standard documents, rescue plan, and geographic information basic data. The dynamic data includes at least one of the following: environmental monitoring data, fire equipment operation data, on-site feedback data, and on-site video or image data.

[0025] In one specific embodiment, a multi-source data acquisition module is used to acquire multi-source heterogeneous emergency rescue data in the emergency rescue scenario. This multi-source heterogeneous emergency rescue data includes static and dynamic data. Static data includes a database of fire rescue cases from the past 20 years, fire investigation reports, the "Fire Emergency Rescue Specification" document, fire scene rescue plans, and basic geographic information data of the fire area. The fire rescue case database includes data on fire type, fire location, cause of fire, fire rescue process, and deployment of rescue forces. The basic geographic information data includes road distribution, fire station locations, and water source distribution. Dynamic data includes real-time environmental monitoring data, fire equipment operation data, on-site firefighter feedback data, and on-site video or image data. Real-time environmental monitoring data may include temperature, humidity, smoke concentration, wind direction, and wind speed, with a sampling frequency of once per minute. Fire equipment operation data may include fire truck location, water pump pressure, and remaining fire extinguisher capacity. On-site firefighter feedback data can be data obtained through text or voice conversion.

[0026] Please see Figure 2 In this embodiment, in step S2, the preprocessing of the multi-source heterogeneous emergency rescue data based on the spatiotemporal graph neural network and large model to obtain multi-source data with deep unified representation includes: S21. The multi-source heterogeneous emergency rescue data is normalized, cleaned, and quality controlled to obtain the first emergency rescue data. S22. Map the first emergency rescue data to a unified spatiotemporal unit, and combine it with a multimodal large model to encode various heterogeneous information and splice it into a disaster vector of a unified dimension. S23. Construct a spatiotemporal correlation model using spatiotemporal units as nodes, disaster vectors as node features, and spatial adjacency and temporal dependency relationships as edges; S24. Based on the spatiotemporal correlation model, perform deep unified representation learning on the first emergency rescue data to output multi-source data with both semantic consistency and spatiotemporal correlation in deep unified representation.

[0027] Specifically, due to the characteristics of multi-source heterogeneous emergency rescue data, such as diversity, heterogeneity, strong spatiotemporal correlation, and significant noise, it is difficult to directly fuse and model these data, thus failing to provide unified data support for emergency decision-making. Therefore, the multi-source heterogeneous emergency rescue data is preprocessed to obtain multi-source data with a deeply unified representation.

[0028] First, in step S21, the multi-source heterogeneous emergency rescue data obtained from different data sources such as sensors, videos, text, and GIS are normalized, cleaned, and quality controlled to obtain the first emergency rescue data, thus solving problems such as inconsistent data formats, high noise, and redundancy.

[0029] We utilize the semantic understanding capabilities of the large-scale model qwen3-32b to clean and extract semantics from disaster-related texts, such as police reports and rescue information. On one hand, it identifies redundant and ambiguous information in the text, achieving accurate deduplication; on the other hand, it replaces traditional BERT encoding, using the large-scale model to generate higher-quality text semantic vectors, which can not only capture the surface semantics of the text but also uncover implicit disaster information within it.

[0030] To address issues such as blurry images, insufficient lighting, and occlusion in emergency scenarios, the image generation and restoration capabilities of large-scale models are utilized to enhance low-quality images, such as deblurring and filling in occluded areas, thereby improving image quality. Simultaneously, the image feature extraction capabilities of multimodal large-scale models are used to replace traditional CNN / ViT encoding, generating more semantically meaningful visual feature vectors. This enables accurate identification of key disaster information in images, such as the extent of collapse, water depth, and the location of rescue personnel, improving the correlation between visual features and disaster semantics.

[0031] To address the lack of key disaster indicators, such as the number of casualties and the severity of the disaster, this system utilizes the reasoning capabilities of a large-scale model, combined with other multi-source data from the same region and time period, to perform intelligent estimation and incomplete filling. The incomplete filling accuracy surpasses traditional neighborhood averaging and interpolation methods, making it particularly suitable for scenarios with severe data gaps. These other multi-source data include meteorological data, image data, and text descriptions.

[0032] A dual strategy is employed to detect outliers: based on statistical methods such as the 3σ principle and box plots, obvious anomalies in numerical data (such as sudden changes in sensor data or temperatures far exceeding normal ranges) are removed; the Isolation Forest algorithm is used to identify latent anomalies such as sensor device drift and false alarms, suitable for anomaly detection in high-dimensional time-series data; combined with strong rule filtering in emergency rescue fields, such anomalous data is directly removed to avoid subsequent characterization biases caused by false alarms and equipment failures. Simultaneously, detected outliers are classified and labeled, including equipment failure, false alarms, and extreme anomalies, and this information is synchronously fed back to the multi-source data acquisition module for subsequent equipment maintenance and data acquisition optimization.

[0033] Based on the spatiotemporal characteristics of emergency rescue data, a deduplication strategy is employed to process multi-source heterogeneous emergency rescue data, avoiding redundant data from consuming storage resources and interfering with the representation effect. For time-series data, a 5-minute sliding window and DBSCAN spatial clustering are used to merge duplicate sensor data and rescue location data within the same time window and spatial cluster unit, retaining the latest valid data. For text data, a large-scale model semantic deduplication is used. A fine-tuned multimodal large-scale model is called to calculate text similarity. A similarity of ≥0.95 is considered duplicate, and duplicate alarm descriptions, social media feedback, and other texts are removed in batches, while retaining the text entries with the most complete semantics and the most comprehensive information. For image data: Image features are extracted based on the SIFT algorithm, and combined with large-scale model image semantic comparison, a feature similarity of not less than 0.9 is considered duplicate, and on-site monitoring images and drone aerial images taken from the same angle and at the same time are removed, prioritizing the retention of image data with high clarity and complete disaster information. For spatial data, duplicate geographic features in the GIS layer are merged to ensure the uniqueness and accuracy of spatial data.

[0034] Next, in S22, the first emergency rescue data is mapped to a unified spatiotemporal unit, and various heterogeneous information is encoded and concatenated into a disaster vector of a unified dimension using a multimodal large model. For numerical features, such as temperature, wind speed, and rainfall, Min-Max normalization is used, mapping them to the [0,1] interval. For categorical features, such as disaster type and rescue level, OneHot encoding is used when the number of categories is less than or equal to 5 and the corresponding rescue level is I-IV; Embedding encoding is used when the number of categories is greater than 5, with a 32-dimensional encoding dimension. For text data encoding, the fine-tuned multimodal large model is called to generate a 768-dimensional text semantic vector, capturing the surface semantics and implicit disaster information of the text. For image data encoding, the image encoding module of the multimodal large model is called to generate a 768-dimensional visual feature vector, accurately identifying key disaster information. For spatial location encoding, the location embedding method is used to map the grid center latitude and longitude and GridID to a 768-dimensional spatial feature vector, preserving spatial adjacency relationships. For disaster vector concatenation and purification, v_str (100-dimensional), v_txt (768-dimensional), v_img (768-dimensional), and v_loc (768-dimensional) are first concatenated and then uniformly adjusted to 512 dimensions through a fully connected layer to achieve disaster vector concatenation. Then, the concatenated vector is denoised and redundantly removed by calling a large model to retain the core disaster features.

[0035] Please see Figure 3In step S23, a spatiotemporal correlation model is constructed using spatiotemporal units as nodes, disaster vectors as node features, and spatial adjacency and temporal dependencies as edges. Each spatiotemporal unit is treated as a node in the spatiotemporal graph, and each node is assigned a unique identifier. The node identifier is formed by combining the time slice identifier (time_slice_id) and the spatial grid identifier (grid_id) of the spatiotemporal unit, denoised as "time_slice_id + grid_id". The feature of each node is a 512-dimensional disaster vector (v_i_refined) refined by the large model. The disaster vector is obtained by encoding and concatenating structured data, text data, image data, and spatial location data, followed by noise reduction and redundancy removal processing by the large model. Spatiotemporal correlation is achieved by constructing spatial and temporal edges. For all nodes within the same time slice, the spatial distance between any two nodes corresponding to spatiotemporal units is calculated. When the spatial distance is less than or equal to a preset distance threshold of 500m, or when the spatial grids corresponding to the two nodes are adjacent grids, a spatial edge is constructed between the two nodes to capture the disaster correlation between different spatial locations at the same time. For all nodes within the same spatial grid, after sorting them in time slice order, time edges are constructed between nodes corresponding to adjacent time slices to capture the evolution pattern of disaster at different times in the same spatial location.

[0036] The weight of each edge is calculated using the following formula: , Where spatial_weight is the spatial weight, calculated using a Gaussian function: , Where distance represents the spatial distance between the corresponding spatiotemporal units of the two nodes; semantic_weight represents the semantic relevance of the large model: First, the disaster vectors (v_i_refined, v_j_refined) of the two nodes are input into the fine-tuned multimodal large model. Through the feature extraction module of the large model, deep semantic mining is performed on the two disaster vectors, and the corresponding high-dimensional semantic feature vectors are output. f i , f j Finally, the calculated cosine similarity value is used as the semantic_weight: , in, This represents the dot product of two semantic feature vectors. These represent the L2 norms of the two semantic feature vectors. The cosine similarity value, semantic_weight, ranges from [0,1], with a larger value indicating a stronger semantic association between the spatiotemporal units corresponding to the two nodes.

[0037] By using a large model as a feature pre-training module and jointly training it with the ST-GNN model, the generalization ability of the large model can be utilized to improve the ST-GNN model's ability to capture spatiotemporal features. Especially in the case of small sample emergency data, it can effectively alleviate the model overfitting problem and improve the generalization ability and accuracy of the unified representation.

[0038] The constructed spatiotemporal correlation model comprises an input layer, an ST-GNN encoding module, a feature bridging layer, a large model interaction module, and an output layer. The input layer receives multi-source disaster vector spatiotemporal graph data, simultaneously receiving multiple heterogeneous spatiotemporal graphs to capture multi-source features. The ST-GNN encoding module consists of spatial and temporal convolutions. Spatial convolution aggregates the neighborhood spatial features of each node through a graph neural network, capturing road network connectivity and node dependencies; temporal convolution uses a temporal convolutional network (TCN) or a gating mechanism to model the dynamic changes of node features over time, completing the initial extraction of spatiotemporal features. The feature bridging layer performs dimensional transformation and mapping on the structured spatiotemporal features output by the ST-GNN model, adapting them to the input format of the large model and converting graph structure features into sequential feature vectors that the large model can understand. The encoder in the large model interaction module performs deep semantic modeling of spatiotemporal features, mining hidden global spatiotemporal patterns in the data, such as the implicit impact of weather on fires. The large model interaction module includes a spatiotemporal attention mechanism. This mechanism strengthens the feature weights for key time periods and nodes, and allows for the interaction and fusion of features from the large model and the ST-GNN model, compensating for the insufficient semantic understanding capabilities of pure graph models. The decoder in the large model interaction module performs prediction or reconstruction tasks based on the fused features, such as predicting traffic flow in future periods or restoring missing spatiotemporal data. After the spatiotemporal correlation model is trained, it processes the input disaster vector and outputs a 512-dimensional deep unified representation vector for each node, i.e., multi-source data, specifically: ,

[0039] Where n represents the number of data samples, corresponding to the total number of nodes in the disaster vector spatiotemporal graph, i.e., the total number of spatiotemporal units in the emergency rescue scenario; d represents the dimension of the unified representation vector. The unified representation vector of the i-th spatiotemporal unit contains only spatiotemporal correlation features and does not incorporate professional knowledge from the field of emergency rescue.

[0040] Please see Figure 4In this embodiment, the method further includes, before step S3, performing emergency domain knowledge enhancement based on the multi-source data to construct the emergency rescue knowledge model: S31. Construct a pre-training dataset based on the multi-source data; S32. Pre-train the large language model based on the pre-trained dataset; S33. Jointly train the pre-trained large language model; S34. Construct an emergency rescue sub-scenario Prompt template and input it into the jointly trained large language model to generate an emergency rescue knowledge model.

[0041] Specifically, the general-purpose large language model qwen3-32b was selected and fine-tuned for the emergency rescue field, with initial weights denoted as follows: The input to the general large model is a unified representation vector. The output is an enhanced representation and structured knowledge fused with domain knowledge. The main structure of the general large model includes a feature encoding layer, a knowledge fusion layer, and a semantic decoding layer. The feature encoding layer adopts a Transformer encoder structure to process the unified representation vector of the input. Perform deep encoding and output the encoded feature matrix. , The encoded feature dimension satisfies , , in, For the input embedding weight matrix, This is the hidden layer weight matrix of the Transformer encoder, which is iteratively updated through subsequent dedicated pre-training tasks.

[0042] The knowledge fusion layer is used to integrate emergency rescue-specific knowledge graphs with domain rules to achieve deep binding between features and knowledge, providing a foundation for subsequent knowledge enhancement.

[0043] The semantic decoding layer performs semantic mapping between the encoded features and the fused domain knowledge, outputting enhanced representations and structured knowledge to complete knowledge understanding and enhancement.

[0044] First, in step S31, a pre-training dataset is constructed based on the multi-source data, including: A corpus is constructed based on the rescue and emergency response professional corpus from the multi-source data, and the rescue and emergency response professional corpus is associated and mapped with the representation vector to construct the first aligned dataset. Based on the emergency rescue-specific knowledge graph in the multi-source data, the entities and relationships in the graph are vector-encoded and associated with the representation vectors to construct a second aligned dataset. Virtual augmented samples are generated based on the small sample representation vectors in the multi-source data to construct a small sample dataset.

[0045] Specifically, we will collect professional language data in the field of emergency rescue to construct a corpus. The corpus is associated with and mapped to representation vectors to construct the first aligned dataset. This is used for subsequent corpus mask pre-training tasks.

[0046] in, The unified representation vector representing the i-th spatiotemporal unit is the multi-source data depth unified representation vector output from step S2. ; Represents the relationship with the i-th representation vector The corresponding professional corpus in the field of emergency rescue, with subscript " "Indicates that the corpus is a language database" The first in Corpus.

[0047] Building a dedicated knowledge graph for emergency rescue After vector encoding of graph entities and relations, they are associated with representation vectors to construct a second aligned dataset. This is used for subsequent knowledge graph alignment pre-training tasks. This represents a knowledge graph specifically designed for emergency response, serving as a core carrier of structured domain knowledge. Representation of knowledge graph The set of entities, including all relevant entities in the field of emergency rescue; Representation of knowledge graph A set of relationships, containing all associations between entities; Representation of knowledge graph The entity attribute set contains basic attributes such as type, parameters, and processing requirements for each entity. The unified representation vector of the i-th spatiotemporal unit is derived from multi-source data of deep unified representation. ; Represents a knowledge graph specifically for emergency rescue. The vector corresponding to the i-th entity in the vector includes entities such as emergency rescue scenario entities, disposal entities, and risk entities, such as "forest fire", "firebreak", and "flammable vegetation". Represents the i-th entity in the knowledge graph With the j-th entity The vectors corresponding to the relationships between them include "prone to triggering", "handling process", "risk level", etc. Represents the relationship between the i-th entity and the knowledge graph. The vector corresponding to the j-th entity with an association relationship.

[0048] Collect small sample representation vectors Through formula Generate virtual augmented samples to construct a small sample dataset. A small sample dataset was formed after expert verification.

[0049] in, This is the virtual augmented sample corresponding to the original representation vector of the i-th small sample; This is the original representation vector of the i-th small sample; This is an enhancement coefficient used to control the intensity of Gaussian noise, with a value ranging from 0.01 to 0.1, to avoid excessive noise damaging the core features of the original sample. To conform to a mean of 0 and a variance of Gaussian noise.

[0050] Next, in step S32, the large language model is pre-trained based on the pre-trained dataset, including: Based on the first aligned dataset, the large language model is pre-trained using corpus masking. Based on the second alignment dataset, knowledge graph alignment pre-training is performed on the large language model; Based on the aforementioned small sample dataset, the large language model is pre-trained with small sample enhancement.

[0051] Specifically, for the first aligned dataset The Chinese corpus is subjected to a 20% random mask, and the masked content is predicted based on the representation vector. The loss function is: , in, The loss value is used for pre-training the corpus mask. The smaller the loss value, the higher the accuracy of the model in predicting the mask content. Aligning datasets for "representation-corpus" The total number of samples; for A single sample, Let i be the i-th representation vector. This is the corresponding corpus for the emergency rescue field; L For a single corpus Length; Let be the real corpus label for the i-th sample at the t-th position; Let be the predicted probability of the model for the mask at position t of the i-th sample; Based on the second alignment dataset The loss function for predicting relationships between entities using a large model is: , in, Align the pre-trained loss value with the knowledge graph. The smaller the loss value, the higher the accuracy of the model's prediction of the relationships between entities, and the better the alignment effect between the representation vector and the knowledge graph. N 2. Representation-Knowledge Alignment Dataset The total number of samples; for A single sample in the middle, For two related emergency rescue knowledge graph entity vectors, This is the relation vector between two entities; This is the Euclidean distance function, used to calculate the spatial deviation between two vectors; Based on small sample datasets The loss function is used to learn the association between the original samples and the augmented samples: , in, The smaller the loss value, the higher the feature similarity between the original small sample and the virtual augmented sample, and the better the model learns the core features of the small sample. N 3 is a small sample pre-trained dataset The total number of samples; for A single sample in the, This is the original representation vector for a small sample. For the corresponding virtual augmented sample; is the cosine similarity function, used to measure the degree of feature matching between two vectors. Its value ranges from [0,1], and the larger the value, the higher the similarity.

[0052] Then, in step S33, the pre-trained large language model is jointly trained. , Calculate total loss ;in, They represent the calculation of total loss respectively. hour The weight it occupies.

[0053] Using the Adam optimizer, the learning rate ,according to Update the model weights and train for 100-200 epochs; when the total loss fluctuation is ≤0.001 for 10 consecutive epochs, stop training and record the model weights after convergence. .

[0054] in, , These are the model weights after training rounds t and t+1, respectively. The learning rate is used to control the step size of weight updates, avoiding model oscillation caused by an excessively large step size and low training efficiency caused by an excessively small step size. Total loss function Weight in round t The gradient at a given point is used to indicate the direction of weight updates.

[0055] Finally, in step S34, emergency rescue sub-scenario Prompt templates are constructed and input into the jointly trained large language model to generate an emergency rescue knowledge model. First, a dynamic Prompt library is constructed for various emergency rescue sub-scenarios. ,in, For dynamic Prompt library, These represent independent Prompt templates for different scenarios in the library; K represents the total number of Prompt templates contained in the dynamic Prompt library.

[0056] Based on formula Identify the optimal scene and Prompt template corresponding to the current representation vector. .in, This serves as the index for the optimal emergency rescue scenario that will ultimately be matched. It represents all preset emergency rescue scenarios in the dynamic Prompt library, such as forest fires, floods, and hazardous chemical leaks; The unified representation vector for the i-th spatiotemporal unit output by S2; This is the encoded feature vector of the Prompt template corresponding to the m-th scene; The feature similarity between the i-th unified representation vector and the m-th scene Prompt template is calculated using cosine similarity.

[0057] representation vector Enter the matching Prompt template placeholder, denoted as .in, This indicates that after scene recognition, it is compared with the current representation vector. The optimal matching emergency rescue sub-scenario-specific Prompt template, among which It is an index of the optimal scenario; This indicates a personalized Prompt input.

[0058] Input personalized prompts into the converged emergency rescue knowledge model This will enhance and improve knowledge comprehension.

[0059] Furthermore, step S35 includes, based on the output of the constructed emergency rescue knowledge model, organizing the three-layer knowledge system to obtain, corresponding to, enhanced representation data, structured knowledge data, and causal relationship data for the multi-source data: Enhanced unified characterization processing is performed to obtain enhanced characterization data. (According to...) Generate an enhanced unified representation ;in, The enhanced unified representation vector representing the i-th spatiotemporal unit is a high-dimensional feature vector after incorporating domain semantics; The semantic decoding layer function of the model maps the fused features encoded by the model into enhanced representation vectors with specialized semantics; These are the optimal weights after the model converges. Enter personalized prompts; Let be the set of all spatiotemporal unit enhanced representation vectors, denoted as It retains the spatiotemporal correlation features of the original representation while injecting professional semantics for emergency rescue.

[0060] Structured Domain Knowledge Layer: Extracting Structured Knowledge Data Displayed in the form of triples and regular expressions; Deepening layer: Mining implicit causal relationships in data It is presented in the form of cause-effect graphs and conditional probabilities; among them, This represents a causal graph where nodes are key elements for emergency rescue, such as rainfall, pipeline aging, and leakage risk, and edges represent causal relationships between elements, such as pipeline aging leading to hazardous chemical leakage. P(Y|X) represents conditional probability, used to quantify the strength of causal association, such as P(increased leakage risk|pipeline aging∧frequent rainfall)=0.85, indicating that under the conditions of pipeline aging and frequent rainfall, the probability of increased hazardous chemical leakage risk is 85%.

[0061] Please see Figure 5 In this embodiment, in step S4, the process of applying a fusion reasoning algorithm to the enhanced representation data, structured knowledge data, and causal correlation data to perform interpretable knowledge reasoning and generate reasoning results includes: In step S41, causal reasoning is performed, and the initial probability of the reasoning target is obtained based on the causal association data processing. In step S42, rule-based reasoning is performed, which involves matching domain rule expressions with structured knowledge data, obtaining reasoning conclusions based on the matched rules, and correcting the initial probability of the reasoning target based on the rules. In step S43, case reasoning is performed. Based on the full knowledge system and the emergency rescue historical case database, the feature vectors of historical case scenarios are extracted, the similarity between the feature vectors of historical case scenarios and the current scenario is calculated, and the handling plan corresponding to the feature vectors of historical case scenarios with high similarity is selected to supplement and optimize the reasoning conclusions obtained based on the rules. In step S44, the reasoning results obtained from causal reasoning, rule-based reasoning, and case-based reasoning are fused and optimized using a counterfactual intervention algorithm and a confidence propagation algorithm.

[0062] Specifically, based on a three-layer knowledge system, an enhanced emergency rescue knowledge reasoning algorithm is constructed by integrating causal reasoning, rule-based reasoning, and case-based reasoning through a three-layer reasoning mechanism, and by using counterfactual intervention and confidence propagation algorithms, to complete multi-dimensional interpretable knowledge reasoning.

[0063] In step S41, causal reasoning is based on causal association. Extract key elements of the current scene, filter causal subgraphs, and construct causal chains. ,in As a key causal element, Let the inference target be the initial probability of the inference target. ;in This indicates the strength of the causal influence of a single factor on the target. This indicates the quantification of the authenticity of the element in the current scenario, ultimately yielding the initial probability of the reasoning target, and providing causal logic support for subsequent reasoning.

[0064] In step S42, rule-based reasoning is based on structured knowledge data. Matching domain rule expressions, according to formulas

[0065] Perform logical judgments and output the reasoning conclusion and the corrected probability. If the condition is met (Judge=1), output the reasoning conclusion corresponding to the rule, such as the risk level and the initial treatment direction, and correct the initial probability of the reasoning target based on the priority weight of the rule. If there are multiple matching rules, integrate the results using a weighted summation method. The weight is set according to the priority of the emergency scenario corresponding to the rule. For example, rules related to life safety have the highest priority to ensure the compliance and rationality of the reasoning conclusion.

[0066] In step S43, case reasoning is based on the full knowledge system. Emergency Rescue Historical Case Database The algorithm extracts feature vectors from historical case scenarios and calculates the similarity between these vectors and the current scenario's feature vectors. Higher similarity indicates greater reference value of historical case handling experience for the current scenario. Several historical cases with the highest similarity are selected, and their handling plans and effects are used to supplement and optimize the conclusions of rule-based reasoning. This includes refining handling parameters and improving handling procedures, thus addressing the problem of rule-based reasoning that only clarifies the direction but not the details.

[0067] Finally, in step S44, the reasoning results obtained from causal reasoning, rule-based reasoning, and case-based reasoning are fused and optimized using a counterfactual intervention algorithm and a confidence propagation algorithm. This addresses potential issues such as causal confusion and outcome uncertainty in the three-layer reasoning process. By employing a dual optimization algorithm of counterfactual intervention and confidence propagation, the three-layer reasoning results are deeply fused and corrected, improving both the accuracy and reliability of the reasoning results while ensuring the traceability of the reasoning process.

[0068] Specifically, the counterfactual intervention algorithm is used to fuse and optimize the inference results obtained above, including: determining the counterfactual intervention object and intervention scenario; and extracting the core causal chain from the causal inference output. Identify all key causal elements in the causal chain. At the same time, it is necessary to identify the core characteristics of the current emergency scenario, such as wind speed and vegetation moisture content in a forest fire scenario, to ensure that the intervention scenario is consistent with the current actual scenario and to avoid intervention that deviates from actual needs.

[0069] Based on the aforementioned counterfactual intervention targets and scenarios, counterfactual hypotheses are constructed and intervention samples are generated. For each key causal element... Construct counterfactual assumptions about "removing the element" one by one, that is, assume If the value is absent or within the normal range, it is based on the implicit causal relationship of the deeper layer. Generate corresponding counterfactual intervention samples, keeping other causal factors unchanged, and only changing the target intervention factor. The value is adjusted to "none" or a normal threshold to form a counterfactual sample set corresponding to the original inference sample.

[0070] The intervention samples are input into the emergency rescue knowledge model for counterfactual reasoning, and the probability change is calculated. All counterfactual intervention samples are then input into the three-layer reasoning mechanism one by one, and the above reasoning process is repeated to obtain the reasoning target probability corresponding to each counterfactual sample. ; Calculate the probability change of each counterfactual sample compared to the original inference sample. ,in The initial probability of the original inference target.

[0071] Key causal elements are selected based on the probability change, and the causal relationship data is corrected based on these key causal elements. A probability change threshold is set. Based on the accuracy requirements of emergency rescue scenarios, when At that time, determine the intervention element. As a "key causal element," its position in the causal chain is preserved; when When this element is identified as a "disruptive element," it is removed from the causal chain, ultimately yielding the corrected core causal chain. At the same time, the initial probability of causal reasoning is corrected. .

[0072] Output counterfactual intervention results. Record the intervention process, probability changes, and correction basis for each key causal element, and form a counterfactual intervention report. This report serves as the core support for subsequent confidence propagation and final inference results, ensuring the traceability of the causal inference process.

[0073] In this embodiment, the inference results obtained above are fused and optimized using the confidence propagation algorithm, including: Obtain the reasoning results and corresponding initial confidence levels of the causal reasoning, rule-based reasoning, and case-based reasoning.

[0074] Extract the initial confidence scores after causal reasoning correction. Corrected confidence level of rule-based reasoning output Optimized confidence level of case reasoning output At the same time, the reliability weights of each reasoning step are clearly defined. The weights are set based on the reliability of each stage and must meet the following requirements. .

[0075] A confidence propagation network is constructed using the reasoning target as the core node and the enhanced representation data, structured knowledge data, and causal relationship data mentioned in the reasoning as child nodes. Using the reasoning target Y as the core node and the three-layer reasoning process as child nodes, the initial confidence level of each reasoning process is used as the input confidence level of the child nodes, and the reliability weight is used as the propagation coefficient from the child nodes to the core node, thus clarifying the confidence transmission path and ensuring the traceability of the propagation process.

[0076] Based on the propagation path, the initial confidence level of each sub-node is multiplied by its corresponding propagation coefficient to obtain the contribution confidence level of each sub-node to the core node. The sum of these contribution confidence levels is then calculated to obtain the initial total confidence level after fusion. Following the propagation path, the initial confidence level of each sub-node is multiplied by its corresponding propagation coefficient to obtain the contribution confidence level of each stage to the core node: causal inference contribution confidence level. Confidence contribution of rule-based reasoning Case reasoning contributes confidence ; Sum the confidence scores of the three contributions to obtain the initial total confidence score after fusion: .

[0077] A confidence correction factor is introduced to adjust the initial total confidence level to obtain the final confidence level. A confidence correction factor γ (ranging from 0.95 to 1.0) is introduced to adjust the fused initial total confidence level to obtain the final confidence level. ; Verify the convergence of the confidence level, when If the confidence propagation is successful, then the confidence propagation is considered to have converged; otherwise, the propagation coefficient is adjusted and recalculated until convergence is achieved.

[0078] In step S45, the causal relationship data corrected by counterfactual intervention and the final confidence level obtained from confidence propagation are integrated, and combined with the specific conclusions of the three-level inference, to output a complete inference result. This complete optimized result includes: emergency risk level, specific response plan, complete inference basis (including the counterfactual intervention process, confidence propagation process, causal chain, matching rules, and similar cases), and result confidence level. Ultimately, this achieves the accuracy and interpretability of the reasoning results.

[0079] In this embodiment, in step S5, the reasoning result is evaluated and processed to output emergency rescue decision-making suggestion data, including: S51. Construct a set of reasoning results based on the reasoning results; S52. A weighted fusion algorithm is used to fuse the inference results based on the reliability weights of the inference results; S53. Evaluate and correct the fused reasoning results; S54. Conflict resolution is performed on the fused inference results based on the resolution rules; S55. After the conflict is resolved, output emergency rescue decision-making recommendations data.

[0080] Specifically, step S51, which constructs a set of inference results based on the inference results, includes: constructing a set of inference results from all the output preliminary inference results. Each reasoning result , Emergency risk levels are categorized as high, medium, and low. An emergency response plan, including the response process and key parameters, The basis for reasoning includes causal chains, matching rules, and similar cases. The initial confidence level for this result is set; all inference results are standardized to unify the risk level classification criteria, the description specifications of the treatment plan, and the confidence level range (0-1) to ensure consistency in subsequent fusion, evaluation, and resolution, and the confidence level is then eliminated. Invalid reasoning results.

[0081] In step S52, a weighted fusion algorithm is used to fuse the preprocessed valid inference results based on the reliability weights of each inference result. The reliability weights of each inference result are then determined. The initial confidence level of the weight calculation combined with the inference results. The completeness of the reasoning basis, such as whether it includes a complete causal chain, rules, and cases, according to the formula. Calculate, where, Let i be the initial confidence level of the i-th inference. Let be the completeness coefficient of the i-th reasoning basis, where 1.0 is for completeness, 0.8 is for missing one item, and 0.5 is for missing two or more items. n For the number of valid reasoning results, satisfying . This is the sum of the corrected confidence scores for all n inferences.

[0082] Regarding risk level Calculate the weighted voting scores for each risk level. The risk level with the highest score is selected as the risk level after fusion. Regarding the handling plan Extract the core handling steps from each solution, and integrate them by weight to form a unified and complete integrated handling solution. Prioritize retaining processing details from high-weight results, such as key parameters and process order; for the reasoning basis... The reasoning basis of all valid results is integrated and deduplicated to form a complete fusion reasoning basis. ; Output fusion inference results ,in This represents the initial confidence level after fusion.

[0083] In step S53, the initial confidence level after fusion is determined. Accurate evaluation and correction are performed to ensure that the confidence level truly reflects the reliability of the reasoning results. First, a confidence level evaluation index system is constructed, selecting three core evaluation indicators, including the suitability of the reasoning basis. Historical case matching degree and compliance with rules The reasoning is based on fit. To measure the degree of matching between the reasoning basis and the current scenario, a value of 0-1 is used; historical case matching degree To measure the alignment between the proposed solution and the outcomes of similar historical cases, a value of 0-1 is used; rule compliance. To integrate the results with the core layer The degree of conformity to the domain rules is measured, ranging from 0 to 1. Next, quantitative calculations of the indicators are performed, and the reasoning is based on the fit. The similarity between the reasoning basis and the key elements of the current scene is calculated to obtain: , in, This is the cosine similarity function. Feature vectors that represent the reasoning basis after multi-source information fusion The enhanced unified representation vector representing the i-th spatiotemporal unit is a high-dimensional feature vector incorporating domain semantics.

[0084] The similarity was calculated by integrating the treatment plan with the average similarity of treatment plans in similar historical cases. The result is judged by rule matching. If the fusion result fully conforms to all relevant rules, it is 1.0; if it violates one core rule, it is 0.6; and if it violates two or more core rules, it is 0.3.

[0085] Then, confidence level adjustment is performed, and the weights of the evaluation indicators are set. Calculate the comprehensive score of the evaluation indicators. According to the formula Calculate the final confidence level, where This is the final confidence level after correction (value 0-1).

[0086] Final confidence level classification: based on the final confidence level. Classify reliability levels and set high reliability ( ), medium reliability ( Low reliability ),like Return to re-execute the three-level reasoning and optimize the reasoning process.

[0087] In step S54, conflict resolution is performed on the fused reasoning results based on the resolution rules. For potential conflicts during the fused reasoning process (including conflicts in risk levels, handling methods, and reasoning basis), the core layer structured knowledge output in S4 is used to resolve these conflicts. Causal relationship at the deeper level The process employs principles of rule priority, causal support, and case evidence to resolve issues, ensuring logical consistency and compliance with emergency response protocols in the final reasoning. Specifically, this includes: Conflict identification and comparison of fusion reasoning results With each preliminary reasoning result The system identifies conflict types, including conflict of risk levels (e.g., the fusion result is high risk while some preliminary results are medium risk); conflict of handling plans (e.g., some plans recommend "immediate evacuation" while others recommend "on-site handling"); and conflict of reasoning basis (e.g., the causal chains of different results contradict each other).

[0088] Conflict resolution rules are set based on the core layer. The domain rules prioritize conflict resolution as follows: Life safety related rules > Emergency response guidelines > Causal logic rules > Historical case experience.

[0089] Conflict resolution is categorized as follows: For conflicts in risk levels, based on the core-level risk grading rules and combined with the causal relationships at the deeper level, the causal probability corresponding to different risk levels is calculated. The risk level with the highest causal probability that conforms to the risk grading rules is selected as the final risk level, while risk level descriptions that do not conform to the rules and causal logic are eliminated. For conflicts in response plans, priority is given to retaining response steps that comply with life safety-related rules and emergency response standards. The rationality of the response plan is analyzed in conjunction with the causal chain. If a plan does not conform to the "risk evolution trend" in the causal chain, it is adjusted. The experience of handling highly similar historical cases is referenced to integrate and form a unified response plan without conflicts. For conflicts in reasoning basis, the causal relationships at the deeper level are used... With this as the core, we verify the causal logic of each reasoning basis, eliminate causal contradictions and bases that are irrelevant to the current scenario, and supplement complete causal chains, matching rules and cases to ensure that the reasoning basis is logically coherent and mutually supportive.

[0090] Conflict resolution verification: After resolution, combine the experience and knowledge system of experts in the field of emergency rescue to verify whether the final reasoning result is free of logical conflicts and conforms to the norms. If conflicts still exist, repeat the above steps until the conflicts are completely resolved.

[0091] In step S55, the optimal emergency rescue decision recommendation is output. This includes: a clearly defined emergency risk level and its final confidence and reliability levels; a complete and operable emergency response plan, including the response process, key parameters, division of responsibilities, and timelines; detailed reasoning basis, including the revised causal chain, matching domain rules, similar historical cases, counterfactual intervention processes, and confidence assessment processes; applicable scenarios and precautions for the decision recommendation, including adjustments for response under special weather and terrain conditions; and compiling the above content into a standardized decision report, which is then output to the emergency rescue decision-making terminal to provide accurate, traceable, and operable decision support for emergency rescue personnel.

[0092] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit of the present invention should fall within the patent scope covered by the present invention.

Claims

1. A method for reasoning emergency rescue knowledge based on multi-source data using a large model, characterized in that, The method includes: Acquire multi-source heterogeneous emergency rescue data in emergency rescue scenarios; Based on spatiotemporal graph neural networks and large models, the multi-source heterogeneous emergency rescue data is preprocessed to obtain multi-source data with deep unified representation. The multi-source data is input into the emergency rescue knowledge model to obtain enhanced representation data, structured knowledge data, and causal relationship data for the multi-source data. A fusion reasoning algorithm is used to perform interpretable knowledge reasoning on the enhanced representation data, structured knowledge data, and causal relationship data to generate reasoning results; The reasoning results are evaluated and processed to output emergency rescue decision-making suggestions.

2. The emergency rescue knowledge reasoning method based on a large model and multi-source data according to claim 1, characterized in that, The multi-source heterogeneous emergency rescue data includes static data and dynamic data. The static data includes at least one of the following: rescue case library, emergency standard documents, rescue plan, and geographic information basic data. The dynamic data includes at least one of the following: environmental monitoring data, fire equipment operation data, on-site feedback data, and on-site video or image data.

3. The emergency rescue knowledge reasoning method based on a large model and multi-source data according to claim 1, characterized in that, The multi-source heterogeneous emergency rescue data is preprocessed based on a spatiotemporal graph neural network and a large model to obtain multi-source data with a deep unified representation, including: The first emergency rescue data is obtained by performing format normalization, cleaning and quality control on the multi-source heterogeneous emergency rescue data; The first emergency rescue data is mapped to a unified spatiotemporal unit, and various heterogeneous information is encoded and spliced ​​into a disaster vector of a unified dimension by combining a large model. A spatiotemporal correlation model is constructed using spatiotemporal units as nodes, disaster vectors as node features, and spatial adjacency and temporal dependency relationships as edges. Based on the spatiotemporal correlation model, deep unified representation learning is performed on the first emergency rescue data to output multi-source data with deep unified representation.

4. The emergency rescue knowledge reasoning method based on a large model and multi-source data according to claim 1, characterized in that, The method also includes enhancing emergency domain knowledge based on the multi-source data to construct the emergency rescue knowledge model: A pre-training dataset is constructed based on the multi-source data; The large language model is pre-trained based on the pre-trained dataset; The pre-trained large language model is then jointly trained. Construct an emergency rescue sub-scenario Prompt template and input it into the jointly trained large language model to generate an emergency rescue knowledge model.

5. The method for emergency rescue knowledge reasoning based on a large model and multi-source data according to claim 4, characterized in that, The construction of the pre-training dataset based on the multi-source data includes: A corpus is constructed based on the rescue and emergency response professional corpus from the multi-source data, and the rescue and emergency response professional corpus is associated and mapped with the representation vector to construct the first aligned dataset. Based on the emergency rescue-specific knowledge graph in the multi-source data, the entities and relationships in the graph are vector-encoded and associated with the representation vectors to construct a second aligned dataset. Virtual augmented samples are generated based on the small sample representation vectors in the multi-source data to construct a small sample dataset.

6. The emergency rescue knowledge reasoning method based on a large model and multi-source data according to claim 5, characterized in that, The pre-training of the large language model based on the pre-trained dataset includes: Based on the first aligned dataset, the large language model is pre-trained using corpus masking. Based on the second alignment dataset, knowledge graph alignment pre-training is performed on the large language model; Based on the aforementioned small sample dataset, the large language model is pre-trained with small sample enhancement.

7. The emergency rescue knowledge reasoning method based on a large model and multi-source data according to claim 1, characterized in that, The enhanced representation data, structured knowledge data, and causal relationship data are processed using a fusion reasoning algorithm to perform interpretable knowledge reasoning and generate reasoning results, including: Causal reasoning obtains the initial probability of the reasoning target based on causal correlation data processing; Rule-based reasoning involves matching structured knowledge data with domain rule expressions, obtaining reasoning conclusions based on the matched rules, and correcting the initial probability of the reasoning target based on the rules. Case-based reasoning, based on a comprehensive knowledge system and a historical case database for emergency rescue, extracts feature vectors from historical case scenarios, calculates the similarity between the feature vectors and the current scenario, selects the handling plan corresponding to the feature vectors of the historical case scenarios with high similarity, and supplements and optimizes the reasoning conclusions obtained based on rules. The reasoning results obtained from causal reasoning, rule-based reasoning, and case-based reasoning are fused and optimized using counterfactual intervention algorithms and confidence propagation algorithms.

8. The emergency rescue knowledge reasoning method based on a large model and multi-source data according to claim 7, characterized in that, The above-obtained reasoning results are fused and optimized using the counterfactual intervention algorithm, including: Identify the targets and scenarios for counterfactual intervention; Based on the aforementioned counterfactual intervention targets and scenarios, counterfactual hypotheses are constructed and intervention samples are generated. The intervention sample is input into the emergency rescue knowledge model for counterfactual reasoning, and the probability change is calculated; Key causal elements are selected based on the probability change, and the causal relationship data is corrected based on the key causal elements. Output counterfactual intervention results.

9. The method for emergency rescue knowledge reasoning based on a large model and multi-source data according to claim 7, characterized in that, The aforementioned inference results are fused and optimized using the confidence propagation algorithm, including: Obtain the reasoning results and corresponding initial confidence levels of the causal reasoning, rule-based reasoning, and case-based reasoning; With the reasoning target as the core node and the enhanced representation data, structured knowledge data, and causal relationship data mentioned in the reasoning as child nodes, a confidence propagation network is constructed. Based on the propagation path, the initial confidence of each sub-node is multiplied by the corresponding propagation coefficient to obtain the contribution confidence of each sub-node to the core node, and the sum of the contribution confidence is calculated to obtain the initial total confidence after fusion. The initial total confidence score is corrected by introducing a confidence score correction factor to obtain the final confidence score; By integrating the causal correlation data corrected by counterfactual intervention, the final confidence level obtained from confidence propagation, and the specific conclusions of the three-level inference, a complete inference result is output.

10. The emergency rescue knowledge reasoning method based on a large model and multi-source data according to claim 1, characterized in that, The evaluation and processing of the reasoning results to output emergency rescue decision-making suggestions includes: Construct a set of reasoning results based on the reasoning results; A weighted fusion algorithm is used to fuse the inference results based on the reliability weights of the inference results; The fused inference results are evaluated and corrected. Conflict resolution is performed on the fused inference results based on resolution rules; Once the conflict is resolved, emergency response decision-making recommendations will be generated.

Images