Multi-level emergency command intelligent decision method and system based on knowledge graph

Abstract

The application discloses a multi-level emergency command intelligent decision-making method and system based on a knowledge graph. The method comprises the following steps: S1, extracting the exclusive features of text modalities, visual modalities and space-time modalities from multi-source heterogeneous data in the whole scene of multi-level emergency command; S2, constructing a cross-modal attention fusion network to output multi-modal fusion features; S3, constructing a multi-modal joint entity relationship Transformer model, taking the multi-modal fusion features as input, and outputting standardized entity-relation triplets; S4, based on the entity-relation triplets, constructing a three-level hierarchical knowledge graph; S5, generating an emergency decision-making scheme by using a three-layer reasoning framework, and outputting the decision-making comprehensive confidence of the scheme; and S6, according to a preset emergency response level, combining the decision-making comprehensive confidence, and executing corresponding human-machine collaborative rules. In this way, intelligent sensing, accurate judgment, collaborative decision-making, efficient disposal and closed-loop optimization of the whole life cycle of an emergency event are realized.

Description

Technical Field

[0001] This invention relates to the fields of intelligent decision-making in emergency command, knowledge graphs, and multimodal deep learning, specifically to a multi-level intelligent decision-making method and system for emergency command based on knowledge graphs. Background Technology

[0002] With the continuous advancement of smart city construction, the emergency command system, as a core guarantee for urban safety operation, faces increasingly higher requirements in terms of intelligence, precision, and collaboration. Currently, emergency command systems still suffer from numerous core technological deficiencies and engineering implementation challenges in practical applications: First, there is insufficient capability for fusion of multi-source heterogeneous data. Data in emergency scenarios encompasses multiple modalities, including text, images, spatiotemporal sensing data, and structured business data. Existing systems struggle to achieve deep semantic alignment across modalities, resulting in large fusion errors that directly impact the accuracy of subsequent knowledge extraction and decision-making. Furthermore, significant barriers exist for cross-departmental data sharing, hindering the implementation of compliant data sharing and collaborative applications.

[0003] Secondly, the knowledge extraction and spatiotemporal modeling capabilities are weak. Existing systems mostly adopt pipeline-style entity-relation extraction schemes, which have significant error propagation problems and low extraction accuracy for overlapping and rare entities in emergency scenarios. Traditional knowledge graphs use static triple representations, which cannot explicitly model the spatiotemporal dynamic evolution characteristics of emergency events and are difficult to adapt to the rapid changes and risk propagation patterns of emergencies.

[0004] Third, multi-level command and coordination is inefficient. The existing system is difficult to adapt to the six-level command structure of national-provincial-municipal-district-street-on-site. Cross-level and cross-departmental instruction transmission is not smooth, the boundaries of authority and responsibility are blurred, and there is a lack of knowledge management and decision-making coordination mechanisms adapted to multi-level command, which easily leads to problems such as decision delays and disconnected handling.

[0005] Fourth, there is no effective control over the accumulation of errors across the entire process. In the multi-module cascading process of existing emergency systems, errors from upstream modules continue to propagate to downstream links, directly affecting the accuracy of the final decision. The lack of uncertainty quantification and error suppression mechanisms throughout the entire process results in insufficient system robustness.

[0006] Fifth, it is difficult to balance generalization ability and real-time performance. Traditional rule-based emergency response systems have poor generalization ability and are difficult to adapt to unknown emergency scenarios. The manual writing and maintenance of rule bases are extremely costly. On the other hand, intelligent solutions based on deep learning often have problems such as a large number of model parameters, difficulty in deployment at the edge, and a prominent contradiction between real-time response and continuous iterative optimization, resulting in high barriers to implementation.

[0007] To address the aforementioned issues, existing technologies largely involve simply migrating general knowledge graphs and deep learning techniques to the emergency response domain, without making substantial technical improvements tailored to the specific business characteristics and core pain points of emergency command. Therefore, these solutions cannot simultaneously meet the multiple requirements of emergency command scenarios regarding accuracy, real-time performance, robustness, collaboration, and feasibility. Summary of the Invention

[0008] To address the aforementioned deficiencies in existing technologies, the present invention aims to provide a knowledge graph-based multi-level emergency command intelligent decision-making method and system. Starting from the core business needs of emergency command, this invention systematically solves the core problems of existing emergency command systems, such as low accuracy of multimodal data fusion, lack of control over error accumulation, weak spatiotemporal dynamic modeling capabilities, low efficiency of multi-level collaboration, insufficient generalization ability and real-time performance, and high implementation thresholds, through core algorithm innovation and architecture design optimization. This enables intelligent perception, accurate judgment, collaborative decision-making, efficient handling, and closed-loop optimization throughout the entire lifecycle of emergency events, while ensuring that the system possesses outstanding creativity, feasibility, and cross-scenario adaptability.

[0009] In a first aspect, embodiments of the present invention provide a multi-level emergency command intelligent decision-making method based on knowledge graphs, comprising the following steps: S1: Collect and standardize the multi-source heterogeneous data in the multi-level emergency command scenario, and extract the exclusive features of text modality, visual modality and spatiotemporal modality respectively; S2: Construct a cross-modal attention fusion network that includes a spatiotemporal-aware multi-head attention mechanism, perform deep semantic alignment and fusion on the extracted text modality, visual modality and spatiotemporal modality-specific features, and output multimodal fusion features and their fusion confidence. S3: Construct a multimodal joint entity relationship Transformer model, using the multimodal fusion features as input, to jointly extract emergency entities and relationships between entities end-to-end, and output standardized entity-relationship triples and their extraction confidence. S4: Based on the entity-relationship triples, a three-level hierarchical knowledge graph adapted to a multi-level command architecture is constructed using the five-tuple spatiotemporal knowledge representation method, and a comprehensive confidence level is assigned to each piece of knowledge in the three-level hierarchical knowledge graph. S5: Based on the three-level hierarchical knowledge graph, an emergency decision-making scheme is generated using a three-layer reasoning framework that integrates rule deduction, analogical reasoning, and graph structure analysis, and the comprehensive confidence level of the emergency decision-making scheme is output. S6: Based on the preset emergency response level and the comprehensive confidence level of the decision, execute the corresponding human-machine collaboration rules to conduct hierarchical review, instruction issuance and execution feedback of the emergency decision-making plan, forming a closed-loop control of the entire disposal process.

[0010] Furthermore, step S1 involves extracting specific features for the text modality, visual modality, and spatiotemporal modality, respectively, specifically including: Text modality feature extraction sub-step: A hierarchical semantic feature extraction architecture is adopted, which is built based on a distilled version of the BERT model and a bidirectional gated recurrent unit. First, the preprocessed text sequence is input into the distilled version of the BERT model to generate context-aware word-level embeddings. Then, the word-level embeddings are input into the bidirectional gated recurrent unit, and sentence-level feature aggregation is performed in combination with the attention mechanism to output the text modality-specific features and their semantic confidence. Visual modality feature extraction sub-steps: Construct a domain-adaptive multi-stage lightweight visual feature extraction framework based on a lightweight convolutional neural network and a feature pyramid network; input preprocessed image or video keyframes into the lightweight convolutional neural network and the feature pyramid network to extract multi-scale visual features; highlight key regions related to emergency events through an attention-based region proposal network, and output the specific features of the visual modality and its visually valid confidence. Spatiotemporal modality feature extraction sub-step: A lightweight tensor modeling method based on one-dimensional convolution and positional encoding is adopted; for the spatial coordinates and timestamps of each emergency event or emergency entity, spatial positional encoding and time encoding are constructed respectively, and the two are concatenated to generate a spatiotemporal embedding vector; all spatiotemporal embedding vectors are organized into a three-dimensional tensor structure that retains the time dimension, spatial dimension and feature dimension, and the exclusive features of spatiotemporal modality are output. The feature extraction networks for the text modality, visual modality, and spatiotemporal modality are all adapted to the lightweight requirements of edge deployment and output confidence scores for the corresponding modality features.

[0011] Furthermore, the cross-modal attention fusion network constructed in step S2 specifically includes: The text-image interaction attention module is used to construct a bidirectional cross-modal attention mechanism. It calculates the attention weights of text features on visual features and the attention weights of visual features on text features, and enhances the text features and visual features based on the attention weights to achieve deep semantic interaction between the text modality and the visual modality. The spatiotemporal-aware multi-head attention module is used to superimpose the spatiotemporal distance matrix constructed based on spatiotemporal location encoding as a bias term onto the attention score for weighted adjustment, so as to constrain the model to prioritize the key features of spatiotemporal coherence and output a multimodal feature representation that integrates spatiotemporal semantics.

[0012] Furthermore, the cross-modal attention fusion network also includes: The feature space alignment module is used to align the feature distributions of text modality and visual modality based on optimal transport theory by minimizing the Wasserstein distance between the feature distributions of text modality and visual modality, and by combining the loss function of downstream tasks, thereby eliminating the distribution heterogeneity between features of different modalities. The fusion feature confidence quantification module is used to comprehensively calculate the overall confidence score of multimodal fusion features based on intermodal semantic matching degree, feature integrity, and feature alignment loss.

[0013] Furthermore, in step S3, the multimodal joint entity relationship Transformer model specifically includes: The Transformer encoder submodule is used to perform deep context encoding on the multimodal fusion features, capture long-distance dependencies between features, and generate context-aware feature representations. The entity boundary detection submodule uses a sequence labeling method based on span annotation to predict the start and end positions of entities from the context-aware feature representation. The entity type classification submodule is used to classify the detected entities by type; The relation classification submodule uses a table-filling method to model the semantic relationships between all entity pairs, thereby achieving joint classification of relationships between entities. The three sub-modules of entity boundary detection, entity type classification, and relation classification serve as a joint prediction head. Based on the shared context-aware feature representation, they perform end-to-end joint training and inference, and synchronously output standardized entity-relation triples and their corresponding extraction confidence scores.

[0014] Furthermore, step S3 also includes: The confidence quantification sub-step of the extraction results calculates the entity confidence and relation confidence respectively; the entity confidence is determined based on the entity boundary prediction probability, the entity type classification probability, and the compliance score of whether the entity conforms to the semantic constraints of the top-level ontology; the relation confidence is determined based on the relation classification probability, the confidence of the subject and object entities, and the compliance score of whether the relation conforms to the semantic constraints of the top-level ontology. The confidence threshold dynamic adjustment sub-step presets a confidence threshold associated with the risk level of the emergency scenario, and automatically adjusts the confidence threshold according to the risk level or data quality of the current emergency event. For extraction results that are lower than the confidence threshold, manual verification or downgrade processing is triggered. The model training optimization sub-step employs a multi-task joint loss function to train the multimodal joint entity-relation Transformer model. This multi-task joint loss function integrates entity boundary detection loss, entity type classification loss, and relationship classification loss, and introduces focus loss to address the imbalance in the distribution of entity and relationship categories. Simultaneously, a semi-supervised few-shot cross-scene transfer learning strategy is adopted, using a three-stage training framework of pre-training, domain adaptation, and fine-tuning to reduce the annotation cost of cross-scene adaptation.

[0015] Furthermore, in step S4, the three-level hierarchical knowledge graph includes: The overall map, deployed at the central node in the cloud, is used to integrate core knowledge across the entire domain and support global situational awareness and cross-regional collaborative decision-making. The sub-maps deployed at regional sub-nodes are used to focus on emergency knowledge in corresponding administrative regions or industry sectors to support command and analysis at the local level. A scenario sub-map deployed at edge-end field response nodes is used to support rapid on-site response and precise handling for specific emergency scenarios.

[0016] Furthermore, in step S5, the three-layer reasoning framework specifically includes: Rule-based deductive reasoning: Based on a pre-built emergency domain reasoning rule base, logical deduction is performed to generate decision recommendations that conform to expert experience and national standards; Embedded analogical reasoning: Based on the embedding of entities and relationships in a knowledge graph, retrieve historical emergency response cases that are most similar to the current event and reuse their successful response plans; Graph structure analysis and reasoning: Analyze the topological structure of the knowledge graph using graph algorithms to identify key nodes, bottlenecks, and potential risk propagation paths in emergency response; The results of the above three reasoning methods are combined to generate the final emergency decision-making plan.

[0017] Furthermore, the method also includes: an automatic rule mining and optimization step, specifically including: An improved rule mining algorithm is used, which combines spatiotemporal constraints with domain ontology semantic constraints, to automatically mine candidate reasoning rules from the knowledge graph of historical emergency cases; The candidate rules are tested for causal effect using a causal inference framework, eliminating spurious rules and retaining rules with causal relationships. To calculate the support and confidence indices of rules that pass the causal test, the rules are screened and then reviewed by experts before being added to the database, so as to realize the semi-automatic construction and low-cost maintenance of the inference rule base.

[0018] Secondly, embodiments of the present invention also provide a multi-level emergency command intelligent decision-making system based on knowledge graphs, including: The multi-source data access and preprocessing module is used to collect and standardize multi-source heterogeneous data in the full scenario of multi-level emergency command, and extract exclusive features of text modality, visual modality and spatiotemporal modality respectively. The cross-modal attention fusion module is used to construct a cross-modal attention fusion network that includes a spatiotemporally aware multi-head attention mechanism. It performs deep semantic alignment and fusion on the extracted text modality, visual modality and spatiotemporal modality-specific features, and outputs multimodal fusion features and their fusion confidence. The multi-scenario joint entity-relationship extraction module is used to construct a multimodal joint entity-relationship Transformer model. Taking the multimodal fusion features as input, it extracts emergency entities and the relationships between entities end-to-end, and outputs standardized entity-relationship triples and their extraction confidence. The spatiotemporal knowledge graph construction and management module is used to construct a three-level hierarchical knowledge graph adapted to a multi-level command architecture based on the entity-relation triplet and using the five-tuple spatiotemporal knowledge representation method, and to assign a comprehensive confidence level to each piece of knowledge in the three-level hierarchical knowledge graph. The multi-level emergency command reasoning and decision-making module is used to generate emergency decision-making schemes based on the three-level hierarchical knowledge graph, using a three-layer reasoning framework that integrates rule deduction, analogical reasoning and graph structure analysis, and outputs the comprehensive confidence level of the emergency decision-making scheme. The multi-scenario application adaptation and human-machine collaboration module is used to execute corresponding human-machine collaboration rules based on the preset emergency response level and the comprehensive confidence level of the decision, to conduct hierarchical review, instruction issuance and execution feedback of the emergency decision plan, forming a closed-loop control of the entire disposal process.

[0019] Compared with existing technologies, the present invention achieves the following beneficial effects: (1) This invention achieves deep semantic alignment of multimodal emergency data by using the self-designed CAFN cross-modal attention fusion network and SAMA spatiotemporal awareness multi-head attention mechanism. Combined with the end-to-end MJERT joint entity-relation extraction model, the entity-relation extraction F1 score reaches 89.0% on the self-built multimodal emergency dataset, which is 8.7% higher than the traditional self-attention method. At the same time, the constructed full-link confidence quantification and control system realizes the quantification of uncertainty and suppression of error propagation throughout the entire process from data input to decision output. It can still maintain an effective accuracy of more than 79% in the single-modal missing scenario. The system robustness is significantly better than the existing technical solutions.

[0020] (2) This invention adopts a four-dimensional spatiotemporal ontology architecture of "general ontology + scenario extension", which is natively adapted to five core emergency scenarios: smart transportation, smart community, smart campus, smart government affairs, and smart tour guide, and supports rapid expansion to other emergency fields; the supporting semi-supervised few-sample learning framework reduces the cost of single-scenario annotation by more than 90%; the improved AMIE+ rule automatic mining mechanism reduces the cost of manual maintenance of the rule base by more than 80%; the horizontal federated learning framework realizes that cross-departmental data is "usable but not visible", effectively breaking the data sharing barrier, while meeting data compliance requirements, and significantly reducing the implementation threshold of the system.

[0021] (3) The three-level hierarchical graph architecture and human-machine collaboration mechanism that are aligned with national standards proposed in this invention realize multi-level efficient collaborative command of "global-local-on-site". On a test set of 1200+ real historical emergency cases, the F1 score of the comprehensive emergency decision support task reached 94.0%, which is 23.7% higher than the traditional rule-based system and 11.5% higher than the conventional knowledge graph scheme. It completely solves the core pain points of information barriers, poor collaboration and decision delay in traditional multi-level command.

[0022] (4) The five-tuple spatiotemporal knowledge representation method proposed in this invention can accurately model the spatiotemporal evolution law of emergency events, breaking through the modeling limitations of traditional static knowledge graphs; the matching "hot data memory graph + cold data disk graph" hybrid storage architecture realizes millisecond-level graph query and second-level dynamic update, which can capture the situational changes of emergency events in real time and provide a reliable knowledge foundation for dynamic decision-making.

[0023] (5) The cloud-edge dual-track operation architecture designed in this invention completely solves the core contradiction between real-time response and continuous iterative optimization. The real-time inference link achieves second-level response, and the offline iterative link does not affect the real-time business operation. The lightweight model design supports edge deployment and can run stably on low-computing-power devices such as NVIDIA Jetson Xavier NX. At the same time, it supports a progressive implementation path, which can start from a single scenario for verification and then gradually expand to the whole scenario, greatly reducing the risk of engineering implementation. Attached Figure Description

[0024] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart illustrating a multi-level emergency command intelligent decision-making method based on knowledge graphs provided in an embodiment of the present invention; Figure 2This is the overall architecture diagram of the multi-level emergency command intelligent decision-making system based on knowledge graphs provided in this embodiment of the invention; Figure 3 This is a schematic diagram of the cross-modal attention fusion network (CAFN) according to an embodiment of the present invention; Figure 4 This is a diagram of the five-tuple spatiotemporal knowledge representation and three-level hierarchical graph architecture of an embodiment of the present invention; Figure 5 This is a schematic diagram of a module of a multi-level emergency command intelligent decision-making system based on knowledge graphs provided in an embodiment of the present invention. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0026] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as being processed sequentially, many of these operations (or steps) may be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations may be rearranged. The process may be terminated when its operation is completed, but may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0027] I. Overview of the Overall Technical Solution of the Invention This invention addresses the core technical shortcomings of existing emergency command systems, such as insufficient multi-source heterogeneous data fusion capabilities, poor cross-scenario adaptability, low entity relationship extraction accuracy, weak spatiotemporal dynamic evolution modeling capabilities, low efficiency of multi-level command and collaborative decision-making, lack of control over the accumulation of multi-module cascaded errors, and insufficient generalization and real-time performance of traditional rule-based systems. It proposes a multi-level emergency command intelligent decision-making method and system that integrates multimodal deep learning and spatiotemporal knowledge graphs.

[0028] This invention constructs a full-link technical architecture encompassing "top-level ontology design, data perception, cross-modal fusion, knowledge extraction, graph construction, reasoning and decision-making, scenario implementation, and closed-loop optimization." At its core, it utilizes the proposed Cross-modal Attention Fusion Network (CAFN), Multi-modal Joint Entity Relation Transformer (MJERT), and spatiotemporal knowledge graph representation and reasoning mechanisms to adapt to the emergency command needs of various scenarios such as smart transportation, smart communities, smart campuses, smart government affairs, and smart tour guides. This architecture builds a hierarchical, scalable, highly real-time, and robust multi-level emergency command system.

[0029] The core inventive points of this invention (the essential difference from existing technologies, the core of patent protection) include: (1) Invent a spatiotemporal-aware multi-head attention (SAMA) mechanism that integrates spatiotemporal distance constraints, and construct a CAFN cross-modal attention fusion network to solve the industry pain points of difficult semantic alignment and large fusion error in multimodal emergency data; (2) A five-tuple spatiotemporal knowledge representation method and a three-level hierarchical graph architecture adapted to multi-level emergency command are proposed to solve the problems that traditional knowledge graphs cannot model the spatiotemporal dynamic evolution of emergency events and the difficulty of cross-level coordination of multi-level command. (3) Construct a full-link confidence quantification and error control system, and give a full-process confidence mathematical model from data quality, feature fusion, knowledge extraction to reasoning decision-making, so as to realize the active suppression of multi-module cascade error and solve the core defect of traditional system error accumulation without control. (4) Design a multi-level emergency reasoning framework that integrates three layers: rule deduction, analogical reasoning, and graph structure analysis. It is equipped with an improved AMIE+ rule automatic mining and causal verification mechanism to significantly reduce the manual maintenance cost of the rule base and solve the problems of poor generalization ability and difficulty in scenario adaptation of traditional emergency systems. (5) A cloud-edge dual-track operation architecture and a horizontal federated learning (FL) collaborative optimization scheme are proposed to completely solve the core contradiction between the emergency system's "real-time response requirements" and "continuous iterative optimization". At the same time, it breaks the barriers to cross-departmental data sharing and adapts to the lightweight deployment requirements of the edge.

[0030] This invention enables intelligent sensing, accurate assessment, collaborative decision-making, efficient handling, and closed-loop optimization throughout the entire lifecycle of emergency events. Its core performance indicators have clear testing, verification, and quantitative bases. (1) Multi-scenario joint entity-relationship extraction task: On the self-built multimodal emergency dataset (integrating 1200+ real emergency event cases from 2018 to 2025 across the country, covering five major scenarios, including 28600+ text reports, 15300+ image data, 32000+ spatiotemporal sensor data, and 5200+ labeled samples), the F1 score reached 89.0%, which is 8.7% higher than the standard Transformer self-attention baseline model (F1=80.3%), and the average F1 score across scenarios is no less than 85.2%; at the same time, cross-domain validation was completed on the public emergency dataset CrisisMMD, with an F1 score of 86.2%, which has good generalization ability.

[0031] (2) Emergency decision support comprehensive tasks (resource allocation, path planning, and response plan generation): On a test set of 1200+ real historical emergency cases, the fully integrated solution achieved an F1 score of 94.0%, which is 23.7% higher than the traditional rule-based emergency system (F1=70.3%) and 11.5% higher than the conventional knowledge graph solution (F1=82.5%).

[0032] II. Top-level Design: Four-dimensional Spatiotemporal Ontology Architecture and Multi-scenario Entity-Relationship System in the Emergency Response Domain This section provides the top-level semantic specifications for the entire solution. It prioritizes the standardized definition of the domain ontology and entity-relationship system, thoroughly resolving the circular dependency problem between entity-relationship extraction and ontology design, and providing a unified semantic benchmark for subsequent knowledge extraction, graph construction, and reasoning decision-making.

[0033] 2.1 Four-dimensional spacetime ontology architecture design This invention constructs a four-dimensional spatiotemporal ontology architecture consisting of a "static concept layer, spatial layer, temporal layer, and scene adaptation layer," achieving semantic unification across scenes and layers, as detailed below: Static Concept Layer: Defines common entity types, relationship types, attribute types and their classification levels in the emergency response field, clarifies the semantic constraints, domains and value ranges of entities and relationships, and provides a unified conceptual benchmark for the entire scenario; Spatial layer: Defines geospatial elements, topological relationships (adjacent, contained, connected, intersecting, etc.), spatial coordinate specifications, and administrative region levels (national-provincial-municipal-district-street-community / park / scenic area), adapting to the spatial hierarchical management needs of multi-level command; Time layer: Defines time points, time intervals, and temporal relationships (before, after, synchronous, continuous, etc.), supports multi-granularity time representation (second-level emergency events, hour-level handling processes, day-level recovery phases), and adapts to the dynamic evolution characteristics of emergency events; Scene Adaptation Layer: Defines exclusive ontology extension rules, emergency response procedures, and decision constraints for five major scenarios: smart transportation, smart community, smart campus, smart government affairs, and smart tour guide. This decouples the general ontology from scene-specific knowledge and supports the rapid expansion of new emergency scenarios.

[0034] 2.2 Definition of Standardized Entity-Relationship System for Multiple Scenarios Based on the aforementioned four-dimensional spatiotemporal ontology architecture, this invention defines a standardized and scalable emergency entity type and relation type system for five core scenarios, clarifies the semantic constraints of each type of entity / relationship, and ensures the consistency of the extraction results with the ontology specification. The core system is shown in Table 1 below: Table 1

[0035] Example 1 like Figure 1 The diagram shown is a flowchart illustrating a multi-level emergency command intelligent decision-making method based on knowledge graphs provided in an embodiment of the present invention; as shown... Figure 2 The diagram shown is the overall architecture of the multi-level emergency command intelligent decision-making system based on knowledge graphs provided in this embodiment of the invention. This method incorporates a full-link confidence quantification and error control system throughout the entire process: each core step outputs a confidence score (with a clearly defined mathematical calculation model), and downstream modules calculate based on the confidence score weights. Results below the confidence threshold trigger manual verification or downgrade processing, fundamentally suppressing the problem of error accumulation in multi-module cascading. Specifically, it includes the following steps: S1: Collect and standardize the multi-source heterogeneous data in the multi-level emergency command scenario, and extract the exclusive features of text modality, visual modality and spatiotemporal modality respectively; Step S1 is used to perform preprocessing and modality-specific feature extraction of multi-source heterogeneous emergency data. This step addresses the multi-modal heterogeneous data across multiple emergency command scenarios by constructing a standardized preprocessing and feature extraction pipeline. Simultaneously, it designs a horizontal FL data processing framework and a lightweight feature extraction network to solve the problems of cross-departmental data sharing barriers and insufficient edge computing power. Specifically, it includes the following steps: S11. Multi-scenario, multi-source emergency data collection and cleaning S111, Comprehensive Collection of Multi-Source Emergency Data Targeting five core application scenarios—smart transportation, smart communities, smart campuses, smart government affairs, and smart tour guides—we have completed the comprehensive collection of emergency data, which is categorized into four main types: (1) Text data: emergency response documents, on-site reporting information, government notices, public opinion information, incident alarm records, emergency plan texts, historical response cases, scenic spot / campus / community announcements, etc.; (2) Visual data: road checkpoint monitoring, campus / community / scenic spot video surveillance, drone aerial footage, on-site law enforcement recorder images, satellite remote sensing images, incident scene pictures, etc.; (3) Spatiotemporal sensing data: road network operation data, personnel positioning data, Internet of Things (IoT) sensor data (fire protection, gas, water quality, earthquake, temperature and humidity, etc.), vehicle global positioning system (GPS) data, passenger flow statistics, meteorological and hydrological data, emergency material positioning data, geographic information system (GIS) data, etc. (4) Structured business data: emergency organizational structure data, emergency team / material / equipment ledgers, personnel information database, basic site information, emergency response process specifications, etc.

[0036] S112, Standardized Pretreatment For the collected multi-source data, this invention performs four standardized preprocessing steps: data cleaning, format standardization, scale normalization, and missing value imputation. Specifically: (1) Text data: Complete deduplication, noise reduction, word segmentation, stop word filtering, entity annotation and standardization processing; (2) Visual data: Complete deblurring, distortion correction, resolution unification, and keyframe extraction processing; (3) Spatiotemporal sensing data: complete outlier removal, time series alignment, spatial coordinate standardization, and missing value interpolation based on spatiotemporal neighborhood; (4) Structured data: Complete field normalization, encoding standardization, and data integrity verification.

[0037] After preprocessing, a data quality confidence score is output for each data point, which serves as the weighting basis for subsequent feature extraction, thus achieving pre-emptive error control.

[0038] S113, Design of Federated Learning Data Processing Framework To address the barriers to cross-departmental data sharing, this invention designs a dual-track FL framework of "offline federated collaborative training + real-time local inference," implemented based on TensorFlow Federated (TFF), while simultaneously resolving the conflict between FL and emergency real-time requirements. The specific design is as follows: Offline training phase: A horizontal FL architecture is adopted. After the local data of each department is preprocessed and features are extracted, only the encrypted model gradient is uploaded. The original data does not leave the local domain throughout the process. At the same time, gradient sparsity compression (compression ratio 10:1) + asynchronous update mechanism is adopted to reduce cross-node communication overhead by more than 85%. The federated collaborative optimization of the model is only completed during non-emergency periods, without occupying real-time computing resources for emergency response. Real-time inference phase: Each node uses a fixed, stable model version. All preprocessing and feature extraction operations are completed locally, with no cross-node federated communication, ensuring a response time within seconds. Privacy compliance design: Gaussian noise is added during the gradient upload process, and a differential privacy (DP) mechanism is adopted to meet the compliance requirements of the "Data Security Law of the People's Republic of China" and the "Personal Information Protection Law of the People's Republic of China".

[0039] S12, Modality-Specific Lightweight Feature Extraction For the three core data types of preprocessed text, visual, and spatiotemporal data, this invention constructs lightweight, dedicated feature extraction networks, clarifies the network architecture and edge computing power assessment, and adapts to the low-computing-power environment of on-site processing. Simultaneously, a semi-supervised learning framework is used to reduce annotation costs. The design of the three feature extraction networks is as follows: S121. Text Semantic Feature Extraction A hierarchical semantic feature extraction architecture is adopted, with the core based on a distilled version of BERT (DistilBERT) model (or: a distilled bidirectional encoder representation model or a lightweight BERT model) - a base version - Chinese (DistilBERT-base-chinese) and a bidirectional gated recurrent unit (BiGRU), balancing lightweight design with feature extraction accuracy. In a specific embodiment of this invention: a hierarchical semantic feature extraction architecture is adopted, based on a distilled version of BERT model and a bidirectional gated recurrent unit; firstly, the preprocessed text sequence is input into the distilled version of BERT model to generate context-aware word-level embeddings; then, the word-level embeddings are input into the bidirectional gated recurrent unit, and sentence-level feature aggregation is performed in conjunction with an attention mechanism to output the text modality-specific features and their semantic confidence. Specifically: (1) Word-level embedding generation: The preprocessed text sequence is input into the DistilBERT-base-chinese model (768 hidden layer dimensions, 6 Transformer layers, 12 attention heads, 66M parameters) to generate context-aware word-level embeddings. The calculation formula is as follows: ; In the formula, For the first The first sentence Word-level embedding vectors of each word; For the first The first sentence One word; For the first One sentence; The total number of sentences in the text sequence; For the first The total number of words in each sentence; This is a distilled version of the BERT model function, pre-trained on Chinese corpus, used to encode input text words and corresponding sentences, generating word-level embedding vectors that integrate contextual semantics. It is the core encoding function for text semantic feature extraction in this invention.

[0040] (2) Sentence-level feature aggregation: The word-level embedding vectors are input into a 2-layer BiGRU (hidden layer dimension 128, dropout rate 0.3), and the sentence-level feature aggregation is completed by combining the attention mechanism. The calculation formula is as follows: ; ; In the formula, For the first The first sentence Attention weights for each word; These are learnable attention parameters; For the first Sentence-level feature vectors of each sentence; The natural exponential function is used for normalization during the attention weight calculation process to ensure that the output attention weights conform to the probability distribution characteristics. It is the core operation function of the attention mechanism. This is a learnable weight parameter matrix in the attention mechanism, used to perform a linear transformation on the word-level embedding vector to calculate the corresponding attention weights, which can be continuously optimized through model training; For learnable attention parameter matrix The transpose of the matrix is ​​used to adapt to the dimensional requirements of matrix operations and complete the core calculation of attention weights.

[0041] (3) Confidence output: Output the semantic confidence of the text features. Text data with a confidence level below 0.75 will trigger a secondary verification. The inference time for a single text is ≤50ms and the memory usage at the edge is ≤200MB.

[0042] S122, Visual Feature Extraction A domain-adaptive, multi-stage lightweight visual feature extraction framework is constructed, with its core based on a lightweight convolutional neural network V2 (MobileNetV2) + feature pyramid network (FPN), adapting to the multi-scale feature extraction requirements of emergency images. In a specific embodiment of this invention, a domain-adaptive, multi-stage lightweight visual feature extraction framework is constructed based on a lightweight convolutional neural network and a feature pyramid network; preprocessed image or video keyframes are input into the lightweight convolutional neural network and feature pyramid network to extract multi-scale visual features; a region proposal network with an attention mechanism highlights key regions related to the emergency event, outputting the specific features of the visual modality and its visually valid confidence score. Specifically: (1) Feature extraction: The preprocessed image / video keyframes are input into the MobileNetV2+FPN backbone network (with only 3.5M parameters) to complete the visual feature extraction. The calculation formula is as follows: ; In the formula, The extracted visual feature vector; For visual feature extraction networks; For the input image / video keyframes; The network parameters are fine-tuned based on a multi-scenario emergency image dataset.

[0043] (2) Highlighting key areas: Through the regional proposal network with attention mechanism, key areas related to emergency events (such as traffic accidents, crowd gatherings, facility damage, etc.) are highlighted. (3) Confidence output: Output the effective confidence of visual features, automatically filter features without critical emergency areas, and calculate the inference time of a single 1080P image. Edge memory usage .

[0044] S123, Spatiotemporal Feature Extraction Step S123 employs a lightweight tensor modeling method based on one-dimensional convolution and positional encoding. For each emergency event or entity, spatial positional encoding and time encoding are constructed based on its spatial coordinates and timestamp, and then concatenated to generate a spatiotemporal embedding vector. All spatiotemporal embedding vectors are organized into a three-dimensional tensor structure that retains the time, spatial, and feature dimensions, outputting the specific features of the spatiotemporal modality. Specifically: A lightweight tensor modeling method using 1D convolution and positional encoding is employed to preserve the spatiotemporal three-dimensional intrinsic structure of emergency data, specifically: (1) Spatiotemporal embedding vector construction: spatial coordinates of each emergency event / entity With timestamp Construct the spatiotemporal embedding vector, and calculate it using the following formula: ; In the formula, It is a spatiotemporal embedding vector; Encoding the position of spatial coordinates; Time encoding for time series; This is a vector concatenation operation.

[0045] (2) Tensor structure organization: Organize all spatiotemporal embedding vectors into tensor structures In the formula, For the time dimension, For spatial dimensions, As a feature dimension, it fully preserves the spatiotemporal relationship between emergency entities and events.

[0046] (3) Inference efficiency: Encoding time for a single spatiotemporal sequence Edge memory usage .

[0047] S124, Actual Evaluation of Edge Computing Power This invention presents a field test of edge computing power for three types of lightweight feature extraction networks. The test hardware is an NVIDIA Jetson Xavier NX edge industrial control machine (21 TOPSAI computing power, 8GB memory). The test results are as follows: when the three lightweight models run in parallel, the total time for a single round of full-process feature extraction is ≤150ms, the peak memory usage is ≤400MB, and the CPU utilization rate is ≤45%, which fully meets the real-time and computing power constraints of edge-side on-site processing. For lower-configuration embedded devices, INT8 quantization can be further adopted, which can improve the inference speed by 2 times and reduce the accuracy loss by ≤2%.

[0048] S125, Cost Optimization Scheme To address the scarcity of labeled data in the emergency response field, this invention provides a semi-supervised self-training + active learning framework to reduce labeling costs, specifically: (1) Sample requirements: Only 500+ high-quality labeled samples are needed for a single scenario (covering more than 80% of the core entity types and relationship types in the scenario, with a single sample labeling time of ≤15 minutes), along with 10,000+ unlabeled scenario data; (2) Training process: pseudo-labels are generated for unlabeled data through self-training, and then high-value samples are selected through active learning to supplement the labels. The model accuracy can reach more than 95% of the fully supervised scheme. (3) Cost quantification: Compared with the traditional full supervision solution that requires 5,000+ labeled samples per scenario, the labeling workload is reduced by 90%, that is, the labeling cost is reduced by more than 90%.

[0049] S2: Construct a cross-modal attention fusion network that includes a spatiotemporal-aware multi-head attention mechanism, perform deep semantic alignment and fusion on the extracted text modality, visual modality and spatiotemporal modality-specific features, and output multimodal fusion features and their fusion confidence. Step S2 is used to achieve cross-modal fusion and feature space alignment based on a cross-modal attention fusion network (CAFN). This invention constructs a hierarchical CAFN cross-modal attention fusion network, incorporating a self-designed SAMA mechanism to achieve deep semantic alignment of textual, visual, and spatiotemporal features, while outputting confidence scores for the fused features, thus solving the problems of semantic fragmentation and large fusion errors in heterogeneous data across multiple scenarios. Figure 3 The diagram shown is a schematic representation of the cross-modal attention fusion network (CAFN) according to an embodiment of the present invention.

[0050] Furthermore, the cross-modal attention fusion network constructed in step S2 specifically includes: a Text-Image Interaction Attention (TIIA) module, used to construct a bidirectional cross-modal attention mechanism, calculating the attention weights of text features on visual features and the attention weights of visual features on text features, and enhancing text features and visual features based on the attention weights to achieve deep semantic interaction between the text modality and the visual modality; and a Spatiotemporal Aware Multi-Head Attention (SAMA) module, used to introduce a spatiotemporal distance matrix constructed based on spatiotemporal position encoding during the multi-head attention calculation process, and superimposing this spatiotemporal distance matrix as a bias term onto the attention score, so that when calculating the similarity between the query and the key, the attention weights are adaptively adjusted according to the proximity of features in the temporal and spatial dimensions, so that the model prioritizes key features with coherence in the spatiotemporal dimension and outputs a multimodal feature representation that integrates spatiotemporal semantics. Furthermore, the cross-modal attention fusion network also includes: a feature space alignment module, which, based on optimal transport theory, aligns the feature distributions of the text and visual modalities by minimizing the Wasserstein distance between them and combining it with the loss function of the downstream task, thereby eliminating the heterogeneity of feature distributions between different modalities; and a fusion feature confidence quantification module, which comprehensively calculates the overall confidence score of the multimodal fusion features based on the semantic matching degree between modalities, feature integrity, and feature alignment loss. Specifically: S21, Text-Image Interaction Attention Module A bidirectional cross-modal attention mechanism is constructed to model the bidirectional semantic association between textual descriptions and visual evidence, achieving mutual enhancement between textual and visual features. The specific calculation is as follows: (1) Similarity score calculation: for the text feature matrix With visual feature matrix The similarity score between text and image pairs is calculated using the following formula: ; In the formula, For the first The text features and the first Similarity score of visual features; It is a learnable projection matrix (768 dimensions). For the first feature sequence of the text One feature; The first of the visual feature sequences One feature; This is the scaling factor; This is a matrix transpose operation.

[0051] (2) Attention weight calculation: The attention weight is calculated based on the similarity score, and the formula is: ; In the formula, For the first The visual feature is related to the first Attention weights for each text feature; The number of features in the visual feature sequence.

[0052] (3) Bidirectional feature enhancement: Text features are enhanced based on attention weights, and the formula is as follows: ; In the formula, For the enhanced first One text feature; The projection matrix is ​​learnable (768 dimensions); similarly, the text-to-image inverse attention is computed to generate enhanced visual features. This enables bidirectional semantic enhancement of both textual and visual features.

[0053] S22, Spatiotemporal Awareness Multi-Head Attention Module This invention, based on the traditional multi-head attention mechanism, introduces spatiotemporal position encoding and a spatiotemporal distance matrix, and designs a SAMA mechanism to integrate spatiotemporal constraints into the attention calculation process. This allows the model to prioritize semantically relevant and spatiotemporally coherent features, adapting to the spatiotemporal evolution characteristics of emergency events. The specific calculation is as follows: (1) Single-head spatiotemporal attention output: for feature sets of multimodal fusion The formula for calculating the spatiotemporal attention output of a single head is: ; In the formula, For the first The spatiotemporal attention output of each attention head; , , For the first The query, key, and value matrix of each attention head; Let be the dimension of the key matrix; This is the spatiotemporal distance balance parameter (default value 0.6, determined based on grid search optimization). This is the spatiotemporal distance matrix; This is the normalization exponential function, used to convert the raw scores of spatiotemporal attention calculation into a probability distribution form, thereby normalizing the attention weights. It is the core normalization operation function in the SAMA module.

[0054] (2) Multi-head attention aggregation: The outputs of multiple attention heads are concatenated and aggregated to obtain the final output of the SAMA module. The formula is as follows: ; In the formula, The number of attention heads (default value 8); The projection matrix is ​​learnable; For vector concatenation operations, the same Symbol meaning; This is the core operation function of the Spatiotemporal-Aware Multi-head Attention module. It is used to perform attention calculation and multi-head aggregation under spatiotemporal constraints on the input query, key and value matrix, and output a feature representation that integrates spatiotemporal semantics. For the 1st to the 1st The spatiotemporal attention output matrix of each attention head is the basic input for multi-head aggregation.

[0055] S23, Feature Space Alignment Module The feature space alignment module is used to achieve feature space alignment based on optimal transport theory. Addressing the issue of heterogeneous feature distributions across different modalities, this invention employs a feature space alignment strategy based on optimal transport theory to minimize the Wasserstein distance between different modal feature distributions. Simultaneously, a domain-adaptive regularization term is introduced to ensure that the fused features retain task information relevant to emergency decision-making. Specifically: (1) Wasserstein distance calculation: Calculate the Wasserstein distance between the text feature distribution and the visual feature distribution. The formula is: ; In the formula, The Wasserstein distance; , The distributions of textual and visual features are respectively; For all , The joint distribution set of marginal distributions; For joint distribution The expected value of the following; Text feature samples Visual feature samples The square of the Euclidean distance between them is used to calculate the similarity between feature samples; The joint probability distribution of text features and visual features is... An element in a set; The infimum (maximum lower bound) operator is used for all possible joint distributions. In the process, find the Wasserstein distance value that minimizes the mathematical expectation.

[0056] (2) Construction of the alignment loss function: Introducing a domain adaptive regularization term, the feature space alignment loss function is constructed, and the formula is: ; In the formula, For feature space alignment loss; This is the weighting parameter (default value 0.2); The loss function for the downstream entity-relationship extraction task.

[0057] S24. Quantitative Calculation and Robust Design of Fusion Feature Confidence Measure S241, Fusion Feature Confidence Measurement Model This invention addresses the uncertainty of fused features by constructing a multi-dimensional confidence quantification model to comprehensively calculate the overall confidence score of the fused features. The formula is as follows: ; In the formula, The range of values ​​for the overall confidence score of the fused features is [not specified]. ; , , The weighting coefficients were determined based on grid search optimization using 1200+ historical cases. The semantic matching degree between modalities is calculated based on the mean cosine similarity of text-visual features, with a value range of [range missing]. ); Feature completeness (calculated based on the ratio of the number of effective modes to the total number of input modes, with a value range of...) ; The feature alignment loss for the current sample; This represents the historical maximum value of the alignment loss.

[0058] S242, Robust Design for Single-Modal Missing This invention designs a single-modal missing degradation operation mechanism to ensure system availability under extreme scenarios: Visual data missing: The system achieves fusion based on text + spatiotemporal dual modality, maintaining an extraction accuracy of over 85%; Text data only: The system completes feature extraction and fusion based on a single text modality, maintaining an extraction accuracy of over 79%; the fused features and the corresponding comprehensive confidence scores are simultaneously transmitted to the downstream entity-relationship extraction module.

[0059] S3: Construct a multimodal joint entity relationship Transformer model, take multimodal fusion features as input, jointly extract emergency entities and relationships between entities end-to-end, and output standardized entity-relationship triples and their extraction confidence; Step S3 is used to implement multi-scenario joint entity-relation extraction based on the Multimodal Joint Entity Relationship Transformer (MJERT). This invention constructs an end-to-end MJERT model, which, based on the multimodal fusion features generated in step S2, simultaneously completes emergency entity identification and semantic relation extraction between entities, solving the error propagation problem of traditional pipeline methods; it also outputs the confidence score of the extraction results, and is equipped with a semi-supervised few-shot learning framework to solve the problems of high annotation costs and difficulty in cross-scenario adaptation. Specifically, it includes the following steps: S31, MJERT model core architecture The MJERT model is built on a Transformer encoder and includes three joint prediction heads for entity boundary detection, entity type classification, and relation classification, achieving end-to-end joint extraction. The core hyperparameters of the model are: 6 Transformer encoder layers, 8 attention heads, 768 hidden layer dimensions, and a dropout rate of 0.2.

[0060] The multimodal joint entity relationship Transformer model specifically includes: a Transformer encoder submodule, used to perform deep context encoding on multimodal fusion features, capture long-distance dependencies between features, and generate context-aware feature representations; an entity boundary detection submodule, which uses a sequence labeling method based on span annotation to predict the start and end positions of entities from the context-aware feature representations; an entity type classification submodule, used to classify the detected entities by type; and a relationship classification submodule, which uses a table filling method to model the semantic relationships between all entity pairs, achieving joint classification of relationships between entities. The entity boundary detection, entity type classification, and relationship classification submodules serve as a joint prediction head, performing end-to-end joint training and inference based on the shared context-aware feature representations, synchronously outputting standardized entity-relationship triples and their corresponding extraction confidence scores. The specific architecture design is as follows: S311, Transformer encoder context coding The multimodal fusion feature matrix output in step S2 Inputting a Transformer encoder generates a deep context-aware feature representation, as shown in the formula: ; In the formula, The Transformer Encoder outputs a context-aware feature representation matrix that carries deep contextual semantic information of the multimodal fusion features. It is a Transformer encoder used to perform deep context encoding on the input multimodal fusion feature matrix, capture long-distance dependencies between features, and generate feature representations containing contextual semantics; This is a multimodal fusion feature matrix. For sequence length, For feature dimensions; Multimodal fusion feature matrix The individual feature vectors in the sequence correspond to the first to the last feature vector in the sequence. Fusion features at each location.

[0061] S312. Entity boundary detection based on span annotation A sequence labeling method is used to identify the start and end positions of entities through a span labeling scheme, solving the problem of identifying overlapping entities in emergency scenarios. The formula is as follows: ; In the formula, For the first Entity boundary prediction probability for each token; For the first The hidden layer representation of each token; , The parameter matrix and bias vector are learnable. This is a normalized exponential function that outputs the boundary location, type, and corresponding probability value of an entity.

[0062] S313. Classification of Union Relationships Based on Table Filling Using a table-filling method, the semantic relationships between all entity pairs are modeled to achieve joint classification of entity relationships. The formula is as follows: ; In the formula, For the first The token and the first The probability of predicting the relationship between entity pairs corresponding to each token; , The parameter matrix and bias vector are learnable. This is a vector concatenation operation; This is an element-wise multiplication operation that outputs the relationship type between entity pairs and their corresponding probability values.

[0063] S32. Calculation of confidence level of extraction results This invention constructs a two-level confidence quantification model for entity-relation extraction results, calculating entity confidence and relation confidence separately to quantify the uncertainty of the extraction results. In step S32, entity confidence and relation confidence are calculated separately; entity confidence is determined based on entity boundary prediction probability, entity type classification probability, and compliance score of whether the entity conforms to the semantic constraints of the top-level ontology; relation confidence is determined based on relation classification probability, confidence of subject and object entities, and compliance score of whether the relation conforms to the semantic constraints of the top-level ontology; the specific model is as follows: S321, Entity Confidence Calculation ; In the formula, For entity confidence, the range of values ​​is... The softmax maximum probability predicted for the entity boundary span; The softmax maximum probability for classifying entity types; Score semantic compliance (1 for conforming to top-level ontology constraints, 0 otherwise).

[0064] S322, Calculation of Relation Confidence ; In the formula, The confidence level of the relationship, with a range of values. The softmax maximum probability for classifying relationships; , These represent the confidence levels of the subject and object entities, respectively. Score the compliance of relational semantic constraints (1 for compliance with top-level ontology constraints, 0 otherwise).

[0065] S33. Confidence threshold setting and dynamic adjustment mechanism In step S33, a confidence threshold associated with the risk level of the emergency scenario is preset, and the confidence threshold is automatically adjusted according to the risk level of the current emergency event or the data quality. Extraction results below the confidence threshold trigger manual verification or downgrade processing. Specifically: S331, Threshold Setting Basis The confidence thresholds set in this invention are all based on a test set of 1200+ historical emergency cases, and were determined by optimizing the cutoff points of the ROC curve and PR curve. The core thresholds are: (1) Entity-relationship extraction confidence threshold: 0.80 (at this threshold, the model precision reaches 91.2% and the recall reaches 86.7%, balancing precision and coverage). (2) Confidence threshold for key decision reasoning: 0.85; (3) Decision confidence threshold for high-level emergency events: 0.90 (the model accuracy reaches 98.5% under this threshold, ensuring the reliability of major decisions); at the same time, 0.75 is set as the minimum confidence red line, and the sampling results below this value will directly trigger manual verification.

[0066] S332, Threshold Sensitivity Analysis The threshold sensitivity test was completed using the controlled variable method. The results showed that when the threshold fluctuated within the range of ±0.05, the F1 score of the model fluctuated by ≤1.2%, with no significant performance change, proving that the confidence threshold set in this invention is robust. When the threshold was lower than 0.75, the model's error sampling rate increased by more than 12%.

[0067] S333, Threshold Dynamic Adjustment Mechanism This invention supports scenario-based dynamic adaptation of thresholds, automatically adjusting thresholds according to the risk level of emergency scenarios: (1) High-risk scenarios (personnel casualties, damage to major facilities): The threshold is automatically increased by 0.05 to improve the reliability of decision-making; (2) Low-risk general events: The threshold can be lowered by 0.03 to improve handling efficiency; (3) Scenarios with poor data quality: The threshold is automatically increased by 0.05 to filter out low-quality extraction results.

[0068] S34 and MJERT model training optimization strategies In step S34, a multi-task joint loss function is used to train the multimodal joint entity-relation Transformer model. This function integrates entity boundary detection loss, entity type classification loss, and relationship classification loss, and introduces a focus loss to address the imbalance in entity and relationship category distributions. Simultaneously, a semi-supervised few-shot cross-scene transfer learning strategy is employed, using a three-stage training framework of pre-training, domain adaptation, and fine-tuning to reduce the annotation cost of cross-scene adaptation. Specifically, the steps include: S341, Multi-task learning loss function Design a multi-task joint loss function that integrates entity boundary detection loss, entity type classification loss, relation classification loss, and regularization term. The contribution of each sub-task is balanced by task weights. The formula is as follows: ; In the formula, This is a combined loss due to multiple tasks; , , , Weighting coefficients for tasks; Loss for detecting entity boundaries; Loss for entity type classification; Loss for classifying relationships; This is a regularization term (including cross-modal feature consistency constraints).

[0069] S342, Focus Loss Resolves Category Imbalance To address the class imbalance issue in emergency scenarios where key entities / relationships have a low proportion of samples, a variant of focus loss is adopted to assign higher weights to rare but critical classes. The formula is as follows: ; In the formula, Loss of focus; The number of samples; For category Weighting factors; For category The predicted probability; For focusing parameters; It is the natural logarithm function.

[0070] S343, Semi-supervised few-shot cross-scenario transfer learning (1) Construct a three-stage training framework of "pre-training-domain adaptation-fine-tuning" to solve the problems of difficult cross-scene adaptation and high annotation cost, specifically: (2) Pre-training stage: Based on massive amounts of publicly available emergency text and image data, comparative learning is used to complete the general semantic pre-training of the model; (3) Domain Adaptation Stage: Cross-scenario domain adaptation is achieved through unlabeled scenario-specific emergency data, reducing performance degradation caused by scenario differences; Fine-tuning phase: Only 500+ labeled data points per scene are needed to complete the scene-specific fine-tuning of the model.

[0071] S344, Details of Model Training and Inference (1) Training parameters: The optimizer used is AdamW, the base learning rate is 2e-5, the batch size is 32, the number of training rounds is 30, the learning rate warm-up ratio is 10%, the weight decay is 1e-4, and gradient clipping (maximum gradient norm is 1.0) is used. (2) Reasoning strategy: The constraint decoding algorithm is adopted to enforce the entity-relation semantic constraints defined in the top-level ontology, filter out semantically invalid extraction results, and output the confidence score of each entity-relation triple. Results below the threshold are triggered for manual verification.

[0072] S4: Based on the entity-relationship triples, construct a three-level hierarchical knowledge graph that adapts to the multi-level command architecture using a five-tuple spatio-temporal knowledge representation method, and assign a comprehensive confidence level to each piece of knowledge in the three-level hierarchical knowledge graph; Step S4 is used to implement the construction and dynamic update of a spatio-temporal decision-making knowledge graph for multi-level emergency command. Based on the entity-relationship triples extracted in step S3, this invention constructs a spatio-temporal decision-making knowledge graph that adapts to multi-level emergency command. The core proposes a five-tuple spatio-temporal knowledge representation method and a three-level hierarchical graph architecture, and supports a hybrid storage architecture of "hot data in-memory graph + cold data disk graph" to solve the problems of weak static modeling ability, large update overhead, and difficult multi-level collaboration of traditional knowledge graphs. The specific steps are as follows: S41. Five-tuple spatio-temporal knowledge representation method This invention extends the triple structure of traditional knowledge graphs and proposes a five-tuple spatio-temporal knowledge representation form of <subject, predicate, object, spatial reference, time reference>, which explicitly models the spatial location and time validity of emergency entities and relationships. At the same time, a confidence score and a source reliability label are attached to each piece of knowledge. Examples are: (1) <Intersection of XX Road and XX Avenue, hasStatus, traffic accident congestion, {lat: 39.12, lng: 117.23}, [2026-03-04T14:30:00, 2026-03-04T15:10:00]>; (2) <Building No. 3 in XX Community, hasStatus, in the process of fire handling, {lat: 39.15, lng: 117.25}, [2026-03-04T16:20:00, now]>.

[0073] This representation form can completely record the state changes of emergency entities, the temporal evolution of events, and the association of spatial locations, adapting to the dynamic characteristics of emergency scenarios. As shown in the left figure of Figure 4 Figure 5 is an example diagram of the five-tuple spatio-temporal knowledge representation method of an embodiment of this invention.

[0074] S42. Three-level hierarchical graph architecture In response to the hierarchical requirements of multi-level emergency command, this invention constructs a three-level hierarchical graph architecture of "total graph - sub-graph - scenario sub-graph" to achieve cross-level knowledge synchronization and permission control, and at the same time adapt to cloud-edge collaborative deployment. Further, the three-level hierarchical knowledge graph includes: a total graph deployed at the cloud center node, which is used to integrate global core knowledge and support global situation awareness and cross-regional collaborative decision-making; a sub-graph deployed at the regional sub-node, which is used to focus on the emergency knowledge of the corresponding administrative region or industry field and support the command and judgment at this level; a scenario sub-graph deployed at the edge-side on-site disposal node, which is used to face specific emergency scenarios and support on-site rapid response and precise disposal. As Figure 4 The right figure in the diagram is a deployment schematic of the three-level hierarchical knowledge graph architecture according to an embodiment of the present invention. The specific architecture is as follows: (1) City / Provincial Overall Map: Integrates core emergency knowledge across the entire region and all scenarios to achieve global situational awareness and cross-regional and cross-departmental collaborative decision-making, and is deployed at the central node in the cloud; (2) District / Industry Sub-map: Focusing on emergency knowledge in the corresponding administrative region / industry field, responsible for the analysis and handling of emergency command at this level, realizing incremental data synchronization and hierarchical permission control with the overall map, and deployed in regional sub-nodes; (3) Scene Sub-graph: Dedicated knowledge graphs for specific scenarios such as smart transportation, smart community, and smart campus are deployed to edge-end on-site processing nodes to achieve rapid response and precise handling within the scenario.

[0075] S43. Knowledge Confidence Measurement and Multi-Source Heterogeneous Knowledge Fusion Framework (MHKFF) S431, Knowledge Confidence Measurement Model For each quintuple of knowledge, this invention constructs a comprehensive confidence calculation model that integrates extraction confidence, source reliability, and time freshness. The formula is as follows: ; In the formula, The range of values ​​for the overall confidence level of knowledge is [not specified]. ; The confidence level of the relationship output in step S3; The reliability weights of information sources are assigned (official authoritative sources 1.0, on-site law enforcement personnel 0.9, social media 0.4). The time decay factor (default for sudden emergency events) Hour); The time elapsed since the information was generated. It is an exponentially decaying function.

[0076] S432, Four-stage process for multi-source heterogeneous knowledge fusion This invention constructs a four-stage MHKFF to address the issues of inconsistency, incompleteness, and redundancy in knowledge from multiple scenarios and sources. It achieves weighted fusion based on knowledge integration confidence. The four-stage process is as follows: (1) Entity normalization: Through a hybrid algorithm of “surface form matching + context embedding similarity + structural similarity”, entity alignment and disambiguation across sources and scenarios are completed, and different representations of the same entity are identified; (2) Relationship Coordination: Based on the semantic constraints of the top-level ontology architecture, unify the relation representations from different sources and solve the problems of relation semantic conflicts and hierarchical inconsistencies; (3) Attribute merging: Based on knowledge confidence, strategies such as weighted average, source authority priority, and latest value priority are adopted to merge the multi-source attributes of entities; (4) Spatiotemporal alignment: Based on the accuracy of spatial coordinates and the freshness of timestamps, the spatial location of entities is aligned with the time interval, and high-precision spatial information and the latest time status information are retained first.

[0077] S433, Rules for Resolving Knowledge Fusion Conflicts To address the conflict problem of multi-source knowledge, this invention formulates clear conflict resolution rules and combines them with knowledge confidence to achieve fusion. The core rules are shown in Table 2 below: Table 2

[0078] S434. Knowledge Completion and Error Correction Mechanism (1) Knowledge completion: Based on domain rules and relational graph convolutional network (RGCN), knowledge completion is completed to predict missing entity relationships and solve the problem of incomplete emergency information; (2) Error correction: Error detection is completed through consistency verification and outlier detection, and conflict correction is completed based on source reliability, time freshness, and neighborhood knowledge consistency; (3) Low confidence knowledge processing: Knowledge with a confidence level below 0.6 is only used as a reference and is not included in the core decision-making process.

[0079] S44, Hybrid Storage Architecture and Dynamic Update Mechanism S441, Hot-Cold Hybrid Data Storage Architecture To address the issues of high overhead in multidimensional index updates and the difficulty in balancing real-time performance and query efficiency, a hybrid storage architecture of "hot data in-memory graph + cold data disk graph" is designed, specifically as follows: (1) Hot data memory map: Stores currently active emergency events, real-time updated entity status, and frequently accessed core knowledge. It adopts a hybrid multidimensional index of R tree spatial index + time interval tree index + semantic predicate index to achieve millisecond-level query and second-level update; (2) Cold data disk map: Stores historical emergency events, inactive entities, and full historical version data. It adopts graph partition storage and batch index update to ensure storage efficiency and historical backtracking capability.

[0080] S442, Multi-strategy knowledge dynamic update mechanism A multi-strategy dynamic knowledge update mechanism is constructed to fully preserve the complete historical record of knowledge evolution, support event tracing and version rollback, and ensure the real-time performance and consistency of the knowledge graph. The specific strategies are as follows: (1) Append-oriented log update: All knowledge update operations are logged to preserve the full history of knowledge evolution and support version rollback and event tracing; (2) Status-based incremental update: Based on real-time access to emergency data, the current state and relationships of entities are updated incrementally, only updating the knowledge that has changed, thus reducing update overhead; (3) Conflict-based belief correction: For conflicting information from multiple sources, knowledge correction is carried out based on knowledge confidence, source reliability, and time freshness to ensure the consistency of the graph.

[0081] S5: Based on the three-level hierarchical knowledge graph, an emergency decision-making scheme is generated using a three-layer reasoning framework that integrates rule deduction, analogical reasoning, and graph structure analysis, and the comprehensive confidence level of the emergency decision-making scheme is output. Step S5 is used to realize knowledge reasoning and intelligent decision generation for multi-level emergency command. This invention, based on a constructed spatiotemporal decision knowledge graph, builds a three-layer integrated reasoning framework of "rule deduction - analogical reasoning - graph structure analysis," and is equipped with an improved AMIE+ automatic rule mining and causal verification mechanism to solve the problems of high maintenance costs and poor generalization ability of traditional rule bases. Simultaneously, based on knowledge confidence, it achieves hierarchical control of decision results, generating intelligent decision-making schemes adapted to multi-level command. Specifically, it includes the following steps: S51, Three-layer integrated emergency knowledge reasoning mechanism S511, Regularized Deductive Reasoning Based on expert knowledge in the emergency response field and national emergency response standards, standardized reasoning rule bases are constructed for five major scenarios. The rule format is "preconditions → reasoning result → application scenario → confidence weight". The core rules cover core scenarios such as resource allocation, risk assessment, path planning, time series prediction, and team assignment, and are fully aligned with the special emergency plans for each scenario.

[0082] S512, Embedded Analogical Reasoning Based on entity and relation embedding in knowledge graphs, a historical case matching model is constructed. For current emergency events, the model retrieves the most similar historical emergency response cases, reuses mature response plans and decision-making experience, and solves the decision support problem in unknown emergency scenarios.

[0083] S513, Graph Structure Analysis and Reasoning By using graph neural networks and graph algorithms, the topological structure of knowledge graphs can be analyzed to identify key nodes, critical paths, bottlenecks, and risk propagation links in emergency response. For example, the road network propagation path of accident impacts in smart transportation scenarios and the risk spread range of fires in smart community scenarios can be identified to provide data support for decision-making.

[0084] S52. Automatic rule discovery and conflict checking mechanism To address the high cost of manual maintenance of rule bases, step S52 of this invention specifically employs an improved AMIE+ frequent pattern mining algorithm + DoWhy causal inference framework to construct a complete process of automatic rule mining and verification, thereby achieving semi-automatic construction and low-cost maintenance of the inference rule base. The specific process includes: S521, Candidate Rule Mining Based on the improved AMIE+ algorithm, this invention mines frequently occurring entity-relationship patterns from historical emergency case knowledge graphs to generate candidate reasoning rules. The improvements of this invention to the AMIE+ algorithm are: the addition of spatiotemporal constraints and ontology semantic constraints for filtering, generating only candidate rules that conform to the ontology specifications of the emergency domain, eliminating invalid patterns, and reducing invalid candidate rules by more than 80%.

[0085] S522, Causality Verification and Confidence Scoring (1) Causal verification: Based on the DoWhy causal inference framework, the preconditions and results of candidate rules are tested for causal effect. Only causal rules that pass the 95% significance test are retained, and spurious correlation rules are eliminated. (2) Confidence score: Calculate the support, confidence and lift for each candidate rule, and only retain the rules with support ≥10 and confidence ≥0.85 to enter the expert review stage.

[0086] S523, Expert Review and Rule Entry Candidate rules are sorted from high to low confidence. Experts only need to review high-confidence rules. At the same time, the system provides historical case matching results for the rules to help experts make quick judgments. Approved rules are added to the reasoning rule base to achieve automatic expansion of the rule base.

[0087] S524, Quantitative Assessment of Maintenance Costs In traditional solutions, it takes an average of 10 person-days for emergency response experts to manually write 100 compliance rules. This solution automatically discovers 100 high-confidence candidate rules, and experts only need 2 person-days to complete the review and fine-tuning, reducing the workload by 80%, which means the rule base maintenance cost is reduced by 80%.

[0088] S53, Calculation of Decision Confidence Measurement For the final emergency decision-making plan, this invention constructs a comprehensive confidence calculation model that integrates rule matching degree, case similarity, and knowledge confidence. The formula is as follows: ; In the formula, The overall confidence level for decision-making, with a range of values. The weighting coefficients are determined based on expert review of emergency management and optimization of historical cases. Rule matching score (calculated based on the average confidence score of the triggered rule, with a value range of...) ); Historical case similarity (calculated based on the matching degree of the most similar case, with a value range of...) ); The average confidence level (range of values) of the core knowledge upon which the decision depends ).

[0089] The overall decision confidence score is directly linked to the emergency response level and the human-machine collaborative verification mechanism, enabling closed-loop control of confidence across the entire process. For example, when the overall decision confidence score is below 0.85, a manual review process for Level III or higher emergency response is automatically triggered.

[0090] S54, Multi-level Collaborative Intelligent Decision-making Core Algorithm This invention is based on a three-layer fusion reasoning mechanism and designs four core intelligent decision-making algorithms to adapt to the needs of multi-level emergency command, specifically: S541, Multi-dimensional Intelligent Assessment Algorithm for Emergency Situation Based on real-time data from knowledge graphs, an emergency situation assessment model is constructed from six dimensions: event level, scope of impact, casualties, facility damage, resource gaps, and risk evolution trends. This model automatically determines the emergency response level, generates situation assessment reports, provides global situation awareness capabilities for multi-level command, and enables real-time updates and visualization of the situation.

[0091] S542, Two-Stage Emergency Resource Global Optimization Allocation Algorithm A two-stage resource allocation algorithm consisting of "global strategic optimization + local tactical adjustment" is constructed to achieve cross-level and cross-scenario emergency resource optimization and scheduling. (1) Global optimization stage: Based on the Constraint Satisfaction Problem (CSP) model, with emergency priority, urgency of handling and resource availability as the core constraints, and with the optimization goal of maximizing rescue efficiency and minimizing casualties, the global allocation of emergency resources across regions and departments is completed, which is adapted to the macro-dispatch at the city-province-national level. (2) Local adjustment stage: Based on deep reinforcement learning, the resource allocation plan is dynamically adjusted according to the real-time feedback of the situation on site, and the micro-level handling at the district-street-site level is adapted.

[0092] S543, Multi-Constraint Emergency Rescue Path Dynamic Programming Algorithm Based on knowledge graph-based road network spatial data, real-time traffic conditions, and event impact range, a multi-objective optimization path planning algorithm is constructed, balancing multiple constraints such as travel time, traffic safety, vehicle carrying capacity, and task priority. Combined with real-time updates of the knowledge graph, rescue routes are dynamically adjusted to avoid newly added road congestion and risk areas, providing optimal route planning for rescue vehicles, material transport vehicles, and personnel evacuation.

[0093] S544, Multi-level Disposal Instruction Generation and Closed-Loop Control Algorithm Based on the emergency decision-making plan and in accordance with the multi-level command organizational structure, the system automatically generates tiered response instructions, clarifying the responsibilities and tasks of command departments at all levels, response units, and on-site teams. At the same time, it uses a knowledge graph to track the execution progress and response effects of instructions in real time, and constructs a closed-loop control mechanism of "instruction issuance - execution feedback - situation update - decision adjustment" to solve the problems of poor instruction transmission and uncontrollable execution progress in multi-level command.

[0094] S6: Based on the preset emergency response level and the comprehensive confidence level of the decision, execute the corresponding human-machine collaboration rules to conduct hierarchical review, instruction issuance and execution feedback of the emergency decision-making plan, forming a closed-loop control of the entire disposal process.

[0095] Step S6 is used to achieve multi-scenario application adaptation and closed-loop management of human-machine collaboration. This step constructs scenario-based adaptation interfaces based on the business characteristics of different emergency scenarios. Simultaneously, it strictly aligns with national emergency management standards, refines the full-process management mechanism for human-machine collaboration, and clarifies emergency response levels, permission matrices, Service Level Agreement (SLA) requirements, and timeout circuit breaker mechanisms. This fully conforms to actual emergency command processes, improving the system's implementability. Specifically, it includes the following steps: S61, Multi-scenario Application Adaptation This invention constructs a scenario-based adaptation module for five major scenarios: smart transportation, smart communities, smart campuses, smart government affairs, and smart tour guides. This module enables seamless integration with existing information systems in each scenario. Specifically: (1) Scenario-based data interface: Adapt to existing information systems in various scenarios (traffic control platform, community smart platform, campus security platform, government emergency platform, scenic area smart tourism platform) to achieve two-way data communication and synchronous issuance of instructions; (2) Scenario-based decision-making interface: Customize the display of situation information, decision-making suggestions and handling instructions for users in different scenarios (traffic police, property management, campus security, government emergency personnel, scenic area staff) to lower the threshold for use; (3) Scenario-based emergency plan adaptation: Digitize and structure the special emergency plans for each scenario and embed them into the reasoning rule base to ensure the consistency between the decision-making plan and the existing plans.

[0096] S62, Refined Human-Machine Collaborative Closed-Loop Management Mechanism S621. Emergency Response Level Classification and Judgment Standards Strictly adhering to the "National General Emergency Response Plan for Public Emergencies," the "Regulations on Reporting and Investigation of Production Safety Accidents," and the "Emergency Management Measures for Emergencies," the plan classifies emergency response levels into four levels, clearly defining judgment criteria, human-machine collaboration rules, confidence thresholds, and SLA requirements. It also includes a parallel consultation mechanism to adapt to the actual procedures of multi-level emergency command, as detailed in Table 3 below. Table 3

[0097] S622, Multi-level Command and Decision-Making Authority Matrix Strictly aligning with the "Emergency Response Law of the People's Republic of China," the decision-making authority of the six levels of command—national, provincial, municipal, district, street, and on-site—is clearly defined. This includes the authority to issue instructions, allocate resources, activate contingency plans, and adjust response levels. This approach aims to prevent unauthorized decision-making and shirking of responsibility, ensuring that the authority and responsibilities of each level of command are compliant and clear.

[0098] S623, SLA timeout circuit breaker and fallback mechanism To address the manual review SLA requirements for each level of emergency incident, a tiered timeout circuit breaker mechanism is designed to prevent decision-making delays. Simultaneously, a backup contingency plan is implemented to ensure the effectiveness of emergency response in extreme situations. Specifically: (1) Level IV event: If the review timeout exceeds 30 minutes, the system will automatically execute the preset backup handling plan and simultaneously push alarm information to the person in charge at this level and the emergency management department at the next higher level; (2) Level III incident: If the review time exceeds 15 minutes, it will be automatically upgraded to the municipal emergency management department and the local emergency response plan will be triggered simultaneously; (3) Level II event: If the review timeout exceeds 5 minutes, it will be automatically synchronized to the provincial emergency command center, and a parallel consultation mechanism will be initiated, with the provincial expert group conducting the review first; (4) Level I events: The system generates alternative solutions within 1 minute and pushes them to the national, provincial and municipal command organizations simultaneously. The entire process is under real-time control with no risk of timeout.

[0099] S624, Closed-loop management of the entire decision-making process A closed-loop control mechanism is constructed, encompassing the entire process of "instruction issuance - execution feedback - situation update - decision adjustment." Relying on a spatiotemporal decision-making knowledge graph and real-time data acquisition channels, this mechanism enables traceability, controllability, and optimization across the entire emergency response chain. The specific process is as follows: (1) Issuance of instructions: Based on hierarchical decision-making authority, the disposal instructions are synchronized to the terminal devices (command screen, mobile APP, law enforcement terminal) of the corresponding execution unit / personnel through the system interface, and a unique instruction ID is generated for tracking; (2) Execution feedback: The executing unit / personnel uploads the execution progress of the instruction (received, in execution, completed, abnormal), on-site handling data (pictures, videos, text records), and resource consumption status in real time through the terminal. The feedback data is automatically associated with the corresponding instruction ID; (3) Situation update: The system will integrate the execution feedback data with the knowledge graph in real time, update the entity status (such as the remaining amount of resources, road traffic status, changes in personnel casualties), event evolution trends, and dynamically adjust the situation assessment results; (4) Decision adjustment: If the situation is updated and the preset conditions are triggered (such as the risk spread speed exceeds expectations, the resource gap expands, or a new emergency event occurs), the system will automatically start decision recalculation, generate an adjustment plan based on the latest knowledge graph and reasoning rules, and issue it for execution after human-machine collaborative verification.

[0100] Preferably, in some embodiments of the present invention, the method adopts a cloud-edge dual-track operating architecture, including: a real-time inference link, in which lightweight feature extraction and scene sub-graph inference modules are deployed at the edge, while complex cross-modal fusion and overall graph inference modules are deployed in the cloud; when an emergency occurs, the edge and cloud collaborate to complete real-time inference, meeting the requirements of second-level response; an offline iteration link, deployed on a dedicated training node in the cloud, is started during non-emergency periods, incrementally fine-tuning the model and updating and optimizing the rule base and knowledge graph based on newly accumulated disposal data; and through a horizontal federated learning framework, cross-departmental and cross-scenario model collaborative optimization is achieved, with the original training data not leaving the local machine, ensuring data privacy and compliance. Specific steps are as follows: S7, Cloud-Edge Dual-Track System Closed-Loop Optimization This invention constructs a dual-track operation architecture that coordinates cloud and edge computing, completely resolving the core contradiction between the "real-time response requirements" and "continuous iterative optimization" of emergency systems. At the same time, it enables cross-scenario and cross-departmental model collaborative optimization through horizontal FL (Flexible Interaction), without leaking original data, ensuring the continuous evolution of system capabilities.

[0101] S71, Cloud-Edge Dual-Track Operation Architecture Design This invention completely decouples real-time inference from offline iteration, constructing a dual-track operating architecture to adapt to the real-time and long-term optimization needs of emergency scenarios. The specific architecture is as follows: S711, Real-time Inference Link (Edge Terminal + Cloud Core Node) (1) Deployment location: Lightweight modules (modal-specific feature extraction, scene sub-graph inference) are deployed to edge-end on-site processing nodes, while complex modules (CAFN cross-modal fusion, MJERT joint extraction, and total graph inference) are deployed on cloud core nodes; (2) Operational logic: When an emergency occurs, the edge device completes real-time data collection, preprocessing, lightweight feature extraction and local sub-graph reasoning to generate preliminary decision suggestions; for complex decisions that require cross-regional / cross-departmental collaboration, key data is uploaded to the core node in the cloud and returned as a decision solution after a complete process reasoning. (3) Performance indicators: No cross-node federated communication, no dynamic model updates, and a single round of full-process inference takes ≤2 seconds, meeting the second-level response requirements for emergency response; (4) Fault-tolerant design: The link has built-in confidence control throughout the entire process. Low confidence results will automatically trigger manual intervention. When the edge end loses connection with the cloud, the edge end can independently complete basic handling decisions based on the local sub-map and synchronize data completion after the connection is restored.

[0102] S712, Offline Iterative Link (Cloud Training Node) (1) Deployment location: Deployed on a dedicated training node in the cloud, physically isolated from the real-time inference link, and does not occupy real-time computing resources; (2) Operation logic: It only starts during non-emergency periods (such as nighttime), and completes incremental fine-tuning of the model, rule base update and knowledge graph improvement based on the labeled data, execution feedback data and historical case data accumulated after the emergency event is handled. (3) Collaborative optimization: Through the horizontal FL framework, local data from various departments and scenarios are combined (the original data does not leave the domain) to complete cross-scenario model collaborative optimization and improve the system's cross-domain generalization ability; (4) Update mechanism: New model versions, rule sets, and knowledge graph increments generated by offline iterations are synchronized to the core nodes and edge terminals in the cloud through incremental updates after passing the test and verification, to ensure that the capabilities of the entire system are consistent.

[0103] S72, System-wide Closed-Loop Optimization Mechanism Based on a dual-track operating architecture, a closed-loop optimization mechanism is constructed for the entire system, encompassing "data accumulation - model iteration - rule optimization - graph improvement - performance enhancement," specifically: S721, Model Iterative Optimization (1) Data accumulation: The manually verified extraction results, decision feedback data, and newly added cases during the emergency response process are labeled as high-quality training data to construct an incremental training dataset; (2) Incremental fine-tuning: The course learning strategy is adopted to fine-tune models such as CAFN, MJERT, and RGCN based on the incremental training dataset, freeze the pre-trained layer parameters, and only update the top task layer to shorten the training time (the time for a single round of incremental training is ≤2 hours). (3) Performance verification: Verify the performance of the fine-tuned model (entity-relation extraction F1 score, inference accuracy) on the test set. The new model will be included in the official version only if the performance improvement is ≥1%.

[0104] S722, Rule Base Optimization (1) Rule evaluation: Based on the historical decision execution effect, calculate the accuracy (number of correct decisions / number of triggers) and coverage (number of emergency event types covered / total number of event types) of each rule, and filter out rules with low accuracy (<80%) and low coverage (<5%). (2) Rule update: For rules with low accuracy, optimize the preconditions / inference results based on actual handling feedback; for rules with low coverage, supplement new scenario rules through the rule automatic mining module; (3) Conflict verification: After the rules are updated, the conflict verification mechanism is used to detect the contradiction between the new rules and the existing rules to ensure the consistency of the rule base.

[0105] S723, Improved Knowledge Graph (1) Entity / relationship supplementation: Supplement the knowledge graph with newly added emergency entities (such as new rescue equipment, new evacuation areas) and relationships (such as cross-departmental cooperation relationships, new risk transmission paths) during the disposal process; (2) Historical case database: The emergency events that have been handled are organized into historical case knowledge in the format of five tuples, and associated with the corresponding entities and relationships to enrich the case database of analogical reasoning; (3) Knowledge quality optimization: Regularly clean up redundancy (delete duplicate knowledge) and correct errors (correct outdated knowledge based on new disposal data) of the knowledge graph to ensure the accuracy of the knowledge base.

[0106] S724, System Performance Optimization (1) Response speed optimization: Based on actual operating data, analyze the bottleneck links of the system (such as spectrum query and feature fusion), optimize the algorithm complexity (such as adjusting the index structure and simplifying redundant calculations), and continuously reduce the inference time; (2) Resource consumption optimization: Quantization compression (INT8 quantization) and pruning are performed on the edge model to further reduce memory consumption and computing power requirements, and adapt to edge devices with lower configurations; (3) Concurrency optimization: Adopting a distributed deployment architecture, expanding the computing resources of the core nodes in the cloud, supporting the simultaneous processing of multiple emergency events (concurrency ≥ 100 events / second), adapting to large-scale emergency scenarios.

[0107] The above-described embodiment of the present invention provides a multi-level emergency command intelligent decision-making method based on knowledge graphs. It employs core inventions such as SAMA mechanism, five-tuple spatiotemporal knowledge representation, full-link confidence control, three-level hierarchical graph architecture, and cloud-edge dual-track architecture. These are all substantial improvements addressing pain points in the field of emergency command. The synergistic effect of these innovations achieves a comprehensive improvement in system accuracy, efficiency, robustness, and real-time performance.

[0108] Example 2 Figure 5 This is a schematic diagram of a multi-level emergency command intelligent decision-making system based on knowledge graphs, provided in Embodiment 2 of the present invention. This system is used to implement the aforementioned multi-level emergency command intelligent decision-making method based on knowledge graphs. It adopts a microservice architecture and a cloud-edge collaborative deployment mode, and is divided into eight core functional modules. Each module has clear functional boundaries and standardized interfaces, supporting independent development, testing, and deployment, thus reducing overall implementation risks. The specific modules are as follows: The multi-source data access and preprocessing module 210 is used to collect and standardize multi-source heterogeneous data in the multi-level emergency command scenario, and extract exclusive features of text modality, visual modality and spatiotemporal modality respectively. Furthermore, the core functions of the multi-source data access and preprocessing module 210 are: to provide standardized access interfaces for multi-scenario, multi-source emergency data (supporting access methods such as databases, files, APIs, and IoT devices); to perform data cleaning, format standardization, scale normalization, and missing value imputation preprocessing; and to build a horizontal FL data processing submodule and a data quality scoring submodule to output data quality confidence labels.

[0109] The cross-modal attention fusion module 220 is used to construct a cross-modal attention fusion network that includes a spatiotemporally aware multi-head attention mechanism. It performs deep semantic alignment and fusion on the extracted text modality, visual modality and spatiotemporal modality-specific features, and outputs multimodal fusion features and their fusion confidence. Furthermore, the core function of the cross-modal attention fusion module 220 is to achieve bidirectional cross-modal attention fusion of text, visual, and spatiotemporal features; it includes the TIIA text-image interaction attention submodule, the SAMA spatiotemporal awareness multi-head attention submodule, and the feature space alignment submodule; and it outputs unified multimodal fusion features and fusion confidence scores.

[0110] The multi-scenario joint entity-relationship extraction module 230 is used to construct a multimodal joint entity-relationship Transformer model. Taking the multimodal fusion features as input, it extracts emergency entities and the relationships between entities end-to-end, and outputs standardized entity-relationship triples and their extraction confidence. Furthermore, the core function of the multi-scenario joint entity-relation extraction module 230 is to achieve end-to-end entity recognition and relation extraction based on the MJERT model; it includes a Transformer encoder submodule, an entity boundary detection submodule, a relation classification submodule, a constraint decoding submodule, and a semi-supervised few-shot learning submodule; and it outputs standardized entity-relation triples and extraction confidence scores.

[0111] The spatiotemporal knowledge graph construction and management module 240 is used to construct a three-level hierarchical knowledge graph adapted to a multi-level command architecture based on the entity-relation triplet and using the five-tuple spatiotemporal knowledge representation method, and to assign a comprehensive confidence level to each piece of knowledge in the three-level hierarchical knowledge graph. Furthermore, the core functions of the spatiotemporal knowledge graph construction and management module 240 are: responsible for the creation, fusion, updating and storage of the three-level hierarchical graph; including the ontology mapping submodule, the MHKFF multi-source knowledge fusion submodule, the knowledge completion and error correction submodule, the hybrid storage and multi-dimensional indexing submodule, and the graph dynamic update submodule; providing a structured knowledge foundation for decision-making.

[0112] The multi-level emergency command reasoning and decision-making module 250 is used to generate emergency decision-making schemes based on the three-level hierarchical knowledge graph, using a three-layer reasoning framework that integrates rule deduction, analogical reasoning and graph structure analysis, and output the comprehensive confidence level of the emergency decision-making scheme. Furthermore, the multi-level emergency command reasoning and decision-making module 250 serves as the core module of the system. Its core functions are: to realize three-layer fusion reasoning based on knowledge graphs; including rule reasoning engine submodule, case matching submodule, graph structure analysis submodule, situation assessment submodule, resource allocation optimization submodule, path planning submodule, instruction generation and closed-loop control submodule; and to output emergency decision-making schemes and decision confidence levels.

[0113] The multi-scenario application adaptation and human-machine collaboration module 260 is used to execute corresponding human-machine collaboration rules based on the preset emergency response level and the comprehensive confidence level of the decision, to conduct hierarchical review, instruction issuance and execution feedback of the emergency decision plan, and form a closed-loop control of the entire disposal process.

[0114] Furthermore, the core functions of the multi-scenario application adaptation and human-machine collaboration module 260 are: responsible for application adaptation in five major scenarios, including smart transportation, smart community, smart campus, smart government affairs, and smart tour guide subsystems; responsible for the full-process control of human-machine collaboration, including hierarchical review submodule, permission management submodule, timeout circuit breaker submodule, and decision tracing submodule; and providing scenario-based operation interfaces and interfaces.

[0115] In some preferred embodiments of the present invention, the system further includes: a top-level ontology and rule management module 270, whose core functions are: responsible for the creation, editing, and version management of the four-dimensional spatiotemporal ontology architecture in the emergency domain; maintenance of the multi-scenario entity-relationship system; editing, reviewing, and conflict verification of the reasoning rule base; and operation of the improved AMIE+ rule automatic mining and causal verification submodule, which is the semantic benchmark module of the system.

[0116] In some preferred embodiments of the present invention, the system further includes: a cloud-edge collaboration and system operation and maintenance module 280, whose core function is to be responsible for the management and control of the dual-track operation architecture, realize data synchronization, model updates and permission control between the cloud and the edge; including a lightweight deployment submodule for the edge, a model incremental iteration submodule, a FL collaborative optimization submodule, a system monitoring submodule, a security audit submodule and a log management submodule; to ensure the safe and stable operation and continuous optimization of the system.

[0117] The knowledge graph-based multi-level emergency command intelligent decision-making system provided in Embodiment 2 of the present invention can execute the operation steps of the knowledge graph-based multi-level emergency command intelligent decision-making method provided in any of the embodiments of the present invention, and has the corresponding functions and beneficial effects of the knowledge graph-based multi-level emergency command intelligent decision-making method. For detailed process, please refer to the relevant operations of the knowledge graph-based multi-level emergency command intelligent decision-making method in the foregoing embodiments.

[0118] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0119] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A multi-level emergency command intelligent decision-making method based on knowledge graphs, characterized in that, Includes the following steps: S1: Collect and standardize the multi-source heterogeneous data in the multi-level emergency command scenario, and extract the exclusive features of text modality, visual modality and spatiotemporal modality respectively; S2: Construct a cross-modal attention fusion network that includes a spatiotemporal-aware multi-head attention mechanism, perform deep semantic alignment and fusion on the extracted text modality, visual modality and spatiotemporal modality-specific features, and output multimodal fusion features and their fusion confidence. S3: Construct a multimodal joint entity relationship Transformer model, using the multimodal fusion features as input, to jointly extract emergency entities and relationships between entities end-to-end, and output standardized entity-relationship triples and their extraction confidence. S4: Based on the entity-relationship triples, a three-level hierarchical knowledge graph adapted to a multi-level command architecture is constructed using the five-tuple spatiotemporal knowledge representation method, and a comprehensive confidence level is assigned to each piece of knowledge in the three-level hierarchical knowledge graph. S5: Based on the three-level hierarchical knowledge graph, an emergency decision-making scheme is generated using a three-layer reasoning framework that integrates rule deduction, analogical reasoning, and graph structure analysis, and the comprehensive confidence level of the emergency decision-making scheme is output. S6: Based on the preset emergency response level and the comprehensive confidence level of the decision, execute the corresponding human-machine collaboration rules to conduct hierarchical review, instruction issuance and execution feedback of the emergency decision-making plan, forming a closed-loop control of the entire disposal process.

2. The multi-level emergency command intelligent decision-making method based on knowledge graphs according to claim 1, characterized in that, Step S1 involves extracting specific features for the text modality, visual modality, and spatiotemporal modality, respectively, including: Text modality feature extraction sub-step: A hierarchical semantic feature extraction architecture is adopted, which is built based on a distilled version of the BERT model and a bidirectional gated recurrent unit. First, the preprocessed text sequence is input into the distilled version of the BERT model to generate context-aware word-level embeddings. Then, the word-level embeddings are input into the bidirectional gated recurrent unit, and sentence-level feature aggregation is performed in combination with the attention mechanism to output the text modality-specific features and their semantic confidence. Visual modality feature extraction sub-steps: Construct a domain-adaptive multi-stage lightweight visual feature extraction framework based on a lightweight convolutional neural network and a feature pyramid network; input preprocessed image or video keyframes into the lightweight convolutional neural network and the feature pyramid network to extract multi-scale visual features; highlight key regions related to emergency events through an attention-based region proposal network, and output the specific features of the visual modality and its visually valid confidence. Spatiotemporal modality feature extraction sub-step: A lightweight tensor modeling method based on one-dimensional convolution and positional encoding is adopted; for the spatial coordinates and timestamps of each emergency event or emergency entity, spatial positional encoding and time encoding are constructed respectively, and the two are concatenated to generate a spatiotemporal embedding vector; all spatiotemporal embedding vectors are organized into a three-dimensional tensor structure that retains the time dimension, spatial dimension and feature dimension, and the exclusive features of spatiotemporal modality are output. The feature extraction networks for the text modality, visual modality, and spatiotemporal modality are all adapted to the lightweight requirements of edge deployment and output confidence scores for the corresponding modality features.

3. The multi-level emergency command intelligent decision-making method based on knowledge graphs according to claim 1, characterized in that, The cross-modal attention fusion network constructed in step S2 specifically includes: The text-image interaction attention module is used to construct a bidirectional cross-modal attention mechanism. It calculates the attention weights of text features on visual features and the attention weights of visual features on text features, and enhances the text features and visual features based on the attention weights to achieve deep semantic interaction between the text modality and the visual modality. The spatiotemporal-aware multi-head attention module is used in the multi-head attention mechanism to superimpose the spatiotemporal distance matrix constructed based on spatiotemporal position encoding as a bias term onto the attention score for weighted adjustment, so as to constrain the model to prioritize the key features of spatiotemporal coherence and output a multimodal feature representation that integrates spatiotemporal semantics.

4. The multi-level emergency command intelligent decision-making method based on knowledge graphs according to claim 3, characterized in that, The cross-modal attention fusion network also includes: The feature space alignment module is used to align the feature distributions of text modality and visual modality based on optimal transport theory by minimizing the Wasserstein distance between the feature distributions of text modality and visual modality, and by combining the loss function of downstream tasks, thereby eliminating the distribution heterogeneity between features of different modalities. The fusion feature confidence quantification module is used to comprehensively calculate the overall confidence score of multimodal fusion features based on intermodal semantic matching degree, feature integrity, and feature alignment loss.

5. The multi-level emergency command intelligent decision-making method based on knowledge graphs according to claim 1, characterized in that, In step S3, the multimodal joint entity relationship Transformer model specifically includes: The Transformer encoder submodule is used to perform deep context encoding on the multimodal fusion features, capture long-distance dependencies between features, and generate context-aware feature representations. The entity boundary detection submodule uses a sequence labeling method based on span annotation to predict the start and end positions of entities from the context-aware feature representation. The entity type classification submodule is used to classify the detected entities by type; The relation classification submodule uses a table-filling method to model the semantic relationships between all entity pairs, thereby achieving joint classification of relationships between entities. The three sub-modules of entity boundary detection, entity type classification, and relation classification serve as a joint prediction head. Based on the shared context-aware feature representation, they perform end-to-end joint training and inference, and synchronously output standardized entity-relation triples and their corresponding extraction confidence scores.

6. The multi-level emergency command intelligent decision-making method based on knowledge graphs according to claim 5, characterized in that, Step S3 further includes: The confidence quantification sub-step of the extraction results calculates the entity confidence and relation confidence respectively; the entity confidence is determined based on the entity boundary prediction probability, the entity type classification probability, and the compliance score of whether the entity conforms to the semantic constraints of the top-level ontology; the relation confidence is determined based on the relation classification probability, the confidence of the subject and object entities, and the compliance score of whether the relation conforms to the semantic constraints of the top-level ontology. The confidence threshold dynamic adjustment sub-step presets a confidence threshold associated with the risk level of the emergency scenario, and automatically adjusts the confidence threshold according to the risk level or data quality of the current emergency event. For extraction results that are lower than the confidence threshold, manual verification or downgrade processing is triggered. The model training optimization sub-step employs a multi-task joint loss function to train the multimodal joint entity-relation Transformer model. This multi-task joint loss function integrates entity boundary detection loss, entity type classification loss, and relationship classification loss, and introduces focus loss to address the imbalance in the distribution of entity and relationship categories. Simultaneously, a semi-supervised few-shot cross-scene transfer learning strategy is adopted, using a three-stage training framework of pre-training, domain adaptation, and fine-tuning to reduce the annotation cost of cross-scene adaptation.

7. The multi-level emergency command intelligent decision-making method based on knowledge graphs according to claim 1, characterized in that, In step S4, the three-level hierarchical knowledge graph includes: The overall map, deployed at the central node in the cloud, is used to integrate core knowledge across the entire domain and support global situational awareness and cross-regional collaborative decision-making. The sub-maps deployed at regional sub-nodes are used to focus on emergency knowledge in corresponding administrative regions or industry sectors to support command and analysis at the local level. A scenario sub-map deployed at edge-end field response nodes is used to support rapid on-site response and precise handling for specific emergency scenarios.

8. The multi-level emergency command intelligent decision-making method based on knowledge graphs according to claim 1, characterized in that, In step S5, the three-layer reasoning framework specifically includes: Rule-based deductive reasoning: Based on a pre-built emergency domain reasoning rule base, logical deduction is performed to generate decision recommendations that conform to expert experience and national standards; Embedded analogical reasoning: Based on the embedding of entities and relationships in a knowledge graph, retrieve historical emergency response cases that are most similar to the current event and reuse their successful response plans; Graph structure analysis and reasoning: Analyze the topological structure of the knowledge graph using graph algorithms to identify key nodes, bottlenecks, and potential risk propagation paths in emergency response; The results of the above three reasoning methods are combined to generate the final emergency decision-making plan.

9. The multi-level emergency command intelligent decision-making method based on knowledge graphs according to claim 8, characterized in that, The method further includes: an automatic rule mining and optimization step, specifically including: An improved rule mining algorithm is used, which combines spatiotemporal constraints with domain ontology semantic constraints, to automatically mine candidate reasoning rules from the knowledge graph of historical emergency cases; The candidate rules are tested for causal effect using a causal inference framework, eliminating spurious rules and retaining rules with causal relationships. To calculate the support and confidence indices of rules that pass the causal test, the rules are screened and then reviewed by experts before being added to the database, so as to realize the semi-automatic construction and low-cost maintenance of the inference rule base.

10. A multi-level emergency command intelligent decision-making system based on knowledge graphs, characterized in that, include: The multi-source data access and preprocessing module is used to collect and standardize the multi-source heterogeneous data in the multi-level emergency command scenario, and extract the exclusive features of text modality, visual modality and spatiotemporal modality respectively. The cross-modal attention fusion module is used to construct a cross-modal attention fusion network that includes a spatiotemporally aware multi-head attention mechanism. It performs deep semantic alignment and fusion on the extracted text modality, visual modality and spatiotemporal modality-specific features, and outputs multimodal fusion features and their fusion confidence. The multi-scenario joint entity-relationship extraction module is used to construct a multimodal joint entity-relationship Transformer model. Taking the multimodal fusion features as input, it extracts emergency entities and the relationships between entities end-to-end, and outputs standardized entity-relationship triples and their extraction confidence. The spatiotemporal knowledge graph construction and management module is used to construct a three-level hierarchical knowledge graph adapted to a multi-level command architecture based on the entity-relation triplet and using the five-tuple spatiotemporal knowledge representation method, and to assign a comprehensive confidence level to each piece of knowledge in the three-level hierarchical knowledge graph. The multi-level emergency command reasoning and decision-making module is used to generate emergency decision-making schemes based on the three-level hierarchical knowledge graph, using a three-layer reasoning framework that integrates rule deduction, analogical reasoning and graph structure analysis, and outputs the comprehensive confidence level of the emergency decision-making scheme. The multi-scenario application adaptation and human-machine collaboration module is used to execute corresponding human-machine collaboration rules based on the preset emergency response level and the comprehensive confidence level of the decision, to conduct hierarchical review, instruction issuance and execution feedback of the emergency decision plan, forming a closed-loop control of the entire disposal process.

Images