An Emergency Command Decision Evaluation Method Based on a Four-Dimensional Visualization Cloud Platform

Abstract

This invention relates to an emergency command decision-making and evaluation method based on a four-dimensional visualization cloud platform. The method involves structured collection of historical cases and heterogeneous sensor data, extracting multi-dimensional features such as event type, influencing factors, response process, spatial-temporal and causal chains, and mapping these features into a dynamic ontology library after normalization. This enables structured expression and semantic association of cross-case knowledge. Based on this, AI-driven semantic structure transfer and graph neural network characteristic transfer are utilized to quickly match the most relevant knowledge fragments for new events and adaptively generate predictive models. Real-time sensing data is then combined to dynamically optimize the model. This technology effectively improves the accuracy and generalization ability of knowledge transfer during emergency response, achieving continuous adaptive situation prediction and highly reliable decision support, thus enhancing the real-time performance and intelligence of the emergency command system.

Description

Technical Field

[0001] This invention relates to the field of "intelligent decision-making and situation prediction technology in emergency management", and in particular to an emergency command decision-making evaluation method based on a four-dimensional visualization cloud platform. Background Technology

[0002] Currently, intelligent emergency command and situation prediction technologies are rapidly developing towards high integration, multi-source information fusion, and intelligent analysis and decision-making. In traditional emergency management systems, the assessment and decision-making of event situations rely on expert experience, single data sources, and rule-driven reasoning. With the increasing maturity of artificial intelligence, big data, cloud computing, and ontology modeling technologies, the industry is gradually exploring how to deeply integrate historical emergency cases and real-time sensing data to improve the ability to analyze, predict, and make decisions about emergencies.

[0003] The existing technology has the following prominent problems:

[0004] (1) Most current adaptive situation prediction models are susceptible to the influence of event type and scenario specialization when using historical case knowledge for model generalization, resulting in poor transfer effect under new categories or heterogeneous events and a decrease in prediction output accuracy.

[0005] (2) Due to the complexity of case data sources, heterogeneity of structure, and many semantic ambiguities, traditional data-driven or shallow rule-based methods are difficult to achieve deep-level structured extraction and flexible reorganization of event elements, response processes and causal chain relationships, making it difficult for the model to fully absorb and transfer historical experience to new scenarios.

[0006] (3) Most existing solutions exhibit problems such as model rigidity, poor generalization ability, and weak dynamic adaptability in dealing with category migration and unknown risk scenarios, making it difficult to make timely and accurate predictions and adaptive adjustments to the diversified evolution of emergencies.

[0007] (4) Few publicly available solutions establish multidimensional ontology dynamic libraries, and there is insufficient structural reuse and modeling of causal chains, multi-source triggers and spatial-temporal evolution laws of historical cases, which seriously restricts the expansion and generalization of models for complex scenarios. Summary of the Invention

[0008] This application provides an emergency command decision-making and evaluation method based on a four-dimensional visualization cloud platform, which aims to solve one of the problems or issues of the existing technology mentioned in the background.

[0009] This application provides an emergency command and decision-making assessment method based on a four-dimensional visualization cloud platform, which specifically includes:

[0010] S1: Based on emergency command scenarios, collect historical emergency response case data and real-time multi-source sensing data to record event types, influencing factors, response processes, spatial-temporal labels, and dynamic characteristics of causal chains.

[0011] S2: Perform heterogeneous data normalization preprocessing on the collected historical emergency response case data and real-time multi-source sensing data to unify the multi-dimensional event feature vector format.

[0012] S3: Based on the normalized event feature vector, through domain knowledge sorting and feature label extraction, a multi-dimensional dynamic ontology library is constructed, which includes event type, influencing factors, response process, spatial-temporal labels, and dynamic causal chain, to realize the structured knowledge mapping of historical cases.

[0013] S4: Input the scene characteristics of newly emerging emergency events into a multi-dimensional dynamic ontology library, and retrieve historical case semantic groups with multiple cross-dimensional and different scene distributions based on spatial-temporal labels, event types and related causal chains.

[0014] S5: Applying an AI-driven semantic structure transfer algorithm, semantic segmentation and component reorganization are performed on the response sequences and evolution processes in the retrieved historical case semantic groups to extract key adaptive knowledge fragments, including resource deployment, decision nodes, and turning points.

[0015] S6: Based on the context information of the multidimensional dynamic ontology library, the above adaptive knowledge fragments are input into the graph neural network for feature transfer learning, and an adaptive state prediction model structure and parameter set that integrates the dynamic features of the current new event are automatically generated.

[0016] S7: During the model inference and prediction process, the changes in real-time multi-source sensing data are dynamically monitored, triggering the state update and adaptive loading of relevant feature factors in the multi-dimensional dynamic ontology library, thereby achieving continuous adaptive optimization of the prediction model parameters and structure.

[0017] S8: Combine real-time multi-source sensing data to perform online error judgment on the dynamic adaptive state prediction model. If the feature transfer effect of historical cases does not meet the preset conditions for prediction accuracy, the feature transfer weights will be automatically adjusted or the ontological structure fragments will be reorganized.

[0018] S9: The output of the adaptive state prediction model after multidimensional dynamic ontology enhancement and feature transfer optimization is input into the ontology-constrained decision simulation and confidence assessment process to generate multi-path situation prediction output, providing intelligent assessment support for emergency command decision-making.

[0019] This application provides an emergency command and decision-making evaluation method based on a four-dimensional visualization cloud platform, which has the following beneficial effects:

[0020] (1) By constructing a dynamic ontology library containing multi-dimensional elements such as event type, influencing factors, response process, spatial-temporal labels, and causal chain dynamics, the deep structuring and semantic label normalization of historical cases are effectively realized, enabling the situation prediction model to achieve cross-domain and cross-temporal knowledge transfer and generalization for various emergencies based on a unified ontology expression. In actual case evaluation, the model improves prediction confidence by more than 15% when facing new events of unknown type or with complex evolution paths, and the model's applicability has expanded from covering a single event category to covering multiple categories and heterogeneous scenarios.

[0021] (2) This invention innovatively introduces an AI-driven semantic structure transfer and component-based knowledge reorganization mechanism, which can automatically extract high-value elements such as response sequences, decision nodes, and resource allocation from historical cases to form reusable knowledge fragments and support dynamic assembly in new event contexts.

[0022] (3) By combining a multidimensional ontology, hierarchical matching and retrieval of multi-level semantics such as event type, space-time, and causal chain are realized, overcoming the transfer failure problem caused by relying on a single-dimensional label or coarse-grained case retrieval in existing technologies. The ontology flexibly supports quantitative evaluation of similarity in multiple scenarios, effectively filtering out the case groups with the most transfer value.

[0023] (4) During model inference, the dynamic linkage between real-time multi-source perception data and ontology feature factors ensures that the prediction structure and parameters can be continuously optimized according to changes in the on-site situation. Through online error judgment mechanism and adaptive adjustment of transfer weights, ontology structure reorganization can be triggered even when historical data feature transfer is insufficient or the scene changes drastically, avoiding model rigidity or transfer distortion. Experimental results show that under abnormal scenarios or extreme events, the system can complete ontology reorganization and model parameter adjustment within 2 seconds, and the key prediction error is significantly reduced compared with traditional static transfer.

[0024] (5) This invention organically combines multi-path decision simulation, ontology consistency verification, and confidence interval assessment, effectively eliminating problems such as semantic conflicts in historical cases and contradictions in reasoning paths, and significantly reducing the risk blind spots in decision simulation. System testing shows that the detection rate of abnormal paths has been significantly improved, and multi-path situation prediction can provide uncertainty quantification for emergency command covering multiple options, significantly improving the scientificity and comprehensiveness of emergency plan formulation.

[0025] In summary, this invention not only achieves a comprehensive breakthrough in the generalization, intelligence, knowledge reuse, and risk prediction software of existing emergency command and situation prediction technologies, but also provides an innovative theoretical and engineering foundation for intelligent decision-making, combat command, and the construction of large-scale response systems in various industries. Its core technological features and methods have significant application value for enhancing my country's intelligent emergency command capabilities and strengthening public safety prevention and control. Attached Figure Description

[0026] Appendix Figure 1 This is the main flowchart of an emergency command decision-making and evaluation method based on a four-dimensional visualization cloud platform.

[0027] Appendix Figure 2 This is a sub-flowchart of an emergency command decision-making and evaluation method based on a four-dimensional visualization cloud platform.

[0028] Appendix Figure 3 This is another sub-flowchart of an emergency command decision-making and evaluation method based on a four-dimensional visualization cloud platform. Detailed Implementation

[0029] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0030] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0031] As attached Figure 1 As shown, this application provides an emergency command decision-making and evaluation method based on a four-dimensional visualization cloud platform, specifically including:

[0032] S1: Based on emergency command scenarios, collect historical emergency response case data and real-time multi-source sensing data to record event types, influencing factors, response processes, spatial-temporal labels, and dynamic characteristics of causal chains.

[0033] S2: Perform heterogeneous data normalization preprocessing on the collected historical emergency response case data and real-time multi-source sensing data to unify the multi-dimensional event feature vector format.

[0034] S3: Based on the normalized event feature vector, through domain knowledge sorting and feature label extraction, a multi-dimensional dynamic ontology library is constructed, which includes event type, influencing factors, response process, spatial-temporal labels, and dynamic causal chain, to realize the structured knowledge mapping of historical cases.

[0035] S4: Input the scene characteristics of newly emerging emergency events into a multi-dimensional dynamic ontology library, and retrieve historical case semantic groups with multiple cross-dimensional and different scene distributions based on spatial-temporal labels, event types and related causal chains.

[0036] S5: Applying an AI-driven semantic structure transfer algorithm, semantic segmentation and component reorganization are performed on the response sequences and evolution processes in the retrieved historical case semantic groups to extract key adaptive knowledge fragments, including resource deployment, decision nodes, and turning points.

[0037] S6: Based on the context information of the multidimensional dynamic ontology library, the above adaptive knowledge fragments are input into the graph neural network for feature transfer learning, and an adaptive state prediction model structure and parameter set that integrates the dynamic features of the current new event are automatically generated.

[0038] S7: During the model inference and prediction process, the changes in real-time multi-source sensing data are dynamically monitored, triggering the state update and adaptive loading of relevant feature factors in the multi-dimensional dynamic ontology library, thereby achieving continuous adaptive optimization of the prediction model parameters and structure.

[0039] S8: Combine real-time multi-source sensing data to perform online error judgment on the dynamic adaptive state prediction model. If the feature transfer effect of historical cases does not meet the preset conditions for prediction accuracy, the feature transfer weights will be automatically adjusted or the ontological structure fragments will be reorganized.

[0040] S9: The output of the adaptive state prediction model after multidimensional dynamic ontology enhancement and feature transfer optimization is input into the ontology-constrained decision simulation and confidence assessment process to generate multi-path situation prediction output, providing intelligent assessment support for emergency command decision-making.

[0041] Step S1: Based on the emergency command scenario, collect historical emergency response case data and real-time multi-source sensing data, and record event type, influencing factors, response process, spatial-temporal labels, and dynamic characteristics of causal chains. Specifically, this includes:

[0042] S1.1: Perform structured retrieval on the historical emergency response case database, and use case label parsing and scene feature extraction algorithms to obtain a historical case dataset containing clear event types, so as to generate original historical case data with labeled event types.

[0043] The dataset from the historical emergency response case database serves as the input for this step. It includes a raw case library containing multi-dimensional descriptions of event types, affected elements, response process information, spatial-temporal labels, and potential causal chain dynamics.

[0044] A structured retrieval method (parameters: event type label, response time sequence index, and influencing factor classification code) is adopted to achieve batch filtering and rapid retrieval of case entries with data integrity and detailed scene description in the historical emergency case database.

[0045] Furthermore, through the case label parsing algorithm (parameters: natural language label matching rules, category weight setting, and deambiguity segmentation template), the automatic parsing and standardized labeling of key information such as event type, main affected objects, and occurrence scenarios in candidate historical case data are achieved.

[0046] Furthermore, relying on the scene feature extraction algorithm (parameters: semantic vector embedding model, context feature induction network, cross-domain knowledge matching rules), the algorithm progressively extracts spatiotemporal factors such as environmental features, response nodes, and decision-making turning points in structured tagged cases and aggregates them according to preset fields.

[0047] By combining the above algorithms, a historical case dataset is generated that includes clear event type labels, main influencing factors, and basic hierarchical response processes, providing highly consistent and scalable raw input for subsequent element identification and dynamic causal chain modeling.

[0048] By employing structured retrieval and automated tag parsing and scene feature extraction processes, the data structuring level of historical cases is significantly improved, enabling accurate labeling of event type information and reducing the risk of tag ambiguity and structural misalignment in subsequent feature extraction and ontology mapping processes.

[0049] For example, targeting a database of 423 historical emergency response cases of hazardous chemical leaks in a certain city from 2020 to 2022, a structured search method based on the Elasticsearch search engine was used. The event type was set as "hazardous chemical leak," the response time range was set as "January 2020 to December 2022," and the impact factor classification codes included "personnel evacuation" and "firefighting resource deployment," quickly yielding 392 original cases. A BERT-BiLSTM-CRF model was used to parse the case labels, setting category weights of 0.4 for casualties, 0.3 for environmental impact, 0.2 for equipment damage, and 0.1 for social impact, and parsing standardized labels. The scene feature extraction algorithm set spatial labels to street-level geographical division, temporal granularity to hourly level, and response node priority to rescue—control—aftermath. Ultimately, 392 structured case data items were obtained, each with event type, key control element annotations, response process nodes, and spatial-temporal labels. The label parsing accuracy rate reached over 96%, and the average retrieval and parsing time per case was 0.28 seconds, which strongly supported feature recognition and model generalization input in subsequent processes.

[0050] S1.2: Based on multi-source sensing terminals (such as video surveillance, sensor clusters, remote sensing devices, etc.), real-time environmental sensing data of the emergency site is acquired. High-frequency acquisition of heterogeneous real-time sensing data is achieved through automated acquisition protocols to obtain real-time multi-source sensing raw data covering spatial-temporal labels.

[0051] Multi-source sensing terminals (such as video surveillance, sensor clusters, remote sensing devices, etc.) are deployed and configured to cover key geographical areas and response links at emergency sites, and to acquire raw environmental sensing signals with high spatial-temporal resolution.

[0052] Employing a multi-channel high-frequency data acquisition protocol (parameters: timing sampling frequency ≥ 10Hz, spatial coverage density > 20 acquisition units / hectare, communication protocol compatible with RTSP / MQTT / LoRaWAN, etc.), it achieves near real-time acquisition of data from multi-modal environmental sensing terminals on-site, covering multiple types of data streams such as video streams, sensor physical quantities, and remote sensing data.

[0053] Furthermore, by using an intelligent synchronization calibration method (parameters: timestamp alignment difference ≤ 500ms, spatial location tolerance deviation < 3m), the time axis and geographic coordinates of sensing data from different terminals are automatically aligned, eliminating the spatiotemporal offset of heterogeneous signal acquisition and improving the integration consistency of multi-source sensing data.

[0054] Furthermore, through node health self-check and abnormal signal elimination algorithms (parameters: online rate threshold of acquisition nodes ≥95%, abnormal data detection standard is 3σ criterion for elimination), the integrity and real-time performance of data of each node in the acquisition link are actively monitored, while invalid or fluctuating abnormal data packets are screened out to ensure the high validity of the acquired raw data.

[0055] Furthermore, through a multi-source data buffering and fragmented archiving mechanism (parameters: data fragment duration 5s-10s, archiving cache time ≤30s), high-speed caching, segmented archiving, and on-demand backtracking of various types of raw sensing data are achieved, ensuring that the dynamic environmental data on site can be continuously traced and accessed for subsequent processing.

[0056] Through the aforementioned high-frequency, multi-source, and automated acquisition strategy, the multi-dimensional sensing signals of the on-site environment are transformed in real time into multi-source sensing raw data covering space-time labels. This provides highly consistent, high-resolution, and low-latency underlying data support for subsequent identification of influencing factors, analysis of response processes, and dynamic modeling of causal chains, achieving full-domain coverage and efficient acquisition of dynamic holographic data of on-site emergency events.

[0057] For example, in a hazardous chemical leak emergency drill conducted in a district of a city in November 2022, 12 high-definition video surveillance systems, 23 sets of wireless environmental temperature and humidity / toxic gas sensor modules, 2 sets of satellite remote sensing mobile base stations, and 1 drone aerial photography terminal were deployed. A multi-channel RTSP streaming protocol was configured, with a video stream sampling frequency of 25 frames per second, a sensor data acquisition frequency of 20Hz, a remote sensing mobile platform resolution of 0.5m / pixel, and a drone timed transmission interval of 30 seconds. NTP clock synchronization was used to ensure that the timestamp accuracy of all terminal data acquisition was within 300 milliseconds. Node health status detection ensured a 97% online rate for on-site data acquisition nodes, automatically removing short-term offline node data and electromagnetic interference background noise packets. Multimodal raw data was fragmented, stored, and indexed every 10 seconds, supporting rapid retrieval and backtracking of all-time sensing data. During the testing period, the total amount of raw data collected exceeded 8GB, the average latency of video content aggregation was 350 milliseconds, and the sensor data integrity rate reached 99.2%, providing high-density and high-continuity data support for subsequent automatic feature recognition and response evaluation.

[0058] S1.3: Apply the element recognition algorithm to the original data of the above historical cases and the original data of real-time multi-source perception to identify and extract influencing elements (such as personnel, equipment, environmental indicators, etc.) to generate a structured feature vector of influencing elements.

[0059] S1.4: Based on the extracted feature vectors of influencing factors, apply the response process analysis algorithm to the historical case data to analyze the event response process and key node time sequence in stages, and obtain the response process sequence with causal chain clues.

[0060] S1.5: For the obtained response flow sequence and its associated spatial-temporal labels, a causal chain dynamic modeling algorithm is used to extract the inherent spatiotemporal development law and causal relationship of the event, so as to form an event knowledge causal chain labeled with dynamic characteristics of the causal chain.

[0061] Step S2: Perform heterogeneous data normalization preprocessing on the collected historical emergency response case data and real-time multi-source sensing data to unify the multi-dimensional event feature vector format. Specifically, this includes:

[0062] S2.1: Classify historical emergency response case data and real-time multi-source sensing data by type. Use data pattern recognition algorithm to classify unstructured text, structured tables, time-series sensor signals and other data into specific data types. This will unify the technical labels of different input data during the acquisition stage and provide an input basis for subsequent categorized feature preprocessing.

[0063] The system categorizes historical emergency response case data and real-time multi-source sensing data by type. Based on the heterogeneous data normalization process, the input includes multi-dimensional structured historical case raw data containing event type, influencing factors, response process, spatial-temporal labels and causal chain dynamic characteristics, as well as real-time multi-source sensing raw data such as time-series sensor signals, video streams, and remote sensing data acquired through multi-channel acquisition.

[0064] A data pattern recognition algorithm (parameters: feature space distribution, acquisition protocol encoding, data file header format feature set) is employed to achieve initial pattern detection and recognition of historical emergency response case data and real-time multi-source sensing data. Furthermore, a hybrid structured and unstructured discrimination method (parameters: regular expression matching rules, natural language processing (NLP) entity boundary discrimination, metadata index mapping) is used to classify unstructured text, structured database tables, and time-series signal data with specific tags. Further, an automatic data type tag mapping mechanism (parameters: tag set covering "text", "table", "time-series signal", "image / video", and "geospatial") maps identified data entries one-to-one with a technical tag library, ensuring each data source obtains a unique technical tag during the acquisition phase. Finally, a data source tag index engine (parameters: hash mapping, tag priority sorting function, index refresh cycle) is used to centrally manage the technical tag table of all input data, providing a standardized tag index input for categorized feature normalization and subsequent feature extraction sub-processes. Furthermore, the automatic label allocation results are subjected to multi-source label consistency verification (parameters: cross-validation matrix, data content redundancy detection ratio, label conflict discrimination rules) to identify and correct label conflicts and redundant allocations under the same event or the same collection period, ensuring the semantic uniqueness and integrity of the label classification results.

[0065] By using data pattern recognition algorithms and technology label normalization mapping, the original data of the aforementioned historical cases and the original data of real-time perception are clearly stratified, enabling targeted data classification during the data collection stage, unifying the technology labels of heterogeneous input data, significantly reducing input ambiguity in subsequent feature preprocessing steps, and improving the foundation for data standardization.

[0066] For example, based on historical cases of a hazardous chemical leak emergency drill in a certain district of a city in November 2022 and real-time multi-source sensing data, 494 event records were exported from the historical case database, using ".csv" and ".docx" as data containers. The on-site multi-source sensing system provided 12 high-definition video streams (".mp4"), 23 environmental sensor time-series signal data (".dat"), and 2 sets of remote sensing sampling images (GeoTIFF format), realizing five types of raw data input. A hybrid algorithm of content-based file header recognition and NLP entity extraction was adopted, setting file header keywords such as "ID, Type, Time", "docx:case", and "SIG:ENV" as features. Regular expression matching and BERT boundary discrimination were used to achieve an accuracy of 99.7% in distinguishing between structured data and unstructured text. Based on a label mapping table, five technical labels were assigned to each type of data: "tabular", "text", "time-series signal", "video", and "geospatial". The output data label table is uniquely indexed across all data sources, employing a hash mapping function (mapping rate of 30,000 records / second). A multi-source label consistency check identifies one misassigned label, which is automatically corrected using cross-validation rules. This achieves automatic classification and label normalization of the entire dataset, ensuring fundamental consistency in subsequent feature normalization and semantic label mapping processes. The above process executes in an average of 0.85 seconds per 100 data records, providing an efficient and accurate data foundation for the subsequent generation of multi-dimensional event feature normalized vectors.

[0067] S2.2: For classified historical emergency response case data, ontology label normalization mapping technology is applied to encode inconsistent text or symbols such as event type, influencing factors, response process, spatial-temporal labels and causal chain dynamic features into multi-dimensional event feature labels, laying the foundation for feature expression in subsequent vectorization conversion.

[0068] S2.3: For real-time multi-source sensing data, a multimodal feature extraction algorithm is adopted to summarize the feature patterns of time-series sensing signals, image monitoring streams, spatial geographic information, etc., and transform their original data features into feature labels of the same dimension as historical case data to ensure the consistency of semantic representation of multi-source data.

[0069] S2.4: Based on the above multi-dimensional event feature labels, a standardized coding strategy is implemented. Professional preprocessing algorithms such as numerical normalization and principal component compression are used to scale all data features and correct outliers, thereby obtaining event feature vectors with uniform dimensions and distribution, and improving the accuracy of cross-type data mapping.

[0070] S2.5: For the normalized multidimensional event feature vector, missing value imputation and noise reduction filtering algorithms are used to correct incomplete entries and perceptual redundancy that may exist in the data, ensuring that the final output event feature vector has high completeness and robustness, serving as the input basis for the multidimensional dynamic ontology library and supporting the subsequent semantic structure mapping and transfer learning process.

[0071] Step S3: Based on the normalized event feature vector, through domain knowledge organization and feature tag extraction, a multi-dimensional dynamic ontology library is constructed, including event type, influencing factors, response process, spatial-temporal labels, and causal chain dynamics, to achieve structured knowledge mapping of historical cases. For example... Figure 2 As shown, it specifically includes:

[0072] S3.1: Apply the feature decomposition algorithm to the normalized multidimensional event feature vector to extract the event type feature subset, so as to obtain the feature label set for different types of sudden events, providing basic input for domain rule selection and semantic mapping.

[0073] S3.2: Based on the feature subset of event type, feature component matching analysis is performed in conjunction with the domain knowledge base to extract feature labels of influencing factors, thereby realizing the structural representation and label mapping between event type and key influencing factors.

[0074] The normalized multidimensional event feature vector is used as input, and the event type feature subset (such as event category number, event attribute vector, etc.) is used as the analysis object.

[0075] A domain knowledge base retrieval method (parameters: event type feature subset, knowledge base indexing rules, and influencing element naming table) is adopted to realize domain knowledge node retrieval based on event type features and obtain a set of influencing element entries that are highly related to specific types of events.

[0076] Furthermore, by using a feature component matching algorithm (parameters: vector similarity threshold, influencing element classification criteria, and feature matching scoring function), component-level matching analysis is performed on the vector attributes in the feature subset of event types and the influencing element tuples in the domain knowledge base, thereby enabling candidate association screening of event types towards key influencing elements.

[0077] Furthermore, feature label extraction and merging techniques (parameters: hierarchical level of influencing element labels, redundant label removal rules, attribute association mapping table) are used to perform multi-level label discrimination and aggregation on the influencing elements in the matching results, generating a standardized list of influencing element feature labels covering categories such as technical equipment, personnel configuration, environmental parameters, and resource scheduling.

[0078] Furthermore, an event-feature structure representation algorithm (parameters: label network structure, type-feature mapping matrix, similarity weighting coefficient) is applied to achieve bidirectional association between nodes by structured mapping of event type feature labels and influencing feature labels, generating a type-feature relationship network with structural attributes.

[0079] Through the ontology label normalization mapping mechanism, the aforementioned output event type feature labels and influencing element feature labels are incorporated into the multidimensional dynamic ontology library in a node-edge mode, realizing the structural representation and label mapping between event types and key influencing elements, and providing a correlation basis for subsequent response process, spatial-temporal label and causal chain feature modeling.

[0080] For example, in a sample set of emergency response cases for hazardous chemical leaks, the normalized subset of event type features includes "leakage event [number A01]" and "toxic gas leak," etc.; the domain knowledge base is configured with a set of impact factor nodes for hazardous chemical events, including "type of hazardous materials," "leakage level," "weather parameters," "personnel emergency response group," and "rescue material grouping," etc. The system is configured with a feature component matching algorithm, with a similarity threshold set to 0.85, referring to the impact factor classification standard GB / T28001-2011, and the feature matching score function is calculated using weighted Jaccard distance. Through matching analysis, it was found that the coupling degree between "leakage event [number A01]" and the related features of "type of hazardous materials," "leakage level," and "weather parameters" exceeded the threshold, and it was selected as a key impact factor for the current event type. Using the impact factor label merging technology and the type-factor relationship network construction algorithm, a type-factor structure network is automatically generated, in which the event type node and the nodes of the five types of impact factors are significantly connected, forming a structured ontology label set, realizing the standardized mapping between historical case event types and key impact factors. In a sample of 989 historical cases, the accuracy of identifying influencing factors reached 97.3%, with an average of 4.7 key influencing factor labels mapped per event type, laying a precise foundation for subsequent response process feature modeling and contextual adaptive transfer.

[0081] S3.3: Apply the response flow parsing algorithm to the feature labels of influencing factors, analyze the response flow nodes at each stage of event evolution, generate response flow feature sequences, and ensure that the event response structure can be standardized and mapped to relevant attributes in the ontology library.

[0082] The input is a normalized set of feature labels for influencing factors, with data types covering multidimensional event feature vectors, and possessing event type markers, influencing factor structure, and label mapping attributes.

[0083] A response flow parsing algorithm (parameters: response node template library, flow segmentation discrimination rules, node timestamp threshold, association level weight) is adopted to realize the structured retrieval and matching of the feature labels of influencing elements and the candidate set of response flow nodes in each event sample.

[0084] Furthermore, an event phase segmentation method (parameters: domain event phase dictionary, feature state transition matrix, event node state label) is applied to realize the phased analysis of the entire event sample process, splitting the original response process into several process node segments with temporal and causal labels, and obtaining the process node segmentation results.

[0085] Furthermore, through a time-series aggregation algorithm (parameters: node time series, response duration statistics, stage transition criteria), the system integrates and summarizes the temporal logic and interrelationships between process nodes, determines the triggering, transition, and termination conditions of each stage node, and establishes a node-level time sequence arrangement.

[0086] Furthermore, a process feature sequence generation method (parameters: node identifier, logical relationship matrix between nodes, process standardization template) is adopted to connect the segmented response process nodes in series according to the triggering order and response logic to generate a complete response process feature sequence, and to realize the standardized mapping from process nodes to ontology attributes.

[0087] Through the above-mentioned response process parsing algorithm, stage segmentation and feature sequence generation mechanism, the normalized influence element feature labels are transformed into response process feature sequences with temporal logic and structured node attributes, realizing the standardized representation of the event response structure and supporting the structural mapping of related attributes in the ontology library.

[0088] For example, in a hazardous chemical leak case in a certain district of a city in 2022, the characteristic labels of influencing factors were obtained after normalization: "Type of hazardous material (Class A)", "Leakage intensity (Level II)", "Emergency team response level (High)", "Equipment deployment strategy (fixed-point coverage)", and "Meteorological conditions (southeast wind direction)". A response process analysis algorithm was adopted, and a response node template library was configured with reference to the national standard GB / T 29639-2013 "Guidelines for the Preparation of Emergency Response Plans for Production Safety Accidents in Production and Business Units". The process segmentation judgment rule adopted the sequential response principle and the emergency preparedness-preliminary handling-main response-post-remediation retrospective stage method. The node timestamp threshold was selected with a 5-minute granularity, and the stage transition criteria were based on sampling time and accident level determination. In practical application, five node segments are obtained in stages: "Emergency Preparation (0-5min)," "Alert and Evacuation (5-13min)," "Material Deployment and Initial Enclosure (13-22min)," "Main Response and Handling (22-65min)," and "Aftermath Management (65-95min)." The logic between process nodes is established through event stage status labels and adjacent node triggering criterion matrices. After standardization and template merging, the process feature sequence is expressed as: P1 (Emergency Preparation) → P2 (Alert and Evacuation) → P3 (Enclosure and Deployment) → P4 (Main Response) → P5 (Aftermath Management), with a one-to-one mapping between node attributes and ontology attributes. Verification with 990 historical hazardous material cases shows a 96.1% accuracy rate in response process segmentation and a 98.3% consistency rate in standardized node mapping, providing high-precision and highly consistent feature data input for the organization and construction of response process attributes in the ontology library.

[0089] S3.4: Based on the response process feature sequence, extract and aggregate spatial-temporal labels from event samples, and form a multi-dimensional spatial-temporal feature label group through a spatial-temporal association modeling algorithm to enrich the cross-scene indexing capability of the ontology library.

[0090] The input is a sequence of response flow features generated by the response flow parsing algorithm. This sequence carries standardized flow node attributes and time sequence information, and combines multi-source information such as spatial coordinates, geographic labels and timestamps from corresponding historical event samples.

[0091] A spatial information extraction algorithm (parameters: geographic coordinate system mapping rules, administrative division coding table, site geographic object database) is used to accurately extract spatial hierarchical information such as geographic spatial coordinates, event location, regional affiliation, and indoor and outdoor identification from process feature sequences and their original data, generating a structured set of spatial labels.

[0092] Furthermore, by using a time information segmentation and classification method (parameters: global time base, time segment division granularity, event node time concentration threshold), the event timestamps in the process feature sequence are summarized, statistically analyzed, and clustered to generate multi-granularity time feature labels such as the start and end times of events, duration, and key node time points.

[0093] Furthermore, a spatial-temporal label aggregation algorithm (parameters: spatial neighborhood criterion, spatiotemporal joint similarity measure, label redundancy discrimination rule) is applied to fuse the aforementioned structured spatial labels and temporal labels at multiple levels and scales to form a spatial-temporal composite feature set of event samples, thereby achieving deep coupling of spatial and temporal information in the same historical case.

[0094] Furthermore, by utilizing a spatial-temporal correlation modeling algorithm (parameters: spatial neighborhood threshold, temporal overlap index, and spatiotemporal interaction coefficient), the spatial-temporal correlation between different event samples and process nodes is modeled, and event distribution patterns and node evolution paths across geographical units and time periods are explored, ultimately generating a multi-dimensional spatial-temporal feature label group.

[0095] By normalizing and storing spatial-temporal feature labels, the aforementioned multi-dimensional spatial-temporal features are used as key dimensions of ontology attributes, enabling rapid cross-scenario indexing and high-level semantic retrieval capabilities for historical cases.

[0096] For example, in the historical case database of hazardous chemical leaks in a certain district of a certain city in 2022, the response process feature sequence node segments correspond to location codes (such as "116.48°E, 39.93°N"), geographical objects "Northeast Test Site of a Certain Chemical Industrial Park", and administrative division codes "11010506", with timestamp sequences of "2022-11-08 14:05", "2022-11-08 14:13", and "2022-11-08 14:22", etc. A spatial information extraction algorithm is implemented, and based on the WGS-84 geographic coordinate system and street division, the spatial label is precisely located to "Chemical Industrial Park - Northeast District of a Certain District". By segmenting and classifying time information, setting the granularity to 5 minutes, the node timestamps are collected, and the event start, phased response, and termination times are output, forming a total of 5 key time tags. The spatial-temporal tag aggregation algorithm, based on a spatial neighborhood radius of 200 meters and a temporal overlap threshold of 2 minutes, merges associated process nodes into composite tags such as "Emergency Preparedness - Northeast Test Site - 14:05-14:10" and "Alert and Evacuation - Main Road of the Park - 14:13-14:22". The spatial-temporal association modeling algorithm further analyzes and identifies three main spatial nodes involved in the accident response and their temporal evolution paths, forming multi-dimensional spatial-temporal feature tag groups. On average, each case outputs 4.8 sets of spatial-temporal composite tags. The output tag groups are embedded into the spatial-temporal attribute domain of a multi-dimensional dynamic ontology library, enabling cross-scene, multi-node, multi-exponential retrieval of historical cases, greatly improving the spatial-temporal indexing capability and accurate hierarchical retrieval efficiency of the ontology library.

[0097] S3.5: The spatial-temporal feature label group and the response flow feature sequence are jointly input into the causal chain dynamic modeling algorithm to discover dynamic features from the temporal and causal changes from the occurrence of an event to the response, construct a causal chain dynamic feature mapping structure, and empower the multi-dimensional dynamic ontology library with hierarchical dynamic knowledge expression capabilities.

[0098] S3.6: Based on the above causal chain dynamic feature mapping structure, execute the multi-label semantic fusion algorithm to restructure the event type, influencing factors, response process, spatial-temporal labels, and causal chain dynamic multi-dimensional features to form standardized multi-dimensional dynamic ontology nodes, providing basic node instances for the semantic network of the ontology library.

[0099] S3.7: Based on event category attributes, hierarchical labels and temporal logic, the standardized multidimensional dynamic ontology nodes are used to generate an ontology network mapping structure through an ontology library relationship construction algorithm, so as to organize and store the structured knowledge of historical cases according to multidimensional attributes and semantic relationships.

[0100] S3.8: Apply ontology update and consistency verification algorithms to dynamically update and redundancy verify the newly constructed ontology network mapping structure, ensuring that the multidimensional dynamic ontology library continuously maintains structural integrity and real-time accuracy of knowledge mapping, laying a solid knowledge foundation for generalization transfer and adaptive tuning of prediction models.

[0101] Step S4: Input the scene features of the newly emerging emergency event into a multi-dimensional dynamic ontology library, and retrieve multiple cross-dimensional, scene-distributed historical case semantic groups based on spatial-temporal labels, event types, and associated causal chains. For example... Figure 3 As shown, it specifically includes:

[0102] S4.1: Normalize the format matching process of the feature vector of the newly emerging emergency event scenario to ensure that it can be consistently compared with the structured knowledge of historical cases in the multidimensional dynamic ontology library, and obtain standardized scenario feature input.

[0103] The input is a feature vector of a newly emerging emergency event scenario. This feature vector includes a normalized event type identifier, influencing element labels, response process stages, spatial-temporal labels, and potential causal chain attributes. The data format is consistent with the structured knowledge specifications of historical cases in the multidimensional dynamic ontology library.

[0104] A feature format normalization algorithm (parameters: feature dimension alignment template, encoding specification table, data missing completion rules) is adopted to adjust the dimensional consistency of each element in the feature vector of new event scenarios, and to perform dimension-by-dimensional completion and standardized encoding of data fields that deviate from the ontology standard.

[0105] Furthermore, by using a feature vector regularization method (parameters: normalized scale interval [0,1], principal component mapping weights, and category label alignment function), the scene features are standardized in terms of numerical scale, distribution pattern, and category label level, providing a standardized expression for subsequent semantic matching retrieval.

[0106] Furthermore, a multi-dimensional label consistency discrimination algorithm (parameters: historical case feature label set, event attribute mapping table, label semantic similarity threshold) is adopted to compare all key label items of the new event feature vector and verify their standardization in the label dictionary item by item, so as to realize automatic completion of missing labels and error correction of abnormal labels.

[0107] Furthermore, through the structured data verification and repair module (parameters: structural consistency constraints, data integrity threshold, repair priority list), automatic structural verification is performed on the processed scene feature vectors. For feature components with missing or multiple meanings, model-based repair is performed according to the ontology library completion rules to ensure that the feature vectors meet the retrieval input requirements of the multidimensional dynamic ontology library.

[0108] Through the above-mentioned normalization format matching process, the feature vector of the newly emerging emergency event scenario is transformed into a highly consistent standardized scenario feature input, which enables high-precision consistency comparison with the structured knowledge of historical cases in the multi-dimensional dynamic ontology library, supporting multi-level semantic retrieval of event type, spatial-temporal labels and causal chains.

[0109] For example, in a sudden meteorological disaster event in a certain city in 2023, the feature vector collected from the original site included the event type "rainstorm and waterlogging," influencing factors such as "high precipitation rate, poor drainage system, and low-lying urban terrain," response process nodes such as "waterlogging warning → traffic control → emergency drainage → rescue and evacuation → recovery assessment," spatial-temporal labels such as "a certain city entrance in a certain urban area - 2023-07-19-16:30-18:10," and causal clues such as "precipitation impact → road waterlogging → traffic congestion → stranded people." A feature format normalization algorithm was used to perform feature dimension template alignment, filling in some missing elements such as "emergency team deployment," encoding all labels into the ontology standard format, and setting the normalization parameter range to [0,1]. Through a multi-dimensional label consistency discrimination algorithm, the feature label set of historical cases was compared with a semantic similarity threshold of 0.90, automatically merging "low-lying urban terrain" under the standard influencing factor label in the ontology library. After verification by the structured data repair module, a missing timestamp was found for the "traffic control" node. Based on the process stage rules, it was completed to "2023-07-19-16:40". The final output is a highly consistent standardized scene feature input vector, achieving 100% label completeness and a structural consistency score of 0.98. This results in high-precision format matching with historical cases in the multi-dimensional dynamic ontology library, providing a standard input basis for subsequent event type retrieval, spatial-temporal label analysis, and causal chain semantic retrieval.

[0110] S4.2: Based on the standardized scene features of the current input, use the event type label matching algorithm to perform a preliminary search of the multidimensional dynamic ontology library in terms of event type dimension, filter out the set of historical case semantic nodes that meet the event type requirements, and obtain the event type filtering results.

[0111] The input is a scene feature vector of a newly emerging emergency event after S4.1 standardized format matching processing. This vector contains normalized event type labels, influencing factors, response process nodes, spatial-temporal labels, and causal chain attributes.

[0112] An event type label matching algorithm (parameters: event type feature vector, event type matching rule set, label similarity threshold) is used to perform a preliminary comparison between the event type labels of historical case nodes in the multidimensional dynamic ontology library and the features of the current input scene, and to screen the set of historical case semantic nodes with similar event attributes.

[0113] Furthermore, by using the event type feature vector space mapping method (parameters: event category principal component matrix, category hierarchical weight, category normalized distance metric), the mathematical distance between the current event type features and the event type nodes of each historical case in the ontology is calculated, and the similarity score between event type labels is quantified.

[0114] The following formula is used to measure label similarity:

[0115]

[0116] in, This is the current event type label vector. This is a vector of historical case event type labels. For feature dimension weights, This is the label consistency discrimination function.

[0117] Furthermore, a threshold filtering algorithm is used (parameter: similarity discrimination threshold). (Based on the label saliency stratification standard), historical case nodes with similarity scores higher than the preset threshold are filtered to obtain a subset of candidate historical case semantic nodes that meet the event type requirements.

[0118] Furthermore, a multi-label consistency verification mechanism (parameters: cross-label coverage threshold, label overlap index, historical case label diversity discrimination rule) is adopted to perform a re-check on the consistency of event type label structure of candidate historical case node set, remove nodes with low consistency or insufficient label overlap, and output the final event type screening result set.

[0119] Through chain-like derivation processing of event type label matching algorithm, similarity measurement, threshold filtering and multi-label consistency verification, the standardized scene features of the current input are used to perform a high-precision preliminary retrieval of event type dimension with historical cases in the multi-dimensional dynamic ontology library. The output is a set of semantic nodes of historical cases that are highly related to the new event type. This provides the basic screening results for industry and context adaptation for the subsequent spatial-temporal label dynamic matching and causal chain feature deep retrieval, realizing the effective retrieval of historical case knowledge and semantic transfer entry point.

[0120] For example, in a scenario of a sudden "urban flooding" incident in a city in 2023, the input normalized event type features are "rainstorm waterlogging event [category C03]" and "urban flooding emergency [subcategory C03-1]", with influencing factor labels such as "poor drainage", "low-lying terrain", and "traffic node paralysis". The parameter configuration event type matching rule set includes 17 main types of emergencies and 64 sub-category sub-labels, with a label similarity threshold set to 0.88. The category principal component matrix is ​​established according to the national emergency event standard (GB / T39016-2020), and the weights are allocated according to the importance of the event. The system calculates the similarity between all event type label vectors in the semantic nodes of 1256 historical cases in the multi-dimensional dynamic ontology and the input event feature vector, and compares each category using the above formula, judging the Jaccard consistency based on weighted average. When the matching scores of "Rainstorm Flooding Event [Category C03]" with historical cases such as "Urban Flooding [Category C03-2]", "Drainage System Emergencies [Category C03-3]", and "Extreme Weather - Urban Emergency Group [Category C02]" were 0.96, 0.93, and 0.77 respectively, only the first two categories, with a discrimination threshold above 0.88, were initially selected. The label structure consistency review used an overlap index criterion of 0.78, ultimately retaining 87 historical case semantic nodes with highly consistent label structures and distributions. This provides a high-relevance event type screening set for the next step of spatial-temporal correlation matching. Actual application testing shows that the event type screening accuracy reached 98.1%, and the average scene type coverage was 96.8%, laying a high-quality semantic foundation for subsequent multi-dimensional retrieval of spatial-temporal and causal features, and enabling intelligent adaptation and retrieval of historical case knowledge for new emergencies.

[0121] S4.3: Based on the event type filtering results, the spatial-temporal label association analysis mechanism is used to dynamically match the spatial distribution labels and temporal evolution labels in the filtered historical cases, further narrowing down the candidate historical case set and outputting the event candidate set under spatial-temporal labels.

[0122] The input is a set of historical case nodes filtered by event type label matching. This set has achieved a preliminary match with the standardized scenario feature vector of the newly emerging emergency event in the event type dimension.

[0123] A spatial distribution label dynamic matching algorithm (parameters: spatial coordinates of historical cases, geographic partition code, spatial neighborhood radius) is adopted to achieve data-level comparison between spatial labels in the selected historical cases and spatial labels of new events. Based on spatial distance measurement and geographic hierarchical index, historical case nodes with spatial coupling degree lower than the preset threshold are filtered out.

[0124] Furthermore, using a temporal evolution label association algorithm (parameters: event occurrence timestamp, response node time series, temporal overlap rate threshold), dynamic time label matching is performed on historical case nodes that have passed the initial spatial label screening. The following temporal overlap index formula is used to calculate the overlap between the key time segments of the current new event and the historical cases:

[0125]

[0126] in, The time period in which the new event occurs. For key time periods of historical case events, Indicates the length of the time interval.

[0127] Furthermore, a multi-scale spatial-temporal label joint matching algorithm (parameter: spatial weight) is used. Time weight Association weighted threshold For historical case nodes that simultaneously satisfy both spatial and temporal effective matching, a joint score is calculated to determine the spatial-temporal coupling score.

[0128]

[0129] in, Let be the spatial similarity between the i-th historical case and the new event. For time overlap, and These are the weighting parameters for space and time. .

[0130] Furthermore, for joint scores exceeding a preset association weighting threshold... The historical case nodes are entered into the subsequent spatial-temporal candidate set, and their spatial and temporal labels are stored and indexed for downstream causal chain dynamic feature retrieval.

[0131] By using the aforementioned spatial-temporal label association analysis algorithm, the selected set of historical case nodes is transformed into a candidate set of events under spatial-temporal labels, which effectively improves the adaptability of historical cases and current new events in geographical and temporal dimensions, and realizes a hierarchical improvement in multi-dimensional dynamic ontology semantic indexing capabilities.

[0132] For example, in a new event scenario of a large-scale thunderstorm disaster in a certain district of a city in August 2023, 192 historical case semantic nodes that have passed the initial screening by the event type label "thunderstorm disaster" are input. A dynamic matching algorithm for spatial distribution labels is configured, along with parameters such as the radius of the spatial range of historical cases. The geographic zoning coding precision is "district-street-location", and the new event spatial label is "stadium on a certain street in a certain district". After actual filtering of spatial neighbors, the spatial similarity threshold for historical cases is set to... Initial screening narrowed down to 52 cases. A temporal evolution label association algorithm was used to associate the current event with its corresponding time period. Historical case time period Clustering based on event node time. Temporal overlap threshold. After screening, 32 cases of effective spatial-temporal overlap were obtained. Let the spatial weight parameter be... Time weight Joint threshold Ultimately, 19 historical events were included in the spatial-temporal label candidate set. This spatial-temporal candidate set of historical case nodes, output through a standardized index, effectively supported subsequent deep causal chain feature retrieval and multi-level semantic group clustering. Evaluation of this processing flow showed a spatial label selection accuracy of 98.3%, improved temporal overlap retrieval efficiency to 0.13 seconds per case, an average spatial-temporal joint coupling score of 0.92, and a case fit improvement of approximately 31% compared to the baseline without joint selection.

[0133] S4.4: For the event candidate set under the space-time label, the causal chain dynamic feature matching algorithm is used to analyze the dynamic coupling degree between the causal chain logical structure of each historical case and the current new event features, realize the deep retrieval of the causal chain semantic level, and form a set of cases related to the causal chain structure.

[0134] S4.5: Semantically hierarchically cluster the set of cases related to the causal chain structure. By using distance measurement and semantic similarity calculation in the multi-dimensional feature space, identify and generate multiple historical case semantic groups with different scene distributions and multi-level representativeness, forming a high-potential case library for subsequent AI-driven semantic transfer.

[0135] Step S5: Applying an AI-driven semantic structure transfer algorithm, semantic segmentation and component reorganization are performed on the response sequences and evolution processes in the retrieved historical case semantic groups, extracting key adaptive knowledge fragments, including resource deployment, decision nodes, and turning points. Specifically, this includes:

[0136] S5.1: Perform multi-dimensional relation extraction processing on the response sequences in the semantic groups of historical cases to identify and label the logical causal chains, time labels and spatial distribution labels between event nodes, and obtain a structured response process flow.

[0137] S5.2: Based on the structured response process flow, a multi-granularity semantic segmentation algorithm is used to segment the core event state to generate a semantic representation set of each response stage in the scene, including feature label sets such as event triggers, response nodes and turning points.

[0138] The input is structured response process flow data of historical case semantic groups output by S5.1, which includes the time labeling, spatial distribution label and causal chain attributes of event nodes.

[0139] A multi-granularity semantic segmentation algorithm (parameters: semantic granularity level configuration, process stage weight matrix, event cause category library) is adopted to decompose the overall process flow semantically based on the data hierarchy of the structured response process flow, so as to realize the segmented identification of the core event state.

[0140] Furthermore, through the event-triggered interval partitioning algorithm (parameters: time window threshold, causal node affinity parameter, spatial clustering radius setting), typical scenarios such as the occurrence of event triggers, the initiation of response nodes, and changes in turning points are automatically identified within the decomposed process flow. Based on time series and spatial location, the process flow sequence is divided into different semantic stages.

[0141] Furthermore, a semantic label extraction method (parameters: trigger semantic label set, response action label library, and transition criterion rule) is adopted to automatically identify and label event trigger nodes, key response nodes, and important transition nodes for each segment interval, generating semantic feature label sets for different stages.

[0142] Furthermore, a multi-label nested representation encoding algorithm (parameters: stage label nesting depth, feature diversity measure, node redundancy filtering function) is used to implement nested structure encoding and filtering for potentially overlapping or redundant label sets within the same stage, ensuring that the semantic representation of each response stage is accurate and does not overlap with each other.

[0143] Furthermore, a stage-based semantic sparsity and hierarchical summarization algorithm (parameters: semantic sparsity threshold, hierarchical mapping table) is adopted to compress high-frequency weakly correlated labels and retain only the core feature labels that make important contributions to subsequent transfer learning, thereby optimizing the conciseness and representativeness of semantic expression.

[0144] Through the above multi-granularity semantic segmentation process, the structured response process flow is transformed into a set of semantic representations of each response stage in the scene, including feature tag sets such as event triggers, response nodes and turning points, so as to realize the standardized segmented semantic expression of complex emergency process flow and lay the foundation for component-based knowledge reorganization and transfer.

[0145] For example, in an emergency response case of rainstorm flooding in a certain district of a city in 2023, the input structured response process flow includes a time-series process: flood warning (16:30), traffic control (16:40), emergency drainage (16:50), rescue and evacuation (17:10), and recovery assessment (18:00); the spatial label is "a certain city intersection section", and the causal chain is "sudden rainfall → urban flooding → traffic paralysis → stranded personnel → increased rescue demand". A multi-granularity semantic segmentation algorithm is adopted, with the semantic granularity level set to 3 levels (cause / response / turning point). The process stage weight matrix corresponds to a cause level of 0.4, a response level of 0.4, and a turning point level of 0.2. When dividing the event trigger interval, a time-series window threshold of 10 minutes is set, and the spatial clustering radius is 500 meters. The system automatically identifies "sudden rainfall" as the cause node, "traffic control" at 16:40 and "emergency drainage" at 16:50 as response nodes, and "rescue and evacuation" at 17:10 as the turning point node. A semantic label extraction method is employed to extract "high rainfall rate" and "poor urban drainage" as causal labels from the process flow, "traffic control" and "emergency team deployment" as response labels, and "personnel relocation" and "outbreak of rescue needs" as inflection point labels. After multi-label nesting representation encoding, the redundant label "rainfall causing road flooding" is merged into the main causal label "high rainfall rate". Stage semantic sparsity optimization, using a sparsity threshold of 0.75, reduces the number of core feature labels from the initial 12 to 5 in the optimized label set, effectively improving representativeness. The final output segmented semantic set is used for subsequent component-based knowledge reorganization, significantly improving the semantic transfer compatibility and transfer learning efficiency of the case studies.

[0146] S5.3: For each segment semantic representation set, a component-based reorganization strategy is executed. Through an ontology-context-assisted information fusion method, event elements such as resource deployment, decision nodes, and turning conditions are transformed into reusable adaptive knowledge components.

[0147] The input is the segmented semantic representation set obtained in the previous step. This set is derived from the multi-granular semantic segmentation of the semantic group response process of historical cases and already includes feature labels such as event triggers, response nodes, and turning points.

[0148] A component-based reorganization strategy (parameters: semantic representation set, component reorganization rule base, feature cluster index table) is adopted to achieve the structural extraction and standardized assembly of event elements in the segmented semantic set.

[0149] Furthermore, through an ontology context-assisted information fusion method (parameters: multi-dimensional dynamic ontology library context, spatial-temporal attribute mapping matrix, causal chain logical association weights), cross-segment semantic completion and contextual relationship enhancement are performed on resource deployment elements, decision node labels and transition conditions within the segmented semantic set, and a candidate set of components with high scenario reusability is generated.

[0150] Furthermore, an event element fusion algorithm (parameters: resource deployment matching degree threshold, decision node saliency screening rule, and turning point condition logical consistency criterion) is adopted to aggregate the event elements in the candidate set and automatically generate a componentized knowledge structure containing resource deployment blocks, key decision nodes, and typical turning point conditions, ensuring that each component has the ability to be independently migrated and reused across scenarios.

[0151] Furthermore, through the standardized coding method of adaptive knowledge components (parameters: component standard specification table, semantic unique identifier generation rules, attribute reduction mapping template), the above-mentioned aggregated componentized knowledge structure is subjected to standardized attribute marking, unique identification number and cross-ontology level tag connection, and is formally transformed into a set of adaptive knowledge components that can be retrieved and reused on demand in multiple scenarios.

[0152] Through the aforementioned algorithms and processing methods, event elements such as resource deployment, decision nodes, and turning conditions in the segmented semantic representation set are efficiently transformed into reusable and adaptive knowledge components with contextual semantic enhancement and reusability and transferability capabilities, thereby achieving modular output of knowledge structures and improved operability of semantic transfer.

[0153] For example, in the 2023 urban subway fire emergency response case, the input segmented semantic representation set contains five process segments: "alarm triggering → smoke spread → passenger evacuation → rescue deployment → resumption of traffic." Each segment contains corresponding influencing factors and event tags. A component-based reorganization strategy is adopted, setting the minimum included items of the resource deployment block to "personnel, equipment, and command vehicle," with the priority in the feature cluster index table being "rescue personnel > ventilation equipment > medical unit." Applying ontology context, it was found that the spatial tag "Station Hall B1" associated with the "rescue deployment" node is consistent in historical cases and new events, and the decision nodes "subway signal interruption" and "emergency ventilation activation" have interactive logic. The event element fusion algorithm sets the saliency threshold of the decision node to 0.8, and both "passenger evacuation decision" and "signal switching linkage" are included. After the component-based knowledge structure is aggregated, the component "B1 area rescue deployment - passenger evacuation node - ventilation activation transition" is automatically generated and encoded with a unique tag ID "MetroFire2023-Comp#011". The resulting knowledge components exhibit a 30% increase in attribute set reusability across multiple scenarios within the ontology library, and their consistency score with actual decision-making models improves to 0.95. This process enables the structured decomposition, standardized assembly, and transferable output of knowledge for complex event responses, significantly enhancing the model's adaptability to new emergency scenarios and improving intelligent scheduling efficiency.

[0154] S5.4: Apply a semantic constraint consistency verification algorithm to the generated adaptive knowledge components to ensure their high compatibility and transfer applicability under dynamic ontology reduction, and obtain a set of verified knowledge fragments.

[0155] S5.5: Semantically associate and match the validated set of knowledge fragments with the multidimensional feature vectors of existing new emergency events to select the most relevant key adaptive knowledge fragments for the scenario, providing an input basis for subsequent transfer learning.

[0156] Step S6: Based on the context information of the multi-dimensional dynamic ontology library, the aforementioned adaptive knowledge fragments are input into a graph neural network for feature transfer learning, automatically generating an adaptive state prediction model structure and parameter set that integrates the dynamic features of the current new event. Specifically, this includes:

[0157] S6.1: Perform multi-dimensional dynamic ontology context representation mapping on adaptive knowledge fragments. Based on structured knowledge such as spatial-temporal labels, event types, and influencing factors in the ontology library, generate semantically enhanced knowledge fragment embedding vectors so that subsequent graph neural network feature transfer inputs have complete ontology context.

[0158] S6.2: Based on the knowledge fragment embedding vectors from the previous step, construct a multi-dimensional feature migration graph structure that includes historical case nodes, feature label edges, and ontology semantic relationships. Utilize the ontology library topology and causal chain dynamic information to perform node association and edge weight initialization, thereby realizing cross-case multi-dimensional connection modeling between knowledge fragments.

[0159] S6.3: For the dynamic feature vector of new events, use the multi-dimensional dynamic ontology context to perform cross-domain node matching and feature alignment, embed the feature nodes of new events into the feature migration graph structure, and realize high-dimensional interactive fusion of new scenarios and historical case knowledge fragments.

[0160] S6.4: For the feature transfer graph structure containing new event nodes and historical knowledge fragments after fusion, apply graph neural network (such as graph convolutional network, graph attention network, etc.) algorithm to perform multi-layer graph embedding propagation and feature aggregation, and learn a comprehensive feature transfer expression that is compatible with the dynamics of new events.

[0161] S6.5: Based on the aforementioned fusion characteristic transfer expression, the structural parameter combination of the adaptive state prediction model is automatically generated through the global convergence layer of the graph neural network, so as to realize the embedded integration of the dynamic characteristics of new events into the model, thereby optimizing the generalization performance and adaptability of the obtained model.

[0162] Step S7: During the model inference and prediction process, dynamically monitor changes in real-time multi-source sensing data, triggering state updates and adaptive loading of relevant feature factors in the multi-dimensional dynamic ontology library, thereby achieving continuous adaptive optimization of the prediction model parameters and structure. Specifically, this includes:

[0163] S7.1: Continuously acquire real-time multi-source sensing data, and use a multi-channel data fusion algorithm to perform feature flow reduction processing on the sensing data to generate high-dimensional event dynamic feature vectors, providing standardized input for dynamic monitoring.

[0164] S7.2: Based on high-dimensional event dynamic feature vectors, feature similarity calculation and mutation detection algorithms are used to determine the temporal change trend of event dynamic feature vectors and detect the state mutation points of key feature factors in real time, providing a trigger basis for the state update of ontology feature factors.

[0165] S7.3: Map the detected key feature factor changes into a multidimensional dynamic ontology library, retrieve the affected feature factor set based on the ontology indexing mechanism, and use the ontology structure state update mechanism to dynamically assign attribute values ​​to the corresponding ontology feature nodes, thereby realizing the synchronous state update of the multidimensional dynamic ontology library.

[0166] For the changes in key feature factors identified by the mutation detection algorithm, an ontology index mapping method (parameters: ontology library classification label, factor unique identifier, time sequence index) is used to realize the dynamic semantic address location of key feature factors.

[0167] Furthermore, by utilizing the feature factor relationship table of the multidimensional dynamic ontology library, the feature factor affected domain retrieval algorithm (parameters: change marker, event category, influence relationship weight) is invoked to retrieve the set of affected feature factors associated with the current state mutation, and to determine the set of ontology structure nodes that need to be updated.

[0168] Furthermore, through the ontology structure state update mechanism (parameters: node attribute vector, change event type, and dynamic update of spatial-temporal labels), an attribute assignment algorithm is used to dynamically overwrite and complete the core structural attributes, state labels, and association fields of the affected ontology feature nodes, and to perform incremental state synchronization of node attributes.

[0169] Furthermore, if dynamic dependency coupling relationships between node attributes change, an ontology causal constraint consistency verification algorithm is adopted to automatically detect whether attribute assignment operations disrupt the topological consistency of the original ontology network. If a conflict is found, dynamic conflict resolution is performed according to ontology evolution rules to restore network integrity.

[0170] After assigning dynamic attribute values, an ontology node state change recording mechanism is adopted to persistently store the attribute differences before and after the change of all updated nodes, generating a state change log of the ontology feature factors, which provides a traceable knowledge base for subsequent dynamic ontology association activation and adaptive adjustment of model parameters.

[0171] Through the chain-processing of the above algorithm, the key feature factor mutation detection results of the previous step are transformed into attribute updates of each relevant feature node in the multidimensional dynamic ontology structure, realizing the synchronous structural evolution and real-time reflection of knowledge state of the ontology under the dynamic changes of emergency scenarios, and significantly improving the model's adaptive generalization foundation.

[0172] For example, in a sudden flood emergency command scenario, a sudden increase in river water level (such as a sudden increase in water level from 2.5m to 4.3m, with a set mutation threshold of 0.5m) is collected by a multi-source sensing terminal, triggering a mutation in the characteristic factor "river water level".

[0173] Using ontology index mapping, the change in "river water level" is mapped to ontology node ID HZ01. Further, the set of feature factors associated with it is retrieved, such as "floodgate status", "downstream warning water level", and "rescue resource allocation" nodes. It is determined that these nodes need to adjust their attributes due to the drastic change in the current main feature factor.

[0174] Applying the ontology structure state update algorithm, the "water level" attribute of node HZ01 is updated from 2.5m to 4.3m. At the same time, based on the causal chain constraint, the "opening ratio" attribute of the "floodgate status" node is automatically increased from 30% to 85%, the "downstream warning water level" node is switched from "not on alert" to "red alert", and the "number of available assault boats" attribute in the "rescue resource allocation" node is increased by 20 according to the emergency plan.

[0175] During attribute assignment, a causal constraint verification algorithm is used to check the consistency of the "flood-gate-downstream" ternary causal chain to confirm that the new attribute assignment does not cause topological breaks. If a loop anomaly occurs due to assignment conflict, the ontology evolution mechanism is used to automatically downgrade the "opening ratio" attribute to 75% to balance downstream flow.

[0176] All node attribute state changes are recorded in real time, generating state change logs such as "HZ01: Water level 2.5→4.3m" and "ZH05: Status not alert → Red alert," which are then used for subsequent model structure adaptation and event evolution analysis. Ultimately, this achieves synchronous dynamic evolution of the multi-dimensional dynamic ontology structure with the real-world emergency situation, laying a solid data and knowledge foundation for model generalization and adaptive adjustment.

[0177] S7.4: After the state of the multidimensional dynamic ontology library is updated, the ontology feature factors required for the current emergency scenario are conditionally activated based on the adaptive loading algorithm, and the latest semantic set of ontology feature factors is exported as the driving input for the structural and parameter changes of the adaptive state prediction model.

[0178] S7.5: Utilizing the updated semantic set of ontology feature factors, the structural adaptive parameter optimization algorithm is applied to the current adaptive state prediction model to reconstruct or incrementally adjust the corresponding structural parameters, weighting coefficients, and nested modules of the prediction model, thereby achieving continuous adaptive optimization of the prediction model parameters and structure.

[0179] Step S8: Online error assessment of the dynamic adaptive state prediction model is performed using real-time multi-source sensing data. If the feature transfer effect of historical cases does not meet the preset conditions for prediction accuracy, the feature transfer weights are automatically adjusted or the ontology structure fragments are reorganized. Specifically, this includes:

[0180] S8.1: Accurately compare the output of the dynamic adaptive state prediction model with real-time multi-source sensing data, and use the error judgment algorithm to calculate the residual error vector between the predicted output and the actual sensing observation value to obtain the online prediction deviation evaluation index of the state prediction model.

[0181] S8.2: Based on the online prediction deviation evaluation index, the output of the historical case feature transfer module is dynamically judged to determine whether the feature transfer effect reaches the preset threshold of prediction accuracy, so as to obtain the feature transfer reliability judgment result.

[0182] S8.3: If the reliability judgment result of feature transfer does not meet the accuracy requirements, based on the current ontology structure fragment and feature transfer weight configuration, call the adaptive weight adjustment algorithm to optimize the historical case feature transfer weight parameter group, so as to dynamically adjust the influence of key features and achieve the optimal selection of knowledge transfer path.

[0183] The output of the historical case feature transfer module is optimized using an adaptive weight adjustment algorithm (parameters: current ontology structure fragment, initial distribution of feature transfer weights, online prediction deviation).

[0184] Furthermore, by using a weighted sensitivity analysis algorithm (parameters: key feature influence coefficient, error sensitivity matrix), the correlation weight gradient between each migration feature and the current model prediction error is calculated to obtain the ranking of key feature migration influence.

[0185] Furthermore, a weighted gradient reduction and momentum update method (parameters: step size adjustment coefficient, historical momentum factor) is adopted to adaptively and dynamically adjust the feature transfer weights that are highly sensitive to errors, thereby optimizing the updated weight parameter set.

[0186] Furthermore, based on the semantic constraints of the dynamic ontology structure fragments, the feature consistency verification algorithm (parameters: ontology node constraint matrix, weight consistency threshold) is invoked to perform compliance checks on the weight-optimized feature transfer paths and identify the optimal knowledge transfer path set.

[0187] Based on the path optimization results, the feature transfer path activation algorithm (parameters: optimal weight path set, dynamic scene features) is used to dynamically activate and load the knowledge associations of the optimized transfer paths, and finally obtain the optimal knowledge transfer weight configuration that is adapted to the current emergency scenario.

[0188] Through the chain-like processing of the above algorithm, the reliability judgment result of feature transfer in the previous step is transformed into targeted optimization of transfer weight parameters and dynamic optimization of transfer path, so as to realize the adaptive adjustment of the state prediction model to historical knowledge transfer, thereby effectively improving the model's generalization ability and prediction accuracy in new emergency scenarios.

[0189] For example, in an urban fire emergency command scenario, assume that the online prediction deviation of the dynamic adaptive state prediction model for a "high-rise building fire" scenario reaches 0.18, and this scenario is transferred from historical case knowledge such as "fires in old residential areas" and "fires in commercial complexes". An adaptive weight adjustment algorithm is adopted, inputting the current ontology structure fragment and the initial distribution of the transfer weights for each feature (e.g., resource deployment weight 0.25, command node weight 0.18, evacuation condition weight 0.20, environmental impact weight 0.15, and rescue means weight 0.22). Through weight sensitivity analysis, it is found that the "evacuation condition" weight has the largest impact on the prediction error, with a weight gradient of 0.087. Using a weighted gradient reduction method, the transfer weight of "evacuation condition" is increased from 0.20 to 0.32, while the weights of other features are reduced proportionally to ensure the total weight sums to 1. A semantic consistency verification algorithm is used to confirm that the new weight combination does not violate ontology constraints (e.g., resource-personnel relationship, environment-evacuation level constraints). After passing this verification, the new transfer weight configuration is activated. Ultimately, in the weighted prediction model, the online prediction bias was reduced to 0.09. The model dynamically adapts to new fire scenarios, improves the accuracy of multi-source data prediction, and effectively optimizes emergency command and decision support capabilities.

[0190] S8.4: For migration modules that still cannot meet the accuracy requirements after weight adjustment, the structure fragment reorganization process is automatically triggered based on the ontology structure reorganization mechanism. The relevant semantic structure fragments in the multidimensional dynamic ontology library are combined and replaced to generate a new ontology structure configuration template.

[0191] S8.5: The reorganized and optimized ontology structure configuration template and its updated feature transfer weights are re-inputted into the dynamic adaptive state prediction model to perform model retraining or parameter fine-tuning, so as to form a new round of state prediction output and continuously iterate to improve the model's generalization performance and dynamic adaptability.

[0192] Step S9: The output of the adaptive state prediction model, after multi-dimensional dynamic ontology enhancement and feature transfer optimization, is input into the ontology-constrained decision simulation and confidence assessment process to generate multi-path situation prediction output, providing intelligent assessment support for emergency command decision-making. Specifically, this includes:

[0193] S9.1: Perform ontology label mapping on the multidimensional prediction feature vectors output by the adaptive state prediction model to establish semantic consistency between the prediction results and the multidimensional dynamic ontology constraints, providing standardized input features for subsequent decision simulation and confidence assessment.

[0194] S9.2: Based on the mapped predicted feature vector, the multi-path decision simulation module under ontology constraints is called. By combining ontology knowledge reasoning rules and causal chain dynamics, the prediction evolution process under each path is simulated to obtain the multi-path state evolution sequence.

[0195] S9.3: The confidence interval analysis algorithm is used for the multi-path state evolution sequence. Combined with the statistical distribution and dynamic weights of historical case groups, the probability confidence intervals and uncertainty evaluation indicators of key decision nodes under each decision path are calculated to quantify the reliability of the prediction results.

[0196] S9.4: Perform risk aggregation and multidimensional correlation verification on all decision paths within the confidence interval. By executing the ontology consistency test algorithm, identify and eliminate abnormal prediction paths with semantic conflicts or causal contradictions to ensure the ontology consistency and validity of the output path set.

[0197] S9.5: The multi-path situation prediction results, after passing ontology consistency checks and confidence interval assessments, will be output in a standardized intelligent assessment report format and synchronized in real time to the emergency command and decision-making visualization interface, providing high-reliability decision support for intelligent command and decision-making and subsequent assessment feedback.

[0198] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0199] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0200] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An emergency command decision-making and evaluation method based on a four-dimensional visualization cloud platform, specifically including: S1: Based on emergency command scenarios, collect historical emergency response case data and real-time multi-source sensing data to record event types, influencing factors, response processes, spatial-temporal labels, and dynamic characteristics of causal chains; S2: Perform heterogeneous data normalization preprocessing on the collected historical emergency response case data and real-time multi-source sensing data to unify the multi-dimensional event feature vector format; S3: Based on the normalized event feature vector, construct a multi-dimensional dynamic ontology library that includes event type, influencing factors, response process, spatial-temporal labels, and causal chain dynamics; S4: Input the scene characteristics of newly emerging emergency events into a multi-dimensional dynamic ontology library, and retrieve multiple cross-dimensional and different scene distribution historical case semantic groups based on spatial-temporal labels, event types and related causal chains; S5: Semantically segment and reorganize the response sequences and evolution processes in the retrieved historical case semantic groups, extracting key adaptive knowledge fragments; specifically including: Multidimensional relation extraction processing is performed on the response sequences in the semantic group of historical cases to identify and label the logical causal chain, time label and spatial distribution label between event nodes, and obtain a structured response process flow. The structured response process flow data covers the time label, spatial distribution label and causal chain attributes of event nodes. Based on the data hierarchy of the structured response process flow, the overall process flow is decomposed into semantic granularity, and the core event states are segmented to generate a set of semantic representations for each response stage in the scenario. For each segmented semantic representation set, a component-based reorganization strategy is executed. Through an ontology-context-assisted information fusion method, resource deployment, decision nodes, and turning point event elements are used to perform cross-segment semantic completion and contextual relationship enhancement on resource deployment elements, decision node tags, and turning point conditions within the segmented semantic set, generating a candidate set of components with high scenario reusability. The event elements in the candidate set are aggregated to automatically generate a componentized knowledge structure containing resource deployment blocks, key decision nodes, and typical turning point conditions. The aggregated componentized knowledge structure is then standardized with attribute marking, unique identification numbers, and cross-ontology hierarchical tag connections, transforming it into a reusable adaptive knowledge component. For the generated adaptive knowledge components, a semantic constraint consistency verification algorithm is applied to ensure that they have high compatibility and transfer applicability under dynamic ontology reduction, and a set of verified knowledge fragments is obtained. The validated set of knowledge fragments is semantically associated and matched with the multidimensional feature vectors of existing new emergency events in order to select the key adaptive knowledge fragments that are most relevant to the scenario. S6: Based on the context information of the multidimensional dynamic ontology library, the above adaptive knowledge fragments are input into the graph neural network for feature transfer learning, and an adaptive state prediction model structure and parameter set that integrates the dynamic features of the current new event are automatically generated.

2. The emergency command decision-making and evaluation method based on a four-dimensional visualization cloud platform according to claim 1, characterized in that, In step S3, a multi-dimensional dynamic ontology library is constructed by sorting out domain knowledge and extracting feature tags, which includes event types, influencing factors, response processes, spatial-temporal labels, and causal chain dynamics.

3. The emergency command decision-making and evaluation method based on a four-dimensional visualization cloud platform according to claim 1, characterized in that, Key adaptive knowledge fragments in step S5 include resource deployment, decision nodes, and turning points.

4. The emergency command decision-making and evaluation method based on a four-dimensional visualization cloud platform according to claim 1, characterized in that, In step S1, a structured search is performed on the historical emergency response case database. Using case label parsing and scene feature extraction algorithms, a historical case dataset containing clear event types is obtained to generate original historical case data with labeled event types.

5. The emergency command decision-making and evaluation method based on a four-dimensional visualization cloud platform according to claim 1, characterized in that, Step S2 specifically includes: Historical emergency response case data and real-time multi-source sensing data are classified by type, and unstructured text, structured tables, and time-series sensor signals are classified into specific data types respectively; For classified historical emergency response case data, the inconsistent text or symbol encoding of event type, influencing factors, response process, spatial-temporal labels and causal chain dynamic characteristics is mapped into multi-dimensional event feature labels. For real-time multi-source sensing data, feature patterns are summarized for time-series sensing signals, image monitoring streams, and spatial geographic information, and their original data features are transformed into feature labels of the same dimension as historical case data. Based on the above multidimensional event feature labels, a standardized coding strategy is implemented, and numerical normalization and principal component compression professional preprocessing algorithms are used to scale all data features and correct outliers. For the normalized multidimensional event feature vector, corrections are made to remove incomplete entries and perceived redundancy that may exist in the data.

6. The emergency command decision-making and evaluation method based on a four-dimensional visualization cloud platform according to claim 1, characterized in that, Step S4 specifically includes: The feature vectors of newly emerging emergency events are normalized and matched to ensure that they can be consistently compared with the structured knowledge of historical cases in the multidimensional dynamic ontology library. Based on the standardized scene features of the current input, a preliminary search is performed on the event type dimension of the multidimensional dynamic ontology library; Based on the results of the event type screening, a spatial-temporal label association analysis mechanism is used to dynamically match the spatial distribution labels and temporal evolution labels in the screened historical cases; For the event candidate set under spatial-temporal labels, analyze the dynamic coupling degree between the causal chain logic structure of each historical case and the characteristics of the current new event; By performing semantic hierarchical clustering on the set of cases related to the causal chain structure, and by using distance measurement and semantic similarity calculation in the multidimensional feature space, multiple semantic groups of historical cases with different scene distributions and multi-level representativeness are identified and generated.

Images