A public safety event recognition method based on small sample learning and knowledge graph

Abstract

The present invention discloses a method for identifying public safety events based on small sample learning and knowledge graphs, comprising the following steps: generating a virtual scene similar to a real event scenario by combining small sample data with a knowledge graph; dynamically activating nodes related to the event by a knowledge activation threshold, freezing and storing non-related nodes, and dynamically releasing frozen nodes; identifying the dynamic evolution of events by monitoring the fluctuations of node relationships in the knowledge graph; generating multidimensional event analysis results by mutually reinforcing a graph fission mechanism and an interactive training mechanism; generating risk path identification results by event aggregation trigger points; this scheme has the characteristics of realizing efficient learning of small samples, enhancing semantic understanding capabilities, dynamically adapting to diversified scenarios, improving event identification flexibility, and building a global risk assessment framework.

Description

Technical Field

[0001] The present invention relates to the field of public network security, and in particular to a public security incident identification method based on small sample learning and knowledge graph. Background Art

[0002] In the field of public safety, the core of emergency response is how to quickly and accurately identify potential risks and situation changes, especially in emergencies. In recent years, with the popularization of information technology and social networks, new intelligence sources such as social media data and public surveillance have provided new data sources for emergency identification. However, such data are often small samples, incomplete or unstructured, which makes traditional deep learning models that rely on large-scale data training face many challenges when applied in these scenarios. This data scarcity easily leads to overfitting problems in the model, thereby weakening its generalization ability in practical applications. Therefore, how to efficiently use small sample data to identify and analyze emergencies has become an important research direction that needs to be solved. Emergencies usually occur in an environment where data is scarce and information is chaotic, which puts higher requirements on the response speed and processing capabilities of the emergency response system. In the early stages of an emergency, the initial information provided by social media, eyewitness reports, and public monitoring systems is usually limited and fragmented, making it difficult to quickly integrate, analyze, and form effective intelligence through traditional methods. At the same time, historical emergency plans and processing records of similar events can often provide valuable knowledge support for the identification of new events. How to effectively combine these historical data and apply them to real-time intelligence analysis is one of the important motivations of this invention.

[0003] In current research and applications, emergency identification mainly relies on the following methods:

[0004] (1) Large-scale deep learning models: These models achieve deep pattern recognition and identification of data by building complex neural network structures, and are suitable for high-dimensional and complex data sets. However, such methods require a large amount of labeled data for training. When the data is insufficient, the model performance will drop significantly, and overfitting or instability problems will occur. In addition, deep learning models require high computing resources during training and operation, which is limited in actual deployment.

[0005] (2) Traditional statistical analysis and clustering models: These methods identify abnormal behaviors and events by analyzing the distribution and trends of data. For example, models based on time series analysis can detect trend changes and anomalies in events. However, statistical models often show low flexibility and adaptability when dealing with highly dynamic and complex scenarios, and are difficult to cope with rapidly changing or unexpected emergencies.

[0006] (3) Rule-based detection systems: This type of method analyzes and detects intelligence data using predefined rules and event patterns. For example, a monitoring system may trigger an alarm by identifying preset behavior patterns or abnormal situations. However, this type of system has poor adaptability to new or rapidly changing event types, and rules are difficult to update and expand quickly, resulting in a slow response of the model to new types of events.

[0007] Although the above methods can work effectively in certain specific situations, they generally have the following defects when dealing with small sample data. Dependence on large-scale data: Deep learning models are prone to overfitting in small sample scenarios and cannot be effectively generalized to practical applications. Insufficient data will seriously affect the stability and recognition accuracy of the model. Lack of dynamic scene adaptability: Traditional statistical and rule-based methods lack flexibility and find it difficult to quickly respond to and adapt to new or dynamically changing events, which restricts their application in rapidly changing emergency environments. Lack of contextual information: Existing statistical and rule-based methods rarely utilize the semantic associations of historical event data or plans, resulting in a lack of deep understanding and background support when identifying new events. High computing resource requirements: Large-scale deep learning model training requires powerful computing resources, and actual deployment may be constrained by hardware and time costs, especially in emergencies, where computing resources and time are extremely limited. Summary of the invention

[0008] To achieve the above objectives, the inventor provides a public safety event recognition method based on small sample learning and knowledge graph, comprising the following steps:

[0009] S1, through the combination of small sample data and knowledge graph, generates virtual scenes similar to real event scenes, and combines the cross-modal contrastive learning framework to reconstruct the performance of small sample data in virtual scenes;

[0010] S2, dynamically activates event-related nodes through knowledge activation thresholds, freezes and stores non-related nodes, and dynamically releases frozen nodes;

[0011] S3, identifies the dynamic evolution of events by monitoring the fluctuations of node relationships in the knowledge graph;

[0012] S4, through the graph fission mechanism, a single graph is disassembled into several sub-graph dimensions. Each sub-graph dimension focuses on a specific attribute of the event. Each sub-graph dimension reinforces each other through an interactive training mechanism, generating multi-dimensional event analysis results in a fission-reconstruction manner;

[0013] S5, through event aggregation trigger points, dynamically tracks the evolution law of emergencies and generates risk path identification results.

[0014] As a preferred embodiment of the present invention, S1 comprises the steps of:

[0015] S101 extracts event-related nodes and relationships from the knowledge graph to construct an abstract semantic model. Through the mimicry generation module, the semantic model is mapped to the semantic space of the small sample data to form a reverse enhancement process of the data to generate a virtual scene. The expression is:

[0016] ;

[0017] in, Represents event nodes and relationships in the public security knowledge graph, It is a semantic pattern extraction function that extracts semantic features from event nodes and relationships. represents small sample data, Generate a module for mimicry. The virtual scene generated by the mimicry generation module;

[0018] S102, by comparing the learning loss function, the characteristics of small sample data are dynamically optimized, and the expression is:

[0019] ;

[0020] in, The feature representation generated for the virtual scene, is the positive sample feature representation of small sample data in the virtual scene, is the feature representation of other negative samples, is the cross-modal contrastive learning loss function;

[0021] S103, through the information density difference calculation function, the key information in the virtual scene is refined, and the redundancy of low-value information is eliminated to form intelligence output with high information density. The expression is:

[0022] ;

[0023] in, The virtual scene generated by the mimic generation module. is the information density difference calculation function, is the information enhancement function, For enhanced intelligence information.

[0024] As a preferred embodiment of the present invention, S2 comprises the steps of:

[0025] S201, by dynamically setting the knowledge activation threshold, activate the nodes that are highly relevant to the current event scenario. The relevance calculation function comprehensively considers the semantic distance, connection weight and scenario factor information of the node, and the expression is:

[0026] ;

[0027] in, is a node in the knowledge graph, For Node The correlation calculation function with the current event scenario C, is the knowledge activation threshold, which is used to filter nodes that are highly relevant to the event scenario. For Node The activation status, 1 means activated, 0 means frozen;

[0028] S202, through the priority scoring function , to evaluate the potential use value of frozen nodes in future scenarios, the expression is:

[0029] ;

[0030] in, A node in a frozen state The priority score of For Node Potential value score at time t;

[0031] S203, by detecting the change in the correlation between the frozen nodes and the new scenario, the frozen information is dynamically released and the relevant nodes are activated. The sensitivity of information reactivation is controlled by the knowledge release threshold to avoid the reactivation of invalid information. The expression is:

[0032] ;

[0033] in, For new scenarios, is the change in correlation between the frozen node and the new scenario, is the knowledge release threshold, Update activation status for the node.

[0034] As a preferred embodiment of the present invention, unactivated frozen nodes are stored as frozen information. The storage of frozen information adopts a dynamic priority management mechanism, based on the principle of high priority nodes, retaining important but unactivated information while further releasing storage space.

[0035] As a preferred embodiment of the present invention, S3 comprises the steps of:

[0036] S301, calculate the change rate of node relationship through the fluctuation monitoring module to capture the hidden feature change trend, the expression is:

[0037] ;

[0038] in, is a node in the public security knowledge graph, for The relationship weight reflects the semantic or event correlation between two nodes. is the volatility of node relationships, Indicates trends over time;

[0039] S302, the monitored fluctuation signal is processed by the intelligence depth amplifier, and high-resolution virtual intelligence is generated by combining the dominant characteristics of small sample data. The expression is:

[0040] ;

[0041] Among them, W is the node relationship fluctuation matrix, which records the fluctuation information of all node relationships in the knowledge graph. It is a small sample of basic data, used as a reference for intelligence amplification. It is a deep intelligence amplifier that extracts and enhances hidden information through the graph neural network of wave signals. For the generated high-resolution virtual intelligence;

[0042] S303, by combining virtual intelligence and the change matrix of the current node relationship, the deduction and recognition module Predict is used to simulate the dynamic evolution process of the event, and based on time series analysis or dynamic graph neural network, the event changes caused by hidden features in the future are mined. The expression is:

[0043] ;

[0044] in, To identify future event scenarios, It is virtual intelligence, including hidden features and their changing trends. is the relationship change matrix at the current time t, Identify models for scenarios of future events.

[0045] As a preferred embodiment of the present invention, S4 comprises the steps of:

[0046] S401, through the graph fission function, based on the specific dimensional attributes of the event, the original single knowledge graph is decomposed to generate a multi-dimensional sub-graph, the expression is:

[0047] ;

[0048] in, is the original single knowledge graph, A collection of multidimensional subgraphs generated for fission, is the splitting criterion of the i-th dimension, is the spectrum fission function, is a multidimensional subgraph;

[0049] S402, introduce interactive reinforcement training, calculate the similarity and interaction weight between sub-graphs, and strengthen the complementary characteristics between sub-graphs by optimizing the loss function. The expression is:

[0050] ;

[0051] in, is the sub-graph interaction reinforcement loss function, is a multidimensional subgraph, is the similarity measure between subgraphs, is the interaction weight;

[0052] S403, through the graph reconstruction function, each sub-graph is merged into a multi-dimensional event analysis result according to the reconstruction criterion, and the expression is:

[0053] ;

[0054] in, is a multidimensional subgraph set, It is a graph reconstruction function that generates event analysis results by fusing multi-dimensional sub-graphs. is the reconstruction criterion, which is used to control the weight of each sub-graph contribution. Analyze the results for the generated multidimensional events.

[0055] As a preferred embodiment of the present invention, the sub-atlas includes:

[0056] Time sub-graph, used to reveal the dynamic evolution of events;

[0057] A spatial submap to show the geographical distribution of events;

[0058] Sentiment sub-graph, used to capture semantic sentiment changes.

[0059] As a preferred embodiment of the present invention, S5 comprises the steps of:

[0060] S501, through the event aggregation trigger function, the changes in the small sample data are monitored in real time, and the expression is:

[0061] ;

[0062] in, is the event scenario at the current time t, is a multidimensional subgraph set, Analyze results for multi-dimensional events; is the event aggregation trigger function, is the amount of intelligence change detected at time t;

[0063] S502, using the intelligence changes captured by the trigger and the structural characteristics of the knowledge graph through the recursive model function, a recursive intelligence chain is generated. The recursive model function expands the intelligence chain through multiple rounds of iterations to dynamically present future risk nodes and propagation paths. The expression is:

[0064] ;

[0065] in, is the recursive intelligence chain at time t+1, and Rec is the recursive model function;

[0066] S503, the risk path identification result is generated by integrating the recursive intelligence chain generated by the risk path simulation function, and the expression is:

[0067] ;

[0068] in, It is a risk path simulation function, which is used to deduce possible high-risk scenarios in the future. Identify outcomes for the generated risk pathways.

[0069] As a preferred embodiment of the present invention, the risk path identification result includes high-risk nodes, key time points and potential propagation directions.

[0070] Different from the prior art, the above technical solution has the following beneficial effects:

[0071] (1) This method combines small sample learning and knowledge graphs to achieve the recognition and analysis of emergencies. That is, by introducing knowledge graphs, the semantic relationship between historical plan data and real-time intelligence information is associated, and the understanding of event context and semantic mining capabilities are enhanced, thereby improving the generalization and reliability of small sample learning. Knowledge graphs can provide important context and background support by integrating data from different sources into a unified semantic framework, enabling efficient recognition of event patterns in small sample environments.

[0072] (2) This method, combined with the knowledge graph, can understand the hidden semantics and logical relationships in small sample data and improve the ability to identify complex scenarios. By utilizing the rich context provided by the knowledge graph, it can make up for the performance degradation caused by insufficient data and improve the generalization ability in small sample scenarios. Combining historical plan data and current intelligence information, it can identify potential risks in advance and provide real-time warnings and decision support for emergency response;

[0073] (3) This method breaks through the bottleneck of data dependence and achieves efficient learning with small samples: In view of the limitation of traditional deep learning models that rely on large-scale data training, this method innovatively introduces small sample learning and designs an event recognition method that can quickly extract key features from limited intelligence data. By optimizing the small sample learning algorithm, it still has high generalization performance in scenarios where data is scarce and unevenly distributed;

[0074] (4) This method integrates knowledge graphs to enhance semantic understanding capabilities: To make up for the a priori deficiencies of small sample data, this method proposes to deeply integrate knowledge graphs with the event recognition process, and use historical plans, event handling experience, and semantic relationships in the knowledge graph to build an architecture with contextual association capabilities. The introduction of knowledge graphs can provide semantic extension and logical association for small sample data, improve semantic understanding and reasoning capabilities, and thus achieve accurate characterization of events.

[0075] (5) This method dynamically adapts to diverse scenarios and improves the flexibility of event recognition: In view of the complex and rapidly changing characteristics of emergency situations, this method designs a dynamic adaptation mechanism that can intelligently adjust model parameters and recognition strategies in different scenarios. Combined with the dynamic update capability of the knowledge graph, this method achieves rapid adaptation to new event patterns and improves the breadth of application of the model in diverse scenarios.

[0076] (6) This method integrates multi-source intelligence and constructs a global risk assessment framework: This method integrates multi-source small sample data information such as social media, surveillance data, and eyewitness reports, and combines it with the knowledge graph of historical plans to construct a unified risk assessment framework. It innovatively proposes a method based on multimodal data fusion to achieve comprehensive analysis of multi-dimensional intelligence data, thereby improving the accuracy of event identification and risk identification as a whole. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 The following is a flow chart of the method described in the specific implementation mode. DETAILED DESCRIPTION

[0078] In order to explain the technical content, structural features, achieved objectives and effects of the technical solution in detail, the following is a detailed description in conjunction with specific embodiments and accompanying drawings.

[0079] like Figure 1 As shown, this embodiment provides a public safety event identification method based on small sample learning and knowledge graph, including the following steps:

[0080] S1, through the combination of small sample data and knowledge graph, generates virtual scenes similar to real event scenes, and combines the cross-modal contrastive learning framework to reconstruct the performance of small sample data in virtual scenes;

[0081] S2, dynamically activates event-related nodes through knowledge activation thresholds, freezes and stores non-related nodes, and dynamically releases frozen nodes;

[0082] S3, identifies the dynamic evolution of events by monitoring the fluctuations of node relationships in the knowledge graph;

[0083] S4, through the graph fission mechanism, a single graph is disassembled into several sub-graph dimensions. Each sub-graph dimension focuses on a specific attribute of the event. Each sub-graph dimension reinforces each other through an interactive training mechanism, generating multi-dimensional event analysis results in a fission-reconstruction manner;

[0084] S5, through event aggregation trigger points, dynamically tracks the evolution law of emergencies and generates risk path identification results.

[0085] In the specific implementation process of the above embodiment, step S1 includes the steps of:

[0086] S101, semantic pattern extraction and mimetic generation based on small sample data and public safety knowledge graph; that is, extracting event-related nodes and relationships from the knowledge graph to construct abstract semantic patterns. Through the mimetic generation module, these semantic patterns are mapped to the semantic space of small sample data, forming a reverse enhancement process of data and generating virtual scenes; the generated virtual scenes not only contain the characteristics of small sample data, but also absorb the potential event relationships in the knowledge graph, thereby expanding the semantic scope of the original data. This generation method can reconstruct small sample data and generate virtual event scenes with situational and panoramic features, enhancing the depth and breadth of understanding of events. The expression is:

[0087] ;

[0088] in, Represents event nodes and relationships in the public safety knowledge graph, which is the core semantic information of public safety events; It is a semantic pattern extraction function that extracts semantic features from event nodes and relations; It represents small sample data, limited intelligence from an emergency; The mimicry generation module fuses the semantic patterns of the knowledge graph with small sample data to generate virtual scenes. The virtual scene generated by the mimic generation module.

[0089] S102, after the virtual scene is generated, it is necessary to ensure its semantic consistency and information alignment with the small sample data; this step introduces a cross-modal contrastive learning framework to conduct contrastive learning training on different modalities in the virtual scene, such as text, images, and audio information, to strengthen its internal semantic connection. By designing a contrast loss function, the features of the small sample data are dynamically optimized, enabling it to be expressed more richly and accurately in the virtual scene. This process not only enhances the relevance of the virtual scene to real events, but also improves the expressive power of virtual intelligence, laying the foundation for further improvement of information density. The expression is:

[0090] ;

[0091] in, Generate feature representations for virtual scenes, such as images or text; It represents the positive sample features of small sample data in virtual scenes, and is used to enhance key intelligence features; It is the feature representation of other negative samples and is used as the optimization target for contrastive learning; It is a cross-modal contrastive learning loss function used to align virtual scenes with small sample data features.

[0092] S103, after the virtual scene is generated and optimized, its internal information density may be redundant or insufficient; this step uses the information density difference calculation function to analyze the importance and semantic contribution of each element in the virtual scene. That is, using the information enhancement function, the key information in the virtual scene is refined, and the redundancy of low-value information is eliminated to form intelligence output with high information density; this enhancement process effectively compensates for the lack of intelligence caused by data scarcity, and provides more valuable event scenario simulation for actual decision-making. The expression is:

[0093] ;

[0094] in, The virtual scene generated by the mimicry generation module; It is a function for calculating the difference in information density, which is used to evaluate the distribution and importance of information in the scene; It is an information enhancement function that generates high-information intelligence based on density difference optimization scenes; For enhanced intelligence information.

[0095] In the specific implementation process of the above embodiment, step S2 includes the steps of:

[0096] S201, in the knowledge graph, many nodes may not be directly related to the current event scenario, resulting in information redundancy and waste of computing resources. This step dynamically sets the knowledge activation threshold to activate only the nodes that are highly relevant to the current event scenario. The relevance calculation function comprehensively considers information such as the semantic distance, connection weight, and scenario factor of the node; in actual operation, by adjusting the activation threshold, the knowledge scope can be flexibly selected according to the complexity of the event, thereby improving reasoning efficiency and avoiding redundant information interference. The expression is:

[0097] ;

[0098] in, is a node in the knowledge graph; For Node A function for calculating the correlation with the current event scenario C; is the knowledge activation threshold, which is used to filter nodes that are highly relevant to the event scenario. For Node The activation status, 1 means activated, 0 means frozen;

[0099] S202: Inactivated nodes are not discarded directly, but are temporarily stored as "frozen information". , evaluate the potential use value of frozen nodes in possible future scenarios. For example, some nodes may be irrelevant in the current event, but may be reactivated during event scenario changes or reasoning. The storage of frozen information adopts a dynamic priority management mechanism, with high-priority nodes as the core, retaining important but unactivated information while further freeing up storage space. This design not only saves computing and storage resources, but also provides the possibility for real-time expansion and reasoning optimization of event scenarios, expressed as:

[0100] ;

[0101] in, A node in a frozen state Priority score; For Node The activation status of For Node The potential value score at time t is based on the historical usage frequency of the node, the strength of the connection with other nodes, etc.

[0102] S203, when the event scenario changes, for example, new data is input or context switches, the relevance of the frozen nodes may increase significantly. This step dynamically releases the frozen information and activates the relevant nodes by detecting the change in the relevance of the frozen nodes to the new scenario. The knowledge release threshold controls the accuracy of the release process to avoid the reactivation of invalid information. Through this mechanism, the knowledge graph has the ability to adaptively update in real time, and can dynamically adjust the reasoning scope according to the changes in events, thereby improving the overall efficiency and reasoning quality. The expression is:

[0103] ;

[0104] in, For new scenarios; is the change in correlation between the frozen node and the new scenario; is the knowledge release threshold, which is used to control the sensitivity of information reactivation; Update activation status for the node.

[0105] In the specific implementation process of the above embodiment, step S3 includes the steps of:

[0106] S301, the relationship weights between nodes in the public security knowledge graph will fluctuate with the event scenario and data changes; this step calculates the change rate of node relationships through the fluctuation monitoring module to capture the hidden feature change trend. The volatility can reflect the potential dynamics in the event. For example, the sudden increase or decrease in the relationship between certain key nodes and other nodes may indicate an important turning point in the event. This real-time monitoring mechanism can identify the deep features hidden in the surface explicit data and establish a preliminary map of the evolution of emergencies. The expression is:

[0107] ;

[0108] in, It is a node in the public security knowledge graph; for The relationship weight reflects the semantic or event correlation between two nodes; is the volatility of node relationships; Indicates trends over time;

[0109] S302, fluctuation monitoring alone is not enough to directly reveal the hidden information of the event. Therefore, the intelligence depth amplifier deeply processes the monitored fluctuation signal and generates high-resolution virtual intelligence by combining the explicit features of small sample data. This process can mine the potential patterns in the fluctuation signal and amplify these patterns to reveal information that is not explicitly expressed by the small sample data. The generated virtual intelligence not only fills the gaps in the original data, but also provides a more comprehensive perspective for event analysis, expressed as:

[0110] ;

[0111] Among them, W is the node relationship fluctuation matrix, which records the fluctuation information of all node relationships in the knowledge graph; It is a small sample of basic data, used as a reference for intelligence amplification; As a deep intelligence amplifier, it extracts and enhances hidden information through graph neural networks of fluctuation signals; For generating high-resolution virtual intelligence.

[0112] S303, by combining virtual intelligence and the change matrix of the current node relationship, this step uses the deduction and recognition module Predict to simulate the dynamic evolution process of the event. The recognition model can be based on time series analysis or dynamic graph neural network to explore the changes in events that may be caused by hidden features in the future. For example, when certain node relationships continue to fluctuate or form new patterns, the recognition model can generate possible event scenarios in advance to achieve forward-looking judgments on the evolution of events. This capability enables the analysis model to move from "relying on explicit data" to "revealing hidden information", expressed as:

[0113] ;

[0114] in, for identifying future event scenarios; It is virtual intelligence, including hidden features and their changing trends; is the relationship change matrix at the current time t; Identify models for scenarios of future events.

[0115] In the specific implementation process of the above embodiment, step S4 includes the steps of:

[0116] S401, in the scenario of small sample data, the traditional public security knowledge graph is often difficult to support multi-dimensional event analysis due to its single structure and low information density. To this end, this step proposes a graph fission mechanism to decompose the original single knowledge graph based on the specific attributes of the event, such as time, space, emotion, etc., to generate a multi-dimensional sub-graph. The fission function comprehensively analyzes the attributes of the nodes and edges in the original graph, and divides the subsets with common characteristics into sub-graphs of different dimensions. This fission mechanism ensures that each sub-graph can focus on the specific attributes of the event, providing a basis for subsequent in-depth analysis. The expression is:

[0117] ;

[0118] in, is the original single knowledge graph; A collection of multidimensional subgraphs generated for fission; is the splitting criterion of the i-th dimension, such as time, space, emotion, semantics, etc.; is the graph fission function, which decomposes the original graph according to a specific dimensionality criterion; It is a multidimensional sub-graph that focuses on specific dimensional attributes.

[0119] S402, the generated multi-dimensional sub-graphs do not exist independently, but influence each other through an interactive reinforcement mechanism to improve the overall semantic expression ability. This step introduces interactive reinforcement training, calculates the similarity and interaction weights between sub-graphs, and strengthens the complementary characteristics between sub-graphs by optimizing the loss function. Specifically, sub-graphs can share some node or edge information, enhance these commonalities through interactive training, and mine new features in differences. This mechanism not only improves the overall information density of the graph, but also captures the deep features of events in multiple dimensions. The expression is:

[0120] ;

[0121] in, is the sub-graph interaction reinforcement loss function; It is a multidimensional subgraph; It is the similarity measure between subgraphs, such as shared nodes, edge weight distribution, etc. is the interaction weight, which is used to measure the importance of the interaction between two sub-graphs.

[0122] S403, after interactive reinforcement, the multidimensional sub-graphs generated by fission need to be reconstructed to form the final event analysis results. This step uses the graph reconstruction function to merge each sub-graph into a comprehensive analysis result according to specific reconstruction criteria. This process combines the advantages of each sub-graph. For example, the time sub-graph can reveal the dynamic evolution of events, the spatial sub-graph can show the geographical distribution of events, and the emotional sub-graph can capture semantic emotional changes. Through the "fission-reconstruction" method, the final analysis result has information density and multi-dimensional characteristics that exceed the original small sample data, providing support for the comprehensive understanding of complex events. The expression is:

[0123] ;

[0124] in, is a multidimensional sub-graph set; It is a graph reconstruction function that generates event analysis results through multi-dimensional sub-graph fusion; is the reconstruction criterion, used to control the weight of each sub-graph contribution; The multi-dimensional event analysis results are finally generated.

[0125] In the specific implementation process of the above embodiment, step S5 includes the steps of:

[0126] S501, by integrating the above steps, graph fission, virtual intelligence generation, and multi-dimensional event analysis results, this step sets up event aggregation trigger points to monitor subtle changes in small sample data in real time. The expression is:

[0127] ;

[0128] in, is the event scenario at the current time t; It is a multi-dimensional sub-graph collection of the public security knowledge graph; Analyze results for multi-dimensional events; It is an event aggregation trigger function, which is used to detect small changes in event intelligence; is the amount of intelligence change detected at time t.

[0129] S502, using the recursive model function to capture the intelligence changes and the structural characteristics of the knowledge graph captured by the trigger, a recursive intelligence chain is generated. The recursive model expands the intelligence chain through multiple rounds of iterations, dynamically presenting possible risk nodes and propagation paths in the future, and the expression is:

[0130] ;

[0131] in, is the recursive intelligence chain at time t+1; Rec is the recursive model function used to generate future scenarios and risk paths.

[0132] S503, through the recursively generated intelligence chain, the system integrates multi-dimensional information to generate the final dynamic risk path identification results. The identification results include high-risk nodes, key time points and potential transmission directions, providing real-time support for emergency decision-making. The expression is:

[0133] ;

[0134] in, It is a risk path simulation function used to deduce possible high-risk scenarios in the future; The final risk path identification result includes high-risk nodes and evolution paths.

[0135] In order to prove the effectiveness of this method, this embodiment uses a public data set for verification experiments. Specifically, CrisisLexT26 is a public social media data set focusing on crisis event research. It contains about 60,000 tweets collected from online platforms, covering 26 types of emergencies, including natural disasters, man-made accidents, and social events. Each tweet is accompanied by detailed annotation information, including event relevance labels, such as "relevant" or "irrelevant" and specific types, such as: event description, humanitarian needs, infrastructure damage, etc., and also contains original data such as time, geographic location, and user ID. This data set has the characteristics of diversity and small samples. Most tweets are short texts, such as less than 140 characters, which are suitable for text classification, crisis event identification, and small sample learning research. The small sample characteristics and high-quality annotations of CrisisLexT26 make it widely used in the field of crisis management research, especially in event classification, correlation analysis, and intelligence extraction. It provides an ideal experimental benchmark for testing and verifying public safety event identification technology based on small sample learning and knowledge graphs. The comparative experimental results are shown in Table 1:

[0136] Table 1: Performance comparison experimental results of different methods

[0137]

[0138] Experimental results show that this method performs well on the CrisisLexT26 dataset, and all core indicators are superior to traditional machine learning and deep learning methods, including 92.5% accuracy, 91.7% precision, 93.0% recall, and 92.4% F1 score, which are all at the best level. Traditional methods such as support vector machine (SVM), random forest, and logistic regression rely on manually designed features, such as TF-IDF, which are difficult to effectively capture complex semantic relationships and context information in small sample data scenarios, resulting in limited performance, with accuracy rates of only 82.3%, 80.1%, and 78.5%. Deep learning methods such as convolutional neural networks (CNN) and long short-term memory networks (LSTM) significantly improve feature extraction and context understanding capabilities after the introduction of word embedding, with F1 scores of 84.8% and 87.6% respectively; while the Transformer-based BERT model further enhances the ability to express text semantics through pre-training and bidirectional context modeling, with an F1 score of 90.0%. Although deep learning methods have achieved good performance, they rely on large-scale data and have limited generalization capabilities in small sample scenarios. This method introduces a dynamic knowledge graph combined with small sample learning technology, and uses the semantic enhancement function and causal reasoning ability of the knowledge graph to supplement the missing background information and semantic associations in small sample data. At the same time, the feature extraction module realizes the effectiveness and superiority of the key feature method in identifying emergencies in small sample scenarios through an efficient small sample learning mechanism, providing reliable technical support for the intelligent identification and emergency management of public safety incidents.

[0139] It should be noted that, although the above embodiments have been described in this article, the patent protection scope of the present invention is not limited thereby. Therefore, based on the innovative concept of the present invention, changes and modifications made to the embodiments described herein, or equivalent structures or equivalent process changes made using the contents of the present invention specification and drawings, directly or indirectly applying the above technical solutions to other related technical fields, are all included in the patent protection scope of the present invention.

Claims

1. A public safety event recognition method based on small sample learning and knowledge graph, characterized in that: The following steps are involved: S1, through the combination of small sample data and knowledge graph, generates virtual scenes similar to real event scenes, and combines the cross-modal contrastive learning framework to reconstruct the performance of small sample data in virtual scenes; S2, dynamically activates event-related nodes through knowledge activation thresholds, freezes and stores non-related nodes, and dynamically releases frozen nodes; S3, identifies the dynamic evolution of events by monitoring the fluctuations of node relationships in the knowledge graph; S4, through the graph fission mechanism, a single graph is disassembled into several sub-graph dimensions. Each sub-graph dimension focuses on a specific attribute of the event. Each sub-graph dimension reinforces each other through an interactive training mechanism, generating multi-dimensional event analysis results in a fission-reconstruction manner; S5, through event aggregation trigger points, dynamically tracks the evolution law of emergencies and generates risk path identification results.

2. The public safety event identification method based on small sample learning and knowledge graph according to claim 1 is characterized in that: S1 includes the steps: S101 extracts event-related nodes and relationships from the knowledge graph to construct an abstract semantic model. Through the mimicry generation module, the semantic model is mapped to the semantic space of the small sample data to form a reverse enhancement process of the data to generate a virtual scene. The expression is: ； in, Represents event nodes and relationships in the public security knowledge graph, It is a semantic pattern extraction function that extracts semantic features from event nodes and relationships. Represents small sample data, Generate a module for mimicry. The virtual scene generated by the mimicry generation module; S102, by comparing the learning loss function, the characteristics of small sample data are dynamically optimized, and the expression is: ； in, The feature representation generated for the virtual scene, is the positive sample feature representation of small sample data in the virtual scene, is the feature representation of other negative samples, is the cross-modal contrastive learning loss function; S103, through the information density difference calculation function, the key information in the virtual scene is refined, and the redundancy of low-value information is eliminated to form intelligence output with high information density. The expression is: ； in, The virtual scene generated by the mimic generation module. is the information density difference calculation function, is the information enhancement function, For enhanced intelligence information.

3. The public safety event identification method based on small sample learning and knowledge graph according to claim 2 is characterized in that: S2 includes the steps: S201, by dynamically setting the knowledge activation threshold, activate the nodes that are highly relevant to the current event scenario. The relevance calculation function comprehensively considers the semantic distance, connection weight and scenario factor information of the node, and the expression is: ； in, is a node in the knowledge graph, For Node The correlation calculation function with the current event scenario C, is the knowledge activation threshold, which is used to filter nodes that are highly relevant to the event scenario. For Node The activation status, 1 means activated, 0 means frozen; S202, through the priority scoring function , to evaluate the potential use value of frozen nodes in future scenarios, the expression is: ； in, A node in a frozen state The priority score of For Node Potential value score at time t; S203, by detecting the change in the correlation between the frozen nodes and the new scenario, the frozen information is dynamically released and the relevant nodes are activated. The sensitivity of information reactivation is controlled by the knowledge release threshold to avoid the reactivation of invalid information. The expression is: ； in, For new scenarios, is the change in correlation between the frozen node and the new scenario, is the knowledge release threshold, Update activation status for the node.

4. The public safety event identification method based on small sample learning and knowledge graph according to claim 3 is characterized by: Unactivated frozen nodes are stored as frozen information. The storage of frozen information adopts a dynamic priority management mechanism, based on the principle of high-priority nodes, to retain important but unactivated information while further releasing storage space.

5. The public safety event identification method based on small sample learning and knowledge graph according to claim 3 is characterized in that: S3 includes the following steps: S301, calculate the change rate of node relationship through the fluctuation monitoring module to capture the hidden feature change trend, the expression is: ； in, is a node in the public security knowledge graph, for The relationship weight reflects the semantic or event correlation between two nodes. is the volatility of node relationships, Indicates trends over time; S302, the monitored fluctuation signal is processed by the intelligence depth amplifier, and high-resolution virtual intelligence is generated by combining the dominant characteristics of small sample data. The expression is: ； Among them, W is the node relationship fluctuation matrix, which records the fluctuation information of all node relationships in the knowledge graph. It is a small sample of basic data, used as a reference for intelligence amplification. It is a deep intelligence amplifier that extracts and enhances hidden information through the graph neural network of wave signals. For the generated high-resolution virtual intelligence; S303, by combining virtual intelligence and the change matrix of the current node relationship, the deduction and recognition module Predict is used to simulate the dynamic evolution process of the event, and based on time series analysis or dynamic graph neural network, the event changes caused by hidden features in the future are mined. The expression is: ； in, To identify future event scenarios, It is virtual intelligence, including hidden features and their changing trends. is the relationship change matrix at the current time t, Identify models for scenarios of future events.

6. The public safety event identification method based on small sample learning and knowledge graph according to claim 5 is characterized in that: S4 includes the steps: S401, through the graph fission function, based on the specific dimensional attributes of the event, the original single knowledge graph is decomposed to generate a multi-dimensional sub-graph, the expression is: ； in, is the original single knowledge graph, A collection of multidimensional subgraphs generated for fission, is the splitting criterion of the i-th dimension, is the spectrum fission function, It is a multidimensional subgraph; S402, introduce interactive reinforcement training, calculate the similarity and interaction weight between sub-graphs, and strengthen the complementary characteristics between sub-graphs by optimizing the loss function. The expression is: ； in, is the sub-graph interaction reinforcement loss function, is a multidimensional subgraph, is the similarity measure between subgraphs, is the interaction weight; S403, through the graph reconstruction function, each sub-graph is merged into a multi-dimensional event analysis result according to the reconstruction criterion, and the expression is: ； in, is a multidimensional subgraph set, It is a graph reconstruction function that generates event analysis results by fusing multi-dimensional sub-graphs. is the reconstruction criterion, which is used to control the weight of each sub-graph contribution. Analyze the results for the generated multidimensional events.

7. The method for identifying public safety events based on small sample learning and knowledge graph according to claim 6 is characterized in that: The sub-atlas includes: Time sub-graph, used to reveal the dynamic evolution of events; A spatial submap to show the geographical distribution of events; Sentiment sub-graph, used to capture semantic sentiment changes.

8. The method for identifying public safety events based on small sample learning and knowledge graph according to claim 6 is characterized in that: S5 includes the steps of: S501, through the event aggregation trigger function, the changes in the small sample data are monitored in real time, and the expression is: ； in, is the event scenario at the current time t, is a multidimensional subgraph set, Analyze results for multi-dimensional events; is the event aggregation trigger function, is the amount of intelligence change detected at time t; S502, using the intelligence changes captured by the trigger and the structural characteristics of the knowledge graph through the recursive model function, a recursive intelligence chain is generated. The recursive model function expands the intelligence chain through multiple rounds of iterations to dynamically present future risk nodes and propagation paths. The expression is: ； in, is the recursive intelligence chain at time t+1, and Rec is the recursive model function; S503, the risk path identification result is generated by integrating the recursive intelligence chain generated by the risk path simulation function, and the expression is: ； in, It is a risk path simulation function, which is used to deduce possible high-risk scenarios in the future. Identify outcomes for the generated risk pathways.

9. The method for identifying public safety events based on small sample learning and knowledge graph according to claim 8, characterized in that: The risk path identification results include high-risk nodes, key time points and potential transmission directions.

Images