5G network slice dynamic scheduling method and system based on multi-modal space-time perception and event knowledge graph

Abstract

The invention relates to a 5G network slice dynamic scheduling method and system based on multi-modal space-time perception and an event knowledge graph, and belongs to the technical field of mobile communication network resource management. According to the method, the change of a physical scene is sensed in real time by constructing a dynamically evolved event knowledge graph and designing a double-flow space-time cross network in combination with visual semantic analysis; dynamically adjusting the resource prediction model by adopting an event-scene dual-drive mechanism, dynamically adjusting parameters of the gated recurrent neural network through an elastic adjustment factor, and optimizing a multi-target resource allocation strategy based on a reinforcement learning algorithm; a two-stage resource scheduling mode is adopted, non-preemptive resource allocation of priority guarantee is implemented in an event triggering stage, and an optimization strategy of continuous adjustment is deployed in a steady-state stage. According to the method, the resource utilization efficiency and the service quality in a high-concurrency scene are remarkably improved, the method is compatible with an O-RAN standard interface, and the method is suitable for high-reliability and low-delay communication scenes such as smart cities and industrial internet.

Description

Technical Field

[0001] The present invention relates to a 5G network slice dynamic scheduling method and system based on multi-modal spatio-temporal perception and event knowledge graph, belonging to the technical field of mobile communication network resource management. Background Art

[0002] Existing 5G network slice resource allocation methods mainly rely on historical traffic statistics or static service priority strategies, and it is difficult to effectively cope with sudden changes in resource requirements caused by sudden events. The traditional solutions have the following deficiencies:

[0003] Sudden traffic prediction depends on historical statistical laws, and it is difficult to quantify the spatio-temporal correlation between sudden events (such as large-scale events, traffic accidents) and network traffic; existing resource allocation strategies mostly adopt fixed threshold triggering mechanisms and cannot adapt to dynamic physical scenarios and user behavior patterns; resource prediction methods based on long short-term memory networks have high prediction errors during peak hours; static resource allocation strategies have insufficient resource utilization in sudden traffic scenarios. Therefore, it is urgent to construct a dynamic scheduling mechanism integrating multi-modal perception. Summary of the Invention

[0004] The purpose of the present invention is to provide a 5G network slice dynamic scheduling method and system based on multi-modal spatio-temporal perception and event knowledge graph, aiming to solve the technical problems of low accuracy of sudden traffic prediction and rigid resource allocation in the prior art.

[0005] To achieve the above purpose, the technical solution of the present invention is: a 5G network slice dynamic scheduling method and system based on multi-modal spatio-temporal perception and event knowledge graph, and the specific steps are as follows:

[0006] Step1: Construct a dynamically evolving event knowledge graph, generate an event-traffic probability matrix through multi-source heterogeneous event data fusion and spatio-temporal correlation modeling, and obtain network traffic characteristics predicted according to events;

[0007] Step2: Based on the obtained network traffic characteristics, perform multi-modal feature collaborative fusion, and generate a high-precision traffic feature vector based on cellular-level geospatial modeling, user behavior pattern analysis, and real-time visual semantic perception;

[0008] Step3: Based on the high-precision traffic feature vector, adopt an event-scenario dual-drive mechanism to dynamically adjust the resource prediction model, and optimize the multi-objective resource allocation strategy by combining a reinforcement learning algorithm to obtain a dynamic resource adjustment strategy that takes into account service quality, resource usage efficiency, and overage penalty;

[0009] Step 4: Based on the obtained dynamic resource adjustment strategy, perform two-level resource scheduling according to the dynamic environmental state. Implement elastic resource allocation with non-preemptive priority guarantee during the event trigger phase and an optimization strategy with continuous adjustment during the steady state phase to optimize the network resource allocation strategy.

[0010] The specific content of Step 1 includes:

[0011] Step 1.1: Parse the unstructured text events obtained through the natural language processing model, combine with optical character recognition technology to extract physical space structured event information, and eliminate false event noise.

[0012] Step 1.2: Construct a dynamic evolution model based on the graph attention network, where:

[0013] The node feature vector includes three dimensions: event propagation heat, cellular-level user density, and remaining duration.

[0014] Calculate the edge weight through the spatio-temporal proximity function, use the Gaussian kernel function to measure spatial correlation, and suppress the interference of expired events with time inverse weighting.

[0015] Step 1.3: Design an incremental graph update mechanism, perform local topology reconstruction based on a configurable time window, and establish a historical version snapshot rollback system.

[0016] The specific content of Step 2 includes:

[0017] Step 2.1: Divide the target area into a cellular grid system with a preset accuracy and establish a base station topology relationship graph.

[0018] Step 2.2: Construct a two-stream spatio-temporal cross network, including:

[0019] Spatial flow module: Use an improved spatio-temporal graph convolutional network to process historical traffic time series data and extract spatio-temporal features.

[0020] Role flow module: Extract the spatial pattern features of user equipment distribution based on the transfer learning framework.

[0021] Step 2.3: Dynamically aggregate the two-stream fusion features through a learnable cross-attention mechanism, where the spatial flow output is used as the query vector and the role flow generates the key-value pair matrix.

[0022] Step 2.4: Deploy a lightweight visual semantic analysis model, map the scene semantic information to a traffic correction coefficient, and optimize the light sensitivity through an adaptive light optimization algorithm.

[0023] The specific content of Step 3 includes:

[0024] Step3.1: Define the dynamic adjustment factor , and its calculation formula is:

[0025]

[0026] where is the maximum value of the event correlation probability matrix, is the visual semantic traffic correction coefficient, , are configurable weight parameters;

[0027] Step3.2: Embed the dynamic adjustment factor into the gated recurrent neural network as the dynamic adjustment parameter of the gated recurrent neural network;

[0028] Step3.3: Construct the state space and action space of the reinforcement learning model:

[0029] The state space encodes multi-dimensional features of predicted traffic, resource utilization rate, and event risk level;

[0030] The action space is defined as the continuous adjustment amounts of bandwidth, computing cores, and cache;

[0031] Step3.4: Design a multi-objective reward function , and the reward function includes a weighted combination of service quality compliance rate, resource utilization efficiency, and over-allocation penalty term, and the expression is:

[0032]

[0033] where , , are adjustable weight coefficients, is a non-linear mapping function of the service quality compliance rate, is the resource utilization efficiency index, is the over-allocation penalty term.

[0034] The specific steps of Step4 include:

[0035] Step4.1: When the dynamic adjustment factor exceeds the preset threshold, trigger the event response mode, and the system is in the event trigger stage:

[0036] Shorten the prediction time window to the first preset duration;

[0037] Reserve the dynamic resource pool and enable the non-preemptive scheduling policy;

[0038] Deploy the multi-user MIMO beamforming technology;

[0039] Step 4.2: When the system is in the steady state phase:

[0040] Use a multi-objective optimization reinforcement learning model optimized by proximal policy to generate continuous resource adjustment instructions;

[0041] Set a resource lock period;

[0042] Implement an operation log tracking and exception rollback mechanism.

[0043] To achieve the above object, the present application also provides a 5G network slice dynamic scheduling system based on multi-modal spatio-temporal perception and event knowledge graph, including:

[0044] Event perception unit: Configure multi-source data acquisition interfaces and event parsing engines;

[0045] Feature fusion unit: Deploy a dual-stream spatio-temporal cross network and a visual semantic analysis module;

[0046] Decision execution unit: Integrate an elastic prediction model and a two-level resource scheduler;

[0047] System control unit: Implement resource status monitoring, policy version management, and exception recovery functions.

[0048] The event perception unit specifically includes:

[0049] Social media event capture sub-module: Support dual-channel data acquisition of API interfaces and web crawlers;

[0050] Physical space event parsing sub-module: Integrate an OCR engine and a spatio-temporal consistency verification algorithm;

[0051] Event feature encoding sub-module: Use a graph neural network to implement dynamic relationship modeling.

[0052] The innovation points of the present invention are:

[0053] Dynamic event knowledge graph: Achieve multi-source event fusion through unstructured text parsing and OCR technology, and construct a graph evolution model with spatio-temporal decay characteristics;

[0054] Cross-modal feature fusion: Design a dual-stream spatio-temporal cross network to jointly process cellular-level geographical features and user behavior patterns, and introduce visual semantic analysis to correct traffic prediction biases;

[0055] Elastic resource regulation mechanism: Dynamically adjust the parameters of the prediction model based on event-scenario dual drive, and combine reinforcement learning to achieve online learning of multi-objective optimization strategies;

[0056] Intelligent two-level scheduling: Adopt a non-preemptive priority guarantee strategy in the event trigger stage, and implement continuous optimization with resource lock period protection in the steady state stage.

[0057] The beneficial effects of the present invention are as follows: The present invention significantly improves the resource scheduling efficiency in scenarios of burst traffic, with prediction accuracy and response speed superior to traditional methods. The resource fragmentation rate is reduced to a negligible level, while ensuring user privacy and service quality, providing key technical support for the intelligent evolution of 5G networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is the system architecture diagram of the present invention;

[0059] Figure 2 is the flowchart of the steps of the present invention;

[0060] Figure 3 is the two-stream spatio-temporal network structure diagram of the present invention;

[0061] Figure 4 is the interaction diagram of the reinforcement learning model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0062] The present invention will be further described below in conjunction with the drawings and specific embodiments.

[0063] Embodiment 1: The system architecture diagram is as Figure 1 shown.

[0064] Event data is transmitted to the Flink stream processing engine through the Kafka message queue (Topic: event_raw).

[0065] Traffic data is stored in the InfluxDB time series database (sampling interval: 5 minutes).

[0066] Decision execution layer:

[0067] After the scheduling instruction is generated, it is sent to the base station through the NETCONF protocol (Yang model: ietf-network-slice).

[0068] The configuration effective delay ≤ 500ms, supporting atomic operations (all successful or all rolled back).

[0069] Monitoring and feedback layer:

[0070] Prometheus collects metrics: CPU utilization rate, packet loss rate, slice SLA compliance rate.

[0071] The Grafana monitoring panel presets threshold alarms (CPU > 80% triggers capacity expansion).

[0072] The implementation process is as Figure 2 shown, and the specific steps are as follows:

[0073] Step1: Construct a dynamically evolving event knowledge graph. Through multi-source heterogeneous event data fusion and spatio-temporal correlation modeling, generate an event-traffic probability matrix, and obtain the network traffic characteristics predicted according to the events.

[0074] Step1.1: Parse the unstructured text events obtained through natural language processing models, combine with optical character recognition technology to extract structured event information in the physical space, and eliminate false event noises;

[0075] Specifically, obtain social media data. Through the Weibo open platform interface or crawler, configure filtering keywords (such as "exhibition", "traffic congestion", "concert", "accident"), and obtain event text data in real time.

[0076] Use the BERT-Base multilingual model to extract event semantic features (output dimension 768).

[0077] Video surveillance data: Deploy an image semantic recognition engine (such as Tesseract OCR 5.0) to parse the traffic camera video stream, and identify license plate numbers (such as "Yun A-XXXXX") and roadside electronic screen information (such as "road congestion situation").

[0078] Perform timestamp annotation (accuracy ±50ms) on video frames through OpenCV, and generate structured event records in combination with GPS / Beidou coordinates (WGS-84 standard).

[0079] Step1.2: Construct a dynamically evolving model based on the graph attention network, where:

[0080] The node feature vector includes three dimensions: event propagation heat, cellular-level user density, and remaining duration;

[0081] Calculate the edge weight through the spatio-temporal proximity function, use the Gaussian kernel function to measure spatial correlation, and use time inverse weighting to suppress the interference of expired events;

[0082] Specifically, perform node feature encoding. Each event node contains the following feature vector:

[0083]

[0084] Among them, Node i represents the i-th event node, and the event heat is calculated as:

[0085]

[0086] For edge weight calculation, use the spatio-temporal decay model to calculate event correlation:

[0087]

[0088] Among them, is the edge weight, is the spatial decay coefficient, and the value of the present invention is 500 m; is the time decay coefficient, and the value of the present invention is 0.95, and respectively represent the state vectors of nodes and node ; is the time difference, in hours.

[0089] Step1.3: Design an incremental graph update mechanism, perform local topology reconstruction based on a configurable time window, and establish a historical version snapshot rollback system;

[0090] Specifically, perform graph update once every 15 minutes and retain historical snapshots of the most recent 24 hours.

[0091] Step2: Based on the obtained network traffic characteristics, perform multi-modal feature collaborative fusion, and generate high-precision traffic feature vectors based on cellular-level geospatial modeling, user behavior pattern analysis, and real-time visual semantic perception.

[0092] Step2.1: Divide the target area into a cellular grid system with a preset precision and establish a base station topology relationship graph;

[0093] Specifically, divide the target area into 50 m × 50 m cellular grids (configurable range is 30 - 100 m), and establish a Voronoi diagram with the base station as the center. Each grid is associated with a feature vector as:

[0094]

[0095] Among them, the base station traffic statistically calculates the historical average with a 1-hour window size, the user density is in people per square kilometer, and the signal strength is in dBm.

[0096] Step2.2: Construct a two-stream spatio-temporal cross network, including:

[0097] Spatial flow module: Use an improved spatio-temporal graph convolutional network (ASTGCN) to process historical traffic time series data and extract spatio-temporal features. The improved spatio-temporal graph convolutional network includes spatio-temporal convolution and temporal convolution;

[0098] Furthermore, the spatio-temporal convolution is:

[0099]

[0100] Among them, is the output feature matrix of the th layer, is a non-linear activation function used to enhance the model's expressive power, and the Sigmoid function is adopted; N is the order of the Chebyshev polynomial, which controls the receptive field size of the graph convolution. A higher order can capture neighborhood information at a farther distance; is the k-th order term of the Chebyshev polynomial based on the normalized Laplacian matrix. By approximating the graph convolution kernel with a polynomial, the computational complexity is reduced. L is the Laplacian matrix, is the normalized Laplacian matrix, which describes the topological relationship of the graph structure and is used to capture spatial dependence, is the largest eigenvalue of the Laplacian matrix, is the identity matrix; is for the input feature matrix of the layer.

[0101] Furthermore, the temporal convolution is:

[0102]

[0103] where, is the hidden state at time step ; LSTM is the Long Short-Term Memory network, which captures long-term dependencies in the time series, such as the periodic changes in traffic data; is the input feature at time step t or the hidden state at the previous moment.

[0104] The time series data of the base station traffic taken every 5 minutes is used as the input of the temporal convolution.

[0105] Role flow module: Extract the spatial pattern features of the user equipment distribution based on the transfer learning framework;

[0106] Specifically, for device distribution pattern recognition, a pre-trained ResNet-18 model (ImageNet weights) is used to analyze the user equipment density heat map and adjust the last layer;

[0107] Video semantic analysis identifies the scene semantic labels in the surveillance video by deploying the YOLOv5s model (input resolution 640×640). An example of the mapping relationship is shown in Table 1:

[0108] Table 1 Example of mapping relationship

[0109] Detection category Flow correction coefficient Traffic congestion +0.3 Crowd gathering +0.5 Normal passage -0.1

[0110] Step2.3: Dynamically aggregate the dual-stream fusion features through a learnable cross-attention mechanism, where the spatial stream output serves as the query vector, and the role flow generates the key-value pair matrix;

[0111] Specifically, the cross-attention mechanism is asFigure 3 as shown

[0112] The feature fusion process is as follows:

[0113]

[0114] where is the fused attention feature is the query matrix, representing the attention query vector of the spatial feature is the key matrix, representing the attention key vector of the behavioral feature is the value matrix, representing the attention value vector of the behavioral feature is the spatial feature matrix, representing features related to geographical space or topological structure is the behavioral feature matrix, representing the dynamic behavior patterns of users or devices , , are learnable parameters. Specifically is the query (Query) weight matrix, mapping the spatial feature matrix to the query space is the key (Key) weight matrix, mapping the behavioral feature matrix to the key space; is the value (Value) weight matrix, mapping the behavioral feature matrix to the value space; is the transpose matrix of, used to calculate the similarity between the query and the key is the scaled dot-product attention score matrix, preventing gradient vanishing through the scaling factor ; is the softmax function, converting the attention scores into a probability distribution for determining the attention weights of each spatial position to the behavioral feature;

[0115] The final fused feature is:

[0116]

[0117] where is layer normalization, stabilizing the training process and accelerating convergence; is the dropout operation, preventing overfitting; is the final fused feature matrix, containing both spatial and behavioral information for dynamic resource scheduling.

[0118] Step2.4: Deploy a lightweight visual semantic analysis model, map the scene semantic information to a traffic correction coefficient, and optimize the light sensitivity through an adaptive light optimization algorithm.

[0119] Step 3: Based on the high-precision traffic feature vectors, adopt an event-scenario dual-driven mechanism to dynamically adjust the resource prediction model, and optimize the multi-objective resource allocation strategy in combination with the reinforcement learning algorithm to obtain a dynamic resource adjustment strategy that takes into account service quality, resource utilization efficiency, and overage penalties.

[0120] Step 3.1: Define the dynamic adjustment factor , and its calculation formula is:

[0121]

[0122] where is the maximum value of the event correlation probability matrix, is the visual semantic traffic correction coefficient, , are configurable weight parameters. In the present invention = 0.5, = 0.3. When = 0.8, = 0.4, ;

[0123] Step 3.2: Embed the dynamic adjustment factor into the gated recurrent neural network as the dynamic adjustment parameter of the gated recurrent neural network;

[0124] Step 3.3: Construct the state space and action space of the reinforcement learning model:

[0125] The state space encodes multi-dimensional features of predicted traffic, resource utilization rate, and event risk level;

[0126] Specifically, the state space is:

[0127]

[0128] The action space is defined as the continuous adjustment amounts of bandwidth, computing cores, and cache;

[0129] Step 3.4: Design a multi-objective reward function , and the reward function includes a weighted combination of service quality compliance rate, resource utilization efficiency, and over-allocation penalty terms, and the expression is:

[0130]

[0131] where , , are adjustable weight coefficients, is a non-linear mapping function of the service quality compliance rate, is the resource utilization efficiency index, is the over-allocation penalty term. In the present invention = 0.6, = 0.3, = 0.1.

[0132] Step4: Based on the obtained dynamic resource adjustment strategy, perform two-level resource scheduling according to the dynamic environmental state. Implement non-preemptive priority-guaranteed elastic resource allocation during the event-triggering phase and an optimization strategy of continuous adjustment during the steady-state phase to optimize the network resource allocation strategy.

[0133] Step4.1: When the dynamic adjustment factor exceeds the preset threshold, trigger the event response mode, and the system is in the event-triggering phase;

[0134] Specifically, as shown in reinforcement learning Figure 4 in the present invention, the preset threshold is taken as 0.7, that is, the triggering condition is that the dynamic adjustment factor > 0.7;

[0135] The execution policy is specifically:

[0136] Resource reservation: Allocate 20% bandwidth and 15% computing core from the common resource pool to the dedicated slice.

[0137] Priority scheduling:

[0138] URLLC service: Priority 1 (delay ≤ 5ms)

[0139] eMBB service: Priority 2 (delay ≤ 20ms)

[0140] mMTC service: Priority 3 (delay ≤ 100ms)

[0141] Massive MIMO optimization: Activate 32-antenna beamforming and adjust the downtilt angle by ±5° to enhance the coverage of the target area.

[0142] Step4.2: When the system is in the steady-state phase, the triggering condition is that the dynamic adjustment factor ≤ 0.7;

[0143] The execution policy is specifically:

[0144] Resource lock-in period: Lock the resources for 5 minutes after each adjustment to prevent service interruption caused by frequent switching.

[0145] Rollback mechanism: If the QoS compliance rate drops by more than 15% after the adjustment, automatically roll back to the previous stable configuration within 10 seconds.

[0146] Logging: Record the operation timestamp, adjusted parameters, and execution results, and store them in the Elasticsearch cluster.

[0147] The specific embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the gist of the present invention.

Claims

1. A 5G network slice dynamic scheduling method based on multi-modal spatio-temporal perception and event knowledge graph, characterized in that, It includes the following steps: Step1: Construct a dynamically evolving event knowledge graph. Through multi-source heterogeneous event data fusion and spatio-temporal correlation modeling, generate an event-traffic probability matrix, and obtain the network traffic characteristics predicted according to the events; Step2: Based on the obtained network traffic characteristics, perform multi-modal feature collaborative fusion. Based on cellular-level geospatial modeling, user behavior pattern analysis, and real-time visual semantic perception, generate high-precision traffic feature vectors; Step3: Based on the high-precision traffic feature vectors, adopt an event-scenario dual-driven mechanism to dynamically adjust the resource prediction model, and combine a reinforcement learning algorithm to optimize the multi-objective resource allocation strategy to obtain a dynamic resource adjustment strategy that takes into account service quality, resource usage efficiency, and overage penalties; Step4: Based on the obtained dynamic resource adjustment strategy, perform two-level resource scheduling according to the dynamic environmental state. Through the event-triggered phase, implement elastic resource allocation with non-preemptive priority guarantee and an optimization strategy of continuous adjustment in the steady state phase to optimize the network resource allocation strategy.

2. The method according to claim 1, characterized in that, The specific content of Step1 includes: Step1.1: Parse the unstructured text events obtained through a natural language processing model, and combine optical character recognition technology to extract physical space structured event information to eliminate false event noise; Step1.2: Construct a dynamically evolving model based on a graph attention network, where: The node feature vector includes three dimensions: event propagation heat, cellular-level user density, and remaining duration; Calculate the edge weights through a spatio-temporal proximity function, use a Gaussian kernel function to measure spatial correlation, and use time inverse weighting to suppress the interference of expired events; Step1.3: Design an incremental graph update mechanism, perform local topology reconstruction based on a configurable time window, and establish a historical version snapshot rollback system.

3. The method according to claim 1, characterized in that, The specific content of Step2 includes: Step2.1: Divide the target area into a cellular grid system with a preset accuracy, and establish a base station topology relationship graph; Step2.2: Construct a two-stream spatio-temporal cross network, including: Spatial flow module: Use an improved spatio-temporal graph convolutional network to process historical traffic time series data and extract spatio-temporal features; Role flow module: Extract spatial pattern features of user equipment distribution based on a transfer learning framework; Step2.3: Dynamically aggregate the two-stream fusion features through a learnable cross-attention mechanism, where the spatial flow output is used as the query vector, and the role flow generates a key-value pair matrix; Step2.4: Deploy a lightweight visual semantic analysis model, map the scene semantic information to a traffic correction coefficient, and optimize the light sensitivity through an adaptive light optimization algorithm.

4. The method according to claim 1, wherein The specific content of Step3 includes: Step 3.1: Define the dynamic adjustment factor , and its calculation formula is: ； Among them, is the maximum value of the event correlation probability matrix, is the visual semantic traffic correction coefficient, , are configurable weight parameters; Step3.2: Embed the dynamic adjustment factor into a gated recurrent neural network as the dynamic adjustment parameter of the gated recurrent neural network; Step3.3: Construct the state space and action space of the reinforcement learning model: The state space encodes multi-dimensional features of predicted traffic, resource utilization rate, and event risk level; The action space is defined as the continuous adjustment amounts of bandwidth, computing cores, and caches; Step3.4: Design a multi-objective reward function , and the reward function includes a weighted combination of the service quality compliance rate, resource utilization efficiency, and over-allocation penalty term, and the expression is: ； Among them, , , are adjustable weight coefficients, is a non-linear mapping function of the service quality compliance rate, is a resource utilization efficiency indicator, is an over-allocation penalty term.

5. The method according to claim 1, wherein The specific content of Step4 includes: Step4.1: When the dynamic adjustment factor exceeds the preset threshold, the event response mode is triggered, and the system is in the event trigger phase: Shorten the prediction time window to the first preset duration; Reserve a dynamic resource pool and enable a non-preemptive scheduling policy; Deploy multi-user MIMO beamforming technology; Step 4.2: When the system is in the steady state: Use a multi-objective optimization reinforcement learning model optimized by proximal policy to generate continuous resource adjustment instructions; Set a resource lock period; Implement an operation log tracking and exception rollback mechanism.

6. A system for implementing the 5G network slice dynamic scheduling method based on multi-modal spatio-temporal perception and event knowledge graph as described in claim 1, characterized in that, Including: Event perception unit: Configure multi-source data acquisition interfaces and event parsing engines; Feature fusion unit: Deploy a dual-stream spatio-temporal cross network and a visual semantic analysis module; Decision execution unit: Integrate an elastic prediction model and a two-level resource scheduler; System management and control unit: Implement resource status monitoring, policy version management, and exception recovery functions.

7. The system according to claim 6, wherein, The event perception unit specifically includes: Social media event capture sub-module: Support dual-channel data acquisition of API interfaces and web crawlers; Physical space event parsing sub-module: Integrate an OCR engine and a spatio-temporal consistency verification algorithm; Event feature encoding sub-module: Use a graph neural network to implement dynamic relationship modeling.

Images