5G network slicing dynamic scheduling method and system based on multimodal spatiotemporal perception and event knowledge graph

Abstract

The present invention relates to a 5G network slicing dynamic scheduling method and system based on multimodal spatiotemporal perception and event knowledge graph, and belongs to the technical field of mobile communication network resource management. The present invention constructs a dynamically evolving event knowledge graph and designs a dual-stream spatiotemporal cross network, and combines visual semantic analysis to perceive physical scene changes in real time; adopts an event-scenario dual-drive mechanism to dynamically adjust the resource prediction model, dynamically adjusts the gated recurrent neural network parameters through elastic adjustment factors, and optimizes the multi-objective resource allocation strategy based on the reinforcement learning algorithm; adopts a two-level resource scheduling mode, implements priority-guaranteed non-preemptive resource allocation in the event triggering phase, and deploys a continuously adjusted optimization strategy in the steady-state phase. This method significantly improves resource utilization efficiency and service quality in high-concurrency scenarios, is compatible with the O-RAN standard interface, and is suitable for high-reliability and low-latency communication scenarios such as smart cities and industrial Internet.

Description

Technical Field

[0001] The present invention relates to a 5G network slicing dynamic scheduling method and system based on multimodal spatiotemporal 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, which are difficult to effectively respond to sudden changes in resource demand caused by unexpected events. Traditional solutions have the following shortcomings:

[0003] Burst traffic prediction relies on historical statistical patterns, making it difficult to quantify the spatiotemporal correlation between sudden events (such as large-scale events and traffic accidents) and network traffic. Existing resource allocation strategies often use fixed threshold trigger mechanisms, which cannot adapt to dynamically changing 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 burst traffic scenarios. Therefore, there is an urgent need to build a dynamic scheduling mechanism that integrates multimodal perception. Summary of the Invention

[0004] The purpose of the present invention is to provide a 5G network slicing dynamic scheduling method and system based on multimodal spatiotemporal perception and event knowledge graph, aiming to solve the technical problems of low burst traffic prediction accuracy and rigid resource allocation in the existing technology.

[0005] To achieve the above objectives, the technical solution of the present invention is: a 5G network slicing dynamic scheduling method and system based on multimodal spatiotemporal perception and event knowledge graph, the specific steps are as follows:

[0006] Step 1: Build a dynamically evolving event knowledge graph. By integrating multi-source heterogeneous event data and modeling spatiotemporal correlations, we generate an event-traffic probability matrix and obtain network traffic characteristics predicted based on events.

[0007] Step 2: Based on the obtained network traffic characteristics, perform multimodal feature collaborative fusion to generate high-precision traffic feature vectors based on cellular-level geospatial modeling, user behavior pattern analysis, and real-time visual semantic perception;

[0008] Step 3: Based on high-precision traffic feature vectors, we use an event-scenario dual-driven mechanism to dynamically adjust the resource prediction model. We then combine this with a reinforcement learning algorithm to optimize the multi-objective resource allocation strategy, resulting in a dynamic resource adjustment strategy that balances service quality, resource efficiency, and overage penalties.

[0009] Step 4: Based on the obtained dynamic resource adjustment strategy, two-level resource scheduling is performed according to the dynamic environment status. The network resource allocation strategy is optimized by implementing elastic resource allocation with non-preemptive priority guarantee in the event trigger phase and continuous adjustment optimization strategy in the steady state phase.

[0010] The Step 1 specifically includes:

[0011] Step 1.1: Analyze the unstructured text events obtained through natural language processing models and combine them 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 graph attention network, where:

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

[0014] Edge weights are calculated using a spatiotemporal proximity function, spatial correlation is measured using a Gaussian kernel function, and inverse time weighting is used to suppress interference from expired events.

[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 Step 2 specifically includes:

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

[0018] Step 2.2: Build a dual-stream spatiotemporal cross network, including:

[0019] Spatial flow module: uses an improved spatiotemporal graph convolutional network to process historical traffic time series data and extract spatiotemporal features;

[0020] Role Stream Module: Extracts spatial pattern features of user device distribution based on a transfer learning framework;

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

[0022] Step 2.4: Deploy a lightweight visual semantic analysis model, map scene semantic information into flow correction coefficients, and optimize lighting sensitivity through an adaptive lighting optimization algorithm.

[0023] The Step 3 specifically includes:

[0024] Step 3.1: Define dynamic adjustment factors , and its calculation formula is:

[0025]

[0026] in, is the maximum value of the event association probability matrix, is the visual semantic flow correction coefficient, 、 is a configurable weight parameter;

[0027] Step 3.2: Embed the dynamic adjustment factor into the gated recurrent neural network as a dynamic adjustment parameter of the gated recurrent neural network;

[0028] Step 3.3: Construct the state space and action space of the reinforcement learning model:

[0029] The state space encodes multidimensional features of predicted traffic, resource utilization, and event risk level;

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

[0031] Step 3.4: Design a multi-objective reward function , the reward function includes a weighted combination of service quality compliance rate, resource utilization efficiency and over-allocation penalty, and the expression is:

[0032]

[0033] in, 、 、 is the adjustable weight coefficient, is the nonlinear mapping function of the service quality compliance rate, is the resource utilization efficiency indicator, Penalty for over-allocation.

[0034] The Step 4 specifically includes:

[0035] Step 4.1: When the dynamic adjustment factor When the preset threshold is exceeded, the event response mode is triggered and the system is in the event triggering stage:

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

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

[0038] Deploy multi-user MIMO beamforming technology;

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

[0040] A multi-objective optimization reinforcement learning model using proximal policy optimization generates continuous resource adjustment instructions;

[0041] Set resource lock period;

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

[0043] To achieve the above objectives, this application also provides a 5G network slicing dynamic scheduling system based on multimodal spatiotemporal perception and event knowledge graph, including:

[0044] Event perception unit: configures multi-source data collection interface and event parsing engine;

[0045] Feature fusion unit: deploys a dual-stream spatiotemporal cross network and visual semantic analysis module;

[0046] Decision execution unit: integrates elasticity prediction model and two-level resource scheduler;

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

[0048] The event sensing unit specifically includes:

[0049] Social media event capture submodule: supports dual-channel data collection via API interface and web crawler;

[0050] Physical space event analysis submodule: integrates OCR engine and spatiotemporal consistency verification algorithm;

[0051] Event feature encoding submodule: uses graph neural network to realize dynamic relationship modeling.

[0052] The innovation of the present invention is:

[0053] Dynamic event knowledge graph: This technology integrates multi-source events through unstructured text parsing and OCR technology, and builds a graph evolution model that includes spatiotemporal attenuation characteristics.

[0054] Cross-modal feature fusion: A dual-stream spatiotemporal cross network is designed to collaboratively process cellular-level geographic features and user behavior patterns, and visual semantic analysis is introduced to correct traffic prediction bias.

[0055] Elastic resource control mechanism: Dynamically adjusts prediction model parameters based on event-scenario dual-driven, and combines reinforcement learning to achieve online learning of multi-objective optimization strategies;

[0056] Intelligent two-level scheduling: A non-preemptive priority guarantee strategy is adopted in the event triggering phase, and continuous optimization of resource lock-up period protection is implemented in the steady-state phase.

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

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

[0059] Figure 2 It is a flow chart of the steps of the present invention;

[0060] Figure 3 This is a diagram of the dual-stream spatiotemporal network structure of the present invention;

[0061] Figure 4 It is an interaction diagram of the reinforcement learning model of the present invention. DETAILED DESCRIPTION

[0062] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0063] Example 1: System architecture diagram 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 is 5 minutes).

[0066] Decision-making and 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 takes effect within 500ms and supports atomic operations (all successful or all rolled back).

[0069] Monitoring feedback layer:

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

[0071] The Grafana monitoring panel sets the preset threshold alarm (CPU > 80% triggers capacity expansion).

[0072] Implementation process such as Figure 2 As shown, the specific steps are:

[0073] Step 1: Build a dynamically evolving event knowledge graph, generate an event-traffic probability matrix through multi-source heterogeneous event data fusion and spatiotemporal correlation modeling, and obtain network traffic characteristics predicted based on events.

[0074] Step 1.1: Analyze the unstructured text events obtained through natural language processing models and combine them with optical character recognition technology to extract physical space structured event information and eliminate false event noise;

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

[0076] The BERT-Base multilingual model is used 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 traffic camera video streams and identify license plates (such as "Cloud A-XXXXX") and roadside electronic screen information (such as "Road Congestion").

[0078] Video frames are timestamped using OpenCV (with an accuracy of ±50ms) and combined with GPS / Beidou coordinates (WGS-84 standard) to generate structured event records.

[0079] Step 1.2: Construct a dynamic evolution model based on graph attention network, where:

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

[0081] Edge weights are calculated using a spatiotemporal proximity function, spatial correlation is measured using a Gaussian kernel function, and inverse time weighting is used to suppress interference from expired events.

[0082] Specifically, node feature encoding is performed, and 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] Edge weight calculation, using the spatiotemporal decay model to calculate event correlation:

[0087]

[0088] in, is the edge weight, is the spatial attenuation coefficient, which is taken as 500m in the present invention; is the time attenuation coefficient, which is 0.95 in the present invention. and Represents nodes respectively and nodes The state vector of is the time difference in hours.

[0089] 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;

[0090] Specifically, the graph is updated every 15 minutes, and the historical snapshots of the last 24 hours are retained.

[0091] Step 2: Based on the obtained network traffic characteristics, perform multimodal 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] Step 2.1: Divide the target area into a cellular grid system with a preset accuracy and establish a base station topology diagram;

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

[0094]

[0095] The base station traffic is calculated as the historical average with a window size of 1 hour, the user density is measured in people / square kilometer, and the signal strength is measured in dBm.

[0096] Step 2.2: Build a dual-stream spatiotemporal cross network, including:

[0097] Spatial flow module: uses the improved space-time graph convolutional network (ASTGCN) to process historical traffic time series data and extract spatiotemporal features. The improved space-time graph convolutional network includes space-time convolution and time convolution;

[0098] Furthermore, the spatiotemporal convolution is:

[0099]

[0100] in, For the The output feature matrix of the layer, is a nonlinear activation function used to enhance the expressiveness of the model, and takes the Sigmoid function; N is the order of the Chebyshev polynomial, which controls the receptive field size of the graph convolution. A high order can capture more distant neighborhood information; is the k-th order term of the Chebyshev polynomial based on the normalized Laplace matrix. The polynomial is used to approximate the graph convolution kernel to reduce the computational complexity. L is the Laplace matrix. is the normalized Laplace matrix, which describes the topological relationship of the graph structure and is used to capture spatial dependencies. is the largest eigenvalue of the Laplacian matrix, is the identity matrix; For the The input feature matrix of the layer.

[0101] Furthermore, the temporal convolution is:

[0102]

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

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

[0105] Role Stream Module: Extracts spatial pattern features of user device distribution based on a transfer learning framework;

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

[0107] Video semantic analysis uses the YOLOv5s model (input resolution 640×640) to identify scene semantic labels in surveillance videos. An example of the mapping relationship is shown in Table 1:

[0108] Table 1 Mapping relationship example

[0109] Detection Category Flow correction factor Traffic congestion +0.3 Crowds gather +0.5 Normal traffic -0.1

[0110] Step 2.3: Dynamically aggregate the two-stream fusion features through a learnable cross-attention mechanism, where the spatial stream output is used as the query vector and the role stream generates a key-value pair matrix;

[0111] Specifically, the cross attention mechanism is as follows Figure 3 shown.

[0112] The feature fusion process is:

[0113]

[0114] in, is the fused attention feature, The query matrix represents the attention query vector of the spatial features, is the key matrix, representing the attention key vector of the behavioral feature, is a value matrix, representing the attention value vector of the behavioral feature, is a spatial feature matrix, which represents features related to geographic space or topological structure. is a behavioral feature matrix, which represents the dynamic behavior pattern of users or devices. , , are learnable parameters, specifically, is the query weight matrix, and the spatial feature matrix Mapped to the query space, is the key weight matrix, and the behavior feature matrix Map to key space; is the value weight matrix, and the behavior feature matrix Mapping to value space; for The transposed matrix of is used to calculate the similarity between the query and the key. To scale the dot product attention score matrix, we use the scaling factor Prevent gradients from vanishing; The normalized exponential function converts the attention score into a probability distribution, which is used to determine the attention weight of each spatial position on the behavioral feature;

[0115] The final fusion features are:

[0116]

[0117] in, Normalize the layers to stabilize the training process and accelerate convergence; It is a random dropout operation to prevent overfitting; It is the final fused feature matrix, which contains both spatial and behavioral information and is used for dynamic resource scheduling.

[0118] Step 2.4: Deploy a lightweight visual semantic analysis model, map scene semantic information into flow correction coefficients, and optimize lighting sensitivity through an adaptive lighting optimization algorithm.

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

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

[0121]

[0122] in, is the maximum value of the event association probability matrix, is the visual semantic flow correction coefficient, 、 To configure the weight parameters, 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 a 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 multidimensional features of predicted traffic, resource utilization, and event risk level;

[0126] Specifically, the state space for:

[0127]

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

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

[0130]

[0131] in, 、 、 is the adjustable weight coefficient, is the nonlinear mapping function of the service quality compliance rate, is the resource utilization efficiency indicator, is the over-allocation penalty term, in the present invention =0.6, =0.3, =0.1.

[0132] Step 4: Based on the obtained dynamic resource adjustment strategy, two-level resource scheduling is performed according to the dynamic environment status. The network resource allocation strategy is optimized by implementing elastic resource allocation with non-preemptive priority guarantee in the event trigger phase and continuous adjustment optimization strategy in the steady state phase.

[0133] Step 4.1: When the dynamic adjustment factor When the preset threshold is exceeded, the event response mode is triggered and the system is in the event triggering stage;

[0134] Specifically, reinforcement learning Figure 4 As shown, the preset threshold of the present invention is 0.7, that is, the trigger condition is the dynamic adjustment factor >0.7;

[0135] The specific execution strategies are:

[0136] Resource reservation: Allocate 20% of bandwidth and 15% of computing cores from the public resource pool to dedicated slices.

[0137] Priority Scheduling:

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

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

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

[0141] Massive MIMO optimization: Activates 32-antenna beamforming and adjusts downtilt angles by ±5° to enhance coverage in the target area.

[0142] Step 4.2: When the system is in the steady state, the trigger condition is the dynamic adjustment factor ≤0.7;

[0143] The specific execution strategies are:

[0144] Resource lock period: After each adjustment, resources are locked for 5 minutes to prevent business interruption caused by frequent switching.

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

[0146] Logging: records operation timestamps, adjustment parameters, and execution results, and stores them in the Elasticsearch cluster.

[0147] The above describes the specific embodiments of the present invention in detail with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A 5G network slice dynamic scheduling method based on multimodal spatiotemporal perception and event knowledge graph, characterized by: The following steps are involved: Step 1: Build a dynamically evolving event knowledge graph. By integrating multi-source heterogeneous event data and modeling spatiotemporal correlations, we generate an event-traffic probability matrix and obtain network traffic characteristics predicted based on events. Step 2: Based on the obtained network traffic characteristics, perform multimodal feature collaborative fusion to generate high-precision traffic feature vectors based on cellular-level geospatial modeling, user behavior pattern analysis, and real-time visual semantic perception; Step 3: Based on high-precision traffic feature vectors, we use an event-scenario dual-driven mechanism to dynamically adjust the resource prediction model. We then combine this with a reinforcement learning algorithm to optimize the multi-objective resource allocation strategy, resulting in a dynamic resource adjustment strategy that balances service quality, resource efficiency, and overage penalties. Step 4: Based on the obtained dynamic resource adjustment strategy, two-level resource scheduling is performed according to the dynamic environment status. The network resource allocation strategy is optimized by implementing elastic resource allocation with non-preemptive priority guarantee in the event-triggered phase and continuous adjustment optimization strategy in the steady-state phase. The Step 3 specifically includes: Step 3.1: Define dynamic adjustment factors , and its calculation formula is: ； in, is the maximum value of the event association probability matrix, is the visual semantic flow correction coefficient, 、 is a configurable weight parameter; Step 3.2: Embed the dynamic adjustment factor into the gated recurrent neural network as a dynamic adjustment parameter of the gated recurrent neural network; Step 3.3: Construct the state space and action space of the reinforcement learning model: The state space encodes multidimensional features of predicted traffic, resource utilization, and event risk level; The action space is defined as the continuous adjustment of bandwidth, computing cores and cache; Step 3.4: Design a multi-objective reward function , the reward function includes a weighted combination of service quality compliance rate, resource utilization efficiency and over-allocation penalty, and the expression is: ； in, 、 、 is the adjustable weight coefficient, is the nonlinear mapping function of the service quality compliance rate, is the resource utilization efficiency indicator, Penalty for over-allocation.

2. The method according to claim 1, characterized in that The Step 1 specifically includes: Step 1.1: Analyze the unstructured text events obtained through natural language processing models and combine them with optical character recognition technology to extract physical space structured event information and eliminate false event noise; Step 1.2: Construct a dynamic evolution model based on graph attention network, where: The node feature vector includes three dimensions: event propagation heat, cell-level user density, and remaining duration; Edge weights are calculated using a spatiotemporal proximity function, spatial correlation is measured using a Gaussian kernel function, and inverse time weighting is used to suppress interference from expired events. 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.

3. The method according to claim 1, characterized in that The Step 2 specifically includes: Step 2.1: Divide the target area into a cellular grid system with a preset accuracy and establish a base station topology diagram; Step 2.2: Build a dual-stream spatiotemporal cross network, including: Spatial flow module: uses an improved spatiotemporal graph convolutional network to process historical traffic time series data and extract spatiotemporal features; Role Stream Module: Extracts spatial pattern features of user device distribution based on a transfer learning framework; Step 2.3: Dynamically aggregate the two-stream fusion features through a learnable cross-attention mechanism, where the spatial stream output is used as the query vector and the role stream generates a key-value pair matrix; Step 2.4: Deploy a lightweight visual semantic analysis model, map scene semantic information into flow correction coefficients, and optimize lighting sensitivity through an adaptive lighting optimization algorithm.

4. The method according to claim 1, wherein The Step 4 specifically includes: Step 4.1: When the dynamic adjustment factor When the preset threshold is exceeded, the event response mode is triggered and the system is in the event triggering stage: Shorten the prediction time window to a first preset duration; Reserve a dynamic resource pool and enable non-preemptive scheduling policy; Deploy multi-user MIMO beamforming technology; Step 4.2: When the system is in steady state: A multi-objective optimization reinforcement learning model using proximal policy optimization generates continuous resource adjustment instructions; Set resource lock period; Implement operation log tracking and exception rollback mechanism.

5. A system for implementing the 5G network slice dynamic scheduling method based on multimodal spatiotemporal perception and event knowledge graph as claimed in claim 1, characterized in that: include: Event perception unit: configures multi-source data collection interface and event parsing engine; Feature fusion unit: deploys a dual-stream spatiotemporal cross network and visual semantic analysis module; Decision execution unit: integrates elasticity prediction model and two-level resource scheduler; System management and control unit: realizes resource status monitoring, policy version management and abnormal recovery functions.

6. The system according to claim 5, characterized in that The event sensing unit specifically includes: Social media event capture submodule: supports dual-channel data collection via API interface and web crawler; Physical space event analysis submodule: integrates OCR engine and spatiotemporal consistency verification algorithm; Event feature encoding submodule: uses graph neural network to realize dynamic relationship modeling.

Images