Edge network task flow cross-layer dynamic offloading method fusing semantics and computing power prediction

Abstract

This application provides a method for dynamic cross-layer offloading of edge network task flows that integrates semantics and computing power prediction. The method includes: acquiring a real-time task flow to be processed in the edge network; extracting the inherent semantic feature vector for each task; and forming quantifiable semantic description data by analyzing task content and resource requirements. Based on a determined task execution plan, a scheduling plan containing specific node sequences and cross-protocol layer conversion instructions is generated, and the plan is distributed to edge nodes and forwarding devices to trigger task transmission along a specified path. During task transmission, end-to-end latency and node resource consumption data are continuously collected, and a heterogeneous data fusion method is used to integrate multi-source feedback to obtain real-time updates of the network state. For the real-time updated network state data, morphism mapping rules and manifold space parameters are dynamically adjusted, and abnormal pattern recognition is used to detect sudden load changes, determining the optimization direction for semantic mapping and path prediction.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for dynamic offloading of edge network task flows across layers by integrating semantics and computing power prediction. Background Technology

[0002] With the rapid development of edge networks, especially in highly dynamic scenarios such as intelligent manufacturing, connected vehicles, and smart cities, the processing requirements of real-time task flows place extremely high demands on network resource scheduling and allocation.

[0003] How to achieve precise matching of tasks and resources in complex network environments has become a pressing technical challenge.

[0004] Specifically, in scenarios involving cross-level task transitions and multi-node collaboration, the extraction and adaptation of task semantic features become a major challenge. Meanwhile, dynamic fluctuations in network load and resource competition among multiple tasks further exacerbate the complexity of the problem.

[0005] For example, in intelligent transportation systems, real-time vehicle data tasks need to frequently switch between edge nodes. If load changes cannot be accurately predicted or resource contention cannot be resolved, it may lead to delays in critical tasks, affecting system stability and efficiency.

[0006] This issue involves not only a deep understanding of the task content, but also the real-time perception and dynamic adjustment of the network status.

[0007] Existing technology: An asset data acquisition method based on the Internet of Things (IoT) is proposed in CN202411275947.3. This method optimizes node collaboration by constructing a task dependency model and adjusts resource allocation by combining data model updates and incentive mechanisms.

[0008] However, this technology mainly focuses on inter-node collaboration processes and data updates, and it is difficult to solve the complex needs of matching task semantics with physical resources in edge networks. Especially in high-load scenarios, it cannot effectively deal with the problems of load fluctuation prediction and resource competition contradictions.

[0009] Therefore, this patent focuses on the core issue of task semantics and resource matching in edge networks, aiming to solve the semantic adaptation problem in task cross-level jumps, address the challenges brought by dynamic fluctuations in network load, and mitigate the impact of resource competition conflicts among multiple tasks on system stability.

[0010] Especially in business scenarios with extremely high real-time requirements, how to ensure the efficient execution of tasks and the reasonable allocation of resources in complex network topologies and heterogeneous resource environments is a technical bottleneck that urgently needs to be overcome.

[0011] Solving this problem will directly affect the application performance and system reliability of edge networks in highly dynamic scenarios.

[0012] Existing edge network task offloading technologies only perform single-dimensional scheduling based on network load or computing resources, without combining the inherent semantic features of tasks to achieve accurate matching of business intent and physical resources. They cannot cope with the nonlinear loss of cross-protocol layer conversion and the node association characteristics of non-Euclidean space. In highly dynamic, multi-task concurrent scenarios, they are prone to problems such as path conflicts, load imbalance, and semantic distortion, making it difficult to meet the low latency and high reliability requirements of scenarios such as intelligent manufacturing, vehicle networking, and smart healthcare. Summary of the Invention

[0013] This invention provides a method for dynamic cross-layer offloading of edge network task flows that integrates semantics and computing power prediction, mainly including: The system acquires the real-time task flow generated in the edge network, extracts the inherent semantic feature vector for each task, and forms quantifiable semantic description data by analyzing the task content and resource requirements. Based on the extracted semantic feature vectors, a morphic association is constructed between the task semantic space and the physical network topology. An algebraic structure is used to characterize the nonlinear transformation in cross-protocol layer transitions, and the semantic adaptation rules of tasks in network layer jumps are determined. For possible network hierarchical jump sequences of a task, a dynamic topology model in non-Euclidean space is constructed. The spatial correlation between nodes is quantified by the Laplace operator to obtain the load fluctuation prediction results for each jump sequence. If the load fluctuation prediction result of a certain jump sequence shows that the load of the endpoint node exceeds the preset threshold, then the semantic conflict detection mechanism is combined to identify the semantic competition contradictions between multiple tasks and determine the jump path that needs to be adjusted. By performing time-series trend analysis on the adjusted jump path, we can capture the evolution of node load over time and determine the path with the lowest overall overhead and that meets the requirements for heterogeneous computing power adaptation as the task execution scheme. Based on the determined task execution plan, a scheduling plan containing specific node sequences and cross-protocol layer conversion instructions is generated, and the plan is sent to edge nodes and forwarding devices to trigger the task to be transmitted along the specified path; During task transmission, end-to-end latency and node resource consumption data are continuously collected, and heterogeneous data fusion method is used to integrate multi-source feedback to obtain real-time updates of network status. Based on real-time updated network status data, the morphological mapping rules and manifold space parameters are dynamically adjusted. Abnormal pattern recognition is used to detect sudden load changes and determine the optimization direction of semantic mapping and path prediction.

[0014] Furthermore, the step of generating a scheduling plan containing specific node sequences and cross-protocol layer conversion instructions based on the determined task execution scheme, distributing the plan to edge nodes and forwarding devices, and triggering task transmission along the specified path includes: By analyzing the task execution requirements, we can obtain the node sequence information required for task driving and determine the distribution location of each node on the specified path. Based on the node sequence information, a cross-protocol layer conversion instruction is constructed. If the conversion instruction does not match the device interaction requirements, the instruction format is adjusted according to the preset mapping rules to obtain a suitable instruction set. Using an adapted instruction set, a scheduling plan is generated for edge nodes and forwarding devices, the execution sequence of each device in the plan is obtained, and the priority order of task transmission is determined. Based on priority order, task-driven transmission resources are allocated. By monitoring the data flow status in real time during transmission, if the data flow is interrupted, the system switches to a backup path to obtain a stable transmission channel. Extract feedback data from device interactions from stable transmission channels, adjust the execution parameters of the scheduling plan based on the feedback data, and determine the transmission efficiency of the task on the specified path; The transmission efficiency evaluation results are obtained. By comparing them with a preset threshold, if the efficiency is lower than the threshold, the execution logic of the cross-protocol layer conversion instruction is optimized to obtain an improved transmission scheme. Based on the improved transmission scheme, update the task-driven configuration of edge nodes and forwarding devices to determine the final transmission status of the task along the specified path.

[0015] Furthermore, during task transmission, end-to-end latency and node resource consumption data are continuously collected, and a heterogeneous data fusion method is used to integrate multi-source feedback to obtain real-time updates of the network status, including: By collecting data during the task transmission process, we can obtain real-time information on end-to-end latency and node resources to determine the current load status of network transmission. Based on changes in load conditions, a preset threshold is used for comparison to determine whether there is any abnormal transmission delay, thus obtaining a preliminary location of the delay problem. If the transmission delay exceeds the preset threshold, a deep analysis is performed on the node resource consumption data to obtain the specific location of the resource bottleneck and determine the range of affected nodes. By integrating the resource consumption data and multi-source feedback information of the affected nodes, the support vector machine algorithm is used to classify the heterogeneous data and obtain the priority ranking of resource allocation. Based on the priority ranking results, the resource configuration scheme of each node in the network status is dynamically adjusted, the adjusted transmission path is obtained, and it is determined whether the latency optimization requirements are met. If the adjusted transmission path still does not meet expectations, historical records are extracted from the status monitoring data to analyze the fluctuation trend of network status and determine the node combination for further optimization. By collecting data in real time from the optimized node combination, and updating monitoring data on end-to-end latency and resource consumption, the latest dynamic results of the network status can be obtained.

[0016] Furthermore, the dynamic adjustment of morphic mapping rules and manifold space parameters based on real-time updated network state data, the detection of sudden load changes through abnormal pattern recognition, and the determination of optimization directions for semantic mapping and path prediction include: By collecting real-time updated network status data and using a preset threshold comparison method, abnormal situations of sudden load changes are detected, and preliminary judgment results of load changes are obtained. If the load change exceeds the preset threshold, pattern recognition processing is performed on the abnormal pattern data to analyze the specific time point and node location of the sudden change and determine the affected network area range. Based on the affected network area, extract relevant data on morphological mapping, dynamically adjust the content of the mapping rules, and obtain the adjusted mapping model structure. By adjusting the mapping model structure and combining it with the existing configuration of the manifold space, the distribution of spatial parameters is updated to obtain the optimized spatial model framework. Based on the optimized spatial model framework, the relevant data of semantic mapping are processed to generate a preliminary path prediction scheme and determine alternative directions for path planning. By comparing the alternative directions of the path prediction with the latest network status updates, the path scheme that meets the load balancing requirements is selected, and the final path planning result is obtained. Based on the final path planning results, update the network status monitoring data, record the handling process of load surges, and obtain the latest dynamic information on network operation.

[0017] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a dynamic cross-layer unloading method for edge network task flows that integrates semantics and computing power prediction. Addressing the core problem of cross-layer jump path conflicts and node load imbalances caused by semantic heterogeneity and resource competition during multi-task concurrency in edge networks, the method extracts task semantic features and constructs morphological associations between them and the physical topology. It predicts load fluctuations for each jump sequence in non-Euclidean space. When overload of the endpoint node is predicted, it identifies competition conflicts and adjusts paths based on semantic conflict detection. Furthermore, it selects the optimal execution scheme with the lowest overall cost and meeting computing power adaptation requirements through time-series trend analysis, generates and distributes a scheduling plan, and dynamically adjusts mapping rules and spatial parameters based on real-time network status data during task execution. This continuously optimizes semantic mapping and path prediction, thereby achieving efficient and stable cross-layer unloading and resource coordination of the task flow.

[0018] Compared to traditional edge network offloading solutions, this method can reduce end-to-end latency by more than 30%, improve node load balancing by 40%, and reduce cross-protocol layer semantic loss rate to below 1%. It can be adapted to heterogeneous computing scenarios such as 5G edge networks, industrial IoT, and vehicle-road collaboration, and still maintain stable operation even with a 10-fold increase in task concurrency.

[0019] like Figure 8 As shown, this invention conducted comparative tests on end-to-end latency under different concurrency levels, covering a range of 50 to 5000 concurrent tasks. The figure uses a line graph combined with area filling to illustrate the latency differences between the proposed method and traditional load scheduling methods, single semantic matching methods, and static path allocation methods. Experimental results show that the proposed method maintains the lowest latency across all concurrency levels, and its latency growth curve is the flattest. With 200 concurrent tasks, the proposed method reduces latency by 57% compared to traditional load scheduling methods; with 1000 concurrent tasks, it reduces latency by 75%; and with a high concurrency scenario of 5000, it reduces latency by 84%, far exceeding the expected 30% reduction. The area filling visually demonstrates the latency difference between the proposed method and the static path allocation method, which significantly widens with increasing concurrency, fully validating the significant advantages of the proposed cross-layer dynamic offloading strategy that integrates semantics and computing power prediction in high-concurrency scenarios.

[0020] like Figure 9As shown, this invention uses heatmaps to compare and contrast the CPU utilization distribution of 12 edge nodes over 10 consecutive time windows. The upper heatmap represents the load distribution of the proposed method, while the lower heatmap represents the load distribution of the traditional load scheduling method. It can be clearly observed from the figures that the heatmap of the proposed method exhibits a uniform color distribution, with the CPU utilization of each node concentrated within a reasonable range of 40% to 70% across different time windows, indicating that the load is effectively and evenly distributed among the edge nodes. In contrast, the heatmap of the traditional method shows obvious load hotspots, with some nodes achieving CPU utilization exceeding 90% during specific time periods, while other nodes have utilization below 20%, exhibiting severe load imbalance. By calculating the load standard deviation, the load standard deviation of the proposed method is reduced by more than 40% compared to the traditional method, verifying the significant effect of the proposed cross-layer dynamic offloading strategy in achieving node load balancing.

[0021] like Figure 10 As shown, this invention comprehensively evaluates the system stability under high-concurrency scenarios from four dimensions. Subfigure (a) shows the trend of semantic loss rate changing with the concurrency multiple. Throughout the process of increasing concurrency from 1x to 15x, the semantic loss rate of the method in this invention remains below the 1% threshold, while the semantic loss rate of the comparative method increases sharply after 5x concurrency. Subfigure (b) uses a stacked bar chart to show the decomposition of CPU, memory, and bandwidth resource utilization of the method in this invention under four typical application scenarios: smart factory, vehicle-road cooperation, smart healthcare, and power grid inspection. The total resource utilization in each scenario is controlled within a reasonable range. Subfigure (c) presents the joint distribution of task completion rate and response time as a scatter plot. The scatter plots of the method in this invention are concentrated in the optimal region of low latency and high completion rate. Subfigure (d) shows the evolution trend of system throughput during a 60-minute long-term operation. The throughput of the method in this invention remains stable at around 850 tasks / s, while the throughput of the traditional method continuously decreases over time, fully verifying the long-term operational stability of this invention under high-concurrency scenarios. Attached Figure Description

[0022] Figure 1 is a flowchart of the edge network task flow cross-layer dynamic unloading method that integrates semantics and computing power prediction according to the present invention.

[0023] Figure 2 is a schematic diagram of the cross-layer dynamic unloading method for edge network task flow that integrates semantics and computing power prediction according to the present invention.

[0024] Figure 3 is another schematic diagram of the edge network task flow cross-layer dynamic unloading method that integrates semantics and computing power prediction according to the present invention.

[0025] Figure 4 is a visualization of the morphic mapping relationship between the semantic space and physical topology of this invention.

[0026] Figure 5 is a visualization of the propagation of network node load fluctuations in non-Euclidean space according to the present invention.

[0027] Figure 6 is a diagram of the multi-layer physical deployment architecture of the edge network of the present invention.

[0028] Figure 7 shows the deployment diagram of the smart factory edge computing scenario of the present invention.

[0029] Figure 8 is a comparison analysis diagram of end-to-end delay of the present invention.

[0030] Figure 9 is a heat map of the node load balance of the present invention.

[0031] Figure 10 This is a comprehensive evaluation diagram of system stability under high concurrency scenarios in this invention. Detailed Implementation

[0032] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0033] Technical Terminology Definitions and Explanations Semantic feature vector: refers to a high-dimensional numerical vector extracted from the content description of the task to be processed through tools such as deep learning models, text parsing or entity recognition, which can quantitatively represent the task's business intent, key attributes and resource preferences.

[0034] Task semantic space: An abstract mathematical space composed of the semantic feature vectors of all tasks to be processed, used to analyze the business similarity, correlation and potential network resource requirements of tasks at the logical level.

[0035] Physical network topology: A graph structure model that describes the physical nodes (such as edge servers, forwarding devices, gateways, etc.) in the edge network and their physical connections and link strengths.

[0036] Phenomorphic association: refers to a structure-preserving mapping relationship established between the task semantic space and the physical network topology, aiming to accurately map the characteristic requirements of upper-layer business logic to the specific distribution of underlying physical resources.

[0037] Algebraic structure: Here it specifically refers to the mathematical structure used to characterize the logical relationships between network nodes, protocol conversion logic, or mapping rules (such as homomorphisms in algebraic group theory), and is used to handle complex nonlinear logical operations.

[0038] Cross-protocol layer conversion: refers to the encapsulation, decapsulation, format mapping, or field conversion processes performed between different network protocol stacks (such as application layer, transport layer, network layer, etc.) during data hopping across network layers.

[0039] Nonlinear transformation: refers to the non-proportional fluctuation relationship between network performance indicators (such as latency and throughput) and load during network hopping or protocol conversion, which is usually caused by protocol stack processing overhead or complex network environment.

[0040] Semantic adaptation rules: These are guidelines that instruct task data on how to adjust parameters, convert formats, or mark priorities based on their inherent semantic characteristics when the data jumps between different protocol levels, so as to ensure that the task intent does not deviate during transmission.

[0041] Non-Euclidean space: refers to a mathematical space (such as hyperbolic space or graph space) that can characterize non-linear and non-flat geometric features, and is used to model the network node relationships with complex business affinity rather than simple physical distance.

[0042] Dynamic topology model: refers to an evolutionary topology model that can reflect the real-time changes in the connection relationship of edge network nodes, real-time computing power distribution, business affinity and resource status over time.

[0043] Laplace operator: A mathematical operator applied to discrete graph structures to quantify the state differences (such as load deviation) or spatial correlation strength between a network node and its logical neighbor nodes.

[0044] Spatial correlation: describes the degree of mutual influence and synchronization trend of logically connected or geographically adjacent nodes in the edge network in terms of status indicators such as load fluctuation and transmission delay.

[0045] Jump sequence: refers to the set of path nodes that a task flow passes through in sequence through a series of network layers, protocol stacks, and physical nodes during the execution of the task flow in the edge network.

[0046] Load fluctuation: refers to the fluctuations in the utilization of computing, storage, and bandwidth resources of network nodes or transmission links within a unit of time due to the influx of tasks or changes in processing status.

[0047] Semantic conflict detection mechanism: a processing logic that identifies when multiple tasks compete for the same resource node, resulting in task blocking or reduced efficiency due to mismatches in business priorities, semantic relevance, or resource demand characteristics.

[0048] Semantic competition conflict: Different task flows may compete for the same physical resource within a specific time window due to similar resource requirement labels or logical mutual exclusion.

[0049] Jump path: refers to the complete trajectory of the task flow in the physical network, which is composed of the selected jump sequence and its corresponding transmission strategy.

[0050] Time series trend analysis: The process of using time series algorithms (such as variational mode decomposition or regression analysis) to capture the patterns, periodicity and future fluctuation trends of network node states (such as load and delay) over time.

[0051] Cross-protocol layer conversion instructions: Specific operation commands that are explicitly specified in the scheduling plan and guide network devices to perform protocol layer conversions, including format conversion, field mapping, and resource allocation parameters.

[0052] Phenomenological mapping rules: A dynamically adjusted logical criterion used to correct the mapping strength from semantic space to physical topology in real-time network conditions in order to achieve optimal resource matching.

[0053] Manifold space parameters: Geometric parameters that describe the local low-dimensional topological features of network states in high-dimensional space, and are used to reflect the underlying evolutionary laws of complex network dynamics.

[0054] Semantic mapping: The decision-making process of associating and matching the semantic features extracted from a task with specific network jump paths, resource allocation strategies, or protocol conversion parameters.

[0055] Sudden load change: refers to a sudden change in resource usage at an edge network node within a unit of time, where the CPU utilization, memory usage, or bandwidth usage increases by more than 50% compared to the previous statistical period, and the change lasts for more than 1 second.

[0056] Heterogeneous data fusion refers to the process of aligning, normalizing, and splicing features of latency data, load data, and semantic feature data from edge nodes, forwarding devices, and terminal sensors according to a unified timestamp and node ID.

[0057] Manifold space: refers to a high-dimensional topological space composed of edge network node states, task semantic features, and link parameters, used to characterize the continuous evolution relationship between network states and task mapping.

[0058] Example 1: High-precision collaboration and safe production monitoring scenario in a smart factory In this embodiment, the method provided in this application is applied to a smart factory that has deployed 5G edge computing nodes. The factory operates a large number of mobile robots, automated assembly lines, and high-definition security monitoring cameras. The system first acquires the task flow generated in the edge network in real time. Here, the task flow refers to a series of logically related data packet sequences emitted by sensors or terminal devices. For each task, the system extracts its semantic feature vector. This vector is generated by using a pre-trained deep learning model to transform the task's business intent (such as "detecting whether workers are wearing safety helmets" or "synchronous dual-camera welding angle") into numerical features in a high-dimensional space. By further analyzing the task's content complexity and required computing resources (such as computing power and bandwidth), quantifiable semantic description data is formed. During this process, the system also performs real-time detection through timestamp comparison to ensure that outdated instruction flows do not consume valuable computing resources.

[0059] Based on the extracted feature vectors, the system constructs morphic relationships between the task semantic space (a virtual mathematical set describing the logical relationships and priority distribution of tasks) and the factory's physical network topology (an entity connection structure composed of switches, routers, and edge servers). A morphic relationship is a mathematical mapping that preserves the structural properties before and after the mapping, ensuring that logical urgency is accurately mapped to low latency in the physical link. To achieve this, the system uses algebraic structures (such as homomorphic mappings in algebraic group theory) to characterize the nonlinear transformations that occur when data undergoes cross-protocol layer conversions. Cross-protocol layer conversions refer to the process of data encapsulation from application-layer industrial protocols (such as OPCUA) to transport-layer (TCP / UDP) and network-layer (IP) encapsulation. Changes in the physical environment can cause unpredictable latency fluctuations, i.e., nonlinear transformations, in this conversion. Therefore, the system uses kernel methods (a mathematical technique for mapping low-dimensional inseparable data to high-dimensional linearly separable data) to characterize this transformation, thereby determining the semantic adaptation rules for task transitions.

[0060] For possible jump sequences (the logical order of data packets moving between network nodes), the system constructs a dynamic topology model in non-Euclidean space. Non-Euclidean space (such as Riemannian geometric space) can better describe the non-equidistant and non-linear logical relationships between nodes. The system uses the Laplace operator (a second-order differential operator used to measure the difference between a node and its neighbors in a graph structure) to quantify the spatial correlation between nodes. Through calculation, the system can predict the load fluctuations that each jump sequence may generate when the assembly line enters full-load mode. If the prediction shows that the load of the endpoint node will exceed a preset warning threshold, the system activates a semantic conflict detection mechanism. This mechanism can identify semantic competition conflicts between the "high-definition video backhaul" task and the "core control signaling" task. Based on this, the system determines the jump path that needs to be adjusted and diverts non-critical traffic.

[0061] Subsequently, the system performs time-series trend analysis on the adjusted paths (using regression models to predict load evolution patterns over time) to determine the path with the lowest overall overhead that meets the requirements for adapting to heterogeneous computing power (such as the computing power provided by different hardware architectures like FPGAs and GPUs) as the execution plan. Based on this plan, a scheduling plan containing specific node sequences and cross-protocol layer conversion instructions is generated and issued. During task transmission, the system continuously collects end-to-end latency (the total time for data to travel from the sender to the receiver) and resource consumption data through probes deployed on each node. The system uses a heterogeneous data fusion method to integrate multi-source feedback to obtain real-time updates of the network status. For these updates, the system dynamically adjusts the morphological mapping rules and manifold space parameters (a manifold is a topological space with local Euclidean space properties, whose parameters determine the form of space compression and expansion). Finally, abnormal pattern recognition detects load bursts and determines the optimization direction of semantic mapping (the process of mapping task intent to network paths), ensuring production safety.

[0062] like Figure 7 As shown, in a typical edge computing application scenario in a smart factory, an assembly line 201 is deployed in the workshop, with an industrial robot arm 202 installed above it for automated assembly operations. Multiple surveillance cameras 203 are installed on the workshop ceiling, providing comprehensive video surveillance coverage of the production area. An edge server rack 204 is located on one side of the workshop, providing nearby computing power support for various computing tasks within the workshop. A workshop gateway 205 is installed next to the power distribution cabinet, responsible for aggregating data from various terminal devices within the workshop and forwarding it to the edge server. An inspection robot 206 moves along a preset path to inspect the workshop floor, collecting environmental data and equipment status information in real time. Operators 207 perform manual operations in the operating area next to the production line. The wireless coverage area of ​​the workshop gateway 205 is marked by a dashed circle, covering the main operating areas of the workshop. The task data transmission path is as follows: the surveillance cameras 203 and production line sensors wirelessly transmit data to the workshop gateway 205, which then forwards the data to the edge server rack 204 via a wired connection for processing, forming a complete task offloading path.

[0063] Example 2: Smart City Autonomous Driving Vehicle-to-Infrastructure (V2X) Scenario In this embodiment, the method provided in this application is applied to a traffic management system at a busy urban intersection. The edge computing unit acquires task flows generated in real time by autonomous vehicles, roadside radar, and streetlight cameras. For tasks such as "collision warning" or "blind spot compensation," the system extracts their semantic feature vectors. These vectors capture the urgency and spatial orientation information of the task. The system analyzes the task content to form quantifiable semantic description data. To ensure the rationality of task allocation, the system also performs resource analysis to determine whether the resource consumption of the current task will cause delays in other safety tasks. Tasks such as "obstacle avoidance warning" are marked as high-priority tasks, and a priority ranking tool is used to generate a processing order list.

[0064] Subsequently, the system constructs morphic associations between the task semantic space and the physical network topology composed of multiple base stations and edge servers. Considering the frequent base station handovers (i.e., dynamic changes in the jump sequence) that occur when vehicles move at high speeds, the system employs the homomorphic mapping method in algebraic structures to characterize the nonlinear transformation of the vehicle-to-everything (V2X) protocol during cross-protocol layer conversions. Through defined semantic adaptation rules, the system ensures that the "safety-critical" attribute of the warning information is not diluted during hierarchical jumps from the vehicle unit to the roadside unit and then to the cloud. For complex intersection environments, the system constructs a dynamic topology model based on non-Euclidean space and uses the Laplace operator to calculate the spatial correlation between nodes in different road segments. By analyzing the correlation strength matrix, the system can predict load fluctuations caused by sudden traffic convergence.

[0065] If the prediction results show that the load of the core switching node exceeds the threshold, the system, combined with a semantic conflict detection mechanism, identifies semantic competition conflicts between a large number of in-vehicle video download tasks and intersection obstacle avoidance tasks. Based on this, the system determines the jump path that needs to be adjusted, switching entertainment data streams to the backup small base station path. Next, the system performs time-series trend analysis on the path, capturing the temporal patterns of load evolution during peak traffic hours, and determines the solution with the lowest overall overhead and suitable for the heterogeneous computing power on the vehicle side (such as dedicated AI acceleration chips) as the task execution solution.

[0066] The system then generates a scheduling plan containing node sequences and cross-protocol layer conversion instructions. These instructions specify the field mappings for data packets when converting between different communication standards (such as C-V2X and Ethernet). During task transmission, the system utilizes heterogeneous data fusion methods to classify and process feedback information from devices from different vendors. If abnormal transmission delays are detected, the system performs in-depth analysis of the resource consumption of affected nodes to identify resource bottlenecks (such as a queuing buffer overflow at a gateway). Based on these real-time dynamics, the system adjusts the morphic mapping rules and the manifold space parameters describing the network's logical geometry. By recognizing abnormal patterns in traffic bursts, the system determines the direction for semantic mapping optimization, guiding subsequent path predictions to avoid congestion points.

[0067] Example 3: Smart Energy Transmission Line Fitting Defect Identification Scenario In this embodiment, the method is applied to a power line inspection edge network to process task flows of images of transmission line fittings captured by drones. The system first acquires the task flow and extracts semantic feature vectors for either "fitting corrosion identification" or "pin missing detection." These vectors transform the severity of the defects and the required detection accuracy into numerical features. The system then combines computing power requirements to form quantified semantic description data and determines the processing order of the task flow based on the real-time requirements of the inspection.

[0068] The system constructs morphological associations between the task semantic space and the physical network topology composed of tower-based edge nodes. A specific algebraic structure is used to characterize the nonlinear transformations of the power grid during cross-protocol layer conversions (such as from encrypted tunneling protocols to internal switching protocols). Through defined semantic adaptation rules, the system ensures that the subtle semantic features of high-resolution images are not lost in compression algorithms when traversing multiple nodes. For the dynamic connections generated by UAVs flying between different towers, the system constructs a dynamic topology model in non-Euclidean space. The Laplacian operator is used to quantify the spatial correlation between nodes on each tower side, thereby predicting load fluctuations caused by large-scale concurrent image transmission.

[0069] When the forecast indicates that the backhaul center node is overloaded, the system identifies a semantic conflict between the "environmental meteorological data monitoring" task and the "fitting defect image recognition" task through a semantic conflict detection mechanism. The system identifies the jump path that needs to be adjusted, either delaying the transmission of meteorological data packets or sending them via a narrowband path. Subsequently, the system performs time-series trend analysis on the path, capturing the evolution pattern of the inspection workload caused by changes in lighting or weather, and determines the execution scheme with the lowest overall overhead that meets the adaptation requirements of heterogeneous computing power at the edge (such as computing modules specifically for image recognition).

[0070] The system generates a scheduling plan containing precise cross-protocol layer conversion instructions. During transmission, the system monitors stable transmission channels and extracts feedback data from device interactions. If transmission efficiency is found to be lower than expected, the system optimizes the execution logic of the conversion instructions. Based on real-time network updates, the system dynamically adjusts morphological mapping rules and manifold space parameters. By identifying abnormal patterns to detect load bursts caused by channel interference, the system ultimately determines the optimization direction for semantic mapping and path prediction, ensuring the timeliness and accuracy of power defect identification.

[0071] Example 4: Smart Healthcare Remote Expert Consultation and AR Surgical Navigation Scenarios In this embodiment, the method is applied to an edge medical cloud network connecting multiple hospitals. The system acquires task flows generated in real time by operating room AR glasses, surgical robots, and 4K endoscopes. For tasks such as "surgical pose correction" or "lesion feature extraction," its inherent semantic feature vectors are extracted. This vector not only contains image features but also incorporates logical information about the surgical steps. The system forms quantified semantic description data and marks control flows involving life safety as the highest priority.

[0072] The system constructs morphic relationships between the task semantic space reflecting the importance of medical logic and the cross-regional physical network topology. The system employs an algebraic structure (such as isomorphic mapping) to characterize the nonlinear transformation impact of data when crossing protocol layers between the public network and the medical private network. The semantic adaptation rules determined accordingly ensure that the temporal semantics of surgical control commands remain absolutely accurate when traversing multiple gateways. For the multiple physical links between the expert and surgical ends, the system constructs a dynamic topology model based on non-Euclidean space. The Laplace operator is used to quantify the spatial correlation between routing nodes and predict potential load fluctuations during critical periods of remote surgery.

[0073] If the prediction results indicate that the latency of a certain relay node will exceed a threshold, the system activates a semantic conflict detection mechanism to identify semantic competition conflicts between the large-scale medical record backup task within the hospital and the remote surgical workflow. Based on this, the system determines the jump path that needs to be adjusted and temporarily caches non-real-time image data. Subsequently, the system performs time-series trend analysis to capture the patterns of traffic fluctuations during the surgery and determine the path with the lowest overhead that meets the heterogeneous computing power adaptation requirements of AI-assisted diagnosis.

[0074] The system generates scheduling plans and cross-protocol layer conversion instructions to ensure data format adaptation under different network environments. During transmission, the system continuously collects end-to-end latency and uses heterogeneous data fusion methods to analyze multi-source feedback from various medical terminals. If the adjusted path still does not meet the requirements, the system extracts network fluctuation trends from historical records to determine further optimized node combinations. Based on the latest state, the system dynamically adjusts morphological mapping rules and manifold space parameters, and uses anomaly pattern recognition to determine the optimization direction of semantic mapping, ensuring high reliability of telemedicine.

[0075] like Figure 1As shown, the edge network task flow cross-layer dynamic offloading method provided by the present invention, which integrates semantics and computing power prediction, includes the following steps: First, step S101 is executed to obtain the task flow to be processed and extract semantic feature vectors; then step S102 is executed to construct morphological associations and determine semantic adaptation rules; next step S103 is executed to construct a non-Euclidean space dynamic topology and predict load fluctuations; then step S104 is executed to perform semantic conflict detection and adjust the jump path, wherein if the load does not exceed a preset threshold, it directly proceeds to the next step; step S105 analyzes the time series trend to determine the optimal execution scheme; step S106 generates a scheduling plan and distributes it to the edge nodes; step S107 collects end-to-end latency and performs heterogeneous data fusion update; step S108 dynamically adjusts the mapping rules and performs abnormal pattern recognition optimization. After step S108 is completed, the dynamically adjusted morphological mapping rules are fed back to step S102 to form a closed-loop optimization mechanism to continuously improve the accuracy and efficiency of task flow offloading.

[0076] like Figure 2 As shown, the method architecture of this invention includes five core functional modules. The upper layer is the task semantic space, which sequentially executes task flow acquisition, semantic feature extraction, and semantic analysis model processing; the middle layer on the left is the morphological association and cross-layer processing module, which includes four sub-modules: morphological association construction, algebraic structure characterization, cross-protocol layer conversion, and semantic adaptation rules, responsible for mapping semantic information to the network protocol layer; the middle layer on the right is the dynamic optimization module, which includes four sub-modules: non-Euclidean space modeling, load fluctuation prediction, semantic conflict detection, and time-series trend analysis, working in conjunction with the morphological association module and providing optimization feedback; the lower layer is the physical network topology, covering edge nodes, forwarding devices, core switching nodes, and terminal devices; the bottom layer is the execution and feedback module, which includes four sub-modules: scheduling plan generation, task transmission execution, heterogeneous data fusion, and real-time status update. The layers are vertically connected through data flow, and the execution and feedback module feeds real-time status information back to the morphological association module, forming a complete closed-loop optimization system.

[0077] like Figure 6As shown, the edge network task flow cross-layer dynamic offloading system involved in this invention adopts a three-layer physical deployment architecture. The bottom layer is the terminal device layer, including terminal sensors 101, industrial cameras 102, mobile robots 103, and vehicle terminals 109. These terminal devices are responsible for collecting environmental data and executing end tasks. The middle layer is the edge computing layer, including edge servers 104, base stations 105, and gateway devices 106. This layer serves as the task offloading decision point, undertaking the reception, scheduling, and local processing of computing tasks. The top layer is the cloud / core layer, including a cloud platform 107 and a core switch 108, providing global resource scheduling and large-scale computing capability support. Data transmission between the terminal device layer and the edge computing layer is achieved through wireless connections (shown by dashed lines), while high-bandwidth data exchange is achieved between devices in the edge computing layer and between the edge computing layer and the cloud / core layer through wired connections (shown by solid lines). This three-layer physical architecture provides a spatial distribution basis for multi-level computing resources for cross-layer dynamic offloading of task flows, enabling tasks to flexibly migrate and offload between different physical layers according to real-time load conditions and network conditions.

[0078] Furthermore, the method for dynamic offloading of edge network task flows across layers, which integrates semantics and computing power prediction, may specifically include: S101. Obtain the task flow to be processed generated in real time in the edge network, extract the inherent semantic feature vector of each task, and form quantifiable semantic description data by analyzing the task content and resource requirements.

[0079] The system acquires task flow data generated in the edge network and performs real-time detection on each task flow. Through timestamp comparison and transmission latency analysis, it determines whether a task is in a pending state. For pending task flows, key information is extracted from the task content, and a text parsing tool is used to separate the core descriptive part of the task, obtaining a structured representation of the task content. The semantic features of the structured task content are analyzed, and a pre-established semantic analysis model is invoked to determine the specific vector representation of the semantic features. Based on the vector representation of the semantic features and combined with resource usage data in the task flow, resource analysis is performed to determine whether resource usage exceeds a preset threshold. If it does, the task is marked as a high-priority task. For task flows marked as high-priority tasks, quantified descriptive data is constructed by associating and mapping semantic feature vectors with resource usage data to obtain quantifiable semantic data. Based on the quantifiable semantic data, a processing order list for the task flows is generated, and a priority sorting tool is used to arrange the task flows to determine the final processing order. Based on the final processing order, scheduling instructions for the task flows are output and issued to the node devices in the edge network to complete the allocation and execution of the task flows.

[0080] In this step, the semantic analysis model adopts the lightweight BERT model. The model input is the task text description, and the output is a 256-dimensional semantic feature vector. The priority ranking tool adopts a priority scheduling algorithm based on heap sort. The processing weight of high-priority tasks is 3 times that of ordinary tasks, and the timestamp difference threshold for real-time detection is set to 5 seconds.

[0081] In one embodiment, for task flow data generated in the edge network, the data is first acquired through a data acquisition module. For example, in a smart manufacturing scenario, the task flow may include equipment monitoring instructions and production scheduling requests, and the data comes from the time-series records of sensor nodes.

[0082] For example, in the real-time detection step, the system uses a timestamp comparison method.

[0083] Specifically, the generation timestamp of the task flow is calculated by subtracting the current system time from the time of generation. If the difference is less than a preset threshold, such as 5 seconds, it is judged as a real-time task. Simultaneously, transmission latency analysis is combined with calculations of the average latency from the source node to the edge server, such as the round-trip time obtained using a ping test. If the latency exceeds 100 milliseconds, it is marked as a potentially delayed task, thus determining the task flow in the pending state. This detection mechanism ensures the timeliness of tasks, effectively filtering out outdated tasks and reducing system load in an industrial IoT environment. Furthermore, for pending tasks, when extracting key information, text parsing tools such as NLP-based entity recognition algorithms are used to separate the core descriptive parts. For example, from "Assembly Line A needs urgent repair," "Assembly Line A" and "urgent repair" are extracted as structured fields, resulting in a representation such as {Object:Assembly Line A, Action: Repair, Priority: Urgent}, thus providing basic data for subsequent semantic analysis.

[0084] In one embodiment, semantic features are analyzed through structured representation. A pre-established semantic analysis model, such as a BERT-based pre-trained model, is invoked to process the vector representation. Specifically, the structured text is input into the model to generate embedded vectors, such as mapping "emergency repair" to a high-dimensional vector [0.8, 0.2, ...] to represent urgency and type. Then, resource usage data in the task flow, such as CPU utilization and memory usage, is combined to perform resource analysis. If the utilization rate exceeds a threshold, such as 80%, it is marked as a high-priority task. This association is achieved through weighted vector summation, thereby quantifying resource pressure.

[0085] For example, for high-priority tasks, quantified descriptive data is constructed, and semantic vectors are mapped to resource data, such as by obtaining a comprehensive score through linear transformation. This score is used to generate a processing order list, and a priority sorting tool, such as heap sort, is used to arrange the task flow, determine the final order, and issue scheduling instructions to node devices to complete the allocation and execution. Through this approach, efficient task flow management is achieved, improving the response speed of the edge network.

[0086] S102. Based on the extracted semantic feature vectors, construct morphic associations between the task semantic space and the physical network topology, use algebraic structures to characterize the nonlinear transformations in cross-protocol layer transitions, and determine the semantic adaptation rules for task jumps in network layers.

[0087] The task data stream is acquired, and semantic features are extracted from the data packet payload to obtain the task semantic feature vector. A physical network topology graph is constructed based on the physical connections between network nodes. The task semantic feature vector is input into a pre-defined graph neural network model, which outputs the node embedding representation in the task semantic space. Homomorphic mappings from algebraic group theory are used to establish morphic relationships between the semantic embedding space and the physical topology graph structure. If the order-preserving property of the morphic relationship is lower than a pre-defined threshold, a nonlinear transformation is determined, and a kernel method is used to characterize the transformation process. Based on the characterized nonlinear transformation rules, semantic adaptation rules are determined when task data jumps between protocol layers.

[0088] In one possible implementation, the task data stream in the edge network is first acquired. For example, in an intelligent transportation system, the task data stream may include vehicle location updates and road condition monitoring data. Semantic features are extracted by parsing the data packet payload, such as using natural language processing tools to identify keywords such as "emergency braking" or "traffic congestion", thereby obtaining a task semantic feature vector, which is represented as a multi-dimensional numerical array to quantify the task intent.

[0089] For example, the next step is to construct a physical network topology map based on the physical connections between network nodes.

[0090] Specifically, in the aforementioned intelligent transportation scenario, nodes include roadside units and vehicle terminals. By scanning wireless link and wired interface data, a graph structure is generated, where nodes represent devices and edges represent connection strength, thus forming a complete topology representation and providing a foundation for subsequent analysis.

[0091] In one possible implementation, the semantic feature vector of the task is input into a pre-defined graph neural network model, such as the GraphSAGE model. This model performs message passing by aggregating features from neighboring nodes. Specifically, this process involves initializing node embeddings and then iteratively updating the representation of each node. For example, in a traffic task, after the vector is input, the model calculates the semantic similarity between nodes and outputs the embedding representation of each node in the task's semantic space. This could be a low-dimensional vector capturing semantic associations related to "emergency response," thus achieving a deep representation of the task distribution within the network. This embedding helps reveal the implicit relationships between tasks and nodes.

[0092] For example, homomorphic mapping methods from algebraic group theory can be used to establish morphic relationships between semantic embedding space and physical topological graph structures.

[0093] Specifically, homomorphic mapping refers to functions that preserve structure operations. For example, it maps vector addition in semantic space to edge connection operations in topological graphs. In transportation systems, if semantic embedding represents task priority, the mapping checks whether the order is preserved, i.e., high-priority tasks correspond to strongly connected paths. If the mapping does not completely preserve the order, it indicates that the structure does not directly correspond.

[0094] In one possible implementation, if the order-preserving property of the morphic correlation is lower than a preset threshold, for example, the threshold is set to 0.8, then a nonlinear transformation is determined to exist, and a kernel method is used to characterize the transformation process.

[0095] Specifically, kernel methods handle nonlinearity by introducing a high-dimensional feature space, such as using radial basis function kernels to project semantic vectors into a new space and compute transformation matrices. In traffic tasks, this can characterize semantic distortions from low-level protocols to high-level applications, thereby quantifying the effects of nonlinearity.

[0096] For example, based on the nonlinear transformation rules, semantic adaptation rules are determined when task data jumps between protocol layers. For instance, when jumping from TCP to the application layer, the rules adjust the feature vectors to match the protocol specifications, thereby ensuring efficient task transmission. Through this method, the optimized integration of task semantics and network structure is achieved.

[0097] In practical applications, the homomorphic mapping method in algebraic group theory is implemented by defining the task attribute set as group elements. For example, different protocol types of task flows (such as MQTT, Modbus) are mapped to specific transformation matrices, and task priorities are mapped to the eigenvalues ​​of the matrices. The cross-protocol layer transformation process is characterized as a composite algebraic operation between these matrices; when data jumps from the transport layer to the application layer, if the matrix structure preservation (i.e., order preservation) before and after the transformation is lower than a preset threshold of 0.8, it is determined to be a semantic distortion caused by nonlinear transformation. At this time, a high-dimensional radial basis function is introduced through the kernel method to perform spatial compensation on the distorted eigenvector, thereby ensuring the accurate restoration of the semantic intent of the task during hierarchical jumps.

[0098] like Figure 4 As shown, this invention achieves morphological mapping between the task semantic space and the physical network topology by constructing a two-layer network structure. The upper layer in the diagram represents the task semantic space, showcasing multiple task nodes including emergency braking detection, video stream transcoding, hardware defect recognition, surgical posture correction, ambient temperature monitoring, voice command parsing, and path planning optimization. The size of each node reflects its priority weight, and the thickness of the lines connecting nodes represents semantic similarity. The lower layer represents the physical network topology, containing heterogeneous network device nodes such as edge gateways, edge servers, core switches, and terminal clusters. Each node is identified by a different shape, and physical connections are represented by topological lines. The intermediate morphological mapping layer uses lines of varying thickness and line types to represent the structural mapping relationship between semantic nodes and physical nodes. Solid lines represent strong mapping relationships with a mapping strength greater than or equal to 0.8, dashed lines represent medium mapping relationships with a mapping strength between 0.5 and 0.8, and dotted lines represent weak mapping relationships with a mapping strength less than 0.5. The key nonlinear transformation points are marked with stars in the figure, and the order preservation threshold is set to 0.8. Mapping relationships below this threshold need to be compensated for by nonlinear transformation using the kernel method in algebraic group theory to ensure the structural preservation of cross-protocol layer mappings.

[0099] S103. For possible network hierarchical jump sequences of the task, construct a dynamic topology model in non-Euclidean space, quantify the spatial correlation between nodes through the Laplace operator, and obtain the load fluctuation prediction results for each jump sequence.

[0100] By analyzing task data to obtain jump sequence data in the network hierarchy, an initial sequence processing framework is constructed, resulting in a preliminary classification set of jump sequences. Based on this preliminary classification set, a dynamic topology model is constructed using non-Euclidean space geometric constraints to determine the node distribution pattern within the network hierarchy. For the node distribution pattern, the Laplace operator is used to perform spatial calculations on node correlations, obtaining the correlation strength matrix between nodes. Load fluctuation characteristics of key nodes are extracted from the correlation strength matrix. Combined with the jump sequence classification set, the potential trend of load fluctuations is determined. If the correlation strength exceeds a preset threshold, key nodes are prioritized, identifying key monitoring targets for load fluctuations. Through these key monitoring targets, abnormal jump patterns in sequence processing are analyzed to determine the impact range of abnormal jumps on the overall network hierarchy. Based on the impact range of abnormal jumps, a fluctuation distribution map of the predicted results is generated, obtaining the propagation path of load fluctuations in the dynamic topology. Based on the propagation path data and the constraints of the structural model, the node-related resource allocation strategy is adjusted to obtain the final load fluctuation prediction result.

[0101] In this step, the non-Euclidean space adopts a hyperbolic space model, the Laplace operator calculation step size is set to 0.1, the correlation strength threshold is set to 0.7, nodes exceeding this threshold are marked as key monitoring objects, and the time window for load fluctuation prediction is set to 5 seconds.

[0102] In one possible implementation, the jump sequence data obtained by task analysis can specifically refer to the sequential record of the network device ports through which a user request passes from the access layer switch to the application server in a data center network.

[0103] For example, the jump sequence of a video stream request might be "user terminal - edge router - core switch - load balancer - video server". The initial sequence processing framework is constructed by designing a data processing pipeline. This pipeline first cleans the raw logs, removing invalid records. Then, it segments consecutive jump records into independent sequence units according to time windows. Finally, based on fundamental characteristics such as the starting device type and jump step size, it initially categorizes them into sets such as "external access," "internal service call," or "data synchronization."

[0104] Specifically, when constructing a dynamic topology model based on the aforementioned preliminary classification set, the geometric constraints of non-Euclidean space can be embodied in a specific definition of the logical distance between nodes. For example...

[0105] In one embodiment, the network hierarchy is viewed as a hyperbolic space model, where the logical distance between nodes depends not only on the hop count of physical connections but, more importantly, on the similarity of the service types they carry. Two server nodes that frequently co-occur in the same "internal service call class" sequence, even if physically far apart, are assigned a closer logical distance under this model. Through this constraint, the model can dynamically form clustered or community-like node distribution patterns centered on service affinity, rather than a simple physical connection graph. For the obtained node distribution pattern, spatial computation is performed using the Laplace operator, essentially quantifying the differences between each node and its logical neighbors.

[0106] Understandably, the application of the Laplace operator to a discrete graph calculates a value for each node, reflecting the degree to which the node's state deviates from the average state of all its neighboring nodes. In a business scenario, a node's "state" can be defined as its real-time CPU utilization. By calculating the correlation of this deviation between all pairs of nodes, a correlation strength matrix can be obtained. High values ​​in this matrix indicate that the load fluctuation patterns of two nodes are highly synchronized.

[0107] For example, when the CPU utilization of application server A increases, the CPU utilization of database server B also shows a closely following upward trend, suggesting that they are on the same tightly coupled business chain. Key nodes are extracted from the association strength matrix, such as those nodes that have high-strength associations with more than three other nodes, and these are marked as hub nodes. Subsequently, the historical load fluctuation characteristics of these hub nodes are analyzed, such as the amplitude and frequency of fluctuations, and classified in conjunction with their respective jump sequences.

[0108] For example, if a storage node marked as a hub and belonging to the "data synchronization" category exhibits periodic spikes in its load fluctuations, it can be determined that these fluctuations are strongly correlated with scheduled data backup tasks. If the correlation strength between this storage node and a certain compute node exceeds a preset threshold, then that compute node is also marked as a key monitoring target. By monitoring these targets, abnormal redirection patterns can be more accurately detected, such as requests that should access node A abnormally flooding node B. This allows for an assessment of the impact of this anomaly on related business links and even the entire network area, and ultimately, prediction of the propagation path of load fluctuations, providing a basis for elastic resource scheduling.

[0109] The geometric constraints of the non-Euclidean space dynamically define logical distances through "business affinity." Specifically, the system continuously monitors the frequency with which nodes jointly process the same business flow within a preset time window; the higher the frequency, the closer the logical distance between the two nodes in the hyperbolic geometric model, rather than being limited by physical fiber length. The Laplace operator operation on this topology essentially calculates the gradient of the deviation between a node's current CPU utilization and the average load of its logical neighbors. This non-Euclidean space modeling method can identify "logical congestion" caused by business coupling earlier than traditional Euclidean space, thus enabling advanced prediction of load fluctuation paths.

[0110] like Figure 5 As shown, this invention uses a Poincaré disk model to map network nodes to hyperbolic space for visualization. The disk is divided into a core layer, an edge layer, and a terminal layer by three concentric rings from the inside out, each layer distinguished by a different line type. The position of a node in the Poincaré disk is determined by the logical distance calculated from business affinity; nodes with higher business affinity are closer in hyperbolic space. The size of each node represents the current load level, and the shade of gray filled in the nodes represents the load state, with dark filling indicating high load and light filling indicating low load. The lines connecting nodes represent the spatial correlation calculated by the Laplace operator; solid lines connect highly correlated node pairs with a correlation greater than or equal to 0.7 and are labeled with specific values; dashed and dotted lines represent medium and low correlation connections, respectively. The thick arrow curve in the figure indicates the propagation path of load fluctuations, demonstrating the dynamic process of overloaded nodes spreading load pressure to adjacent nodes. By setting a high correlation threshold of 0.7, the system can quickly identify the potential impact range of load fluctuations, providing spatial correlation basis for dynamic unloading decisions.

[0111] S104. If the load fluctuation prediction result of a certain jump sequence shows that the load of the endpoint node exceeds the preset threshold, then the semantic conflict detection mechanism is combined to identify the semantic competition contradictions between multiple tasks and determine the jump path that needs to be adjusted.

[0112] By processing load prediction data, the load status information of the endpoint nodes in the jump sequence is obtained to determine whether it exceeds a preset threshold. If the load status information of the endpoint nodes exceeds the preset threshold, a pre-established analysis model is used to identify resource allocation conflict points among multiple tasks based on the node load data in the jump sequence. Based on the distribution of resource allocation conflict points, semantic conflict data among multiple tasks is obtained, and specific content of competition conflict is obtained through semantic analysis tools. For the content of competition conflict, a semantic conflict detection mechanism is used to analyze the semantic correlation of each task in the jump sequence and determine the path direction that needs to be adjusted. From the adjusted path direction, jump sequence segments related to the endpoint node load are obtained, and the optimal path combination for load balancing is determined through comparative analysis. If the load distribution of the optimal path combination meets the preset threshold requirement, the combination is applied to the update processing of the jump sequence to obtain the final path adjustment scheme. Through the implementation of the path adjustment scheme, the updated node load data is obtained, and it is determined whether the load has recovered to within the preset threshold range, thus completing the load optimization process.

[0113] In this step, the node load threshold is set to 85% CPU utilization and 90% memory usage; semantic conflict detection uses cosine similarity calculation based on semantic tags, and if the similarity is less than 0.3, it is judged as a semantic competition conflict.

[0114] In one possible implementation, the data processing for load forecasting can be performed based on a time series analysis model.

[0115] Specifically, the system continuously collects historical load data from each network node, such as CPU utilization, memory usage, or network bandwidth utilization, and uses exponential smoothing or ARIMA models to predict the load status of the endpoint nodes in the next time period. Preset thresholds are typically set based on the physical performance limits of the nodes or the Service Quality Agreement (SLA); for example, defining a sustained CPU utilization exceeding 85% as an overload state. When the load of an endpoint node exceeds the threshold, the system invokes a pre-established analysis model to locate resource allocation conflicts. This analysis model can be a graph-based task dependency resolver.

[0116] For example, in a cloud computing task scheduling scenario, multiple computing tasks may compete for the read and write bandwidth of the same data storage node. The model constructs a task-resource contention graph by analyzing the access patterns, frequency, and resource requirements of each task in the jump sequence to the shared node (such as database instance DB1). This identifies the core set of competing tasks that cause overload on the endpoint node, such as the high-frequency write conflict between task A and task B to DB1. Based on the identified conflict points, the system further obtains semantic conflict data. This requires the use of semantic analysis tools, which can parse the business attributes of the tasks themselves.

[0117] For example, in video stream processing, task A is "real-time transcoding," and task B is "content moderation." Both require reading the original video stream, but transcoding demands high bandwidth and low latency, while moderation can tolerate some latency but requires high computing resources. Semantic analysis tools extract semantic tags such as "latency-sensitive" and "computation-intensive" by parsing task descriptions or resource request tags, thus clarifying that the specific content of the competitive conflict is "bandwidth resource contention" rather than "computation resource contention." For clearly defined competitive conflicts, the system employs a semantic conflict detection mechanism. The core of this mechanism is to assess the semantic relevance of tasks in the jump sequence and plan and adjust the path accordingly.

[0118] For example, the mechanism will determine whether adjusting the redirection path of the "content moderation" task from passing through the currently overloaded storage node N1 to a node N2 that stores the same data replica but has a lower load will affect the fulfillment of its "computation-intensive" semantics. By analyzing the matching degree between the task semantics and node attributes (such as N2 also providing the required computing units), feasible path adjustment directions are determined, i.e., the moderation task is redirected to N2. From the feasible adjustment directions, the system selects redirection sequence segments directly related to the load of the endpoint node for comparative analysis.

[0119] For example, comparing the predicted load distribution of the endpoint node (such as the transcoding server) under two scenarios: keeping the original path for the "real-time transcoding" task and rerouting the "content moderation" task to N2. Through simulation calculations, the optimal path combination that balances the load on the endpoint node and all related nodes is determined. If this combination meets the preset threshold requirements of all nodes, a path adjustment plan is generated, such as updating the routing table to redirect data requests from the moderation task to N2. Finally, the system implements this plan and monitors the updated node load data. By continuously collecting the actual load metrics of each node after adjustment and comparing them with preset thresholds, the success of the load optimization is verified, thus forming a closed-loop management process from prediction, analysis, decision-making to verification.

[0120] The semantic conflict detection mechanism employs a label-based scoring logic: the system pre-assigns semantic labels such as "latency-sensitive," "computation-intensive," or "reliability-priority" to tasks. When two or more tasks are predicted to compete for bandwidth on the same physical node, if all task labels include "latency-sensitive," it is determined to be a serious semantic contention conflict. The path adjustment scheme sorts tasks according to their overall scores, prioritizing the highest-scoring task to be transmitted along the original path, while redirecting the second-highest-scoring task ("computation-intensive" but with higher latency tolerance) to redundant nodes with equivalent computing power (e.g., equipped with the same type of GPU) but currently lower bandwidth load, thereby eliminating resource contention at the logical level.

[0121] S105. By performing time-series trend analysis on the adjusted jump path, the evolution pattern of node load in the time dimension is captured, and the path with the lowest overall cost and meeting the requirements of heterogeneous computing power adaptation is determined as the task execution scheme.

[0122] By collecting data from the jump paths, the changes in node load at different time points are recorded. Time-series analysis tools are used to obtain the evolution pattern of node load over time. Based on the results of the evolution pattern, a pre-established analysis model is used to determine the basis for calculating the overall cost, considering the fluctuations in node load over time. From the basis for calculating the overall cost, path segment data related to the lowest cost is obtained. By comparing the cost values ​​of different path segments, the path combination with the lowest cost is determined. For the path combination with the lowest cost, its compatibility data with heterogeneous computing power is obtained. If the compatibility data meets the computing power requirements, the path combination is marked as a candidate execution scheme. Based on the marking results of the candidate execution schemes, path allocation information related to task execution is obtained. By integrating the path allocation information, the final task execution scheme is determined. Through the output of the final task execution scheme, the load allocation data of each node is obtained. A logic verification tool is used to determine whether the load allocation meets the preset balance requirements.

[0123] In this step, the time series trend analysis adopts the variational mode decomposition algorithm with a decomposition layer of 3; the overall overhead includes latency overhead, computing power overhead, and bandwidth overhead, with a weight ratio of 4:3:3; and the heterogeneous computing power adaptation includes matching verification of three hardware types: GPU, FPGA, and CPU.

[0124] In one possible implementation, the time series analysis tool can employ a prediction method based on state-space decomposition.

[0125] Specifically, the node load time series data collected by the system is first decomposed into multiple eigenmode components with different center frequencies using a variational mode decomposition algorithm.

[0126] For example, in edge computing scenarios, the CPU utilization sequence of an edge gateway node may be decomposed into a low-frequency component representing long-term trends, a mid-frequency component reflecting daily cyclical business fluctuations, and a high-frequency noise component characterizing random interference. By establishing autoregressive models for each component and making predictions, and then superimposing the prediction results, the evolution pattern of node load in future periods can be obtained. This pattern can clearly reveal the periodic peaks and troughs of the load. Based on the evolution pattern, a pre-established analytical model is used to determine the overall overhead. This model can be a multi-objective weighted evaluation function.

[0127] For example, in data center network traffic scheduling, the calculation of overall overhead includes not only path transmission latency but also node processing energy consumption and link leasing costs. The model dynamically adjusts the weighting coefficients of each item based on predicted high-load periods according to load evolution patterns. For instance, during predicted peak load periods, the weight of the latency item is significantly increased to prioritize the response speed of critical services. From this calculation, the system selects the path combination with the lowest overhead value.

[0128] For example, for a query task from a user terminal to a cloud database, the predicted overall cost is calculated by comparing two main path segments passing through core switch A and core switch B within a specific future time window. If path A has fewer hops, but the nodes it passes through experience high load during the prediction period, leading to a surge in processing latency, while path B bypasses a less loaded backup link, resulting in lower overall cost, then the system marks the combination of path B as a candidate. Subsequently, the system needs to evaluate the compatibility of this candidate path with heterogeneous computing power.

[0129] In one embodiment, this involves checking whether nodes along the path have the specific computing resources required for the task.

[0130] For example, if the above query task includes an AI inference subtask, it is necessary to ensure that the node or neighboring node ultimately reached by the path has a GPU accelerator card deployed. Adaptation data can be obtained by querying the node resource registry. If the endpoint node of path B happens to be equipped with a GPU instance that meets the computing power requirements, the adaptation is successful, and the path is officially adopted. Finally, the system integrates the path allocation information of all tasks to form a global execution plan. A logic verification tool is then activated, its core function being to check whether the implementation of the plan will lead to new load hotspots. This tool can perform rapid simulations based on the maximum flow minimum cut theorem, calculating the variance between the predicted load of all nodes and the system average load under the plan. If this variance is lower than a preset balancing threshold, the load allocation is deemed to meet the requirements, and the plan is finally output and executed.

[0131] S106. Based on the determined task execution plan, generate a scheduling plan containing specific node sequences and cross-protocol layer conversion instructions, distribute the plan to edge nodes and forwarding devices, and trigger the task to be transmitted along the specified path.

[0132] By analyzing task execution requirements, the node sequence information needed for task driving is obtained, and the distribution location of each node on the specified path is determined. Based on the node sequence information, cross-protocol layer conversion instructions are constructed. If the conversion instructions do not match the device interaction requirements, the instruction format is adjusted according to preset mapping rules to obtain an adapted instruction set. Using the adapted instruction set, a scheduling plan for edge nodes and forwarding devices is generated, and the execution sequence of each device in the plan is obtained to determine the priority order of task transmission. Based on the priority order, task-driven transmission resources are allocated. By monitoring the data flow status during transmission in real time, if the data flow is interrupted, a switch to a backup path is obtained to obtain a stable transmission channel. Feedback data of device interaction is extracted from the stable transmission channel, and the execution parameters of the scheduling plan are adjusted based on the feedback data to determine the transmission efficiency of the task on the specified path. The evaluation results of the transmission efficiency are obtained, and by comparing them with preset thresholds, if the efficiency is lower than the threshold, the execution logic of the cross-protocol layer conversion instructions is optimized to obtain an improved transmission scheme. Based on the improved transmission scheme, the task-driven configuration of edge nodes and forwarding devices is updated to determine the final transmission status of the task along the specified path.

[0133] In this step, cross-protocol layer conversion includes format mapping for four protocols: MQTT, Modbus TCP, C-V2X, and OPC UA; the backup path switching trigger condition is a data stream packet loss rate greater than 5%, and the transmission efficiency threshold is set to 90%.

[0134] By analyzing the task execution requirements, the node sequence information required for task driving is obtained, and the distribution location of each node on the specified path is determined.

[0135] For example, in an Industrial Internet of Things (IIoT) scenario, the task requirement is to transmit real-time quality inspection images from production line A to edge computing center B for processing. The node sequence information includes production line sensor nodes, workshop gateway nodes, factory edge server nodes, and finally, the computing center node. These nodes are distributed sequentially along a specified physical and logical path from the workshop to the server room. Based on the node sequence information, cross-protocol layer conversion instructions are constructed.

[0136] Specifically, cross-protocol layer translation commands refer to the conversion of application layer data requests into a series of control commands suitable for network layer routing and link layer transmission.

[0137] In one possible implementation, task data is initially encapsulated as MQTT protocol messages, but upon passing through the workshop gateway, it needs to be converted to the Modbus TCP frame format suitable for industrial Ethernet. The system determines whether this conversion instruction matches the interaction requirements of the next-hop device (such as a PLC that only supports the OPC UA protocol). If they do not match, the instruction format is adjusted according to preset mapping rules. The preset mapping rules can be a protocol field mapping table, for example, mapping MQTT topic names to OPC UA node identifiers, thus obtaining a set of instructions adapted to the entire link from sensor to PLC. Using the adapted instruction set, a scheduling plan is generated for edge nodes and forwarding devices. This scheduling plan needs to obtain the execution timing of each device and determine the priority order of task transmission.

[0138] For example, high-priority real-time alarm data is scheduled to be transmitted before low-priority log data, and a fixed transmission time slot is allocated to high-priority data. Based on this priority order, the system allocates corresponding network bandwidth and computing resources. By monitoring the data flow status during transmission in real time, if a data flow interruption is detected, the system switches to a backup path. The backup path may be a redundant link connected via a wireless mesh network, thus obtaining a stable transmission channel. Feedback data from device interactions, such as network latency, packet loss rate, and device processing heartbeats, is extracted from the stable transmission channel. Based on this feedback data, the execution parameters of the scheduling plan can be adjusted, such as dynamically adjusting the data transmission window size or retransmission timeout, thereby determining the actual transmission efficiency of the task on the specified path. After obtaining the evaluation results of the transmission efficiency, it is compared with a preset threshold. If the efficiency is lower than the threshold, the execution logic of the cross-protocol layer conversion instruction is optimized.

[0139] In one embodiment, the optimization method includes merging fragmented conversion steps or prefetching frequently used mapping relationships and caching them locally, thereby obtaining an improved transmission scheme. Based on the improved transmission scheme, the task-driven configurations of edge nodes and forwarding devices are updated, and the transmission status of tasks along the specified path is finally determined, achieving efficient and reliable data transmission.

[0140] S107. During the task transmission process, end-to-end latency and node resource consumption data are continuously collected, and heterogeneous data fusion method is used to integrate multi-source feedback to obtain real-time update results of network status.

[0141] By collecting data during task transmission, real-time information on end-to-end latency and node resources is obtained to determine the current network load status. Based on changes in load status, a preset threshold is used for comparison to determine if there are any abnormal transmission delays, thus providing an initial location of the latency problem. If the transmission delay exceeds the preset threshold, in-depth analysis of node resource consumption data is performed to pinpoint the specific location of resource bottlenecks and determine the range of affected nodes. By integrating the resource consumption data of affected nodes and multi-source feedback information, a support vector machine algorithm is used to classify heterogeneous data and obtain a priority ranking of resource allocation. Based on the priority ranking results, the resource configuration scheme of each node in the network is dynamically adjusted to obtain the adjusted transmission path and determine whether it meets the latency optimization requirements. If the adjusted transmission path still does not meet expectations, historical records are extracted from the status monitoring data to analyze the fluctuation trend of the network status and determine further optimized node combinations. By collecting data on the optimized node combinations in real time and updating the end-to-end latency and resource consumption monitoring data, the latest dynamic results of the network status are obtained.

[0142] In this step, the end-to-end latency threshold is set to 100 milliseconds; the support vector machine algorithm uses a linear kernel function; the heterogeneous data classification features include three categories: latency, load, and semantic priority; and the resource allocation priority is divided into three levels: high, medium, and low.

[0143] like Figure 3 As shown, the real-time feedback optimization closed-loop mechanism of this invention includes three collaborative functional groups. The left side is the data acquisition group, responsible for end-to-end latency acquisition, node resource consumption monitoring, and aggregation of multi-source feedback information. The middle side is the data fusion and analysis group, which fuses the acquired multi-source heterogeneous data, performs resource bottleneck location and fluctuation trend analysis after SVM classification. The right side is the dynamic adjustment group, which adjusts morphological mapping rules and identifies abnormal patterns based on the analysis results, while updating manifold space parameters, ultimately converging to the path prediction optimization module. The path prediction optimization result is output to the upper-level decision module as the semantic mapping optimization direction, and also fed back to the data acquisition group through the feedback closed loop, enabling continuous monitoring and dynamic optimization of end-to-end latency, ensuring that the entire offloading system can continuously adapt and adjust according to the real-time network status during operation.

[0144] In one possible implementation, data acquisition during task transmission can be accomplished through probe agents deployed on end devices and network nodes.

[0145] Specifically, these agents periodically capture the timestamps of data packet transmission and reception, and collect metrics such as node CPU utilization, memory usage, and network interface queue depth, thereby forming a real-time information set of end-to-end latency and node resource consumption.

[0146] For example, in a scenario where a video stream analysis task transmits data from an edge camera to a regional server, the probe records the time difference between each frame leaving the camera and arriving at the server, while monitoring the CPU load of each switch along the way to determine the overall congestion level of the current network.

[0147] It should be noted that the preset threshold comparison mechanism is usually set based on historical operational data or service level agreement requirements. The system compares the real-time calculated average end-to-end latency with a dynamic baseline threshold, which may be adaptively adjusted according to the expected load of different times of day or network areas. When the system determines that the latency exceeds the threshold, it triggers the anomaly localization process.

[0148] For example, in an Industrial Internet of Things (IIoT) scenario, if the latency of control commands being sent from the cloud platform to the field-programmable logic controller (FPGA) consistently exceeds a threshold of 50 milliseconds, it initially indicates an anomaly in the path from the core network to the edge access segment. For in-depth analysis of resource bottlenecks, correlation graph analysis of resource consumption data can be introduced. The system integrates the CPU, memory, and bandwidth data of all relevant nodes during the period exceeding the latency threshold and constructs a correlation graph showing their consumption trends. By identifying nodes in the graph whose consumption curves show a synchronous sharp increase and are on the critical path, the bottleneck can be accurately located.

[0149] For example, in the multi-channel video aggregation and transmission of a smart city, if it is found that the central processing unit utilization of a certain edge gateway is consistently above 95%, while the resource consumption of its upstream and downstream nodes is relatively stable, then it can be determined that the gateway is the resource bottleneck node causing the overall latency to increase.

[0150] Specifically, the Support Vector Machine (SVM) algorithm is used here to classify heterogeneous, multi-source node state data to determine resource allocation priorities. The algorithm treats each node as a data sample, whose feature vector includes its real-time resource consumption metrics, historical failure frequency, and business importance level. Through a trained classification model, the algorithm can categorize these samples into different priority classes.

[0151] For example, in a vehicle-to-everything (V2X) edge computing scenario, a node processing emergency braking warning messages, even if its current CPU utilization is only 70%, may be classified as the highest priority by the support vector machine model due to the high real-time requirements of its service, thus receiving priority access to additional computing resource quotas. The core of dynamically adjusting resource allocation based on priority ranking lies in a centralized resource orchestrator. This orchestrator receives the priority list and issues instructions to the virtualization infrastructure manager or software-defined network controller according to the policy.

[0152] For example, for nodes identified as having the highest priority, the orchestrator might instruct the server hosting that node to dynamically increase the allocation of virtual CPU cores, or reserve a dedicated bandwidth queue for its traffic on the switch via software-defined networking. After the adjustment, the system will recalculate the theoretical transmission path of the task, for example, switching some traffic from an overloaded path to a backup path with more hops but sufficient resources, and immediately collect new latency data for verification. If the latency still does not meet the standard after path adjustment, trend analysis and node combination optimization based on historical data are required. The system will extract records of network state fluctuations under similar load patterns over a period of time from the state monitoring database, including the latency contribution and resource release efficiency of each node. Through time series analysis, it can predict which node combinations can produce more stable low-latency performance when working together.

[0153] For example, in distributed machine learning training tasks, the system might discover that when the parameter server and two specific worker nodes are deployed in the same rack, even with a higher overall network load, the synchronization latency is significantly lower than in a cross-rack deployment mode. Based on this, the system can optimize its node combination selection strategy. Finally, through continuous real-time data collection of the optimized node combination, the system achieves closed-loop state updates. Probes deployed on newly selected nodes report their latency and resource data more frequently. These latest dynamic results are fed back into the system's state database to correct the classification boundaries of the support vector machine model and the parameters of the resource prediction model, thus enabling the entire resource scheduling and path optimization mechanism to continuously adapt to changes in network state.

[0154] S108. For real-time updated network status data, dynamically adjust the morphological mapping rules and manifold space parameters, detect sudden load changes through abnormal pattern recognition, and determine the optimization direction of semantic mapping and path prediction.

[0155] By collecting real-time network status updates and using a preset threshold comparison method, abnormal load surges are detected to obtain preliminary judgments on load changes. If the load change exceeds the preset threshold, pattern recognition processing is performed on the abnormal pattern data to analyze the specific time points and node locations of the sudden changes and determine the affected network area. Based on the affected network area, relevant data on morphological mapping is extracted, the mapping rules are dynamically adjusted, and the adjusted mapping model structure is obtained. Using the adjusted mapping model structure and the existing configuration of the manifold space, the distribution of spatial parameters is updated to obtain an optimized spatial model framework. Based on the optimized spatial model framework, relevant data on semantic mapping is processed to generate a preliminary path prediction scheme and determine alternative path planning directions. By comparing the alternative path prediction directions with the latest network status updates, path schemes that meet load balancing requirements are selected to obtain the final path planning result. Based on the final path planning result, the network status monitoring data is updated, the handling process of load surges is recorded, and the latest dynamic information on network operation is obtained.

[0156] In this step, the threshold for sudden load changes is set to a 50% increase in load; the K-nearest neighbor algorithm is used for abnormal pattern recognition, and the number of nearest neighbors k is set to 5; the principal component analysis method is used for updating manifold space parameters, and the dimensionality reduction dimension is set to 8.

[0157] In one embodiment, load fluctuations are initially identified by collecting real-time updates of network status data, such as in a wireless sensor network where distributed sensor nodes collect data packet transmission time and node CPU utilization once per second.

[0158] In one embodiment, a preset threshold comparison method is used to detect abnormal situations of load bursts. For example, the end-to-end delay threshold is set to 50 milliseconds. When the collected data exceeds this value, the system obtains a preliminary judgment result of load change by comparing the current load with the historical average value. If the load change exceeds the preset threshold, pattern recognition processing is performed on the abnormal pattern data.

[0159] Specifically, the K-nearest neighbor algorithm is used to analyze the data sequence and identify the specific time point of sudden changes, such as the 10th second, and the node location, such as node A to C, thereby determining the affected network area. For example, in a data center network, this may cover the link area from router 1 to switch 3. In this way, the system can accurately locate the source of the problem and avoid blind adjustments.

[0160] In one embodiment, relevant data of morphological mapping is extracted based on the affected network region. Here, morphological mapping refers to a function that maps network state variables such as latency and bandwidth to an abstract state space. For example, the latency value is converted from actual milliseconds to a standardized score of 0 to 1 through a linear transformation. The content of the mapping rules is dynamically adjusted, such as modifying the weight coefficients to emphasize high-load nodes, to obtain the adjusted mapping model structure. The adjusted mapping model structure is combined with the existing configuration of the manifold space, where the manifold space is a low-dimensional embedded geometric structure representing the network topology. For example, principal component analysis is used to reduce the dimensionality of high-dimensional network data to a 2-dimensional plane, and the distribution of spatial parameters is updated, such as uniformizing the distance between nodes, to obtain an optimized spatial model framework, thereby making the model more adaptable to dynamic loads.

[0161] In one embodiment, semantic mapping data is processed according to the optimized spatial model framework. Here, semantic mapping is the process of converting network path attributes into predictable semantic labels, such as mapping "high-latency path" to the label "avoid". A preliminary path prediction scheme is generated, and alternative paths for path planning are determined, such as detouring from node B to E. By comparing the alternative paths for path prediction with the latest updated information of the network status, a path scheme that meets the load balancing requirements is selected, such as comparing the expected delays of alternative paths and selecting the one with the lowest delay, to obtain the final path planning result.

[0162] In one embodiment, the network status monitoring data is updated based on the final path planning result, the handling process of load bursts is recorded, and the latest dynamic information of network operation is obtained, thereby improving the overall transmission efficiency.

[0163] When adjusting the manifold space parameters, the system uses principal component analysis to project high-dimensional edge network state data (including latency, jitter, packet loss rate, and temperature of nodes at each level) into a low-dimensional embedding space, forming a manifold framework describing the network state evolution. Anomaly pattern recognition is achieved by monitoring the offset velocity of data points on the manifold surface: if a data point representing the current state momentarily deviates from the envelope of its historical normal trajectory, it is determined to be a load burst. At this point, by increasing the local curvature parameter of the manifold space, the system can more sensitively capture subtle trends in load evolution, thereby guiding the path prediction module to avoid potential congestion singularities and achieving closed-loop optimization of the network operating state.

[0164] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for dynamic cross-layer unloading of edge network task flows that integrates semantics and computational power prediction, characterized in that, The method includes: The system acquires the real-time task flow generated in the edge network, extracts the inherent semantic feature vector for each task, and forms quantifiable semantic description data by analyzing the task content and resource requirements. Based on the extracted semantic feature vectors, a morphic association is constructed between the task semantic space and the physical network topology. An algebraic structure is used to characterize the nonlinear transformation in cross-protocol layer transitions, and the semantic adaptation rules of tasks in network layer jumps are determined. For possible network hierarchical jump sequences of a task, a dynamic topology model in non-Euclidean space is constructed. The spatial correlation between nodes is quantified by the Laplace operator to obtain the load fluctuation prediction results for each jump sequence. If the load fluctuation prediction result of a certain jump sequence shows that the load of the endpoint node exceeds the preset threshold, then the semantic conflict detection mechanism is combined to identify the semantic competition contradictions between multiple tasks and determine the jump path that needs to be adjusted. By performing time-series trend analysis on the adjusted jump path, we can capture the evolution of node load over time and determine the path with the lowest overall overhead that meets the requirements for heterogeneous computing power adaptation as the task execution scheme.

2. The method according to claim 1, characterized in that, Also includes: Based on the determined task execution plan, a scheduling plan containing specific node sequences and cross-protocol layer conversion instructions is generated, and the plan is sent to edge nodes and forwarding devices to trigger the task to be transmitted along the specified path; During task transmission, end-to-end latency and node resource consumption data are continuously collected, and heterogeneous data fusion method is used to integrate multi-source feedback to obtain real-time updates of network status. Based on real-time updated network status data, the morphological mapping rules and manifold space parameters are dynamically adjusted. Abnormal pattern recognition is used to detect sudden load changes and determine the optimization direction of semantic mapping and path prediction.

3. The method according to claim 1, characterized in that, The process involves acquiring the real-time task flow generated in the edge network, extracting the inherent semantic feature vector for each task, and analyzing the task content and resource requirements to form quantifiable semantic description data, including: Acquire task flow data generated in the edge network, perform real-time detection on each task flow, and determine whether a task is in a pending state by comparing timestamps and analyzing transmission delays. For a task flow in a pending state, key information is extracted from the task content, and a text parsing tool is used to separate the core description of the task to obtain a structured representation of the task content. By analyzing the semantic features of the task content represented in a structured manner, and calling a pre-established semantic analysis model, the specific vector representation of the semantic features is determined. Based on the vector representation of semantic features and combined with resource usage data in the task flow, a resource analysis operation is performed to determine whether the resource usage exceeds a preset threshold range. If it does, it is marked as a high-priority task. For task flows marked as high-priority tasks, quantitative descriptive data is constructed. By associating and mapping semantic feature vectors with resource consumption data, quantifiable semantic data is obtained. Based on quantifiable semantic data, a list of task flow processing order is generated, and a priority sorting tool is used to arrange the task flow to determine the final processing order. Based on the final processing order, the task flow scheduling instructions are output, and instructions are issued to the node devices in the edge network to complete the allocation and execution of the task flow.

4. The method according to claim 1, characterized in that, The process involves constructing morphological associations between the task semantic space and the physical network topology based on the extracted semantic feature vectors, using an algebraic structure to characterize the nonlinear transformations in cross-protocol layer transitions, and determining the semantic adaptation rules for tasks during network layer jumps, including: Obtain the task data stream, extract semantic features from the data packet payload, and obtain the task semantic feature vector; Construct a physical network topology diagram based on the physical connections between network nodes; Input the task semantic feature vector into a pre-defined graph neural network model, and the model outputs the embedding representation of the node in the task semantic space; Using the homomorphism method in algebraic group theory, a morphic relationship is established between the semantic embedding space and the physical topological graph structure; If the order-preserving property of the morphic correlation is lower than a preset threshold, it is determined that there is a nonlinear transformation, and the kernel method is used to characterize the transformation process. Based on the nonlinear transformation rules, semantic adaptation rules are determined when task data jumps between protocol layers.

5. The method according to claim 1, characterized in that, For the possible network hierarchical jump sequences of the task, a dynamic topology model in non-Euclidean space is constructed. The spatial correlation between nodes is quantified using the Laplace operator to obtain the load fluctuation prediction results for each jump sequence, including: By analyzing tasks, we can obtain jump sequence data in the network layers, construct an initial sequence processing framework, and obtain a preliminary classification set of jump sequences. Based on the preliminary classification set of jump sequences, a dynamic topology model is constructed using geometric constraints in non-Euclidean space to determine the node distribution pattern within the network hierarchy. Based on the node distribution pattern, the Laplacian operator is used to perform spatial calculations on node correlations to obtain the correlation strength matrix between nodes; The load fluctuation characteristics of key nodes are extracted from the correlation strength matrix. Combined with the classification set of jump sequences, the potential trend of load fluctuation is judged. If the correlation strength exceeds the preset threshold, the key nodes are prioritized to obtain the key monitoring objects of load fluctuation. By focusing on key monitoring targets of load fluctuations, analyze abnormal jump patterns in sequence processing to determine the scope of impact of abnormal jumps on the overall network hierarchy. Based on the impact range of abnormal jumps, a fluctuation distribution map of the prediction results is generated to obtain the propagation path of load fluctuations in the dynamic topology; Based on the data from the propagation path and the constraints of the structural model, the resource allocation strategy related to the nodes is adjusted to obtain the final load fluctuation prediction result.

6. The method according to claim 1, characterized in that, If the load fluctuation prediction result of a certain jump sequence shows that the load of the endpoint node exceeds a preset threshold, then, in conjunction with the semantic conflict detection mechanism, the semantic competition contradictions between multiple tasks are identified, and the jump path that needs to be adjusted is determined, including: By processing the load prediction data, the load status information of the endpoint node in the jump sequence is obtained to determine whether it exceeds the preset threshold. If the load status information of the endpoint node exceeds the preset threshold, a pre-established analysis model is used to determine the resource allocation conflict points among multiple tasks based on the node load data in the jump sequence. Based on the distribution of resource allocation conflict points, semantic conflict data between multiple tasks is obtained, and specific content of competitive conflict is obtained through semantic analysis tools; For content with conflicting objectives, a semantic conflict detection mechanism is used to analyze the semantic relationships between tasks in the jump sequence and determine the path direction that needs to be adjusted. From the adjusted path directions, obtain the jump sequence fragments related to the load of the endpoint node, and determine the optimal path combination for load balancing through comparative analysis; If the load distribution of the optimal path combination meets the preset threshold requirement, the combination is applied to the update process of the jump sequence to obtain the final path adjustment scheme. By implementing the path adjustment scheme, the updated node load data is obtained, and it is determined whether the load has recovered to the preset threshold range, thus completing the load optimization process.

7. The method according to claim 1, characterized in that, The process involves performing time-series trend analysis on the adjusted jump paths to capture the evolution of node load over time, and determining the path with the lowest overall overhead and that meets the requirements for heterogeneous computing power adaptation as the task execution scheme. This includes: By collecting data from the jump path, the changes in node load at different points in time are recorded. Using time series analysis tools, the evolution pattern of node load over time is obtained. Based on the results of the evolution model, and considering the fluctuations in node load over time, a pre-established analytical model is used to determine the basis for calculating the overall overhead. From the basis of the overall cost calculation, obtain the path segment data related to the lowest cost, and determine the path combination with the lowest cost by comparing the cost values ​​of different path segments. For the path combination with the lowest cost, obtain its compatibility data with heterogeneous computing power. If the compatibility data meets the computing power requirements, mark the path combination as a candidate execution plan. Based on the labeling results of candidate execution schemes, obtain path allocation information related to task execution, and determine the final task execution scheme by integrating the path allocation information; By analyzing the output of the final task execution plan, load distribution data for each node is obtained, and logical verification tools are used to determine whether the load distribution meets the preset balance requirements.

8. An electronic device, comprising: A processor, a memory, and a program stored in the memory and executable on the processor, characterized in that, when executed by the processor, the program implements the method as described in any one of claims 1 to 6.

Images