An artificial intelligence assisted edge computing task scheduling method

Abstract

The application discloses an artificial intelligence assisted edge computing task scheduling method, relates to the technical field of data processing, and comprises the following steps: collecting environment data by deploying a monitoring agent on an edge node; processing the environment data by adopting a space-time joint coding mechanism to generate a unified node state feature vector; constructing a local digital twin in each edge domain based on the node state feature vector; performing privacy protection aggregation and updating on the local digital twin model parameters of each edge domain to generate a global digital twin model; optimizing a scheduling strategy based on a multi-objective deep reinforcement learning algorithm; issuing the optimized scheduling strategy to a physical edge node for execution; continuously comparing the actual running state of the physical system with the simulation state of the digital twin by adopting a virtual-real comparison monitoring mechanism; and when it is monitored that the state deviation exceeds an early warning threshold, triggering a dynamic compensation mechanism to match and execute a compensation action from a preset compensation strategy library.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an artificial intelligence-assisted edge computing task scheduling method. Background Technology

[0002] In the current field of edge computing task scheduling, existing technical solutions suffer from systemic shortcomings. The core issue lies in the difficulty of effectively coordinating and optimizing multiple key performance indicators in dynamic and heterogeneous environments, resulting in a trade-off between task execution efficiency, energy consumption control, and system reliability. This problem is particularly pronounced in typical edge scenarios characterized by significant network state fluctuations, frequent changes in computing load, limited node resources, and real-time changes in task priorities. From the perspectives of different stakeholders, multiple conflicting demands exist: end users seek low-latency and highly reliable service experiences; system operators need to strictly control energy consumption and operational costs while ensuring performance; and deployment parties require scheduling systems with strong cross-domain coordination capabilities and rapid anomaly recovery characteristics. Practical applications show that existing methods generally suffer from long task latency, high energy consumption, and unstable reliability, and recover slowly from node failures or network fluctuations, failing to meet the comprehensive requirements of real-time performance, energy efficiency, and robustness in edge computing scenarios. Summary of the Invention

[0003] This application provides an AI-assisted edge computing task scheduling method that addresses the challenges of existing technologies in collaboratively optimizing task latency, energy consumption, and reliability in dynamic heterogeneous environments, as well as their lack of adaptive fault tolerance and cross-domain collaborative capabilities. By constructing a privacy-preserving federated digital twin, employing intelligent decision-making based on multi-objective reinforcement learning and look-ahead optimization, and implementing a virtual-real comparison monitoring and dynamic compensation mechanism, the method achieves global optimization of the scheduling strategy, rapid fault tolerance, and continuous self-evolution, thereby improving the overall performance, resilience, and scalability of the system in complex edge environments.

[0004] This application provides an artificial intelligence-assisted edge computing task scheduling method, including: S1: By deploying monitoring agents on edge nodes, environmental data including network traffic, computing load, energy consumption, and task queue status is collected; the environmental data is processed using a spatiotemporal joint coding mechanism to generate a unified node status feature vector; S2: Based on node state feature vectors, construct local digital twins in each edge domain; through a federated coordination server, aggregate and update the parameters of the local digital twin model in each edge domain with privacy protection, and generate a global digital twin model. S3: In the virtual environment constructed by the global digital twin model, optimize the scheduling strategy based on the multi-objective deep reinforcement learning algorithm; and distribute the optimized scheduling strategy to the physical edge nodes for execution. S4: Employs a virtual-real comparison monitoring mechanism to continuously compare the actual operating state of the physical system with the simulated state of the digital twin; when the detected state deviation exceeds the warning threshold, a dynamic compensation mechanism is triggered to match and execute compensation actions from a pre-set compensation strategy library.

[0005] Furthermore, the node state feature vector includes: performing temporal interpolation on the collected discrete environmental data using a preset time window to complete it into continuous temporal data; splicing and fusing the spatial attribute features of the node with the continuous temporal data to obtain spatiotemporal fusion features; and inputting the spatiotemporal fusion features into the coding layer for nonlinear mapping to obtain a node state feature vector of a unified dimension.

[0006] Furthermore, the global digital twin model is constructed using a federated learning framework, where differential privacy technology is used to perturb the gradient during the aggregation of model parameters; the aggregated model is refined into a lightweight inference model through model distillation technology; and the federated learning framework dynamically adjusts the contribution weight of each edge domain in the model aggregation based on the amount and integrity of the data.

[0007] Furthermore, the multi-objective deep reinforcement learning algorithm employs a proximal policy optimization algorithm and includes a look-ahead optimization mechanism; the look-ahead optimization mechanism generates environmental state variants based on Monte Carlo sampling, simulates the impact of the current task scheduling decision on the execution of several future tasks, and evaluates the long-term effect of the scheduling strategy.

[0008] Furthermore, the optimized scheduling strategy includes constructing a multi-objective reward function: , in, For a multi-objective reward function, Due to task delay, For the maximum allowable delay, For energy consumption, For maximum allowable energy consumption, Assign a reliability score to the task, with a value ranging from [0, 1]. , and For the corresponding weight coefficients, and satisfying 1.

[0009] Furthermore, the virtual-real comparison monitoring mechanism includes an adaptive early warning threshold setting; the initial value of the adaptive early warning threshold is set based on the percentile of the prediction error of the digital twin model, and the current threshold is dynamically adjusted according to the standard deviation of the system's historical performance data, while setting upper and lower limits for the threshold to prevent excessive fluctuations.

[0010] Furthermore, the dynamic compensation mechanism includes: a strategy library of three types of compensation strategies: pre-set resource reallocation, task migration, and computation unloading; a strategy matching module matches the optimal compensation scheme from the strategy library based on the real-time collected node CPU utilization, network latency, and task priority environment characteristics; and an execution and feedback module issues the matched compensation strategy for execution, records the system state changes before and after compensation, and stores them in the experience playback buffer.

[0011] Furthermore, the method also includes establishing a unified heterogeneous resource description framework, abstracting computing units into computing power vectors that include computing power, parallel processing capabilities, and instruction set adaptability, and abstracting network protocols into communication quality indicators that include bandwidth stability, latency jitter rate, and transmission energy consumption coefficient; generating and continuously updating dynamic capability profiles for each edge node, the profiles including computing characteristics, communication characteristics, energy consumption characteristics, and hardware lifespan degradation rate.

[0012] Furthermore, the method also includes: generating a task resource requirement profile through static code scanning and dynamic sampling; constructing an uncertain heterogeneity association model using a graph neural network to predict the execution indicators of tasks under heterogeneous resource combinations; and making scheduling decisions using a fuzzy matching algorithm based on the task requirement profile and dynamic capability profile, and dynamically adjusting the task decomposition granularity according to resource load and network conditions.

[0013] Furthermore, the method also includes: collecting health indicators containing a security threat index through a monitoring agent, setting an adaptive health baseline based on historical data; constructing a threat assessment matrix to quantify risk levels; simulating node replacement, path switching, and functional degradation reconstruction schemes in parallel in a digital twin environment, and triggering adaptive reconstruction decisions based on risk levels; initiating closed-loop optimization after reconstruction, and using deep reinforcement learning to fine-tune the reconstruction strategy.

[0014] One or more technical solutions provided in this application have at least the following technical effects or advantages: High-fidelity collaborative simulation of heterogeneous edge environments is achieved by constructing a privacy-preserving federated digital twin; scheduling strategies that balance latency, energy consumption, and reliability are generated based on multi-objective deep reinforcement learning and look-ahead optimization; accurate execution and rapid fault tolerance of strategies are ensured by virtual-real comparison monitoring and dynamic compensation mechanisms; intelligent matching and adaptive scheduling of resources and tasks are achieved by combining unified resource description and task profiling; the system's elastic recovery capability under fault and attack scenarios is enhanced through anomaly monitoring and pre-simulation reconstruction; and continuous system evolution and rapid deployment in new scenarios are achieved by relying on experience playback and knowledge transfer, thereby comprehensively improving the intelligence, resilience, and scalability of edge computing task scheduling. Attached Figure Description

[0015] Figure 1This is a flowchart of an AI-assisted edge computing task scheduling method according to an embodiment of the present invention. Detailed Implementation

[0016] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0018] Example 1: As Figure 1 As shown, an AI-assisted edge computing task scheduling method is presented.

[0019] S1: By deploying monitoring agents on edge nodes, environmental data including network traffic, computing load, energy consumption, and task queue status is collected; the environmental data is processed using a spatiotemporal joint coding mechanism to generate a unified node status feature vector; Specifically, a lightweight monitoring agent is launched on the edge nodes of each edge domain, with a data collection frequency of 10Hz, continuously collecting four-dimensional environmental data covering network traffic patterns, computing load fluctuations, energy consumption trends, and task queue status. Network traffic patterns focus on uplink bandwidth utilization, downlink bandwidth utilization, and packet loss rate; computing load fluctuations focus on CPU, memory, and GPU utilization; energy consumption trends focus on real-time node power consumption and remaining battery power; and task queue status records the number, type, and priority of tasks.

[0020] The node state feature vector includes: performing temporal interpolation on the collected discrete environmental data using a preset time window to complete it into continuous temporal data; splicing and fusing the spatial attribute features of the node with the continuous temporal data to obtain spatiotemporal fusion features; and inputting the spatiotemporal fusion features into the coding layer for nonlinear mapping to obtain a node state feature vector of a unified dimension.

[0021] Specifically, to preserve the temporal correlation of environmental data and eliminate the fragmentation of discrete data, a spatiotemporal joint coding mechanism is adopted: using a 100ms sliding time window, 10 discrete data points (such as CPU utilization) collected within the window are supplemented into 60 continuous temporal points using a cubic spline interpolation algorithm. The spatial attribute features of the node (including its deployment location's regional coding, physical distance to key gateways, and its degree centrality and proximity centrality indices in the network topology, totaling 4 dimensions) are concatenated and fused with the temporal data. The fused 64-dimensional spatiotemporal features are input into a lightweight fully connected coding layer, which nonlinearly maps them into a 64-dimensional continuous feature vector, which serves as the unified feature vector for the node's state. To address data heterogeneity, a multimodal data fusion channel maps the data to a 64-dimensional unified vector space; simultaneously, an attention-weighted feature selection mechanism dynamically allocates data weights, and the attention weights are normalized using a softmax function to ensure that the sum of all weights is 1.

[0022] S2: Based on node state feature vectors, construct local digital twins in each edge domain; through a federated coordination server, aggregate and update the parameters of the local digital twin model in each edge domain with privacy protection, and generate a global digital twin model. Specifically, based on a unified feature vector, each edge domain's digital twin server constructs a local digital twin using an LSTM neural network. Each twin integrates three core modules: an environment simulator, a performance predictor, and a decision evaluator. The environment simulator can simulate node performance changes under different network traffic conditions, the performance predictor can predict node load change trends within the next 5 minutes, and the decision evaluator can evaluate the node's operating status under different scheduling strategies.

[0023] The global digital twin model is constructed using a federated learning framework, where differential privacy technology is used to perturb the gradient during the aggregation of model parameters; the aggregated model is refined into a lightweight inference model through model distillation technology; and the federated learning framework dynamically adjusts the contribution weight of each edge domain in the model aggregation based on the amount and integrity of the data.

[0024] Specifically, digital twins in each edge domain exchange privacy-preserving model parameters through a federated coordination server. Differential privacy technology is used to perturb the model gradients before uploading to avoid leakage of raw data. The federated coordination server dynamically adjusts aggregation weights based on a contribution evaluation algorithm. For example, the security monitoring area, which accounts for 45% of the data volume and has ≥98% integrity, has a contribution of 0.4, while the office area and public service facility area both have a contribution of 0.3. Model distillation technology refines the aggregated model obtained from federated learning into a lightweight inference model, improving the model inference speed and meeting the real-time requirements of edge nodes.

[0025] Random perturbations are injected into the twin to simulate typical abnormal scenarios such as network latency fluctuations of 10-100ms and a 5% node failure probability. Monte Carlo sampling generates 1000 possible environmental states, which the system uses to evaluate the adaptability of different scheduling strategies. For example, for the scheduling strategy of "assigning security monitoring video analysis tasks to security zone edge node A," in 92% of the 1000 simulated environmental states, the scenario with task completion latency ≤200ms accounts for a significant proportion, indicating excellent robustness of this strategy. To obtain the globally optimal scheduling strategy, a multi-objective reward function is constructed in the deep reinforcement learning engine of the federated coordination server, aiming at minimizing latency, energy consumption, and reliability. , in, For a multi-objective reward function, Due to task delay, For the maximum allowable delay, For energy consumption, For maximum allowable energy consumption, Assign a reliability score to the task, with a value ranging from [0, 1]. , and For the corresponding weight coefficients, and satisfying It can be set based on prior knowledge of the domain, for example, .

[0026] S3: In the virtual environment constructed by the global digital twin model, optimize the scheduling strategy based on the multi-objective deep reinforcement learning algorithm; and distribute the optimized scheduling strategy to the physical edge nodes for execution. The multi-objective deep reinforcement learning algorithm employs a near-end policy optimization algorithm and includes a look-ahead optimization mechanism. The look-ahead optimization mechanism generates environmental state variants based on Monte Carlo sampling, simulates the impact of the current task scheduling decision on the execution of several future tasks, and evaluates the long-term effect of the scheduling strategy.

[0027] Specifically, policy exploration is conducted in the digital twin environment based on the Proximal Policy Optimization (PPO) algorithm. To avoid the adverse impact of short-term decisions on the execution of subsequent tasks, a 5-step look-ahead optimization mechanism is further adopted. While optimizing the current task scheduling decision, the execution effect of the next 5 tasks is fully considered. For example, when scheduling high-priority office tasks at the edge of the office area, the deep reinforcement learning engine selects the edge node with the lowest CPU utilization based on the current state. Using the performance predictor in the digital twin, it simulates the load impact of this node on the execution of subsequent tasks (such as the next 5 tasks) within the next 5 minutes after executing the current task, thereby avoiding long-term performance degradation due to short-term optimization.

[0028] S4: Employs a virtual-real comparison monitoring mechanism to continuously compare the actual operating state of the physical system with the simulated state of the digital twin; when the detected state deviation exceeds the warning threshold, a dynamic compensation mechanism is triggered to match and execute compensation actions from a pre-set compensation strategy library.

[0029] The virtual-real comparison monitoring mechanism includes an adaptive early warning threshold setting; the initial value of the adaptive early warning threshold is set based on the percentile of the prediction error of the digital twin model, and the current threshold is dynamically adjusted according to the standard deviation of the system's historical performance data, while setting upper and lower limits for the threshold to prevent excessive fluctuations.

[0030] Specifically, to ensure consistency between virtual strategies and physical execution results, a virtual-physical comparison monitoring mechanism is initiated. The federated coordination server continuously compares the actual operating status of physical edge nodes with the simulated status of the digital twin, focusing on three core indicators: task completion latency, node energy consumption, and task success rate. The system sets adaptive warning thresholds. Taking task completion latency deviation as an example, its initial threshold ( The initial threshold is set based on the average absolute prediction error of the digital twin during the model validation phase. For example, if the twin's average error for delayed prediction is 15ms, but its 95th percentile error is 20ms, then the initial threshold is set accordingly. This provides a reasonable tolerance buffer for prediction errors.

[0031] The system dynamically adjusts the current effective threshold based on performance fluctuations over the past hour. If the volatility is high, such as the standard deviation being greater than 50% of the initial threshold, then... If the volatility is low, such as standard deviation < 20% of the initial threshold, then If the volatility is within the normal range, then To prevent the threshold from deviating from a reasonable range in extreme cases, upper and lower limits are set for the threshold, for example, .

[0032] The dynamic compensation mechanism includes: a strategy library of three types of compensation strategies: pre-set resource reallocation, task migration, and computation unloading; a strategy matching module matches the optimal compensation scheme from the strategy library based on the real-time collected node CPU utilization, network latency, and task priority environment characteristics; and an execution and feedback module issues the matched compensation strategy for execution, records the changes in system state before and after compensation, and stores them in the experience playback buffer.

[0033] Specifically, when the task completion delay deviation between the physical system and the digital twin exceeds the current dynamic threshold... Upon such an event, the compensation process is immediately triggered. The compensation strategy library pre-configures three types of strategies: resource reallocation, task migration, and compute offloading. The system quickly matches the optimal solution based on the current environmental characteristics: for example, when a node's CPU utilization is detected to be ≥85% and network latency is ≤30ms, the resource reallocation strategy is matched, and the federated coordination server sends an instruction to the corresponding edge node to increase its memory resources from 2GB to 4GB; if network latency is detected to be ≥80ms, the task migration strategy is matched, migrating the current node's low-priority tasks to edge nodes within the same domain with better network conditions; the compute offloading strategy dynamically adjusts task allocation based on real-time load.

[0034] Construct and maintain an experience replay buffer (with a capacity of 10,000 records) to record the execution effect of the compensation strategy. Each experience record is a 5-tuple. ,in, To determine the environmental state before compensation, For the compensation action performed, To obtain the actual new state from the physical system after compensation is performed. As a vector for improving core performance indicators, This is a timestamp. The system initiates a model update cycle every 24 hours. During the update, data is sampled from the buffer using a priority experience replay strategy, and the LSTM digital twin model is iteratively updated by minimizing the following composite loss function: , in, This represents the total loss value. This represents the number of empirical data points sampled in a training batch. For digital twins based on the state before compensation The predicted subsequent state, To determine the new state actually observed from the physical system after implementing the compensation strategy. The square norm of the difference between the predicted state and the true state. This is the performance deviation vector, recording the improvement of key indicators before and after compensation. Let be the square norm of the performance deviation vector, which measures the severity of performance fluctuations in that rule of thumb. The regularization coefficient is greater than 0, and its value ranges from (0, 1]. The larger the value, the more emphasis is placed on correcting scenarios where significant performance deviations have occurred during model updates.

[0035] For multi-domain experience buffers Cluster analysis (such as the DBSCAN algorithm) is performed to identify common state patterns across domains and their corresponding effective scheduling actions. For example, the analysis found that the peak load pattern on weekday mornings frequently occurs in both the office area and the security monitoring domain, and the corresponding efficient actions are task sharding and parallel processing.

[0036] The identified patterns are formalized into interpretable and executable scheduling rules and stored in a knowledge base. The rule format is: IF state features satisfy the condition THEN recommended action set. For example: IF {time period = 9:00-11:00, average CPU > 70%} THEN {action: task fragmentation and parallelism, success rate: 95%}.

[0037] When initializing the system for a new edge domain, the federated coordination server not only provides a pre-trained base model, but also injects a knowledge base of scheduling rules that best matches the characteristics of its current environment through a domain adaptation mechanism. This knowledge serves as prior knowledge to guide the initial strategy or optimization direction of the new domain digital twin model, which can improve the model's convergence speed and significantly shorten the deployment and tuning cycle of the new environment.

[0038] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application achieves high-fidelity dynamic modeling and privacy-preserving collaborative simulation of heterogeneous, distributed edge environments by constructing a collaborative digital twin based on federated learning. It generates a global scheduling strategy that balances latency, energy consumption, and reliability while possessing good robustness through multi-objective deep reinforcement learning and look-ahead optimization mechanisms. Through virtual-real comparison monitoring and a dynamic compensation mechanism based on adaptive thresholds, it ensures the accurate execution and rapid fault tolerance of the scheduling strategy in the physical system. Furthermore, by leveraging experience-replay-driven online evolution of the twin model and cross-domain rule knowledge transfer, it enables autonomous optimization of the system during long-term operation and rapid deployment in new edge scenarios, thereby improving the intelligence level of task scheduling, the overall resilience of the system, and its scalability.

[0039] Example 2: Example 1 implemented edge task scheduling based on digital twins and federated learning, but it still has shortcomings such as inconsistent descriptions of heterogeneous resources and inaccurate task resource matching. This example further supplements the content of Example 1.

[0040] The method also includes establishing a unified heterogeneous resource description framework, abstracting computing units into computing power vectors that include computing power, parallel processing capabilities, and instruction set adaptability, and abstracting network protocols into communication quality indicators that include bandwidth stability, latency jitter rate, and transmission energy consumption coefficient; generating and continuously updating dynamic capability profiles for each edge node, the profiles including computing characteristics, communication characteristics, energy consumption characteristics, and hardware lifespan degradation rate.

[0041] Specifically, a unified heterogeneous resource description framework is established, abstracting CPU, GPU, and FPGA computing units into a three-dimensional computing power vector that includes computing power, parallel processing capability, and instruction set adaptability. Computing power is quantified by floating-point operations per second, parallel processing capability is represented by the product of the number of cores and threads, and instruction set adaptability is the number of dedicated instruction sets supported by the node divided by the total number of dedicated instruction sets required by the task. Dedicated instruction sets include, but are not limited to, SIMD instruction sets, AI acceleration instructions, and encryption instruction sets. If the task has no dedicated instruction requirements, the default value for instruction set adaptability is 1.

[0042] For 5G, Wi-Fi 6, and LoRa network protocols, a unified communication quality index is defined, consisting of bandwidth stability, latency jitter rate, and transmission energy consumption coefficient. Among them, bandwidth stability is the standard deviation of bandwidth fluctuation within 1 hour divided by the average bandwidth during that period. Latency jitter rate is the variance of 100 transmission delays. Transmission energy consumption coefficient is obtained by measuring the energy consumption of different network protocols under different distances and load conditions under standard test tasks and normalizing it to the same benchmark.

[0043] A dynamic capability profile is generated for each edge node, including computing characteristics, communication characteristics, energy consumption characteristics, and hardware lifespan degradation rate. The profile update frequency is consistent with the data acquisition frequency (10Hz).

[0044] Static code scanning is used to parse the task abstract syntax tree (AST) to identify features such as loop structures and matrix operation interfaces to determine the computation mode. Dynamic sampling is performed within 100ms before the task trial run to collect data transmission volume, execution latency, and energy consumption changes, determining communication requirements and energy sensitivity. Energy sensitivity is obtained by dividing the energy consumption change by the execution latency. Based on an experience replay buffer, features, suitable resource types, and execution performance data of historical tasks are extracted to construct a task resource matching knowledge base. The k-nearest neighbor algorithm (k=5) is used to match resource adaptation schemes of similar historical tasks for new tasks, generating an optimal resource requirement profile. For example, for video encoding and decoding tasks, historical data is matched to identify their requirements for GPU parallel processing capabilities and 5G high bandwidth, prioritizing edge nodes with corresponding characteristics.

[0045] A new uncertainty heterogeneity correlation model has been introduced. This model, constructed using a graph neural network (GNN), models the complex relationships between heterogeneous resources and environmental dynamics. Each edge node is treated as a node in a graph, and network connections, communication protocol differences, and task dependencies between nodes are considered as edges. The model input includes the dynamic capability profiles of each node, real-time environmental data, and their correlations; the output is a joint prediction of key performance indicators (KPIs) such as task execution latency and success rate under given resource combinations and uncertainties. The uncertainty heterogeneity correlation model serves as a core component of the performance predictor, working in conjunction with the existing LSTM model. The LSTM model analyzes the temporal trends of resource indicators for individual nodes, while the GNN model learns the complex relationships between heterogeneous resources and environmental uncertainty by abstracting edge nodes as nodes in a graph and network connections, protocol differences, and task dependencies between nodes as edges. The model's input is a graph structure composed of the dynamic capability profiles of each node and real-time environmental data, and the output is a joint prediction of key task execution indicators (such as latency and success rate) under specific heterogeneous resource combinations and uncertainties. This prediction result is fused with the prediction result of the LSTM model to provide a more comprehensive and accurate simulation environment feedback for the deep reinforcement learning decision engine, enabling the deep reinforcement learning decision engine to more reliably evaluate the long-term effects of different scheduling strategies in the virtual environment.

[0046] By leveraging real-time heterogeneous resource data and historical task logs collected through digital twins, a large number of joint simulation scenarios covering different resource loads, network latency ranges, and node availability states are generated using the Monte Carlo method for local pre-training. The training objective is to enable the model to predict key performance indicators for task execution in that environment based on the input real-time resource profile and uncertainty parameters. The local model parameters of each edge node are periodically and securely aggregated and updated through a federated learning framework, thereby constructing a more generalized global prediction model without centralizing the original data.

[0047] Based on a multi-objective reward function framework, a weighted decision tree is established: when the latency jitter rate is <5%, the energy consumption optimization weight is adjusted. Adjusted to 0.4, delay weight Reduced to 0.3, reliability weight Keep the reliability weight at 0.3; when the latency jitter rate is ≥10%, adjust the reliability weight. Increased to 0.45, delay weight Maintain at 0.4, energy consumption weight The threshold is reduced to 0.15; the optimal weight combination is dynamically selected based on real-time environmental characteristics. The latency jitter rate threshold is set based on historical network performance data analysis and can be dynamically adjusted through the federated coordination server.

[0048] The method further includes: generating a task resource requirement profile through static code scanning and dynamic sampling; constructing an uncertain heterogeneity association model using a graph neural network to predict the execution indicators of tasks under heterogeneous resource combinations; and making scheduling decisions using a fuzzy matching algorithm based on the task requirement profile and dynamic capability profile, and dynamically adjusting the task decomposition granularity according to resource load and network conditions.

[0049] Specifically, a heterogeneous resource fuzzy matching algorithm is developed, which makes decisions based on three dimensions: task requirement profile, resource dynamic capability profile, and real-time network conditions. The task requirement profile and resource dynamic capability profile are converted into feature vectors. Cosine similarity is used to calculate the matching degree between the task vector and each available resource vector. Weights are assigned to different feature dimensions according to the current optimization objective, and a weighted similarity score is calculated. The resource with the highest score is selected as the approximate optimal solution. The message passing mechanism of a Generative Neural Network (GNN) is used to learn the association strength between task feature nodes and resource type nodes in the graph, thereby identifying which resource types have historically performed best for tasks with specific characteristics. Every 24 hours, the graph is updated with new scheduling data, and the GNN model is retrained.

[0050] When exact matching (a perfect match between task requirements and resource capabilities) is not feasible, an approximate optimal solution is provided by calculating the similarity of task and resource characteristics (such as the compatibility between task computing mode and hardware architecture, and the compatibility between task communication requirements and network protocol capabilities). Appropriate combinations of computing architecture and communication paths are allocated to each task. For example, GPU nodes are prioritized for matrix operation tasks, paired with 5G communication paths; CPU nodes are allocated for lightweight stream processing tasks, using Wi-Fi 6 communication; and LoRa protocol and edge gateway CPU resources are adapted for sensor data acquisition tasks with low bandwidth requirements.

[0051] The task decomposition granularity is dynamically adjusted based on the heterogeneous resource distribution characteristics: when computing resources are abundant (overall resource utilization <40%), fine-grained decomposition is adopted, splitting the task into 20-30 sub-tasks to improve parallel execution efficiency; when communication is limited (network bandwidth utilization ≥80%), coarse-grained decomposition is adopted, splitting the task into 3-5 sub-tasks to reduce data transmission overhead; when resource load is balanced (utilization 40%-80%), medium-grained decomposition is adopted, splitting the task into 8-12 sub-tasks to balance parallel efficiency and transmission cost.

[0052] Establish a joint optimization framework for resource allocation, task scheduling, and data transmission, and design a collaborative optimization objective function: , in, To optimize the objective function, This represents the number of tasks completed. The total time of the scheduling cycle. For average delay, The required latency threshold for the application. Total energy consumption, The periodic performance cost budget set for the system, , and These are the corresponding weight coefficients, and their sum is 1. Iterative optimization is performed using the Alternating Direction Multiplier Method (ADMM), updating the optimization results every 100ms to find the Pareto optimal solution set.

[0053] The task execution environment is encapsulated based on containerization technology, supporting rapid task transfer between CPU / GPU / FPGA nodes; a fault prediction module is added to the digital twin, using the gradient boosting tree (GBT) model to analyze node load trends, energy consumption anomalies and other characteristics, and identify potential failure nodes 3 minutes in advance; a heterogeneous backup resource pool is established to allocate suitable backup nodes for critical tasks, ensuring that tasks can switch to backup resources within 50ms when a node fails.

[0054] A heterogeneous scheduling knowledge graph is constructed, with task characteristics, resource types, environmental states, scheduling strategies, and execution effects as nodes to establish multi-dimensional relationships. The GNN learns the deep relationship between task characteristics and resource requirements to discover the optimal scheduling mode in a heterogeneous environment, and the knowledge graph is updated every 24 hours.

[0055] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application employs a unified heterogeneous resource description framework and dynamic capability profile, achieving standardized modeling and real-time perception of diverse heterogeneous resources in edge computing environments, thus improving the accuracy of resource description and system compatibility. Through a task profiling method combining static code analysis and dynamic sampling, and an uncertain heterogeneous association model based on graph neural networks, it achieves precise matching and joint prediction between task requirements and resource capabilities, enhancing the environmental adaptability and reliability of scheduling decisions. Leveraging collaborative training and knowledge transfer mechanisms within a federated learning framework, it realizes privacy-preserving collaborative optimization and model generalization for cross-domain heterogeneous resources. Through dynamic weight adjustment, fuzzy matching algorithms, and adaptive task granularity decomposition strategies, it achieves intelligent resource adaptation and improved scheduling efficiency under multi-objective constraints. Combined with containerized rapid migration, fault prediction, and heterogeneous backup mechanisms, it constructs a highly available, low-latency edge task scheduling system, improving the overall performance, resilience, and maintainability of the system in complex heterogeneous environments.

[0056] Example 3: Example 2 achieved standardized modeling and precise task scheduling for heterogeneous resources, but still suffers from insufficient system anomaly recovery and security protection capabilities. This example further supplements and explains the content of Example 2.

[0057] By monitoring agents to collect health indicators including security threat indices, an adaptive health baseline is set based on historical data; a threat assessment matrix is ​​constructed to quantify risk levels; in the digital twin environment, parallel simulations of node replacement, path switching, and functional degradation refactoring schemes are performed, and adaptive refactoring decisions are triggered based on risk levels; after refactoring, closed-loop optimization is initiated, and deep reinforcement learning is used to fine-tune the refactoring strategy.

[0058] Specifically, a lightweight monitoring agent on each edge node collects four types of health indicators in real time at a frequency of 10Hz: node liveness status, resource utilization, network connection quality, and security threat index. The lightweight monitoring agent is deployed as a Docker container, with memory usage controlled to within 50MB. The security threat index is calculated using an anomaly detection model based on a one-dimensional convolutional neural network (1D-CNN). The model's input is a feature vector of network traffic over the past 10 seconds, including: connections per second, average packet size, SYN packet ratio, frequency of abnormal port access, and access time distribution entropy. After pre-training in the cloud, the model is distributed to each node via a federated coordination server and incrementally fine-tuned every 7 days based on local traffic data. Local differential privacy protection technology is used during fine-tuning to ensure that traffic characteristics are not leaked.

[0059] The node's liveness status is determined through a periodic heartbeat mechanism: the agent sends a heartbeat probe every second, and if the target node fails to respond three times consecutively, it is marked as inactive. Resource utilization metrics include real-time CPU utilization, memory utilization, and current bandwidth utilization. Network connection quality is quantified through high-frequency probing: the agent sends a set of probe data packets every 100 milliseconds, and calculates the average communication latency and packet loss rate accordingly. The security threat index is based on real-time traffic characteristic analysis, outputting a normalized score within the range [0,1] through a threat assessment model to dynamically reflect the node's security risk level. The threat assessment model is constructed using an ensemble learning method and includes three sub-models: a baseline detection module based on statistical rules, an anomaly detection module based on isolated forests, and a temporal anomaly detection module based on LSTM. The outputs of the three sub-models are weighted and voted on (with weights of 0.4, 0.3, and 0.3, respectively) to synthesize the final security threat index score.

[0060] Based on historical data from the past 72 hours, a sliding window (window size 1 hour, step size 10 minutes) was used to calculate the statistical characteristics of each indicator: For continuous indicators (such as CPU utilization), the mean of the data within the window was calculated ( ) and standard deviation ( The normal range is defined as follows: For discrete indicators (such as heartbeat status), the proportion of normal responses within the calculated window is used; if the proportion is below 95%, an alert is triggered. The baseline is automatically updated every 6 hours to avoid misjudgments due to environmental changes. For indicators with bounded ranges, such as CPU utilization and memory usage (which have inherently bounded ranges of 0%-100%), the baseline range should intersect with the physical boundary.

[0061] Construct a threat assessment matrix, defining two assessment dimensions: anomaly type and impact scope. Anomaly type includes permanent failures (value 3), transient fluctuations (value 1), and malicious attacks (value 5); impact scope includes local nodes (value 1), single edge domains (value 3), and cross-domain attacks (value 5). Calculate the risk level: , in, The result is the risk level, and is rounded to the nearest integer, ranging from [1, 5]. Score the severity of the abnormality type. To score the scope of influence, This represents the theoretical maximum score. For example, the priority of a cross-domain malicious attack is 5×5 / 25×5=5.

[0062] In a digital twin environment, three core refactoring schemes are simulated in parallel: node replacement, path switching, and function degradation. The node replacement scheme simulates selecting and enabling nodes from the resource pool, recording task migration time, resource usage changes, and the stability performance of the new nodes. The path switching scheme simulates switching to a backup network path, measuring path switching latency, bandwidth changes, and data transmission stability. The function degradation scheme simulates tiered suspension of non-core services (from low-priority logging to medium-priority backup tasks), assessing the performance loss of core services and the integrity of system functionality.

[0063] After each candidate solution is run in various environmental variations, an evaluation report is generated, containing multi-dimensional indicators including recovery timeliness, resource economy, service availability, and solution robustness scores. Recovery timeliness indicators include recovery times at the 50th, 90th, and 99th percentiles, reflecting recovery speed at different confidence levels. Resource economy indicators record the proportion of additional CPU, memory, and bandwidth resources consumed during solution execution. Service availability indicators quantify the increase in latency and changes in success rate of core services (such as security video analytics) during reconstruction. Solution robustness scores calculate the proportion of stable performance across different environmental variations. The evaluation report is differentiated based on risk level: reports for high-risk scenarios highlight worst-case performance indicators, while reports for low-risk scenarios focus more on resource consumption and long-term stability.

[0064] Based on the assessment report and risk level, the system constructs an adaptive reconstruction decision-making mechanism. When the node replacement scheme is activated, a backup node with matching capabilities is selected from the heterogeneous backup resource pool, and the switchover is completed through a containerized task migration mechanism; when When the task traffic is redirected to a backup network path with lower latency and higher stability, the switching latency is controlled within 50ms. When necessary, a functional degradation plan is initiated, suspending non-core services in order of priority from low to high to ensure that the latency increase of core tasks does not exceed 20%. All refactoring decisions are pre-run and verified in a digital twin environment, and their stability under 1000 random perturbations is evaluated through Monte Carlo simulation. The random perturbations are generated based on the following preset probability distribution models: network latency fluctuations follow a normal distribution with a mean of 50ms and a standard deviation of 15ms; the probability of instantaneous node failures follows a Poisson distribution with a parameter of 0.01; and the occurrence of malicious attack events follows a Poisson distribution with a parameter of 0.005. Based on these distributions, the system generates 1000 independent environmental parameter variations to comprehensively evaluate the robustness of the refactoring scheme. After confirmation, the federated coordination server issues the refactoring instructions to the physical edge nodes for execution, while simultaneously updating the state models of the digital twins in each domain.

[0065] After refactoring, a continuous monitoring and closed-loop optimization process is initiated. The state deviation between physical nodes and the digital twin is compared in real time. If the deviation of core indicators (such as task latency and node survival rate) exceeds a dynamic threshold, a secondary refactoring or policy rollback is triggered. The environmental characteristics, decision actions, and actual effects of each refactoring event are recorded, forming a triplet of event, action, and result experience data, which is stored in a cross-domain experience buffer. Every 12 hours, based on recent experience data, deep reinforcement learning is used to fine-tune the refactoring strategy model, optimizing its decision accuracy and response speed in complex and abnormal scenarios. The deep reinforcement learning fine-tuning process uses the Proximal Policy Optimization (PPO) algorithm. Its state space S is defined as the four health indicator vectors at the current moment, the calculated risk level, and the system load state; the action space A is defined as the selection of three refactoring schemes and their key parameters; the reward function R is designed as follows: , Where R is the reward function, For recovery time, As a time threshold, For additional resource consumption, For resource budgeting, For core service success rate, , and For the corresponding weights, and satisfying The goal of fine-tuning is to maximize long-term cumulative rewards.

[0066] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application achieves intelligent quantitative identification and graded early warning of abnormal risks by establishing an adaptive health baseline and multi-dimensional threat assessment matrix based on historical data; it realizes the pre-evaluation and optimization selection of elastic recovery strategies by simulating three types of reconstruction schemes—node replacement, path switching, and function degradation—in a digital twin environment, and by combining the Monte Carlo method to verify robustness under various random perturbations; and it realizes the adaptive elastic recovery and continuous enhancement of the system under various anomaly and attack scenarios by constructing an adaptive reconstruction decision mechanism based on risk level, and closed-loop optimization and experience accumulation based on deep reinforcement learning, thereby improving the availability and security of the edge computing system in highly dynamic and multi-threat environments.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI-assisted edge computing task scheduling method, characterized in that, include: S1: By deploying monitoring agents on edge nodes, environmental data including network traffic, computing load, energy consumption, and task queue status is collected; the environmental data is processed using a spatiotemporal joint coding mechanism to generate a unified node status feature vector; S2: Based on node state feature vectors, construct local digital twins in each edge domain; through a federated coordination server, aggregate and update the parameters of the local digital twin model in each edge domain with privacy protection, and generate a global digital twin model. S3: In the virtual environment constructed by the global digital twin model, optimize the scheduling strategy based on the multi-objective deep reinforcement learning algorithm; and distribute the optimized scheduling strategy to the physical edge nodes for execution. S4: Employs a virtual-real comparison monitoring mechanism to continuously compare the actual operating state of the physical system with the simulated state of the digital twin; when the detected state deviation exceeds the warning threshold, a dynamic compensation mechanism is triggered to match and execute compensation actions from a pre-set compensation strategy library.

2. The AI-assisted edge computing task scheduling method as described in claim 1, characterized in that, The node state feature vector includes: performing temporal interpolation on the collected discrete environmental data using a preset time window to complete it into continuous temporal data; splicing and fusing the spatial attribute features of the node with the continuous temporal data to obtain spatiotemporal fusion features; and inputting the spatiotemporal fusion features into the coding layer for nonlinear mapping to obtain a node state feature vector of a unified dimension.

3. The AI-assisted edge computing task scheduling method as described in claim 1, characterized in that, The global digital twin model is constructed using a federated learning framework, where differential privacy technology is used to perturb the gradient during the aggregation of model parameters; the aggregated model is refined into a lightweight inference model through model distillation technology; and the federated learning framework dynamically adjusts the contribution weight of each edge domain in the model aggregation based on the amount and integrity of the data.

4. The AI-assisted edge computing task scheduling method as described in claim 1, characterized in that, The multi-objective deep reinforcement learning algorithm employs a near-end policy optimization algorithm and includes a look-ahead optimization mechanism. The look-ahead optimization mechanism generates environmental state variants based on Monte Carlo sampling, simulates the impact of the current task scheduling decision on the execution of several future tasks, and evaluates the long-term effect of the scheduling strategy.

5. The AI-assisted edge computing task scheduling method as described in claim 1, characterized in that, The optimized scheduling strategy includes constructing a multi-objective reward function: ， in, For a multi-objective reward function, Due to task delay, For the maximum allowable delay, For energy consumption, For maximum allowable energy consumption, Assign a reliability score to the task, with a value ranging from [0, 1]. , and For the corresponding weight coefficients, and satisfying 1.

6. The AI-assisted edge computing task scheduling method as described in claim 1, characterized in that, The virtual-real comparison monitoring mechanism includes an adaptive early warning threshold setting; the initial value of the adaptive early warning threshold is set based on the percentile of the prediction error of the digital twin model, and the current threshold is dynamically adjusted according to the standard deviation of the system's historical performance data, while setting upper and lower limits for the threshold to prevent excessive fluctuations.

7. The AI-assisted edge computing task scheduling method as described in claim 1, characterized in that, The dynamic compensation mechanism includes: a strategy library of three types of compensation strategies: pre-set resource reallocation, task migration, and computation unloading; a strategy matching module matches the optimal compensation scheme from the strategy library based on the real-time collected node CPU utilization, network latency, and task priority environment characteristics; and an execution and feedback module issues the matched compensation strategy for execution, records the changes in system state before and after compensation, and stores them in the experience playback buffer.

8. The AI-assisted edge computing task scheduling method as described in claim 1, characterized in that, The method also includes establishing a unified heterogeneous resource description framework, abstracting computing units into computing power vectors that include computing power, parallel processing capabilities, and instruction set adaptability, and abstracting network protocols into communication quality indicators that include bandwidth stability, latency jitter rate, and transmission energy consumption coefficient; generating and continuously updating dynamic capability profiles for each edge node, the profiles including computing characteristics, communication characteristics, energy consumption characteristics, and hardware lifespan degradation rate.

9. The AI-assisted edge computing task scheduling method as described in claim 8, characterized in that, The method further includes: generating a task resource requirement profile through static code scanning and dynamic sampling; constructing an uncertain heterogeneity association model using a graph neural network to predict the execution indicators of tasks under heterogeneous resource combinations; and making scheduling decisions using a fuzzy matching algorithm based on the task requirement profile and dynamic capability profile, and dynamically adjusting the task decomposition granularity according to resource load and network conditions.

10. The AI-assisted edge computing task scheduling method as described in claim 1, characterized in that, The method further includes: collecting health indicators containing security threat indices through monitoring agents, setting an adaptive health baseline based on historical data; constructing a threat assessment matrix to quantify risk levels; simulating node replacement, path switching, and functional degradation reconstruction schemes in parallel in a digital twin environment, and triggering adaptive reconstruction decisions based on risk levels; initiating closed-loop optimization after reconstruction, and using deep reinforcement learning to fine-tune the reconstruction strategy.

Images