An AI intelligent gateway edge computing resource dynamic scheduling method and system

Abstract

The application relates to an AI intelligent gateway edge computing resource dynamic scheduling method and system. The method comprises the following steps: performing semantic decoupling and feature extraction on task descriptions in AI tasks to generate a task semantic unit set and a semantic feature vector; collecting real-time performance data of edge nodes to generate a dynamic resource portrait vector and a resource opportunity window prediction sequence; based on the above data, a collaborative decision model is used to generate an executable scheduling scheme containing a semantic unit and edge node binding relationship, a calculation accuracy parameter and an expected execution time window; an instruction set is executed and runtime performance data is collected to update a timing prediction model and a collaborative decision model. Through the closed-loop optimization of task semantic decoupling and feature extraction, resource dynamic sensing and prediction and the collaborative decision model, the accuracy and adaptability of edge computing resource scheduling are improved, the low delay, high precision and high reliability requirements of multi-modal AI tasks are better met, and the large-scale application of edge intelligence is promoted.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a method and system for dynamic scheduling of edge computing resources for AI smart gateways. Background Technology

[0002] With the deep integration and large-scale application of IoT and AI technologies, multimodal AI tasks generated by terminal devices, such as object detection, sentiment analysis, and temporal anomaly prediction, are experiencing explosive growth. These tasks typically place stringent demands on end-to-end latency, inference accuracy, and operational reliability. To address the high latency issues caused by long data transmission distances and network congestion in cloud computing, edge computing architecture has emerged. By deploying computing resources at edge nodes close to terminal devices, localized and rapid processing of AI tasks can be achieved. The AI ​​smart gateway, as the core hub of the edge computing system, is responsible for task access, resource management, and scheduling decisions; its scheduling performance directly determines the service quality of the entire edge system.

[0003] Specifically, existing edge computing resource scheduling methods can be mainly divided into two categories: traditional rule-based scheduling and preliminary intelligent scheduling. Traditional rule-based scheduling methods schedule tasks as a whole based on preset static strategies such as round-robin, load balancing, and proximity allocation. They do not consider the differences in the internal semantic structure of AI tasks, such as computationally intensive, communication-intensive, and precision-sensitive characteristics, nor do they fully perceive the dynamic changes in edge node resources. As a result, scheduling decisions are coarse and it is difficult to match the differentiated needs of multimodal tasks. While preliminary intelligent scheduling methods incorporate techniques such as reinforcement learning and multi-agent collaboration, they still have significant drawbacks: First, they treat AI tasks as indivisible black boxes, failing to semantically decouple and extract features from tasks, thus hindering refined matching between tasks and resources. Second, they rely solely on real-time performance data from edge nodes for resource status assessment, lacking the ability to predict resource availability in the short term, leading to short-sighted scheduling decisions and increasing the risk of subsequent resource conflicts and task blockages. Third, the collaborative decision-making mechanism is poorly designed, failing to effectively address resource competition and collaborative incentives among multiple agents, and lacking a closed-loop optimization mechanism based on actual operational data, making it difficult for the scheduling model to adapt to the dynamic changes in resource status and task load in edge environments.

[0004] Furthermore, in real-world applications, edge nodes typically possess heterogeneous computing capabilities (such as CPUs, GPUs, and NPUs), with their available computing power, memory, and network bandwidth fluctuating dynamically over time. High-concurrency multimodal AI tasks exhibit significantly different requirements for computational intensity, memory, and communication overhead, and are subject to clear upper and lower limits on latency and accuracy. However, existing scheduling methods cannot establish a precise match between task demands and resource supply, making it difficult to simultaneously meet the low-latency, high-precision, and high-reliability requirements of multimodal AI tasks in dynamically changing edge environments. This becomes a core bottleneck restricting the large-scale application of edge intelligence. Summary of the Invention

[0005] This application provides a method and system for dynamic scheduling of edge computing resources for AI smart gateways, in order to improve the accuracy and flexibility of resource scheduling, enhance the overall performance and adaptability of the system, meet the low latency, high precision and high reliability requirements of multimodal AI tasks in dynamic edge environments, and thus promote the large-scale application of edge intelligence.

[0006] Firstly, this application provides a method for dynamic scheduling of edge computing resources for AI smart gateways, including: The system acquires AI tasks containing task descriptions, performs semantic decoupling and feature extraction on the task descriptions, generates corresponding task semantic unit sets and corresponding semantic feature vectors; collects real-time performance data of each edge node, quantifies resource status based on real-time performance data, generates dynamic resource profile vectors for each edge node, and uses a time-series prediction model to perform short-term prediction based on real-time performance data to generate resource opportunity window prediction sequences. Based on the task semantic unit set, semantic feature vector, dynamic resource profile vector, and resource opportunity window prediction sequence, a collaborative decision-making model is used to make decisions and generate an executable scheduling scheme. The executable scheduling scheme includes the binding relationship between each semantic unit and the target edge node, the calculation accuracy parameters, and the expected execution time window. Based on the executable scheduling scheme, an instruction set including computation and communication instructions is generated and executed. Runtime performance data is collected during the execution process, and a runtime performance feedback data packet is generated based on the runtime performance data. Based on the runtime performance feedback data packet, the parameters of the time series prediction model and the collaborative decision-making model are updated to obtain the updated time series prediction model and the updated collaborative decision-making model.

[0007] In one embodiment, the task description is semantically decoupled and feature extracted to generate a corresponding set of task semantic units and a corresponding semantic feature vector; real-time performance data of each edge node is collected, resource status is quantified based on the real-time performance data, dynamic resource profile vectors of each edge node are generated, and a time-series prediction model is used to perform short-term prediction based on the real-time performance data to generate a resource opportunity window prediction sequence, including: Retrieve a predefined model knowledge base, which stores computational graph models corresponding to various task types. Each computational graph model contains a set of operators and the data dependencies between operators. Based on the task type information in the task description, load the corresponding computational graph model from the predefined model knowledge base. The computation patterns and data dependencies of each operator in the computation graph model are analyzed. Operators that are continuous and have the same computation pattern are aggregated to generate multiple semantic units that constitute the AI ​​task. The semantic units are then integrated to obtain a task semantic unit set. For each semantic unit in the task semantic unit set, feature quantization is performed based on the corresponding aggregation operator type, data flow scale and topological position in the computation graph model, and combined with the dependency information in the task description, to generate a corresponding semantic feature vector. The semantic feature vector includes computational intensity, memory requirements, communication overhead, latency sensitivity, accuracy sensitivity and dependency identifier. The monitoring agents deployed on each edge node periodically collect raw performance monitoring data of each edge node according to a preset cycle. The raw performance monitoring data includes computing resource utilization, memory usage, network transmission rate and energy consumption data. Based on the original performance monitoring data, sliding window statistics are performed to calculate the real-time available resources and performance status parameters of each edge node in terms of computing, memory, network and energy dimensions, and to generate a dynamic resource profile vector of each edge node. The historical dynamic resource profile vector and time-series prediction model of each edge node are obtained. The dynamic resource profile vector corresponding to the real-time performance data and the historical dynamic resource profile vector are input into the time-series prediction model for forward inference processing to generate a resource opportunity window prediction sequence for each edge node at multiple consecutive time points within a predetermined future time interval. Each prediction element in the resource opportunity window prediction sequence includes the predicted value of available computing power, the predicted value of network quality, and the prediction confidence.

[0008] In one embodiment, an executable scheduling scheme is generated by making decisions through a collaborative decision-making model based on a set of task semantic units, semantic feature vectors, dynamic resource profile vectors, and resource opportunity window prediction sequences, including: For each semantic unit in the task semantic unit set, each target edge node and its corresponding execution start time point are traversed. Based on the features in the semantic feature vector, the dynamic resource profile vector corresponding to the target edge node, and the resource opportunity window prediction sequence corresponding to the target edge node, multi-dimensional adaptation degree calculation is performed to generate a composite decision state feature containing the demand-supply matching coefficient. Each edge node is virtualized as an intelligent agent, and a collaborative decision-making model based on a policy neural network is configured for each intelligent agent. Through each intelligent agent, based on the dynamic resource profile vector, resource opportunity window prediction sequence and globally shared composite decision state features, the collaborative decision-making model performs reasoning processing to generate action proposals that include resource pricing and collaborative resource requirements. The central coordination node pre-installed in the AI ​​smart gateway is invoked to collect action proposals from each intelligent agent. Conflict detection processing is performed on each action proposal to identify the set of conflicting action proposals and the set of conflict-free action proposals. Priority calculation processing is performed on each conflict-free action proposal in the set of conflict-free action proposals to obtain the sorted set of conflict-free action proposals. Based on a preset global optimization objective, and combined with a sorted set of conflict-free action proposals, decisions are made regarding the relationship between each semantic unit and the target edge node to obtain the binding relationship. Based on the dependency identifier in the semantic feature vector, the execution time window order of each semantic unit is calibrated to obtain the expected execution time window. According to the precision sensitivity in the semantic feature vector and the computing power of the target edge node, corresponding computational precision parameters are configured. The binding relationship, expected execution time window, and computational precision parameters are integrated to form an executable scheduling scheme.

[0009] In one embodiment, based on an executable scheduling scheme, an instruction set including computation and communication instructions is generated and executed; runtime performance data during execution is collected; a runtime performance feedback data packet is generated based on the runtime performance data; and based on the runtime performance feedback data packet, the parameters of the time-series prediction model and the collaborative decision-making model are updated to obtain the updated time-series prediction model and the updated collaborative decision-making model, including: The AI ​​smart gateway's instruction distribution module parses executable scheduling schemes and generates corresponding computation instructions for each target edge node, including semantic unit identifiers, computation precision parameters, and data addresses. For edge node pairs with data interaction requirements, communication instructions containing data transmission source nodes, target nodes, and transmission priorities are generated. The computation instructions and communication instructions are integrated to obtain an instruction set. The instruction set is distributed to the corresponding edge nodes, and the corresponding instructions in the instruction set are executed by each edge node. The runtime performance data of the instruction execution process is collected by the monitoring module deployed locally on each edge node. The runtime performance data includes actual computing time, peak memory usage, network transmission time, actual energy consumption, and execution result accuracy. Based on runtime performance data, the corresponding semantic unit identifier, target edge node identifier, and scheduling scheme entries are associated, and the data is encapsulated in a structured manner to generate a runtime performance feedback data packet. The actual performance parameters in the runtime performance feedback data packet are compared with the corresponding prediction parameters in the resource opportunity window prediction sequence to calculate the resource prediction error. Based on the resource prediction error, the parameters of the time series prediction model are updated through the backpropagation algorithm to obtain the updated time series prediction model. For the scheduling process of each semantic unit, the composite decision state features corresponding to the semantic unit, the action proposals corresponding to the agent executing the semantic unit, and the reward values ​​calculated based on the actual performance data in the runtime performance feedback data packet are extracted. The composite decision state features, action proposals, and reward values ​​are integrated to obtain the scheduling decision instance corresponding to each semantic unit. An experience replay buffer is constructed, and each scheduling decision instance is stored in the experience replay buffer. A preset number of scheduling decision instances are randomly sampled from the experience replay buffer, and the parameters of the collaborative decision model are updated by the policy gradient algorithm to obtain the updated collaborative decision model.

[0010] In one embodiment, priority calculation is performed on each conflict-free action proposal in the conflict-free action proposal set to obtain a sorted conflict-free action proposal set, including: For each conflict-free action proposal in the conflict-free action proposal set, extract the latency sensitivity and precision sensitivity values ​​of the semantic units corresponding to each conflict-free action proposal in the semantic feature vector, and extract the resource quotation value in each conflict-free action proposal. Based on the preset numerical normalization algorithm, the latency sensitivity value, accuracy sensitivity value and resource price value are standardized respectively to obtain normalized latency sensitivity, normalized accuracy sensitivity and normalized resource price in the same numerical range. Based on the current overall load status of edge nodes and the distribution of task types to be scheduled, set dynamic weight coefficients for normalized latency sensitivity, normalized accuracy sensitivity, and normalized resource pricing; the sum of the dynamic weight coefficients is 1. The normalized latency sensitivity, normalized accuracy sensitivity, and normalized resource bid are multiplied by their respective dynamic weight coefficients and then summed to obtain the comprehensive priority score for each conflict-free action proposal. The conflict-free action proposals in the conflict-free action proposal set are sorted according to their comprehensive priority scores from high to low, resulting in a sorted set of conflict-free action proposals.

[0011] In one embodiment, the reward value is calculated using the following formula: ; in, For the first In this scheduling, semantic units Assigned to edge nodes The aforementioned reward value; , and All are target weight coefficients, and ; For the first The actual execution accuracy of this scheduling; For the first The preset precision threshold for this scheduling; The actual execution delay of the t-th scheduling; Execute in each of the aforementioned edge nodes The minimum theoretical delay; Execute in each of the aforementioned edge nodes The maximum theoretical delay; The actual energy consumption of the t-th scheduling; For edge nodes The energy consumption threshold; This is the penalty coefficient for violations, and 1; For indicator functions, This indicates that there is a defect. =1, otherwise =0.

[0012] Secondly, this application also provides an AI smart gateway edge computing resource dynamic scheduling system, including: The task parsing and resource assessment module is used to acquire AI tasks containing task descriptions, perform semantic decoupling and feature extraction on the task descriptions, generate corresponding task semantic unit sets and corresponding semantic feature vectors; collect real-time performance data of each edge node, quantify resource status based on real-time performance data, generate dynamic resource profile vectors of each edge node, and use a time series prediction model to perform short-term prediction based on real-time performance data to generate resource opportunity window prediction sequences. The collaborative decision-making and scheduling scheme generation module is used to make decisions based on the task semantic unit set, semantic feature vector, dynamic resource profile vector and resource opportunity window prediction sequence through a collaborative decision-making model to generate an executable scheduling scheme. The executable scheduling scheme includes the binding relationship between each semantic unit and the target edge node, the calculation accuracy parameters and the expected execution time window. The scheduling execution and feedback update module is used to generate and execute an instruction set including computation and communication instructions based on an executable scheduling scheme, collect runtime performance data during the execution process, generate runtime performance feedback data packets based on the runtime performance data, and update the parameters of the time series prediction model and the collaborative decision-making model based on the runtime performance feedback data packets to obtain the updated time series prediction model and the updated collaborative decision-making model.

[0013] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0014] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0015] The aforementioned AI intelligent gateway edge computing resource dynamic scheduling method and system first performs semantic decoupling and feature extraction on the AI ​​task description, and generates dynamic resource profile vectors and resource opportunity window prediction sequences by combining real-time performance data of edge nodes. This solves the problems of traditional black-box task scheduling and short-sighted resource status assessment, improving the accuracy and foresight of task-resource matching. Secondly, based on multi-dimensional task and resource data, an executable scheduling scheme is generated through a collaborative decision-making model, achieving refined adaptation of task requirements and resource supply, and improving the scientific nature and dynamic adaptability of scheduling decisions. Furthermore, an instruction set is generated based on the scheduling scheme, and runtime performance data is collected, ensuring the orderly execution of scheduling and improving task execution reliability and data support capabilities. Finally, a closed-loop optimization mechanism is constructed through runtime performance feedback data packet update timing prediction models and collaborative decision-making models, effectively solving the problem that scheduling models are difficult to adapt to dynamic edge environments, and improving the system's continuous optimization capabilities and service quality stability. Attached Figure Description

[0016] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart of a dynamic scheduling method for edge computing resources of an AI smart gateway is provided as an exemplary embodiment of the present invention; Figure 2 A flowchart illustrating a method for obtaining an updated time-series prediction model and an updated collaborative decision-making model, provided as an exemplary embodiment of the present invention; Figure 3 This is a schematic diagram of an AI smart gateway edge computing resource dynamic scheduling system structure provided as an exemplary embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] In one embodiment, such as Figure 1 As shown, a method for dynamic scheduling of edge computing resources for an AI smart gateway is provided. This embodiment illustrates the method by applying it to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps: S101: Obtain AI tasks containing task descriptions, perform semantic decoupling and feature extraction on the task descriptions respectively, generate corresponding task semantic unit sets and corresponding semantic feature vectors; collect real-time performance data of each edge node, quantify resource status based on real-time performance data, generate dynamic resource profile vectors of each edge node, and use a time-series prediction model to perform short-term prediction based on real-time performance data to generate resource opportunity window prediction sequences.

[0020] Specifically, the internal execution logic of multimodal AI tasks such as object detection and sentiment analysis exhibits significant heterogeneity. Different computational stages have vastly different requirements for computing power, memory, and communication. Treating the task as a whole as a scheduling object makes it impossible to achieve a precise match between resources and task requirements. Therefore, complex tasks can be decomposed into more granular and clearly defined scheduling units through semantic decoupling. Subsequently, each unit can be assigned a quantifiable requirement identifier through feature extraction. For example, core information in the task description, such as task type, input data format, and QoS constraints, can be analyzed first. Based on the general execution logic of AI tasks, the boundaries of computational stages within the task can be identified. Computational stages that are continuous and have similar resource requirements can be aggregated into semantic units, forming a task semantic unit set. For each semantic unit, its computational complexity, peak memory usage, communication overhead, latency sensitivity, and accuracy sensitivity can be quantified to generate a structured semantic feature vector, giving each semantic unit's requirements quantifiable and matchable attributes. Furthermore, since the availability of heterogeneous computing power, memory, network bandwidth, and other resources at edge nodes dynamically changes with task load and device operating status, relying solely on real-time data cannot avoid the short-sightedness of scheduling decisions and can easily lead to subsequent resource conflicts. Therefore, after collecting real-time performance data from each edge node, resource status can be quantified to generate a dynamic resource profile vector for each edge node. This dynamic resource profile vector is a multi-dimensional vector that comprehensively reflects the resource status of the edge node at the current moment. In addition, to predict resource availability in the short term, a time-series prediction model can be used to analyze real-time performance data. Based on historical data and current trends, this model can predict resource changes at edge nodes over a future period and generate a resource opportunity window prediction sequence. This sequence can inform in advance which resources are available within a specific future time window, thus providing forward-looking support for scheduling decisions.

[0021] S102: Based on the task semantic unit set, semantic feature vector, dynamic resource profile vector and resource opportunity window prediction sequence, a collaborative decision-making model is used to make decisions and generate an executable scheduling scheme; the executable scheduling scheme includes the binding relationship between each semantic unit and the target edge node, the calculation accuracy parameters and the expected execution time window.

[0022] Specifically, in edge computing, multiple semantic units and multiple heterogeneous edge nodes form complex many-to-many matching relationships, and there are multiple objective constraints such as latency, accuracy, and energy consumption. However, traditional single-objective optimization or static rules cannot balance local demand and global efficiency. Therefore, we can first incorporate the task semantic unit set, semantic feature vector, dynamic resource profile vector, and resource opportunity window prediction sequence into the decision input space to construct a demand-supply dual-dimensional decision matrix. For each semantic unit, we can also traverse all potential target edge nodes and their corresponding execution start times to calculate multi-dimensional fit. For example, we can use cosine similarity based on demand features such as computational intensity and memory requirements in the semantic feature vector, and perform similarity matching with the corresponding supply features in the dynamic resource profile vector (current supply) and resource opportunity window prediction sequence (future supply) of the target edge node. At the same time, we can introduce penalty terms such as latency constraints (e.g., expected execution time ≤ maximum tolerable latency of the task) and accuracy constraints (e.g., configuration accuracy ≥ minimum accuracy requirement of the task) to form a comprehensive fit score for each combination of semantic unit-edge node-execution time. Subsequently, the collaborative decision-making model can use the aforementioned comprehensive fit score to solve decisions through multi-objective optimization, generating an executable scheduling scheme. This scheme specifies in detail the binding relationship between each semantic unit and the target edge node, which may include specific computing nodes and communication paths. It can also assign computational precision parameters to each task, ensuring that tasks can be executed efficiently while meeting accuracy requirements. Furthermore, the scheme can plan expected execution time windows to ensure that tasks are completed within the specified time.

[0023] S103: Based on the executable scheduling scheme, generate and execute an instruction set including computation instructions and communication instructions, collect runtime performance data during the execution process, generate runtime performance feedback data packets based on the runtime performance data, and update the parameters of the time series prediction model and the collaborative decision-making model based on the runtime performance feedback data packets to obtain the updated time series prediction model and the updated collaborative decision-making model.

[0024] Specifically, based on an executable scheduling scheme, an instruction set containing computation and communication instructions can be generated. The computation instructions specify core information such as the semantic unit identifier, computation precision parameters, and data storage address to be processed by each target edge node, ensuring that nodes execute computations according to predetermined logic. The communication instructions specify the transmission path, data block identifier, and transmission priority for edge node pairs with data interaction needs, guaranteeing data consistency between semantic units. The instruction set is transmitted to the corresponding edge nodes through the distribution module of the AI ​​smart gateway. After receiving the instructions, each node can apply for dedicated resource quotas through its local resource management component and execute computation and data transmission operations as required by the instructions. During execution, the monitoring modules deployed on each edge node can collect runtime performance data in real time throughout the entire instruction execution process, including actual computation time, peak memory usage, network transmission time, actual energy consumption, and execution result accuracy. Based on this data, the corresponding unique semantic unit identifier, target edge node identifier, and scheduling scheme entries can be associated to generate runtime performance feedback data packets, ensuring that the execution results of each scheduling decision are traceable and correlated. Based on this feedback data packet, machine learning algorithms can be used to update the parameters of the time-series prediction model and the collaborative decision-making model. This allows the models to adjust their parameters according to actual operational data, thus better adapting to dynamic changes in the edge environment. The updated time-series prediction model can more accurately predict resource status, and the updated collaborative decision-making model can generate more optimized scheduling schemes. This optimization process ensures continuous system improvement, enhancing the accuracy and adaptability of resource scheduling.

[0025] The aforementioned method first generates a set of task semantic units and semantic feature vectors through semantic decoupling and feature extraction of task descriptions. Simultaneously, it collects real-time performance data from edge nodes to generate dynamic resource profile vectors and resource opportunity window prediction sequences, achieving refined modeling and dynamic perception of tasks and resources, thus enhancing the accuracy and foresight of scheduling decisions. Secondly, based on the task semantic unit set, semantic feature vectors, dynamic resource profile vectors, and resource opportunity window prediction sequences, an executable scheduling scheme is generated through a collaborative decision-making model. This further overcomes the problem of coarse task-resource matching in traditional methods, improving the comprehensiveness and adaptability of scheduling. Finally, by executing instruction sets and collecting runtime performance data, a runtime performance feedback data package is generated to update the parameters of the time-series prediction model and the collaborative decision-making model, enhancing the system's adaptability and improving resource utilization and scheduling reliability.

[0026] In one embodiment, the task description is semantically decoupled and feature extracted to generate a corresponding set of task semantic units and corresponding semantic feature vectors; real-time performance data of each edge node is collected, resource status is quantified based on the real-time performance data, dynamic resource profile vectors of each edge node are generated, and a time-series prediction model is used to perform short-term prediction based on the real-time performance data to generate a resource opportunity window prediction sequence, including: Retrieve a predefined model knowledge base, which stores computational graph models corresponding to various task types. Each computational graph model contains a set of operators and the data dependencies between operators. Based on the task type information in the task description, load the corresponding computational graph model from the predefined model knowledge base. The computation patterns and data dependencies of each operator in the computation graph model are analyzed. Operators that are continuous and have the same computation pattern are aggregated to generate multiple semantic units that constitute the AI ​​task. The semantic units are then integrated to obtain a task semantic unit set. For each semantic unit in the task semantic unit set, feature quantization is performed based on the corresponding aggregation operator type, data flow scale and topological position in the computation graph model, and combined with the dependency information in the task description, to generate a corresponding semantic feature vector. The semantic feature vector includes computational intensity, memory requirements, communication overhead, latency sensitivity, accuracy sensitivity and dependency identifier. The monitoring agents deployed on each edge node periodically collect raw performance monitoring data of each edge node according to a preset cycle. The raw performance monitoring data includes computing resource utilization, memory usage, network transmission rate and energy consumption data. Based on the original performance monitoring data, sliding window statistics are performed to calculate the real-time available resources and performance status parameters of each edge node in terms of computing, memory, network and energy dimensions, and to generate a dynamic resource profile vector of each edge node. The historical dynamic resource profile vector and time-series prediction model of each edge node are obtained. The dynamic resource profile vector corresponding to the real-time performance data and the historical dynamic resource profile vector are input into the time-series prediction model for forward inference processing to generate a resource opportunity window prediction sequence for each edge node at multiple consecutive time points within a predetermined future time interval. Each prediction element in the resource opportunity window prediction sequence includes the predicted value of available computing power, the predicted value of network quality, and the prediction confidence.

[0027] Specifically, the predefined model knowledge base is a structured data storage system built upon common AI task types in edge computing scenarios. It provides standardized computational logic references for different types of AI tasks, avoiding repetitive parsing of task execution processes. The computational graph models corresponding to various task types stored in this knowledge base are formal descriptions of the AI ​​task execution process. The operator set covers various basic computational units required for task execution, including operators with different functions such as arithmetic operations, logical operations, data transmission, and model inference. Each operator contains clearly defined computational rules, input / output data formats, and resource requirement thresholds. The data dependencies between operators are represented through a topological structure, clarifying the source of input data and the recipient of output data for each operator, forming a complete task execution chain. Illustratively, based on the task type information carried in the task description, the corresponding computational graph model can be retrieved and loaded from the predefined model knowledge base through keyword matching or type identifier mapping, ensuring that subsequent decoupling and feature extraction processes accurately match the actual execution logic of the task.

[0028] Specifically, by parsing the loaded computation graph model, the computation modes and data dependencies of each operator can be obtained. The computation mode is a classification identifier based on the operator's operational characteristics, resource consumption type, and execution logic. For example, convolution operations and matrix multiplication can be classified as computationally intensive modes, data copying and network transmission as communication-intensive modes, and threshold judgment and logical branching as control-intensive modes. By traversing the topology of the computation graph, the data flow between operators can be tracked, identifying the pre-dependent and post-dependent operators of each operator, thus recognizing the continuous computation links in the task execution process. Subsequently, operators with no data dependency conflicts and consistent resource requirements can be grouped together. Aggregation processing is performed on continuous operators with the same computation mode, and key information such as the overall input / output interfaces, total computational load, and total data transmission volume of the group of operators must be retained during the aggregation process. Ultimately, multiple semantic units constituting the AI ​​task can be generated. Each semantic unit maintains the independence of its computational logic, and aggregation reduces scheduling granularity and overhead. Integrating all semantic units forms a task semantic unit set.

[0029] Furthermore, based on the type of aggregation operator corresponding to a semantic unit, its core resource requirements can be determined. For example, semantic units corresponding to computationally intensive aggregation operators can have computational power requirements as their core feature. Combined with the data flow scale—that is, the total amount of data input, output, and intermediate variables during the semantic unit's execution—its requirements for memory storage and network transmission can be quantified. Furthermore, based on the semantic unit's topological position in the computational graph model, its importance in the entire task execution chain can be determined. For instance, semantic units located on the critical path (the chain with the longest task execution time) have a more significant impact on the overall task latency. In addition, the dependency information in the task description can be combined to clarify the pre-dependencies and post-associations between the semantic unit and other semantic units. Based on the above, the parameters of each dimension of the generated semantic feature vector can be obtained through quantitative calculation. For example, the total computational load of the aggregation operator can be calculated to characterize the computational intensity, such as using FLOPs (floating-point operations) as a quantitative indicator. The calculation method is as follows: ; in, The computational intensity of a semantic unit, measured in FLOPs (floating-point operations), is used to quantify the overall computational requirements of that semantic unit. Use the index of the operator to traverse all operators contained in the semantic unit; This represents the total number of operators contained in this semantic unit; For the first The number of operations per operator, that is, the number of floating-point operations contained in one execution of a single operator; For the first The execution count of an operator is the total number of times that operator is called during the execution of a semantic unit. Memory requirements can be quantified by the maximum simultaneous storage space occupied by the semantic unit during execution, including the storage requirements for input data, intermediate variables, and output data. Communication overhead can be calculated based on the total amount of data interaction between the semantic unit and other semantic units or external data nodes. Latency sensitivity can be determined by the weighting of the semantic unit's execution delay on the overall task completion time; for example, semantic units on the critical path are given higher weighting values. Accuracy sensitivity can be quantified by the influence coefficient of the semantic unit's computational accuracy on the accuracy of the final task output, where the influence coefficient can be obtained by comparing the correlation between the unit's output error under different accuracy configurations and the overall task error. Dependency identifiers can use a unique encoding form to clearly identify the identifiers of the semantic units that pre-depend on and post-associate with the semantic unit, thereby forming a dependency network between semantic units.

[0030] Specifically, raw performance monitoring data can be collected through monitoring agents deployed on each edge node. These monitoring agents are lightweight software modules embedded in the local system of the edge node, possessing low resource consumption and high real-time data acquisition capabilities. Their deployment can adapt to the hardware architecture of the edge node, such as heterogeneous computing platforms like CPUs, GPUs, and NPUs, and their corresponding operating system types. The monitoring agents can collect data periodically according to a preset cycle, which can be determined based on the characteristic cycle of resource fluctuations on the edge node to ensure timely capture of dynamic changes in resource status. The collected raw performance monitoring data includes core dimensions of the edge node's operating status. Computational resource utilization can be obtained by statistically analyzing the percentage of busy time for computing units (such as CPU cores and GPU stream processors); memory usage can be obtained by reading the used storage space data from the system memory management module; network transmission rate can be calculated based on the ratio of the total amount of data sent and received by the network card within a preset cycle to the cycle duration; and energy consumption data can be collected through the edge node's power management module or external energy sensors, reflecting the energy consumption rate during node operation.

[0031] Subsequently, based on the collected raw performance monitoring data, a sliding window statistical method can be used for data preprocessing and feature extraction. The window size can be set according to the fluctuation frequency of the raw performance data, typically selecting a time length encompassing 3-5 collection cycles. Statistical analysis of the raw data within the window effectively smooths out instantaneous data fluctuations and improves the stability of resource status assessment. During the statistical process, for the computation dimension, the average computing resource utilization, peak utilization, and utilization variance within the window can be calculated. Combined with the node's total computing power, the real-time available computing power is obtained. For the memory dimension, the average memory usage and maximum memory usage within the window can be calculated. Combined with the node's total memory capacity, the real-time available memory amount is determined. For the network dimension, the average network transmission rate, transmission latency, and latency jitter within the window can be statistically analyzed to quantify the real-time performance of network communication. For the energy dimension, the average energy consumption rate and energy consumption stability coefficient within the window can be calculated to reflect the node's energy supply status. By structurally integrating the real-time available resource amount and performance status parameters from each dimension, a dynamic resource profile vector can be formed. This vector can comprehensively and accurately characterize the service capacity and operating status of edge nodes at the current moment.

[0032] Specifically, the historical dynamic resource profile vectors of each edge node are stored in order by timestamp, and an adaptive historical period pruning mechanism is introduced. This mechanism dynamically adjusts the retention length of historical data by monitoring the frequency and amplitude of resource fluctuations in real time. For nodes with drastic resource fluctuations, the historical period is shortened to focus on recent changes, while for nodes with stable resources, the historical period is extended to capture long-term trends. This avoids information redundancy or loss of key trends caused by fixed preset periods, and ensures that the model can accurately learn the time-varying patterns of resource status. The time-series prediction model can adopt a deep learning model architecture suitable for multi-dimensional time-series data. Specifically, it can be a hybrid model of LSTM / GRU and convolutional neural network that introduces a multi-scale spatiotemporal attention mechanism. In this architecture, the convolutional neural network (CNN) layer not only extracts local features, but also assigns dynamic weights to local fluctuation features in different time windows through a multi-scale attention weighting mechanism, giving priority to key local patterns such as resource mutations and load peaks. The LSTM / GRU layer, on the basis of capturing long-term dependencies in time series, adds a dynamic weight adjustment module of the gating unit, which adaptively updates the coefficients of the forget gate and the input gate according to the mutation intensity of resource fluctuations. After sudden fluctuations in resources, it can quickly adjust the memory weights of long-term dependencies, avoiding the lagging response of traditional gating units to sudden changes. Furthermore, the dimension of the input layer of this model is consistent with the dimension of the dynamic resource profile vector, the hidden layer can capture the temporal correlation and nonlinear features of the data through multi-layer neural networks, and the dimension of the output layer corresponds to the dimension of the predicted resource parameters.

[0033] In a schematic representation, the dynamic resource profile vector corresponding to real-time performance data and the historical dynamic resource profile vector are concatenated chronologically using a dynamic temporal alignment and padding strategy. Addressing the issues of non-uniform sampling and occasional missing data in edge node resource data, a context-aware interpolation completion method is employed to complete missing data based on resource fluctuation trends at adjacent time points, generating an equidistant temporal input sequence. This avoids information distortion caused by fixed-length concatenation and is input into a temporal prediction model for forward inference processing. This model learns the changing patterns of resource status in historical temporal data, such as periodic fluctuations, trend changes, and recovery characteristics after sudden fluctuations. It also incorporates a scenario-aware transfer learning module. When an edge node connects to a new business scenario, it quickly adapts using a pre-trained general resource fluctuation model and fine-tunes it with a small amount of local running data, significantly shortening the model's cold start time in heterogeneous scenarios. This allows the model to predict resource status at multiple consecutive time points within a predetermined future time interval, generating a resource opportunity window prediction sequence. Each prediction element in this sequence includes a predicted value for available computing power, a predicted value for network quality, and a prediction confidence level. The available computing power prediction value represents the total amount of computing resources that a node can provide at a corresponding time point. The network quality prediction value can be obtained by quantifying parameters such as network transmission rate, latency, and jitter through weighted summation. The prediction confidence can be calculated based on the statistical results of the model's prediction error on historical data, reflecting the reliability of the prediction value. For example, the prediction confidence can be calculated by comparing the deviation between the model's historical prediction values ​​and the actual observed values ​​and using probabilistic statistical methods to obtain the probability that the prediction value falls within a preset interval near the true value.

[0034] In one embodiment, an executable scheduling scheme is generated by making decisions through a collaborative decision-making model based on a set of task semantic units, semantic feature vectors, dynamic resource profile vectors, and resource opportunity window prediction sequences, including: For each semantic unit in the task semantic unit set, each target edge node and its corresponding execution start time point are traversed. Based on the features in the semantic feature vector, the dynamic resource profile vector corresponding to the target edge node, and the resource opportunity window prediction sequence corresponding to the target edge node, multi-dimensional adaptation degree calculation is performed to generate a composite decision state feature containing the demand-supply matching coefficient. Each edge node is virtualized as an intelligent agent, and a collaborative decision-making model based on a policy neural network is configured for each intelligent agent. Through each intelligent agent, based on the dynamic resource profile vector, resource opportunity window prediction sequence and globally shared composite decision state features, the collaborative decision-making model performs reasoning processing to generate action proposals that include resource pricing and collaborative resource requirements. The central coordination node pre-installed in the AI ​​smart gateway is invoked to collect action proposals from each intelligent agent. Conflict detection processing is performed on each action proposal to identify the set of conflicting action proposals and the set of conflict-free action proposals. Priority calculation processing is performed on each conflict-free action proposal in the set of conflict-free action proposals to obtain the sorted set of conflict-free action proposals. Based on a preset global optimization objective, and combined with a sorted set of conflict-free action proposals, decisions are made regarding the relationship between each semantic unit and the target edge node to obtain the binding relationship. Based on the dependency identifier in the semantic feature vector, the execution time window order of each semantic unit is calibrated to obtain the expected execution time window. According to the precision sensitivity in the semantic feature vector and the computing power of the target edge node, corresponding computational precision parameters are configured. The binding relationship, expected execution time window, and computational precision parameters are integrated to form an executable scheduling scheme.

[0035] Specifically, a precise mapping relationship between task requirements and resource supply can be established first to provide a quantitative basis for subsequent decision-making. For example, for each semantic unit, demand features such as computational intensity, memory requirements, communication overhead, latency sensitivity, precision sensitivity, and dependency identifiers can be extracted from its semantic feature vector. Secondly, for each potential target edge node, current supply features such as available computing power, memory, and network bandwidth in its dynamic resource profile vector can be obtained, as well as future supply features such as predicted available computing power and network quality at the corresponding execution start time point in the resource opportunity window prediction sequence. Based on these features, a weighted summation algorithm can be used to calculate the matching coefficient between demand features and supply features. The weight of different features is determined by the degree of influence of the feature on the execution of the semantic unit; for example, the matching weight of computational intensity and available computing power can be higher than the matching weight of communication overhead and network bandwidth. Illustratively, in addition to calculating the matching coefficient between demand features and supply features using a weighted summation algorithm, the demand-supply matching coefficient can also be calculated using the following formula: ; in, For the first A semantic unit, , The total number of semantic units; For the first Target edge nodes, , This represents the total number of edge nodes. For the first Weight coefficients for resource-related dimensions. These correspond to support for computing, memory, networking, energy, and precision, respectively. ; For semantic units In the Dimensional requirements parameters, such as When calculating strength, This is due to memory requirements; For edge nodes In the Available resource parameters for a dimension, such as At that time, computing power was available. Available memory at that time; For semantic units The latency sensitivity threshold; For edge nodes The expected response time can be calculated based on the resource opportunity window prediction sequence; Let be the penalty coefficient for latency mismatch, and ; This is the communication overhead penalty coefficient. It is positively correlated with the network distance between edge nodes; For semantic units With edge nodes Unit data transfer overhead; Match The matching coefficient ranges from 0 to 1, with values ​​closer to 1 indicating a higher degree of fit. By structurally integrating this matching coefficient with demand and supply features, a composite decision state feature containing the demand-supply matching coefficient can be generated. This feature comprehensively characterizes the adaptability of semantic units at specific nodes and time points.

[0036] Specifically, each edge node is virtualized as an agent, abstracting the resource decision-making capabilities of the edge nodes into the action generation capabilities of the agents to achieve distributed collaborative decision-making. Subsequently, a collaborative decision-making model based on a policy neural network can be configured for each agent. This policy neural network can employ a multilayer perceptron or Transformer architecture, with its input layer dimension consistent with the dimension of the composite decision state features. The hidden layers can capture complex correlations between features through nonlinear transformations, and the output layer corresponds to the probability distribution of the output action proposals. Illustratively, each agent, based on its corresponding dynamic resource profile vector, resource opportunity window prediction sequence, and globally shared composite decision state features, can perform inference processing through the policy neural network. During inference, the agent can generate action proposals containing resource quotes and collaborative resource requirements based on locally observed resource states and globally shared adaptability features. The resource quotes are the resource exchange conditions proposed by the agent for a specific semantic unit, such as estimated processing time and energy consumption costs. The collaborative resource requirements are the external resource requests from other nodes, such as communication and computing power support, needed by the agent to execute the semantic unit.

[0037] Furthermore, a central coordination node pre-installed on the AI ​​smart gateway can be invoked. This node, a software module deployed on the AI ​​smart gateway, is responsible for coordinating global decisions and resolving conflicts. The central coordination node collects action proposals submitted by all agents through a reliable communication link, performs conflict detection on each action proposal, and identifies resource contention, time overlap, or dependency constraint conflicts within the action proposals. For example, for each action proposal's corresponding semantic unit and target edge node, it can be checked whether the node's resources at the corresponding execution time point have been occupied by other action proposals; if resource overlap exists, it is determined to be a resource conflict. Secondly, the dependency identifiers of the semantic unit can be checked; if the execution time of a preceding dependent semantic unit is later than the execution start time of the semantic unit itself, it is determined to be a dependency conflict. Finally, conflicting action proposals can be categorized into a conflicting action proposal set, and conflict-free action proposals into a conflict-free action proposal set.

[0038] Specifically, the preset global optimization objective can comprehensively consider multiple dimensions of indicators such as end-to-end latency, inference accuracy achievement rate, system energy consumption, and load balancing. Combined with a sorted set of conflict-free action proposals, it makes decisions regarding the relationship between each semantic unit and the target edge node. During decision-making, the binding relationship corresponding to the action proposal with the highest priority score can be prioritized to ensure that the semantic unit is bound to the node with the optimal resource supply. Furthermore, the execution time window order of each semantic unit can be calibrated based on the dependency identifiers in the semantic feature vector. For example, if semantic unit A is a prerequisite for semantic unit B, the execution end time of semantic unit A is set as the lower limit of the execution start time of semantic unit B, thus obtaining the expected execution time window that satisfies the dependency constraints. In addition, corresponding computational precision parameters can be configured based on the precision sensitivity in the semantic feature vector and the computational capabilities of the target edge node. For example, FP16 precision can be configured for semantic units with high precision sensitivity, while INT8 precision can be configured for scenarios with low precision sensitivity and limited node computing power to balance computational precision and execution efficiency. Finally, the binding relationships, expected execution time windows, and computational precision parameters are structurally integrated to form an executable scheduling scheme.

[0039] In one embodiment, priority calculation is performed on each conflict-free action proposal in the conflict-free action proposal set to obtain a sorted conflict-free action proposal set, including: For each conflict-free action proposal in the conflict-free action proposal set, extract the latency sensitivity and precision sensitivity values ​​of the semantic units corresponding to each conflict-free action proposal in the semantic feature vector, and extract the resource quotation value in each conflict-free action proposal. Based on the preset numerical normalization algorithm, the latency sensitivity value, accuracy sensitivity value and resource price value are standardized respectively to obtain normalized latency sensitivity, normalized accuracy sensitivity and normalized resource price in the same numerical range. Based on the current overall load status of edge nodes and the distribution of task types to be scheduled, set dynamic weight coefficients for normalized latency sensitivity, normalized accuracy sensitivity, and normalized resource pricing; the sum of the dynamic weight coefficients is 1. The normalized latency sensitivity, normalized accuracy sensitivity, and normalized resource bid are multiplied by their respective dynamic weight coefficients and then summed to obtain the comprehensive priority score for each conflict-free action proposal. The conflict-free action proposals in the conflict-free action proposal set are sorted according to their comprehensive priority scores from high to low, resulting in a sorted set of conflict-free action proposals.

[0040] Specifically, for each conflict-free action proposal in the conflict-free action proposal set, the latency sensitivity and precision sensitivity values ​​of the corresponding semantic unit in the semantic feature vector can be extracted, and the resource price value in the action proposal can also be extracted. The latency sensitivity value represents the weight of the semantic unit's execution delay on the overall task completion time; the precision sensitivity value represents the weight of the semantic unit's computational precision on the final output precision of the task; and the resource price value represents the resource cost required for the agent to execute the semantic unit. Subsequently, a min-max normalization algorithm can be used to standardize the values ​​of the above three dimensions to eliminate differences in units and make the parameters of different dimensions comparable, obtaining normalized latency sensitivity, normalized precision sensitivity, and normalized resource price values ​​in the [0, 1] interval.

[0041] Furthermore, based on the current overall load status of edge nodes and the distribution of task types to be scheduled, dynamic weight coefficients corresponding to the above-mentioned normalized parameters can be set, and the sum of the dynamic weight coefficients is 1. For example, when the overall load of edge nodes is at its peak, the weight coefficient of normalized latency sensitivity is increased to prioritize low-latency tasks; when the proportion of high-precision tasks among the tasks to be scheduled exceeds a preset ratio, the weight coefficient of normalized precision sensitivity is increased to prioritize high-precision tasks; and when system energy consumption is too high, the weight coefficient of normalized resource pricing can be increased to prioritize action proposals with low resource costs. Subsequently, the normalized latency sensitivity, normalized precision sensitivity, and normalized resource pricing can be multiplied by their corresponding dynamic weight coefficients and summed to obtain the comprehensive priority score of each conflict-free action proposal, the formula of which is: ; in, To determine the overall priority score, , , These are the dynamic weighting coefficients for normalized latency sensitivity, normalized accuracy sensitivity, and normalized resource pricing, respectively. To normalize the latency sensitivity, To normalize the accuracy sensitivity, To normalize resource pricing, all action proposals in the conflict-free action proposal set are sorted according to their calculated comprehensive priority scores from highest to lowest, resulting in a sorted set of conflict-free action proposals. This sorting provides a clear priority basis for subsequent scheduling decisions, ensuring that high-priority action proposals are selected for execution first.

[0042] In one embodiment, such as Figure 2 As shown, based on an executable scheduling scheme, an instruction set including computation and communication instructions is generated and executed. Runtime performance data is collected during execution, and a runtime performance feedback data packet is generated based on this data. Based on the runtime performance feedback data packet, the parameters of the time-series prediction model and the collaborative decision-making model are updated to obtain the updated time-series prediction model and the updated collaborative decision-making model, including: S201: Through the instruction distribution module of the AI ​​smart gateway, the executable scheduling scheme is parsed, and for each target edge node, a corresponding calculation instruction containing semantic unit identifier, calculation precision parameters, and data address is generated. For edge node pairs with data interaction requirements, a communication instruction containing data transmission source node, target node, and transmission priority is generated. The calculation instruction and communication instruction are integrated to obtain the instruction set. S202: Distribute the instruction set to the corresponding edge nodes, execute the corresponding instructions in the instruction set through each edge node, and collect runtime performance data of the instruction execution process through the monitoring module deployed locally on each edge node. The runtime performance data includes actual calculation time, peak memory usage, network transmission time, actual energy consumption, and execution result accuracy. S203: Based on runtime performance data, associate the corresponding semantic unit identifier, target edge node identifier and scheduling scheme entry, perform data structured encapsulation processing, and generate runtime performance feedback data package; compare the actual performance parameters in the runtime performance feedback data package with the corresponding prediction parameters in the resource opportunity window prediction sequence, calculate the resource prediction error, and update the parameters of the time series prediction model based on the resource prediction error through the backpropagation algorithm to obtain the updated time series prediction model. S204: For the scheduling process of each semantic unit, extract the composite decision state features corresponding to the semantic unit, the action proposal corresponding to the agent executing the semantic unit, and the reward value calculated based on the actual performance data in the runtime performance feedback data packet. Integrate the composite decision state features, action proposal, and reward value to obtain the scheduling decision instance corresponding to each semantic unit. S205: Construct an experience replay buffer, store each scheduling decision instance in the experience replay buffer, randomly sample a preset number of scheduling decision instances from the experience replay buffer, and update the parameters of the collaborative decision model through the policy gradient algorithm to obtain the updated collaborative decision model.

[0043] Specifically, the instruction distribution module of the AI ​​smart gateway is a software component with structured data parsing and instruction generation capabilities. It can identify core information such as binding relationships, computational precision parameters, execution time windows, and communication paths in the scheduling scheme. Therefore, an executable scheduling scheme can be parsed based on this module. During the parsing process, for each target edge node, information such as the semantic unit identifier bound to the node, computational precision parameters, and data storage address can be extracted to generate computational instructions. These computational instructions adopt a standardized machine-readable format and can include a unique instruction identifier, execution priority, and resource reservation requirements, ensuring that edge nodes can directly parse and execute them. For edge node pairs with data interaction needs, information such as source node identifier, target node identifier, transmission priority, and data block identifier can be extracted according to the communication path planning in the scheduling scheme to generate communication instructions. The transmission priority is determined by the communication overhead and latency sensitivity of the semantic unit. High-priority communication instructions can preempt network bandwidth resources to ensure the transmission efficiency of critical data. By integrating various computational and communication instructions, a structured instruction set containing a global scheduling identifier can be formed. The global scheduling identifier is used for the association and tracing of subsequent execution results.

[0044] Furthermore, the instruction set can be distributed to the corresponding edge nodes via a reliable transmission protocol and a verification mechanism to ensure the integrity and reachability of the instructions. After receiving the instructions, each edge node can apply for a dedicated resource quota through its local resource management component. This resource quota is dynamically allocated based on parameters such as computational intensity and memory requirements in the computation instructions to avoid resource interference between different tasks. Subsequently, the edge nodes can execute computation and data transmission operations as required by the instructions. The computation operations can follow the computational precision parameters. For example, FP16 precision can be used to perform operator operations on semantic units with high precision sensitivity, while INT8 precision can be used for scenarios with low precision sensitivity and limited node computing power to balance computational precision and execution efficiency. The data transmission operations can follow the transmission priority in the communication instructions, with high-priority data occupying network bandwidth first to ensure data consistency between semantic units. During execution, the monitoring modules deployed locally on each edge node can collect runtime performance data in real time. The collected data may include actual computation time (the duration from the start of the instruction to the completion of the computation), peak memory usage (the maximum amount of memory used during the computation), network transmission time (the transmission time of data from the source node to the target node), actual energy consumption (the integral value of the energy consumption rate during the execution of the instruction), and execution result accuracy (model inference accuracy or data processing correctness).

[0045] Specifically, based on the collected runtime performance data, corresponding semantic unit identifiers, target edge node identifiers, and scheduling scheme entries can be associated to ensure that each performance data entry is traceable to a specific scheduling decision. Subsequently, an efficient serialization format can be used to encapsulate this data into a runtime performance feedback data packet. This packet contains a global scheduling identifier, semantic unit identifier, node identifier, numerical values ​​of each performance parameter, and timestamps, facilitating data analysis and matching during subsequent model updates. The actual performance parameters (such as actual available computing power and network transmission rate) in the runtime performance feedback data packet are compared with the corresponding predicted parameters in the resource opportunity window prediction sequence. The mean absolute error (MAE) metric can be used to calculate the resource prediction error, quantifying the degree of deviation between the predicted and actual values. Based on this resource prediction error, the parameters of the time-series prediction model can be updated using the backpropagation algorithm. During backpropagation, the prediction error is used as a loss function, and gradient descent is used to adjust the model's weights and biases, optimizing the model's prediction accuracy for future resource states and reducing the impact of prediction deviations on subsequent scheduling decisions, thus obtaining an updated time-series prediction model.

[0046] Indicatively, for the scheduling process of each semantic unit, the following can be extracted: the composite decision-making state features corresponding to that semantic unit (i.e., demand-supply matching coefficients, resource status, etc. at the time of decision-making), the action proposals corresponding to the agent executing that semantic unit (i.e., resource quotes, collaborative resource requirements, etc.), and the reward value calculated based on the actual performance data in the runtime performance feedback data packet. The reward value can be calculated using the following formula: ; in, For the first In this scheduling, semantic units Assigned to edge nodes The aforementioned reward value; , and All are target weight coefficients, and , can be based on Type dynamically adjusted, such as precision-sensitive tasks. Can be taken ; For the first The actual execution accuracy of the sub-scheduling, such as model inference accuracy and data processing accuracy; For the first The preset precision threshold for this scheduling; The actual execution delay of the t-th scheduling; Execute in each of the aforementioned edge nodes The minimum theoretical delay; Execute in each of the aforementioned edge nodes The maximum theoretical delay; The actual energy consumption of the t-th scheduling; For edge nodes The energy consumption threshold is dynamically generated based on the resource profile vector; This is the penalty coefficient for violations, and 1; For indicator functions, This indicates that there is a timeout, insufficient accuracy, or excessive power consumption. =1; otherwise =0.

[0047] Specifically, by integrating composite decision state features, action proposals, and reward values, a scheduling decision instance corresponding to each semantic unit can be obtained. Subsequently, an experience replay buffer can be constructed. This buffer uses a first-in, first-out queue structure to store a certain number of scheduling decision instances, avoiding data redundancy and overfitting. A predetermined number of scheduling decision instances can be randomly sampled from the experience replay buffer, and the parameters of the collaborative decision model can be updated using a policy gradient algorithm. That is, the reward value can be used as the optimization objective, and the weights of the policy neural network can be adjusted by calculating the gradient, so that the action proposals generated by the model can obtain higher long-term cumulative rewards, thereby improving the scheduling performance of the collaborative decision model and obtaining an updated collaborative decision model.

[0048] Based on the same inventive concept, this application also provides an AI smart gateway edge computing resource dynamic scheduling system for implementing the aforementioned AI smart gateway edge computing resource dynamic scheduling method. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the AI ​​smart gateway edge computing resource dynamic scheduling system provided below can be found in the limitations of the AI ​​smart gateway edge computing resource dynamic scheduling method described above, and will not be repeated here.

[0049] In one exemplary embodiment, such as Figure 3 As shown, an AI intelligent gateway edge computing resource dynamic scheduling system 300 is provided, including: The task parsing and resource assessment module 301 is used to acquire AI tasks containing task descriptions, perform semantic decoupling and feature extraction on the task descriptions respectively, generate corresponding task semantic unit sets and corresponding semantic feature vectors; collect real-time performance data of each edge node, quantify resource status based on real-time performance data, generate dynamic resource profile vectors of each edge node, and use a time series prediction model to perform short-term prediction based on real-time performance data to generate resource opportunity window prediction sequences. The collaborative decision-making and scheduling scheme generation module 302 is used to make decisions based on the task semantic unit set, semantic feature vector, dynamic resource profile vector and resource opportunity window prediction sequence through a collaborative decision-making model to generate an executable scheduling scheme; the executable scheduling scheme includes the binding relationship between each semantic unit and the target edge node, the calculation accuracy parameters and the expected execution time window; The scheduling execution and feedback update module 303 is used to generate and execute an instruction set including computation instructions and communication instructions based on an executable scheduling scheme, collect runtime performance data during the execution process, generate runtime performance feedback data packets based on the runtime performance data, and update the parameters of the time series prediction model and the collaborative decision-making model based on the runtime performance feedback data packets to obtain the updated time series prediction model and the updated collaborative decision-making model.

[0050] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the AI ​​smart gateway edge computing resource dynamic scheduling method of this application. A multi-core processor is preferred to improve the system's parallel processing capability. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0051] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the AI ​​smart gateway edge computing resource dynamic scheduling method of this application. The computer-readable storage medium may include: read-only memory, random access memory (RAM), solid-state drive (SSD), or optical disk, etc.

[0052] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

[0053] For ease of explanation, the above description has been provided in conjunction with specific embodiments. However, the discussion in some embodiments is not intended to be exhaustive or to limit the embodiments to the specific forms disclosed above. Various modifications and variations can be obtained based on the above teachings. The selection and description of the above embodiments are for the purpose of better explaining the contents of this disclosure, thereby enabling those skilled in the art to better utilize the embodiments.

Claims

1. A method for dynamic scheduling of edge computing resources for an AI smart gateway, characterized in that, The method includes: The system acquires AI tasks containing task descriptions, performs semantic decoupling and feature extraction on the task descriptions, generates corresponding task semantic unit sets and corresponding semantic feature vectors; collects real-time performance data of each edge node, quantifies resource status based on the real-time performance data, generates dynamic resource profile vectors for each edge node, and uses a time-series prediction model to perform short-term prediction based on the real-time performance data to generate resource opportunity window prediction sequences. Based on the task semantic unit set, the semantic feature vector, the dynamic resource profile vector, and the resource opportunity window prediction sequence, a collaborative decision-making model is used to generate an executable scheduling scheme; the executable scheduling scheme includes the binding relationship between each semantic unit and the target edge node, the calculation precision parameters, and the expected execution time window; Based on the executable scheduling scheme, an instruction set including computation instructions and communication instructions is generated and executed. Runtime performance data during the execution process is collected, and a runtime performance feedback data packet is generated based on the runtime performance data. Based on the runtime performance feedback data packet, the parameters of the time series prediction model and the collaborative decision-making model are updated to obtain the updated time series prediction model and the updated collaborative decision-making model.

2. The method according to claim 1, characterized in that, The task descriptions are semantically decoupled and feature extracted to generate corresponding task semantic unit sets and corresponding semantic feature vectors; real-time performance data of each edge node is collected, resource status is quantified based on the real-time performance data, dynamic resource profile vectors of each edge node are generated, and a time-series prediction model is used to perform short-term prediction based on the real-time performance data to generate a resource opportunity window prediction sequence, including: Obtain a predefined model knowledge base, which stores computational graph models corresponding to various task types. The computational graph models include sets of operators and data dependencies between operators. Based on the task type information in the task description, load the corresponding computational graph model from the predefined model knowledge base. The computation patterns and data dependencies of each operator in the computation graph model are analyzed. Operators that are continuous and have the same computation pattern are aggregated to generate multiple semantic units that constitute the AI ​​task. The semantic units are integrated to obtain the task semantic unit set. For each semantic unit in the task semantic unit set, feature quantization is performed based on the corresponding aggregation operator type, data flow scale, and topological position in the computation graph model, combined with the dependency information in the task description, to generate the corresponding semantic feature vector. The semantic feature vector includes computational intensity, memory requirements, communication overhead, latency sensitivity, accuracy sensitivity, and dependency identifier. The monitoring agents deployed on each edge node periodically collect raw performance monitoring data of each edge node according to a preset period. The raw performance monitoring data includes computing resource utilization, memory usage, network transmission rate and energy consumption data. Based on the original performance monitoring data, sliding window statistics are performed to calculate the real-time available resources and performance status parameters of each edge node in terms of computing, memory, network and energy dimensions, and to generate a dynamic resource profile vector for each edge node. The historical dynamic resource profile vectors of each edge node and the time-series prediction model are obtained. The dynamic resource profile vectors corresponding to the real-time performance data and the historical dynamic resource profile vectors are input into the time-series prediction model for forward inference processing to generate a resource opportunity window prediction sequence for each edge node at multiple consecutive time points within a predetermined future time interval. Each prediction element in the resource opportunity window prediction sequence includes a predicted value of available computing power, a predicted value of network quality, and a prediction confidence level.

3. The method according to claim 1, characterized in that, The step of generating an executable scheduling scheme by making decisions through a collaborative decision-making model based on the task semantic unit set, the semantic feature vector, the dynamic resource profile vector, and the resource opportunity window prediction sequence includes: For each semantic unit in the task semantic unit set, each target edge node and its corresponding execution start time point are traversed. Based on the features in the semantic feature vector, the dynamic resource profile vector corresponding to the target edge node, and the resource opportunity window prediction sequence corresponding to the target edge node, multi-dimensional adaptation calculation is performed to generate a composite decision state feature containing demand-supply matching coefficients. Each edge node is virtualized as an agent, and a collaborative decision-making model based on a policy neural network is configured for each agent. Through each agent, based on the dynamic resource profile vector, the resource opportunity window prediction sequence, and the globally shared composite decision state features, the collaborative decision-making model performs reasoning processing to generate action proposals that include resource pricing and collaborative resource requirements. The central coordination node preset in the AI ​​smart gateway is invoked to collect action proposals from each of the smart agents. Conflict detection processing is performed on each of the action proposals to identify the conflict action proposal set and the non-conflict action proposal set. Priority calculation processing is performed on each non-conflict action proposal in the non-conflict action proposal set to obtain the sorted non-conflict action proposal set. Based on a preset global optimization objective, and combined with the sorted set of conflict-free action proposals, a decision is made on the relationship between each semantic unit and the target edge node to obtain the binding relationship. Then, based on the dependency identifier in the semantic feature vector, the execution time window order of each semantic unit is calibrated to obtain the expected execution time window. According to the precision sensitivity in the semantic feature vector and the computing power of the target edge node, corresponding computational precision parameters are configured. Finally, the binding relationship, the expected execution time window, and the computational precision parameters are integrated to form the executable scheduling scheme.

4. The method according to claim 3, characterized in that, The process involves generating and executing an instruction set including computation and communication instructions based on the executable scheduling scheme, collecting runtime performance data during execution, generating a runtime performance feedback data packet based on the runtime performance data, and updating the parameters of the time-series prediction model and the collaborative decision-making model based on the runtime performance feedback data packet to obtain updated time-series prediction models and updated collaborative decision-making models, including: The AI ​​smart gateway's instruction distribution module parses the executable scheduling scheme and generates a corresponding calculation instruction containing a semantic unit identifier, calculation precision parameters, and data address for each target edge node. For edge node pairs with data interaction requirements, a communication instruction containing a data transmission source node, target node, and transmission priority is generated. The calculation instructions and the communication instructions are integrated to obtain the instruction set. The instruction set is distributed to the corresponding edge nodes, and the corresponding instructions in the instruction set are executed by each edge node. The runtime performance data of the instruction execution process is collected by the monitoring module deployed locally on each edge node. The runtime performance data includes actual computing time, peak memory usage, network transmission time, actual energy consumption, and execution result accuracy. Based on the runtime performance data, the corresponding semantic unit identifier, target edge node identifier, and scheduling scheme entry are associated, and the data is encapsulated in a structured manner to generate the runtime performance feedback data packet. The actual performance parameters in the runtime performance feedback data packet are compared with the corresponding prediction parameters in the resource opportunity window prediction sequence to calculate the resource prediction error. Based on the resource prediction error, the parameters of the time series prediction model are updated using the backpropagation algorithm to obtain the updated time series prediction model. For the scheduling process of each semantic unit, the composite decision state feature corresponding to the semantic unit, the action proposal corresponding to the agent executing the semantic unit, and the reward value calculated based on the actual performance data in the runtime performance feedback data packet are extracted. The composite decision state feature, the action proposal, and the reward value are integrated to obtain the scheduling decision instance corresponding to each semantic unit. An experience replay buffer is constructed, and each scheduling decision instance is stored in the experience replay buffer. A preset number of scheduling decision instances are randomly sampled from the experience replay buffer, and the parameters of the collaborative decision model are updated using a policy gradient algorithm to obtain the updated collaborative decision model.

5. The method according to claim 3, characterized in that, The step of prioritizing each conflict-free action proposal in the conflict-free action proposal set to obtain a sorted conflict-free action proposal set includes: For each conflict-free action proposal in the conflict-free action proposal set, extract the latency sensitivity value and precision sensitivity value of the semantic unit corresponding to each conflict-free action proposal in the semantic feature vector, and extract the resource quotation value in each conflict-free action proposal. Based on a preset numerical normalization algorithm, the latency sensitivity value, the precision sensitivity value, and the resource price value are standardized respectively to obtain normalized latency sensitivity, normalized precision sensitivity, and normalized resource price within the same numerical range. Based on the current overall load status of edge nodes and the distribution of task types to be scheduled, dynamic weight coefficients are set for the normalized latency sensitivity, the normalized accuracy sensitivity, and the normalized resource price; the sum of the dynamic weight coefficients is 1. The normalized latency sensitivity, the normalized accuracy sensitivity, and the normalized resource price are multiplied by their respective dynamic weight coefficients and then summed to obtain the comprehensive priority score for each conflict-free action proposal. The conflict-free action proposals in the conflict-free action proposal set are sorted according to the comprehensive priority score from high to low to obtain the sorted conflict-free action proposal set.

6. The method according to claim 4, characterized in that, The reward value is calculated using the following formula: ； in, For the first In this scheduling, semantic units Assigned to edge nodes The aforementioned reward value; , and All are target weight coefficients, and ; For the first The actual execution accuracy of this scheduling; For the first The preset precision threshold for this scheduling; The actual execution delay of the t-th scheduling; Execute in each of the aforementioned edge nodes The minimum theoretical delay; Execute in each of the aforementioned edge nodes The maximum theoretical delay; The actual energy consumption of the t-th scheduling; For edge nodes The energy consumption threshold; This is the penalty coefficient for violations, and 1; For indicator functions, This indicates that there is a defect. =1, otherwise =0.

7. A dynamic scheduling system for edge computing resources of an AI intelligent gateway, characterized in that, The system includes: The task parsing and resource assessment module is used to acquire AI tasks containing task descriptions, perform semantic decoupling and feature extraction on the task descriptions respectively, generate corresponding task semantic unit sets and corresponding semantic feature vectors; collect real-time performance data of each edge node, quantify resource status based on the real-time performance data, generate dynamic resource profile vectors of each edge node, and use a time-series prediction model to perform short-term prediction based on the real-time performance data to generate resource opportunity window prediction sequences. The collaborative decision-making and scheduling scheme generation module is used to make decisions based on the task semantic unit set, the semantic feature vector, the dynamic resource profile vector, and the resource opportunity window prediction sequence through a collaborative decision-making model to generate an executable scheduling scheme; the executable scheduling scheme includes the binding relationship between each semantic unit and the target edge node, the calculation accuracy parameters, and the expected execution time window; The scheduling execution and feedback update module is used to generate and execute an instruction set including computation instructions and communication instructions based on the executable scheduling scheme, collect runtime performance data during the execution process, generate runtime performance feedback data packets based on the runtime performance data, and update the parameters of the time series prediction model and the collaborative decision model based on the runtime performance feedback data packets to obtain the updated time series prediction model and the updated collaborative decision model.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images