Cloud edge collaboration-oriented model evolution task scheduling strategy generation method and system

Abstract

This invention provides a method and system for generating scheduling strategies for model evolution tasks in a cloud-edge collaborative manner. The method includes: determining scheduling requests for decomposed subtasks corresponding to the model evolution task; wherein, each decomposed subtask corresponds one-to-one with a data slice on the scheduled node; the scheduling request includes the model evolution task type and scheduling resource requirements including network performance requirements and model efficiency requirements; based on the scheduling request and pre-stored node resource status information, obtaining multiple candidate scheduling schemes using a multi-objective optimization algorithm; wherein, each candidate scheduling scheme includes candidate scheduling nodes for all decomposed subtasks; and selecting one of the multiple candidate scheduling schemes as the scheduling strategy for the model evolution task. This invention can find a balance between model performance and resource utilization between edge computing and cloud computing.

Description

Technical Field

[0001] This invention relates to the field of cloud-edge collaboration technology, and in particular to a method and system for generating scheduling strategies for model evolution tasks oriented towards cloud-edge collaboration. Background Technology

[0002] With the deepening of the construction of new power systems, the increase in heterogeneous edge devices has led to a high degree of uncertainty and dynamism in the operation of the power grid, posing unprecedented challenges to real-time perception, rapid decision-making, and adaptive control. A large number of heterogeneous edge devices (such as smart gateways, substation monitoring devices, distribution terminals, smart meters, and online monitoring sensors) are deployed as edge nodes at the power system edge. Their computing capabilities vary greatly, ranging from TOPS (Tera Operations Per Second) level smart gateways to MOPS (Million Operations Per Second) level distribution terminals. Their communication capabilities range from 5G power private networks to high-speed power line carriers (HPLC, typically with a bandwidth of 1-2 Mbps) and narrowband Internet of Things (NB-IoT, typically with a bandwidth of tens of Kbps). Furthermore, they rely on various power supply devices such as batteries or solar power. Simultaneously, multiple cloud devices are deployed as cloud nodes in the power system cloud. Through cloud-edge collaboration mechanisms, not only can bidirectional data flow be achieved, but the power system can also be continuously optimized.

[0003] In power systems, each cloud node deploys a cloud-based model, and each edge node deploys an edge-based model. These cloud and edge models are collectively referred to as cloud-edge collaborative models, which support critical tasks such as fault diagnosis, equipment monitoring, and condition assessment. Cloud models (such as deep learning models like Transformer-based transformer fault diagnosis models and ResNet-based transmission line defect identification models) typically have 100MB to several GB of parameters and an accuracy of 0.9–0.95, possessing powerful global computing capabilities and high-precision analysis capabilities, but they need to continuously adapt to new fault modes and operating conditions. Edge models (such as lightweight deep learning models like MobileNet and EfficientNet) typically have 1–10MB of parameters and an accuracy of 0.8–0.9, possessing rapid response capabilities, but there are significant differences in model accuracy and parameter size among models deployed on different edge nodes. Furthermore, edge models need to periodically retrieve updated parameters from the cloud to improve performance. Therefore, the continuous evolution of cloud-edge collaborative models is a core requirement for the long-term operation of power systems.

[0004] The evolutionary process of cloud-edge collaborative models involves three co-evolutionary tasks: incremental model learning (typically, the cloud model learns incrementally based on accumulated data from edge nodes, or local edge nodes learn based on fault samples, to continuously optimize model performance and adapt to dynamic evolution scenarios), knowledge fusion (aggregating feature knowledge scattered across multiple edge nodes to the cloud for knowledge refinement and dimensionality enhancement, forming more comprehensive and efficient global knowledge), and parameter update (the cloud model distributes updated parameters to other nodes, and multiple nodes coordinate to synchronously update model parameters, achieving dimensionality reduction and empowerment of knowledge from the cloud to the edge). These three types of model evolution tasks have characteristics different from ordinary computing tasks: ① They need to consider model-specific attributes such as model accuracy, knowledge fusion efficiency, and the number of model parameters; ② The task execution involves the transmission and updating of a large amount of data, requiring high network communication; ③ Different evolutionary tasks have significantly different resource requirements, such as incremental learning requiring high model accuracy, knowledge fusion requiring high-bandwidth networks, and parameter update being sensitive to latency. Therefore, how to fully consider the characteristics of the task and reasonably allocate the model evolution task to cloud nodes and edge nodes in the power system to achieve model optimization has gradually become the research focus of the continuous model evolution task.

[0005] Existing model evolution task scheduling strategies are mainly divided into two categories: "edge-end model evolution task scheduling mechanism based on static resource balancing" and "evolution task scheduling scheme based on centralized cloud processing". The static resource-based allocation strategy is as follows: The power system adopts a decentralized distributed architecture, with each edge node independently handling the evolution tasks of the edge model, and employing heuristic task allocation strategies (round-robin scheduling, random allocation, etc.). For example, the scheduling strategy generator periodically collects metrics such as CPU utilization and memory usage of each edge node, and distributes the model evolution tasks to be processed to edge nodes with relatively low loads based on the instantaneous load status. This makes scheduling decisions mainly based on static metrics such as periodically collected node resource utilization, and cannot adaptively allocate tasks according to the specific needs of the model evolution task, such as model accuracy and knowledge fusion efficiency. Moreover, even for computationally intensive tasks that require processing large-scale model parameters or knowledge fusion, the power system uses an atomic task allocation strategy, allocating the entire model evolution task as an indivisible execution unit to a single edge node. Furthermore, the power system adopts a single-objective optimization strategy, mainly optimizing node load balancing or task response latency, without considering the impact of model-specific attributes such as model parameter transmission volume and knowledge fusion efficiency during model evolution task execution.

[0006] The cloud-based centralized processing-based evolutionary task management scheme is as follows: Existing power systems employ a centralized computing architecture, with edge nodes acting only as data acquisition terminals and data transmission agents. All evolution-related data generated by edge devices (such as new fault samples and edge feature knowledge) is uploaded to a cloud data center for unified processing. The cloud data center adopts a single-objective optimization strategy, using a queue-based task scheduling mechanism to manage model evolutionary tasks. Furthermore, the system utilizes a high-performance cloud server cluster and large-scale cloud models (such as Transformer-based transformer fault diagnosis models and ResNet-based transmission line defect identification models, with model parameters typically ranging from 100MB to several GB), providing high computational accuracy (usually above 90%).

[0007] However, current model evolution task scheduling suffers from several shortcomings: ① Lack of task decomposition and distributed parallel processing capabilities for model evolution tasks: Existing solutions use a single scheduling node to execute a complete model evolution task. For example, allocating complete incremental model learning tasks and knowledge fusion tasks as atomic units to a single edge node can easily lead to low resource utilization and difficulty in meeting real-time requirements. Moreover, when transmitting data related to the evolution task (especially a large number of model parameters or high-dimensional feature knowledge), data transmission capacity can easily become a system performance bottleneck in bandwidth-constrained HPLC or NB-IoT environments. ② Existing single-objective optimization methods cannot balance the multiple mutually restrictive indicators involved in model evolution tasks, such as network communication resources (e.g., communication overhead and latency), making them unsuitable for real-world scenarios. ③ Rigid intelligent matching schemes: Most existing methods determine task scheduling nodes based on static indicators without considering the characteristics of model evolution tasks (e.g., model evolution tasks have special requirements for model-specific attributes such as the amount of model parameter transmission and knowledge fusion efficiency). ④ Lack of collaborative computing mechanisms between nodes: Current power system model evolution tasks are completed in the cloud or at the edge, resulting in the underutilization of computing resources and deployed models at the edge and cloud.

[0008] Therefore, there is an urgent need for a scheduling strategy generation scheme suitable for power system model evolution tasks. In a cloud-edge collaborative scenario, this scheme can meet real-time requirements by decomposing the complete model evolution task into parallel sub-tasks. Furthermore, through multi-objective optimization, it can match the node resource status and task resource requirements, which contain various mutually restraining indicators, thereby generating an evolution task scheduling strategy that improves the overall resource utilization of the power system. Summary of the Invention

[0009] In view of this, embodiments of the present invention provide a method and system for generating scheduling strategies for model evolution tasks oriented towards cloud-edge collaboration. By using distributed processing to decompose subtasks and multi-objective optimization algorithms, an evolution task scheduling strategy that improves the overall resource utilization of the power system is generated.

[0010] One aspect of the present invention provides a method for generating scheduling strategies for cloud-edge collaborative model evolution tasks, the method comprising the following steps: Determine the scheduling requests for the decomposed subtasks corresponding to the model evolution task; wherein, the data transmitted from the scheduled node to the scheduling node during the scheduling of the model evolution task is in the form of data slices, and the decomposed subtasks correspond one-to-one with the data slices; the scheduling requests for the decomposed subtasks include the type of the model evolution task to which they belong and the scheduling resource requirements, the dimensions of which include network performance and model effectiveness. Based on scheduling requests and pre-stored node resource status information, a multi-objective optimization algorithm is used to obtain multiple candidate scheduling schemes. Each candidate scheduling scheme includes a candidate scheduling node for all decomposed subtasks corresponding to the model evolution task. The fitness evaluation of the multi-objective algorithm is based on scheduling resource requirements. The scheduled node, the scheduling node, and the candidate scheduling node are cloud nodes or edge nodes. One of the multiple candidate scheduling schemes is selected as the scheduling strategy for the model evolution task and output. The scheduling strategy is then sent to the scheduled node, and the scheduled node sends the data slices to the corresponding scheduling node.

[0011] In some embodiments of the present invention, determining the scheduling request for the decomposed subtask corresponding to the model evolution task includes: Receive scheduling requests for model evolution tasks from edge nodes or cloud nodes; wherein, the scheduling request for model evolution tasks includes the type of model evolution task, the size of each initial data slice on the scheduled node, the network performance dimension requirements of the model evolution task, and the model efficiency dimension requirements of the model evolution task; wherein, the initial data slice is determined based on the type of model evolution task; The number of task decompositions and the merging strategy of the initial data slices are determined based on the size of the initial data slices and the pre-stored node resource status. Based on the number of task decompositions and the merging strategy of the initial data slices, the scheduling request of the model evolution task is divided into scheduling requests of multiple decomposed subtasks. The merging strategy of the initial data slices is the strategy of merging the initial data slices to obtain data slices. The scheduling request of each decomposed subtask is determined based on the scheduling request of the model evolution task.

[0012] In some embodiments of the present invention, the multi-objective function of the multi-objective optimization algorithm includes an execution delay function and a model accuracy function; The value of the execution delay function is determined based on the number of decomposed subtasks assigned to the candidate scheduling node and the base delay of the candidate scheduling node; the value of the model accuracy function is determined based on the number of decomposed subtasks assigned to the candidate scheduling node and the accuracy of the model deployed on the candidate scheduling node.

[0013] In some embodiments of the present invention, the scheduling request for decomposing subtasks may also include the data slice size corresponding to the decomposing subtasks; The multi-objective function of the multi-objective optimization algorithm also includes a communication cost function; the value of the communication cost function is determined based on the data slice size corresponding to the decomposed subtasks assigned to the candidate scheduling node and the physical regions to which the scheduled node and the scheduling node belong.

[0014] In some embodiments of the present invention, the multi-objective optimization algorithm is a genetic algorithm; Based on scheduling requests and pre-stored node resource status information, a multi-objective optimization algorithm is used to obtain multiple candidate scheduling schemes, including: Initialize the population: For each decomposed subtask, based on the scheduling resource requirements of the decomposed subtask and the pre-stored node resource status information, select nodes that meet the preset node matching rules, and generate an initial population based on the preset probability of randomly selecting individual solutions from the selected nodes and the preset probability of randomly selecting individual solutions from all nodes. Multi-objective fitness assessment: The fitness of each individual solution in the initial population is assessed using a multi-objective function; Non-dominated ordination and crowding distance calculation: Based on the fitness assessment results, individuals in the population are divided into different frontiers according to non-dominated relationships, and the crowding distance of each individual in the population is calculated; Selection operation: Select a new population based on non-dominated ordering and crowding distance; Crossover operation: Select parent individuals from the new population and generate new offspring individuals based on a preset crossover probability; Mutation operation: Mutate the generated offspring individuals based on the preset mutation probability; wherein, the mutation probability of offspring individuals that conform to the preset node matching rules is greater than the mutation probability of offspring individuals that do not conform to the preset node matching rules. Merging and sorting: Merge parent and child individuals, perform non-dominated sorting and crowding distance calculation on the merged population, and select the next generation population from the merged population; Termination condition judgment: If the iteration termination condition is met, output the non-dominated frontier in the next generation population as the scheduling solution set; otherwise, repeat the crossover operation, mutation operation, and merging and sorting.

[0015] In some embodiments of the present invention, model performance requirements include model accuracy requirements, and network performance requirements include latency requirements. The preset node matching rules also include: If the model accuracy requirement of the decomposed subtask meets the preset accuracy threshold, the node matching order of the decomposed subtask is arranged in descending order of the accuracy of the models deployed on the nodes. If the latency requirements of the decomposed subtasks meet the preset latency threshold, then the node matching order of the decomposed subtasks is arranged in ascending order of node latency; and If the size of the data slice corresponding to the decomposed subtask meets the preset data volume threshold, then the highest priority matching node for the decomposed subtask is a node whose bandwidth meets the preset bandwidth threshold and / or a node belonging to the same physical region as the scheduled node.

[0016] In some embodiments of the present invention, the model evolution task is of the type of incremental learning task, knowledge fusion task, or parameter update task. The initial data slices were obtained in the following way: For incremental learning tasks, each batch of data on the scheduled node is used as an initial data slice; For knowledge fusion tasks, the knowledge at each edge node to be fused is treated as a data slice; and For parameter update tasks, the updated parameters of each model layer on the cloud model after parameter update are taken as a data slice; The preset node matching rules also include: the scheduling nodes for all decomposed subtasks corresponding to the knowledge fusion task are the same cloud node; the scheduling nodes for the parameter update task are other nodes besides the scheduled node.

[0017] In some embodiments of the present invention, the method further includes: when an anomaly is detected in a specific decomposed subtask, reselecting an available scheduling node for the specific decomposed subtask from other candidate scheduling schemes besides the selected candidate scheduling scheme.

[0018] Another aspect of the present invention provides a scheduling strategy generation system for cloud-edge collaborative model evolution tasks, including a processor, a memory, and a computer program / instructions stored in the memory. The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method described in any of the above embodiments.

[0019] Another aspect of the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method described in any of the above embodiments.

[0020] The proposed method and system for generating scheduling strategies for cloud-edge collaborative model evolution tasks can meet real-time requirements in cloud-edge collaborative scenarios by decomposing the complete model evolution task into parallel-processable sub-tasks. Furthermore, through multi-objective optimization algorithms, it can match node resource status and task resource requirements that include various mutually constraining indicators (such as some network performance indicators and model efficiency indicators), thereby generating an evolution task scheduling strategy that improves the overall resource utilization of the power system. This aims to find a balance between model performance and resource utilization between edge computing and cloud computing.

[0021] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0022] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings: Figure 1 This is a schematic diagram of the systematic path of the scheduling strategy generation method in one embodiment of the present invention.

[0024] Figure 2 This is a flowchart illustrating a scheduling strategy generation method in one embodiment of the present invention.

[0025] Figure 3 This is a flowchart illustrating the scheduling strategy generation method in another embodiment of the present invention.

[0026] Figure 4 This is a flowchart illustrating a multi-objective optimization algorithm in one embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0028] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0029] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0030] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0031] In the model evolution scenario of power systems, cloud-based models need to continuously adapt to new fault modes and operating conditions, while edge models need to obtain updated parameters from the cloud to improve performance. However, current power system task scheduling mainly targets ordinary computing tasks, and there is a lack of in-depth research on scheduling schemes for cloud-edge collaborative model evolution tasks. Existing scheduling strategies for model evolution tasks are mainly divided into two categories: "edge-side model evolution task scheduling based on static resource balancing" and "evolution task scheduling scheme based on centralized cloud processing." However, their shortcomings, such as a single scheduling node executing a complete evolution task, single-objective optimization, and single-side node task execution, make it difficult to adapt to the increasingly complex power systems.

[0032] Based on this, this application proposes a scheduling strategy generation method for model evolution tasks in power system cloud-edge collaboration. This method not only overcomes the shortcomings of traditional single-objective optimization in not being able to simultaneously consider multiple mutually constraining indicators such as model accuracy and latency, but also adaptively adjusts the task decomposition and scheduling scheme according to the resource requirements of the model evolution task and the resource status of nodes. This allows for finding a balance between model performance and resource utilization between edge computing and cloud computing, thereby realizing power cloud-edge collaborative computing.

[0033] Specifically, such as Figure 1As shown, the technical implementation of this application follows a systematic path of "task decomposition - multi-objective optimization combined with intelligent matching - task strategy generation" to support the evolution of cloud-edge collaborative models: ① Constructing a unified measurement system for state and task requirements covering multiple dimensions of resources such as network communication and model performance. This accurately characterizes the resource characteristics of cloud nodes and edge nodes (such as the accuracy and number of model parameters on the cloud node, the node's computing and storage resources, network bandwidth, node communication latency, and device energy consumption), and determines the resource requirements of the model evolution task for the cloud-edge collaborative network (such as computational load, latency, energy consumption, and requirements for model-specific attributes such as accuracy and knowledge fusion efficiency of the cloud-edge collaborative model) through received task scheduling requests. ② This application can use a task decomposition mechanism to decompose the complete model evolution task into multiple independent and parallel sub-tasks, laying the foundation for the distributed execution of the model evolution task. For example, this application can design incremental learning tasks to be decomposed according to data batches, knowledge fusion tasks to be decomposed according to the number of edge nodes to be fused, and parameter update tasks to be decomposed according to model layers or parameter groups. ③ This application simultaneously optimizes key objectives (such as network performance and model efficiency) through a multi-objective optimization algorithm. Furthermore, during the multi-objective optimization process, multiple candidate scheduling schemes can be obtained by matching the resource status information of nodes with the scheduling resource requirements of the model evolution task, thus achieving a reasonable allocation of the model evolution task. ④ This application can select the most suitable scheduling scheme from multiple candidate schemes based on user-defined preferences or scenario requirements, thereby allocating the decomposed subtasks to corresponding edge nodes and cloud nodes. This achieves intelligent task allocation tailored to the characteristics of the model evolution task and promotes the collaborative evolution of cloud and edge models.

[0034] The scheduling strategy generation method for model evolution tasks proposed in this application can be implemented through a scheduling strategy generator. This generator adopts a hierarchical architecture, comprising three core modules: a demand awareness and resource measurement module, a multi-objective optimization engine module containing preset node matching rules, and a solution set selection module. These modules work collaboratively to generate the scheduling strategy for model evolution tasks. Specifically, the demand awareness and resource measurement module periodically / in real-time collects and measures the resource status information of edge nodes and cloud nodes in the cloud-edge collaborative network, and obtains the scheduling resource requirements of the model evolution task from the received scheduling requests, providing a foundation for subsequent multi-objective optimization. The multi-objective optimization engine module is the core algorithm module, used to intelligently match the scheduling nodes of each decomposed subtask with the resource status information of cloud nodes and edge nodes in the cloud-edge collaborative network and the scheduling resource requirements of the decomposed subtasks during multi-objective optimization. It obtains multiple candidate scheduling schemes that satisfy the constraints under multiple conflicting objective functions, thereby allocating the subtasks to computing nodes with matching capabilities. The solution set selection module is used to select the optimal solution set from multiple scheduling solution sets (i.e., candidate scheduling schemes) obtained by multi-objective optimization algorithms. For example, it selects the scheduling scheme most suitable for the current power application scenario through a weighted adaptive fusion strategy, generating an executable scheduling policy. Further, if the scheduling request transmitted by the scheduled node to the scheduling policy generator is a scheduling request for a model evolution task, the scheduling policy generator may also include a task decomposition module, used to decompose the complete model evolution task into multiple independently executable subtasks. Each subtask can be independently assigned to different computing nodes for parallel processing. In addition, after the scheduling policy is generated, the scheduling of the model evolution task can be executed with the help of a scheduler: the scheduling policy generator transmits the scheduling policy to the scheduler, which can transmit the scheduling policy (e.g., the scheduling policy may contain the identification information of the scheduling nodes for each decomposed subtask) to the scheduled node, enabling the scheduled node to assign the subtask to the corresponding scheduling node. Furthermore, the scheduler can monitor the execution status of each decomposed subtask in real time and collect the task execution results of each node, thereby aggregating and integrating the task execution results to present the final result to the user.

[0035] As an example, the resource status information and scheduling resource requirements perceived by the demand perception and resource measurement module can theoretically include one or more of the following dimensions: computing and storage resources (such as CPU usage, memory usage, and GPU computing power), network performance resources (such as network bandwidth, latency, network jitter, and packet loss rate), model performance resources (such as model accuracy, model parameter count, and knowledge fusion efficiency), communication overhead resources, and energy resources that nodes can consume, etc., but this invention is not limited to these. Since this application targets model evolution task scheduling, the resource dimensions in the resource status information and scheduling resource requirements must include both network performance and model performance dimensions. That is, the resource status information includes resources corresponding to at least one of the network performance indicators (such as bandwidth and latency resources) and resources corresponding to at least one of the model performance indicators, and the scheduling resource requirements include resource requirements corresponding to at least one of the network performance indicators and resource requirements corresponding to at least one of the model performance indicators. For example, scheduling resource requirements may include response latency requirements and model accuracy requirements. The demand perception and resource measurement module can collect computing and storage resource information from hardware and operating systems through monitoring systems and API interfaces, and perceive network performance indicators by interfacing with the cloud-edge collaborative network controller. In addition, both the scheduling policy generator and the scheduler can be deployed in the cloud.

[0036] In some embodiments of the present invention, the scheduling strategy generation method proposed in this application is applicable to incremental learning tasks, knowledge fusion tasks, and parameter update tasks. The specific mechanisms by which scheduling these three types of evolutionary tasks promotes the evolution of cloud-edge collaborative models are as follows: ① Incremental learning task: After the edge model accumulates a certain amount of data through data collection or collects new fault sample data, the scheduled node at the edge transmits the accumulated data or fault sample data to the scheduling node, so that the model deployed on the scheduling node can learn based on the data fed back by the scheduled node; ② Knowledge fusion task: The edge node inputs the collected data into the edge model on the node, and can output a feature vector as the knowledge of the edge node; when an edge node determines that knowledge fusion is required, other edge nodes (such as other edge nodes) that will be fused with that node can be further determined. (This can be any two edge nodes located in the same physical region as the edge node to be fused with knowledge, and the method can be designed and determined independently.) The edge node can acquire knowledge from other edge nodes and transmit its own knowledge, along with the acquired knowledge from other edge nodes, to the scheduling node in the cloud (or all edge nodes to be fused can each transmit their knowledge to the scheduling node), enabling the scheduling node to aggregate and refine the feature knowledge of multiple edge nodes. ③ After the model on the cloud node updates its model parameters based on data transmitted from the incremental learning task and / or knowledge fusion task, it can transmit the updated model parameters to other nodes (which can be cloud nodes or edge nodes) through a parameter update task, allowing other nodes to update their models based on the received model parameters, thus realizing data transfer from the cloud to the edge model. These three types of cloud-edge collaborative model evolution tasks can form a collaborative evolution closed loop of "edge feedback → cloud learning → edge empowerment," achieving continuous evolution of the cloud-edge collaborative model. The edge nodes located in the same physical region mentioned in this application refer to local nodes.

[0037] It should be noted that, unlike knowledge fusion tasks and parameter update tasks, the scheduling nodes in the incremental learning task of this application can change according to the data type transmitted during the scheduling of the incremental learning task: the scheduled node in the incremental learning task is an edge node, and the scheduling node can be a cloud node or an edge node, for the following reasons: when a new fault sample is collected by an edge node, since the fault occurs urgently, the learning and processing process usually needs to be completed in a short time. This application designs, through preset node matching rules, that when the type of data transmitted by the incremental learning task is determined to be fault sample data based on the scheduling request (at this time, whether it is the scheduling request of the model evolution task or the scheduling request of the decomposition subtask, it includes the type of data transmitted from the scheduled node to the scheduling node), the scheduling nodes of all decomposition subtasks corresponding to this incremental learning task are edge nodes (such as any other edge node in the same physical area as the edge node that collected the new fault sample); this application also designs, through preset node matching rules, that when the type of data transmitted by the incremental learning task is determined to be accumulated normal node operation data based on the scheduling request (similarly, the scheduling request includes the type of data transmitted from the scheduled node to the scheduling node), the scheduling nodes of all decomposition subtasks corresponding to this incremental learning task are cloud nodes. In the knowledge fusion task, the scheduled node is an edge node, and the scheduling node is a cloud node. In the parameter update task, the scheduled node is a cloud node, and the scheduling node includes all cloud nodes and edge nodes except for the node that generates the model and updates the parameters (i.e., the scheduled node). In addition, the preset node matching rules may also include: the scheduling node for all decomposed subtasks corresponding to the knowledge fusion task is the same cloud node.

[0038] As an example, the triggering conditions for edge nodes to initiate knowledge fusion tasks are as follows: ① When the model deployed on the edge node determines that the current fault data does not meet the preset confidence threshold, for example, when the classification probability distributions are very close (e.g., the probability of judging one fault is 40%, and the probability of judging another fault is 60%), and the edge node cannot give a definite conclusion, the node will actively initiate a knowledge fusion task, requesting that the feature vector be uploaded to the cloud for more accurate fusion judgment. ② When multiple edge nodes in the same physical area alarm simultaneously, or when the fault affects multiple different areas causing concurrent anomalies, since the data on a single node cannot reconstruct the full picture of the fault, it is necessary to aggregate features from multiple sources through knowledge fusion to confirm the type of problem (e.g., sensor false alarms, regional faults, or cross-regional system-level problems).

[0039] Figure 2 This is a flowchart illustrating a method for generating a scheduling strategy for a model evolution task according to an embodiment of the present invention. Figure 2As shown, the method includes steps S110 to S130. Specifically, the demand perception and resource measurement module and the task decomposition module can execute step S110, the multi-objective optimization engine module can execute step S120, and the solution set selection module can execute step S130. The multi-objective optimization engine module can obtain information contained in the scheduling requests of the decomposed subtasks from the task decomposition module, and obtain resource status information of each node in the cloud-edge collaborative network from the demand perception and resource measurement module; the solution set selection module selects one of the multiple scheduling solutions obtained from the multi-objective optimization engine module as the scheduling strategy; the solution set selection module transmits the scheduling strategy to the scheduler. The details are as follows: Step S110: Determine the scheduling request of the decomposed subtask corresponding to the model evolution task. The scheduling request of the decomposed subtask may include the type of model evolution task to which the decomposed subtask belongs (incremental learning task, knowledge fusion task, or parameter update task) and the scheduling resource requirements of the decomposed subtask. The scheduling resource requirements include at least one of the network performance requirements and at least one of the model efficiency requirements (so that the scheduling request of each decomposed subtask includes the task's requirements for model-specific attributes such as model accuracy and / or knowledge fusion efficiency).

[0040] Considering the large amount of data in the complete model evolution task, this application designs a data slice format for the data transmitted from the scheduled node to the scheduling node during model evolution task scheduling (hereinafter referred to as task transmission data). Each subtask corresponds one-to-one with a data slice, and this correspondence can be stored in the scheduled node. The scheduling resource requirements of each subtask are the resource requirements when the corresponding data slice is transmitted to the scheduling node. Decomposing the complete model evolution task into multiple independent, parallelizable subtasks allows for parallel processing of these subtasks across multiple nodes, significantly shortening the co-evolution cycle and improving the real-time performance and efficiency of co-evolution.

[0041] It should be noted that the data slices corresponding to each decomposed subtask are obtained based on the initial data slices. These initial data slices can be determined based on the type of model evolution task. For example, for incremental learning tasks, the task transmission data on the scheduled nodes is divided into multiple initial data slices according to data batches; for knowledge fusion tasks, the task transmission data is divided into multiple initial data slices according to the edge nodes where knowledge fusion will take place; for parameter update tasks, the task transmission data is divided into multiple initial data slices according to model layers or parameter groups, thus achieving the decomposition of the complete model evolution task. The above-mentioned methods for dividing the task transmission data for different types of evolution tasks are merely examples, and this invention is not limited to them.

[0042] In some embodiments of the present invention, step S110 may include the following two determination methods. The multiple decomposed subtasks obtained through these two methods do not have data synchronization or dependency relationships, and each decomposed subtask contains complete task configuration information and can be independently allocated to different computing nodes for processing: The scheduled node uses the initial data slice as a data slice, generates scheduling requests corresponding to each data slice (i.e., generates scheduling requests for multiple decomposed subtasks), and transmits it to the scheduling policy generator (specifically, the demand awareness and resource measurement module); or the scheduled node generates a complete scheduling request for the model evolution task and transmits it to the scheduling policy generator, whereby the task decomposition module divides it into multiple decomposed subtask scheduling requests. The second determination method can adjust the decomposition granularity of the task transmission data on the scheduled node according to the current cloud-edge collaborative network environment and the evolution task. Compared with the first determination method, it can reduce the number of decomposed subtasks and improve the resource utilization of the cloud-edge collaborative network. The second determination method will be described in detail below.

[0043] More specifically, determining the scheduling request for the decomposed subtasks corresponding to the model evolution task may include: the scheduling policy generator (specifically the task decomposition module) receiving the scheduling request for the model evolution task from edge nodes or cloud nodes; wherein, the scheduling request for the model evolution task includes the type of the model evolution task, the size of each initial data slice on the scheduled node, the network performance dimension requirement of the model evolution task, and the model efficiency dimension requirement of the model evolution task; wherein, the network performance dimension requirement and the model efficiency dimension requirement of the model evolution task refer to the resource requirements corresponding to at least one of the network performance indicators and at least one of the model efficiency indicators for the model evolution task; the number of task decompositions and the merging strategy of the initial data slices are determined based on the size of the initial data slices and the pre-stored node resource status (factors such as the number of cloud-edge collaborative network nodes may also be considered); the scheduling request for the model evolution task is divided into multiple scheduling requests for decomposed subtasks based on the number of task decompositions and the merging strategy of the initial data slices, so that each decomposed subtask corresponds to the data slice obtained after merging the initial data slices. The initial data slice merging strategy refers to the strategy for merging initial data slices to obtain new data slices. This strategy can be determined using classification models (such as support vector machines) or deep learning models. This application does not specify the number of task decompositions or the method for determining the merging strategy. For example, when determining the number of task decompositions based solely on the initial slice size, the following principles can be followed: when the task size is large, a smaller decomposition granularity is used to obtain more sub-tasks, thereby improving parallelism and resource utilization; when the task size is small, a larger decomposition granularity is used to obtain fewer sub-tasks, thereby reducing communication overhead.

[0044] As an example, the scheduling resource requirements in the scheduling requests of each decomposed subtask are determined based on the network performance and model efficiency requirements of the model evolution task in the scheduling request of the model evolution task. For instance, when multiple decomposed subtask scheduling requests are obtained based on the scheduling request of the model evolution task, the scheduling resource requirements in the scheduling requests can be decomposed as follows: For metrics related to the amount of data transmitted by the task (such as bandwidth and node energy consumption requirements), the corresponding requirements in each decomposed subtask are determined based on the size of the data slices corresponding to the decomposed subtask (obtained through the merging strategy of the initial data slices and the size of each initial data slice). For metrics unrelated to the amount of data transmitted by the task (such as latency and model accuracy), the relevant requirement information in the scheduling requests of the decomposed subtasks can be consistent with the relevant requirement information in the scheduling requests of the model evolution task. Furthermore, after the task decomposition module determines the task decomposition granularity, the size of each data slice can be determined based on the merging strategy of the initial data slices and the size of each initial data slice, and the size of each data slice can be added to the scheduling requests of each decomposed subtask. Alternatively, the identification information of the initial data slices corresponding to each decomposed subtask can be added to the scheduling requests of each decomposed subtask.

[0045] like Figure 3 As shown, before task decomposition, node resource status information and scheduling resource requirements are converted into the parameter format required by the multi-objective optimization algorithm. The process of converting perceived measurement data into parameters required by the multi-objective optimization algorithm is a mapping process that reconstructs qualitative business semantics into quantitative mathematical vectors. Through the demand perception and resource measurement module, features are extracted from the original scheduling requests and network-wide node resource status information, extracting core load indicators at the task level (such as specific data volume values ​​and hard thresholds for accuracy constraints) and real-time physical profiles at the node level (such as the current computing power load of each node and historical model accuracy). Furthermore, to address the problem of inconsistent physical units of these indicators preventing direct weighting by the algorithm, a linear normalization method can be used to map all values ​​to the dimensionless interval [0, 1], eliminating the interference of dimensional differences on the calculation weights. After converting the parameter format, these standardized values ​​can be encapsulated into node resource status vectors. Using these vectors as the direct input interface for the genetic algorithm allows the multi-objective optimization engine module to perform purely mathematical weighted calculations and iterative optimization without needing to understand complex power fault business logic; it only needs to substitute the values ​​in the vectors into the calculation formula of the multi-objective function for purely mathematical weighted calculations and iterative optimization. Through this format conversion, the method proposed in this application can quickly evaluate thousands of scheduling combinations in a multidimensional solution space, and finally accurately output a node allocation scheme that can satisfy both the soft and hard constraints of service quality and maximize the overall benefits of the system.

[0046] Step S120: Based on the scheduling requests of each decomposed subtask and the pre-stored node resource status information, multiple scheduling solutions are obtained using a multi-objective optimization algorithm, rather than simple load balancing or random allocation.

[0047] It should be noted that the resource types included in the pre-stored node resource status mentioned in this application must include the indicators corresponding to the scheduling resource requirements in the scheduling request, so as to ensure resource matching. For example, if the scheduling resource requirements include latency requirements, bandwidth requirements, and model accuracy requirements, the pre-stored node resource status must include node latency, network bandwidth, and the accuracy of the model deployed on the node. In addition, the pre-stored node resource status must also include resource status information needed for multi-objective function calculation, which can be designed independently and will not be elaborated here. Furthermore, the pre-stored node resource status includes the resource status information of edge nodes and cloud nodes in the cloud-edge collaborative network.

[0048] Model evolution involves multiple metrics that may have mutual constraints. This application designs and utilizes a genetic algorithm as a multi-objective optimization algorithm to generate a Pareto optimal set among multiple mutually constraining objectives, achieving a balance between model performance and resource utilization. This application may also employ other types of multi-objective optimization algorithms, such as multi-objective artificial bee colony optimization or multi-objective particle swarm optimization, but is not limited to these. The Pareto optimal set is an important concept in multi-objective optimization problems, used to describe the optimal set of solutions when multiple objectives cannot be optimized simultaneously. In this application, the genetic algorithm may include operations such as population initialization, fitness evaluation of multiple objectives, selection, crossover, and mutation. Furthermore, this application introduces preset node matching rules in the population initialization and mutation stages, thereby obtaining a population that conforms to the preset node matching rules in these two stages.

[0049] More specifically, this application also designs an intelligent matching process based on preset node matching rules during multi-objective optimization to ensure that model evolution tasks are executed on appropriate scheduling nodes. Intelligent matching is performed based on pre-stored node resource status information (such as the accuracy of the model deployed on the node and the number of model parameters) and scheduling resource requirements in the scheduling request (such as the accuracy requirements of sub-tasks and the efficiency requirements of knowledge fusion), achieving task allocation tailored to the characteristics of the model evolution task. Through this intelligent matching process, it can be ensured that model evolution tasks are executed efficiently on nodes with corresponding capabilities. For example, model evolution tasks with high accuracy requirements (such as some knowledge fusion tasks) can be preferentially matched to nodes with higher accuracy (such as cloud nodes where the deployed model accuracy is typically 0.9-0.95), while model evolution tasks with low latency requirements (such as some parameter update tasks) can be preferentially matched to nodes with lower latency (such as edge nodes where the network latency is typically 5-20ms).

[0050] As an example, if the model performance requirement in the scheduling resource requirements is a model accuracy requirement, and the network performance requirement is a latency requirement, the preset node matching rules may include: if the model accuracy requirement of the decomposed subtask meets a preset accuracy threshold (e.g., the model accuracy requirement is not less than 0.9), then the node matching order of the decomposed subtask is arranged in descending order of the accuracy of the models deployed on the nodes; if the latency requirement of the decomposed subtask meets a preset latency threshold (e.g., the latency requirement is not greater than 20ms), then the node matching order of the decomposed subtask is arranged in ascending order of the node latency. If the scheduling request of the decomposed subtask includes the size of each corresponding data slice, or the scheduling resource requirements include indicators related to the amount of data transmitted by the task (e.g., bandwidth), then the preset node matching rules may also include: if the data slice size corresponding to the decomposed subtask meets a preset data volume threshold (e.g., the data slice size is not less than 10MB), then the highest priority matching node for the decomposed subtask is a node whose bandwidth meets a preset bandwidth threshold (e.g., the node bandwidth is not less than 100Mbps) and / or a node belonging to the same physical region as the scheduled node (in this case, the communication cost between the matching node and the scheduled node is 0).

[0051] The above-described intelligent matching process based on resource requirements for model evolution tasks is merely an example. This application can also employ differentiated matching schemes when generating scheduling strategies based on different model evolution task types. For instance, the scheduling nodes for incremental learning tasks are mainly cloud nodes, which have high requirements for model accuracy and relatively low requirements for response latency. Therefore, the sub-tasks of incremental learning tasks can be designed to be preferentially matched to cloud nodes with higher accuracy. Similarly, the computation of knowledge fusion tasks mainly involves cloud nodes aggregating knowledge from multiple edge nodes, which also has high requirements for model accuracy. Therefore, the sub-tasks of knowledge fusion tasks can be designed to be preferentially assigned to cloud nodes with higher accuracy. Parameter update tasks have high latency requirements and need to quickly transmit updated parameters to the model on edge nodes. Therefore, the sub-tasks of parameter update tasks can be designed to be preferentially matched to edge nodes with lower latency. The relevant rules for matching scheduling resource requirements and node resource status information in the preset node matching rules can be designed independently, and this invention is not limited to these rules.

[0052] In some embodiments of the present invention, for the sake of simplicity, the NSGA-II algorithm, which integrates non-dominated sorting and congestion distance calculation, is used as an example to describe the specific process of step S120. Other genetic algorithms can be described by analogy. The NSGA-II algorithm can handle multiple conflicting objectives and find the Pareto optimal solution set through the non-dominated sorting mechanism, thereby providing a variety of optional scheduling strategies for the power system to support intelligent scheduling of co-evolutionary tasks. Figure 4 As shown, the multi-objective optimization engine module implements its functions through the following steps: Step S01, Population Initialization: For each decomposition subtask, based on the scheduling resource requirements of the decomposition subtask and the pre-stored node resource status information, nodes that meet the preset node matching rules are selected, and an initial population is generated based on the preset probability of randomly selecting individual solutions from the selected nodes and the preset probability of randomly selecting individual solutions from all nodes.

[0053] For example, for each decomposed subtask, the resource requirement information in its scheduling request can be judged, and nodes that meet the rules can be selected from the candidate node set according to the preset node matching rules. If there are nodes that meet the rules, a node is randomly selected from the nodes that meet the rules and assigned to the task (preset probability is 0.8). Otherwise, a node is randomly selected from all nodes (preset probability is 0.2) as the individual solution of the task.

[0054] Step S02, Multi-objective fitness evaluation: Multi-objective fitness evaluation is based on scheduling resource requirements, meaning the fitness evaluation metrics for multiple objectives must be related to the metrics included in the resource scheduling requirements. For example, when the model performance requirement is model accuracy, the network performance requirement is latency, and the scheduling request for decomposing subtasks also includes the data slice size corresponding to the decomposition subtasks, this application can design to evaluate the execution latency, model accuracy, and communication cost for each individual solution in the initial population to achieve multi-objective fitness evaluation. This application's design only needs to be able to evaluate the fitness of the population based on the information in the scheduling request. The evaluation metrics are not limited to these; according to the resource scheduling requirements, the fitness evaluation objectives can be set based on at least one of the model performance metrics and at least one of the network performance metrics, such as constructing a multi-objective fitness evaluation function based on model accuracy evaluation, bandwidth evaluation, latency evaluation, and energy consumption evaluation.

[0055] Step S03, Non-dominated sorting and crowding distance calculation: Based on the fitness assessment results, individuals in the population are divided into different frontiers according to non-dominated relationships, and the crowding distance of each individual solution in the population is calculated.

[0056] Step S04, selection operation: select a new population based on non-dominated sorting and crowding distance.

[0057] Step S05, crossover operation: Select parent individuals from the new population and generate new offspring individuals based on a preset crossover probability. For example, it can be set to prioritize retaining parent individuals that meet preset node matching rules.

[0058] Step S06, Mutation operation: Mutate the generated offspring individuals based on the preset mutation probability; wherein, the mutation probability of offspring individuals that conform to the preset node matching rules is greater than the mutation probability of offspring individuals that do not conform to the preset node matching rules.

[0059] For example, in the mutation operation, for each offspring individual to be mutated, the probability that each node (i.e., the individual solution) conforms to the preset node matching rule is calculated; the mutation probability of the node that conforms to the matching rule can be set to 0.7, and the mutation probability of the node that does not conform to the matching rule can be set to 0.3. Mutation selection can be performed according to the normalized probability.

[0060] Step S07, Merging and Sort: Merge parent and child individuals, perform non-dominated sorting and crowding distance calculation on the merged population, and select the next generation population from the merged population.

[0061] Step S08, Termination condition judgment: If the iteration termination condition is met, the non-dominated frontier in the next generation population is output as the scheduling solution set; otherwise, the crossover operation, mutation operation, and merging and sorting operation are repeated to finally output the Pareto optimal solution set.

[0062] As an example, the following describes step S02 using the multi-objective functions in multi-objective fitness evaluation, including the execution delay function, model accuracy function, and communication cost function. The value of the execution delay function is determined based on the number of decomposed subtasks assigned to the candidate scheduling node and the base delay of the candidate scheduling node; the value of the model accuracy function is determined based on the number of decomposed subtasks assigned to the candidate scheduling node and the accuracy of the model deployed on the candidate scheduling node; the value of the communication cost function is determined based on the data slice size corresponding to the decomposed subtasks assigned to the candidate scheduling node and the physical regions to which the scheduled node and the scheduling node belong.

[0063] More specifically, the allocation problem of multiple decomposition subtasks in step S120 is modeled as a multi-objective optimization problem. Assume the number of decomposition subtasks corresponding to the model evolution task is... The number of nodes in the cloud-edge collaborative network (including the number of edge nodes and the number of cloud nodes) is: This application designs a communication cost objective function. Model accuracy objective function and the time delay objective function We construct a multi-objective function to select a suitable subtask scheduling node from the Pareto optimal solution set by simultaneously minimizing communication cost, maximizing model accuracy, and minimizing maximum response latency. This ensures efficient execution of co-evolution while controlling communication cost and response latency.

[0064] ① Communication cost objective function The goal is to minimize the total communication cost. In this application, the communication cost primarily comes from the data transmitted from the scheduled node to the scheduling node during model evolution task scheduling. Therefore, if a certain decomposed subtask A is assigned to a node... Then the communication cost objective function of the decomposed subtask A is... It can be defined as: ; in, Represents a node The number of subtasks assigned. This represents the size of the data slice corresponding to each decomposition subtask (in reality, the size of the data slice corresponding to each subtask may be different; this is a theoretical value). Represents a node The communication factor. For example, if a node The scheduled node corresponding to subtask A is located in the same physical region (local node). If node If the scheduled node corresponding to subtask A is not located in the same physical region, then the settings are based on the communication network type. (such as 5G power private networks) HPLC network NB-IoT network ).

[0065] ② Model accuracy objective function The goal is to maximize the overall model accuracy, a unique optimization objective of model evolution tasks. This application designs a multi-objective optimization algorithm to minimize the multi-objective function, thus negating the model accuracy objective function. If a certain decomposed subtask A is assigned to a node... Then the objective function for the model accuracy of the decomposed subtask A is... It can be defined as: ; in, This represents the number of all subtasks corresponding to the model evolution task. For nodes The accuracy of the model deployed on it.

[0066] ③Time Delay Objective Function The goal is to minimize the maximum latency of a node. If a subtask A is assigned to a node... Then the time delay objective function of the decomposed subtask A is... It can be defined as: ; in, For nodes The base latency is determined based on network transmission latency and queue waiting time.

[0067] Step S130: Select one of the multiple candidate scheduling schemes as the scheduling strategy for the model evolution task and output it. The scheduling strategy is then sent to the scheduled nodes, and the scheduled nodes send data slices to the corresponding scheduling nodes. Furthermore, when the scheduling strategy is generated and distributed to the scheduled nodes, the scheduling strategy (e.g., using a correspondence between scheduling nodes and data slices) and the initial data slice merging strategy can be distributed to the scheduled nodes to support the co-evolution of cloud and edge models. Additionally, for the knowledge fusion task, if the scheduling request contains the identification information of all edge nodes to be fused, the relevant information of the scheduling nodes can be distributed to these edge nodes, enabling them to transmit feature vectors (i.e., knowledge) to the scheduling node. If the scheduling request does not contain the identification information of the nodes to be fused, the scheduling strategy can be distributed to the edge node that sent the scheduling request, allowing that node to obtain knowledge from other nodes and transmit it to the scheduling node.

[0068] As an example, this application can introduce a weighted adaptive fusion strategy to support the customization of weight factors for different objectives in different power application scenarios, so as to select the most suitable task allocation strategy for the current user needs from multiple candidate scheduling schemes obtained by multi-objective optimization algorithms. For example, latency is given priority in fault emergency response scenarios, and model accuracy is given priority in equipment status assessment scenarios, thereby further improving the adaptability of co-evolution.

[0069] Based on a multi-objective function and a communication cost objective function Model accuracy objective function and delay objective function Taking the multiple scheduling solutions obtained using a genetic algorithm as an example, step S130 can perform multi-objective fitness evaluation operations on the multiple scheduling solutions respectively, and select one solution set from the multiple scheduling solutions as the scheduling strategy for the model evolution task based on the evaluation results. Specifically, the weight adaptive fusion strategy can introduce communication cost weights according to the type of objective function. Accuracy weighting and latency Weights (the sum of these three weights is 1). For each scheduling solution set obtained by the NSGA-II algorithm, after obtaining the multi-objective fitness evaluation results, calculate the comprehensive score of each scheduling solution set. The solution set with the highest overall score is selected as the scheduling strategy for the model evolution task. Among these, , and These are standardized communication costs, model accuracy, and response latency, respectively. Indicates the first The overall score of each scheduling solution set Indicates the first The sum of the standardized communication costs of all individual solutions in a given scheduling solution set. Indicates the first The sum of the standardized model accuracies of all individual solutions in a given scheduling solution set. Indicates the first The sum of the standardized response delays of all individual solutions in a scheduling solution set.

[0070] The weight factors can also be set according to the type of model evolution task. For example, the weights for incremental learning tasks are biased towards accuracy, the weights for knowledge fusion tasks are biased towards accuracy, and the weights for parameter update tasks are biased towards latency. The weight factors of the weight adaptive fusion strategy can be pre-stored in the scheduling strategy generator. The above method of selecting a scheduling strategy based on multi-objective evaluation through the weight adaptive fusion strategy is only an example, and other methods can also be used.

[0071] In some embodiments of the present invention, the scheduler can monitor the execution status of each decomposed subtask (the execution status may include not executed, executing, completed, or abnormal). When a node fails or is overloaded, a dynamic task migration mechanism will be triggered, so that the task of that node is reassigned to other available nodes. Therefore, the method further includes: when the scheduler detects an abnormality in a specific decomposed subtask, it can send an execution abnormality instruction to the scheduling policy generator, so that the scheduling policy generator reselects available scheduling nodes for the specific decomposed subtask from other candidate scheduling schemes besides the selected candidate scheduling schemes. For example, the scheduling policy generator performs a multi-objective fitness evaluation operation on the candidate scheduling nodes of the specific decomposed subtask in other scheduling solution sets besides the selected solution set, thereby reselecting available scheduling nodes for the specific decomposed subtask based on the evaluation results.

[0072] The scheduling strategy generation method proposed in this application can adopt differentiated task decomposition methods, node matching rules, and select appropriate scheduling strategies from multiple scheduling solution sets for different types of model evolution tasks, thereby achieving unified scheduling of model evolution tasks.

[0073] The scheduling strategy generation method proposed in this application has the following significant advantages: ① Achieving intelligent matching and adaptive scheduling for model evolution tasks: Existing technologies primarily rely on static metrics (such as CPU utilization and memory usage) for scheduling decisions, lacking awareness of the performance of deployed models on nodes and failing to match tasks based on model-specific attributes such as model accuracy and the number of model parameters. This invention, through a demand awareness and resource measurement module, can intelligently match and adaptively allocate tasks based on node resource status (including model performance) and resource requirements of model evolution tasks (such as requirements for model accuracy and knowledge fusion efficiency). The unified measurement system for resource status and task requirements, covering multiple dimensions such as computing power, network, and business, constructed in this invention can accurately characterize the resource requirements of model evolution tasks on the cloud-edge collaborative network and the heterogeneous resource characteristics of edge nodes. Furthermore, this application also designs an intelligent matching strategy to ensure that each task is assigned to a scheduling node suitable for its characteristics; for example, high-precision model evolution tasks are preferentially assigned to nodes with higher precision, and low-latency model evolution tasks are preferentially assigned to nodes with lower latency.

[0074] ② This invention achieves task decomposition and subtask allocation for model evolution tasks, improving resource utilization in cloud-edge collaborative networks: It decomposes different types of complete model evolution tasks into multiple independently executable subtasks according to specific decomposition methods (e.g., incremental model learning tasks decomposed by data batches, knowledge fusion tasks decomposed by edge nodes, parameter update tasks decomposed by model layers or parameter groups). These parallel-executable subtasks are then allocated to different edge and cloud nodes. Compared to existing schemes that allocate complete model evolution tasks as atomic units to single nodes, which fail to fully utilize the collaborative computing capabilities of multiple nodes, this invention, through its task decomposition mechanism, fully utilizes heterogeneous computing resources from TOPS to MOPS, improving the resource utilization of power systems. Furthermore, through multi-node parallel processing, it significantly shortens task execution time.

[0075] ③ Balancing Multiple Mutually Constraining Objectives in Model Evolution Tasks Using Multi-Objective Optimization Algorithms: Most existing task scheduling schemes employ single-objective optimization methods, which struggle to balance multiple factors, making them unsuitable for real-world scenarios. This application provides a multi-objective optimization algorithm based on genetic algorithms (such as the NSGA-II algorithm based on non-dominated sorting and crowding distance calculation). This algorithm can simultaneously optimize multiple key objectives such as communication cost, model accuracy, and response latency, obtaining Pareto optimal solution sets under multiple different optimization scenarios. This allows multiple subtasks to be assigned to different edge nodes and cloud nodes. Furthermore, this invention can also adjust the weights of different optimization objectives according to actual needs through a weight adaptive fusion strategy, thereby selecting the most suitable scheduling scheme from the Pareto optimal solution set and achieving balanced optimization of multiple objectives.

[0076] ④ Achieving Cloud-Edge Collaborative Optimization for Model Evolution Tasks: Existing model evolution tasks are completed in the cloud or at the edge, failing to fully utilize the node resources of cloud-edge collaborative models and networks. This invention, through its proposed scheduling strategy generation method, can rationally allocate model evolution tasks to edge nodes and cloud nodes for scheduling. This invention, through resource-demand matching, task decomposition, and adaptive scheduling mechanisms, dynamically adjusts the task allocation strategy based on the real-time resource status of nodes (such as computing power, network performance, and model accuracy) and the requirements of the model evolution task. This fully leverages the advantages of edge computing and deployed models to achieve load sharing and reduced response latency.

[0077] ⑤ By incorporating communication cost and response latency into the objective function of the multi-objective optimization process, it is possible to reduce communication cost to improve data transmission efficiency while simultaneously reducing response latency to meet the real-time requirements of model evolution: Through a task decomposition mechanism and a multi-objective optimization algorithm, under the constraint of optimizing communication cost, this invention can prioritize the allocation of model evolution tasks to edge nodes for local processing, reducing data transmission's consumption of network bandwidth. Furthermore, under latency constraints, latency-sensitive model evolution tasks (such as parameter update tasks, which typically require fast responses) can be allocated to local edge nodes or nodes with low network latency, achieving millisecond-level response latency and reducing end-to-end network transmission time.

[0078] Corresponding to the above method, the present invention also provides a scheduling strategy generation system for cloud-edge collaborative model evolution tasks. The system includes a computer device, which includes a processor and a memory. The memory stores computer programs / instructions, and the processor is used to execute the computer programs / instructions stored in the memory. When the computer programs / instructions are executed by the processor, the system implements the steps of the method described above.

[0079] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0080] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0081] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0082] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for generating scheduling strategies for cloud-edge collaborative model evolution tasks, characterized in that, The method includes the following steps: Determine the scheduling requests for the decomposed subtasks corresponding to the model evolution task; wherein, during the scheduling of the model evolution task, the data transmitted from the scheduled node to the scheduling node is in the form of data slices, and the decomposed subtasks correspond one-to-one with the data slices; the scheduling request for the decomposed subtasks includes the type of the model evolution task to which it belongs and the scheduling resource requirements, and the dimensions of the scheduling resource requirements include network performance dimension and model efficiency dimension. Based on the scheduling request and pre-stored node resource status information, a multi-objective optimization algorithm is used to obtain multiple candidate scheduling schemes; wherein, each candidate scheduling scheme includes a candidate scheduling node for all decomposed subtasks corresponding to the model evolution task; the fitness evaluation of the multi-objective is based on the scheduling resource requirements; the scheduled node, the scheduling node, and the candidate scheduling node are cloud nodes or edge nodes; One of the multiple candidate scheduling schemes is selected as the scheduling strategy for the model evolution task and output. The scheduling strategy is then sent to the scheduled node, and the scheduled node sends the data slice to the corresponding scheduling node.

2. The method according to claim 1, characterized in that, The scheduling request for determining the decomposed subtasks corresponding to the model evolution task includes: Receive scheduling requests for model evolution tasks from edge nodes or cloud nodes; wherein, the scheduling request for the model evolution task includes the type of model evolution task, the size of each initial data slice on the scheduled node, the network performance dimension requirements of the model evolution task, and the model efficiency dimension requirements of the model evolution task; wherein, the initial data slice is determined based on the type of model evolution task; The number of task decompositions and the merging strategy of the initial data slices are determined based on the size of the initial data slices and the pre-stored node resource status. Based on the number of task decompositions and the merging strategy of the initial data slices, the scheduling request of the model evolution task is divided into scheduling requests of multiple decomposed subtasks. The merging strategy of the initial data slices is a strategy of merging the initial data slices to obtain the data slices. The scheduling request of each decomposed subtask is determined based on the scheduling request of the model evolution task.

3. The method according to claim 1, characterized in that, The multi-objective function of the multi-objective optimization algorithm includes an execution delay function and a model accuracy function; The value of the execution delay function is determined based on the number of decomposed subtasks assigned to the candidate scheduling node and the base delay of the candidate scheduling node; the value of the model accuracy function is determined based on the number of decomposed subtasks assigned to the candidate scheduling node and the accuracy of the model deployed on the candidate scheduling node.

4. The method according to claim 3, characterized in that, The scheduling request for the decomposed subtask also includes the data slice size corresponding to the decomposed subtask; The multi-objective function of the multi-objective optimization algorithm also includes a communication cost function; the value of the communication cost function is determined based on the data slice size corresponding to the decomposed subtasks assigned to the candidate scheduling node and the physical regions to which the scheduled node and the scheduling node belong.

5. The method according to claim 4, characterized in that, The multi-objective optimization algorithm is a genetic algorithm; Based on the scheduling request and pre-stored node resource status information, a multi-objective optimization algorithm is used to obtain multiple candidate scheduling schemes, including: Initialize the population: For each decomposed subtask, based on the scheduling resource requirements of the decomposed subtask and the pre-stored node resource status information, select nodes that meet the preset node matching rules, and generate an initial population based on the preset probability of randomly selecting individual solutions from the selected nodes and the preset probability of randomly selecting individual solutions from all nodes. Multi-objective fitness assessment: The fitness of each individual solution in the initial population is assessed using a multi-objective function; Non-dominated ordination and crowding distance calculation: Based on the fitness assessment results, individuals in the population are divided into different frontiers according to non-dominated relationships, and the crowding distance of each individual in the population is calculated; Selection operation: Select a new population based on non-dominated ordering and crowding distance; Crossover operation: Select parent individuals from the new population and generate new offspring individuals based on a preset crossover probability; Mutation operation: Mutate the generated offspring individuals based on the preset mutation probability; wherein, the mutation probability of offspring individuals that conform to the preset node matching rules is greater than the mutation probability of offspring individuals that do not conform to the preset node matching rules. Merging and sorting: Merge parent and child individuals, perform non-dominated sorting and crowding distance calculation on the merged population, and select the next generation population from the merged population; Termination condition judgment: If the iteration termination condition is met, output the non-dominated frontier in the next generation population as the scheduling solution set; otherwise, repeat the crossover operation, mutation operation, and merging and sorting.

6. The method according to claim 5, characterized in that, Model performance requirements include model accuracy requirements, and network performance requirements include latency requirements. The preset node matching rules also include: If the model accuracy requirement of the decomposed subtask meets the preset accuracy threshold, the node matching order of the decomposed subtask is arranged in descending order of the accuracy of the models deployed on the nodes. If the latency requirements of the decomposed subtasks meet the preset latency threshold, then the node matching order of the decomposed subtasks is arranged in ascending order of node latency; and If the size of the data slice corresponding to the decomposed subtask meets the preset data volume threshold, then the highest priority matching node for the decomposed subtask is a node whose bandwidth meets the preset bandwidth threshold and / or a node belonging to the same physical region as the scheduled node.

7. The method according to claim 6, characterized in that, Model evolution tasks can be categorized as incremental learning tasks, knowledge fusion tasks, or parameter update tasks. The initial data slices were obtained in the following way: For incremental learning tasks, each batch of data on the scheduled node is used as an initial data slice; For knowledge fusion tasks, the knowledge on each edge node to be fused is treated as a data slice. as well as For parameter update tasks, the updated parameters of each model layer on the cloud model after parameter update are taken as a data slice; The preset node matching rule also includes: the scheduling nodes for all decomposed subtasks corresponding to the knowledge fusion task are the same cloud node; The scheduling node for the parameter update task is any node other than the node being scheduled.

8. The method according to claim 1, characterized in that, The method further includes: when an anomaly is detected in a specific decomposed subtask, reselecting an available scheduling node for the specific decomposed subtask from other candidate scheduling schemes besides the selected candidate scheduling scheme.

9. A scheduling strategy generation system for cloud-edge collaborative model evolution tasks, comprising a processor, a memory, and a computer program / instructions stored in the memory, characterized in that, The processor is configured to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 8.

Images