Cloud edge end collaborative lightweight AI intelligent scheduling method and system and storage medium

Abstract

This invention relates to the field of artificial intelligence technology, and discloses a cloud-edge-device collaborative lightweight AI intelligent scheduling method, system, and storage medium. The method includes the following steps: Step S1: Acquire the data to be processed on the terminal side, and perform preliminary inference based on the lightweight AI model on the terminal side. If the preliminary inference result does not meet preset conditions, generate a task to be scheduled; Step S2: Acquire the node status information of cloud nodes and edge nodes, and generate candidate execution configurations; Step S3: Filter the candidate execution configurations to determine the executable configuration; Step S4: Determine the target execution configuration based on the executable configuration, and control the corresponding target node to call the corresponding lightweight model to execute the task to be scheduled. Update the node status information and lightweight model configuration based on the execution result. This invention reduces the execution delay and scheduling failure caused by unreasonable task allocation, and improves the efficiency, stability, and reliability of lightweight AI task execution.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a lightweight AI intelligent scheduling method, system and storage medium for cloud-edge-device collaboration. Background Technology

[0002] With the development of artificial intelligence technology, an increasing number of intelligent analysis tasks are being applied to scenarios such as industrial inspection, intelligent monitoring, and edge sensing. In these application scenarios, terminal devices typically need to process the collected data in real time to reduce data transmission pressure and meet the requirements for rapid response. However, limited by the computing power, storage capacity, and power consumption of terminal devices, the terminal side is usually unable to undertake the inference tasks of complex artificial intelligence models for extended periods. Therefore, a data processing model involving terminal devices, edge nodes, and cloud nodes has gradually emerged.

[0003] In existing cloud-edge-device collaborative processing methods, some AI tasks are typically migrated to edge nodes or cloud nodes for execution, depending on the task scale or device resources. For terminal devices with limited computing resources, some preliminary analysis can be completed by deploying lightweight AI models to reduce data transmission volume and improve response efficiency. However, in actual operation, different AI tasks have different requirements for inference accuracy, execution latency, and computing resources. The computing power and task load status of different edge nodes and cloud nodes are also dynamically changing, making fixed task allocation methods difficult to adapt to complex application environments.

[0004] Furthermore, existing task scheduling methods typically focus on selecting task execution locations based on node computing power or network status, without fully considering the matching relationship between the lightweight model configuration corresponding to the task and the execution node status. When there is a mismatch between the task execution node, model configuration, and node resource status, problems such as insufficient model inference accuracy, excessive node resource consumption, or reduced task execution efficiency can easily occur. At the same time, when node status changes dynamically, the execution plan determined during the scheduling process may fail to meet the task requirements in the actual execution stage, affecting the stability of cloud-edge-device collaborative processing.

[0005] Therefore, how to rationally determine the execution method of a task in a cloud-edge-device collaborative environment, taking into account the needs of artificial intelligence tasks and the resource status of different nodes, and improve the resource matching capability and execution reliability in the lightweight AI task scheduling process, has become an urgent problem to be solved. Summary of the Invention

[0006] This invention provides a lightweight AI intelligent scheduling method, system, and storage medium for cloud-edge-device collaboration, which at least solves the problems of insufficient AI task scheduling and node resource adaptation in the cloud-edge-device environment in the prior art, resulting in low task execution efficiency and insufficient scheduling stability.

[0007] To achieve the above objectives, a first aspect of the present invention provides a lightweight AI intelligent scheduling method for cloud-edge-device collaboration, the method comprising the following steps: Step S1: Obtain the data to be processed on the terminal side and perform preliminary reasoning based on the lightweight AI model on the terminal side. If the preliminary reasoning result does not meet the preset conditions, generate a task to be scheduled. Step S2: Obtain the node status information of cloud nodes and edge nodes, and generate candidate execution configurations based on the node status information and the lightweight model configuration corresponding to the task to be scheduled; Step S3: Based on the task constraint information corresponding to the task to be scheduled, filter the candidate execution configurations to determine the executable configuration; Step S4: Determine the target execution configuration based on the executable configuration, and control the corresponding target node to call the corresponding lightweight model to execute the scheduled task, and update the node status information and lightweight model configuration based on the execution result.

[0008] To achieve the above objectives, a second aspect of the present invention provides a cloud-edge-device collaborative lightweight AI intelligent scheduling system, the system being used to execute the above-described cloud-edge-device collaborative lightweight AI intelligent scheduling method, the system comprising: The acquisition module is used to acquire data to be processed on the terminal side and perform preliminary inference based on the lightweight AI model on the terminal side. When the preliminary inference results do not meet the preset conditions, a task to be scheduled is generated. The candidate module is used to obtain node status information of cloud nodes and edge nodes, and generate candidate execution configurations based on the node status information and the lightweight model configuration corresponding to the task to be scheduled. The configuration module is used to filter the candidate execution configurations based on the task constraint information corresponding to the task to be scheduled, and determine the executable configuration; The execution module is used to determine the target execution configuration based on the executable configuration, control the corresponding target node to call the corresponding lightweight model to execute the scheduled task, and update the node status information and lightweight model configuration based on the execution result.

[0009] The beneficial effects of this invention are as follows: By performing preliminary inference on the data to be processed on the terminal side and generating tasks to be scheduled when processing requirements cannot be met, the computing power of the terminal side is rationally utilized; simultaneously, candidate execution configurations are generated by combining the state information of cloud nodes and edge nodes with lightweight model configurations, and filtered according to task constraints, thereby improving the matching degree between tasks and execution resources; furthermore, by updating node state information and lightweight model configurations based on execution results, the scheduling process can adapt to changes in node state, improving the stability and reliability of AI task execution in a cloud-edge-device collaborative environment. This method can reduce task execution delays and scheduling failures caused by resource mismatches, and improve the execution efficiency of lightweight AI tasks in complex computing environments. Attached Figure Description

[0010] Figure 1 This is a schematic diagram illustrating the execution flow of the cloud-edge-device collaborative lightweight AI intelligent scheduling method in an embodiment of the present invention. Figure 2 This is a schematic diagram of the cloud-edge-device collaborative lightweight AI task scheduling process in an embodiment of the present invention; Figure 3 This is a schematic diagram of the operational architecture of the cloud-edge-device collaborative lightweight AI intelligent scheduling system in an embodiment of the present invention. Detailed Implementation

[0011] 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 accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0012] This invention provides a lightweight AI intelligent scheduling method for cloud-edge-device collaboration, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the lightweight AI intelligent scheduling method for cloud-edge-device collaboration according to an embodiment of the present invention.

[0013] In this embodiment, a lightweight AI intelligent scheduling method for cloud-edge-device collaboration includes the following steps: S1: Obtain the data to be processed on the terminal side and perform preliminary reasoning based on the lightweight AI model on the terminal side. If the preliminary reasoning result does not meet the preset conditions, generate a task to be scheduled.

[0014] Specifically, the system acquires the data to be processed collected from the terminal side, preprocesses the data, inputs the preprocessed data into the lightweight AI model on the terminal side for inference, and obtains the corresponding inference result and confidence level. It then determines whether the confidence level meets the preset confidence conditions. When the confidence level meets the preset confidence conditions, the inference result is output as the terminal processing result. When the confidence level does not meet the preset confidence conditions, a task to be scheduled is generated based on the data to be processed, the inference result, and the corresponding task identifier, and the task to be scheduled is sent to the scheduling node.

[0015] Specifically, the process involves inputting the preprocessed data into a lightweight AI model on the input side for inference to obtain corresponding inference results and confidence levels. This includes: extracting features from the preprocessed data into the lightweight AI model on the input side to obtain corresponding target feature information; wherein the lightweight AI model is an AI inference model pre-configured based on the terminal-side computing resources, and the lightweight AI model corresponds to a preset model parameter scale and inference accuracy level, used for feature extraction and preliminary inference on the preprocessed data, and outputting corresponding inference results and confidence levels; performing inference based on the target feature information to obtain the inference results corresponding to the data to be processed, and determining the corresponding confidence level based on the inference results.

[0016] In this embodiment of the invention, data to be processed is acquired from the terminal side and preprocessed to obtain target input data that meets the requirements of terminal-side inference. The data to be processed includes at least one of image data, audio data, and sensor-acquired data. The preprocessing includes data format conversion, size adjustment, and noise filtering to make the target input data conform to the data input format of the lightweight AI model on the terminal side.

[0017] Furthermore, the target input data is input into a lightweight AI model on the terminal side for inference, obtaining the corresponding inference results and confidence levels. The lightweight AI model is an AI inference model pre-configured based on the terminal-side computing resources. This lightweight AI model corresponds to a preset model parameter scale and inference accuracy level, used to complete the preliminary analysis of the data to be processed on the terminal side. The model parameter scale characterizes the model's computational and storage requirements, and the inference accuracy level characterizes the model's ability to meet the task recognition requirements.

[0018] Specifically, after the target input data is input into the lightweight AI model, feature extraction is first performed on the target input data to obtain target feature information. Inference processing is then performed based on the target feature information to obtain the corresponding inference result, and the corresponding confidence level is determined based on the inference result.

[0019] For example, in an industrial equipment image inspection scenario, an image of the equipment surface captured by a terminal-side camera is acquired as data to be processed. After size normalization and image noise suppression, it is input into a lightweight terminal-side AI model. The lightweight terminal-side AI model extracts features from the equipment image to obtain target feature information including equipment area features and defect texture features. Based on the target feature information, it outputs a defect category judgment result and its corresponding confidence level. When the output result is "defect exists" and the confidence level is 0.96, the inference result is determined to meet the preset confidence condition; when the output result corresponds to a confidence level of 0.62, while the confidence threshold in the preset confidence condition is 0.85, the current terminal-side inference result is determined to not meet the preset condition.

[0020] The system determines whether the confidence level meets a preset confidence condition. If the confidence level meets the preset confidence condition, the inference result is determined as the edge processing result, and the scheduling process for the current data to be processed ends. If the confidence level does not meet the preset confidence condition, a task to be scheduled is generated based on the data to be processed, the inference result, and the corresponding task identifier. The task to be scheduled is then sent to the scheduling node to proceed to the subsequent execution configuration determination process based on node status information and lightweight model configuration.

[0021] The pre-set confidence conditions are determined based on specific application scenarios and are used to determine whether the output results of the lightweight AI model on the edge meet the requirements for direct application. These pre-set confidence conditions are not limited to a single confidence threshold, but also include requirements for accurate identification, anomaly identification, or result completeness corresponding to the task type. Through this method, the subsequent scheduling process only proceeds when the edge-side inference results cannot meet the task requirements, reducing unnecessary data transmission and computational scheduling, and improving resource utilization efficiency in the cloud-edge-edge collaborative processing process. In the above embodiments, the lightweight AI model on the edge, the preprocessing method, and the pre-set confidence conditions are all specific configuration forms for implementing preliminary edge-side inference, used to limit the generation conditions of the tasks to be scheduled, but do not constitute limitations on specific application scenarios and model structures.

[0022] Step S2: Obtain the node status information of cloud nodes and edge nodes, and generate candidate execution configurations based on the node status information and the lightweight model configuration corresponding to the task to be scheduled.

[0023] Specifically, the node status information corresponding to the cloud node and the edge node is obtained, and the node status information includes the node computing resource status and task execution status; according to the lightweight model configuration corresponding to the task to be scheduled, multiple model configuration items matching the task to be scheduled are determined; each of the node status information is combined with the corresponding model configuration item to generate a corresponding candidate execution configuration, and the candidate execution configuration includes the execution node and the corresponding lightweight model configuration.

[0024] Specifically, based on the lightweight model configuration corresponding to the task to be scheduled, multiple model configuration items matching the task to be scheduled are determined, including: obtaining a set of model configurations corresponding to the task to be scheduled in the lightweight model configuration according to the data type and task requirements of the task to be scheduled; determining multiple model configuration items from the set of model configurations, each of the model configuration items including at least a model identifier, a model parameter scale, and a corresponding inference accuracy level.

[0025] In this embodiment of the invention, node status information corresponding to cloud nodes and edge nodes is obtained. This node status information characterizes the current task processing capacity and task load of each candidate execution node, including node computing resource status and task execution status. Specifically, the node computing resource status characterizes the node's current computing capacity available for executing artificial intelligence tasks, including the amount of available computing resources, resource occupancy ratio, and remaining computing resources required for model inference. The task execution status characterizes the node's current task processing status, including the number of currently executing tasks, the task waiting queue, and task resource occupancy.

[0026] Furthermore, based on the lightweight model configuration corresponding to the task to be scheduled, multiple model configuration items matching the task to be scheduled are determined. Specifically, based on the data type, task processing requirements, and corresponding inference accuracy requirements of the task to be scheduled, a set of model configurations corresponding to the task to be scheduled is obtained from the pre-configured lightweight model configurations, and multiple model configuration items are determined from the set of model configurations.

[0027] Each of the model configuration items describes a lightweight model execution method suitable for the scheduled task, and includes at least a model identifier, model parameter size, and corresponding inference accuracy level. The model identifier distinguishes different model configuration items, the model parameter size characterizes the computational and storage requirements of the corresponding lightweight model, and the inference accuracy level characterizes the ability of the corresponding lightweight model's output to meet the task requirements. For example, for the same image recognition task, three model configuration items with pre-configured model identifiers M1, M2, and M3 are pre-configured, where M1 corresponds to a smaller model parameter size and a lower inference accuracy level, M2 corresponds to a medium model parameter size and a medium inference accuracy level, and M3 corresponds to a larger model parameter size and a higher inference accuracy level. The corresponding model configuration set is obtained according to the requirements of the scheduled task.

[0028] Furthermore, the node status information corresponding to each cloud node and edge node is combined with the corresponding model configuration item to generate candidate execution configurations. Specifically, a node status information is associated with a model configuration item to form a corresponding candidate execution configuration. Each candidate execution configuration is used to characterize the execution method of the task to be scheduled when the specified lightweight model configuration is invoked at the specified execution node.

[0029] For example, for the same scheduled task, obtain the node status information corresponding to edge node A, edge node B and cloud node C, and combine it with model configuration items M1, M2 and M3 to form candidate execution configurations such as "edge node A calls model M1", "edge node B calls model M2" and "cloud node C calls model M3". Subsequently, each candidate execution configuration is filtered according to the task constraint information.

[0030] The candidate execution configuration provides alternative execution methods during task scheduling. The specific number of nodes, the number of model configurations, and the scale of model parameters are determined based on the actual application scenario. By jointly combining node status information with lightweight model configurations, the task execution method is not determined solely based on node computing power or the accuracy of a single model. This provides a data foundation for subsequently determining the target execution configuration that meets the execution requirements based on task constraints.

[0031] Preferably, a correlation analysis is further performed on the historical task execution records corresponding to different node state information to construct an adaptation relationship between node state changes and lightweight model configuration. Specifically, the changes in node resources, the invoked model configuration, and the task completion results corresponding to the historical task execution process are obtained. Different node states and model configurations under the same type of task are combined to form historical execution samples, and the optimal model configuration range under different node state conditions is determined based on the historical execution samples.

[0032] For example, for the same type of image recognition task, when an edge node has 40% to 60% of available computing resources, historical execution results show that a model with medium parameter size can meet the task accuracy requirements. However, when available computing resources are below 40%, a model configuration with smaller parameter size is preferred. Based on the above adaptation relationship, when generating candidate execution configurations for the task to be scheduled, node and model combinations that conform to historical adaptation patterns are prioritized before proceeding to the subsequent candidate execution configuration selection process.

[0033] Step S3: Based on the task constraint information corresponding to the task to be scheduled, filter the candidate execution configurations to determine the executable configuration.

[0034] Specifically, based on the task constraint information, the lightweight model configurations corresponding to each candidate execution configuration are matched and verified to determine the candidate execution configurations that meet the inference requirements of the task to be scheduled; based on the node status information of the target execution node corresponding to the candidate execution configuration, the resource carrying capacity of the target execution node is verified to determine the candidate execution configurations that meet the task execution resource requirements; based on the task execution conditions corresponding to the candidate execution configuration, the execution conditions of the candidate execution configuration are verified, and the candidate execution configurations that pass all verifications are determined as executable configurations.

[0035] Specifically, the process of performing execution condition verification on the candidate execution configuration includes: determining whether the task completion time limit and resource usage requirements corresponding to the candidate execution configuration meet the preset execution constraints based on the target execution node status information and the task constraint information; and determining that the candidate execution configuration passes the execution condition verification when the task completion time limit and resource usage requirements corresponding to the candidate execution configuration meet the preset execution constraints.

[0036] In this embodiment of the invention, based on the task constraint information corresponding to the task to be scheduled, the lightweight model configurations corresponding to each candidate execution configuration are matched and verified to determine whether the lightweight models in the candidate execution configurations meet the inference requirements of the task to be scheduled. The inference requirements characterize the requirements of the task to be scheduled for the model's processing capabilities, including information such as model inference accuracy level, task processing type, and output result requirements. When the lightweight model configuration corresponding to the candidate execution configuration meets the inference requirements, the corresponding candidate execution configuration is retained; when the lightweight model configuration cannot meet the inference requirements, the corresponding candidate execution configuration is deleted.

[0037] Furthermore, based on the node status information of the target execution node corresponding to the candidate execution configuration after model matching verification, the resource carrying capacity of the target execution node is verified to determine whether the target execution node has the resource conditions to execute the corresponding lightweight model configuration. The resource carrying capacity characterizes the computing resources, storage resources, and task processing capabilities that the target execution node can currently provide. Specifically, the current available resource status of the target execution node is matched with the resources required by the corresponding lightweight model configuration. When the available resources of the target execution node meet the requirements for model invocation and task processing, the corresponding candidate execution configuration is retained; otherwise, the corresponding candidate execution configuration is discarded.

[0038] Furthermore, based on the task execution conditions corresponding to the candidate execution configuration, the candidate execution configuration is validated for execution conditions. Specifically, based on the target execution node status information and the task constraint information corresponding to the candidate execution configuration, the task completion time limit and resource usage requirements corresponding to the candidate execution configuration are determined, and it is judged whether the task completion time limit and resource usage requirements meet the preset execution constraints.

[0039] The preset execution constraints are used to limit the conditions under which candidate execution configurations meet the task execution requirements. For example, for video detection tasks with high real-time requirements, the task completion time is set to no more than 100ms, and the resource usage requirement is limited to no more than 80% of the currently allocable computing resources of the target execution node. When the expected completion time of the task corresponding to the candidate execution configuration is 80ms and the expected resource utilization rate is 70%, the candidate execution configuration is determined to meet the preset execution constraints; when the expected completion time of the task reaches 150ms, or the expected resource utilization rate reaches 90%, the candidate execution configuration is determined to not meet the preset execution constraints.

[0040] When a candidate execution configuration passes the lightweight model matching verification, resource carrying capacity verification, and execution condition verification, the corresponding candidate execution configuration is determined as an executable configuration. The executable configuration is used to characterize the candidate execution method that meets the execution requirements of the currently scheduled task and serves as the data basis for the subsequent target execution configuration determination process.

[0041] In this process, the task constraint information, preset execution constraints and verification parameters are determined according to the specific application scenario. This implementation method avoids task allocation based solely on node computing power or model accuracy by performing multi-dimensional constraint screening on candidate execution configurations. This ensures that the determined executable configurations simultaneously meet task requirements and node resource conditions, thereby improving the reliability of AI task scheduling in a cloud-edge-device collaborative environment.

[0042] Preferably, after completing model matching verification, resource capacity verification, and execution condition verification, the verification results corresponding to each candidate execution configuration are recorded, and configuration adaptation information is generated based on the constraint satisfaction status of different verification items. For example, for the same scheduled task, candidate execution configuration A satisfies the model inference requirements and resource capacity requirements, but the task completion time is close to the preset execution constraint boundary; candidate execution configuration B satisfies all verification conditions, and both task completion time and resource usage have high margins. The corresponding constraint satisfaction status is then recorded. Based on the configuration adaptation information, the candidate execution configurations that pass the verification are further filtered, prioritizing the retention of candidate execution configurations with higher constraint margins, and using these as the data basis for determining the subsequent target execution configuration.

[0043] Step S4: Determine the target execution configuration based on the executable configuration, and control the corresponding target node to call the corresponding lightweight model to execute the scheduled task, and update the node status information and lightweight model configuration based on the execution result.

[0044] Specifically, based on the task execution requirements and node status information corresponding to each executable configuration, a target execution configuration is determined; the target node corresponding to the target execution configuration is controlled to call the corresponding lightweight model to execute the scheduled task and obtain the task execution result; the node status information corresponding to the target node and the lightweight model configuration are updated based on the task execution result.

[0045] In this embodiment of the invention, each executable configuration is compared based on the task execution requirements and node status information corresponding to each executable configuration to determine the target execution configuration. The target execution configuration is used to characterize the final execution method corresponding to the currently scheduled task, including the target execution node and the lightweight model configuration invoked by the target execution node.

[0046] The task execution requirements characterize the execution requirements that the scheduled task must meet to complete processing, including task processing time requirements, model inference requirements, and resource consumption requirements. The node status information characterizes the current available resources of each execution node. Based on the task execution requirements and node status information, a target execution configuration is determined, ensuring that the execution nodes and lightweight model configuration corresponding to the target execution configuration meet the current task processing requirements.

[0047] For example, when there are two executable configurations, edge node A calling model configuration M1 and cloud node B calling model configuration M2, the task completion time for edge node A corresponding to model configuration M1 is 80ms and the resource usage ratio is 60%, while the task completion time for cloud node B corresponding to model configuration M2 is 150ms and the resource usage ratio is 40%. When the completion time limit of the current scheduled task is 100ms, the executable configuration corresponding to edge node A calling model configuration M1 is determined as the target execution configuration.

[0048] Furthermore, based on the target execution configuration, the lightweight model corresponding to the target execution node is invoked to execute the scheduled task, and the corresponding task execution result is obtained. The task execution result is used to characterize the processing output of the currently scheduled task, including information such as model inference results, task completion time, and task execution status.

[0049] Furthermore, the node status information corresponding to the target execution node and the lightweight model configuration are updated based on the task execution results. Specifically, the computing resource status and task execution status of the corresponding node are updated based on the resource consumption, task completion status, and model inference results during task execution; the lightweight model configuration is updated based on the inference effect and execution parameters during the actual execution of the lightweight model, providing a reference for generating candidate execution configurations for subsequent tasks to be scheduled.

[0050] The determination method of the target execution configuration, the specific content of the task execution result, and the update parameters of the node status information and lightweight model configuration are determined according to the actual application scenario. This implementation realizes the state update during the task scheduling process by feeding back the execution result to the node status information and lightweight model configuration, so that subsequent task scheduling can be generated based on the updated execution state, thereby improving the adaptability of the continuous scheduling process of lightweight AI tasks in a cloud-edge-device collaborative environment.

[0051] In one embodiment, such as Figure 2 As shown, this embodiment provides a lightweight AI task scheduling process for cloud-edge-device collaboration. In practical applications, the data to be processed is first acquired from the terminal side, and then preprocessed to obtain data that meets the input requirements of the lightweight AI model on the device side. The preprocessing process includes data format conversion, data size adjustment, and noise interference processing, so that the data to be processed can adapt to the data processing requirements of the lightweight AI model on the device side.

[0052] Furthermore, the preprocessed data to be processed is input into a lightweight AI model on the edge for preliminary inference to obtain the corresponding inference results and confidence levels. The lightweight AI model on the edge is an AI inference model pre-configured based on the computing resources of the terminal side, with preset model parameter scales and inference accuracy levels, used to quickly analyze the data to be processed on the terminal side. For example, in a device defect detection scenario, the terminal side collects image data of the device surface, identifies abnormal areas in the image using the lightweight AI model on the edge, and outputs the defect category and corresponding confidence level. When the confidence level of the output result reaches a preset confidence condition, such as a confidence level of 0.9 or higher, it is determined that the current inference result meets the edge processing requirements, and the edge processing result is directly output, ending the current task flow.

[0053] When the confidence level of the output result of the lightweight AI model on the edge does not reach the preset confidence condition, it means that the current processing result on the edge cannot meet the task requirements. At this time, a task to be scheduled is generated based on the data to be processed, the inference result and the corresponding task identifier, and enters the cloud-edge collaborative scheduling process.

[0054] For the generated tasks to be scheduled, the node status information corresponding to the cloud nodes and edge nodes is obtained. This node status information includes the node's computing resource status and task execution status, representing the current resource availability and task load of different nodes. Simultaneously, based on the data type and requirements of the task to be scheduled, lightweight model configuration items matching the task are obtained. Each model configuration item corresponds to different model parameter scales and inference accuracy levels; for example, for the same image recognition task, multiple lightweight models with different computational loads and accuracy levels can be configured.

[0055] Furthermore, the acquired node status information is combined with the corresponding model configuration items to generate multiple candidate execution configurations. Each candidate execution configuration corresponds to a task execution method, including the target execution node and the lightweight model configuration to be invoked. For example, edge node A invokes a low-parameter-scale model configuration, edge node B invokes a medium-parameter-scale model configuration, and cloud node invokes a high-precision model configuration, thus forming different candidate execution configurations.

[0056] After generating candidate execution configurations, each candidate configuration is filtered based on the task constraint information corresponding to the task to be scheduled. First, the lightweight model configurations in the candidate execution configurations are matched and verified to determine whether the inference capability of the corresponding model meets the task requirements. Then, based on the target execution node status information, the node resource carrying capacity is verified to determine whether the current node can support the operation of the corresponding model. Finally, the execution conditions are verified based on the task completion time limit and resource usage requirements. The candidate execution configurations that pass all the verifications are determined as executable configurations.

[0057] Finally, based on the task execution requirements and node status information corresponding to each executable configuration, the target execution configuration is determined, and the target node is controlled to call the corresponding lightweight model to execute the scheduled task and obtain the task execution result. Based on the actual resource consumption, task completion status, and model execution effect during task execution, the node status information and lightweight model configuration corresponding to the target node are updated. This allows subsequent task scheduling processes to generate and filter candidate execution configurations based on the updated status information, thereby improving the adaptability and execution reliability of the lightweight AI task scheduling process in a cloud-edge-device environment.

[0058] Reference Figure 3 , Figure 3 This is a schematic diagram of the structure of the cloud-edge-device collaborative lightweight AI intelligent scheduling system according to an embodiment of the present invention.

[0059] like Figure 3 As shown, the cloud-edge-device collaborative lightweight AI intelligent scheduling system proposed in this embodiment of the invention includes: the system is used to execute the above-described cloud-edge-device collaborative lightweight AI intelligent scheduling method, and the system includes: The acquisition module is used to acquire data to be processed on the terminal side and perform preliminary inference based on the lightweight AI model on the terminal side. When the preliminary inference results do not meet the preset conditions, a task to be scheduled is generated. The candidate module is used to obtain node status information of cloud nodes and edge nodes, and generate candidate execution configurations based on the node status information and the lightweight model configuration corresponding to the task to be scheduled. The configuration module is used to filter the candidate execution configurations based on the task constraint information corresponding to the task to be scheduled, and determine the executable configuration; The execution module is used to determine the target execution configuration based on the executable configuration, control the corresponding target node to call the corresponding lightweight model to execute the scheduled task, and update the node status information and lightweight model configuration based on the execution result.

[0060] It is understood that in the description of this specification, references to terms such as "one embodiment," "another embodiment," "other embodiments," or "first embodiment to Nth embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0061] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0062] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A lightweight AI intelligent scheduling method for cloud-edge-device collaboration, characterized in that, The method includes the following steps: Step S1: Obtain the data to be processed on the terminal side and perform preliminary reasoning based on the lightweight AI model on the terminal side. If the preliminary reasoning result does not meet the preset conditions, generate a task to be scheduled. Step S2: Obtain the node status information of cloud nodes and edge nodes, and generate candidate execution configurations based on the node status information and the lightweight model configuration corresponding to the task to be scheduled; Step S3: Based on the task constraint information corresponding to the task to be scheduled, filter the candidate execution configurations to determine the executable configuration; Step S4: Determine the target execution configuration based on the executable configuration, and control the corresponding target node to call the corresponding lightweight model to execute the scheduled task, and update the node status information and lightweight model configuration based on the execution result.

2. The lightweight AI intelligent scheduling method for cloud-edge-device collaboration as described in claim 1, characterized in that, Acquire the data to be processed on the terminal side and perform preliminary inference based on the lightweight AI model on the terminal side. If the preliminary inference result does not meet the preset conditions, generate a task to be scheduled, including: The system acquires the data to be processed collected from the terminal side, preprocesses the data, inputs the preprocessed data into the lightweight AI model on the terminal side for inference, and obtains the corresponding inference results and confidence levels. Determine whether the confidence level meets the preset confidence conditions; When the confidence level meets the preset confidence conditions, the inference result is output as the end-side processing result; When the confidence level does not meet the preset confidence conditions, a task to be scheduled is generated based on the data to be processed, the inference result, and the corresponding task identifier, and the task to be scheduled is sent to the scheduling node.

3. The lightweight AI intelligent scheduling method for cloud-edge-device collaboration as described in claim 2, characterized in that, The preprocessed data is input into a lightweight AI model on the input side for inference, obtaining the corresponding inference results and confidence levels, including: The preprocessed data to be processed is input into a lightweight AI model on the terminal side for feature extraction to obtain the corresponding target feature information; wherein, the lightweight AI model is an AI inference model pre-configured according to the computing resources on the terminal side, the lightweight AI model corresponds to a preset model parameter scale and inference accuracy level, and is used to perform feature extraction and preliminary inference on the preprocessed data to be processed, and output the corresponding inference results and confidence level; Reasoning is performed based on the target feature information to obtain the reasoning result corresponding to the data to be processed, and the corresponding confidence level is determined based on the reasoning result.

4. The lightweight AI intelligent scheduling method for cloud-edge-device collaboration as described in claim 1, characterized in that, Obtain node status information from cloud nodes and edge nodes, and generate candidate execution configurations based on the node status information and the lightweight model configuration corresponding to the task to be scheduled, including: Obtain the node status information corresponding to the cloud node and the edge node, the node status information including the node computing resource status and task execution status; Based on the lightweight model configuration corresponding to the task to be scheduled, determine multiple model configuration items that match the task to be scheduled. The node status information is combined with the corresponding model configuration items to generate a corresponding candidate execution configuration, which includes the execution node and the corresponding lightweight model configuration.

5. The lightweight AI intelligent scheduling method for cloud-edge-device collaboration as described in claim 4, characterized in that, Based on the lightweight model configuration corresponding to the task to be scheduled, determine multiple model configuration items that match the task to be scheduled, including: Based on the data type and task requirements of the task to be scheduled, obtain the set of model configurations corresponding to the task to be scheduled in the lightweight model configuration; Multiple model configuration items are determined from the model configuration set, and each model configuration item includes at least a model identifier, a model parameter scale, and a corresponding inference accuracy level.

6. The lightweight AI intelligent scheduling method for cloud-edge-device collaboration as described in claim 1, characterized in that, Based on the task constraint information corresponding to the task to be scheduled, the candidate execution configurations are filtered to determine the executable configuration, including: Based on the task constraint information, the lightweight model configurations corresponding to each candidate execution configuration are matched and verified to determine the candidate execution configurations that meet the inference requirements of the task to be scheduled. Based on the node status information of the target execution node corresponding to the candidate execution configuration, the resource carrying capacity of the target execution node is verified to determine the candidate execution configuration that meets the task execution resource requirements; Based on the task execution conditions corresponding to the candidate execution configuration, the candidate execution configuration is validated for execution conditions, and the candidate execution configuration that passes all validations is determined as an executable configuration.

7. The lightweight AI intelligent scheduling method for cloud-edge-device collaboration as described in claim 6, characterized in that, The process of validating the execution conditions of the candidate execution configuration includes: Based on the target execution node status information and the task constraint information corresponding to the candidate execution configuration, determine whether the task completion time limit and resource usage requirements corresponding to the candidate execution configuration meet the preset execution constraints. When the task completion time limit and resource usage requirements corresponding to the candidate execution configuration meet the preset execution constraints, the candidate execution configuration is determined to have passed the execution condition verification.

8. The lightweight AI intelligent scheduling method for cloud-edge-device collaboration as described in claim 1, characterized in that, The target execution configuration is determined based on the executable configuration, and the corresponding target node is controlled to call the corresponding lightweight model to execute the scheduled task. The node status information and lightweight model configuration are updated based on the execution result, including: Based on the task execution requirements and node status information corresponding to each executable configuration, the target execution configuration is determined; The target node corresponding to the target execution configuration is controlled to call the corresponding lightweight model to execute the scheduled task and obtain the task execution result. Update the node status information corresponding to the target node and the lightweight model configuration based on the task execution result.

9. A lightweight AI intelligent scheduling system for cloud-edge-device collaboration, characterized in that, The system is used to execute the cloud-edge-device collaborative lightweight AI intelligent scheduling method according to any one of claims 1-8, and the system includes: The acquisition module is used to acquire data to be processed on the terminal side and perform preliminary inference based on the lightweight AI model on the terminal side. When the preliminary inference results do not meet the preset conditions, a task to be scheduled is generated. The candidate module is used to obtain node status information of cloud nodes and edge nodes, and generate candidate execution configurations based on the node status information and the lightweight model configuration corresponding to the task to be scheduled. The configuration module is used to filter the candidate execution configurations based on the task constraint information corresponding to the task to be scheduled, and determine the executable configuration; The execution module is used to determine the target execution configuration based on the executable configuration, control the corresponding target node to call the corresponding lightweight model to execute the scheduled task, and update the node status information and lightweight model configuration based on the execution result.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the cloud-edge-device collaborative lightweight AI intelligent scheduling method as described in any one of claims 1-8.

Images