Cloud edge-end collaborative control method and system based on remote communication

Abstract

Embodiments of the present application provide a cloud edge end collaborative control method and system based on remote communication, the method comprising: generating task scheduling input data set according to real-time running state data generated by terminal equipment in the remote communication process, current computing resource information of edge computing node and global optimization target parameters issued by cloud end management system; calling a preset task scheduling decision model to structurally split the cloud edge end collaborative control task, obtaining original task structure information; combining multiple information to label priority for subtasks to generate task structure labeling information and perform cyclic matching deduction processing to obtain hierarchical resource allocation results; distributing the subtasks to corresponding cloud edge end nodes (including a first node corresponding to the terminal equipment, a second node corresponding to the edge computing node and a third node corresponding to the cloud end management system) for execution, and configuring a task execution process adjustment mechanism based on real-time resource changes to adapt to resource fluctuations.

Description

Technical Field

[0001] This invention relates to the field of cloud-edge-device data processing technology, and more specifically, to a cloud-edge-device collaborative control method and system based on remote communication. Background Technology

[0002] With the rapid development of IoT, 5G communication, and AI technologies, cloud-edge-device collaborative control has become a core architecture supporting complex application scenarios such as intelligent transportation, smart security, and the industrial internet. In this collaborative system, the cloud management system is responsible for global decision-making and resource coordination, edge computing nodes handle localized data processing and low-latency services, and terminal devices act as the data acquisition and execution endpoints. These three components interact and collaborate on tasks through remote communication. Currently, the key to cloud-edge-device collaborative control lies in how to rationally split, efficiently allocate, and dynamically schedule tasks based on the real-time status of terminal devices, the resource availability of edge nodes, and the global objectives of the cloud. However, the main problem with existing technologies is that the overall efficiency and stability of cloud-edge-device collaborative control are insufficient to meet the demands of complex scenarios. Summary of the Invention

[0003] This invention provides a cloud-edge-device collaborative control method based on remote communication, applied to a cloud-edge-device collaborative control system. The method includes: generating a task scheduling input dataset based on real-time operating status data generated by the terminal device during remote communication, current computing resource information of the edge computing node, and global optimization target parameters issued by the cloud management system; based on the task scheduling input dataset, calling a preset task scheduling decision model to perform structured decomposition of the cloud-edge-device collaborative control task to obtain original task structure information including sub-task types, data dependencies, and resource requirements; and combining the terminal device's operating status priority ranking results, edge computing node resource load rate analysis data, and cloud global optimization target weight configuration. The system performs a collaborative control priority labeling operation on each subtask in the original task structure information to generate task structure labeling information carrying priority identifiers. Based on the priority identifiers of the subtasks, resource requirements, and resource availability status of cloud-edge-device nodes, the system performs cyclic matching and deduction processing, including multi-round resource allocation simulation, on the task structure labeling information to obtain hierarchical resource allocation results. Based on the hierarchical resource allocation results, each subtask is distributed to the corresponding cloud-edge-device nodes for execution, and a task execution process adjustment mechanism based on real-time resource changes is configured to adapt to resource fluctuations. The cloud-edge-device nodes include at least a first node corresponding to the terminal device, a second node corresponding to the edge computing node, and a third node corresponding to the cloud management system.

[0004] This invention also provides a cloud-edge-device collaborative control system, including: a memory and a processor; the memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the aforementioned cloud-edge-device collaborative control method based on remote communication.

[0005] This invention also provides a computer storage medium storing a computer program / instruction, which, when executed by a processor, implements the aforementioned cloud-edge-device collaborative control method based on remote communication. Attached Figure Description

[0006] Figure 1 This is a flowchart illustrating the steps of a cloud-edge-device collaborative control method based on remote communication, as provided in an embodiment of the present invention.

[0007] Figure 2 This is a schematic diagram of the functional modules of a cloud-edge-device collaborative control system provided in an embodiment of the present invention.

[0008] Figure 3 This is a structural block diagram of a cloud-edge-device collaborative control system provided in an embodiment of the present invention. Detailed Implementation

[0009] See Figure 1 , Figure 1 The flowchart of a cloud-edge-device collaborative control method based on remote communication provided by an embodiment of the present invention is shown, specifically including the following steps 110-140.

[0010] Step 110: Generate a task scheduling input dataset based on the real-time operating status data generated by the terminal device during remote communication, the current computing resource information of the edge computing node, and the global optimization target parameters issued by the cloud management system.

[0011] In this embodiment of the invention, under a cloud-edge-device collaborative control scenario, the terminal device is a smart surveillance camera, the edge computing node is an edge server cluster within the region, and the cloud management system is a smart park-level smart security management platform. During the remote transmission of video streams, the smart surveillance camera generates real-time operational status data, including CPU utilization, memory usage, and network upload speed. The edge server cluster reports its current computing resource information in real time, such as the number of remaining CPU cores, available memory capacity, and remaining storage space. The cloud management system issues global optimization target parameters based on overall security needs. These parameters include optimization indicators such as task execution efficiency, resource utilization, and communication latency. This data is integrated according to a preset data format. For example, the real-time operational status data of the smart surveillance camera is converted into a standardized state vector, the computing resource information of the edge computing node is organized into a resource matrix, and the cloud-based global optimization target parameters are set as a weight vector. Finally, these are combined to generate a task scheduling input dataset, which serves as the basic data source for task scheduling decisions.

[0012] Step 120: Based on the task scheduling input dataset, call the preset task scheduling decision model to perform structured decomposition of the cloud-edge-device collaborative control task to obtain the original task structure information containing sub-task types, data dependencies and resource requirements.

[0013] Step 121: Input the task scheduling input dataset into the task decomposition module of the task scheduling decision model. The task feature extraction unit of the task decomposition module divides the control task into functional modules, identifies the basic operation units that constitute the control task, and classifies the basic operation units into different sub-task types.

[0014] In this embodiment of the invention, after receiving the task scheduling input dataset, the task decomposition module of the task scheduling decision model performs in-depth analysis of the video stream data processing task through its task feature extraction unit. By breaking down the functional requirements of the task, basic operation units such as video decoding, target detection, feature extraction, abnormal behavior recognition, and result reporting are identified. Specifically, the video decoding unit converts compressed video stream data into raw image frame data; the target detection unit analyzes the raw image frames to identify target objects such as pedestrians and vehicles in the scene (e.g., a smart factory park); the feature extraction unit extracts key feature information, such as contours, colors, and motion trajectories, from the detected target objects; the abnormal behavior recognition unit determines whether the target objects exhibit abnormal behavior, such as running or gathering, based on the extracted feature information; and the result reporting unit uploads the abnormal behavior recognition results and related data to the cloud management system. These basic operation units are categorized into different sub-task types according to their functional attributes, such as video decoding sub-tasks, target detection sub-tasks, feature extraction sub-tasks, abnormal behavior recognition sub-tasks, and result reporting sub-tasks.

[0015] Step 122: The dependency analysis unit of the task scheduling decision model tracks the data flow of the basic operation units corresponding to different sub-task types, records the input and output data association between sub-tasks, and generates a data dependency network represented in the form of a directed graph.

[0016] In this embodiment of the invention, the dependency analysis unit tracks the data flow of the basic operation units corresponding to the categorized subtask types. The output data of the video decoding subtask is the original image frame, which serves as the input data for the object detection subtask. The target object information output by the object detection subtask is the input for the feature extraction subtask. The target feature data obtained by the feature extraction subtask is input to the abnormal behavior recognition subtask. The recognition result of the abnormal behavior recognition subtask is the input data for the result reporting subtask. By recording the relationship between the input and output data of these subtasks, a data dependency network is constructed in the form of a directed graph, where each node represents a subtask type, and directed edges represent the data flow between subtasks. For example, a directed edge from the video decoding subtask node to the object detection subtask node indicates that the output data of the video decoding subtask flows to the object detection subtask.

[0017] Step 123: Invoke the resource requirement estimation unit of the task scheduling decision model, and calculate the amount of computing resources, storage resources and data transmission resources required for each subtask during execution based on the historical execution records of the subtask type and the current transmission bandwidth status of the remote communication network, to form a resource requirement descriptor.

[0018] In this embodiment of the invention, the resource requirement estimation unit first queries the historical execution records of each subtask type in the video stream data processing task, such as the average CPU usage time and memory usage of the video decoding subtask when processing a video stream at a set resolution. Simultaneously, it obtains the current transmission bandwidth status of the remote communication network, such as the network bandwidth between the intelligent surveillance camera and the edge computing node, and the network bandwidth between the edge computing node and the cloud management system. Combining the historical execution records and the current network status, it calculates the resource requirements for each subtask. For the video decoding subtask, based on the encoding format and resolution of the video stream, it calculates the required computational resources, such as the number of CPU processing cycles; storage resources, such as the storage space of the decoded image frames; and data transmission resources, such as the bandwidth requirement for the video stream transmission from the camera to the edge node. For the target detection subtask, it estimates its computational resources and memory usage based on the complexity of the detection algorithm and the size of the image frames. These calculated resource quantities are integrated into a resource requirement descriptor, with each subtask type corresponding to a unique resource requirement descriptor, clearly indicating the various resource requirements during the execution of that subtask.

[0019] Step 124: Integrate the data dependency network and the resource requirement descriptor, convert the data dependency network into a linear execution sequence according to the execution order constraints of the subtasks, and associate and bind the resource requirement descriptor with the corresponding subtask type to generate the original task structure information containing the subtask type field, the directed graph of data dependency, and the resource requirement descriptor.

[0020] In this embodiment of the invention, the data dependency network represents the execution order of each subtask in the form of a directed graph. For example, the video decoding subtask must be completed before the object detection subtask can be performed. Following this execution order constraint, the directed graph is transformed into a linear execution sequence: "video decoding - object detection - feature extraction - abnormal behavior recognition - result reporting". Simultaneously, a resource requirement descriptor corresponding to each subtask type is associated and bound to it. For example, the video decoding subtask type is bound to its corresponding descriptors for computational resources, storage resources, and data transmission resources. In the final generated original task structure information, the subtask type field clearly lists the type name of each subtask, the directed graph of data dependencies visually displays the dependencies between subtasks, and the resource requirement descriptors detail the resource requirements of each subtask.

[0021] Step 130: Combining the priority ranking results of the terminal device's operating status, the resource load rate analysis data of the edge computing node, and the global optimization target weight configuration information in the cloud, perform a collaborative control priority labeling operation on each subtask in the original task structure information to generate task structure labeling information carrying priority identifiers.

[0022] Step 131: Obtain the operating status parameters of the terminal devices and perform normalization processing. Calculate the operating health index of each terminal device through the status evaluation function, and obtain the priority ranking result of the operating status of the terminal devices by sorting them from high to low according to the health index.

[0023] In this embodiment of the invention, the intelligent surveillance camera serves as a terminal device, and its operating status parameters include CPU utilization, memory usage, network upload speed, and device temperature. Since these parameters have different dimensions and value ranges—for example, CPU utilization ranges from 0-100%, memory usage also ranges from 0-100%, while network upload speed is measured in Mbps and device temperature in degrees Celsius—they need to be normalized to map all parameter values ​​to the 0-1 range. Then, a status evaluation function is constructed. This function takes the normalized operating status parameters as input and uses the analytic hierarchy process (AHP) to determine the influence weight of each parameter. For example, the weight of CPU utilization is 0.3, memory usage is 0.2, network upload speed is 0.3, and device temperature is 0.2. The normalized parameter values ​​are multiplied by their corresponding weights and then summed to obtain the operating health index of each intelligent surveillance camera. For example, if a camera's normalized parameter values ​​are 0.6 (CPU utilization), 0.5 (memory utilization), 0.8 (network upload speed), and 0.4 (device temperature), then its operational health index is 0.6×0.3+0.5×0.2+0.8×0.3+0.4×0.2. Finally, all smart surveillance cameras are sorted from highest to lowest based on their health index to obtain the priority ranking of the terminal devices' operational status. Cameras with higher health indices have higher operational priority, and their related subtasks will be prioritized for allocation to appropriate nodes during task assignment.

[0024] Step 1311: Collect the operating status parameters of the terminal device during the remote communication process. The operating status parameters reflect the processor operating status, memory usage status and network connection status of the device.

[0025] In this embodiment of the invention, during remote video stream transmission communication with edge computing nodes, the intelligent surveillance camera collects operational status parameters in real time through its built-in sensors and monitoring programs. Parameters reflecting processor operation include CPU utilization and CPU core temperature; parameters reflecting memory usage include memory utilization and cache hit rate; and parameters reflecting network connectivity include network upload speed, network latency, and packet loss rate. These parameters comprehensively reflect the real-time operational status of the intelligent surveillance camera during remote communication, providing data support for calculating operational health indicators. For example, CPU utilization is collected every 10 seconds, and the average value is used as the CPU utilization parameter value for that time period; the network upload speed is calculated by statistically analyzing the size of data packets transmitted per unit time.

[0026] Step 1312: Perform dimensional unification processing on the collected operating status parameters to map the value ranges of different types of parameters to the same interval.

[0027] In this embodiment of the invention, the collected operating status parameters of the intelligent monitoring camera have different dimensions and value ranges. For example, the CPU utilization rate ranges from 0 to 100%, the network upload rate may be between 0 and Mbps, and the device temperature is between - degrees Celsius and 1 degree Celsius. To eliminate the influence of the difference in dimensions on the calculation, the min-max normalization method is used to process these parameters. For proportional parameters such as CPU utilization rate and memory utilization rate, their values ​​are directly divided by 100 to convert them into values ​​between 0 and 1. For the network upload rate, the normalized value is calculated based on its historical maximum and minimum upload rates using the formula (current value - minimum value) / (maximum value - minimum value). For the device temperature, a reasonable temperature range is set, such as 0-80 degrees Celsius. Temperature values ​​outside this range are treated as 0 or 1 respectively, and then normalization calculation is performed. Through this processing, the value range of all operating status parameters is uniformly mapped to the 0-1 interval, ensuring that each parameter has the same weight and influence in the status evaluation.

[0028] Step 1313: Construct a state evaluation function. The state evaluation function takes the normalized operating state parameters as input, determines the influence weight of each operating state parameter through the analytic hierarchy process, and sums the parameter values ​​with their corresponding weights to obtain the operating health index.

[0029] In this embodiment of the invention, the constructed state evaluation function is a multi-input single-output function model. First, the influence weights of each normalized operating state parameter are determined using the analytic hierarchy process (AHP). Domain experts are invited to compare the importance of each parameter pairwise, constructing a judgment matrix. Then, the eigenvectors of the judgment matrix are calculated, and after a consistency check, the weight values ​​of each parameter are obtained. For example, the weight of CPU utilization is determined to be 0.25, memory utilization to be 0.2, network upload speed to be 0.3, device temperature to be 0.15, and network latency to be 0.1. The normalized parameter values ​​are multiplied by their corresponding weights, and then all products are summed to obtain the operating health index. For example, if the normalized parameter values ​​of a certain smart surveillance camera are CPU utilization 0.6, memory utilization 0.5, network upload speed 0.7, device temperature 0.3, and network latency 0.4, then its operational health index is 0.6×0.25+0.5×0.2+0.7×0.3+0.3×0.15+0.4×0.1. The larger this operational health index value, the healthier the smart surveillance camera's operating status.

[0030] Step 1314: Compare the operational health indicators of all terminal devices, sort the terminal devices in descending order of indicator values, and generate a priority ranking result of the terminal devices' operational status. Terminal devices ranked higher have higher task execution priority.

[0031] In this embodiment of the invention, for example, there are 5 smart surveillance cameras. Their operational health indicators, calculated using the aforementioned state evaluation function, are 0.85, 0.72, 0.91, 0.68, and 0.79, respectively. Comparing these indicator values, they are sorted from highest to lowest as 0.91, 0.85, 0.79, 0.72, and 0.68, corresponding to smart surveillance camera numbers 3, 1, 5, 2, and 4. Therefore, the generated terminal device operational status priority ranking result is: Camera 3 > Camera 1 > Camera 5 > Camera 2 > Camera 4. During task allocation, subtasks related to high-priority cameras, such as the video stream data processing subtask acquired by that camera, will be prioritized for resource matching and allocation to ensure timely processing.

[0032] Step 132: Collect the resource usage of edge computing nodes at the current moment, and combine the load rate analysis to determine the real-time occupancy ratio of computing resources, storage resources and network resources, and generate resource load rate analysis data for edge computing nodes.

[0033] In this embodiment of the invention, the edge computing node is an edge server cluster within a region, comprising multiple edge servers. A resource monitoring agent deployed on each edge server collects real-time data on its resource usage, such as the number of CPU cores used, memory usage, used storage space, and network bandwidth. Based on this data, the real-time utilization rate of various resources is calculated: CPU utilization is calculated by dividing the number of used cores by the total number of cores; memory utilization is calculated by dividing the used memory by the total memory capacity; storage utilization is calculated by dividing the used storage space by the total storage space; and network utilization is determined by the ratio of current network bandwidth usage to total bandwidth. This real-time utilization rate data, along with basic information about the edge servers, such as server ID and IP address, is integrated to generate resource load analysis data for the edge computing node. This data clearly reflects the current resource load status of each server in the edge server cluster, providing a basis for task allocation and resource scheduling. For example, if an edge server has a total of 16 CPU cores and is currently using 8 cores, the CPU utilization rate is 50%; if the total memory is 64GB and 32GB is used, the memory utilization rate is 50%, etc.

[0034] Step 1321: Collect multi-dimensional resource usage data of edge computing nodes at the current moment. The resource usage data includes the task queue length and process blocking status of computing resources, the file handle activity and cache hit rate trend of storage resources, and the connection session establishment frequency and data packet retransmission events of network resources.

[0035] In this embodiment of the invention, multi-dimensional resource usage data of edge computing nodes are collected through specialized resource monitoring tools. For computing resources, the task queue length is obtained by monitoring the operating system's process scheduling queue, recording the number of tasks currently waiting to be executed; process blocking status is determined by analyzing process status information, statistically analyzing the number of processes in a blocked state and the reasons for blocking, such as waiting for I / O or waiting for locks. Regarding storage resources, file handle activity is measured by statistically analyzing the number of file handles opened and closed per unit time; cache hit rate trends are calculated by continuously collecting cache hit counts and total access counts, analyzing their changing trends, such as calculating the cache hit rate every 5 minutes to form time-series data. The connection session establishment frequency of network resources is obtained by statistically analyzing the number of newly established network connections per unit time; packet retransmission events are obtained by monitoring the retransmission packet counter of the network interface, recording the number of retransmitted packets and the corresponding application. This multi-dimensional resource usage data comprehensively reflects the resource operation status of the edge computing nodes.

[0036] Step 1322: Extract time-series features from the collected multi-dimensional resource usage data to capture the usage fluctuation patterns of each resource dimension at different time granularities, and identify sudden peak patterns, periodic fluctuation patterns, and trend growth patterns of resource usage.

[0037] In this embodiment of the invention, time-series features are extracted from the collected multi-dimensional resource usage data of edge computing nodes according to different time granularities, such as 1 minute, 5 minutes, and 15 minutes. A sliding window method is used to segment the time-series data, calculating statistical characteristics such as mean, variance, maximum, and minimum values ​​within each window to capture fluctuations in resource usage. For sudden peak patterns, a sudden peak is identified when the usage of a certain resource dimension increases sharply within a short period, exceeding the normal fluctuation range by several times the standard deviation. For periodic fluctuation patterns, Fourier transform is used to perform spectral analysis on the time-series data to identify the main periodic components, such as a daily cycle pattern where usage peaks occur once per hour. For trend-growth patterns, linear regression analysis is used to analyze the changing trend of the time-series data; when the regression coefficient is positive and passes the significance test, a trend-growth is identified. Through such time-series feature extraction and pattern recognition, a deeper understanding of the usage patterns of various resource dimensions of edge computing nodes can be achieved.

[0038] Step 1323: Based on the preset resource status determination rules, and combined with the identified fluctuation patterns, classify the status of each resource dimension, mark computing resources as idle, stable or strained, storage resources as sufficient, saturated or overflowing, and network resources as unobstructed, congested or interrupted.

[0039] In this embodiment of the invention, the preset resource status judgment rules are formulated based on the resource configuration and historical operating experience of edge computing nodes. For computing resources, when the CPU utilization rate is below 30% and the task queue length is less than 5, it is marked as idle; when the CPU utilization rate is between 30% and 70%, the task queue length is between 5 and 20, and there is no sudden peak pattern, it is marked as stable; when the CPU utilization rate is above 70%, or the task queue length is greater than 20, or a sudden peak pattern occurs, it is marked as strained. Regarding storage resources, when the storage space utilization rate is below 50%, file handle activity is normal, and the cache hit rate trend is stable, it is marked as sufficient; when the storage space utilization rate is between 50% and 80%, file handle activity is high, or the cache hit rate trend is declining, it is marked as saturated; when the storage space utilization rate is above 80%, or a file handle overflow occurs, it is marked as overflowed. In network resources, when the network bandwidth utilization rate is below 40%, the connection session establishment frequency is stable, and there are no data packet retransmission events, it is marked as unobstructed; when the network bandwidth utilization rate is between 40% and 80%, the connection session establishment frequency fluctuates greatly, or there are a few data packet retransmission events, it is marked as congested; when the network bandwidth utilization rate is above 80%, or there are a large number of data packet retransmission events, or the connection session establishment failure rate is high, it is marked as interrupted.

[0040] Step 1324: Perform association rule mining on the classification status of different resource dimensions, identify the co-occurrence pattern of computing resource shortage and network resource congestion, the impact weight of storage resource saturation on computing resource processing latency, and establish a collaborative impact relationship model of resource load.

[0041] In this embodiment of the invention, an association rule mining algorithm is used to analyze the classification status data of edge computing nodes across different resource dimensions. The classification statuses of computing resources, storage resources, and network resources are treated as item sets. By statistically analyzing the frequent item sets of different state combinations, association rules are mined. For example, historical data analysis reveals that when computing resources are under strain and network resources are congested, this state combination occurs frequently, forming a co-occurrence pattern with support and confidence reaching preset thresholds. Regarding the impact weight of storage resource saturation on computing resource processing latency, sample data of computing resource processing latency under saturated storage conditions and processing latency data under unsaturated storage conditions are collected. A regression analysis method is used to establish a relationship model between the two, calculating the impact weight value of storage resource saturation on computing resource processing latency. For example, saturated storage resources will increase computing resource processing latency by a certain percentage. Based on these association rules and impact weights, a collaborative impact model of resource load is established, which can describe the mutual influence relationships between different resource dimension states.

[0042] Step 1325: Combine the classification status, fluctuation pattern description and collaborative influence relationship model of each resource dimension to generate resource load rate analysis data containing resource status labels, feature pattern vectors and association rule sets.

[0043] In this embodiment of the invention, the classification status of each resource dimension of the edge computing node, such as computing resource strain, storage resource saturation, and network resource congestion, is used as resource status labels. Identified fluctuation patterns, such as sudden peak patterns, periodic fluctuation patterns, and trend growth patterns, are converted into feature pattern vectors through feature encoding. Each fluctuation pattern corresponds to a dimension in the vector, and the dimension value represents the intensity or frequency of occurrence of the pattern. The association rule set includes rules for the co-occurrence of computing resource strain and network resource congestion, as well as rules on the impact of storage resource saturation on computing resource processing latency. Integrating these data generates resource load rate analysis data, which not only includes the current status labels of each resource dimension but also its fluctuation characteristics and rules on the collaborative influence between resources.

[0044] Step 133: Receive the global optimization target weight configuration information issued by the cloud management system. The global optimization target weight configuration information includes task execution efficiency weight, resource utilization weight, and communication latency weight.

[0045] In this embodiment of the invention, the cloud management system dynamically distributes global optimization target weight configuration information based on the overall needs and current operational status of smart park-level smart security. In the intelligent surveillance camera video stream data processing task, the task execution efficiency weight reflects the importance placed on video processing speed, the resource utilization weight reflects the requirement for efficient resource utilization of edge computing nodes, and the communication latency weight focuses on the timeliness of video stream data transmission between the cloud and the edge. For example, during periods of high incidence of security incidents, the cloud management system may increase the task execution efficiency weight (e.g., setting it to 0.5), decrease the resource utilization weight to 0.3, and the communication latency weight to 0.2 to ensure rapid processing of video stream data and timely detection of anomalies. During periods of low incidence, the resource utilization weight may be increased (e.g., 0.4), while the task execution efficiency weight may be appropriately decreased to 0.3, and the communication latency weight to 0.3, to save resource costs. The edge computing nodes receive and parse this global optimization target weight configuration information through periodic communication with the cloud management system, storing it in a local configuration file for use in task priority calculation.

[0046] Step 134: Based on the priority ranking result of the running status, determine the basic value of the execution priority of the subtask on the terminal device, and correct the basic value by combining the resource load rate analysis data of the edge computing node. When the load rate of the edge node exceeds the preset load rate, reduce the priority correction coefficient of the corresponding subtask.

[0047] In this embodiment of the invention, the higher the ranking of the intelligent monitoring camera in the priority ranking result, the higher the base priority value of its related subtasks on the terminal device. For example, the base priority value of the subtask related to the first-ranked camera is set to 10, the second-ranked camera to 9, and so on. Then, based on the resource load rate analysis data of the edge computing nodes, when the load rate indicators such as CPU utilization and memory utilization of the edge nodes exceed the preset load rate, such as when the CPU utilization exceeds 70%, the priority correction coefficient of the corresponding subtask is reduced. For example, the initial value of the priority correction coefficient is 1.0. When the load rate of the edge node exceeds the preset value, the correction coefficient is reduced according to the proportion of the load rate exceeding the preset value, such as reducing the correction coefficient by 0.1 for every 10% increase in load rate. The base priority value of the subtask is multiplied by the correction coefficient to obtain the corrected priority value, realizing dynamic adjustment of priority to avoid assigning high-priority tasks to the edge node when the load is too high, which would cause task execution delays.

[0048] Step 135: Based on the weights in the global optimization target weight configuration information in the cloud, perform weighted fusion processing on the corrected priority. Set a time-sensitive factor according to the real-time requirements of the sub-task, a transmission cost factor according to the data transmission volume, and a computation time factor according to the computational complexity characteristics. Multiply these factors with the weighted fusion result to obtain the final priority value.

[0049] In this embodiment of the invention, the task execution efficiency weight in the cloud-based global optimization target weight configuration information is WE, the resource utilization weight is WR, and the communication latency weight is WL. The corrected priority value is multiplied by each of these three weights, and then summed to obtain the weighted fusion result, i.e., Weighted Fusion Result = Corrected Priority × WE + Corrected Priority × WR + Corrected Priority × WL (in practical applications, the calculation method for each weight can be combined with the task's performance on each optimization target). Simultaneously, a time-sensitive factor is set according to the real-time requirements of the subtask; the higher the real-time requirement, the larger the factor value, such as 0.8-1.2. A transmission cost factor is set according to the data transmission volume; the smaller the data transmission volume, the larger the factor value, such as 0.5-1.5. A computation time factor is set according to the computational complexity characteristics; the lower the computational complexity, the larger the factor value, such as 0.6-1.4. The weighted fusion result is multiplied by the time-sensitive factor, transmission cost factor, and computation time factor to obtain the final priority value, i.e., Final Priority Value = Weighted Fusion Result × Time-Sensitive Factor × Transmission Cost Factor × Computation Time Factor. For example, if the weighted fusion result of a certain subtask is 8, the time sensitivity factor is 1.1, the transmission cost factor is 1.2, and the computation time factor is 1.0, then the final priority value is 8×1.1×1.2×1.0.

[0050] Step 136: Convert the final priority value into a priority identifier, reorder the subtasks in the original task structure information in descending order of priority, and associate the priority identifier with the corresponding subtask type, data dependency and resource requirements to generate task structure annotation information carrying the priority identifier.

[0051] In this embodiment of the invention, the final priority value is converted into a priority identifier according to a preset grading standard. For example, a final priority value of 10 or above is assigned a priority identifier P0 (highest), 8-9 is P1, 6-7 is P2, 4-5 is P3, and below 4 is P4 (lowest). Then, the subtasks in the original task structure information are reordered according to the priority identifiers from high to low, so that the subtasks with higher priority are listed first. The priority identifier of each subtask is associated and stored with its subtask type, data dependency relationship directed graph, and resource requirement descriptor. For example, a record is made for each subtask in a database table, containing a subtask type field, a data dependency relationship field, a resource requirement field, and a priority identifier field. The resulting task structure annotation information carrying priority identifiers can clearly indicate the execution priority of each subtask.

[0052] Step 140: Based on the priority identifier, resource requirements, and resource availability status of cloud-edge-end nodes of the subtasks, perform cyclic matching and deduction processing on the task structure annotation information, including multi-round resource allocation simulation, to obtain hierarchical resource allocation results. Based on the hierarchical resource allocation results, distribute each subtask to the corresponding cloud-edge-end nodes for execution, and configure a task execution process adjustment mechanism based on real-time resource changes to adapt to resource fluctuations.

[0053] In this embodiment of the invention, the cloud-edge-device nodes include intelligent surveillance cameras (first node in the terminal layer), edge server clusters (second node in the edge layer), and a cloud management system (third node in the cloud layer). Based on the priority identifiers of each subtask in the task structure annotation information, higher-priority subtasks are processed first. Resource matching is performed based on the resource requirements of the subtasks, such as computing and storage resources, combined with the current resource availability status of the cloud-edge-device nodes, such as the remaining computing power of the terminal devices, the idle memory of the edge nodes, and the processing capacity of the cloud. Through multiple rounds of resource allocation simulation—that is, preliminary allocation followed by resource conflict detection and resolution, and multiple iterative matching and deduction—the final allocation scheme for each subtask in the cloud-edge-device nodes is determined, forming a hierarchical resource allocation result. Subtasks are distributed to the corresponding nodes for execution according to this result, and a real-time resource monitoring mechanism is established. When the resource availability status of a node changes, such as a sudden decrease or increase in resources, a task execution process adjustment mechanism is triggered to dynamically adjust the task allocation to adapt to resource fluctuations and ensure continuous and stable task execution. For example, when the resources of an edge node suddenly become scarce, its low-priority subtasks are migrated to other nodes with sufficient resources.

[0054] Step 141: Initialize the resource availability status table of cloud-edge-device nodes, and record the current available computing resources, storage resources, and network resources of each node in the terminal layer, edge layer, and cloud layer.

[0055] In this embodiment of the invention, the resource availability status table for cloud-edge nodes is a two-dimensional table, where rows represent different nodes and columns represent different resource types. During initialization, resource query requests are sent to each node to obtain the remaining computing resources (e.g., remaining CPU processing power), storage resources (e.g., remaining local storage space), and network resources (e.g., remaining upload bandwidth) of the intelligent surveillance cameras at the terminal layer; the available computing resources (e.g., remaining CPU cores, GPU computing power), storage resources (e.g., available hard disk space), and network resources (e.g., remaining network bandwidth) of each server in the edge server cluster at the edge layer; and the available computing resources (e.g., remaining number of virtual machine instances), storage resources (e.g., cloud storage capacity), and network resources (e.g., bandwidth for communication with edge nodes) of the cloud management system at the cloud layer. This obtained resource availability data is then filled into the corresponding cells of the resource availability status table to complete the initialization. For example, the available computing resources of terminal layer node 1 (a smart camera) are 0.3 (relative value), the available storage resources are 10GB, and the available network resources are 2Mbps; the available computing resources of edge layer node A (an edge server) are 8-core CPU, 100GB of storage resources, and 100Mbps of network resources, etc.

[0056] Step 142: Extract the list of subtasks carrying priority identifiers from the task structure annotation information, and select subtasks in order of priority from high to low for resource matching processing.

[0057] In this embodiment of the invention, the task structure annotation information is stored in a structured data format, such as XML or JSON, which includes fields such as subtask type, data dependency, resource requirements, and priority identifier. By parsing this structured data, all subtasks and their corresponding priority identifiers are extracted, forming a subtask list carrying priority identifiers. This list is then sorted from highest to lowest priority identifier; for example, the subtask with priority identifier P0 is at the top, followed by P1, P2, and so on. Following this sorting order, subtasks are selected sequentially for resource matching processing, meaning the highest priority subtask is processed first to ensure it receives the necessary resources to meet its real-time or importance requirements. For example, if the subtask with priority identifier P0 is the abnormal behavior recognition subtask, and priority P1 is the object detection subtask, then resource matching is performed first on the abnormal behavior recognition subtask.

[0058] Step 143: For the currently selected subtask, based on the computing resources, storage resources, and data transmission resources in the corresponding resource requirements, traverse the resource availability status table of cloud-edge nodes and filter out a set of candidate nodes whose resource availability meets the requirements of the subtask.

[0059] In this embodiment of the invention, the currently selected subtask is the abnormal behavior recognition subtask, which requires 2 CPU cores for computing resources, 10GB of storage space, and 5Mbps of network bandwidth for data transmission. The resource availability status table of the cloud-edge nodes is traversed to check whether the available computing resources of each node are greater than or equal to 2 CPU cores, 10GB or more of storage resources, and 5Mbps or more of network bandwidth. All nodes that meet these resource requirements are selected to form a candidate node set. For example, edge servers A and B in the edge layer, and a virtual machine instance in the cloud management system, all have sufficient available resources to meet the requirements of the abnormal behavior recognition subtask; these nodes together constitute the candidate node set.

[0060] Step 144: Calculate the matching degree between the subtask and each node in the candidate node set, and determine the node with the highest matching degree as the optimal allocation node.

[0061] In this embodiment of the invention, when calculating the matching degree between a subtask and a candidate node, multiple factors are comprehensively considered, including the node's remaining resources, the degree of matching between the node and the subtask's resource requirements, the node's historical task execution efficiency, and the communication latency between the subtask and the node. A certain weight is assigned to each factor, such as a weight of 0.3 for remaining resources, 0.4 for the degree of matching between resource requirements, 0.2 for historical execution efficiency, and 0.1 for communication latency. For remaining resources, the more of a node's remaining resources of a certain type, the higher the score for that item; the degree of matching between resource requirements is calculated by the ratio of the node's available resources to the subtask's resource requirements, with a higher score for a ratio closer to 1; historical execution efficiency is scored based on the average time and success rate of the node in executing similar subtasks in the past; and the lower the communication latency, the higher the score for that item. The matching degree between the subtask and the node is obtained by multiplying the scores of each factor by their corresponding weights and summing them. For example, if candidate node A has scores of 80 for remaining resources, 90 for resource demand matching, 85 for historical execution efficiency, and 75 for communication latency, with weights of 0.3, 0.4, 0.2, and 0.1 respectively, then the matching degree is 80×0.3 + 90×0.4 + 85×0.2 + 75×0.1. After calculating the matching degree of all candidate nodes, the node with the highest matching degree is selected as the optimal allocation node.

[0062] Step 145: Assign the subtask to the optimal allocation node, update the resource availability status table of the optimal allocation node, and record the subtask type, allocation node identifier, and resource usage of the current assignment.

[0063] In this embodiment of the invention, the abnormal behavior identification subtask is assigned to the calculated optimal allocation node, such as edge server A. Then, based on the resource requirements of this subtask, the resource availability status table of edge server A is updated by subtracting 2 CPU cores from its available computing resources, 10GB from its available storage resources, and 5Mbps from its available network resources. Simultaneously, a record is added to the task allocation record, containing information such as the subtask type (abnormal behavior identification), the allocation node identifier (IP address or device number of edge server A), resource usage (2 CPU cores, 10GB storage, 5Mbps network), and allocation time. Thus, the resource availability status table can reflect the real-time changes in resource usage of each node, and the task allocation record provides a basis for task tracking and management.

[0064] Step 146: Repeat the steps of selecting subtasks, filtering candidate nodes, calculating matching degree and allocating resources until all subtasks have completed the initial allocation and the first round of resource allocation results are formed.

[0065] In this embodiment of the invention, after the allocation of the abnormal behavior recognition subtask is completed, the next subtask, such as the target detection subtask, is selected according to priority. Then, the same resource matching process as the abnormal behavior recognition subtask is executed: based on the resource requirements of the target detection subtask, a set of candidate nodes that meet the requirements is selected from the updated resource availability status table; the matching degree between the subtask and each candidate node is calculated; the subtask is assigned to the node with the highest matching degree; the resource availability status table of that node is updated; and the allocation information is recorded. All subtasks are processed sequentially in this manner until the subtask list in the task structure annotation information is empty, completing the initial allocation. The first round of resource allocation results formed at this time contains the initial allocation node information of all subtasks, but resource conflicts may exist, such as multiple subtasks being assigned to the same node, resulting in insufficient resources for that node.

[0066] Step 147: Perform conflict detection on the first round of resource allocation results. When multiple subtasks compete for the same node resource, resulting in insufficient resources, resolve the conflict according to the subtask priority identifier. Subtasks whose priority does not meet the preset index will re-enter the candidate node screening process.

[0067] In this embodiment of the invention, conflict detection compares the total resource requirements of all subtasks allocated to each node in the first round of resource allocation with the total resource availability table initialized at that node. A resource conflict is determined when the total resource requirements exceed the total resource availability table. For example, edge server A is allocated 3 subtasks in the first round of allocation, with a total CPU resource requirement of 10 cores, while the server has a total of 8 CPU cores, resulting in a CPU resource conflict. During conflict resolution, the lowest priority subtask is selected for processing according to its priority identifier from low to high. A preset metric is priority identifier P2. When a subtask's priority identifier is lower than P2 (e.g., P3, P4), it is removed from the current allocation node and re-added to the candidate node selection process to find other nodes with sufficient resources. During re-selection, the updated resource availability table is used to avoid reassigning to a node with resource conflicts.

[0068] Step 1471: Traverse the list of allocated subtasks for each cloud-edge node in the first round of resource allocation results, and calculate the total resource requirements of all subtasks allocated to each node.

[0069] In this embodiment of the invention, the results of the first round of resource allocation are stored in list form. Each node corresponds to a list of allocated subtasks, and each element in the list contains the resource requirements of the subtask. By traversing the allocated subtask list of each node, the computing resources, storage resources, and network resources of all subtasks in the list are summed to obtain the total resource requirements of each node. For example, if the allocated subtask list of edge server A contains subtask 1 (CPU requirement 2 cores, storage requirement 10GB, network requirement 5Mbps) and subtask 2 (CPU requirement 3 cores, storage requirement 15GB, network requirement 8Mbps), then its total resource requirements are 5 CPU cores, 25GB storage, and 13Mbps network.

[0070] Step 1472: Compare the total resource demand with the initial resource availability in the node's resource availability status table. If the total resource demand exceeds the initial resource availability, it is determined that the node has a resource allocation conflict.

[0071] In this embodiment of the invention, the initial total resources of each node are recorded during the initialization of the node's resource availability status table. For example, the initial total CPU resources of edge server A are 8 cores, the total storage resources are 100GB, and the total network resources are 100Mbps. The total resource demand (5 CPU cores, 25GB storage, 13Mbps network) calculated in step 1471 is compared with the initial total resources. At this point, it does not exceed the limit, and there is no conflict. If another edge server B has an initial total CPU resource of 4 cores and the total CPU demand of the allocated subtasks is 5 cores, then the total resource demand exceeds the initial available resources, and it is determined that there is a CPU resource allocation conflict for this node.

[0072] Step 1473: Extract the list of subtasks assigned to nodes with resource allocation conflicts, sort them from low to high according to the priority of the subtasks, and select the subtask with the lowest priority as the conflict resolution target.

[0073] In this embodiment of the invention, the allocated subtask list of the edge server B with resource conflicts includes subtask 3 (priority P2), subtask 4 (priority P3), and subtask 5 (priority P1). These subtasks are sorted from low to high priority as subtask 4 (P3), subtask 3 (P2), and subtask 5 (P1). Subtask 4, with the lowest priority, is selected as the conflict resolution target and is prepared to be migrated out of edge server B.

[0074] Step 1474: Remove the conflict resolution object from the list of assigned subtasks of the original node, restore the corresponding amount of resources occupied by the node, and update the node's resource availability status table.

[0075] In this embodiment of the invention, subtask 4 is removed from the allocated subtask list of edge server B. Based on the resource requirements of subtask 4, such as 2 CPU cores, 5GB storage, and 3Mbps network speed, these resources are added back to the resource availability table of edge server B. That is, the available CPU cores increase by 2, the available storage increases by 5GB, and the available network speed increases by 3Mbps. The updated resource availability table accurately reflects the resource availability of edge server B after removing subtask 4.

[0076] Step 1475: Add the conflict resolution object back to the list of subtasks to be assigned, and re-enter the candidate node filtering process according to priority order. Recalculate the matching degree and assign it to other candidate nodes with available resources.

[0077] In this embodiment of the invention, the removed subtask 4 is re-added to the list of subtasks to be assigned. Since its priority identifier is P3, it is ranked after other higher-priority subtasks in the assignment list according to priority order. When it is subtask 4's turn to be processed, a candidate node screening process is executed. At this time, the resources of edge server B, even after recovery, may still be insufficient (if there are other conflicts), but other edge servers or cloud nodes may have available resources. The matching degree between subtask 4 and these candidate nodes is recalculated, and it is assigned to the node with the highest matching degree and sufficient resources, such as edge server C, thus resolving the conflict.

[0078] Step 148: After multiple rounds of resource allocation simulation and conflict resolution, a stable correspondence between each subtask and cloud-edge-device nodes is generated. The stable correspondence is then organized into layers according to the terminal layer, edge layer, and cloud layer to obtain a hierarchical resource allocation result containing a list of subtasks allocated to each layer of nodes.

[0079] In this embodiment of the invention, the multi-round resource allocation simulation and conflict resolution is an iterative process. After each round of allocation, conflict detection and resolution are performed, and conflicting subtasks are redistributed until the total resource demand of all nodes does not exceed their total resource capacity, and each subtask finds a stable allocation node. At this point, the stable correspondence between each subtask and the cloud-edge-device nodes is a mapping table, recording the unique node identifier corresponding to each subtask. Then, these stable correspondences are hierarchically organized according to the terminal layer, edge layer, and cloud layer. Subtasks allocated to intelligent monitoring cameras at the terminal layer are grouped into one category, forming a terminal layer node allocation subtask list; similarly, allocation subtask lists for the edge layer and cloud layer are formed. These lists together constitute the hierarchical resource allocation result, clearly representing the subtasks undertaken by nodes at different levels.

[0080] Step 149: Based on the hierarchical resource allocation results, distribute the execution instructions and related data of each subtask to the corresponding cloud-edge-device nodes through a remote communication protocol, and simultaneously establish a resource monitoring mechanism to collect the resource status change data of each node in real time.

[0081] In this embodiment of the invention, the hierarchical resource allocation result clearly defines the target execution node for each subtask. Based on the node's network address, the execution instructions and related data for the subtask are sent to the corresponding cloud-edge node via remote communication protocols such as HTTP, MQTT, or RPC. The execution instructions include information such as the subtask type, execution parameters, and target output; the related data includes the input data required for the subtask execution, such as video stream segments and model parameters. For example, the execution instructions for the target detection subtask and the corresponding video frame data are sent to edge server A via the MQTT protocol. Simultaneously, a resource monitoring mechanism is established, deploying resource monitoring agents on each cloud-edge node. These agents collect data on changes in node resource status, such as CPU utilization, memory usage, and network bandwidth usage, at preset time intervals (e.g., 5 seconds), and report this data to the task scheduling center in real time via remote communication protocols to promptly detect resource fluctuations.

[0082] Step 1410: When the resource monitoring mechanism detects that the available node resources are lower than the preset available resources, the task execution process adjustment mechanism is triggered. The optimal allocation node for the subtask is recalculated based on the current resource status. The task migration instruction is generated and sent to the original allocation node and the new allocation node to realize the dynamic migration of the subtask between cloud, edge and terminal nodes to adapt to resource fluctuations.

[0083] In this embodiment of the invention, the resource monitoring mechanism collects real-time resource availability data for each node. When it detects that the CPU availability of a node (such as edge server D) is lower than the preset availability (e.g., 1 core remaining), a task execution flow adjustment mechanism is triggered. This mechanism first analyzes the subtasks currently being executed by the node and selects migrateable subtasks (e.g., subtasks with low execution progress and low priority) based on their execution progress and importance. Then, based on the current resource availability status of all cloud-edge-end nodes, it recalculates the optimal allocation node for the migrateable subtask, such as edge server E. A task migration instruction is generated, including the subtask ID, original node ID, new node ID, migration data list, etc., and sent to edge servers D and E respectively. After receiving the instruction, edge server D stops the subtask execution, packages the relevant data, and sends it to E; edge server E, after receiving the data, restores the subtask execution environment and continues execution. Through this dynamic migration, flexible scheduling of subtasks between cloud-edge-end nodes is achieved to adapt to resource fluctuations.

[0084] Step 1410-1: Collect resource availability data of each cloud-edge node at preset time intervals through the resource monitoring mechanism. When it is detected that any one of the available computing resources, available storage resources, or available network resources of the target node is lower than the corresponding preset value, a resource warning signal is generated.

[0085] In this embodiment of the invention, the preset time interval is 1 second. Every second, the resource monitoring mechanism sends a resource query request to each cloud-edge node, collecting its available computing resources (e.g., remaining CPU cores), available storage resources (e.g., available storage space), and available network resources (e.g., remaining bandwidth). Preset values ​​are set for each resource type; for example, the preset value for available computing resources is 2 CPU cores, the preset value for available storage resources is 20GB, and the preset value for available network resources is 10Mbps. When the available computing resources of a target node (e.g., edge server D) are 1 core (less than 2 cores), or the available storage resources are 15GB (less than 20GB), or the available network resources are 5Mbps (less than 10Mbps), the resource monitoring mechanism immediately generates a resource warning signal. This signal includes information such as node identifier, resource type, and current availability.

[0086] Step 1410-2: After receiving a resource warning signal using the task execution process adjustment mechanism, extract the list of subtasks currently being executed by the node, filter out the relocatable subtasks based on their execution progress, and mark the subtasks with execution progress below a preset percentage as relocatable objects.

[0087] In this embodiment of the invention, after receiving a resource warning signal from edge server D, the task execution process adjustment mechanism queries the task execution status database to obtain a list of subtasks currently being executed by that node, such as subtask 6 (execution progress 30%), subtask 7 (execution progress 70%), and subtask 8 (execution progress 50%). With a preset percentage of 50%, subtask 6 (30%), whose execution progress is below 50%, is marked as a migrateable object. Subtasks 7 and 8, whose execution progress is above or equal to 50%, are not marked as migrateable objects due to higher migration costs.

[0088] Step 1410-3: For each migrateable object, re-execute the resource matching process, traverse the resource availability status table of the current cloud-edge nodes, and filter out a new set of candidate nodes whose resource availability meets the remaining resource requirements of the subtask.

[0089] In this embodiment of the invention, the remaining resource requirements for the migrated object subtask 6 are 1.5 CPU cores, 8GB of storage, and 4Mbps of network. The resource availability status table of the current cloud-edge nodes is traversed to check whether the available computing resources of each node are greater than or equal to 1.5 cores, the available storage resources are greater than or equal to 8GB, and the available network resources are greater than or equal to 4Mbps. For example, the available computing resources of edge server E are 3 cores, the available storage resources are 50GB, and the available network resources are 50Mbps; the available computing resources of cloud nodes are 10 cores, the available storage resources are 1000GB, and the available network resources are 100Mbps, all of which meet the remaining resource requirements of subtask 6. Therefore, these nodes constitute a new candidate node set.

[0090] Step 1410-4: Combine the remaining resources of the node, the communication bandwidth with the original assigned node, and the network transmission cost of the subtask migration to calculate the migration matching degree between the migrateable object and each node in the new candidate node set, and determine the node with the highest migration matching degree as the new assigned node.

[0091] In this embodiment of the invention, the migration matching degree is calculated considering multiple factors: the more abundant the remaining resources of the node, the higher the matching degree; the greater the communication bandwidth with the original assigned node (edge ​​server D), the faster the data transmission speed, and the higher the matching degree; the network transmission cost of the subtask migration, i.e., the product of the amount of data to be transmitted for migration and the unit data transmission cost, the lower the cost, the higher the matching degree. Weights are assigned to each factor, such as a resource remaining amount weight of 0.4, a communication bandwidth weight of 0.3, and a transmission cost weight of 0.3. The migration matching degree of the migrateable object subtask 6 with the edge server E and the cloud node in the new candidate node set is calculated separately. For example, if the edge server E has a resource remaining amount score of 80, a communication bandwidth score of 90, and a transmission cost score of 70, the migration matching degree is 80×0.4+90×0.3+70×0.3; the corresponding scores of the cloud node are 90, 70, and 60, and the migration matching degree is 90×0.4+70×0.3+60×0.3. Therefore, edge server E had a higher migration match and was identified as the newly assigned node.

[0092] Step 1410-5: Generate a task migration instruction, which includes the subtask identifier of the migrateable object, the original assigned node identifier, the new assigned node identifier, the current execution progress data, and a list of data files to be transferred.

[0093] In this embodiment of the invention, the task migration instruction is a structured message encoded in JSON format. Its subtask identifier is "Task_6", the original assigned node is the IP address of edge server D "192.168.1.10", the new assigned node is the IP address of edge server E "192.168.1.11", the current execution progress is "30%", and the list of data files to be transmitted includes the intermediate result file "result_temp.dat" of subtask 6, the configuration file "config.xml", and the dependency library file "lib.tar.gz". This information ensures that the original and new assigned nodes accurately identify and process the migration task.

[0094] Step 1410-6: Send the task migration command to both the original and new allocation nodes. Upon receiving the command, the original allocation node will stop executing the subtask, package the current execution progress data and related data files, and transmit them to the new allocation node via remote communication. After receiving the transmitted data, the new allocation node will restore the subtask execution environment based on the execution progress data, continue executing the subtask until completion, and synchronously update its own resource availability status table and the list of allocated subtasks.

[0095] In this embodiment of the invention, the task migration instruction is sent to both the original allocation node edge server D and the newly allocated node edge server E via the TCP protocol. Upon receiving the instruction, edge server D immediately pauses the execution of subtask 6, compresses and packages the current execution progress data (such as program counter status and memory data snapshot) and files in the data file list, generating a "Task_6_migrate.tar" file, which is then transferred to edge server E via the FTP protocol. Upon receiving the file, edge server E decompresses it and verifies its integrity. Then, based on the execution progress data, it restores the execution environment of subtask 6, such as loading the memory data snapshot and setting the program counter, and resumes execution of subtask 6 from 30% progress. During execution, edge server E updates its own resource availability status table, marking the resources allocated to subtask 6 as occupied, and adds the information of subtask 6 to the list of allocated subtasks.

[0096] Step 1410-7: Use the task execution process adjustment mechanism to record the migration time, node identifiers before and after migration, resource changes, and subtask execution status corresponding to the current task migration.

[0097] In this embodiment of the invention, after the task migration is completed, the task execution process adjustment mechanism creates a migration record, which records the start and end times of the migration (migration time), the original assigned node identifier (edge ​​server D) and the newly assigned node identifier (edge ​​server E), the resource changes (edge ​​server D releases 1.5 CPU cores, 8GB of storage, and 4Mbps of network; edge server E occupies the corresponding resources), and the subtask execution status (30% before migration, continuing execution after migration). This record is stored in the task migration log and is used for migration effect analysis and system optimization.

[0098] In an optional embodiment, the method further includes: Step 210: Establish a cross-node task execution status synchronization mechanism during the cloud-edge-device collaborative control task execution process, and collect and summarize the sub-task execution progress data, intermediate result data and resource consumption data of each node in the terminal layer, edge layer and cloud layer in real time.

[0099] In this embodiment of the invention, the cross-node task execution status synchronization mechanism adopts a publish-subscribe model. The terminal layer intelligent monitoring camera, the edge layer edge server cluster, and the cloud layer cloud management system act as subscribers, while the task scheduling center acts as the publisher. During the execution of a subtask, each node publishes its subtask execution progress data (e.g., completion percentage), intermediate result data (e.g., intermediate image features of target detection), and resource consumption data (e.g., CPU usage, memory usage) to a message queue at preset time intervals (e.g., 1 second). The task scheduling center subscribes to and receives this data from the message queue, then performs aggregation processing, associating data from different nodes according to their subtask IDs and storing it in a centralized task execution status database. For example, the intelligent monitoring camera publishes that the video decoding subtask execution progress is 50%, the intermediate result is the decoded image frame data, and the resource consumption is 20% CPU; the edge server publishes that the target detection subtask execution progress is 30%, the intermediate result is the detected target coordinates, and the resource consumption is 40% CPU. After aggregating this data, the task scheduling center can comprehensively grasp the execution status of the entire cloud-edge-device collaborative control task.

[0100] Step 220: Perform time series analysis on the execution progress data, calculate the deviation between the actual and planned execution rates, and identify influencing factors, including node resource fluctuations, changes in remote communication latency, and data interaction blockage between subtasks, when the deviation exceeds the preset range.

[0101] In this embodiment of the invention, time series analysis is performed on the summarized subtask execution progress data, and the curve of actual execution progress over time is compared with the planned execution progress curve. The planned execution rate is calculated based on the estimated execution time of the subtask set during the task planning stage. For example, if a subtask is planned to be completed in 5 minutes, the planned execution rate is 20% per minute. The actual execution rate is obtained by dividing the change in actual execution progress by the corresponding time interval. The deviation value is the percentage of the difference between the actual execution rate and the planned execution rate divided by the planned execution rate. When the deviation value exceeds a preset range (e.g., ±10%), the influencing factor identification process is initiated. By analyzing resource consumption data, it is determined whether there are fluctuations in node resources, such as a sudden increase or decrease in CPU utilization; by checking the latency indicators in network transmission data, it is determined whether there are changes in remote communication latency; by analyzing the transmission timestamps of intermediate result data, it is determined whether there is data interaction blockage between subtasks, such as the intermediate result of subtask MA not being transmitted to subtask MB on time, causing subtask MB to wait.

[0102] Step 230: Invoke the corresponding collaborative optimization strategy according to the type of influencing factor: when node resources fluctuate, trigger the dynamic resource scheduling strategy to adjust the node resource allocation ratio; when remote communication latency changes, start the communication path optimization strategy to reselect the latency mitigation link; when data interaction between subtasks is blocked, execute the data transmission priority adjustment strategy to transmit the subtask interaction data of the target priority according to the priority identifier.

[0103] In this embodiment of the invention, when a node resource fluctuation is identified as an influencing factor, such as a sudden decrease in the CPU availability of the current subtask due to the CPU resources of the edge server being occupied by other high-priority tasks, a dynamic resource scheduling strategy is triggered. This strategy adjusts the node's resource scheduling algorithm, such as increasing the CPU scheduling priority of the current subtask or allocating CPU resources from other low-load nodes, to increase the CPU resource allocation ratio of the current subtask. If the influencing factor is a change in remote communication latency, such as an increase in network latency from the smart surveillance camera to the edge server from 50ms to 200ms, a communication path optimization strategy is initiated. By querying the network topology and real-time link status, a communication path with lower latency is reselected, such as switching to another base station or using a backup network interface. When the influencing factor is data interaction blockage between subtasks, such as subtask MC waiting for the intermediate result of subtask MD for more than a preset time, a data transmission priority adjustment strategy is executed. Based on the priority identifier of the subtask, the data interaction of subtasks with target priorities (such as P0, P1) is given higher transmission priority and transmitted in the network first, ensuring smooth data interaction of high-priority subtasks.

[0104] Step 240: Convert the collaborative optimization strategy into control commands and send them to the corresponding nodes so that the nodes can adjust the subtask execution parameters after receiving them.

[0105] In this embodiment of the invention, the collaborative optimization strategy is a high-level strategy description that needs to be converted into specific control instructions before it can be executed by the nodes. For example, the resource dynamic scheduling strategy is converted into control instructions such as "adjust the CPU scheduling priority of subtask X to high" or "allocate 2 CPU cores from node A to node B"; the communication path optimization strategy is converted into instructions such as "switch the communication path of subtask Y to link L" or "set the network transmission bandwidth to Mbps"; and the data transmission priority adjustment strategy is converted into instructions such as "set the data transmission priority of the interaction between subtasks MC and MD to P0". These control instructions adopt a standardized format, such as JSON, and include information such as instruction type, target node, subtask ID, and specific parameters. The control instructions are sent to the corresponding nodes through a remote communication protocol. After receiving the instructions, the nodes parse the parameters and adjust the execution parameters of the subtasks, such as modifying process priority, switching network configuration, or adjusting the data transmission queue.

[0106] Step 250: Monitor the execution effect of the optimization strategy in real time, compare the changes in the deviation value of the subtask execution rate before and after the strategy execution, and if the change in the deviation value does not meet expectations, re-identify the influencing factors and adjust the strategy.

[0107] In this embodiment of the invention, after the control command corresponding to the collaborative optimization strategy is issued, the task scheduling center continues to collect the execution progress data of subtasks through the task execution status synchronization mechanism and calculates the actual execution rate deviation value after the strategy is executed. This deviation value is compared with the deviation value before the strategy is executed to determine whether the deviation value is changing in a decreasing direction and achieving the expected improvement effect, such as the expected deviation value decreasing from 20% to within 5%. If the actual deviation value change meets expectations, such as decreasing from 20% to 8%, the optimization strategy is considered effective; if the deviation value change does not meet expectations, such as only decreasing from 20% to 18%, the influencing factor identification process is re-executed, as the previous influencing factor identification may have been inaccurate or there may be other unidentified factors. Then, the collaborative optimization strategy is adjusted based on the newly identified influencing factors, such as simultaneously adjusting resource allocation and communication paths.

[0108] Step 260: Establish a feedback mechanism to record the types of influencing factors, strategies adopted, execution effects, and related parameter data for each optimization, and store them in the experience knowledge base; based on the experience knowledge base, mine the mapping relationship between influencing factors and the optimal collaborative optimization strategy to generate an adaptive collaborative optimization decision model.

[0109] In this embodiment of the invention, after each execution of a collaborative optimization strategy, the feedback mechanism records the type of influencing factor (e.g., node resource fluctuation), the adopted collaborative optimization strategy (e.g., dynamic resource scheduling strategy), the execution effect (e.g., deviation value decreasing from 20% to 8%), and related parameter data (e.g., the number of CPU cores allocated, the adjusted priority value) in the experience knowledge base. The experience knowledge base is stored using a relational database, and each record contains the above information. As optimization cases accumulate, data mining algorithms (e.g., association rule mining, decision tree learning) are used to analyze the data in the experience knowledge base to uncover the mapping relationship between the type of influencing factor and the optimal collaborative optimization strategy. For example, when the influencing factor is node CPU resource fluctuation and the fluctuation range exceeds 30%, the optimal strategy is a dynamic resource scheduling strategy combined with task migration. Based on these mapping relationships, an adaptive collaborative optimization decision model is trained and generated. This model can automatically recommend the optimal collaborative optimization strategy based on the real-time identified influencing factors, thereby improving the system's adaptive optimization capability.

[0110] As another optional embodiment, the method further includes: Step 310: Collect historical execution data of cloud-edge-device collaborative control tasks within a preset period. The historical execution data includes task migration records, resource fluctuation characteristic data, and task execution efficiency indicators.

[0111] In this embodiment of the invention, a preset period of one month is used to collect historical execution data of cloud-edge-device collaborative control tasks by querying task migration logs, resource monitoring databases, and task execution status databases. Task migration records include migration time, nodes before and after migration, migration reasons, and migration data volume; resource fluctuation characteristic data includes curves showing the usage rates of CPU, memory, network, and other resources of each node over time, fluctuation amplitude, and fluctuation frequency; task execution efficiency indicators include task completion time, average execution rate, resource utilization, and communication latency. For example, 100 task migration records were collected in the past month, of which 80 were caused by node resource fluctuations; resource fluctuation characteristic data shows that the CPU utilization rate of the edge server fluctuates significantly between 9:00 AM and 11:00 AM on weekdays; the average task completion time in the task execution efficiency indicators is 10 minutes, and the average resource utilization rate is 60%.

[0112] Step 320: Extract features from the historical execution data. Extract the migration triggering reason, node level before and after migration, and migration duration from the task migration record. Identify the resource fluctuation cycle, fluctuation amplitude, and fluctuation correlation from the resource fluctuation feature data. Extract the deviation rate between actual execution time and planned time, data transmission success rate, and resource utilization change from the task execution efficiency indicators. Generate a historical feature dataset containing migration feature vectors, fluctuation pattern matrix, and efficiency deviation sequence.

[0113] In this embodiment of the invention, when extracting features from historical execution data, for task migration records, the migration triggering reasons (such as insufficient resources, communication failures) are encoded, the node levels before and after migration (terminal-edge, edge-cloud, etc.) are classified and marked, and the migration duration is converted into numerical features, which are combined to form a migration feature vector, such as [insufficient resources encoding, terminal-edge marking, 5 seconds]. From the resource fluctuation feature data, the resource fluctuation cycle (such as a 24-hour cycle) is identified through Fourier transform, the fluctuation amplitude (such as the difference between the maximum and minimum values) is calculated, the fluctuation correlation between different resources (such as the correlation coefficient between CPU utilization and network bandwidth utilization) is analyzed, and a fluctuation pattern matrix is ​​constructed. The rows of the matrix represent resource types, and the columns represent fluctuation features (cycle, amplitude, correlation). From the task execution efficiency indicators, the deviation rate between the actual execution time and the planned time ((actual - planned) / planned) is calculated, the statistical data transmission success rate (number of successful transmissions / total number of transmissions) is calculated, and the change in resource utilization (current utilization - historical average utilization) is calculated to form an efficiency deviation sequence, such as [deviation rate 0.2, success rate 0.95, change 0.1]. The migration feature vector, fluctuation pattern matrix, and efficiency deviation sequence are integrated to generate a historical feature dataset.

[0114] Step 330: Input the historical feature dataset into the preset strategy evaluation model, determine the mapping relationship between the migration feature vector and the efficiency deviation sequence through correlation analysis, and identify the strategy defect patterns that cause the efficiency deviation to exceed the set value. The strategy defect patterns include the frequent migration pattern of priority tasks, the resource prediction deviation pattern of target edge nodes, and the cross-layer data transmission blocking pattern.

[0115] In this embodiment of the invention, the preset strategy evaluation model employs an association rule mining algorithm. The migration feature vectors from the historical feature dataset are used as the first term, and the efficiency deviation sequence as the second term. The mapping relationship between the two is determined by calculating support and confidence. When the deviation rate in the efficiency deviation sequence exceeds a set value (e.g., 0.15), an efficiency problem is considered to exist. At this point, the corresponding migration feature vectors are backtracked, and their common features are analyzed to identify strategy defect patterns. The frequent migration pattern of priority tasks is characterized by high-priority (e.g., P0, P1) subtasks migrating multiple times within a short period (e.g., within 10 minutes), leading to execution interruptions and resource waste, thus causing efficiency deviations. The target edge node resource estimation deviation pattern refers to a large discrepancy between the estimated and actual values ​​of edge node resource availability (e.g., estimated availability of 4 CPU cores, actual availability of only 2 cores), resulting in insufficient allocated resources for subtasks and low execution efficiency. The cross-layer data transmission blocking pattern is characterized by data transmission between the cloud and edge nodes or between edge nodes and terminals not completing for a long time (e.g., more than 10 seconds), causing subtasks to wait for data, resulting in efficiency deviations.

[0116] Step 340: Generate corresponding strategy adjustment instructions based on the strategy defect pattern. When a frequent migration pattern of priority tasks is identified, adjust the time-sensitive factor calculation logic in the collaborative control priority labeling operation. When a target edge node resource estimation deviation pattern is identified, correct the network transmission resource quantity calculation parameters in the resource demand estimation unit. When a cross-layer data transmission blocking pattern is identified, optimize the directed graph topology of subtask data dependencies in the task structure labeling information.

[0117] In this embodiment of the invention, for the frequent migration mode of priority tasks, the calculation logic of the time-sensitive factor is adjusted, and the weight of the time-sensitive factor for high-priority subtasks is increased, so that they are more likely to be allocated to nodes with stable resources during resource allocation, reducing the possibility of migration. For example, the original time-sensitive factor = real-time requirement × 0.5 is adjusted to time-sensitive factor = real-time requirement × 0.7 + priority coefficient × 0.3. When the resource estimation deviation mode of the target edge node is identified, the calculation parameter of network transmission resource volume in the resource demand estimation unit is corrected. For example, the original network transmission resource volume = data volume / transmission rate is corrected to network transmission resource volume = data volume / (transmission rate × resource fluctuation coefficient), where the resource fluctuation coefficient is calculated based on historical resource fluctuation characteristic data to improve the accuracy of resource estimation. For the cross-layer data transmission blocking mode, the directed graph topology of the data dependency relationship of subtasks is optimized to reduce the number of subtasks for cross-layer data transmission. Data interaction subtasks that can be completed at the same level are adjusted to the same level, or data caching nodes are added. For example, intermediate data that originally needs to be directly transmitted from the terminal to the cloud is first transmitted to the edge node for caching, and then the edge node transmits it to the cloud in batches, optimizing the data flow.

[0118] Step 350: Feed the strategy adjustment instruction back to the task scheduling decision model, update the generation rules of the task scheduling input dataset, the splitting logic of the original task structure information, and the weighted fusion algorithm of the collaborative control priority labeling, and form a collaborative strategy evolution loop based on historical execution data.

[0119] In this embodiment of the invention, the generated strategy adjustment instructions are converted into configuration parameters that the task scheduling decision model can recognize. For example, instructions to adjust the calculation logic of time-sensitive factors are converted into parameter updates for the time-sensitive factor module in the model; instructions to correct the calculation parameters of network transmission resource quantity update the configuration file of the resource demand estimation unit; instructions to optimize the topology of the directed graph of data dependencies modify the splitting rules of the task decomposition module. These configuration parameters are fed back to the task scheduling decision model, which updates the generation rules of the task scheduling input dataset according to the new parameters, such as adding a resource fluctuation coefficient when generating the dataset; updates the splitting logic of the original task structure information, such as considering the optimization of data transmission paths during splitting; and adjusts the weighted fusion algorithm for collaborative control priority labeling, such as increasing the weight of time-sensitive factors. Through such updates, the task scheduling decision model can continuously optimize its decision logic based on historical execution data, forming an evolutionary cycle of collaborative strategies and continuously improving the efficiency and stability of cloud-edge-device collaborative control.

[0120] This invention improves the efficiency of task execution, the accuracy of resource utilization, and the stability of system operation in complex scenarios. Specifically, by integrating real-time operating status data of terminal devices, computing resource information of edge nodes, and global optimization targets in the cloud to generate task scheduling input datasets, deep collaboration of multi-source heterogeneous data is achieved. Based on a preset task scheduling decision model, collaborative control tasks are structurally decomposed to accurately identify sub-task types, data dependencies, and resource requirements, ensuring the scientific and logical nature of task decomposition. By combining terminal priority ranking, edge load rate analysis, and cloud weight configuration for collaborative control priority labeling, dynamic quantification and global optimization of sub-task priorities are achieved, avoiding blind task execution. Through multi-round resource allocation simulation and conflict resolution mechanisms, hierarchical resource allocation results are generated, and a task execution process adjustment mechanism based on real-time resource changes is configured, effectively solving the problems of resource competition and fluctuation adaptation between cloud, edge, and terminal nodes, and improving resource utilization and task execution efficiency. In summary, the embodiments of the present invention break through the limitations of static task allocation and isolated resource scheduling in traditional cloud-edge-device collaboration. Through dynamic collaboration and closed-loop optimization across the entire link, the collaborative control capability of cloud-edge-device systems in complex environments is qualitatively improved, providing a solid technical guarantee for the efficient operation of large-scale distributed intelligent systems.

[0121] See Figure 2 , Figure 2The diagram illustrates a block diagram of a cloud-edge-device collaborative control system according to an embodiment of the present invention. The system includes: a data generation module, used to generate a task scheduling input dataset based on real-time operating status data generated by the terminal device during remote communication, current computing resource information of the edge computing node, and global optimization target parameters issued by the cloud management system; a task splitting module, used to perform structured splitting of the cloud-edge-device collaborative control task based on the task scheduling input dataset and by calling a preset task scheduling decision model to obtain original task structure information including sub-task types, data dependencies, and resource requirements; and a control annotation module, used to combine the terminal device's operating status priority ranking results, edge computing node resource load rate analysis data, and cloud global optimization target parameters. The system optimizes target weight configuration information and performs collaborative control priority labeling operations on each subtask in the original task structure information to generate task structure labeling information carrying priority identifiers. The task allocation module performs cyclic matching and deduction processing, including multi-round resource allocation simulation, on the task structure labeling information based on the subtask's priority identifier, resource requirements, and the resource availability status of cloud-edge-device nodes to obtain hierarchical resource allocation results. Based on these hierarchical resource allocation results, each subtask is distributed to its corresponding cloud-edge-device node for execution, and a task execution process adjustment mechanism based on real-time resource changes is configured to adapt to resource fluctuations. The cloud-edge-device nodes include at least a first node corresponding to the terminal device, a second node corresponding to the edge computing node, and a third node corresponding to the cloud management system.

[0122] See Figure 3 , Figure 3 A structural block diagram of a cloud-edge-device collaborative control system 1200 is shown, including a memory 1210 and a processor 1220. The memory 1210 is used to store computer programs / instructions, and the processor 1220 is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, they implement the steps of the cloud-edge-device collaborative control method based on remote communication described above.

[0123] Furthermore, a computer storage medium is provided that stores a computer program / instruction, which, when executed by a processor, implements the aforementioned cloud-edge-device collaborative control method based on remote communication.

[0124] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals. The preferred embodiments of the present invention disclosed above are only for the purpose of illustrating the present invention. Optional embodiments do not describe all details exhaustively, nor do they limit the invention to only the specific implementations described. Obviously, many modifications and variations can be made according to the content of the embodiments of the present invention. The present invention selects and specifically describes these embodiments to better explain the principles and practical applications of the embodiments of the present invention, so that those skilled in the art can better understand and utilize the present invention. The above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this invention shall be included within the scope of protection of this invention.

Claims

1. A cloud-edge-device collaborative control method based on remote communication, characterized in that, The method includes: Based on the real-time operating status data generated by the terminal device during remote communication, the current computing resource information of the edge computing node, and the global optimization target parameters issued by the cloud management system, a task scheduling input dataset is generated. Based on the task scheduling input dataset, a preset task scheduling decision model is invoked to perform structured decomposition of the cloud-edge-device collaborative control task to obtain original task structure information including sub-task types, data dependencies, and resource requirements. The task scheduling input dataset is then input into the task decomposition module of the task scheduling decision model. The task feature extraction unit of the task decomposition module divides the control task into functional modules, identifies the basic operation units constituting the control task, and categorizes these basic operation units into different sub-task types. The dependency analysis unit of the task scheduling decision model tracks the data flow of the basic operation units corresponding to different sub-task types, recording the input-output data correlation between sub-tasks. Furthermore, a data dependency network represented as a directed graph is generated; the resource requirement estimation unit of the task scheduling decision model is invoked to calculate the amount of computing resources, storage resources, and data transmission resources required for each subtask during execution, based on the historical execution records of subtask types and the current transmission bandwidth status of the remote communication network, forming a resource requirement descriptor; the data dependency network and the resource requirement descriptor are integrated and processed, and the data dependency network is converted into a linear execution sequence according to the execution order constraints of the subtasks, and the resource requirement descriptor is associated and bound with the corresponding subtask type to generate the original task structure information containing the subtask type field, the directed graph of data dependency, and the resource requirement descriptor; Combining the priority ranking results of the terminal device's operating status, the resource load rate analysis data of the edge computing node, and the global optimization target weight configuration information in the cloud, a collaborative control priority labeling operation is performed on each subtask in the original task structure information to generate task structure labeling information carrying priority identifiers; Based on the priority identifiers, resource requirements, and resource availability status of cloud-edge-device nodes for subtasks, the task structure annotation information is subjected to cyclic matching and deduction processing based on multi-round resource allocation simulation to obtain hierarchical resource allocation results. According to the hierarchical resource allocation results, each subtask is distributed to the corresponding cloud-edge-device nodes for execution, and a task execution process adjustment mechanism based on real-time resource changes is configured to adapt to resource fluctuations. The cloud-edge-device nodes include at least a first node corresponding to the terminal device, a second node corresponding to the edge computing node, and a third node corresponding to the cloud management system.

2. The method according to claim 1, characterized in that, The process combines the priority ranking results of the terminal device's operating status, the resource load rate analysis data of the edge computing node, and the global optimization target weight configuration information in the cloud. It then performs a collaborative control priority labeling operation on each subtask in the original task structure information to generate task structure labeling information carrying priority identifiers, including: The operating status parameters of the terminal devices are obtained and normalized. The operating health index of each terminal device is calculated through the status evaluation function. The terminal devices are sorted from high to low according to the health index to obtain the priority ranking result of the operating status. Collect the resource usage of edge computing nodes at the current moment, and combine the load rate analysis to determine the real-time occupancy ratio of computing resources, storage resources and network resources, and generate resource load rate analysis data for edge computing nodes; Receive global optimization target weight configuration information from the cloud management system. The global optimization target weight configuration information includes task execution efficiency weight, resource utilization weight, and communication latency weight. Based on the priority ranking results of the running status, the basic value of the execution priority of the subtask on the terminal device is determined. The basic value is then corrected by combining the resource load rate analysis data of the edge computing node. When the load rate of the edge node exceeds the preset load rate, the priority correction coefficient of the corresponding subtask is reduced. Based on the weights in the global optimization target weight configuration information in the cloud, the corrected priority is weighted and fused. A time-sensitive factor is set according to the real-time requirements of the sub-task, a transmission cost factor is set according to the data transmission volume, and a computation time factor is set according to the computational complexity characteristics. The weighted fusion result is then multiplied to obtain the final priority value. The final priority value is converted into a priority identifier. The subtasks in the original task structure information are reordered in descending order of priority. The priority identifier is associated with the corresponding subtask type, data dependency, and resource requirements and stored to generate task structure annotation information carrying the priority identifier.

3. The method according to claim 2, characterized in that, The process of acquiring and normalizing the operating status parameters of the terminal devices, calculating the operating health index of each terminal device through a status evaluation function, and sorting the terminal devices by health index from high to low to obtain the priority ranking result of their operating status includes: The system collects operational status parameters of the terminal device during remote communication, which reflect the device's processor operating status, memory usage status, and network connection status. The collected operating status parameters are processed to unify the dimensions, mapping the value ranges of different types of parameters to the same interval; A state evaluation function is constructed. The state evaluation function takes the normalized operating state parameters as input, determines the influence weight of each operating state parameter through the analytic hierarchy process, and obtains the operating health index by multiplying the parameter value by the corresponding weight and summing the results. The health indicators of all terminal devices are compared, and the terminal devices are sorted in descending order of indicator values ​​to generate a priority ranking result of the terminal devices' operating status. The terminal devices ranked higher have higher task execution priority. The data collection process gathers information on the resource usage of edge computing nodes at the current moment. Combined with load rate analysis, it identifies the real-time occupancy ratios of computing, storage, and network resources to generate resource load rate analysis data for the edge computing nodes. This data includes: Collect multi-dimensional resource usage data of edge computing nodes at the current moment. The resource usage data includes the task queue length and process blocking status of computing resources, the file handle activity and cache hit rate trend of storage resources, and the connection session establishment frequency and data packet retransmission events of network resources. Time-series feature extraction is performed on the collected multi-dimensional resource usage data to capture the usage fluctuation patterns of each resource dimension at different time granularities, and to identify sudden peak patterns, periodic fluctuation patterns, and trend growth patterns of resource usage. Based on the preset resource status determination rules, the status of each resource dimension is classified according to the identified fluctuation patterns. Computing resources are marked as idle, stable or strained, storage resources are marked as sufficient, saturated or overflowed, and network resources are marked as unobstructed, congested or interrupted. Association rule mining is performed on the classification status of different resource dimensions to identify the co-occurrence pattern of computing resource shortage and network resource congestion, the impact weight of storage resource saturation on computing resource processing latency, and to establish a collaborative impact model of resource load. By combining the classification status, fluctuation pattern description, and collaborative influence relationship model of each resource dimension, resource load rate analysis data containing resource status labels, feature pattern vectors, and a set of association rules is generated.

4. The method according to claim 1, characterized in that, The process involves performing a cyclic matching and deduction process based on multi-round resource allocation simulation on the task structure annotation information according to the priority identifier, resource requirements, and resource availability status of cloud-edge-device nodes to obtain hierarchical resource allocation results. Based on these results, each subtask is distributed to its corresponding cloud-edge-device node for execution. A task execution process adjustment mechanism based on real-time resource changes is configured to adapt to resource fluctuations, including: Initialize the resource availability status table for cloud-edge-device nodes and record the current computing resource availability, storage resource availability, and network resource availability for each node in the terminal layer, edge layer, and cloud layer. Extract a list of subtasks carrying priority identifiers from the task structure annotation information, and select subtasks in descending order of priority for resource matching processing. For the currently selected subtask, based on the computing resources, storage resources, and data transmission resources in the corresponding resource requirements, the resource availability status table of cloud-edge-device nodes is traversed to filter out a set of candidate nodes whose resource availability meets the requirements of the subtask. Calculate the matching degree between the subtask and each node in the candidate node set, and determine the node with the highest matching degree as the optimal allocation node; Assign the subtask to the optimal allocation node, update the resource availability status table of the optimal allocation node, and record the subtask type, allocation node identifier, and resource usage of the current assignment. Repeat the steps of selecting subtasks, filtering candidate nodes, calculating matching degree and allocating resources until all subtasks have completed the initial allocation and the first round of resource allocation results are formed. Conflict detection is performed on the first round of resource allocation results. When multiple subtasks compete for the same node resource, resulting in insufficient resources, conflict resolution is carried out according to the subtask priority identifier. Subtasks whose priority does not meet the preset index are re-entered into the candidate node screening process. After multiple rounds of resource allocation simulation and conflict resolution, a stable correspondence between each subtask and cloud-edge-device nodes is generated. The stable correspondence is then organized into layers according to the terminal layer, edge layer, and cloud layer to obtain a hierarchical resource allocation result containing a list of subtasks allocated to each layer of nodes. Based on the hierarchical resource allocation results, the execution instructions and related data of each subtask are distributed to the corresponding cloud-edge-device nodes through a remote communication protocol, and a resource monitoring mechanism is established simultaneously to collect the resource status change data of each node in real time. When the resource monitoring mechanism detects that the available resources of a node are lower than the preset available resources, the task execution process adjustment mechanism is triggered. Based on the current resource status, the optimal allocation node for the subtask is recalculated, a task migration instruction is generated and sent to the original allocation node and the new allocation node, so as to realize the dynamic migration of the subtask between cloud, edge and terminal nodes to adapt to resource fluctuations.

5. The method according to claim 4, characterized in that, The calculation of the matching degree between the subtask and each node in the candidate node set, and the determination of the node with the highest matching degree as the optimal allocation node, includes: Obtain the remaining resources of each node in the candidate node set, calculate the ratio of the remaining resources to the resource requirements of the subtask. When the ratio is greater than 1, the ratio is used to calculate the matching degree. When the ratio is less than 1, the ratio is converted into a penalty coefficient to reduce the matching degree. Analyze the collaborative execution efficiency between the current subtask and other subtasks already assigned to candidate nodes. Query the efficiency improvement value corresponding to the subtask type combination through the collaboration coefficient matrix. The higher the efficiency improvement value, the greater the matching degree. The communication delay time between the subtask data source and the candidate node is measured by the remote communication delay detection module. The communication delay time is converted into a delay impact factor. The longer the delay time, the smaller the delay impact factor, and the greater the negative impact on the matching degree. Based on the set weights for remaining resources, collaborative execution efficiency, and remote communication latency, the ratio of remaining resources, the improvement in collaborative execution efficiency, and the latency impact factor are weighted and summed to obtain the comprehensive matching degree between the subtask and the candidate node. Compare the overall matching degree of each candidate node, and select the node with the highest overall matching degree as the optimal allocation node for the current subtask. When there are multiple nodes with the same matching degree, select the node with the lowest level.

6. The method according to claim 4, characterized in that, The process of conflict detection in the first round of resource allocation results involves resolving conflicts based on the priority of each subtask when multiple subtasks compete for the same node resource, leading to resource shortages. Subtasks whose priorities do not meet preset criteria re-enter the candidate node selection process, including: Iterate through the list of allocated subtasks for each cloud-edge node in the first round of resource allocation results, and calculate the total resource requirements of all subtasks allocated to each node. The total resource demand is compared with the initial resource availability in the node's resource availability status table. When the total resource demand exceeds the initial resource availability, the node is determined to have a resource allocation conflict. Extract the list of subtasks assigned to nodes with resource allocation conflicts, sort the subtasks from low to high priority, and select the subtask with the lowest priority as the conflict resolution target. Remove the conflict resolution object from the list of assigned subtasks of the original node, restore the corresponding amount of resources occupied by the node, and update the node's resource availability status table. The conflict resolution object is added back to the list of subtasks to be assigned, and then re-enters the candidate node filtering process according to priority order. The matching degree is recalculated and the object is assigned to other candidate nodes with available resources. Repeat the steps of conflict detection, conflict resolution, and reallocation until the total resource requirements of all nodes do not exceed their initial resource availability. Once the conflict detection is passed, the resource allocation result is generated.

7. The method according to claim 4, characterized in that, When the resource monitoring mechanism detects that the available resources of a node are lower than the preset available resources, the task execution process adjustment mechanism is triggered. Based on the current resource status, the optimal allocation node for the subtask is recalculated, a task migration instruction is generated and sent to both the original and new allocation nodes, enabling dynamic migration of subtasks between cloud, edge, and endpoint nodes to adapt to resource fluctuations. This includes: The resource monitoring mechanism collects resource availability data of each cloud-edge node at preset time intervals. When any one of the available computing resources, available storage resources, or available network resources of the target node is detected to be lower than the corresponding preset value, a resource warning signal is generated. After receiving a resource warning signal using the task execution process adjustment mechanism, the list of subtasks currently being executed by the node is extracted, and the migrateable subtasks are selected based on their execution progress. Subtasks with execution progress below a preset percentage are marked as migrateable objects. For each migrateable object, the resource matching process is re-executed, the resource availability status table of the current cloud-edge-end nodes is traversed, and a new set of candidate nodes whose resource availability meets the remaining resource requirements of the subtask is selected. By combining the remaining resources of the node, the communication bandwidth with the original assigned node, and the network transmission cost of the subtask migration, the migration matching degree between the migrateable object and each node in the new candidate node set is calculated, and the node with the highest migration matching degree is determined as the new assigned node. Generate a task migration instruction, which includes the subtask identifier of the migrateable object, the original assigned node identifier, the new assigned node identifier, the current execution progress data, and a list of data files to be transferred; The task migration command is sent to both the original and new allocation nodes. Upon receiving the command, the original allocation node stops the execution of the subtask, packages the current execution progress data and related data files, and transmits them to the new allocation node via remote communication. After receiving the transmitted data, the new allocation node restores the subtask execution environment based on the execution progress data, continues to execute the subtask until completion, and synchronously updates its own resource availability status table and the list of allocated subtasks. The task execution process adjustment mechanism is used to record the migration time, node identifiers before and after the migration, resource changes, and subtask execution status corresponding to the current task migration.

8. A cloud-edge-device collaborative control system, characterized in that, include: Memory and processor; The memory is used to store computer programs or instructions, and the processor is used to execute the computer programs or instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 7.

9. A computer storage medium, characterized in that, The device stores a computer program or instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 7.

Images