A cloud edge-based low-energy experimental equipment scheduling method

Abstract

The application discloses a kind of based on cloud edge end's low-energy consumption experimental equipment scheduling method, comprising: initialization system parameter and establish cloud edge end collaborative system model;Real-time sensing system resource state and experimental equipment task request;Task scheduling strategy is calculated based on deep reinforcement learning algorithm, introduce self-attention mechanism to high-dimensional state information is identified and weighted with key features, and energy consumption model is converted into the core penalty function of reinforcement learning, with energy consumption minimization as optimization goal;Task scheduling is executed, for the strong time sequence constraint and physical state dependence of experimental equipment task, special optimization including pre-scheduling mechanism, atomicity protection and end side feature extraction is carried out.The application realizes real-time sensing and dynamic allocation to heterogeneous computing resources through cloud edge end collaboration and intelligent scheduling, significantly reduces the overall energy consumption of system under the satisfaction of task delay constraint, ensures the continuity and data integrity of experimental process.

Description

Technical Field

[0001] This invention relates to the field of cloud-edge-device collaboration technology, and more specifically, to a low-energy experimental equipment scheduling method based on cloud-edge-device. Background Technology

[0002] With the rapid development of technologies such as cloud computing, edge computing, and artificial intelligence, computing power demands are becoming increasingly diversified and dynamic. The integrated management of heterogeneous computing resources has become crucial for improving service efficiency. In cloud-edge collaborative systems, several powerful cloud computing centers and numerous edge clusters deployed across various regions together constitute a new computing architecture. However, existing technologies suffer from several problems: untimely perception of computing resources, inflexible scheduling strategies, insufficient load balancing, and poor energy consumption control. These issues lead to low utilization of heterogeneous computing resources and high task processing latency, making it difficult to meet the dynamic task requirements of complex scenarios.

[0003] In the industrial manufacturing sector, traditional methods for assessing the condition of industrial equipment primarily rely on data-driven approaches based on cloud computing. This involves uploading operational data from industrial equipment to cloud servers and utilizing artificial intelligence and machine learning for evaluation. However, this method has limitations: constrained by bandwidth, computational load, and cost, it cannot meet the demands for high timeliness and information security in industrial equipment assessment. Furthermore, existing edge computing scheduling schemes often fix different computing tasks at different computational layers, resulting in limited system flexibility and resource utilization, making it difficult to adapt to dynamic industrial environments and achieve flexible production.

[0004] To address these issues, the industry has been exploring new technological solutions. For example, some research has proposed using edge computing mechanisms to extract edge features from data using edge algorithms, thereby solving the problem of real-time assessment of industrial equipment status under massive data conditions. Other research suggests offloading the computational load from the central node to the edge, preprocessing raw data through edge computing nodes to reduce the amount of data transmitted to the cloud. However, these methods still cannot meet the complexity and diverse QoS requirements of industrial production systems, particularly in terms of real-time performance and energy consumption. Therefore, there is an urgent need for a solution that comprehensively considers resource utilization, energy consumption, and load to achieve energy minimization and optimized resource scheduling.

[0005] Several invention patents have been published to address the energy consumption optimization problem in cloud-edge-device collaborative systems. For example, CN118394520A discloses a task load balancing scheduling method based on cloud-edge-device collaboration. This method involves a terminal initiating a task scheduling request, transmitting the task to the edge base station responsible for that terminal, and setting up multiple edge servers around the edge base station. Each agent collects the latency and resource requirements of tasks in its domain and communicates with other agents to formulate a task scheduling strategy. During the strategy formulation process, each agent learns and updates based on the reward signal received in the previous cycle, aiming to minimize the load standard deviation. Finally, based on the completed task scheduling strategy, a specific server in the cloud-edge-device system is assigned, and the task is transmitted to that server for processing. However, this method still suffers from insufficient intelligence and adaptability in its load balancing algorithm. CN118018610A proposes a cloud-edge collaborative resource scheduling method, system, electronic device, and readable medium. This method involves establishing energy consumption models for IT devices, power supply devices, and cooling devices in a cloud-edge collaborative system, setting energy consumption optimization targets for the system based on these models, and generating resource scheduling decisions using a reinforcement learning algorithm based on the current resource status, task requests, and energy consumption of the system to minimize the energy consumption optimization targets. However, the accuracy and practicality of this method's energy consumption models still need improvement.

[0006] In summary, the existing technology has the following drawbacks: 1. The lack of timely perception of computing resources and the lack of flexibility in scheduling strategies result in low utilization of heterogeneous computing resources and high task processing latency, making it difficult to meet the dynamic task requirements in complex scenarios. 2. Existing edge computing scheduling schemes often fix different computing tasks at different computing levels, which limits system flexibility and resource utilization, making it difficult to adapt to dynamic industrial environments and achieve flexible production; 3. Traditional industrial equipment condition assessment methods mainly rely on data-driven approaches based on cloud computing, which upload the operating data of industrial equipment to cloud servers and use artificial intelligence and machine learning methods for assessment. However, this approach cannot meet the requirements of high timeliness and information security in industrial equipment assessment. 4. Existing load balancing algorithms are not intelligent or adaptive enough, making it difficult to dynamically adjust and optimize them according to the needs of different task types and application scenarios; 5. Existing energy consumption models and optimization algorithms still need improvement in terms of accuracy and practicality, making it difficult to achieve effective control of system energy consumption and optimization of resource scheduling.

[0007] To address the aforementioned issues, how to solve the problems of low computing resource utilization, lack of flexibility in scheduling strategies, and poor energy consumption control in existing cloud-edge-device collaborative systems has become an urgent technical problem to be solved. Summary of the Invention

[0008] In view of the above-mentioned technical problems in related technologies, the present invention proposes a low-energy experimental equipment scheduling method based on cloud-edge-device, which can overcome the above-mentioned shortcomings of the prior art.

[0009] To achieve the above-mentioned technical objectives, the technical solution of the present invention is implemented as follows: A low-energy experimental equipment scheduling method based on cloud-edge-device integration includes the following steps: S1 initializes system parameters and establishes a cloud-edge-device collaborative system model, which includes an energy consumption model; S2 senses the system resource status in real time and obtains task request information from experimental equipment; S3 calculates a task scheduling strategy based on a deep reinforcement learning algorithm. In the process of calculating the task scheduling strategy, a self-attention mechanism is introduced to identify and weight key features in the state information, and the system energy consumption model is transformed into a reward signal for reinforcement learning, with energy minimization as one of the optimization objectives. S4 performs task scheduling and monitors whether the system status has reached the optimization goal; during the scheduling process, it performs special optimizations based on the characteristics of the experimental equipment tasks, and the special optimizations include at least one or more of the following: a pre-scheduling mechanism based on task dependencies, atomicity protection for high-precision real-time monitoring tasks, and data compression and feature extraction on the edge side. S5 updates system parameters and ends the current scheduling cycle.

[0010] Furthermore, step S1 specifically includes the following steps: S101 establishes an energy consumption model for the cloud-edge-device collaborative system, including an energy consumption model for the cloud computing center, an energy consumption model for the edge cluster, and an energy consumption model for the communication network. S102 constructs a system resource database to record the real-time status information of various resources; S103 initializes the task queue and stores information about tasks to be processed.

[0011] Furthermore, step S2 specifically includes the following steps: S201 collects local resource status information through edge computing nodes; S202 receives a task request from the mobile device and parses the task type and priority; S203 stores the task information in the task queue and waits for scheduling.

[0012] Furthermore, step S3 specifically includes the following steps: S301 uses a deep reinforcement learning network with an Actor-Critic architecture to train a task scheduling model based on historical task data. S302 uses machine learning algorithms to extract features and classify tasks; S303 calculates the fitness score for each task based on the system resource status and task characteristics; S304 employs a deep reinforcement learning algorithm based on an attention mechanism. The input includes the real-time load of the experimental equipment, the task pipeline dependency graph, and the energy consumption model parameters. It uses an attention mask to filter out invalid environmental noise and outputs the optimal node and power configuration for task offloading.

[0013] Furthermore, the reward function of the deep reinforcement learning network in step 301 is designed as a negative weighted cost function: R = -(ω1.E) total +ω2.D latency +ω2.I violation ), where ω1, ω2, and ω3 represent weighting coefficients, which are dynamically adjusted according to the experiment type; E total D represents total energy consumption; latency Indicates the total delay in task completion; I violation This is a mandatory penalty.

[0014] Furthermore, the specific optimizations in step S4 include: The pre-scheduling mechanism is as follows: when an experimental task with a dependency relationship is detected, the computing resources required by the subsequent task are pre-locked through the edge computing node during the execution of the preceding task. The atomicity protection is as follows: a non-interruptible flag is set for tasks identified as high-precision real-time monitoring tasks, prohibiting the tasks from migrating across nodes during execution; The data compression and feature extraction at the endpoint are as follows: a lightweight machine learning model is used at the endpoint to downsample or detect anomalies in the raw data generated by the experimental equipment, and only the data containing changes with features is uploaded.

[0015] Furthermore, step S4 further includes: S401 allocates system resources according to the scheduling policy; The S402 performs experimental task assignment. For tasks identified as critical physics experiments, it activates the atomicity protection protocol to ensure that the tasks are completed continuously within the predetermined energy consumption envelope and provides real-time feedback on the actual power consumption. To determine whether the energy consumption and resource utilization of the S403 monitoring system have reached the optimization target, a dynamic weighted sum method is used. The weights of processing latency and system energy consumption are automatically adjusted according to the experimental stage. Under the condition of meeting the basic QoS threshold, the solution with the lowest energy consumption is searched. If the target is not achieved, adjust the scheduling strategy and return to step 2; if the target is achieved, proceed to step 5.

[0016] Furthermore, step S5 further includes: S501 analyzes system operation data based on machine learning algorithms, including principal component analysis algorithm for dimensionality reduction of high-dimensional experimental environment data, and / or K-Means clustering algorithm for automatic classification of experimental tasks. S502 updates system parameters based on analysis results; S503 optimizes the task scheduling order in the task queue; S504 enters the next scheduling cycle.

[0017] The beneficial effects of this invention are as follows: By introducing a deep reinforcement learning algorithm with a self-attention mechanism, this invention can perceive high-dimensional states in real time and dynamically weight bottleneck features, thereby improving the timeliness of resource perception and scheduling flexibility. By transforming the refined energy consumption model into a reinforcement learning penalty function with the goal of minimizing energy consumption, the system can adaptively search for the load balancing path with the lowest global energy consumption under the constraint of time latency, overcoming the problem of poor energy consumption control in traditional solutions. For the strong temporal constraints and physical state dependencies of experimental equipment, resources are locked through a pre-scheduling mechanism, and atomicity protection is used to prevent critical tasks from migrating midway, thereby ensuring the continuity of experiments and data integrity. At the same time, by performing feature extraction and data cleaning on the edge, communication energy consumption and cloud load are significantly reduced, solving the defects of bandwidth limitation and high latency in traditional methods, and ultimately maximizing the utilization of heterogeneous computing resources and minimizing the overall energy consumption of the system. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0019] A low-energy experimental equipment scheduling method based on cloud-edge-device according to an embodiment of the present invention includes the following steps: Step 1: Initialize system parameters and establish a cloud-edge-device collaborative system model; Step 1 includes: Step 101: Establish an energy consumption model for the cloud-edge-device collaborative system, including an energy consumption model for the cloud computing center, an energy consumption model for the edge cluster, and an energy consumption model for the communication network. Step 102: Construct a system resource database to record the real-time status information of various resources; Step 103: Initialize the task queue and store the task information to be processed.

[0020] Step 2: Real-time monitoring of system resource status and acquisition of task request information; Step 2 includes: Step 201: Collect local resource status information through edge computing nodes; Step 202: Receive task requests from mobile devices and parse the task type and priority; Step 203: Store the task information in the task queue and wait for scheduling.

[0021] Step 3: Calculate the task scheduling strategy based on the deep reinforcement learning algorithm; in the cloud-edge-device collaborative environment, state information has high dimensionality and time-varying characteristics.

[0022] 1. This algorithm introduces a self-attention mechanism to identify key features: (1) Feature weighting: The agent does not simply receive data such as CPU and bandwidth, but automatically identifies the "bottleneck factors" in the current environment by calculating attention scores. For example, when the experimental equipment is in the high-frequency sampling stage, the attention mechanism will give higher weights to "network bandwidth" and "round-trip time (RTT)" while temporarily reducing the weight of "static storage space".

[0023] (2) Spatial perception: In a multi-edge node scenario, the attention mechanism is used to calculate the "matching degree" between each edge server and the current experimental device task, so as to accurately lock the optimal target node in the action space.

[0024] 2. Energy consumption is no longer just a statistical indicator, but serves as a core feedback mechanism (RewardSignal) in reinforcement learning: (1) Penalty function coupling: The energy consumption model established in step 101 is transformed into the instantaneous reward r_t of RL. If an action causes a surge in energy consumption at the edge node (such as triggering a high-power heat dissipation and cooling system), the reward function will drop significantly, forcing the model to learn a more energy-efficient load balancing path.

[0025] (2) Virtual energy state: The device's "remaining power" or "energy efficiency ratio" is used as one dimension of the state space. When the power is below the threshold, the actions explored by the RL strategy through epsilon-greedy will automatically tend to perform "local low-power feature extraction" rather than "full data upload to the cloud".

[0026] Step 3 includes: Step 301: Based on historical task data, train the task scheduling model using a deep reinforcement learning algorithm. Specifically, a deep reinforcement learning network with an Actor-Critic architecture is used: 1) Input (State): Triple vector Among them, T task R represents the urgency of the experimental task. resource P represents the available computing power of an edge node. energy This indicates the current power / battery status.

[0027] 2) Output (Action): Discrete action space, representing the task allocation scheme (local, edge A…N, cloud) and the device execution power level.

[0028] 3) Reward function: Designed as a negative weighted cost function: R = -(ω1.E) total +ω2.D latency +ω2.I violation ), where I violation These are hard constraint penalties (such as equipment downtime risk, experiment timeout); ω1, ω2, and ω3 represent weighting coefficients, which are dynamically adjusted according to the experiment type; E total D represents total energy consumption; latency Indicates the total delay in task completion; In step 301, the deep reinforcement learning algorithm specifically includes: Step 3011: Collect historical task data and construct a task dataset; Step 3012: Use a deep neural network to extract features from the task data; Step 3013: Calculate the priority and completion delay of each task using a deep reinforcement learning algorithm; Step 3014: Optimize task scheduling based on these metrics.

[0029] Step 302: Use machine learning algorithms to extract features and classify the task; Step 303: Calculate the fitness score for each task based on the system resource status and task characteristics; Step 304: Employ a deep reinforcement learning algorithm based on an attention mechanism. The input includes the real-time load of the experimental equipment, the task pipeline dependency graph, and the energy consumption model parameters; use an attention mask to filter out invalid environmental noise, and output the optimal node and power configuration for task offloading.

[0030] Step 4: Execute task scheduling and monitor whether the system status has reached the optimization goal. If not, return to step 2; otherwise, execute step 5. Unlike general internet requests, experimental equipment tasks have the following three core characteristics, which have been specifically optimized in this algorithm: A. Strong timing constraints and physical state dependencies Experimental equipment (such as liquid chromatographs and automated pipetting workstations) often have a streamlined nature.

[0031] Pre-scheduling mechanism: The algorithm recognizes that task A is "device preheating" and task B is "formal sampling". When A is being executed, the system will pre-lock the computing resources required by B through edge computing nodes (Resource Reservation) to ensure that the physical experiment process will not be interrupted due to computing power competition.

[0032] B. Atomicity Protection Mid-course migration is prohibited: When a task is identified as a "high-precision real-time monitoring" task, the scheduling algorithm will set a "non-interruptible" flag. Once the task starts, even if cloud computing costs decrease, cross-node migration will not be performed to prevent data interruption or equipment control failure due to network switching.

[0033] C. Sampling data compression and edge cleaning Feature Extraction: For the massive amounts of raw waveforms or images generated by experimental equipment, the algorithm prioritizes "downsampling" or "anomaly detection" on the edge using a lightweight machine learning model, uploading only the changing data containing experimental features, thereby minimizing communication power consumption.

[0034] Step 4 includes: Step 401: Allocate system resources according to the scheduling strategy; Step 402: Perform experimental task assignment. For tasks identified as "critical physics experiments", enable the atomicity protection protocol to ensure that the task is completed continuously within the predetermined energy consumption envelope, and provide real-time feedback on the actual power consumption to update the experience replay pool of the RL agent.

[0035] In step 402, the task forwarding strategy specifically includes: Step 4021: Calculate the resource utilization rate of each cloud edge; Step 4022: Select the most suitable processing node based on task type and priority; Step 4023: Forward the task to the target node via the communication network.

[0036] Step 403: Monitor whether the system's energy consumption and resource utilization have reached the optimization target. A dynamic weighted sum method is used. The system automatically adjusts the weights according to the experimental stage. High-precision acquisition stage: Automatically increase the weight of "processing latency" to ensure no packet loss in experimental data; Long-term observation phase: The weight of "system energy consumption" is automatically increased, and a low-energy consumption detection mode is entered. In addition, Pareto optimal selection logic is introduced to search for the solution with the lowest energy consumption while meeting the basic QoS threshold.

[0037] In step 403, the optimization objectives specifically include: Step 4031: Minimize system energy consumption; Step 4032: Highest resource utilization rate; Step 4033: Minimize task processing latency; Step 4034: The service quality score is the highest.

[0038] Step 404: If the goal is not achieved, adjust the scheduling strategy and return to step 2; if the goal is achieved, proceed to step 5.

[0039] Step 5: Update system parameters and end the current scheduling cycle.

[0040] Step 5 includes: Step 501: Analyze system operation data based on machine learning algorithms. For example, PCA (Principal Component Analysis): used to reduce the dimensionality of high-dimensional experimental environment data (temperature, humidity, electromagnetic interference, task concurrency, etc.), extract 3-5 principal components that affect energy consumption, and reduce the state space dimension of the RL agent.

[0041] K-Means clustering: used to automatically profile experimental tasks, automatically classifying "quick and easy" data sampling tasks with "computationally demanding" image analysis tasks, and achieving differentiated scheduling.

[0042] In step 501, the machine learning algorithm specifically includes: Step 5011: Collect system operation data, including resource usage, task processing status, and network status; Step 5012: Use machine learning algorithms to extract and analyze features from the data; Step 5013: Update system parameters based on the analysis results, including resource allocation strategies and task scheduling rules.

[0043] Step 502: Update system parameters based on analysis results; Step 503: Optimize the task scheduling order in the task queue; Step 504: Proceed to the next scheduling cycle.

[0044] To facilitate understanding of the above technical solutions of the present invention, the following detailed description of the above technical solutions of the present invention will be provided through specific usage methods. Example

[0045] This invention discloses a low-energy experimental equipment scheduling method based on cloud-edge-device architecture, applicable to a distributed environmental monitoring sensor array powered by batteries. The core logic is as follows: the energy consumption weight β in the reinforcement learning reward function is set to 0.8. The algorithm tends to retain 90% of the computation at the edge nodes for "data cleaning," transmitting only feature values ​​to the cloud to maximize device endurance. Results: Compared to traditional solutions, communication energy consumption is reduced by 45%.

[0046] The specific implementation steps are as follows: Step 1: Initialize system parameters and establish a cloud-edge-device collaborative system model.

[0047] Step 101: Establish an energy consumption model for the cloud-edge-device collaborative system. First, establish an energy consumption model for the cloud computing center, obtaining its energy consumption parameters based on the center's configuration information and operational status. Then, establish an energy consumption model for the edge cluster, obtaining its energy consumption parameters based on the processing power, storage resources, and network connectivity of the edge nodes. Finally, establish an energy consumption model for the communication network, obtaining its energy consumption parameters based on the network topology and data transmission conditions. The energy consumption model is established using a linear regression method, with system energy consumption as the objective function, encompassing computing resource consumption, storage resource consumption, and communication resource consumption.

[0048] Step 102: Construct a system resource database. Collect real-time resource status information from the cloud-edge-device system via API interfaces, including metrics such as CPU utilization, memory utilization, storage space utilization, and network bandwidth utilization. Store the collected resource information in a MySQL database to establish a real-time resource monitoring mechanism.

[0049] Step 103: Initialize the task queue. Establish a task queue data structure to store information about tasks to be processed. Each task information includes fields such as task type, priority, creation time, data volume, and processing time limit.

[0050] Step 2: Real-time monitoring of system resource status and acquisition of task request information.

[0051] Step 201: Collect local resource status information through edge computing nodes. Edge computing nodes run on each device using an agent, responsible for collecting local resource information and reporting it to the central management platform. The collected information includes CPU utilization, memory utilization, network bandwidth usage, and disk I / O performance.

[0052] Step 202: Receive task requests from mobile devices. Receive task requests from mobile devices via the RESTful API interface, and parse information such as task type (AI calculation, data processing, business logic), priority (high priority, medium priority, low priority), creation time, data volume, and processing time limit.

[0053] Step 203: Store task information in the task queue. Classify tasks according to their priority and processing time limit, and place them in the corresponding task queues to await scheduling.

[0054] Step 3: Calculate the task scheduling strategy based on the deep reinforcement learning algorithm.

[0055] Step 301: Based on historical task data, train the task scheduling model using a deep reinforcement learning algorithm. Collect task execution data from the past week, including task type, processing nodes, processing results, processing time, etc., to construct a training sample set. Use a deep neural network for feature extraction and learning, and train the model using the Adam optimization algorithm.

[0056] Step 302: Extract features and classify tasks using machine learning algorithms. The PCA algorithm is used to extract features from the task data, retaining 80% of the information. Tasks are then classified based on the feature values, including data processing tasks, AI computing tasks, and business logic tasks.

[0057] Step 303: Calculate the fitness score for each task based on the system resource status and task characteristics. Using the trained scheduling model, calculate the fitness score for each task under the current system state, including factors such as resource utilization, processing efficiency, and energy consumption.

[0058] Step 304: Optimize the task scheduling strategy using a deep reinforcement learning algorithm based on an attention mechanism. Different weights are assigned to different tasks using the attention mechanism, and the optimal scheduling strategy is calculated based on the current system state and task fitness scores.

[0059] Step 4: Execute task scheduling and monitor whether the system status has reached the optimization goal.

[0060] Step 401: Allocate system resources according to the scheduling strategy. Based on the task scheduling results, allocate the corresponding computing resources, storage resources, and network bandwidth resources.

[0061] Step 402: Process local tasks via edge computing nodes or forward tasks to appropriate cloud-edge processing. Based on the task type and processing capacity, determine the most suitable processing node to assign the task. Use various transmission protocols (such as SFTP, HTTP, MQTT, etc.) for data transmission.

[0062] Step 403: Monitor whether the system's energy consumption and resource utilization have reached the optimization goals. Calculate the system's energy consumption indicators in real time, including overall energy consumption, CPU energy consumption, memory energy consumption, and network energy consumption. Monitor resource utilization indicators, including overall resource utilization, CPU utilization, and memory utilization. Determine whether the preset optimization goals have been achieved, such as a 20% reduction in energy consumption or a 30% increase in resource utilization.

[0063] Step 404: If the goal is not achieved, adjust the scheduling strategy and return to step 2; if the goal is achieved, proceed to step 5.

[0064] Step 5: Update system parameters and end the current scheduling cycle.

[0065] Step 501: Analyze system operation data based on machine learning algorithms. Collect operation data for the current scheduling cycle, including resource usage, task processing status, network status, etc. Use the PCA algorithm to extract features from the data and retain key information.

[0066] Step 502: Update system parameters based on analysis results. Based on the data analysis results, optimize and adjust system parameters, such as adjusting task queue priority rules and resource allocation ratios.

[0067] Step 503: Optimize the task scheduling order in the task queue. Based on the updated parameter rules, reorder the task queue to optimize the task scheduling order.

[0068] Step 504: Proceed to the next scheduling cycle. Apply the updated parameters and scheduling rules to the next round of scheduling, then return to step 2 to continue execution. Example

[0069] This invention discloses a low-energy experimental equipment scheduling method based on cloud-edge-device architecture, applicable to a medical laboratory high-speed centrifuge and robotic arm linkage system. The core logic involves introducing a Directed Acyclic Graph (DAG) for perception. Machine learning is used to predict the completion time of the task chain. If the load on edge nodes fluctuates, a "pre-scheduling" mechanism is used to pre-activate hot standby resources in the cloud to ensure uninterrupted experiments. The result is a reduction in the task deadline violation rate to below 0.1%.

[0070] The specific implementation steps are as follows: Step 1: Initialize system parameters and establish a cloud-edge-device collaborative system model.

[0071] Step 101: Establish an energy consumption model for the cloud-edge-device collaborative system. First, establish an energy consumption model for the cloud computing center, obtaining its energy consumption parameters, including CPU energy consumption, memory energy consumption, and network energy consumption, based on the center's configuration information and operational status. Then, establish an energy consumption model for the edge cluster, obtaining its energy consumption parameters based on the edge nodes' processing power, storage resources, and network connectivity. Finally, establish a communication network energy consumption model, obtaining its energy consumption parameters based on the network topology and data transmission conditions. The energy consumption model is established using a linear regression method, with system energy consumption as the objective function, encompassing computing resource consumption, storage resource consumption, and communication resource consumption.

[0072] Step 102: Construct a system resource database. Collect real-time resource status information from the cloud-edge-device system via the Python API interface, including metrics such as CPU utilization, memory utilization, storage space utilization, and network bandwidth utilization. Store the collected resource information in a MongoDB database to establish a real-time resource monitoring mechanism.

[0073] Step 103: Initialize the task queue. Establish a task queue data structure to store information about tasks to be processed. Each task information includes fields such as task type, priority, creation time, data volume, and processing time limit.

[0074] Step 2: Real-time monitoring of system resource status and acquisition of task request information.

[0075] Step 201: Collect local resource status information through edge computing nodes. Edge computing nodes run on each device using an agent, responsible for collecting local resource information and reporting it to the central management platform. The collected information includes CPU utilization, memory utilization, network bandwidth usage, and disk I / O performance.

[0076] Step 202: Receive task requests from mobile devices. Receive task requests from mobile devices via the RESTful API interface, and parse information such as task type (AI calculation, data processing, business logic), priority (high priority, medium priority, low priority), creation time, data volume, and processing time limit.

[0077] Step 203: Store task information in the task queue. Classify tasks according to their priority and processing time limit, and place them in the corresponding task queues to await scheduling.

[0078] Step 3: Calculate the task scheduling strategy based on the deep reinforcement learning algorithm.

[0079] Step 301: Based on historical task data, train the task scheduling model using a deep reinforcement learning algorithm. Collect task execution data from the past month, including task type, processing nodes, processing results, processing time, etc., to construct a training sample set. Use a deep convolutional neural network for feature extraction and learning, and employ the SGD optimization algorithm to train the model.

[0080] Step 302: Extract features and classify tasks using machine learning algorithms. The PCA algorithm is used to extract features from the task data, retaining 90% of the information. Tasks are then classified based on the feature values, including data processing tasks, AI computing tasks, and business logic tasks.

[0081] Step 303: Calculate the fitness score for each task based on the system resource status and task characteristics. Using the trained scheduling model, calculate the fitness score for each task under the current system state, including factors such as resource utilization, processing efficiency, and energy consumption.

[0082] Step 304: Optimize the task scheduling strategy using a deep reinforcement learning algorithm based on an attention mechanism. Different weights are assigned to different tasks using the attention mechanism, and the optimal scheduling strategy is calculated based on the current system state and task fitness scores.

[0083] Step 4: Execute task scheduling and monitor whether the system status has reached the optimization goal.

[0084] Step 401: Allocate system resources according to the scheduling strategy. Based on the task scheduling results, allocate the corresponding computing resources, storage resources, and network bandwidth resources.

[0085] Step 402: Process local tasks via edge computing nodes or forward tasks to appropriate cloud-edge processing. Based on the task type and processing capacity, determine the most suitable processing node to assign the task. Use various transmission protocols (such as SFTP, HTTP, MQTT, etc.) for data transmission.

[0086] Step 403: Monitor whether the system's energy consumption and resource utilization have reached the optimization goals. Calculate the system's energy consumption indicators in real time, including overall energy consumption, CPU energy consumption, memory energy consumption, and network energy consumption. Monitor resource utilization indicators, including overall resource utilization, CPU utilization, and memory utilization. Determine whether the preset optimization goals have been achieved, such as a 30% reduction in energy consumption or a 40% increase in resource utilization.

[0087] Step 404: If the goal is not achieved, adjust the scheduling strategy and return to step 2; if the goal is achieved, proceed to step 5.

[0088] Step 5: Update system parameters and end the current scheduling cycle.

[0089] Step 501: Analyze system operation data based on machine learning algorithms. Collect operation data for the current scheduling cycle, including resource usage, task processing status, network status, etc. Use the PCA algorithm to extract features from the data and retain key information.

[0090] Step 502: Update system parameters based on analysis results. Based on the data analysis results, optimize and adjust system parameters, such as adjusting task queue priority rules and resource allocation ratios.

[0091] Step 503: Optimize the task scheduling order in the task queue. Based on the updated parameter rules, reorder the task queue to optimize the task scheduling order.

[0092] Step 504: Proceed to the next scheduling cycle. Apply the updated parameters and scheduling rules to the next round of scheduling, then return to step 2 to continue execution.

[0093] In summary, by employing the technical solutions described above, this invention introduces a deep reinforcement learning algorithm with a self-attention mechanism to perceive high-dimensional states in real time and dynamically weight bottleneck features, thereby improving the timeliness of resource perception and scheduling flexibility. By transforming a refined energy consumption model into a reinforcement learning penalty function with the goal of minimizing energy consumption, the system can adaptively search for the load balancing path with the lowest global energy consumption while satisfying latency constraints, overcoming the problem of poor energy consumption control in traditional solutions. For the strong temporal constraints and physical state dependencies of experimental equipment, a pre-scheduling mechanism locks resources, and atomicity protection prevents critical tasks from migrating midway, thus ensuring experimental continuity and data integrity. Simultaneously, by performing feature extraction and data cleaning on the edge, communication energy consumption and cloud load are significantly reduced, solving the bandwidth limitations and high latency defects of traditional methods, ultimately maximizing the utilization of heterogeneous computing resources and minimizing overall system energy consumption.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A low-energy experimental equipment scheduling method based on cloud-edge-device architecture, characterized in that, Includes the following steps: S1 initializes system parameters and establishes a cloud-edge-device collaborative system model, which includes an energy consumption model; S2 senses the system resource status in real time and obtains task request information from experimental equipment; S3 calculates a task scheduling strategy based on a deep reinforcement learning algorithm. In the process of calculating the task scheduling strategy, a self-attention mechanism is introduced to identify and weight key features in the state information, and the system energy consumption model is transformed into a reward signal for reinforcement learning, with energy minimization as one of the optimization objectives. S4 performs task scheduling and monitors whether the system status has reached the optimization goal; during the scheduling process, it performs special optimizations based on the characteristics of the experimental equipment tasks, and the special optimizations include at least one or more of the following: a pre-scheduling mechanism based on task dependencies, atomicity protection for high-precision real-time monitoring tasks, and data compression and feature extraction on the edge side. S5 updates system parameters and ends the current scheduling cycle.

2. The low-energy experimental equipment scheduling method based on cloud-edge-device according to claim 1, characterized in that, Step S1 specifically includes the following steps: S101 establishes an energy consumption model for the cloud-edge-device collaborative system, including an energy consumption model for the cloud computing center, an energy consumption model for the edge cluster, and an energy consumption model for the communication network. S102 constructs a system resource database to record the real-time status information of various resources; S103 initializes the task queue and stores information about tasks to be processed.

3. The low-energy experimental equipment scheduling method based on cloud-edge-device according to claim 1, characterized in that, Step S2 specifically includes the following steps: S201 collects local resource status information through edge computing nodes; S202 receives a task request from the mobile device and parses the task type and priority; S203 stores the task information in the task queue and waits for scheduling.

4. The low-energy experimental equipment scheduling method based on cloud-edge-device according to claim 1, characterized in that, Step S3 specifically includes the following steps: S301 uses a deep reinforcement learning network with an Actor-Critic architecture to train a task scheduling model based on historical task data. S302 uses machine learning algorithms to extract features and classify tasks; S303 calculates the fitness score for each task based on the system resource status and task characteristics; S304 employs a deep reinforcement learning algorithm based on an attention mechanism. The input includes the real-time load of the experimental equipment, the task pipeline dependency graph, and the energy consumption model parameters. It uses an attention mask to filter out invalid environmental noise and outputs the optimal node and power configuration for task offloading.

5. The low-energy experimental equipment scheduling method based on cloud-edge-device according to claim 4, characterized in that, The reward function of the deep reinforcement learning network in step 301 is designed as a negative weighted cost function: R = -(ω1.E total +ω2.D latency +ω2.I violation ), where ω1, ω2, and ω3 represent weighting coefficients, which are dynamically adjusted according to the experiment type; E total D represents total energy consumption; latency Indicates the total delay in task completion; I violation This is a mandatory penalty.

6. The low-energy experimental equipment scheduling method based on cloud-edge-device according to claim 1, characterized in that, The specific optimizations in step S4 include: The pre-scheduling mechanism is as follows: when an experimental task with a dependency relationship is detected, the computing resources required by the subsequent task are pre-locked through the edge computing node during the execution of the preceding task. The atomicity protection is as follows: a non-interruptible flag is set for tasks identified as high-precision real-time monitoring tasks, prohibiting the tasks from migrating across nodes during execution; The data compression and feature extraction at the endpoint are as follows: a lightweight machine learning model is used at the endpoint to downsample or detect anomalies in the raw data generated by the experimental equipment, and only the data containing changes with features is uploaded.

7. The low-energy experimental equipment scheduling method based on cloud-edge-device according to claim 1, characterized in that, Step S4 further includes: S401 allocates system resources according to the scheduling policy; The S402 performs experimental task assignment. For tasks identified as critical physics experiments, it activates the atomicity protection protocol to ensure that the tasks are completed continuously within the predetermined energy consumption envelope and provides real-time feedback on the actual power consumption. To determine whether the energy consumption and resource utilization of the S403 monitoring system have reached the optimization target, a dynamic weighted sum method is used. The weights of processing latency and system energy consumption are automatically adjusted according to the experimental stage. Under the condition of meeting the basic QoS threshold, the solution with the lowest energy consumption is searched. If the target is not achieved, adjust the scheduling strategy and return to step 2; if the target is achieved, proceed to step 5.

8. The low-energy experimental equipment scheduling method based on cloud-edge-device according to claim 1, characterized in that, Step S5 further includes: S501 analyzes system operation data based on machine learning algorithms, including principal component analysis algorithm for dimensionality reduction of high-dimensional experimental environment data, and / or K-Means clustering algorithm for automatic classification of experimental tasks. S502 updates system parameters based on analysis results; S503 optimizes the task scheduling order in the task queue; S504 enters the next scheduling cycle.