Source network load storage heterogeneous device multi-protocol adaptive access method and device based on edge control

Abstract

The invention belongs to the field of power internet of things access control, and particularly relates to a source network load storage heterogeneous device multi-protocol self-adaptive access method and device based on edge control, and the method comprises the steps: carrying out the multi-protocol self-adaptive recognition and access of a source network load storage heterogeneous device through an intelligent gateway as an edge node; task state information of the accessed mobile equipment is collected; based on task state information, a calculation unloading decision is made, a decision instruction is generated, tasks of the mobile equipment are correspondingly distributed according to the decision instruction, in addition, a global optimization strategy and resource prices are provided through interaction of a control plane, a management plane and an infrastructure plane, and task execution and resource distribution can be dynamically coordinated. According to the invention, efficient access and unified management of multi-type and multi-protocol equipment are realized through the intelligent sensing and scheduling capability of the edge node.

Description

Technical Field

[0001] This invention belongs to the field of power Internet of Things access control, specifically relating to a multi-protocol adaptive access method and device for heterogeneous source-grid-load-storage devices based on edge control. Background Technology

[0002] As power systems accelerate their transformation towards digitalization, distributed computing, and intelligence, various types of equipment, including distributed photovoltaic systems, energy storage systems, electric vehicle charging facilities, and smart electricity terminals, are being widely deployed across diverse scenarios. To achieve flexible access, unified management, and intelligent dispatch, power systems are placing higher demands on the intelligence, adaptability, and compatibility of equipment access systems. However, existing equipment access solutions are generally based on static configuration and manual management, relying primarily on manual intervention for communication protocol adaptation, data format mapping, and resource scheduling strategy settings. This makes it difficult to cope with operating environments characterized by multiple equipment types, multiple communication protocols, and dynamic changes. Especially in the context of widespread distributed deployment and edge computing, traditional centralized access systems struggle to meet practical needs in terms of latency, flexibility, and scalability. Issues such as inconsistent communication protocols, incompatible access methods, and diverse data structures among various heterogeneous devices severely restrict rapid system deployment and efficient operation. Furthermore, while edge nodes typically possess certain computing and control capabilities, existing systems have not fully explored their potential in areas such as equipment access decisions, communication method selection, and edge-cloud collaborative dispatch. Summary of the Invention

[0003] (a) Purpose of the invention The purpose of this invention is to provide a method and apparatus for multi-protocol adaptive access of heterogeneous source-grid-load-storage devices based on edge control, which aims to achieve efficient access and unified management of multiple types and multi-protocol devices through the intelligent sensing and scheduling capabilities of edge nodes.

[0004] (II) Technical Solution To address the aforementioned problems, a first aspect of the present invention provides a multi-protocol adaptive access method for heterogeneous source-grid-load-storage devices based on edge control, comprising the following steps: S1: Through the intelligent gateway acting as an edge node, multi-protocol adaptive identification and access are performed on the heterogeneous source-grid-load-storage devices, and task status information of the accessed mobile devices is collected. S2: Based on the task status information, a deep reinforcement learning model deployed on the edge node is used to calculate the unloading decision and generate a decision instruction to instruct the task to be executed locally on the mobile device or unloaded to a specified edge node. S3: According to the decision instruction, the task of the mobile device is assigned to the local computing queue or transmission queue of the mobile device. For the task assigned to the transmission queue, it is forwarded to the edge node or cloud computing service provider specified in step S2 through the smart gateway for processing. S4: The control plane collects resource status and user subscription information from the infrastructure plane, and receives resource pricing strategies and computing request strategies formulated by the management plane based on the resource status and user subscription information; the control plane distributes the resource pricing strategies and computing request strategies to edge computing service providers and cloud computing service providers in the infrastructure plane to dynamically coordinate task execution and resource allocation in step S3.

[0005] Furthermore, in step S1, the step of "performing multi-protocol adaptive identification and access for the heterogeneous source-grid-load-storage devices" specifically includes: The intelligent gateway analyzes the communication characteristics of the heterogeneous source-grid-load-storage devices and automatically identifies their communication protocols. Based on the identified communication protocol type, the corresponding protocol adapter is dynamically invoked to complete data format conversion and communication connection.

[0006] Furthermore, in step S1, the task status information includes: task size, processing density, task deadline, mobile device battery level, and the length of the mobile device's local computing queue and transmission queue.

[0007] Further, in step S2, the deep reinforcement learning model is a DQN-LSTM hybrid network structure, and its decision-making process includes: The task state information is used as the input state of the deep reinforcement learning model; The evaluation network of the deep reinforcement learning model calculates the Q-values ​​of different unloading actions under the current input state. Select the action corresponding to the highest Q value to generate the decision instruction.

[0008] Furthermore, step S3 specifically includes: S31: If the decision instruction indicates that the task is to be executed locally on the mobile device, then the task is assigned to the local computing queue of the mobile device and processed in a first-in-first-out order; S32: If the decision instruction indicates that the task needs to be unloaded, the task is assigned to the transmission queue of the mobile device and forwarded to the edge node or cloud computing service provider specified in the decision instruction via the wireless interface.

[0009] Furthermore, when the decision instruction directs the task to be unloaded to the edge node specified in the decision instruction, the method further includes: The designated edge node maintains an independent first-in-first-out task queue for each associated mobile device to receive unload tasks from that mobile device; By adopting a general processor sharing approach, the computing power of edge nodes is dynamically allocated to all their active first-in-first-out task queues to process tasks from multiple mobile devices in parallel.

[0010] Furthermore, step S4 specifically includes: S41: The control plane collects user subscription information and local computing capabilities of each computing service provider by requesting the analysis controller; S42: Based on the user subscription information and the local computing power, the management plane generates the optimal resource pricing strategy and computing request strategy through the resource pricing module and the request strategy module; S43: The control plane, through the resource allocation controller, distributes the resource pricing strategy and the computing request strategy to the cloud computing service provider and the smart gateway acting as the edge computing service provider.

[0011] Furthermore, the method also includes: S5: Based on the feedback of the task execution results in step S3, update the parameters of the deep reinforcement learning model used in step S2.

[0012] Furthermore, in step S5, the iterative update of the deep reinforcement learning model is completed at the edge node side, specifically including: The states, actions, service quality, and next states generated during task execution are stored as experience data in the experience replay buffer of the deep reinforcement learning model. Periodically sample data from the experience replay buffer and calculate the target Q value; The parameters of the evaluation network are updated by minimizing the difference between the output of the evaluation network and the target Q-value, and the updated parameters are periodically synchronized to the target network of the deep reinforcement learning model.

[0013] A second aspect of the present invention provides a multi-protocol adaptive access device for heterogeneous source-grid-load-storage devices based on edge control, the device comprising: The protocol identification and access module is used to perform multi-protocol adaptive identification and access of the source-grid-load-storage heterogeneous device through a smart gateway that acts as an edge node; The first information acquisition module is used to collect the task status information of the connected mobile device; The decision module is used to calculate the unloading decision based on the task status information by using a deep reinforcement learning model deployed on the edge node, and generate a decision instruction to instruct the task to be executed locally on the mobile device or unloaded to a specified edge node. The task allocation module is used to allocate tasks of the mobile device to the local computing queue or transmission queue of the mobile device according to the decision instruction. For tasks allocated to the transmission queue, the smart gateway forwards them to a designated edge node or cloud computing service provider for processing. The second information acquisition module is used to collect resource status and user subscription information from the infrastructure plane through the control plane, and to receive resource pricing strategies and calculation request strategies formulated by the management plane based on the resource status and user subscription information. The distribution module enables the control plane to distribute the resource pricing strategy and computing request strategy to edge computing service providers and cloud computing service providers in the infrastructure plane, so as to dynamically coordinate task execution and resource allocation.

[0014] (III) Beneficial Effects The above-mentioned technical solution of the present invention has the following beneficial technical effects: The present invention provides a multi-protocol adaptive access method and device for heterogeneous source-network-load-storage devices based on edge control. This method supports multi-protocol identification and adaptation, has adaptive access mode capabilities, and can realize a unified device access system method for edge intelligent collaboration, thereby improving the flexibility of device access, the dynamics of resource allocation, and the overall intelligence level of the network. This method constructs a heterogeneous access collaboration framework that integrates edge computing and cloud computing. Through the collaborative cooperation of the software-defined networking (SDN) control plane, management plane, and infrastructure plane, it realizes the dynamic distribution of device status awareness, resource requests, and scheduling strategies, and introduces distributed... The deep reinforcement learning algorithm optimizes task offloading and resource allocation decisions. Specifically, it first collects task status information from connected mobile devices via intelligent gateways at edge nodes; then, a deep reinforcement learning model deployed on the edge nodes performs computation offloading decisions, generating decision instructions to instruct tasks to be executed locally on the mobile device or offloaded to a designated edge node; based on these instructions, the tasks on the mobile device are allocated to its local computation or transmission queue; finally, the resource pricing and computation request strategies formulated by the management plane are distributed to edge computing service providers and cloud computing service providers in the infrastructure plane via the control plane, dynamically coordinating task execution and resource allocation. This method leverages the protocol parsing, data processing, and access decision-making capabilities of the edge gateway to effectively identify device communication characteristics and dynamically adapt access methods, avoiding deployment delays and configuration complexity issues associated with traditional manual configuration. By introducing adaptive access strategies and communication method selection mechanisms, it can intelligently select the optimal access path and resource allocation strategy based on network load, device priority, and real-time business requirements, improving system operating efficiency and responsiveness. Meanwhile, this method possesses excellent protocol scalability and structural portability, making it suitable for large-scale, dynamically changing multi-source access scenarios. It performs particularly well in practical applications requiring rapid deployment, flexible access, or edge-cloud collaborative task processing. By constructing a unified and open method for heterogeneous device access and edge scheduling, this application can significantly improve the intelligence level, operational stability, and resource scheduling efficiency of device access systems, supporting the digital and ubiquitous interconnection development needs of future energy systems. Attached Figure Description

[0015] Figure 1 This is a flowchart of the multi-protocol adaptive access method for heterogeneous source-grid-load-storage devices based on edge control according to the present invention; Figure 2 This is a schematic diagram of the multi-protocol adaptive access device for heterogeneous source-grid-load-storage devices based on edge control according to the present invention; Figure 3 This is a schematic diagram of the three-layer system structure according to a specific embodiment of the present invention; Figure 4This is a schematic diagram of the three-layer system connection according to a specific embodiment of the present invention; Figure 5 This is a data transmission flowchart of a three-layer system structure according to a specific embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0017] like Figure 1 As shown, the first aspect of the present invention provides a multi-protocol adaptive access method for heterogeneous source-grid-load-storage devices based on edge control, comprising the following steps: S1: Through the intelligent gateway acting as an edge node, multi-protocol adaptive identification and access are performed on the heterogeneous source-grid-load-storage devices, and task status information of the accessed mobile devices is collected; the task status information includes: task size, processing density, task deadline, power level of the mobile device, and length of the local computing queue and transmission queue of the mobile device. S2: Based on the task status information, a deep reinforcement learning model deployed on the edge node is used to calculate the unloading decision and generate a decision instruction to instruct the task to be executed locally on the mobile device or unloaded to a specified edge node. S3: According to the decision instruction, the task of the mobile device is assigned to the local computing queue or transmission queue of the mobile device. For the task assigned to the transmission queue, it is forwarded to the edge node or cloud computing service provider specified in step S2 through the smart gateway for processing. S4: The control plane collects resource status and user subscription information from the infrastructure plane and receives resource pricing and computation request policies formulated by the management plane based on the resource status and user subscription information. The control plane distributes the resource pricing and computation request policies to edge computing service providers and cloud computing service providers in the infrastructure plane to dynamically coordinate task execution and resource allocation in step S3. This application introduces a three-layer system architecture, including an infrastructure plane, a control plane, and a management plane. The three are connected through wired or wireless communication modules to achieve cross-layer collaboration of computing resources and policies. The infrastructure plane consists of multiple mobile devices (MDs) and multiple edge nodes (ENs). The control plane includes a request analysis controller, an access controller, and a resource allocation controller. The management plane includes a resource pricing module and a request policy module. Figure 4As shown, the infrastructure plane and control plane are connected via wireless communication links, while the control plane and management plane are connected via wired connections. The management plane can indirectly issue policies to the infrastructure plane through the control plane. Specifically, a layered game-theoretic resource management mechanism based on SDN is introduced in the control plane; dynamic pricing and policy coordination are introduced in the management plane. The infrastructure plane includes a system model, a communication model, and a computational model. In the system model, an MD–EN task queue one-to-one mapping structure is established, where MD represents mobile devices and EN represents edge nodes. In the communication model, an intelligent gateway is designed to achieve multi-protocol adaptive access. In the computational model, distributed deep reinforcement learning is used to optimize offloading and resource allocation.

[0018] Furthermore, in step S1, the multi-protocol adaptive identification and access of the heterogeneous source-grid-load-storage device specifically includes: S11, the intelligent gateway analyzes the communication characteristics of the heterogeneous source-grid-load-storage device and automatically identifies its communication protocol; S12, based on the identified communication protocol type, dynamically call the corresponding protocol adapter to complete data format conversion and communication connection.

[0019] Further, in step S2, the deep reinforcement learning model is a DQN-LSTM hybrid network structure, and its decision-making process includes: S21, the task state information is used as the input state of the deep reinforcement learning model; S22, calculate the Q-values ​​of different unloading actions under the current input state through the evaluation network of the deep reinforcement learning model; S23, select the action corresponding to the highest Q value and generate the decision instruction.

[0020] Furthermore, step S3 specifically includes: S31: If the decision instruction indicates that the task is to be executed locally on the mobile device, then the task is assigned to the local computing queue of the mobile device and processed in a first-in-first-out order; S32: If the decision instruction indicates that the task needs to be unloaded, the task is assigned to the transmission queue of the mobile device and forwarded to the edge node or cloud computing service provider specified in the decision instruction via the wireless interface. In this step, when the decision instruction indicates that the task should be unloaded to the edge node specified in the decision instruction, the following steps are also included: S321, The specified edge node maintains an independent first-in-first-out task queue for each associated mobile device to receive unloading tasks from that mobile device; The S322 employs a general processor sharing approach, dynamically allocating the computing power of edge nodes to all their active first-in-first-out task queues to process tasks from multiple mobile devices in parallel.

[0021] Furthermore, step S4 specifically includes: S41: The control plane collects user subscription information and local computing capabilities of each computing service provider by requesting the analysis controller; S42: Based on the user subscription information and the local computing power, the management plane generates the optimal resource pricing strategy and computing request strategy through the resource pricing module and the request strategy module; S43: The control plane, through the resource allocation controller, distributes the resource pricing strategy and the computing request strategy to the cloud computing service provider and the smart gateway acting as the edge computing service provider.

[0022] Furthermore, the method also includes: S5: Based on the feedback from the task execution results in step S3, update the parameters of the deep reinforcement learning model used in step S2. In step S5, the iterative update of the deep reinforcement learning model is completed at the edge nodes, specifically including: S51, the states, actions, service quality and the next state generated during the task execution process are stored as experience data in the experience replay buffer of the deep reinforcement learning model. S52, periodically sample data from the experience replay buffer and calculate the target Q value; S53, by minimizing the difference between the output of the evaluation network and the target Q value, the parameters of the evaluation network are updated, and the updated parameters are periodically synchronized to the target network of the deep reinforcement learning model.

[0023] The following is combined with Figure 3 The specific application of the three control planes in this application is explained as follows: 1) Control Plane SDN-based architecture separates computing resource management from the infrastructure, forming a cloud computing resource market based on hierarchical game theory in the control plane. In this architecture, the smart gateway, as a key component of edge computing, connects mobile devices to the cloud. Through data exchange between the infrastructure plane and the control plane, it enables the collection of user subscription information and the distribution of policies by computing service providers (CPs) in the infrastructure plane, as detailed below: 1. Information Collection: Through the request analytics controller, the control plane collects user subscription information across different Content Providers (CPs) and the local computing power of each CP via the smart gateway, and sends the received information to the management plane in real time. Specifically, the request analytics and access control controller can communicate with cloud computing service providers (CCPs) and edge computing service providers (ECPs) through access points, and then select the time-varying number of users for the appropriate CP.

[0024] 2. Policy Distribution: The control plane receives pricing and request policies for cloud computing resources formulated by the upper-layer management plane, and then distributes these policies to the CCP and ECP through the resource allocation controller. The intelligent gateway, as part of the ECP, is responsible for enforcing these policies. In addition, the access controller and core controller are responsible for cloud resource sharing and access control between edge gateways and cloud gateways, i.e., setting access network paths between gateways and managing compute offloading, etc.

[0025] II) Management Plan After receiving user subscription and computing power information from each Content Provider (CP), the optimal pricing and request strategy for cloud resources is determined in the management plane. During this process, the access capabilities and protocol adaptability of the intelligent gateway are taken into account to optimize resource allocation and task scheduling. Specifically, the management layer assists the CCP in making dynamic resource pricing decisions and helps the ECP determine how much computing power it should request. These decisions are then fed back to the control plane to guide cloud resource sharing between the CCP and ECP. Given the wired connections between the SDN controller and the edge and cloud gateways, and the powerful processing capabilities of the SDN server, the above service response can be achieved quickly.

[0026] III) Infrastructure Plan The infrastructure plane includes a system model, a communication model, and a computing model. This application introduces these three models because task offloading and resource allocation involve the entire process from task generation to transmission and execution, requiring hierarchical modeling. The system model includes mobile devices (MD), edge nodes (EN), their task queue structure, and time-slotted arrival processes, representing the device and resource status of the infrastructure layer. The communication model defines the wireless link rate, latency, and energy consumption between the MD and EN, representing the transmission process of the infrastructure layer. The computing model defines the processing method and energy consumption of tasks on the MD or EN side, also representing the execution process of the infrastructure layer. By establishing these three models, precise analysis of the characteristics of different stages can be performed, providing a fine-grained mathematical foundation for subsequent optimization. Specifically: the system model provides the task generation and queue scheduling mechanism, determining whether a task enters the local computing queue or the transmission queue; the communication model handles the transmission queue, providing performance indicators for tasks during transmission from the device to the edge node; and the computing model determines the execution method of tasks on the local CPU or the edge node processor. Each model is described below: 1. System Model This paper studies a MEC system consisting of a set of mobile devices (MDs), denoted as . And a MEC system consisting of a set of ENs, denoted as The smart gateway, as a specific implementation of edge nodes (ENs), supports multi-protocol adaptive access, ensuring that different types of MDs can seamlessly connect to the edge server. and These represent the number of MDs and ENs, respectively. This application considers time as a series of... A specific episode in a time slot is denoted as each Each time slot represents The duration is seconds. This method considers two independent queues for each MD to organize tasks for local processing or to be sent to ENs, and operates in a First-In-First-Out (FIFO) manner. The MD's scheduler is responsible for assigning newly arriving tasks to each queue at the start of the time slot. On the other hand, this method assumes that each EN consists of I FIFO queues, where each queue corresponds to one MD.

[0027] definition To reach MD The computational task in the time slot The above index value makes This indicates the size of the task, in bits. The size of a discrete set Selected from, among which This indicates the number of these values. Therefore, Let's consider the case where no tasks arrive. This method also denotes the task processing density as... , This represents the number of CPU cycles required to complete the execution of one task unit. Furthermore, this method uses... This indicates the deadline for the task. This refers to the number of time slots that must be completed to prevent the task from being discarded.

[0028] This method defines two binary variables. and , These are used to determine the uninstallation decision and the uninstallation target, respectively. Indicates task It is assigned to the computation queue. Or be assigned to the transmission queue , Indicates task Uninstall to If the task is assigned to ENj, then let ;otherwise .

[0029] 2. Communication model This application allocates tasks in the transmission queue to appropriate ENs via the MD wireless interface. The intelligent gateway acts as the access point, responsible for protocol conversion and data forwarding to ensure efficient task transmission. Indicators such as transmission rate, latency, and energy consumption are calculated to determine task allocation or discard. MD and The interface transmission rate during communication, in time slots In the middle, if the task Assigned to the transmission queue for computation offloading, defined To indicate whether a task is assigned to an EN slot or discarded, another definition is needed. For the task The number of time slots that need to be waited for before transmission in the queue. It's worth noting that MD... Calculations are needed before making a decision. The value, The value is calculated as follows: (15) in, , This indicates the historical time slot index considered when calculating the maximum latency. ,when hour, The value depends only on ,if In the time slot Medium scheduling tasks Then it is in the time slot It may be scheduled or discarded.

[0030] (1) in Refers to the task From MD Transmit to EN The number of time slots required. Indicates the deadline for the task. The calculation method is as follows: (2) in, Indicates task Uninstall to , This indicates a situation where no task has been completed. MD and The interface transmission rate during communication. express arrive The energy consumption of transmission is: (3) in express The communication link in the time slot Power consumption.

[0031] 3. Computational Model The computation task can be executed either locally on MD or on EN, as detailed below: 1) Local Execution: Local execution is modeled as a queuing system consisting of a computation queue and a MD processor. This indicates the MDi's processing capacity (one cycle per second). When a task... In the time slot When the initial computation queue is assigned, this method is defined. For the task The time slots that are processed or discarded. If the computation queue is empty, then... .make This indicates the tasks being processed in the computation queue. The remaining time slots from before: (4) In the above equation, Indicates in time slot The time slot in which each existing task in the previously arrived computation queue is processed or discarded. Therefore, Indicates task The number of time slots that need to be waited before processing. This method will handle the task. exist If a task is assigned to a computation queue for local processing at a given time, then... Will be If fully processed, then: (5) If the task In the time slot If processing is not completed before the end, the data is immediately discarded. Furthermore, this method introduces... As a task to be completed on MD The number of time slots required for processing is calculated using the following formula: (6) in, This indicates the processing power of MDi (one cycle per second). This indicates the computational density of the task, expressed in cycles per bit.

[0032] 2) Edge Execution: This method simulates edge execution by deploying MD-related queues on EN. If the computation task... In time Being assigned to EN, and (In bits) Represent the unique index of the task and EN respectively. First The size of the tasks in each queue. This method defines... (in bits) represents the queue in the time slot The length at the end. This method defines a queue as an active queue within a time slot if it is not empty. That is, a queue is active if at least one task from a previous time slot was already in the queue, or if a task has arrived in the queue. This method defines... Indicates time slot EN The set of active queues.

[0033] (7) Introduction This indicates that EN is in the time slot. Number of internal activity queues. In each time slot. Processing capacity is partitioned within its active queue using a common processor sharing method. Let the variable... (in cycle per second) means EN The computing power. Therefore, EN Can be in time slots Will The computing power is allocated to each To calculate MD In EN The method defines the queue length calculation. (In bits) to represent the queue in time slots The number of bits discarded at the end of the task. The queue backlog, denoted as... The calculation formula is as follows: (8) This method also defines To uninstall the task By EN The time slot for processing or discarding. Given EN Uncertain workload ahead, MD and EN There is nothing about Information until the corresponding task To be processed or discarded. Order Indicates task Mathematically, the time slot for starting execution, according to this method, includes: (9) In the formula In fact, the task The initial processing slot for EN should not precede the slots when tasks are queued or when previously arrived tasks are processed or interrupted. Therefore The time slot that satisfies the following constraints is calculated using the following formula: (10) and (11) In the formula: For the characteristic function, EN From time slot Time slot Assigned to MD The total processing capacity should exceed the task The size. Conversely, from the time slot Time slot The total processing capacity should be less than the size of the task.

[0034] In addition, this method defines To indicate the task In EN The number of processing time slots allocated to it during execution is calculated using the following formula: (12) (13) in It is a constant, representing EN The processor's energy consumption when running at full load. (Except for EN) In addition to the energy consumed in task processing, this method also considers MD. The user interface waits for EN in standby mode. Energy consumed in completing the task. This method defines... To be with MD The energy consumption related to the user interface is given by the following formula: (14) in: For MD Backup energy consumption.

[0035] The QECO algorithm is meticulously designed to optimize the distribution of computational tasks between Method Analyzers (MDs) and Entities (ENs). Since training neural networks incurs significant computational burdens for MDs, this method enables MDs to utilize ENs to train their neural networks, effectively reducing their computational workload. For each MD, there exists an associated EN, denoted as . It assists in the training process. This EN has the highest transmission capacity among all ENs. This method defines... Let be a set of MDs, where training is performed by EN, i.e. Because it requires minimal information exchange and processing during training compared to the task of macroscopic method (MD). Its core idea is to train the neural network using MD experience (i.e., state, action, QoE, next state), mapping Q-values ​​to each state-action pair. This mapping allows MD to identify actions in observed states with the highest Q-value and maximizes its long-term QoE.

[0036] Specifically, the EN node maintains an experience replay buffer. And configure two neural networks with identical structures but independent parameters: evaluate the network and target network However, they each have different parameter vectors. and The Q-value of the network computation is expressed as... The Q-value of the target network is expressed as They are based on the current state. The calculations are performed along with action 'a'. The playback buffer records the observation experience of MDi. .in, Used for action selection. This is used to generate the target Q-value, which represents the estimated long-term QoE achievable with a given action in the current state. The update process uses the target Q-value as a reference, updating the network by minimizing the difference between the Q-values ​​output by the evaluation network and the target network. The parameters.

[0037] Furthermore, the specific functional modules such as the offloading decision module structure, task queue structure, and scheduling path design of the edge nodes are all located in the infrastructure plane, because they run directly on the MD and EN sides for task acquisition, queuing, decision-making, and processing. For example... Figure 5 As shown, these modules connect bidirectionally to the control plane: Uplink, the modules report status information such as task size, queue length, and power level to the control plane; downlink, the control plane generates offloading and scheduling instructions based on the resource pricing and request policies of the management plane, and sends them to the EN for execution. The control plane is also responsible for coordinating resource scheduling and task forwarding between the EN and the cloud. The management plane, through interaction with the control plane, provides global optimization strategies and resource prices, indirectly influencing offloading decisions and task scheduling.

[0038] like Figure 2 As shown, a second aspect of the present invention provides a multi-protocol adaptive access device for heterogeneous source-grid-load-storage devices based on edge control, the device comprising: Protocol identification and access module 21 is used to perform multi-protocol adaptive identification and access of the source-grid-load-storage heterogeneous device through a smart gateway that acts as an edge node; The first information acquisition module 22 is used to collect the task status information of the connected mobile device; The decision module 23 is used to calculate the unloading decision based on the task status information by using a deep reinforcement learning model deployed on the edge node, and generate a decision instruction to instruct the task to be executed locally on the mobile device or unloaded to a specified edge node. The task allocation module 24 is used to allocate the tasks of the mobile device to the local computing queue or transmission queue of the mobile device according to the decision instruction. The tasks allocated to the transmission queue are forwarded to the designated edge node or cloud computing service provider for processing through the smart gateway. The second information acquisition module 25 is used to collect resource status and user subscription information from the infrastructure plane through the control plane, and to receive resource pricing strategies and calculation request strategies formulated by the management plane based on the resource status and user subscription information. Distribution module 26 is used to enable the control plane to distribute the resource pricing strategy and computing request strategy to edge computing service providers and cloud computing service providers in the infrastructure plane, so as to dynamically coordinate task execution and resource allocation.

[0039] The technical solution proposed in this application innovatively constructs a collaborative system for the access and computation offloading of heterogeneous source-network-load-storage devices based on edge intelligent control and SDN management mechanisms. It integrates key technologies such as distributed computation offloading, protocol identification, adaptive access, and edge-cloud collaboration. This method leverages the local decision-making capabilities of edge nodes and the centralized scheduling advantages of the SDN control plane to achieve dynamic identification of multiple types of devices, access path optimization, and real-time task distribution, effectively improving overall resource utilization and response efficiency. By introducing deep reinforcement learning algorithms, terminal devices can autonomously make decisions based on task characteristics and device status, significantly reducing latency and energy consumption while improving offloading accuracy and service quality. Facing complex and variable access environments and task loads, the system exhibits high adaptability and robustness, making it particularly suitable for scenarios with multi-source heterogeneity, high concurrency, and real-time computing requirements. The overall solution not only optimizes the processing flow and saves manual configuration costs but also enhances the system's scalability and deployment flexibility.

[0040] It should be understood that the specific embodiments described above are merely illustrative of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, or improvements made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries. Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it includes the flow of the embodiments of the methods described above. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. The steps in the methods of the embodiments of the present invention can be adjusted, merged, or deleted according to actual needs. The modules [units] in the system [terminal or device] of the embodiments of the present invention can be merged, divided, or deleted according to actual needs.

Claims

1. A method for multi-protocol adaptive access based on edge control of source network load heterogeneous devices, characterized in that, The method comprises the following steps: S1: through the intelligent gateway as an edge node, multi-protocol adaptive identification and access to the source heterogeneous device, and collecting the task state information of the accessed mobile device; S2: based on the task state information, a deep reinforcement learning model deployed on the edge node is used to calculate the offloading decision, and a decision instruction indicating whether the task is executed locally or offloaded to a specified edge node is generated; S3: according to the decision instruction, the task of the mobile device is allocated to the computing queue or the transmission queue of the mobile device, wherein for the task allocated to the transmission queue, the intelligent gateway is used to forward the task to the edge node or the cloud computing service provider specified in step S2 for processing; S4: the control plane collects resource state and user subscription information from the infrastructure plane, and receives resource pricing strategies and computing request strategies formulated based on the resource state and user subscription information from the management plane; the control plane distributes the resource pricing strategies and computing request strategies to the edge computing service provider and the cloud computing service provider in the infrastructure plane to dynamically coordinate the task execution and resource allocation in step S3.

2. The method of claim 1, wherein the method further comprises: In step S1, the multi-protocol adaptive identification and access to the source heterogeneous device specifically comprises: The intelligent gateway analyzes the communication characteristics of the source heterogeneous device to automatically identify the communication protocol; According to the type of the identified communication protocol, the corresponding protocol adapter is dynamically called to complete data format conversion and communication connection. 3.The method of claim 1, wherein, In step S1, the task state information includes: task size, processing density, task deadline, power level of the mobile device, length of the local computing queue and the transmission queue of the mobile device.

4. The method of claim 1, wherein the method further comprises: In step S2, the deep reinforcement learning model is a DQN-LSTM hybrid network structure, and the decision process comprises: The task state information is used as the input state of the deep reinforcement learning model; The evaluation network of the deep reinforcement learning model is used to calculate the Q value of different offloading actions under the current input state; The action corresponding to the highest Q value is selected to generate the decision instruction.

5. The method of claim 1, wherein the method further comprises: Step S3 specifically comprises: S31: if the decision instruction indicates that the task is executed locally, the task is allocated to the local computing queue of the mobile device, and the tasks are processed in the first-in-first-out order; S32: if the decision instruction indicates that the task needs to be offloaded, the task is allocated to the transmission queue of the mobile device, and the task is forwarded to the edge node or the cloud computing service provider specified in the decision instruction through the wireless interface.

6. The method of claim 5, wherein the method further comprises: When the decision instruction indicates that the task is offloaded to the edge node specified in the decision instruction, the method further comprises: The specified edge node maintains an independent first-in-first-out task queue for each associated mobile device, which is used to receive offloaded tasks from the mobile device; A general-purpose processor sharing method is used to dynamically allocate the computing capacity of the edge node to all active first-in-first-out task queues to process tasks from multiple mobile devices in parallel.

7. The method of claim 1, wherein the method further comprises: Step S4 specifically comprises: S41: The control plane collects user subscription information and local computing capacity of each computing service provider through a request analysis controller; S42: The management plane generates optimal resource pricing strategy and computing request strategy through a resource pricing module and a request strategy module based on the user subscription information and the local computing capacity; S43: The control plane issues the resource pricing strategy and the computing request strategy to cloud computing service providers and intelligent gateways as edge computing service providers through a resource allocation controller. 8.The method of claim 4, wherein, The method further comprises: S5: updating parameters of the deep reinforcement learning model used in step S2 based on result feedback of task execution in step S3.

9. The method of claim 8, wherein the method further comprises: In step S5, the iterative updating of the deep reinforcement learning model is completed on the edge node side, specifically including: storing state, action, service quality and next state generated during task execution as experience data in an experience replay buffer of the deep reinforcement learning model; periodically sampling data from the experience replay buffer to calculate target Q value; updating parameters of the evaluation network by minimizing the difference between the output of the evaluation network and the target Q value, and periodically synchronizing the updated parameters to a target network of the deep reinforcement learning model.

10. A source network payload heterogeneous device multi-protocol adaptive access device based on edge control, characterized in that, The device comprises: a protocol identification and access module for multi-protocol adaptive identification and access of the source heterogeneous device through an intelligent gateway as an edge node; a first information collection module for collecting task state information of the accessed mobile device; a decision module for computing offloading decision based on the task state information through a deep reinforcement learning model deployed on the edge node to generate a decision instruction indicating local execution or offloading of the task to a designated edge node; a task allocation module for allocating the task of the mobile device to a computing queue or a transmission queue of the mobile device according to the decision instruction, wherein for the task allocated to the transmission queue, the task is forwarded to a designated edge node or a cloud computing service provider through the intelligent gateway for processing; a second information collection module for collecting resource state and user subscription information from an infrastructure plane through a control plane, and receiving resource pricing strategy and computing request strategy formulated by a management plane based on the resource state and the user subscription information; a distribution module for enabling the control plane to distribute the resource pricing strategy and the computing request strategy to edge computing service providers and cloud computing service providers in the infrastructure plane to dynamically coordinate task execution and resource allocation.

Images