A heterogeneous network task scheduling method, device, medium and system

Abstract

This invention discloses a method, device, medium, and system for task scheduling in heterogeneous networks. The method includes: acquiring current operating parameters and constructing a physical device attribute graph; extracting features using a first-path graph attention network to obtain a dimensionality-reduced global device state matrix; parsing flexible manufacturing instructions and converting each subtask into a logical task dependency graph; performing aggregation processing using a second-path graph attention network to output a global task state matrix; concatenating and fusing the global device state matrix and the global task state matrix to generate an environmental state feature vector; inputting this vector into a near-end optimization algorithm model to output a binary mapping matrix from subtasks to computing nodes; and generating control signals based on the binary mapping matrix and sending them to physical devices for execution. This application can accurately filter high-frequency noise interference in the network, reduce the dimensionality of the reinforcement learning state space, and ensure millisecond-level low-latency real-time optimal task scheduling under dynamic changes in heterogeneous network topology.

Description

Technical Field

[0001] This application relates to the field of G06N technology, and specifically to a heterogeneous network task scheduling method, device, medium and system. Background Technology

[0002] In the context of flexible manufacturing within the Industrial Internet of Things (IIoT), production lines consist of heterogeneous nodes such as fixed sensing devices, mobile automated guided vehicles (AGVs), and robotic arms. These heterogeneous nodes exhibit significant physical differences in computing power, communication protocol types, and available link bandwidth. Flexible manufacturing tasks typically need to be decomposed into multiple sub-tasks with a strict execution sequence, such as image acquisition, feature extraction, trajectory planning, and motion execution.

[0003] Currently, existing task scheduling methods mainly employ mixed-integer linear programming algorithms or conventional heuristic algorithms for task allocation. When mobile devices cause dynamic changes in network topology, or when on-site communication links experience latency fluctuations due to electromagnetic interference, heuristic algorithms need to recalculate the global iteratively. The complexity of such calculations increases exponentially with the number of nodes and tasks, causing the solution time to typically exceed the millisecond-level latency threshold required by industrial control. Simultaneously, traditional deep learning models struggle to directly process non-Euclidean space data and cannot accurately represent the physical topological connections between heterogeneous device nodes and the logical dependencies between various manufacturing task nodes, easily leading to scheduling deadlocks or extremely low task execution efficiency. Summary of the Invention

[0004] The technical problem solved by this application is to provide a heterogeneous network task scheduling method that can output the optimal task allocation method with low latency when the heterogeneous node network structure changes dynamically.

[0005] To address the aforementioned technical problems, this application provides a heterogeneous network task scheduling method applied in an industrial Internet of Things (IIoT) edge computing environment. The method includes: acquiring the current operating parameters of each computing node and network communication link in the IIoT, and constructing a physical device attribute graph based on the operating parameters; extracting node features and edge features from the physical device attribute graph using a first-path graph attention network to calculate a dimensionality-reduced global device state matrix; acquiring and parsing flexible manufacturing instructions to generate a directed acyclic graph task set containing multiple sub-tasks, converting the computational and storage requirements of each sub-task and their temporal dependencies into a logical task dependency graph; aggregating the logical task dependency graph using a second-path graph attention network to output a global task state matrix; concatenating and fusing the global device state matrix and the global task state matrix through a fully connected network layer to generate an environment state feature vector; and inputting the environment state feature vector into a proximal optimization algorithm model, preferably, the proximal optimization algorithm model is Proximal Policy Optimization (PPO). The model uses minimizing the weighted sum of completion time and total system energy consumption as the reward objective, and outputs a binary mapping matrix from each subtask in the directed acyclic graph task set to the computing node; it generates a scheduling control signal based on the binary mapping matrix and sends the scheduling control signal to the corresponding physical device to execute the corresponding subtask.

[0006] Furthermore, the specific process of obtaining the current operating parameters of each computing node and network communication link in the Industrial Internet of Things, and constructing a physical device attribute graph based on the operating parameters, includes: extracting the processor operating frequency and available storage capacity of each computing node as the initial node features of the physical device attribute graph; extracting the theoretical network bandwidth and actual communication latency of the network communication links between each computing node as the initial edge features of the physical device attribute graph; and mapping each initial node feature and each initial edge feature to the physical device attribute graph using a graph topology structure.

[0007] Further, the specific process of extracting node features and edge features from the physical device attribute graph using the first path graph attention network and calculating the dimensionality-reduced global device state matrix includes: inputting the mapped initial node features and initial edge features into the feature transformation matrix built into the first path graph attention network for linear mapping operations to obtain the query vector and key vector of each computing node; determining the attention weight allocation coefficient between adjacent nodes by calculating the dot product of the query vector and the key vector, combined with a nonlinear activation function; filtering out network signal fluctuation components greater than a preset frequency threshold by calculating the product of the attention weight allocation coefficient of each node in the physical device attribute graph and a preset smoothing attenuation factor, and outputting the global device state matrix based on the weighted summation processing result.

[0008] Furthermore, the specific process of converting the computational and storage requirements of each subtask and their temporal dependencies into a logical task dependency graph includes: using the floating-point operation count requirement and maximum memory usage of each subtask in the directed acyclic graph task set as the logical node features of the logical task dependency graph; and using the amount of preceding data output and the directed flow and transmission relationship generated after the preceding subtask is completed as the temporal association edge features of the logical task dependency graph.

[0009] Furthermore, the specific process of using the second path graph attention network to aggregate the logical task dependency graph and output the global task state matrix includes: extracting the logical node features and temporal correlation edge features of the current level node layer by layer according to the internal topological flow hierarchy of the logical task dependency graph; inputting the extracted feature data of each dimension into the multi-head attention mechanism processing layer equipped in the second path graph attention network to perform multi-dimensional orthogonal projection feature transformation; calculating the message flow transfer weight matrix between the preceding subtask and the following subtask in multiple independent attention feature subspaces respectively, and performing side-by-side concatenation and layer normalization operations on the feature vectors output from different attention feature subspaces to generate the global task state matrix used to characterize the global instruction flow correlation dependency relationship.

[0010] Furthermore, the specific process of inputting the environmental state feature vector into the proximal optimization algorithm model to minimize the weighted sum of completion time and total system energy consumption as the reward objective includes: in the internal reward constraint function of the proximal optimization algorithm model, setting the weight of the completion time penalty term, which characterizes the total completion time of the scheduling scheme, to 0.75, and setting the weight of the system energy consumption penalty term, which characterizes the total power consumption of the physical equipment execution process, to 0.25; calculating the numerical product of the obtained completion time penalty term and the weight of the completion time penalty term, and performing a linear superposition calculation operation with the numerical product of the obtained system energy consumption penalty term and the weight of the system energy consumption penalty term, to obtain the real-time reward scalar value parameter under the current environmental state.

[0011] Furthermore, the specific process of outputting the binary mapping matrix from each subtask in the directed acyclic graph task set to the computing node includes: using the fused environmental state feature vector as the reconstructed environmental state parameter S corresponding to time t. t The input is fed into the evaluation network execution unit inside the near-end optimization algorithm model; the allocation probability distribution parameters of each subtask to all selectable computing nodes are calculated through the multilayer perceptron; a two-dimensional allocation decision matrix structure with a size defined as N×M is sampled from the allocation probability distribution parameters; the corresponding matrix element values ​​of the selected mapping nodes in the two-dimensional allocation decision matrix structure are reset to logical values ​​1, and the corresponding matrix element values ​​of the other non-selected nodes are reset to logical values ​​0, thereby generating the binary mapping matrix, where parameter N represents the total number of subtasks to be scheduled in the current computing cycle, and parameter M represents the total number of computing nodes that are alive and available in the current network environment.

[0012] This application also provides an electronic device, including: a memory unit for storing computer-readable program code instructions; and a processor unit for establishing a physical communication connection with the memory unit and for calling and executing the computer-readable program code instructions stored in the memory unit to implement the execution steps included in the aforementioned heterogeneous network task scheduling method applied in the edge computing environment of the Industrial Internet of Things.

[0013] This application also provides a non-volatile computer-readable storage medium that stores computer program executable file code. When the computer program executable file code is read and run by a physical computing device configured with computing capabilities, the physical computing device configured with computing capabilities implements the aforementioned heterogeneous network task scheduling method applied in the edge computing environment of the Industrial Internet of Things.

[0014] This application also provides a heterogeneous network task scheduling system applied in an industrial IoT edge computing environment, comprising: an environment attribute mapping module, used to obtain the current operating parameters of each computing node and network communication link in the industrial IoT, and construct a physical device attribute graph based on the operating parameters; a first state aggregation module, used to extract node features and edge features in the physical device attribute graph using a first path graph attention network, and calculate a dimensionality-reduced global device state matrix; a task dependency conversion module, used to obtain and parse flexible manufacturing instructions to generate a directed acyclic graph task set containing multiple sub-tasks, and convert the computational and storage requirements of each sub-task and their temporal dependencies into a logical task dependency graph; and a second state aggregation module, used to... The system utilizes a second-path graph attention network to aggregate the logical task dependency graph and output a global task state matrix. A spatial state stitching module is used to stitch and fuse the global device state matrix and the global task state matrix through a fully connected network layer to generate an environmental state feature vector. An optimized scheduling mapping module inputs the environmental state feature vector into a proximal optimization algorithm model, using the minimization of the weighted sum of completion time and total system energy consumption as the reward objective, and outputs a binary mapping matrix from each subtask in the directed acyclic graph task set to the computing node. A hardware execution dispatch module generates a scheduling control signal based on the binary mapping matrix and dispatches the scheduling control signal to the corresponding physical device to execute the corresponding subtask.

[0015] In summary, this application achieves smooth dimensionality reduction of the high-dimensional state vectors of heterogeneous physical nodes and logical tasks through a dual-path graph attention network, filtering out high-frequency noise interference, and inputting the extracted topological association features into the near-end optimization algorithm model to calculate the allocation matrix. Through the synergistic cooperation of these features and mechanisms, the dimensionality of the state space faced by reinforcement learning is reduced, overcoming the model's non-convergence defect under a large action space, and ensuring millisecond-level real-time scheduling capability under dynamic topology changes. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a heterogeneous network task scheduling method applied in an industrial Internet of Things edge computing environment, as provided in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the process for obtaining parameters and constructing a global device state matrix provided in an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of the process for generating a task set and outputting a global task status matrix provided in an embodiment of the present invention.

[0019] Figure 4This is a schematic diagram of the output mapping matrix based on the near-end optimization algorithm model provided in an embodiment of the present invention.

[0020] Figure 5 This is a structural block diagram of a heterogeneous network task scheduling system applied in an industrial Internet of Things edge computing environment, provided by an embodiment of the present invention. Detailed Implementation

[0021] The technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1

[0023] Figure 1 This is a flowchart illustrating a heterogeneous network task scheduling method applied in an industrial IoT edge computing environment, as provided in this embodiment of the invention. To support the operation of this scheduling method, the overall task scheduling system physical architecture consists of a field terminal layer, an edge gateway layer, and a multi-access edge computing scheduling server layer. The field terminal layer includes specific devices such as industrial cameras, automated guided vehicle controllers, and six-axis robotic arm controllers. Field terminal layer devices connect to the edge gateway layer via network interfaces. The access devices in the edge gateway layer specifically employ time-sensitive network switches supporting the IEEE 802.1Qbv network protocol standard. The multi-access edge computing scheduling server is equipped with a multi-access edge computing core scheduling chip as the main control computing power core and is directly connected to the edge gateway layer via optical fiber. The system's underlying layer also includes a network interface driver module that supports the parsing of the underlying transmission protocol between the time-sensitive network switch and the multi-access edge computing server to ensure correct packet transmission. The edge gateway layer is responsible for collecting the liveness status, processor load rate, and available memory capacity of the field terminal layer devices, and for detecting the link bandwidth and data packet loss rate between devices through underlying heartbeat packets. It then packages and sends this information to the multi-access edge computing scheduling server according to a set data packet sending period of 20 milliseconds.

[0024] Combination Figure 1 As shown, the method in this embodiment executes S101, which involves obtaining operating parameters and constructing a physical device attribute map. Specifically, it obtains the current operating parameters of each computing node and network communication link in the industrial Internet of Things (IIoT) in S201.

[0025] refer to Figure 2 The system specifically constructs and defines the physical device attribute map. Among them, V D This represents a set of nodes in the device graph, referring to various heterogeneous physical computing devices; E DThis represents the set of edges in the device graph, indicating the network communication links between devices. The system extracts the current operating frequency of the processor and the remaining available storage capacity of each computing node as the initial node features of the physical device attribute graph (S202). Simultaneously, the system extracts the theoretical network bandwidth of the network communication links between each computing node and the actual communication latency obtained from heartbeat packet detection as the initial edge features of the physical device attribute graph (S203). The system further maps each initial node feature and each initial edge feature to the physical device attribute graph using a graph topology structure (S204).

[0026] The method then performs feature extraction to calculate the dimensionality-reduced global device state matrix S102, and uses the first path graph attention network to extract node features and edge features in the physical device attribute graph.

[0027] Specifically, the system inputs the mapped initial node features and initial edge features into the feature transformation matrix built into the first path graph attention network for linear mapping operation S205, obtaining the query vector and key vector of each computation node S206. The system then calculates the dot product of the query vector and the key vector, and combines this with a non-linear activation function to determine the attention weight allocation coefficients between adjacent nodes S207. In a specific embodiment, the first path graph attention network uses the following first path graph attention weight calculation formula to obtain the attention score parameter:

[0028]

[0029] Among them, e ij h represents the unnormalized attention score between nodes. i with h j x represents the device node state feature vector of adjacent device node i and device node j in the device graph node set, respectively; ij This represents the link feature vector corresponding to the communication association between adjacent nodes; and || represents the learnable weight matrix that is continuously updated during the training phase of the neural network; || represents performing a matrix concatenation operation on features of different dimensions; Represents the built-in mapping weight row vector; This represents the nonlinear activation function with leakage correction introduced by the system.

[0030] After obtaining the score, the system performs attention coefficient normalization calculation according to the following formula:

[0031]

[0032] in, This represents the attention weight allocation coefficients after normalization. This represents the normalization exponent processing function.

[0033] Furthermore, the system filters out network signal fluctuation components S208 that exceed a preset frequency threshold by calculating the product of the attention weight allocation coefficient of each node in the physical device attribute graph and a preset smoothing attenuation factor. Finally, it outputs the dimensionality-reduced global device state matrix H based on the weighted summation processing result. D S209. Through the calculation of attention weights and the smoothing filtering mechanism, the bottleneck nodes in the communication topology that are most sensitive to latency can be dynamically captured, and high-frequency noise caused by minor environmental interference can be effectively filtered out, thereby providing highly stable low-dimensional representation data for subsequent reinforcement learning.

[0034] The method executes to generate a task set and transforms the logical task dependency graph S103, and obtains and parses flexible manufacturing instructions to generate a directed acyclic graph task set containing multiple sub-tasks.

[0035] like Figure 3 As shown, after receiving the flexible manufacturing instruction, the multi-access edge computing scheduling server parses the flexible manufacturing instruction S1 and generates a directed acyclic graph task set S2 containing multiple subtasks. Subsequently, the computational and storage requirements of each subtask, as well as their temporal dependencies, are converted into a logical task dependency graph S3. This logical graph is defined as... In the logical task dependency graph, V T This represents a set of nodes in the task graph, representing the individual subtasks that need to be executed; E T This represents the set of edges in the task time-series dependency graph, signifying the logical constraints that each subtask must follow regarding execution order. The system uses the floating-point operation requirements and maximum memory usage of each subtask in the directed acyclic graph task set as logical node features. The system also inputs the preceding data output after the completion of the preceding subtask and the directed flow and transmission relationships as time-series association edge features into the dependency graph.

[0036] The method then performs aggregation processing to output the global task state matrix S104, and uses the second path graph attention network to perform aggregation processing on the logical task dependency graph.

[0037] Specifically, the system extracts the logical node features and temporal correlation edge features of the current level nodes layer by layer according to the topological flow hierarchy within the logical task dependency graph, completing the layer-by-layer feature extraction S4. The system inputs the extracted feature data of each dimension into the multi-head attention mechanism S5, i.e., the multi-head attention mechanism processing layer equipped within the second-path graph attention network, to perform multi-dimensional orthogonal projection feature transformation S6. The system calculates the transfer weight matrix S7, i.e., calculates the message flow transfer weight matrix between the preceding and following subtasks in multiple independent attention feature subspaces, and performs side-by-side concatenation and execution layer normalization operations on the feature vectors output from different attention feature subspaces S8, thereby generating a high-dimensional compressed vector representing the global instruction flow correlation dependency relationship, i.e., outputting the global task state matrix H. T S9.

[0038] The method then performs a fusion to generate an environment state feature vector S105, which is then fused with the global device state matrix and the global task state matrix through a fully connected network layer.

[0039] The system uses the data channel of the fully connected layer to concatenate and fuse the global device state matrix and the global task state matrix in terms of vector dimension, thereby obtaining a unified environmental state feature vector node1.

[0040] The method executes the input algorithm model output binary mapping matrix S106, and inputs the environmental state feature vector into the proximal optimization algorithm model, with the goal of minimizing the weighted sum of completion time and total system energy consumption as the reward objective.

[0041] like Figure 4 As shown, the system employs a proximal optimization algorithm model (node2) with an action evaluator architecture to complete the decision output. Internally, the system also includes a dataset caching module (node8) for reinforcement learning training and a corresponding value network backpropagation structure (node9) configured in parallel to assist the backbone network in optimizing weight parameters. The system uses the pre-fused environment state feature vector (node1) as the proximal policy optimization reinforcement learning environment state parameter S corresponding to time t. t The input is fed into the evaluation network execution unit node3 within the near-end optimization algorithm model node2. The system calculates the allocation probability distribution parameters (node5) of each subtask to all available computing nodes through the multilayer perceptron processing network node4. The system then samples and outputs the corresponding action set A from the allocation probability distribution parameters. t .

[0042] The action set is represented by a two-dimensional allocation decision matrix structure with dimensions N×M, defined as the binary mapping matrix node6. In this structure, parameter N represents the total number of subtasks to be scheduled in the current computation cycle, and parameter M represents the total number of live and available computing nodes in the current network environment. The system uses the binary mapping matrix element determination logic: the corresponding matrix element value of the selected mapping node in the two-dimensional allocation decision matrix structure is reset to the logical value 1; if the element value is 1, it means that the corresponding task is assigned to that device; at the same time, the corresponding matrix element values ​​of the remaining non-selected nodes are reset to the logical value 0.

[0043] In this process, the proximal optimization algorithm model guides the evolution of the reinforcement learning model through the calculation formula of the system reward value parameter node7. Specifically, the mathematical calculation logic of the system reward value parameter in the proximal optimization algorithm model is as follows:

[0044]

[0045] Where R represents the system reward value parameter node7, which the system approaches the optimal solution by maximizing; T makespan E represents the time when the last subtask in a directed acyclic graph task set completes, i.e., the task set completion time penalty term. total This parameter represents the total energy consumption of all devices in the current scheduling cycle, i.e., the total energy consumption of all devices. The completion time weight parameter is represented in this embodiment by setting parameters. Highlighting sensitivity to latency; This represents the energy consumption weighting parameter, and in this embodiment, parameters are set simultaneously. To constrain the total heat generation of the equipment, the product of the obtained completion time penalty term and its weight is linearly superimposed with the product of the obtained system energy consumption penalty term and its weight, and then inverted to obtain the real-time reward scalar value parameter under the current environmental condition. It should be noted that 0.75 and 0.25 are only preferred balancing parameters for this application; those skilled in the art can adjust them to other allocation ratios such as 0.6 and 0.4 based on different site conditions to implement the main logic of this application.

[0046] The method generates control signals and sends them out for execution S107. It generates scheduling control signals according to the binary mapping matrix and sends the scheduling control signals to the corresponding physical devices to execute the corresponding sub-tasks.

[0047] The system iterates through the coordinates of all non-zero matrix elements in the binary mapping matrix node6. The system encapsulates the data payload, containing the device execution instructions for the subtask and the specified communication address of the sending computing node, into a network data transmission packet in the format of the underlying industrial communication protocol and executes the physical channel transmission operation. Through this step, the multi-access edge computing scheduling server directly drives the physical execution of the field terminal layer devices across layers, ultimately completing closed-loop scheduling.

[0048] Example 2

[0049] Combination Figure 5 As shown, this embodiment of the invention also provides a heterogeneous network task scheduling system applied in an industrial Internet of Things (IIoT) edge computing environment. This system comprises multiple modules internally interconnected via electrical interfaces and software buses.

[0050] Specifically, the system includes an environment attribute mapping module 501, configured to acquire the current operating parameters of each computing node and network communication link in the Industrial Internet of Things (IIoT), and construct a physical device attribute graph based on the operating parameters. A first state aggregation module 502 is configured to extract node and edge features from the physical device attribute graph using a first-path graph attention network, and calculate a dimensionality-reduced global device state matrix. A task dependency transformation module 503 is configured to acquire and parse flexible manufacturing instructions to generate a directed acyclic graph task set containing multiple sub-tasks, and convert the computational and storage requirements of each sub-task, as well as their temporal dependencies, into a logical task dependency graph. A second state aggregation module 504 is configured to aggregate the logical task dependency graph using a second-path graph attention network, and output a global task state matrix. A spatial state stitching module 505 is configured to stitch and fuse the global device state matrix and the global task state matrix through a fully connected network layer to generate an environment state feature vector. The optimization scheduling mapping module 506 is configured to input the environmental state feature vector into the near-end optimization algorithm model, using the minimization of the weighted sum of completion time and total system energy consumption as the reward objective, and output a binary mapping matrix from each subtask in the directed acyclic graph task set to the computing node. The hardware execution dispatch module 507 is configured to generate scheduling control signals based on the binary mapping matrix and dispatch the scheduling control signals to the corresponding physical devices to execute the respective subtasks.

[0051] The specific implementations of the corresponding process mechanisms in the aforementioned method embodiments of each module are not described in detail here.

[0052] Example 3

[0053] This embodiment provides an electronic device. The electronic device physically manifests as an industrial computing terminal server. The electronic device includes a memory unit and a processor unit. The memory unit is configured to store computer-readable program code instructions, and the memory unit may specifically be a solid-state drive or high-frequency dynamic random access memory. The processor unit establishes a physical communication connection with the memory unit and is configured to call and execute the computer-readable program code instructions stored within the memory unit. When the processor unit executes the instructions, it implements the various execution steps included in the heterogeneous network task scheduling method applied to the edge computing environment of the Industrial Internet of Things.

[0054] Example 4

[0055] This embodiment provides a non-volatile computer-readable storage medium. The non-volatile computer-readable storage medium internally stores computer program executable file code. When the computer program executable file code is read and executed by a computing device configured with computing capabilities, the computing device can losslessly reproduce and implement the various scheduling method calculation logics of the heterogeneous network task scheduling method applied in the edge computing environment of the Industrial Internet of Things.

[0056] Based on the above scheme, this application not only achieves dual-path dimensionality reduction of the high-dimensional matrix of on-site physical equipment and logical task attributes, but also accurately maps the timing constraints and network latency with industrial site characteristics into the policy optimization algorithm. By eliminating the drastic fluctuations in decision weights caused by link noise, the allocation time of complex tasks in a large action space is ultimately controlled at the tens of millisecond level, achieving low-latency collaborative scheduling.

Claims

1. A heterogeneous network task scheduling method applied in an industrial Internet of Things (IIoT) edge computing environment, characterized in that, include: Obtain the current operating parameters of each computing node and network communication link in the Industrial Internet of Things, and construct a physical device attribute map based on the operating parameters; The first path graph attention network is used to extract the node features and edge features of the physical device attribute graph, and the dimension-reduced global device state matrix is ​​calculated. The flexible manufacturing instructions are acquired and parsed to generate a directed acyclic graph task set containing multiple subtasks. The computational and storage requirements of each subtask and their temporal dependencies are converted into a logical task dependency graph. The logical task dependency graph is aggregated using a second-path graph attention network to output a global task state matrix; The global device state matrix and the global task state matrix are spliced ​​and fused by a fully connected network layer to generate an environmental state feature vector. The environmental state feature vector is input into the near-end optimization algorithm model, with the goal of minimizing the weighted sum of completion time and total system energy consumption, and the binary mapping matrix from each subtask in the directed acyclic graph task set to the computing node is output. A scheduling control signal is generated based on the binary mapping matrix, and the scheduling control signal is sent to the corresponding physical device to execute the corresponding sub-task.

2. The heterogeneous network task scheduling method applied to the edge computing environment of the Industrial Internet of Things as described in claim 1, characterized in that, Obtain the current operating parameters of each computing node and network communication link in the Industrial Internet of Things (IIoT), and construct a physical device attribute graph based on the operating parameters, including: Extract the processor operating frequency and available storage capacity of each computing node as the initial node features of the physical device attribute graph; The theoretical network bandwidth and actual communication delay time of the network communication links between each computing node are extracted as the initial edge features of the physical device attribute graph. The initial node features and initial edge features are mapped to the physical device attribute graph using a graph topology structure.

3. The heterogeneous network task scheduling method applied to the edge computing environment of the Industrial Internet of Things as described in claim 2, characterized in that, The first-path graph attention network is used to extract node features and edge features from the physical device attribute graph, and the dimensionality-reduced global device state matrix is ​​calculated, including: The mapped initial node features and initial edge features are respectively input into the feature transformation matrix built into the first road graph attention network for linear mapping operation to obtain the query vector and key vector of each computing node. The attention weight allocation coefficients between adjacent nodes are determined by calculating the dot product of the query vector and the key vector, combined with a nonlinear activation function. By calculating the product of the attention weight allocation coefficient of each node in the physical device attribute graph and the preset smoothing attenuation factor, network signal fluctuation components greater than the preset frequency threshold are filtered out, and the global device state matrix is ​​output based on the weighted summation processing result.

4. The heterogeneous network task scheduling method applied to the edge computing environment of the Industrial Internet of Things as described in claim 1, characterized in that, The computational and storage requirements of each subtask, as well as their temporal dependencies, are converted into a logical task dependency graph, including: The floating-point operation requirements and maximum memory usage of each subtask in the directed acyclic graph task set are used as the logical node features of the logical task dependency graph. The output of preceding data and the directed flow and transmission relationship generated after the preceding subtasks are completed are used as the temporal correlation edge features of the logical task dependency graph.

5. The heterogeneous network task scheduling method applied to the edge computing environment of the Industrial Internet of Things as described in claim 4, characterized in that, The logical task dependency graph is aggregated using a second-path graph attention network to output a global task state matrix, including: According to the internal topological flow hierarchy of the logical task dependency graph, extract the logical node features of the current level node and the temporal association edge features of the connected edge structure layer by layer; The extracted feature data of each dimension are input into the multi-head attention mechanism processing layer equipped inside the second path graph attention network to perform multi-dimensional orthogonal projection feature transformation; In multiple independent attention feature subspaces, the message flow transfer weight matrix between the preceding and following subtasks is calculated respectively. The feature vectors output from different attention feature subspaces are then concatenated side-by-side and subjected to layer normalization to generate the global task state matrix used to characterize the global instruction flow correlation and dependency relationship.

6. The heterogeneous network task scheduling method applied to the edge computing environment of the Industrial Internet of Things as described in claim 1, characterized in that, The environmental state feature vector is input into the proximal optimization algorithm model, with the reward objective being to minimize the weighted sum of completion time and total system energy consumption, including: In the internal reward constraint function of the near-end optimization algorithm model, the weight of the completion time penalty term, which characterizes the total completion time of the scheduling scheme, is set to 0.75, and the weight of the system energy consumption penalty term, which characterizes the total power consumption of the physical equipment during execution, is set to 0.

25. The product of the obtained completion time penalty and its weight is calculated, and then linearly superimposed with the product of the obtained system energy consumption penalty and its weight to obtain the real-time reward scalar value parameter under the current environmental state.

7. The heterogeneous network task scheduling method applied to the edge computing environment of the Industrial Internet of Things as described in claim 6, characterized in that, Output the binary mapping matrix from each subtask in the directed acyclic graph task set to the computing node, including: The fused environmental state feature vector is used as the reconstructed environmental state parameter S corresponding to time t. t The input is fed into the evaluation network execution unit within the proximal optimization algorithm model; The allocation probability distribution parameters of each subtask to all selectable computing nodes are calculated using a multilayer perceptual processing network. A two-dimensional allocation decision matrix structure with a size defined as N×M is obtained by sampling from the allocation probability distribution parameters. The corresponding matrix element values ​​of the selected mapping nodes in the two-dimensional allocation decision matrix structure are reset to logical values ​​1, and the corresponding matrix element values ​​of the other non-selected nodes are reset to logical values ​​0, thereby generating the binary mapping matrix. The parameter N represents the total number of the subtasks to be scheduled in the current computing cycle, and the parameter M represents the total number of the computing nodes that are alive and available in the current network environment.

8. An electronic device, characterized in that, include: Memory units are used to store computer-readable program code instructions; The processor unit establishes a physical communication connection with the memory unit and is used to call and execute the computer-readable program code instructions stored inside the memory unit to implement the execution steps included in the heterogeneous network task scheduling method applied to the edge computing environment of industrial Internet of Things as described in claim 1.

9. A non-volatile computer-readable storage medium, characterized in that, The non-volatile computer-readable storage medium internally stores computer program executable file code. When the computer program executable file code is read and run by a physical computing device configured with computing capabilities, the physical computing device configured with computing capabilities implements the heterogeneous network task scheduling method applied to the edge computing environment of the Industrial Internet of Things as described in claim 1.

10. A heterogeneous network task scheduling system applied in an industrial Internet of Things (IoT) edge computing environment, characterized in that, include: The environment attribute mapping module is used to obtain the current operating parameters of each computing node and network communication link in the industrial Internet of Things, and to construct a physical device attribute map based on the operating parameters. The first state aggregation module is used to extract node features and edge features from the physical device attribute graph using the first path graph attention network, and calculate the dimensionality-reduced global device state matrix. The task dependency conversion module is used to acquire and parse flexible manufacturing instructions to generate a directed acyclic graph task set containing multiple sub-tasks, and convert the computational and storage requirements of each sub-task and their temporal dependencies into a logical task dependency graph. The second state aggregation module is used to aggregate the logical task dependency graph using the second path graph attention network and output a global task state matrix. The spatial state stitching module is used to stitch and fuse the global device state matrix and the global task state matrix through a fully connected network layer to generate an environmental state feature vector. The optimized scheduling mapping module is used to input the environmental state feature vector into the near-end optimization algorithm model, with the goal of minimizing the weighted sum of completion time and total system energy consumption, and outputs the binary mapping matrix from each subtask in the directed acyclic graph task set to the computing node. The hardware execution dispatch module is used to generate a scheduling control signal based on the binary mapping matrix, and dispatch the scheduling control signal to the corresponding physical device to execute the corresponding sub-task.

Images