Vehicle DNN task fusion unloading method based on graph neural network

Abstract

The invention is suitable for the technical field of Internet of Vehicles and edge computing, and provides a vehicle DNN task fusion unloading method based on a graph neural network, and the method comprises the steps: enabling a vehicle to generate a DNN task at a terminal, determining an optimal segmentation point through an MLP model, and dividing the DNN task into a vehicle end sub-graph and an edge sub-graph; the RSU constructs a dynamic heterogeneous graph, and vehicle embedding vectors are extracted through GNN; the vehicle outputs unloading and fusion actions based on the PPO intelligent agent, and feature fusion in the group is supported; the RSU adopts convex optimization and greedy scheduling, resource allocation and task sequence are optimized in a combined manner, and end-to-end delay and energy consumption of the system are minimized. According to the method, the vehicle-side computing resources can be efficiently utilized, the task scheduling performance and the system operation efficiency are improved, and the energy consumption and the communication load of repeated computing are reduced.

Description

Technical Field

[0001] This application belongs to the field of vehicle networking and edge computing technology, and in particular relates to a method for vehicle DNN task fusion and offloading based on graph neural networks. Background Technology

[0002] The rapid development of vehicle-to-everything (V2X) technology has transformed vehicles into mobile intelligent terminals capable of generating and processing massive amounts of data to support deep neural network tasks such as autonomous driving and path planning. However, vehicles themselves have limited computing and storage resources, and processing complex DNN tasks locally can easily lead to high latency and high energy consumption. To address this challenge, edge computing has emerged. By offloading computing tasks from vehicles to nearby roadside units and other edge servers, response latency can be significantly reduced and processing efficiency improved. Currently, vehicle-side collaborative task offloading methods typically make decisions based on the vehicle's real-time communication conditions, computing load, or task priority, employing heuristics or traditional optimization algorithms to allocate computing resources, aiming to achieve load balancing and latency optimization.

[0003] However, current vehicle-side collaborative task offloading methods suffer from problems such as uneven resource allocation, redundant computation due to ignoring the similarity of features between tasks, lack of dynamic adaptability of fixed offloading strategies, and decreased scheduling efficiency under conditions of high-speed vehicle movement and frequent changes in network topology. Summary of the Invention

[0004] This application provides a vehicle DNN task fusion and offloading method based on graph neural networks, which can solve the problems of uneven resource allocation, repeated calculations caused by ignoring the similarity of features between tasks, lack of dynamic adaptability of fixed offloading strategies, and decreased scheduling efficiency under high-speed vehicle movement and frequent changes in network topology in current vehicle-side collaborative task offloading methods.

[0005] In a first aspect, embodiments of this application provide a vehicle DNN task fusion and offloading method based on graph neural networks, including: S1, when the vehicle terminal generates a DNN task, using an MLP model to pre-analyze the network structure of the DNN task and determine the optimal segmentation point. According to the optimal segmentation point The DNN task is divided into vehicle side view sub-images and edge side view sub-images; the attributes of the DNN task are uploaded to the RSU, including task description, task deadline, and optimal split point. S1. The computational resources, storage resources, and communication resources required for the DNN task; S2. The vehicle terminal periodically collects its own status: location. Speed Signal-to-noise ratio And form a state vector The state vector is sent to the RSU, which uses the attributes of the DNN task and the state vector to construct a dynamic heterogeneous graph. Vehicle nodes are clustered, and GNN forward propagation is performed to generate a vehicle embedding representation, which is then sent back to the vehicle terminal. S3: The vehicle terminal inputs the vehicle embedding representation into a PPO-based Actor module and outputs two action parameters: unloading action. 0 represents local uninstallation, 1 represents uninstallation from another vehicle within the group, 2 represents single-hop RSU uninstallation, and 3 represents multi-hop RSU uninstallation; and fusion layer actions. , and when When the vehicle initiates feature fusion at the corresponding layer; S4, when the output action parameter is a fusion layer action, the target node establishes a fusion pool, receives and aggregates the intermediate feature maps of vehicles within the group within a set waiting time slot, generates fused features, and distributes them to each vehicle to continue executing subsequent layer inference; S5, when the output action parameter is an unloading action, the vehicle performs the corresponding unloading operation according to the action type: when Local uninstallation is performed at that time; when When the unloading is performed on another vehicle within the group; Alternatively, single-hop RSU offloading and multi-hop RSU offloading can be performed at time 3. When the latency within the group exceeds a preset threshold, the task is transferred from the group to the RSU for continued execution. S6. After the task is transferred to the RSU, the RSU optimizes the resource quota using the block coordinate descent method (BCD convex approximation) based on the global queue, the remaining computing power of the node, and the link lifetime, and determines the task execution order by combining the earliest deadline first algorithm (EDF). By alternately optimizing the RSU resource quota and the fusion pool waiting time, the joint minimization of end-to-end latency is achieved.

[0006] In one possible implementation of the first aspect, the optimal segmentation point in step S1 above... The specific determination method is as follows:

[0007]

[0008] in, This indicates the execution time for calculating the vehicle profile submap locally. This represents the inference time of the edge subgraph after it has been transmitted over the network;

[0009] Will and The minimum value of the sum corresponds to the layer, which is determined as the optimal split point. .

[0010] Optionally, in another possible implementation of the first aspect, the dynamic heterogeneous graph constructed in step S2 above is represented as ,in For the set of vehicle nodes, This refers to the set of connection relationships between vehicles or between a vehicle and an RSU. The node feature matrix contains location, vehicle speed, signal-to-noise ratio, and task attributes; the update rule for performing GNN forward propagation is as follows:

[0011]

[0012] in, Indicates that node i is at the 1st rank. Layer embedding representation, Let i represent the set of neighbors of node i. This represents the learnable weight matrix of the l-th layer. This represents the learnable bias vector of the l-th layer.

[0013] Optionally, in another possible implementation of the first aspect, step S2 above employs a dynamic clustering method based on embedding similarity for clustering vehicle nodes, as follows:

[0014]

[0015] in, For the m-th group, For the embedding of node i, For the first Cluster centers This is the similarity threshold.

[0016] Optionally, in another possible implementation of the first aspect, step S4 above, when the output action parameter is a fusion layer action, involves the target node establishing a fusion pool, receiving and aggregating intermediate feature maps of vehicles within the group within a set waiting time slot, generating fused features, and then distributing them to each vehicle to continue executing subsequent layer inference. Specifically, this includes:

[0017] The fusion pool receives and aggregates vehicle sets within the group. Intermediate feature map:

[0018]

[0019] in, Indicates vehicle In the Intermediate features of the layer output, This represents a fusion operator, taking the form of a weighted average or attention aggregation.

[0020] The fusion layer corresponding to the fusion pool pairs the intermediate feature maps. Processing is performed to generate fusion features. And by using task identifiers or weight mapping to fuse features Perform feature splitting, and then fuse the split features. Distribute to the computing nodes corresponding to each task, where The feature inverse mapping function ensures that each task continues forward inference of the remaining layers without relying on the original vehicle backhaul.

[0021] For layers that are not involved in the fusion, the current node completes the forward propagation independently.

[0022] Optionally, in another possible implementation of the first aspect, in the PPO-based Actor module of step S3 above, the loss function of PPO is defined as:

[0023]

[0024] in, For strategy ratio, For advantage function estimation, clip(*) is the policy gradient function. Update the step size limit for the strategy.

[0025] Optionally, in another possible implementation of the first aspect, step S6 above uses BCD convex approximation to optimize resource quotas and combines it with EDF to determine the task execution order, as follows:

[0026] RSU by task deadline d j Arrange the queue in ascending order;

[0027] Resource quotas are optimized using a BCD convex approximation, and EDF is used to map continuous decals into task execution order. The specific process of BCD convex approximation optimization is as follows:

[0028]

[0029] in, This indicates that resources are allocated on the j-th RSU. J represents the time required to complete its computational task. RSU This represents the set of tasks that are unloaded to RSU.

[0030] Optionally, in another possible implementation of the first aspect, the joint minimization of end-to-end delay in step S6 above is as follows:

[0031] Determine the optimization problem of overall system latency:

[0032]

[0033]

[0034] in, It is the task completion delay of the group. It refers to the task completion delay on the RSU.

[0035] Secondly, embodiments of this application provide a vehicle-to-everything (V2X) collaborative task offloading and fusion system based on GNN-PPO, including an in-vehicle terminal and an RSU cluster; the in-vehicle terminal and the RSU cluster communicate with each other via a wireless channel; the in-vehicle terminal is used for task generation, segmentation and inference; the RSU is used for heterogeneous graph construction, embedded computing and fusion offloading scheduling.

[0036] Beneficial Effects: This application provides a vehicle DNN task fusion and offloading method based on graph neural networks. The vehicle generates a DNN task at the terminal, uses an MLP model to determine the optimal segmentation point, and divides the system into vehicle-side and edge subgraphs. Responsive Query Unit (RSU) constructs a dynamic heterogeneous graph, and vehicle embedding vectors are extracted using a graph neural network (GNN). The vehicle outputs offloading and fusion actions based on a PPO agent, supporting feature fusion within groups. RSU employs convex optimization and greedy scheduling to jointly optimize resource allocation and task order, minimizing end-to-end latency and energy consumption. This application can efficiently utilize vehicle-side computing resources, improve task scheduling performance and system operating efficiency, and reduce the energy consumption and communication load of repetitive computations. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of a scenario provided in an embodiment of this application;

[0039] Figure 2 This is a flowchart illustrating a vehicle DNN task fusion and unloading method based on a graph neural network according to an embodiment of this application.

[0040] Figure 3 This is a schematic diagram of the structure of a vehicle networking collaborative task offloading and fusion system based on GNN-PPO provided in an embodiment of this application. Detailed Implementation

[0041] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0042] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0043] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0044] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0045] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0046] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0047] In one scenario embodiment, refer to Figure 1The participants include the vehicle generating the task and multiple roadside units (RSUs) deployed along the road, forming a two-layer vehicle-side architecture. Each RSU is equipped with an edge server, distributed along the unidirectional road, and connected via a local area network, responsible for receiving, fusing, and scheduling vehicle tasks. The road is divided into N regions, each covered by one RSU. Let N = {1, 2, ...} represent the set of RSUs; the computational, storage, and communication resources of each RSU are limited, denoted as Fn, Hn, and Un, respectively. The vehicle, as the user terminal, is responsible for generating DNN tasks, collecting status information, and uploading it to the RSU. Considering the high-speed mobility of vehicles, communication between the vehicle and the RSU typically uses a dedicated short-range communication (DSRC) channel. The tasks mainly include DNN inference tasks generated by the vehicle, such as autonomous driving target recognition, path planning, and traffic video processing tasks.

[0048] The following is a detailed description of a vehicle DNN task fusion and unloading method based on graph neural networks provided in this application, with reference to the accompanying drawings.

[0049] Figure 2 The illustration shows a flowchart of a vehicle DNN task fusion and unloading method based on a graph neural network according to an embodiment of this application.

[0050] like Figure 2 As shown, this vehicle DNN task fusion and unloading method based on graph neural networks includes:

[0051] S1. When the vehicle terminal generates a DNN task, it uses an MLP model to pre-analyze the network structure of the DNN task and determine the optimal segmentation point. Based on the optimal split point The DNN task is divided into vehicle side view sub-images and edge side view sub-images; the attributes of the DNN task are uploaded to RSU, including task description, task deadline, and optimal split point. The computational resources required for a DNN task, the storage resources required for a DNN task, and the communication resources required for a DNN task;

[0052] Furthermore, in the embodiments of this application, the optimal segmentation point in step S1 above... The specific determination method is as follows:

[0053]

[0054] in, This indicates the execution time for calculating the vehicle profile submap locally. This represents the inference time of the edge subgraph after it has been transmitted over the network;

[0055] Will and The minimum value of the sum corresponds to the layer, which is determined as the optimal split point. .

[0056] S2. The vehicle terminal periodically collects its own state and forms a state vector. The state vector is sent to the RSU. The RSU combines the attributes of the DNN task and the state vector to construct a dynamic heterogeneous graph. The vehicle nodes are clustered and the GNN forward propagation is performed. The vehicle embedding representation is generated and then sent back to the vehicle terminal.

[0057] Specifically, the vehicle-mounted terminal periodically collects its own status: location Speed Signal-to-noise ratio And form a state vector .

[0058] It should be noted that the dynamic heterogeneous graph constructed in step S2 above is represented as follows: ,in For the set of vehicle nodes, This refers to the set of connection relationships between vehicles or between a vehicle and an RSU. The node feature matrix contains location, vehicle speed, signal-to-noise ratio, and task attributes; the update rule for performing GNN forward propagation is:

[0059]

[0060] in, Indicates that node i is at the 1st rank. Layer embedding representation, Let i represent the set of neighbors of node i. This represents the learnable weight matrix of the l-th layer. This represents the learnable bias vector of the l-th layer.

[0061] Furthermore, in this embodiment, step S2 above employs a dynamic clustering method based on embedding similarity for clustering vehicle nodes, as detailed below:

[0062]

[0063] in, For the m-th group, For the embedding of node i, For the first Cluster centers This is the similarity threshold.

[0064] S3. The vehicle terminal embeds the vehicle's input into the PPO-based Actor module and outputs two action parameters: unloading action. 0 represents local uninstallation, 1 represents uninstallation from another vehicle within the group, 2 represents single-hop RSU uninstallation, and 3 represents multi-hop RSU uninstallation; and fusion layer actions. , and when At that time, the vehicle initiates feature fusion at the corresponding layer;

[0065] It should be noted that in the PPO-based Actor module described above, the loss function for PPO is defined as:

[0066]

[0067] in, For strategy ratio, For advantage function estimation, clip(*) is the policy gradient function. Update the step size limit for the strategy.

[0068] S4. When the output action parameter is a fusion layer action, the target node establishes a fusion pool, receives and aggregates the intermediate feature maps of vehicles in the group within the set waiting time slot, generates fused features, and then distributes them to each vehicle to continue executing subsequent layer inference.

[0069] Furthermore, in this embodiment of the application, step S4 includes:

[0070] The fusion pool receives and aggregates vehicle sets within the group. Intermediate feature map:

[0071]

[0072] in, Indicates vehicle In the Intermediate features of the layer output, This represents a fusion operator, taking the form of a weighted average or attention aggregation.

[0073] The fusion layer corresponding to the fusion pool pairs the intermediate feature maps. Processing is performed to generate fusion features. And by using task identifiers or weight mapping to fuse features Perform feature splitting, and then fuse the split features. Distribute to the computing nodes corresponding to each task, where The feature inverse mapping function ensures that each task continues forward inference of the remaining layers without relying on the original vehicle backhaul.

[0074] For layers that are not involved in the fusion, the current node completes the forward propagation independently.

[0075] In this embodiment, a hierarchical splitting and fusion mechanism for DNN tasks is implemented. The fusion layer shares computation within the group, while non-fusion layers complete independently, ensuring computational continuity, communication efficiency, and task independence. Furthermore, the model only needs a cold start in the system; all subsequent fusion layers and normal layers can share the same model instance for inference, eliminating the need for repeated loading.

[0076] S5. When the output action parameter is an unloading action, the vehicle performs the corresponding unloading operation according to the action type: when Local uninstallation is performed at that time; when When the unloading is performed on another vehicle within the group; Alternatively, single-hop RSU offloading and multi-hop RSU offloading can be performed at time 3; when the latency within the group exceeds the preset threshold, the task will be transferred from the group to RSU for continued execution.

[0077] S6. After the task is transferred to RSU, RSU optimizes the resource quota using BCD convex approximation based on the global queue, the remaining computing power of the node, and the link lifetime, and determines the task execution order in combination with EDF; by alternately optimizing the RSU resource quota and the fusion pool waiting time, the joint minimization of end-to-end latency is achieved.

[0078] Furthermore, in this embodiment, the BCD convex approximation optimization of resource quotas in step S6 above, combined with EDF to determine the task execution order, is as follows:

[0079] RSU by task deadline d j Arrange the queue in ascending order;

[0080] Resource quotas are optimized using a BCD convex approximation, and EDF is used to map continuous decals into task execution order. The specific process of BCD convex approximation optimization is as follows:

[0081]

[0082] in, This indicates that resources are allocated on the j-th RSU. J represents the time required to complete its computational task. RSU This represents the set of tasks that are unloaded to RSU.

[0083] Furthermore, in this embodiment, the joint minimization of end-to-end latency in step S6 is specifically as follows:

[0084] Determine the optimization problem of overall system latency:

[0085]

[0086]

[0087] in, It is the task completion delay of the group. It refers to the task completion delay on the RSU.

[0088] It should be noted that:

[0089]

[0090] in For the task in the layer The calculation time, For the set of layers participating in the fusion, To integrate waiting and calculation time;

[0091]

[0092] in For the transmission latency to upload to RSU, For queuing time, Calculate the delay for RSU;

[0093] The constraints of the overall system latency optimization problem mentioned above include computational resource constraints, communication resource constraints, and storage resource constraints:

[0094]

[0095] Where Jn is the set of tasks executed on RSUn. These represent the computing, communication, and storage resources used by the task.

[0096] As one possible implementation, in this embodiment, the learning and optimization of online distributed task scheduling strategies are achieved through joint training of GNN and PPO. The joint training process is as follows:

[0097] The vehicle generates a DNN task and collects its own state at the onboard terminal, and then converts the state vector... With task attribute vector Uploaded to RSU. RSU constructs a dynamic heterogeneous graph based on vehicle status and mission information. High-dimensional embedding vectors for each vehicle are generated through forward propagation of GNN. :

[0098]

[0099] The generated embedding vector The data is transmitted back to the vehicle as input to the PPO Actor module, used to output the unloading action. Actions with fusion layer .

[0100] The joint training process includes the following steps:

[0101] Step 1: State-Action Sampling: Under the current policy πθ, the vehicle samples the state-action vector... Generate unloading and fusion actions It performs task calculations, records task completion time, fusion delay, and resource consumption, and forms a state-action-reward triplet. .

[0102] Step 2: Policy Optimization: Based on the PPO algorithm, the policy network parameters θ are updated using the Clipped Policy Gradient, allowing the policy to gradually converge while minimizing task completion latency and optimizing resource utilization. The specific gradient calculation is as follows:

[0103]

[0104] in:

[0105]

[0106] For probability ratios, For the estimation of the advantage function, For shearing parameters.

[0107] Step 3: Value Network Update: Leveraging the PPO Value Network To evaluate the value of a state, the parameter ϕ is optimized by minimizing the squared error loss function:

[0108]

[0109] in This is the actual reward accumulated after discounts.

[0110] Step 4: Joint Optimization of GNN Embeddings: During policy training, the GNN embedding network and the PPO policy network are jointly updated. The GNN receives gradient signals from the policy network through backpropagation and adjusts the node feature aggregation weights to optimize the embedding vectors. It can more effectively reflect the correlation of vehicle tasks within the group and the overall status information of the system, thereby improving the accuracy of unloading and fusion decisions.

[0111] Step 5: Fusion Feedback: After the task is completed, the fusion layer latency, task completion time, and resource consumption information are sent back to the RSU as a reward. The components are fed back to the PPO for policy iteration, forming a closed-loop training mechanism.

[0112] Through the aforementioned joint training process, the GNN learns the high-dimensional structural features of the task and the vehicle, while the PPO generates unloading and fusion action strategies based on the embedding vectors. The optimization of these strategies is guided by system-level latency and resource consumption as reward functions, thereby achieving efficient unloading and fusion scheduling of the vehicle's DNN task. During system deployment, the trained GNN and PPO models can be directly used for online scheduling, allowing the vehicle to quickly generate unloading and fusion decisions based on the embedding vectors.

[0113] This application provides a vehicle-side DNN task fusion and offloading method based on graph neural networks. The vehicle generates a DNN task at the terminal, uses an MLP model to determine the optimal segmentation point, and divides the graph into vehicle-side and edge subgraphs. Responsive Query Unit (RSU) constructs a dynamic heterogeneous graph, and extracts vehicle embedding vectors using a graph neural network (GNN). The vehicle outputs offloading and fusion actions based on a PPO agent, supporting feature fusion within groups. RSU employs convex optimization and greedy scheduling to jointly optimize resource allocation and task order, minimizing end-to-end latency and energy consumption. This application can efficiently utilize vehicle-side computing resources, improve task scheduling performance and system operating efficiency, and reduce the energy consumption and communication load of repetitive computations.

[0114] Figure 3 This paper presents a schematic diagram of a vehicle-to-everything (V2X) collaborative task offloading and fusion system based on GNN-PPO, as provided in an embodiment of this application.

[0115] like Figure 3 As shown, the GNN-PPO-based vehicle network collaborative task offloading and fusion system 300 includes an on-board terminal 301 and an RSU cluster 302; the on-board terminal and the RSU cluster communicate with each other via a wireless channel; the on-board terminal is used for task generation, segmentation and inference; the RSU is used for heterogeneous graph construction, embedded computing and fusion offloading scheduling.

[0116] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0117] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A vehicle DNN task fusion offloading method based on a graph neural network, characterized in that, Includes the following steps: S1, when a vehicle-mounted terminal generates a deep neural network (DNN) task, a MLP model is used to pre-analyze the network structure of the DNN task to determine an optimal segmentation point , according to the optimal segmentation point , the DNN task is divided into a vehicle-side subgraph and an edge-side subgraph; and attributes of the DNN task, including a task description, a task deadline, the optimal segmentation point , required computing resources of the DNN task, required storage resources of the DNN task, and required communication resources of the DNN task, are uploaded to a roadside unit (RSU). S2, the vehicle terminal periodically collects its own state: position , vehicle speed , signal-to-noise ratio , and forms a state vector , sends the state vector to the RSU, constructs a dynamic heterogeneous graph using the RSU combining the attributes of the DNN task and the state vector, clusters the vehicle nodes and performs graph neural network GNN forward propagation, and generates vehicle embedding representation back to the vehicle terminal; S3, the vehicle terminal inputs the vehicle embedding representation into an Actor module based on a proximal policy optimization algorithm PPO, and outputs two action parameters: an unloading action , 0 for local unloading, 1 for unloading by another vehicle in the group, 2 for single-hop RSU unloading, and 3 for multi-hop RSU unloading; and a fusion layer action , and when , the vehicle starts feature fusion at the corresponding layer. S4. When the output action parameter is a fusion layer action, the target node establishes a fusion pool, receives and aggregates the intermediate feature maps of vehicles in the group within the set waiting time slot, generates fused features, and then distributes them to each vehicle to continue executing subsequent layer inference. S5、When the output action parameter is an offloading action, the vehicle performs the corresponding offloading operation according to the action type: when , a local offloading is performed; when , an intra-group other-vehicle offloading is performed; when or 3, a single-hop RSU offloading and a multi-hop RSU offloading are respectively performed; when the intra-group delay exceeds a preset threshold, the task is transferred from the intra-group to the RSU for continuous execution; S6. After the task is transferred to RSU, RSU optimizes the resource quota using the block coordinate descent method (BCD convex approximation) based on the global queue, the remaining computing power of the node, and the link lifetime, and determines the task execution order by combining the earliest deadline first algorithm (EDF). By alternately optimizing the RSU resource quota and the fusion pool waiting time, the joint minimization of end-to-end latency is achieved.

2. The method of claim 1, wherein, The optimal split point in the step S1 The specific determination method is as follows: ； wherein, denotes the execution time of the vehicle-side subgraph computed locally, denotes the inference time of the edge-side subgraph after network transmission; Will and The minimum value of the sum corresponds to the layer, which is determined as the optimal split point. .

3. The method of claim 2, wherein, The dynamic heterogeneous graph constructed in the step S2 is represented as wherein is a set of vehicle nodes, is a set of connection relationships between vehicles or between a vehicle and an RSU, is a node feature matrix containing position, vehicle speed, signal-to-noise ratio, and task attributes; and the update rule for performing GNN forward propagation is: ； wherein, represents the embedding of node i at the layer, represents the set of neighbors of node i, represents the learnable weight matrix of the represents the learnable bias vector of the 4. The method of claim 3, wherein, Step S2 employs a dynamic clustering method based on embedding similarity for clustering vehicle nodes, as detailed below: ； wherein, is the mth group, is the embedding of node i, is the mth cluster center is the similarity threshold.

5. The method of claim 4, wherein, In step S4, when the output action parameter is a fusion layer action, the target node establishes a fusion pool, receives and aggregates the intermediate feature maps of vehicles within the group within a set waiting time slot, generates fused features, and distributes them to each vehicle to continue executing subsequent layer inference. Specifically, this includes: The fusion pool receives and aggregates the set of vehicles within the group intermediate feature maps: ； wherein, representing a vehicle In a first layer outputted intermediate features, representing a fusion operator, taking the form of a weighted average or attention aggregation; The fusion layer corresponding to the fusion pool affects the intermediate feature map. Processing is performed to generate fusion features. And by using task identifiers or weight mapping to fuse features Perform feature splitting, and then fuse the split features. Distribute to the computing nodes corresponding to each task, where The feature inverse mapping function ensures that each task continues forward inference of the remaining layers without relying on the original vehicle backhaul. For layers that are not involved in the fusion, the current node completes the forward propagation independently.

6. The method of claim 5, wherein, In the PPO-based Actor module of step S3, the loss function of PPO is defined as: ； where, is the policy ratio, is the advantage function estimate, clip(*) is the policy gradient function, is the policy update step limit.

7. The method of claim 6, wherein, The step S6, which uses BCD convex approximation to optimize resource quotas and combines it with EDF to determine the task execution order, is as follows: RSU by task deadline d j Arrange the queue in ascending order; Resource quotas are optimized using a BCD convex approximation, and EDF is used to map consecutive demaps to the task execution order. The specific process of BCD convex approximation optimization is as follows: ； wherein, denotes the allocated resource on the jth RSU, denotes the time required to complete its computing task, J RSU denotes the set of tasks offloaded to the RSU.

8. The method of claim 1, wherein, The joint minimization of end-to-end delay in step S6 is as follows: Determine the optimization problem of overall system latency: ； ； wherein, is the task completion latency of the group, is the task completion latency on the RSU.

9. A GNN-PPO-based cooperative task offloading and fusion system for Internet of Vehicles, applied to the method of any one of claims 1-8. It includes an onboard terminal and an RSU cluster; the onboard terminal and the RSU cluster communicate with each other via a wireless channel; the onboard terminal is used for task generation, segmentation and inference; the RSU is used for heterogeneous graph construction, embedded computing and fusion offloading scheduling.

Images