A Multi-Dimensional Resource Joint Allocation Method for Cloud-Edge Collaborative Networks Based on Semantic Communication

Abstract

This invention relates to the field of wireless communication and provides a method for joint allocation of multi-dimensional resources in cloud-edge collaborative networks based on semantic communication. The method includes: acquiring task attributes and channel state information in a heterogeneous network; based on the task attributes and channel state information, using a lightweight neural network deployed at edge nodes to adaptively extract semantic features from the original sensing data, generating semantic feature data; inputting the local observation state into an outer decision network, and combining it with an action masking mechanism to generate discrete decision combinations; the decision combinations include computational offloading path decisions and semantic segmentation point decisions; based on the decision combinations, using a closed-form solution based on the Cartesian-Kuhn-Tucker condition at a remote computing center to allocate continuous physical computing resources to concurrent tasks. This invention solves the problem in existing technologies of jointly optimizing discrete decision-making and continuous resource allocation in bandwidth-constrained cloud-edge collaborative networks, achieving efficient joint allocation of multi-dimensional resources based on semantic communication.

Description

Technical Field

[0001] This invention relates to the field of wireless communication, and in particular to a method for joint allocation of multi-dimensional resources in a cloud-edge collaborative network based on semantic communication. Background Technology

[0002] With the rapid development of 6G mobile communication and IoT technologies, cloud-edge collaborative heterogeneous networks have become a key infrastructure supporting the intelligent interconnection of everything. In these networks, massive numbers of edge sensing nodes need to transmit the collected multimodal data back to remote computing centers for processing. However, in actual deployments, the wireless backhaul links between edge nodes and remote computing centers generally suffer from bandwidth limitations and scarce spectrum resources, making it difficult to support the full transmission of high-dimensional sensing data.

[0003] To alleviate bandwidth pressure on communication links, existing technologies attempt to deploy lightweight neural networks at edge nodes. After preliminary processing of the raw data, only intermediate layer features are transmitted to reduce data transmission volume. Meanwhile, regarding resource scheduling issues in cloud-edge collaborative networks, existing solutions typically employ heuristic algorithms or traditional optimization methods to make decisions on the offloading path of computing tasks and resource allocation.

[0004] However, in heterogeneous network environments with highly limited resources and dynamically changing topologies, the existing solutions mentioned above have the following drawbacks: On the one hand, traditional heuristic or numerical optimization algorithms are difficult to adapt to rapid changes in network states and lack timely decision-making; on the other hand, existing resource scheduling methods cannot effectively handle mixed optimization problems that simultaneously include discrete and continuous decision variables, resulting in low system resource utilization and difficulty in meeting the high-efficiency collaboration requirements in multi-task concurrent scenarios.

[0005] Therefore, how to achieve joint optimization and allocation of multi-dimensional resources in cloud-edge collaborative networks with limited bandwidth and computing power has become an important issue that the industry urgently needs to address. Summary of the Invention

[0006] This invention provides a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication, which solves the problem in the prior art of jointly optimizing discrete decision-making and continuous resource allocation in bandwidth-constrained cloud-edge collaborative networks, and realizes efficient joint allocation of multi-dimensional resources based on semantic communication.

[0007] This invention provides a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication, comprising the following steps: Obtain task attributes and channel state information in heterogeneous networks; Based on the task attributes and channel state information, a lightweight neural network deployed at the edge nodes is used to adaptively extract semantic features from the original sensing data to generate semantic feature data. The local observation state is input into the outer decision network, and combined with the action masking mechanism, discrete decision combinations are generated; the decision combinations include calculation of unloading path decisions and semantic segmentation point decisions; Based on the aforementioned decision combination, continuous physical computing resources are allocated to concurrent tasks at the remote computing center using a closed-form solution based on the Caro-Kun-Tucker condition.

[0008] According to the present invention, a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication is provided. The method involves using a lightweight neural network deployed at edge nodes to adaptively extract semantic features from raw sensing data based on task attributes and channel state information, generating semantic feature data. The method includes: obtaining semantic segmentation point instructions from an outer decision network; determining the forward processing layer of the lightweight neural network based on the semantic segmentation point instructions; inputting the raw sensing data into the encoder part of the lightweight neural network and performing forward computation layer by layer according to the determined forward processing layer; stopping the forward computation and extracting the output of the current layer as semantic feature data after reaching the network layer specified by the semantic segmentation point instructions; wherein the dimension of the semantic feature data is lower than the dimension of the raw sensing data, and the density of task-related features in the semantic feature data is higher than the density of task-related features in the raw sensing data.

[0009] According to the present invention, a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication is provided. The method involves inputting local observation states into an outer decision network and generating discrete decision combinations using an action masking mechanism. The method includes: inputting the local observation states into a multi-agent reinforcement learning action network; outputting the initial logarithmic probabilities corresponding to each candidate offloading path and each candidate semantic segmentation point through the forward computation of the action network; obtaining the connectivity status of each wireless communication link at the current time, where the connectivity status includes visibility indicators for edge-to-edge links and edge-to-cloud links; generating an action mask vector for each candidate offloading path based on the connectivity status, wherein the mask value corresponding to a link with a connectivity status of interruption or unreachability is set to negative infinity; superimposing the action mask vector with the initial logarithmic probabilities to obtain a corrected logarithmic probability distribution; performing a flexible maximum normalization operation on the corrected logarithmic probability distribution to obtain a probability distribution of effective actions; and determining the computational offloading path decision and semantic segmentation point decision at the current time based on random sampling or greedy selection according to the probability distribution of effective actions.

[0010] According to the present invention, a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication is provided. The method involves allocating continuous physical computing resources to concurrent tasks at a remote computing center using a closed-form solution based on the Caro-Kuhn-Tucker condition, according to the decision combination. This includes: at the remote computing center, aggregating all tasks selected to be offloaded to the remote computing center; for each aggregated task, obtaining the semantic segmentation point decision corresponding to the current task, and determining the remaining computing requirements of the current task based on the semantic segmentation point decision; obtaining the comprehensive urgency weight of each task; and assigning the weights of each task to the appropriate urgency weights. The remaining computing requirements and the overall urgency weight are substituted into the convex optimization closed-form formula based on the Caro-Kuhn-Tucker condition to calculate the optimal computing power allocation value for each task. According to the optimal computing power allocation value, corresponding continuous physical computing power resources are allocated to each task. The convex optimization closed-form formula uses the product of the square root of the remaining computing requirements of each task and the square root of the overall urgency weight as the numerator, the sum of the numerators of all aggregated tasks as the denominator, and then multiplies it by the upper limit of the total physical computing power capacity of the remote computing center to obtain the optimal computing power allocation value for each task.

[0011] According to the present invention, a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication is provided. After allocating continuous physical computing resources, the method further includes: transmitting the semantic feature data to a remote computing center for task inference based on the allocated physical computing resources to obtain task inference results; and updating the parameters of the outer decision network based on the task inference results.

[0012] According to the present invention, a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication is provided. The step of transmitting semantic feature data to a remote computing center for task inference based on allocated physical computing resources includes: determining the target remote computing center corresponding to the semantic feature data based on the computation offloading path decision; transmitting the semantic feature data to the target remote computing center via a wireless link specified by the computation offloading path decision; determining the corresponding input layer of a deep neural network decoder at the target remote computing center based on the semantic segmentation point decision; inputting the received semantic feature data to the corresponding input layer of the decoder; and controlling the decoder to perform subsequent forward computations starting from the corresponding input layer until the final task inference result is output.

[0013] According to the present invention, a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication is provided. The step of updating the parameters of the outer decision network based on the task inference result includes: obtaining the task execution latency and inference accuracy corresponding to the task inference result; calculating a comprehensive utility reward value based on the task execution latency and the inference accuracy; inputting the comprehensive utility reward value into a centralized evaluation network, and calculating a state value function through the centralized evaluation network; calculating an estimated dominance function using a generalized dominance estimation method based on the comprehensive utility reward value and the state value function; calculating the gradient of the loss function of the action network and the gradient of the loss function of the centralized evaluation network based on the estimated dominance function; updating the network parameters of the action network using a gradient ascent method based on the gradient of the loss function of the action network; and updating the network parameters of the centralized evaluation network using a gradient descent method based on the gradient of the loss function of the centralized evaluation network.

[0014] This invention also provides a cloud-edge collaborative network multi-dimensional resource allocation device based on semantic communication, comprising the following modules: The state awareness module is used to acquire task attributes and channel state information in heterogeneous networks. The semantic feature extraction module, deployed at the edge node, is used to adaptively extract semantic features from the original sensing data based on the task attributes and channel state information, using a lightweight neural network deployed at the edge node, to generate semantic feature data. The outer decision module, deployed at the edge nodes, is used to input the local observation state into the outer decision network and generate discrete decision combinations by combining the action masking mechanism; the decision combinations include calculating the unloading path decision and the semantic segmentation point decision. The resource allocation module, deployed in a remote computing center, is used to allocate continuous physical computing resources to concurrent tasks in the remote computing center based on the decision combination and using a closed-form solution based on the Caro-Kuhn-Tucker condition.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication as described above.

[0016] The present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication as described above.

[0017] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication as described above.

[0018] This invention provides a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication, which has the following beneficial effects: By acquiring task attribute and channel state information in heterogeneous networks, and using lightweight neural networks deployed on edge nodes to adaptively extract semantic features from the original sensing data, compact semantic feature data is generated, thereby reducing the amount of data to be transmitted at the source; by inputting the local observation state into the outer decision network and combining it with an action masking mechanism, a discrete decision combination including computational unloading paths and semantic segmentation points is generated, thereby achieving rapid adaptive decision-making for highly dynamic network environments; by allocating continuous physical computing resources to concurrent tasks at a remote computing center based on the Cartesian-Kun-Tucker condition according to the discrete decision combination, the precise allocation of continuous resources is achieved at the mathematical level. In summary, claim 1 solves the joint optimization problem of discrete decision-making and continuous resource allocation in cloud-edge collaborative networks through a collaborative mechanism of "semantic feature extraction + outer discrete decision + inner closed-form solution continuous allocation", achieving efficient joint allocation of multi-dimensional resources under bandwidth-constrained conditions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication provided by the present invention.

[0021] Figure 2 This is a framework diagram of the cloud-edge collaborative heterogeneous network and semantic communication system provided by the present invention.

[0022] Figure 3 This is a flowchart of the resource joint allocation method provided by the present invention.

[0023] Figure 4 This is an architecture diagram of the two-layer decoupled resource allocation method provided by the present invention.

[0024] Figure 5 This is a comparison chart of the training convergence performance of the algorithm provided by this invention.

[0025] Figure 6This invention provides a comparison of the total system utility and task success rate under different task loads.

[0026] Figure 7 This is a schematic diagram of the structure of the cloud-edge collaborative network multi-dimensional resource joint allocation device based on semantic communication provided by the present invention.

[0027] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] The terminology involved in this invention will be explained below.

[0030] (1) Semantic communication Semantic communication is a novel task-oriented communication paradigm. Unlike traditional communication, which focuses on transmitting raw bitstreams without loss of quality, semantic communication uses deep learning models to extract core features (i.e., semantic information) directly related to downstream specific tasks (such as object recognition and classification) at the sending end, and only transmits these features. It can significantly eliminate redundant information irrelevant to the task from the original data, thereby greatly reducing communication bandwidth requirements and providing an effective way to alleviate the bandwidth bottleneck of wireless backhaul links.

[0031] (2) Joint allocation of resources Joint resource allocation refers to the unified planning and scheduling optimization of system resources across multiple dimensions in a resource-constrained network environment. In this invention, specifically targeting the limited wireless communication bandwidth and physical computing power in cloud-edge collaborative networks, it combines discrete computation offloading path selection, semantic segmentation point selection from deep neural networks, and continuous remote computing resource allocation for global joint optimization. This maximizes the overall system utility while ensuring task latency and success rate. (3) Multi-agent reinforcement learning Multi-agent reinforcement learning (MARL) is an important branch of reinforcement learning, involving multiple agents interacting and learning in a distributed manner within the same environment. In this invention, each edge node is treated as an independent agent. Due to limited edge-to-edge communication resources and dynamic topology changes, each edge node needs to autonomously make resource scheduling decisions based on its local state observations. Through trial-and-error interaction with the network environment, it learns a cooperative strategy that maximizes the long-term cumulative utility of the system.

[0032] (4) Two-layer decoupled architecture The two-layer decoupled architecture is a mechanism that decomposes the highly complex mixed-integer nonlinear programming (MINLP) problem into two cooperating layers for solution. In this invention, the outer layer uses a reinforcement learning network to handle discrete decision variables (such as unloading paths and semantic segmentation points); the inner layer abandons neural network prediction and instead uses a closed-form solution formula based on KKT conditions to precisely allocate continuous physical resources (such as remote computing capacity). This architecture effectively avoids the curse of dimensionality faced by traditional reinforcement learning when dealing with mixed action spaces and mathematically guarantees that it does not violate the hard constraints of physical resources.

[0033] (5) Cloud-edge collaborative heterogeneous network In a broad sense, this refers to a distributed computing and communication network architecture that deeply integrates centralized remote cloud computing power with distributed edge computing power. The "heterogeneity" is reflected in the fact that network nodes encompass a variety of devices with different computing capabilities, storage capacities, and communication protocols (such as terminal sensors, edge gateways, base stations, and cloud servers), and communication links include various media with varying bandwidth, latency, and topology characteristics (such as fiber optics, mobile cellular networks, and microwave links). This network achieves dynamic offloading of complex computing tasks and joint scheduling of multi-dimensional resources through a cloud-edge collaboration mechanism. Applicable scenarios for this architecture include, but are not limited to: Industrial Internet of Things (IIoT), Vehicle-to-Everything (V2X), drone swarm collaboration, smart city sensing networks, and integrated air-space-ground networks.

[0034] This invention discloses a multi-dimensional resource joint allocation method and system for cloud-edge collaborative networks based on semantic communication, relating to the fields of wireless communication and edge computing technology. Addressing the challenges of limited wireless link bandwidth, insufficient computing power at edge nodes, and the tendency of traditional reinforcement learning to violate system physical constraints when solving high-dimensional hybrid resource allocation problems in existing heterogeneous networks for massive sensing data backhaul, this invention proposes a task-oriented, two-layer decoupled resource scheduling scheme. First, a collaborative network scenario including edge nodes and a remote computing center is constructed. Second, an adaptive semantic segmentation model is built using deep neural networks to extract and transmit compact semantic features at edge nodes, significantly reducing communication overhead. Finally, a two-layer decoupled architecture is adopted to perform joint resource optimization: the outer layer uses multi-agent reinforcement learning combined with action masking mechanisms to make discrete computational offloading paths and semantic segmentation point decisions; the inner layer uses a closed-form solution formula based on KKT conditions to accurately allocate continuous computing resources at the remote computing center. This invention not only effectively overcomes the curse of dimensionality in mixed action spaces, but also strictly guarantees at the mathematical level that the system allocation does not exceed the constraints of physical computing power. It can be widely applied to various resource-constrained and highly dynamic congested network scenarios, significantly improving the overall utility and task success rate of the system.

[0035] The following is combined with Figures 1-8 The embodiments of the present invention are described in detail.

[0036] The cloud-edge collaborative network multi-dimensional resource joint allocation method based on semantic communication provided in this embodiment of the invention is executed by a cloud-edge collaborative network multi-dimensional resource joint allocation device based on semantic communication. This device can be configured in a computer, which can be a local computer or a cloud computer. The local computer can be a computer, tablet, etc., and no specific limitation is made here.

[0037] Figure 1 This is a flowchart illustrating the multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps: S110. Obtain task attributes and channel state information in heterogeneous networks.

[0038] S120. Based on the task attributes and channel state information, a lightweight neural network deployed at the edge node is used to adaptively extract semantic features from the original sensing data to generate semantic feature data.

[0039] S130. Input the local observation state into the outer decision network and combine it with the action masking mechanism to generate discrete decision combinations; the decision combinations include calculation of unloading path decisions and semantic segmentation point decisions.

[0040] S140. Based on the decision combination, at the remote computing center, continuous physical computing resources are allocated to concurrent tasks using a closed-form solution based on the Caro-Kun-Tucker condition.

[0041] In one specific embodiment, the task attributes and channel state information of each edge node in the cloud-edge collaborative heterogeneous network are first obtained. The task attributes include the dimension, computational load, and latency tolerance of the raw sensing data, while the channel state information includes the channel gain and connectivity status of the edge-to-edge link and the edge-to-cloud link.

[0042] Based on the information obtained above, each edge node responds to the semantic segmentation point instructions issued by the outer-layer decision network using a locally deployed lightweight deep neural network encoder. The edge node determines the forward processing layer of the neural network according to the semantic segmentation point instructions, inputs the original high-dimensional perceptual data into the lightweight neural network, and performs forward computation layer by layer. When it reaches the network layer specified by the semantic segmentation point instructions, it stops computation and extracts the output of the current layer, using it as compact semantic feature data. This semantic feature data has a lower dimensionality than the original perceptual data and retains core features highly relevant to downstream tasks.

[0043] Each edge node inputs its local observation state into the action network of multi-agent reinforcement learning. This action network outputs the initial log probabilities corresponding to each candidate offloading path and each candidate semantic segmentation point through forward computation. Simultaneously, the system acquires the connectivity state of each wireless communication link at the current moment and generates an action mask vector for the candidate offloading paths based on the connectivity state, setting the mask value corresponding to interrupted or unreachable links to negative infinity. After superimposing this action mask vector with the initial log probabilities, a flexible maximum normalization operation is performed on the corrected log probability distribution to obtain the probability distribution of effective actions. Finally, the offloading path decision and semantic segmentation point decision at the current moment are determined through random sampling or greedy selection.

[0044] The remote computing center aggregates all tasks selected for offloading to it. For each task, it obtains its semantic segmentation point decision and determines the remaining computing demand accordingly, while also acquiring the comprehensive urgency weight of each task. The product of the square root of the remaining computing demand of each task and the square root of the comprehensive urgency weight is used as the numerator, and the sum of the numerators of all aggregated tasks is used as the denominator. This sum is then multiplied by the maximum physical computing capacity of the remote computing center. The optimal computing power allocation value for each task is calculated using a convex optimization closed-form formula based on the Caro-Kuhn-Tucker condition, and corresponding continuous physical computing power resources are allocated to each task according to this allocation value.

[0045] This embodiment compresses high-dimensional raw perceptual data into low-dimensional, compact semantic feature data at edge nodes through adaptive semantic feature extraction, reducing data transmission volume on wireless links. It generates discrete offloading paths and semantic segmentation point decisions through an outer decision network combined with an action masking mechanism, adapting to highly dynamic network topology changes and shielding unreachable links. Furthermore, it allocates continuous physical computing resources to concurrent tasks using a closed-form solution based on the Caro-Kuhn-Tucker condition in the inner layer, achieving precise allocation of computing resources. In a bandwidth-constrained cloud-edge collaborative network, this embodiment achieves joint optimization of discrete decision-making and continuous resource allocation, improving the system's efficiency in jointly allocating multi-dimensional resources.

[0046] According to the present invention, a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication is provided. Based on task attributes and channel state information, a lightweight neural network deployed at edge nodes is used to adaptively extract semantic features from raw sensing data to generate semantic feature data. The method includes: obtaining semantic segmentation point instructions issued by an outer decision network; determining the forward processing layer of the lightweight neural network based on the semantic segmentation point instructions; inputting the raw sensing data into the encoder part of the lightweight neural network and performing forward calculation layer by layer according to the determined forward processing layer; stopping the forward calculation and extracting the output of the current layer after reaching the network layer specified by the semantic segmentation point instructions, as the semantic feature data; wherein the dimension of the semantic feature data is lower than that of the raw sensing data, and the density of task-related features in the semantic feature data is higher than that in the raw sensing data.

[0047] In one specific embodiment, the edge node first receives a semantic segmentation point instruction from the outer decision network. This semantic segmentation point instruction is used to instruct the lightweight neural network to stop computation and output features at a specific network layer position during the forward computation process. Based on this semantic segmentation point instruction, the edge node determines the forward processing layer corresponding to the locally deployed lightweight neural network, that is, it determines which layer of the network needs to be executed.

[0048] Edge nodes input the collected raw perceptual data into the encoder part of the lightweight neural network, and perform forward computation layer by layer, starting from the first layer of the network, according to the determined forward processing levels. During the layer-by-layer computation, when the computation reaches the network level specified by the semantic segmentation point instruction, the edge node immediately stops the forward computation of subsequent layers and extracts the feature data output by the current network level, using this feature data as the semantic feature data to be transmitted. The semantic feature data obtained in this way has a lower dimensionality than the raw perceptual data, and the feature density related to the downstream computation task in this semantic feature data is higher than the task-related feature density in the raw perceptual data.

[0049] This embodiment receives semantic segmentation point instructions from the outer decision network, determines the forward processing layer of the lightweight neural network based on these instructions, and stops forward computation upon reaching the specified network layer, extracting the output of the current layer as semantic feature data. This allows edge nodes to transmit only compressed, compact features instead of the full original data. This embodiment reduces the amount of data transmitted from edge nodes to the remote computing center while maintaining high density of features relevant to downstream tasks.

[0050] According to the present invention, a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication is provided. This method inputs local observation states into an outer decision network and, combined with an action masking mechanism, generates discrete decision combinations. The method includes: inputting local observation states into a multi-agent reinforcement learning action network; outputting the initial log probabilities corresponding to each candidate offloading path and each candidate semantic segmentation point through forward computation of the action network; obtaining the connectivity status of each wireless communication link at the current time, including visibility indicators for edge-to-edge and edge-to-cloud links; generating action mask vectors for candidate offloading paths based on the connectivity status, wherein the mask value corresponding to links with a connectivity status of interruption or unreachability is set to negative infinity; superimposing the action mask vectors with the initial log probabilities to obtain a corrected log probability distribution; performing flexible maximum normalization on the corrected log probability distribution to obtain the probability distribution of effective actions; and determining the computational offloading path decision and semantic segmentation point decision at the current time by performing random sampling or greedy selection based on the probability distribution of effective actions.

[0051] In one specific embodiment, each edge node inputs its local observation state into a multi-agent reinforcement learning action network. This action network is a policy network based on a multi-agent proximal policy optimization algorithm, used to output corresponding action decisions based on the current observation state. Through forward computation, the action network outputs the corresponding initial log probabilities for each candidate offloading path and each candidate semantic segmentation point. Simultaneously, the edge nodes acquire the connectivity status of each wireless communication link at the current time. This connectivity status includes visibility indicators for edge-to-edge and edge-to-cloud links, used to characterize whether each link is currently connected or within communication coverage. Based on the acquired connectivity status, the system generates action mask vectors for candidate offloading paths: for candidate offloading paths corresponding to links with a connectivity status of interrupted or unreachable, the mask value corresponding to that path in the action mask vector is set to negative infinity; for candidate offloading paths corresponding to reachable links, the mask value is set to zero.

[0052] The system element-wise superimposes the action mask vector with the initial log probabilities output by the action network to obtain a corrected log probability distribution. A flexible maximum normalization operation is then performed on the corrected log probability distribution, mapping it to a probability distribution where the sum of the probabilities of all valid actions is 1, thus obtaining the probability distribution of valid actions. In this probability distribution, the probability corresponding to paths masked as negative infinity is reduced to zero, thereby completely blocking unreachable offloading paths at the decision-making level.

[0053] Based on the probability distribution of effective actions, the system uses random sampling or greedy selection (i.e., selecting the action with the highest probability) to determine the computation unloading path decision and semantic segmentation point decision of the edge node at the current moment.

[0054] This embodiment inputs the local observation state into the action network of multi-agent reinforcement learning, outputting an initial logarithmic probability. This initial logarithmic probability is then corrected by combining it with an action mask vector generated based on link connectivity states. After flexible maximum normalization, the probability distribution of effective actions is obtained, ultimately determining the offloading path decision and semantic segmentation point decision. In a cloud-edge collaborative network with highly dynamic topology changes, this embodiment effectively masks invalid offloading paths caused by link interruptions or unreachability, ensuring that the outer decision network outputs only physically feasible decision combinations.

[0055] According to the present invention, a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication is provided. Based on decision combinations, continuous physical computing resources are allocated to concurrent tasks at a remote computing center using a closed-form solution based on the Caro-Kuhn-Tucker condition. The method includes: at the remote computing center, aggregating all tasks selected to be offloaded to the remote computing center; for each aggregated task, obtaining the semantic segmentation point decision corresponding to the current task, and determining the remaining computing demand of the current task based on the semantic segmentation point decision; obtaining the comprehensive urgency weight of each task; substituting the remaining computing demand and comprehensive urgency weight of each task into the convex optimization closed-form solution formula based on the Caro-Kuhn-Tucker condition to calculate the optimal computing power allocation value for each task; and allocating corresponding continuous physical computing resources to each task based on the optimal computing power allocation value. The convex optimization closed-form solution formula uses the product of the square root of the remaining computing demand of each task and the square root of the comprehensive urgency weight as the numerator, the sum of the numerators of all aggregated tasks as the denominator, and multiplying by the upper limit of the total physical computing power capacity of the remote computing center to obtain the optimal computing power allocation value for each task.

[0056] In one specific embodiment, the remote computing center first aggregates all edge node decisions regarding computing tasks to be offloaded to the center. For each aggregated task, the remote computing center obtains the semantic segmentation point decision corresponding to the task and determines the remaining computing requirements that the task still needs to be executed at the remote computing center based on the semantic segmentation point decision. Simultaneously, the remote computing center obtains the comprehensive urgency weight of each task, which characterizes the overall urgency of each task in terms of latency requirements, task priority, and other factors.

[0057] The remote computing center substitutes the remaining computing requirements of each task with the overall urgency weight into the closed-form formula of convex optimization based on the Caro-Kuhn-Tucker conditions to calculate the optimal computing power allocation value for each task. Specifically, for each task, the numerator is obtained by multiplying the square root of its remaining computing requirements by the square root of its overall urgency weight; the numerators of all aggregated tasks are summed to obtain the denominator; the numerator of this task is divided by the denominator, and then multiplied by the maximum physical computing power capacity of the remote computing center to obtain the optimal computing power allocation value for that task. Based on the calculated optimal computing power allocation value, the remote computing center allocates corresponding continuous physical computing power resources to each task.

[0058] This embodiment aggregates all offloaded tasks at a remote computing center, determines the remaining computing requirements based on the semantic segmentation points of each task, and, combined with the comprehensive urgency weight of each task, calculates the optimal computing power allocation value for each task using a convex optimization closed-form formula based on the Caro-Kuhn-Tucker condition. It then allocates corresponding continuous physical computing resources to each task. This embodiment achieves precise allocation of continuous physical computing resources for concurrent tasks at the remote computing center of a cloud-edge collaborative network.

[0059] According to the present invention, a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication is provided. After allocating continuous physical computing resources, semantic feature data is transmitted to a remote computing center for task inference based on the allocated physical computing resources to obtain task inference results; and the parameters of the outer decision network are updated based on the task inference results.

[0060] In one specific embodiment, after the remote computing center allocates continuous physical computing resources to each concurrent task, each edge node determines the target remote computing center corresponding to its semantic feature data based on its respective computing offload path decision. The edge node transmits the extracted semantic feature data to the corresponding target remote computing center through the wireless link specified by the computing offload path decision. After receiving the semantic feature data, the target remote computing center determines the corresponding input layer of the locally deployed deep neural network decoder based on the semantic segmentation point decision corresponding to the task, and inputs the received semantic feature data into the corresponding input layer of the decoder.

[0061] The decoder executes subsequent forward computations starting from the input level until the final task inference result is output. After obtaining the task inference result, the system acquires the corresponding task execution latency and inference accuracy, and calculates the comprehensive utility reward value based on these parameters. This comprehensive utility reward value is then input into the centralized evaluation network, which calculates the state value function. Based on the comprehensive utility reward value and the state value function, the generalized dominance estimation method is used to calculate the dominance function estimate. Based on the dominance function estimate, the loss function gradients of the action network and the centralized evaluation network are calculated respectively. Finally, the gradient ascent method is used to update the action network parameters based on the action network's loss function gradient, and the gradient descent method is used to update the centralized evaluation network parameters based on the centralized evaluation network's loss function gradient.

[0062] In this embodiment, after allocating continuous physical computing resources, semantic feature data is transmitted to the target remote computing center according to the calculation unloading path decision. The corresponding input level of the decoder is determined according to the semantic segmentation point decision to perform subsequent forward calculations to obtain the task inference results. Then, the comprehensive utility reward value is calculated based on the task execution latency and inference accuracy. The generalized advantage estimation method is used to calculate the advantage function estimate value, and the network parameters of the action network and the centralized evaluation network are updated respectively.

[0063] According to the present invention, a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication is provided. Based on the allocated physical computing resources, semantic feature data is transmitted to a remote computing center for task inference. The method includes: determining the target remote computing center corresponding to the semantic feature data based on a computation offloading path decision; transmitting the semantic feature data to the target remote computing center via a wireless link specified by the computation offloading path decision; at the target remote computing center, determining the corresponding input layer of the deep neural network decoder based on a semantic segmentation point decision; inputting the received semantic feature data into the corresponding input layer of the decoder; and controlling the decoder to execute subsequent forward computations starting from the corresponding input layer until the final task inference result is output.

[0064] In one specific embodiment, after the remote computing center allocates continuous physical computing resources to each concurrent task, each edge node first determines the target remote computing center to which the semantic feature data extracted by its node needs to be sent, based on the computation offloading path decision obtained from its own decision. This computation offloading path decision specifies whether the semantic feature data is offloaded to the edge aggregation center or to the remote cloud center. The edge node then transmits the semantic feature data through the wireless link specified by the decision, according to the computation offloading path decision. After the semantic feature data arrives at the target remote computing center, the target remote computing center determines from which network layer the locally deployed deep neural network decoder should begin receiving input, based on the semantic segmentation point decision corresponding to the task.

[0065] The semantic segmentation point decision indicates the hierarchical position where the encoder stops during forward computation, corresponding to the input level where the decoder begins execution. The target remote computing center inputs the received semantic feature data into the corresponding input level of the decoder and controls the decoder to start subsequent forward computation from that input level, processing layer by layer until the decoder's output layer, and finally outputting the task inference result.

[0066] This embodiment determines the target remote computing center based on the computational offloading path decision, transmits semantic feature data through the wireless link specified by the decision, determines the corresponding input level of the decoder based on the semantic segmentation point decision, inputs the semantic feature data into this level, and controls the decoder to perform subsequent forward computation until the task inference result is output. This embodiment realizes the directional transmission of semantic feature data from edge nodes to the remote computing center in a cloud-edge collaborative network, as well as the collaborative docking between the decoder and encoder at the semantic segmentation point, completing the end-to-end task inference process based on semantic communication.

[0067] According to the present invention, a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication updates the parameters of the outer decision network based on task inference results, including: obtaining the task execution latency and inference accuracy corresponding to the task inference results; calculating the comprehensive utility reward value based on the task execution latency and inference accuracy; inputting the comprehensive utility reward value into a centralized evaluation network and calculating the state value function through the centralized evaluation network; calculating the advantage function estimate using the generalized advantage estimation method based on the comprehensive utility reward value and the state value function; calculating the loss function gradient of the action network and the loss function gradient of the centralized evaluation network based on the advantage function estimate; updating the network parameters of the action network using the gradient ascent method based on the loss function gradient of the action network; and updating the network parameters of the centralized evaluation network using the gradient descent method based on the loss function gradient of the centralized evaluation network.

[0068] In one specific embodiment, after the final task inference result is output at the remote computing center, the system obtains the task execution latency and inference accuracy corresponding to the task inference result. The task execution latency includes the transmission latency of semantic feature data from the edge node to the remote computing center and the computation latency of decoding and inference at the remote computing center. The inference accuracy is determined by comparing the decoder output with the true label. The system calculates a comprehensive utility reward value based on the task execution latency and inference accuracy. This comprehensive utility reward value quantifies the overall performance of this task execution. This comprehensive utility reward value is input into a centralized evaluation network, which outputs the state value function corresponding to the current observed state through forward computation. Based on the comprehensive utility reward value and the state value function, the system calculates an estimate of the dominance function using the generalized dominance estimation method. This dominance function estimate measures the superiority or inferiority of the current decision relative to the average level. Based on the calculated dominance function estimate, the system calculates the gradient of the loss function for the action network and the gradient of the loss function for the centralized evaluation network, respectively.

[0069] The system employs gradient ascent to update the network parameters of the action network based on the gradient of the loss function of the action network, thereby increasing the probability of actions with positive advantages. It also employs gradient descent to update the network parameters of the centralized evaluation network based on the gradient of the loss function of the centralized evaluation network, thereby reducing the estimation error of the state value function.

[0070] This embodiment obtains task execution latency and inference accuracy, calculates the comprehensive utility reward value, inputs the comprehensive utility reward value into a centralized evaluation network to obtain the state value function, and uses the generalized dominance estimation method to calculate the dominance function estimate. Based on this, the loss function gradients of the action network and the centralized evaluation network are calculated respectively, and the network parameters of the action network and the centralized evaluation network are updated using gradient ascent and gradient descent methods respectively. This embodiment implements an outer layer decision network parameter update mechanism based on task execution quality in a cloud-edge collaborative network.

[0071] The present invention will be further illustrated below through overall embodiments.

[0072] This invention provides a cloud-edge collaborative network resource allocation method based on semantic communication, implemented through a cloud-edge collaborative network resource allocation system based on semantic communication. This method is used to resolve the "communication bandwidth-physical computing power" conflict and algorithm hard constraint violation problems during massive data backhaul in various resource-constrained and highly dynamic heterogeneous networks (including but not limited to industrial IoT, UAV collaborative swarms, vehicle-to-everything (V2X) networks, and integrated air-space-ground networks). The cloud-edge collaborative heterogeneous network and semantic communication system architecture of this invention is as follows: Figure 2As shown. The various functional modules of this system are physically distributed and deployed in different nodes of the cloud-edge collaborative network (such as edge-side sensing nodes, remote cloud / edge computing centers, etc.), and mainly include the following functional modules: (1) Status and Task Awareness Module: Deployed at edge nodes (such as sensors, drones, satellite terminals, vehicle terminals, etc.). This module is used to acquire the attributes of the currently arriving computing tasks in real time (including raw data dimensions, computational load, latency tolerance, etc.), and collect the status of the local computing queue in real time, as well as the gain and visibility / connectivity information of wireless communication channels (such as edge-to-edge links, edge-to-cloud links), providing real-time network environment observation input for subsequent intelligent decision-making and resource allocation.

[0073] (2) Semantic Feature Extraction Module: Deployed at edge nodes. This module contains the encoder part of a lightweight deep neural network (such as a customized ResNet). This module is used to adaptively process the raw high-dimensional perception data obtained by the perception module according to the "semantic segmentation point" instructions issued by the outer decision module, compress it and extract it into a low-dimensional, compact and task-oriented semantic feature matrix, so as to eliminate a large amount of data redundancy that is irrelevant to the downstream task from the source and significantly reduce the bandwidth overhead of subsequent wireless communication.

[0074] (3) Outer Discrete Decision Module: Deployed on edge nodes. This module runs an Actor network of multi-agent reinforcement learning (such as HMAPPO) combined with an action masking mechanism. It receives the local observation state from the state and task awareness module and outputs a discrete decision combination that maximizes the long-term utility of the system, namely, the calculation offloading path decision (such as choosing local computation, offloading to other edge nodes, or offloading to a remote computation center) and the semantic segmentation point decision. Among them, the action masking mechanism is used to forcibly mask invalid offloading paths that are currently unreachable (or disconnected) by the physical link at the bottom layer of the algorithm.

[0075] (4) Inner Layer Continuous Resource Allocation Module: Deployed in remote computing centers (such as cloud servers, high-performance aggregation nodes, space-based core nodes, etc.). This module is used to aggregate and receive computing requests from all edge nodes that choose to offload tasks to this center, and uses a closed-form convex optimization formula based on KKT conditions to perform deterministic calculations in a centralized manner, thereby accurately allocating continuous computing resources from the remote computing center to each concurrent task. This mechanism strictly guarantees, through rigorous mathematical analytical solutions, that the sum of computing power allocated to all tasks will not exceed the hard constraint upper limit of the underlying physical computing capacity.

[0076] (5) Target Inference Execution and Global Broadcast Module: Deployed in remote computing centers and aggregation nodes with sufficient computing power. This module includes the decoder and classifier / regressor parts of the deep neural network, which are used to receive compact semantic features transmitted by edge nodes through wireless links and complete the final target inference and downstream task processing (such as target recognition, classification, etc.); at the same time, this module is also used to calculate the comprehensive utility reward based on the actual end-to-end latency and inference accuracy of the task, drive the outer reinforcement learning network to iteratively update the model parameters, and broadcast global network load and congestion status indications to the entire network to guide the coordinated scheduling of each edge node in the next time slot.

[0077] Corresponding to the system architecture described above, this invention provides a multi-dimensional resource joint allocation method. The overall steps of this method are as follows: Figure 3 As shown, its core two-layer decoupling algorithm architecture is as follows: Figure 4 As shown, the core of this method lies in employing a decoupled "decision-optimization" two-layer algorithm execution process: the outer layer learns discrete policies through a multi-agent reinforcement learning network, while the inner layer constructs sub-problems for precise allocation of continuous resources. This architecture effectively overcomes the curse of dimensionality and physical constraint overrun problems in reinforcement learning within a mixed-dimensional action space.

[0078] Step S1: Obtain the heterogeneous network state and task attributes, and perform adaptive dynamic semantic feature extraction using a lightweight edge-side neural network. Specifically, considering the high dynamism of the cloud-edge collaborative network topology, each edge node first obtains the current task attributes, local queue status, and channel gain and connectivity / visibility information of various wireless links. In the feature extraction stage, this invention abandons the traditional full-data direct transmission mode and instead adopts a customized lightweight deep neural network as the semantic communication framework. The edge nodes use the front layers of this network as semantic encoders, and according to the segmentation point instructions issued by the outer decision module, map and compress the original high-dimensional sensor data into a low-dimensional, compact semantic feature matrix. The calculation rules are as follows: in, This represents the set of semantic feature data extracted from the k-th edge node at time t. Indicates by parameters Defined semantic encoder function, This represents the raw high-dimensional sensing data acquired by the k-th edge node at time t. For dynamically determined semantic segmentation points, This represents the total number of semantic symbols extracted at this segmentation point. This step eliminates a large amount of redundant information irrelevant to downstream tasks from the source, effectively alleviating communication bandwidth bottlenecks.

[0079] Step S2: Input the local observation state into the outer multi-agent reinforcement learning network, and generate discrete computational offloading paths and semantic segmentation points using an action masking mechanism. This step is executed in parallel by the Actor network distributed across each edge node. The input is the network environment observation state obtained above, and the output is a discrete action combination. .in, Decisions based on semantic segmentation points; For uninstallation path decisions, for example when When the time indicates the option to unload to the edge aggregation center, This indicates that the process has been selected to offload to a remote cloud center across the backhaul link.

[0080] To ensure the effectiveness and feasibility of the decision-making process on the physical link, this invention forcibly superimposes an action masking mechanism at the Softmax output layer of the Actor network. The calculation rules are as follows: in, Indicates by parameters Defined strategy function, This represents the action taken by the k-th edge node at time t. This represents the local observation state of the k-th edge node at time t. This represents the initial log-probability of the action network output with respect to action x. This represents the decision variables for calculating the unloading path. This represents the link connectivity mask vector or visibility indicator of the k-th edge node at time t.

[0081] When a specific wireless communication link is interrupted or goes out of coverage (i.e., visibility / connectivity masking) When the mask term is set to negative infinity, it will shield infeasible paths at the algorithm level, guiding the agent to converge safely and quickly.

[0082] Step S3: Using the closed-form formula of convex optimization based on KKT conditions, continuous physical computing resources are accurately calculated and allocated at the remote computing center. For concurrent tasks selected to be offloaded to computing-constrained nodes, the inner module of this invention avoids using the traditional neural network "trial and error" mechanism for computing power prediction. Instead, the task requirements determined by the outer layer are aggregated by the center. and overall urgency weight The values ​​are directly substituted into the closed-form solution for continuous resource allocation derived in this invention for calculation. The formula for this closed-form solution is as follows: in, To accurately allocate the optimal computing frequency / computing capacity to a specific task, This is the set of all tasks selected at time t to be unloaded to the same remote computing center. Let be the overall urgency weight of the j-th task at time t. For the semantic segmentation points based on the j-th task The determined remaining computational requirements, This represents the upper limit of the total physical computing power capacity available to the remote computing center.

[0083] This step, through rigorous mathematical derivation, not only ensures that the allocation process has extremely low linear computational complexity, but also absolutely guarantees, in terms of physical principles, that the sum of the computing power allocated to all concurrent tasks will not exceed the upper limit of the underlying hardware capacity.

[0084] Step S4: Complete semantic data transmission and remote inference execution, calculate the task-oriented comprehensive utility reward, and update the parameters of the outer reinforcement learning network. Based on the joint decision results of the inner and outer layers, the edge nodes and the remote computing center collaborate via corresponding wireless links to complete the remaining deep neural network decoding and target inference tasks. Subsequently, the system calculates the comprehensive utility reward (Reward) including a congestion penalty term based on the final inference accuracy of the downstream task and whether the maximum tolerable latency is met. The advantage function is calculated using generalized advantage estimation (GAE), and the parameters of the outer Actor and centralized Critic networks are updated using gradient ascent, thereby driving the entire two-layer decoupled architecture to continuously iterate and optimize in interaction with heterogeneous network environments.

[0085] The advantages that can be obtained by using the above method are as follows: (1) Breaking the bandwidth bottleneck: By using a task-oriented semantic communication paradigm, the model segmentation point is dynamically adjusted, which greatly reduces the transmission volume of high-dimensional perception data and effectively alleviates the wireless transmission congestion of massive concurrent data in heterogeneous networks.

[0086] (2) Eliminating the curse of dimensionality: The innovative two-layer decoupled architecture only requires the outer network to explore discrete unloading and splitting strategies, while continuous resource allocation is directly calculated by the inner analytical solution, which effectively overcomes the training difficulties of multi-agent reinforcement learning in mixed-dimensional action space and greatly improves the model convergence efficiency.

[0087] (3) Physical constraint guarantee: The allocation of inner-layer computing resources is directly derived from the KKT closed-form solution, which strictly guarantees the hard constraint boundary of resource allocation at the mathematical level. It effectively avoids the system violation risk of computing power overload and overflow caused by the "random trial and error exploration" of traditional end-to-end neural networks, and shows excellent robustness when facing large-scale sudden tasks and extreme network congestion.

[0088] During simulation verification, this invention built a typical cloud-edge collaborative heterogeneous network simulation environment based on the Gymnasium interface (this embodiment specifically selects the "space-ground integrated telemetry, tracking, and command network" as an extremely resource-constrained test scenario for verification). In this specific implementation scenario, the network topology and channel model referenced the 3GPP TR 38.821 NTN standard. The bandwidth of the wireless backhaul links (i.e., inter-satellite and space-ground links) between the edge and the remote end was strictly limited, and the total computing power capacity of the remote computing center (i.e., the space computing center) was set to 400 GFLOPs. The baseline algorithms for comparison included: the traditional end-to-end hybrid action space reinforcement learning algorithm (HMADRL), as well as the heuristic load-aware algorithm and the random allocation algorithm.

[0089] The first simulation compares the training convergence performance of this invention (HMAPPO) with the baseline algorithm (HMADRL). Figure 5 As shown, because the inner closed-loop solution of this invention strictly guarantees the satisfaction of physical constraints and avoids a large number of invalid explorations, this invention rapidly improves and stabilizes in the high-reward range within a very short training cycle. In contrast, the baseline algorithm relying on the penalty function converges extremely slowly, requiring over 400 rounds to reach a lower performance index, and the process is accompanied by severe oscillations. This result fully demonstrates the efficiency and stability of the two-layer decoupled architecture of this invention.

[0090] The second simulation compares the total system utility and weighted task success rate of the present invention with those of various baseline algorithms under different task arrival rates (network congestion levels). The simulation results are as follows: Figure 6 As shown, under light load, the performance of each algorithm is similar. However, as network load approaches saturation, the baseline algorithm experiences a significant drop in task success rate to approximately 50% or even lower due to severe resource contention and latency timeouts caused by soft penalty mechanisms, resulting in a substantial reduction in system utility. In contrast, this invention can keenly detect network bottlenecks, adaptively select deeper semantic segmentation points to compress transmitted data, and maintain a task success rate of over 90% even under extreme congestion, with overall system utility significantly outperforming existing technologies. This result fully verifies that this invention possesses superior congestion resistance and robustness in resource-constrained extreme environments.

[0091] This invention aims to address the conflict between communication bandwidth and physical computing power in existing cloud-edge collaborative heterogeneous network resource scheduling, as well as the technical shortcomings of traditional pure data-driven reinforcement learning algorithms that easily violate the hard physical constraints of the system. It proposes a multi-dimensional resource dual-layer decoupled joint allocation method based on semantic communication. This invention innovatively introduces a task-oriented adaptive semantic segmentation paradigm, significantly reducing communication bandwidth consumption and transmission latency during the backhaul of massive high-dimensional data by extracting and transmitting only compact semantic features at edge sensing nodes. Simultaneously, to overcome the curse of dimensionality in a mixed-dimensional action space, this invention constructs a dual-layer "decision-optimization" architecture: the outer layer utilizes a multi-agent reinforcement learning network (combined with an action masking mechanism) to make purely discrete decisions on unloading paths and segmentation points, effectively compressing the exploration space and accelerating algorithm convergence under highly dynamic topologies; the inner layer abandons traditional neural network computing power prediction and innovatively adopts a convex optimization closed-form solution based on KKT conditions for precise allocation of continuous computing power at a remote computing center. This joint mechanism not only boasts extremely low linear computational complexity but also rigorously guarantees at the mathematical level that the total computing power allocated to the system will not exceed the underlying physical capacity limit, effectively addressing the pain point that traditional end-to-end algorithms are prone to outputting infeasible solutions. This invention effectively improves the overall system utility, ensuring high reliability and high task success rate for various resource-constrained and highly dynamic network systems under extreme heavy load and congestion environments.

[0092] The following describes the multi-dimensional resource joint allocation device for cloud-edge collaborative networks based on semantic communication provided by the present invention. The multi-dimensional resource joint allocation device for cloud-edge collaborative networks based on semantic communication described below can be referred to in correspondence with the multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication described above.

[0093] like Figure 7 The image shows a cloud-edge collaborative network multi-dimensional resource allocation device based on semantic communication provided by the present invention, comprising: The state awareness module 710 is used to acquire task attributes and channel state information in heterogeneous networks; The semantic feature extraction module 720 is deployed at the edge node and is used to perform adaptive semantic feature extraction on the raw sensing data based on the task attributes and channel state information, using a lightweight neural network deployed at the edge node, to generate semantic feature data. The outer decision module 730, deployed at the edge nodes, is used to input the local observation state into the outer decision network and generate discrete decision combinations by combining the action masking mechanism; the decision combinations include computational unloading path decisions and semantic segmentation point decisions; Resource allocation module 740, deployed in a remote computing center, is used to allocate continuous physical computing resources to concurrent tasks in the remote computing center based on the decision combination and using a closed-form solution based on the Caro-Kuhn-Tucker condition.

[0094] Specifically, the functions of each module in the cloud-edge collaborative network multi-dimensional resource joint allocation device based on semantic communication provided in this embodiment of the invention correspond one-to-one with the operation flow of each step in the above method-like embodiments, and the achieved effects are also the same. For details, please refer to the above embodiments, and this will not be repeated in this embodiment of the invention.

[0095] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840. The processor 810, communications interface 820, and memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication. This method includes: acquiring task attributes and channel state information in the heterogeneous network; based on the task attributes and channel state information, using a lightweight neural network deployed on edge nodes to adaptively extract semantic features from the original sensing data to generate semantic feature data; inputting the local observation state into the outer decision network, and combining it with an action masking mechanism to generate discrete decision combinations; the decision combinations include computational offloading path decisions and semantic segmentation point decisions; and based on the decision combinations, using a closed-form solution based on the Cartesian-Kuhn-Tucker condition at a remote computing center to allocate continuous physical computing resources to concurrent tasks.

[0096] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0097] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication provided by the above methods. The method includes: acquiring task attribute and channel state information in a heterogeneous network; based on the task attribute and channel state information, using a lightweight neural network deployed on edge nodes to adaptively extract semantic features from the original sensing data to generate semantic feature data; inputting the local observation state into an outer decision network and combining it with an action masking mechanism to generate discrete decision combinations; the decision combinations include computational unloading path decisions and semantic segmentation point decisions; and based on the decision combinations, using a closed-form solution based on the Cartesian-Kuhn-Tucker condition to allocate continuous physical computing resources to concurrent tasks at a remote computing center.

[0098] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication provided by the above methods. The method includes: acquiring task attribute and channel state information in a heterogeneous network; based on the task attribute and channel state information, using a lightweight neural network deployed at edge nodes to adaptively extract semantic features from the original sensing data to generate semantic feature data; inputting the local observation state into an outer decision network and combining it with an action masking mechanism to generate discrete decision combinations; the decision combinations include computational unloading path decisions and semantic segmentation point decisions; and based on the decision combinations, allocating continuous physical computing resources to concurrent tasks at a remote computing center using a closed-form solution based on the Caro-Kuhn-Tucker condition.

[0099] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0100] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and these 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 the present invention.

Claims

1. A multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication, characterized in that, include: Obtain task attributes and channel state information in heterogeneous networks; Based on the task attributes and channel state information, a lightweight neural network deployed at the edge nodes is used to adaptively extract semantic features from the original sensing data to generate semantic feature data. The local observation state is input into the outer decision network, and combined with the action masking mechanism, discrete decision combinations are generated; the decision combinations include calculation of unloading path decisions and semantic segmentation point decisions; Based on the aforementioned decision combination, continuous physical computing resources are allocated to concurrent tasks at the remote computing center using a closed-form solution based on the Caro-Kun-Tucker condition.

2. The multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication according to claim 1, characterized in that, The step involves using a lightweight neural network deployed at edge nodes to adaptively extract semantic features from the original sensing data based on the task attributes and channel state information, generating semantic feature data, including: Obtain semantic segmentation point instructions issued by the outer decision network; The forward processing layer of the lightweight neural network is determined based on the semantic segmentation point instructions; The raw sensing data is input into the encoder part of the lightweight neural network, and forward calculation is performed layer by layer according to the determined forward processing level; After reaching the network level specified by the semantic segmentation point instruction, the forward computation is stopped and the output of the current level is extracted as semantic feature data; The semantic feature data has a lower dimension than the original perceptual data, and the task-related feature density in the semantic feature data is higher than that in the original perceptual data.

3. The multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication according to claim 1, characterized in that, The step of inputting the local observation state into the outer decision network and combining it with an action masking mechanism to generate discrete decision combinations includes: The local observation state is input into the action network of multi-agent reinforcement learning; Through the forward computation of the action network, the initial log probability corresponding to each candidate unloading path and each candidate semantic segmentation point is output. Obtain the connectivity status of each wireless communication link at the current moment, the connectivity status including the visibility indication of edge-to-edge links and edge-to-cloud links; Based on the connectivity status, an action mask vector is generated for the candidate offloading path, wherein the mask value corresponding to the link with a connectivity status of interruption or unreachable is set to negative infinity. The action mask vector is superimposed with the initial log probability to obtain the corrected log probability distribution; Perform a flexible maximum normalization operation on the modified logarithmic probability distribution to obtain the probability distribution of effective actions; Based on the probability distribution of the effective actions, random sampling or greedy selection is performed to determine the computational unloading path decision and semantic segmentation point decision at the current moment.

4. The multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication according to claim 1, characterized in that, The process of allocating continuous physical computing resources to concurrent tasks at a remote computing center, based on the decision combination and using a closed-form solution based on the Caro-Kuhn-Tucker condition, includes: At the remote computing center, all tasks that have been selected to be offloaded to the remote computing center are aggregated; For each aggregated task, obtain the semantic segmentation point decision corresponding to the current task, and determine the remaining computational requirements of the current task based on the semantic segmentation point decision; Obtain the overall urgency weight of each task; Substitute the remaining computational requirements and the overall urgency weight of each task into the convex optimization closed-form formula based on the Caro-Kun-Tucker condition to calculate the optimal computing power allocation value for each task. Based on the optimal computing power allocation value, allocate corresponding continuous physical computing power resources to each task; The convex optimization closed-form solution formula uses the product of the square root of the remaining computing demand of each task and the square root of the comprehensive urgency weight as the numerator, the sum of the numerators of all converged tasks as the denominator, and then multiplies it by the upper limit of the total physical computing power capacity of the remote computing center to obtain the optimal computing power allocation value for each task.

5. The multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication according to claim 1, characterized in that, After allocating continuous physical computing resources, the method further includes: Based on the allocated physical computing resources, the semantic feature data is transmitted to a remote computing center for task reasoning to obtain the task reasoning result; The parameters of the outer decision network are updated based on the task reasoning results.

6. The multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication according to claim 5, characterized in that, The step of transmitting the semantic feature data to a remote computing center for task inference based on the allocated physical computing resources includes: Based on the computation unloading path decision, the target remote computing center corresponding to the semantic feature data is determined; The semantic feature data is transmitted to the target remote computing center via the wireless link specified by the computation offload path decision. At the target remote computing center, the corresponding input layer of the deep neural network decoder is determined based on the semantic segmentation point decision; The received semantic feature data is input to the corresponding input level of the decoder; The decoder is controlled to perform subsequent forward computations starting from the corresponding input level until the final task inference result is output.

7. The multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication according to claim 5, characterized in that, Updating the parameters of the outer decision network based on the task inference result includes: Obtain the task execution latency and inference accuracy corresponding to the task inference results; Calculate the comprehensive utility reward value based on the task execution delay and the inference accuracy. The comprehensive utility reward value is input into a centralized evaluation network, and the state value function is calculated through the centralized evaluation network. Based on the comprehensive utility reward value and the state value function, the generalized advantage estimation method is used to calculate the advantage function estimate. Based on the estimated advantage function, calculate the gradient of the loss function of the action network and the gradient of the loss function of the centralized evaluation network; The gradient ascent method is used to update the network parameters of the action network according to the gradient of the loss function of the action network; The gradient descent method is used to update the network parameters of the centralized evaluation network according to the gradient of the loss function of the centralized evaluation network.

8. A cloud-edge collaborative network multi-dimensional resource allocation device based on semantic communication, characterized in that, include: The state awareness module is used to acquire task attributes and channel state information in heterogeneous networks. The semantic feature extraction module, deployed at the edge node, is used to adaptively extract semantic features from the original sensing data based on the task attributes and channel state information, using a lightweight neural network deployed at the edge node, to generate semantic feature data. The outer decision module, deployed at the edge nodes, is used to input the local observation state into the outer decision network and generate discrete decision combinations by combining the action masking mechanism; the decision combinations include calculating the unloading path decision and the semantic segmentation point decision. The resource allocation module, deployed in a remote computing center, is used to allocate continuous physical computing resources to concurrent tasks in the remote computing center based on the decision combination and using a closed-form solution based on the Caro-Kuhn-Tucker condition.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-dimensional resource joint allocation method for cloud-edge collaborative networks based on semantic communication as described in any one of claims 1 to 7.

Images