Edge network resource arrangement system in holographic communication zhi rong identification network and point cloud service arrangement method based on graph attention network

Abstract

This invention provides an edge network resource orchestration architecture and a point cloud service orchestration method based on graph attention networks in a holographic communication intelligent identification network. The architecture includes: an infrastructure network layer for fusing computing and transmission resources based on the data processing and transmission capabilities of the physical infrastructure; a decision control layer for generating joint optimization strategies based on real-time network status; and a holographic service layer for aggregating different types of holographic service requests and triggering orchestration strategies based on the characteristics of the holographic service requests. This invention can solve the problem of low-latency and high-reliability orchestration of existing holographic services in resource-constrained edge networks.

Description

Technical Field

[0001] This invention relates to the field of holographic communication technology, and in particular to an edge network resource orchestration architecture and a point cloud service orchestration method based on graph attention networks in a holographic communication intelligent identification network. Background Technology

[0002] Holographic communication, as a cutting-edge VR technology carrier that integrates the virtual and real worlds, demonstrates significant value in applications such as remote diagnosis and treatment and smart cities by creating highly immersive interactive environments. However, the technical requirements of this type of service have exceeded the performance limits of traditional multimedia services. Taking point cloud video as an example, when a single frame contains millions of point clouds, the uncompressed transmission bandwidth requirement is as high as 3.6Gbps, far exceeding that of conventional streaming media services, and requiring more computing resources and time for processing. At the same time, to avoid user dizziness, the motion-to-image latency must be controlled within 20ms, posing a dual challenge to network bandwidth, latency, and video quality parameters. Point cloud services are both highly computationally intensive and strictly latency-sensitive. Terminal devices are limited by their own computing power and cannot independently complete complex processing tasks, while the cloud cannot meet real-time requirements due to backhaul link latency. Although edge computing nodes can support video transcoding, encoding and decoding to enhance real-time streaming media services, their limited computing power is insufficient to handle the pipeline processing of large-scale point cloud service (PCS) requests. In resource-constrained edge networks, application services compete for limited computing and transmission resources, thus requiring the use of distributed computing resources and dynamic orchestration of network slices to provide deterministic transmission guarantees.

[0003] Existing research on point cloud video (PCV) mainly focuses on encoding method optimization (such as Draco, V-PCC, and G-PCC) and transmission protocol design (such as DASH), but there are still significant shortcomings in computational offloading. Traditional computational offloading research is mostly aimed at independent video tasks, which is difficult to meet the collaborative scheduling requirements of the end-to-end transmission chain of holographic services. PCV transmission involves multiple computationally intensive operations with dependencies. These operations need to dynamically adjust processing parameters according to features such as field of view (FoV) and point density. A fixed processing ratio will cause QoE loss or resource waste. At the same time, the amount of data processed is closely related to the computational demand, and the amount of processed data affects the transmission bandwidth demand, forming a mutually restrictive relationship between computational latency and transmission latency. The limited computing power of edge nodes needs to support the sharing of resources for multiple services. Inappropriate bandwidth allocation can easily lead to network congestion and latency spikes, directly affecting user experience quality and operator costs (OPEX). In summary, traditional algorithms are not adaptable enough to complex PCS architectures and dynamic network environments, and cannot fully utilize point cloud features and network state information for high-quality decision-making. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide an edge network resource orchestration architecture and a point cloud service orchestration method based on graph attention networks in a holographic communication intelligent identification network.

[0005] To achieve the above objectives, the present invention adopts the following technical solution.

[0006] In a first aspect, the present invention provides an edge network resource orchestration architecture in a holographic communication intelligent identification network, the architecture comprising:

[0007] The infrastructure network layer is used to integrate computing and transmission resources based on the data processing and transmission capabilities of the physical infrastructure.

[0008] The decision control layer is used to generate joint optimization strategies based on the real-time collected network status.

[0009] The holographic service layer is used to aggregate different types of holographic service requests and trigger orchestration strategies based on the characteristics of the holographic service requests.

[0010] In some embodiments of the present invention, the architecture further includes:

[0011] The point cloud service model is used to decompose the received point cloud service into multiple sub-tasks to form a task chain; and to obtain the perception information of the point cloud service based on the task chain, point cloud frame group information, consumption information, latency, and processing level.

[0012] In some embodiments of the present invention, the infrastructure network layer is a directed connected graph, which includes a set of servers characterized by computing resources and a set of physical links characterized by load bandwidth resources.

[0013] In some embodiments of the present invention, triggering the orchestration strategy based on the characteristics of the holographic service request includes:

[0014] The joint orchestration process of holographic services is modeled as a Markov decision process (MDP), and then processed using quadruples. It means that, among them, For state space, For the action space, For the transformation probability, This is the reward function.

[0015] Secondly, the present invention also provides a point cloud service orchestration method based on a graph attention network, the method being implemented based on the edge network resource orchestration architecture in the holographic communication intelligent identification network of the first aspect, the method comprising:

[0016] The original state of the edge network is transformed into a structured graph by orchestrating neural networks using the point cloud service PCS.

[0017] A new feature vector is generated by fusing neighbor context information through a graph attention convolutional layer, and the new feature vector is then subjected to a nonlinear transformation by an activation function to obtain the transformed result.

[0018] The transformed result is enhanced by a multi-head attention mechanism to obtain enhanced features, and the final point cloud service orchestration strategy is obtained by splicing and averaging.

[0019] In some embodiments of the present invention, the point cloud service PCS orchestration neural network is obtained through training aided by the PPO algorithm with generalized advantage estimation (GAE) and pruning of surrogate target (CSO).

[0020] In some embodiments of the present invention, the point cloud service PCS orchestration neural network is trained using the following steps:

[0021] Initialize the parameters of the Actor network and Critic network, and initialize the experience replay buffer;

[0022] In each loop, the agent interacts with the environment, constructing a graph representation of the network state under the currently arriving PCS request and sampling actions.

[0023] Under the premise of satisfying the constraints, the action is executed and feedback is obtained, and the graph structure observation trajectory is stored in the experience playback buffer;

[0024] The network parameters are updated using stochastic gradient descent based on a batch of samples randomly selected from the experience replay buffer based on the advantage function calculated by GAE, until all point cloud service PCS orchestration is completed.

[0025] Thirdly, the present invention also provides a point cloud service orchestration device based on a graph attention network, the device being implemented based on the edge network resource orchestration architecture in the holographic communication intelligent identification network of the first aspect, the device comprising:

[0026] The first computing module uses the point cloud service PCS to orchestrate the neural network and convert the original state of the edge network into a structured graph.

[0027] The second calculation module fuses neighbor context information through graph attention convolutional layers to generate new feature vectors, and performs a non-linear transformation on the new feature vectors through an activation function to obtain the transformed result.

[0028] The third calculation module enhances the features of the transformed result through a multi-head attention mechanism to obtain enhanced features, and obtains the final point cloud service orchestration strategy through splicing and averaging.

[0029] Fourthly, 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 program to implement the method described above.

[0030] Fifthly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0031] The beneficial effects of this invention are as follows: The edge network resource orchestration architecture and point cloud service orchestration method based on graph attention networks provided by this invention in the holographic communication intelligent fusion identification network, wherein the architecture supports multi-task chain computation offloading, routing planning and adaptive processing to guide holographic services in resource-constrained environments. Specifically, this invention models the multi-task computation offloading and video processing rate adjustment problem as a stochastic optimization problem considering computation and transmission coordination, and maximizes system utility by jointly optimizing network satisfaction and user QoE utility. In addition, it uses Lyapunov theory to decouple the original long-term global constraint problem into local decision problems for each time slot, and further transforms it into an MDP, combining dynamic perception and reinforcement learning to achieve multi-objective collaborative optimization, eliminate control delay, and improve the quality of personalized services. In addition, a deep reinforcement learning algorithm based on graph attention networks, GCNCO, is proposed: it focuses on key issues such as multi-task chain dependency, dynamic resource allocation, and collaborative optimization of computation and transmission for holographic services in edge networks. GCNCO significantly improves the orchestration capability of holographic services, system utility and user experience quality, obtains the mapping association between nodes and links in the network, fully perceives the status of computing network resources and holographic services, and realizes multi-standard decision analysis.

[0032] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a schematic diagram of the edge network resource orchestration architecture in the holographic communication intelligent identification network provided in an embodiment of the present invention;

[0035] Figure 2 This is a flowchart illustrating the point cloud service orchestration method based on graph attention networks provided in an embodiment of the present invention. Detailed Implementation

[0036] Computation offloading is a technique that transfers computational tasks from resource-constrained devices (such as mobile terminals) to the cloud or edge servers for processing, thereby improving computational efficiency and saving local resources. Yuan et al. designed the AIRA framework, which, combined with improved multi-agent reinforcement learning and smart contract technology, achieved collaborative optimization of computation offloading, resource allocation, and video preprocessing in a blockchain environment, improving system credibility and operational efficiency. Luo et al. developed a video streaming optimization framework based on the A3C algorithm, which effectively optimized energy consumption and QoE metrics in software-defined mobile networks and mobile edge computing environments by comprehensively considering factors such as buffer dynamics, adaptive video quality, edge caching, and transcoding.

[0037] Field of view (FoV) prediction enhancement is a technique that optimizes and enhances the field of view of images or videos through algorithms to improve visual experience and information acquisition efficiency. Sun et al. proposed using the local storage advantages of mobile edge computing (MEC) to pre-cache key content frequently accessed by users, and forming a collaborative link of "edge caching-terminal processing" through post-processing technology of terminal devices, thereby shortening the content transmission distance and computing response time, and effectively reducing end-to-end latency. Meanwhile, Perfecto et al. designed a gated recurrent unit-deep recurrent neural network model to model the spatiotemporal features such as FoV trajectory and attention distribution, targeting the dynamic changes in user field of view (FoV) in scenarios such as virtual reality (VR). Combined with multicast transmission mechanism, the content distribution strategy is dynamically adjusted to maximize user experience quality (QoE) while ensuring low-latency transmission, providing a new technical path for network resource optimization in dynamic interactive scenarios.

[0038] QoE-driven dynamic adjustment technology is a QoE-based feedback technology used to adjust system parameters in real time to optimize service quality and user satisfaction. Li et al. proposed an adaptive encoding and offloading strategy for point cloud video streams, addressing the high-dimensional, high-density data characteristics. This strategy dynamically adjusts the encoding resolution and compression rate of the point cloud by analyzing the computing power and network bandwidth of user devices, and intelligently offloads some encoding tasks to edge servers or the cloud, reducing terminal processing load and transmission traffic while ensuring the semantic integrity of the video. Zhang et al. constructed a collaborative framework for streaming media transmission and super-resolution processing, designing a cross-layer optimization algorithm to dynamically schedule the transmission rate and resolution enhancement strategy of the video stream, effectively balancing real-time playback quality and network transmission efficiency.

[0039] The above solution has the following disadvantages:

[0040] Existing computational offloading technologies focus on resource allocation optimization for traditional independent tasks, or on simple task offloading decisions, but they often overlook the collaborative adaptation among multiple factors. In complex and ever-changing application scenarios, dynamic changes in task requirements and uncertainties in resource supply can lead to resource waste and inefficiency. This limitation makes it difficult for existing technologies to meet the comprehensive requirements of emerging businesses for computing performance, energy consumption, and real-time performance, and hinders efficient deployment and widespread application in diverse real-world scenarios. Therefore, this invention proposes a novel multi-task chain computational offloading and orchestration framework, PCSs, to ensure efficient resource allocation in holographic communication.

[0041] Existing field-of-view (FoV) prediction enhancement technologies, some based on historical data to build statistical models, struggle to adapt to the randomness of user behavior in real-time interactive scenarios and cannot handle sudden viewpoint shifts and dynamic scene transitions. Others employ deep learning-based dynamic prediction, but these fail to accurately capture instantaneous changes in user attention. Furthermore, VR / AR scenarios involve large volumes of visual data and demanding real-time rendering; the conversion between prediction results and rendering instructions is prone to timing misalignments, leading to screen delays and stuttering. Traditional latency compensation mechanisms, while mitigating the problem through preloading content, consume excessive computing resources for pre-rendering due to the large amount of data generated by dynamic FoV changes, failing to meet the stringent real-time and computing resource requirements of mobile VR / AR devices. Therefore, this invention proposes GCNCO, which achieves deep collaboration between FoV prediction and rendering processes by perceiving task relevance and resource allocation.

[0042] Existing QoE-driven dynamic control technologies suffer from a time lag between data processing and control command execution in real-time business scenarios, easily leading to control delays and causing issues such as screen stuttering and service interruptions. Traditional feedback compensation mechanisms, when balancing multiple objectives (such as latency, energy consumption, and image quality), often compromise on some aspects, failing to meet users' comprehensive needs for high-quality, low-latency, and personalized service experiences. Therefore, this invention constructs a stochastic optimization model, transforming it into a Multi-Purpose Optimization (MDP) model using Lyapunov optimization, and combines dynamic perception and reinforcement learning to achieve multi-objective collaborative optimization, eliminating control delays and improving the quality of personalized services.

[0043] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0044] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, and / or modules, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0045] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0046] Before introducing the edge network resource orchestration architecture and point cloud service orchestration method based on graph attention network in a holographic communication intelligent identification network provided by the embodiments of the present invention, the following explanations are provided:

[0047] Holographic-type communication (HTC): Holographic communication is a communication technology that provides users with highly immersive, multi-dimensional interactive application scenarios.

[0048] Point Cloud Service (PCS): Point Cloud Service is an architectural pattern that enables direct interaction between terminal devices and cloud resources.

[0049] Point Cloud Video (PCV): Point cloud video is dynamic visual data composed of continuous three-dimensional point cloud frames, each frame containing a set of spatial coordinate information with a specific timestamp.

[0050] Field of View (FoV): The field of view is the effective observation range of an optical system or biological vision in space.

[0051] Quality of Experience (QoE): Quality of experience is a comprehensive evaluation of user satisfaction with a service, quantified through subjective perception indicators.

[0052] Quality of Service (QoS): QoS is the optimization and control of network transmission parameters achieved through technologies such as traffic prioritization and resource reservation.

[0053] Markov Decision Process (MDP): The Markov Decision Process is designed to solve the problem of making decisions in uncertain environments.

[0054] GraphAttention Mechanism driven Computing-Network Collaboration based ProximalPolicy Optimization Orchestration Algorithm (GCNCO): The graph attention mechanism-driven computation-network collaboration-based algorithm achieves collaborative scheduling of computation and network resources through proximal policy optimization.

[0055] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0056] Example 1

[0057] See Figure 1 This invention provides an edge network resource orchestration architecture in a holographic communication intelligent identification network, which includes an infrastructure network layer, a decision control layer, and a holographic service layer.

[0058] The infrastructure network layer integrates computing and transmission resources based on the data processing and transmission capabilities of the physical infrastructure. In other words, based on hardware facilities with data processing and transmission capabilities, the infrastructure network layer integrates computing power and communication resources, specifically designed for real-time processing and efficient transmission in 3D point cloud applications. For complex computational demands, a dynamic scheduling algorithm matches the optimal computing units and transmission channels from a heterogeneous resource library based on real-time task characteristics and node load. Employing a resource-aware strategy, it monitors node computing power and network bandwidth in real time, constructs a multi-dimensional resource evaluation model, and achieves collaborative optimization of computing load and transmission paths, ensuring minimal service response latency while optimizing infrastructure scalability and network resource utilization.

[0059] The decision control layer is used to generate joint optimization strategies based on real-time network status data. As the core intelligent module, the decision control layer includes status monitoring, strategy optimization, and dynamic orchestration functions. Based on real-time network status data (computing power, link bandwidth, queue backlog) and holographic service request characteristics (such as FoV and point density), a joint optimization strategy is generated using the GCNCO algorithm, including task offloading decisions, processing ratio adjustments, and transmission path planning.

[0060] The holographic service layer aggregates different types of holographic service requests and triggers orchestration strategies based on the characteristics of these requests. Taking HCS1 as an example, point cloud data dynamically changes according to the user's field of view and is encoded into multiple quality levels to meet different QoS and network conditions. Simultaneously, some services involve sub-tasks, such as Simultaneous Localization and Simultaneous Mapping (SLAM), which are dynamically mapped to edge nodes for collaborative processing. Finally, service requests are submitted to the decision control layer for global resource-aware orchestration and strategy deployment.

[0061] In some embodiments, the architecture further includes a point cloud service model, which is used to decompose the received point cloud service into multiple sub-tasks to form a task chain; and to obtain the perception information of the point cloud service based on the task chain, point cloud frame group information, consumption information, latency, and processing level.

[0062] In this embodiment, the system provides high-quality, low-latency PCS to user u∈U, and the contextual characteristics of this type of service are complex and diverse. In the application scenario, PCV is composed of Group of Frames (GoF). u This represents the sequence of frame groups for user u, where each frame group contains |g u | Frame, frame group index is g u ,in The features of points in a point cloud include three-dimensional coordinates (x, y, z) and corresponding color values. Each coordinate value occupies 4 bytes of storage space, and each primary color component (red, green, blue) occupies 1 byte of storage space. Because the point cloud features vary for different users' FoV information, the initial data size of the point cloud frame group is set to e. u,0,g During transmission to the user end, PCV is processed sequentially by the computing functions to reduce service latency while ensuring QoE. During transmission, the point cloud service is broken down into multiple dependent, computationally intensive tasks, forming a delivery chain, described as follows: For example, a PCS can be decomposed into multiple atomic computational functions, including registration, compression, decompression, reconstruction, and rendering. Furthermore, offloading computational tasks to the server leads to increased consumption of computing resources; therefore, C++ is used... u,f,gThis represents the computational resource consumption per voxel unit. Simultaneously, PCV transmission incurs bandwidth resource consumption, therefore b is used. u,f,g This indicates bandwidth consumption. PCV has high latency requirements, therefore T is used. u - This indicates the overall maximum tolerable latency for the service.

[0063] The different data volumes before and after processing by various computational tasks in PCV represent different quantization scales. Since the degree of PCV processing not only reduces the transmission rate but also decreases the processing efficiency of subsequent tasks, to ensure a fast and timely point cloud delivery chain, the optional processing level of the PCS computational function f in u is defined as α. u,f For the calculation function f, the target processing ratio of frame group g is d. u,f,g , where d u,f,g = [0,1]. Because the point cloud service is broken down into a series of ordered computation tasks, the computation task f+1 constituting the point cloud service u will only begin processing PCV after the preceding task f has been completed. Therefore, the data processed by task f is the input of task f+1, and the load after task f is executed satisfies e′. u,f,g =α u,f (d u,f,g )e u,f,g +∈, where, e u,f,g This indicates that computation task f has not been executed, and e is present. u,f+1,g =e′ u,f,g $.$α u,f (·) represents the nonlinear complexity function of computation task f, and ∈ represents the nonlinear quantization error of the keyframe. Therefore, the perceptual information of the point cloud service is

[0064] In some embodiments, the infrastructure network layer is a directed connected graph, which includes a set of servers characterized by computing resources and a set of physical links characterized by load bandwidth resources.

[0065] In this embodiment, to monitor the network status information of point cloud content during computation and network transmission, the system timeline is divided into a set of discrete time slots, denoted by T = {1,...,t,...,|T|}. Each time slot's actions prepare for the subsequent time slot. First, the point cloud content to be transmitted in the previous time slot is sent to the server via the selected path. Then, the current server performs necessary point cloud content computation operations, such as compression / decompression, encoding / decoding, and other computationally intensive tasks. Finally, the processed content is transmitted to the next hop node and prepared to enter the next execution stage.

[0066] For the infrastructure network layer, we define the infrastructure network as a directed connected graph G = {V, ε}, where It is a collection of servers characterized by computing resources. This is a set of physical links characterized by load bandwidth resources, where |·| is the cardinality of the set, and e = (v, v′) ∈ ε, v, v′ ∈ V. To ensure the network transmission of large amounts of point cloud content, computing operations for processing point cloud content are configured on servers in the infrastructure network. Servers within the network each carry limited computing resources and are responsible for processing point cloud data. The computing resources configured on server v ∈ V in time slot t are... The amount of transmission resources configured for link e∈ε in time slot t is

[0067] In some embodiments, triggering the orchestration strategy based on the characteristics of the holographic service request includes:

[0068] The joint orchestration process of holographic services is modeled as a Markov decision process (MDP), and then processed through quadruples. It means that, among them, For state space, For the action space, For the transformation probability, This is the reward function.

[0069] In this embodiment, in order to optimize PCS routing, computation offloading, and processing ratio selection, and to maximize system energy efficiency while ensuring holographic service latency, the following definitions are provided.

[0070] (1) User QoE Utility: Because the higher the processing level, the closer the processed and analyzed PCV will be to the original PCV, that is, if the data volume is larger, the video quality of the display terminal after a series of processing and analysis procedures on the PCV will be better. Therefore, User QoE is related to the point cloud data processing ratio and is a monotonically increasing function.

[0071]

[0072] in, The weight matrix represents the user's QoE, and the utility of the function f is φ. u,f (t)=log2(1+α u,f (d u,f,g (t)).

[0073] (2) Network Satisfaction: To address the resource consumption issues of computing tasks on nodes and links, the following measures are used: and The latency and operational costs of holographic service transmission are used as weights to quantify the latency and operational costs. Based on the heterogeneous characteristics of PCS, network satisfaction is defined as the weighted sum of cost and latency under computation offloading and path planning:

[0074]

[0075] The difference U(t) between system utility identifier user QoE utility and network satisfaction in time slot t is: In future scenarios with limited knowledge, end users will need to consider various factors and make optimal decisions, specifically including the selection of computational methods, routing decision analysis, and processing ratio adjustment. Therefore, an objective function is set. The objective function aims to make optimal decisions for all point cloud holographic services in the current time slot t, minimizing the negative impact on system utility and thus effectively maximizing overall system utility. However, the network is subject to resource constraints and is dynamically changing. Therefore, solving the above problem P1 without prior knowledge of the system state and queue backlog is challenging. Furthermore, the selection of an orchestration strategy leads to an imbalance in computational processing queues and transmission costs within the network. Therefore, Lyapunov optimization methods are used to regulate the selection strategy. The main purpose is to reduce the negative impact on the average system utility while maintaining the stable state of nodes within the network regarding processing and transmission queues. This decomposes the long-term optimization problem into a single-time-slot optimization problem without relying on any prior knowledge.

[0076] Therefore, a computation processing queue is introduced to simulate the dynamic processing flow of PCV tasks. The time-slot task processing queue is as follows:

[0077]

[0078] Simultaneously, a transmission queue is introduced to simulate the PCV transmission process. The time-slot transmission of PCV is as follows:

[0079]

[0080] Let Z(t) = [Λ v (t),Π e [(t)] is the concatenated vector of all queues in time slot t, and and The quadratic Lyapunov function L(Z(t)) is half the sum of squares of the current queue backlog, serving as an index to quantify the queue backlog Z(t) and indicate its stability. L(Z(t)) is... In any time slot t, the conditional expectation Lyapunov drift Δ(L(t)) is the expectation of the difference between two consecutive time slots of the Lyapunov function, i.e., EL(t+1) - L(t)|Z(t). To ensure stable processing and transmission queues in the network and optimize system utility, the drift-plus-penalty technique optimized by Lyapunov is needed, i.e., Δ(L(t)) - VE U(t)|Z(t), where the non-negative constant V is used to minimize the balance between penalty and system stability. A general rule is that a larger V value results in better system utility. However, this can lead to high-level oscillations in queue backlog. By appropriately selecting V, the drift-plus-penalty decoupling time variables in each time slot are minimized, while the negative impact on long-term system utility is gradually minimized by stabilizing the queues. The upper bound for minimizing drift-plus-penalty is given by the following lemma.

[0081] For any queue backlog and orchestration action:

[0082] Δ(L(t))-VE[U(t)]≤BE VU(t)-| υ Λ υ (t)(c υ (t)-C υ )|+ e Π e (t)(b e (t)-B e (5)

[0083] in, and It represents the maximum processing capacity of node v and the maximum transmission capacity of link e across all time slots.

[0084] To address the aforementioned optimization problem, the joint orchestration process of point cloud holographic services is modeled as a Markov decision process (MDP), and then implemented using quadruples. The specific parameters are as follows:

[0085] (1) State space The system state s(t) consists of PCS request, server computation state, network transmission state, and their corresponding queue states, i.e. Specifically

[0086] PCS Information:

[0087]

[0088] Server computing and transmission status:

[0089]

[0090] Network transmission status:

[0091]

[0092] (2) Action space In the current state s(t), the action a(t) is a multidimensional discrete process that includes processing ratio control, computational task offloading, and transmission path selection. For each PCSu, its action is defined as:

[0093]

[0094] (3) Transition probability This indicates that action a is taken in time slot t. t Then by state s t Transition to state s t+1 The probability pr(s) t+1 |s t ,a t ).

[0095] (4) Reward function Map the state-action pair to a value r(s) t ,a t The immediate reward of the system is represented as the negative value of the objective function U′(t), denoted as r(s). t ,a t )=-U′(t). Therefore, maximizing the cumulative reward is equivalent to minimizing the negative impact on system utility. If the scheduling strategy can improve system utility and meet the maximum latency requirements of the point cloud holographic service, U′(t) will approach a smaller value, and the reward will approach a larger value; otherwise, the reward will decrease.

[0096] (5) Discount factor: 0≤γ≤1 indicates the difference in importance between the future reward function value and the current reward function value.

[0097] In some embodiments, whether a frame group g within a single time slot t requires function f analysis is used. It indicates. Among them, It means to execute the computation function f. The computation function f is not executed. Furthermore, computational tasks offloaded to the network server selectively process PCV. Task offload matrix o u (t) is defined as the unloading status information of all computing tasks of service u within time slot t, where o u,f,υ (t) is an indicator variable. If o u,f,υ If (t) = 1, then the computation task f is offloaded to the edge server v for execution in time slot t; otherwise, o u,f,υ(t) = 0. Computational tasks are hosted on a server, and in each time slot t, the task offloading matrix needs to ensure that PCV is processed by only one server's computation function in any given time slot t. Simultaneously, the total computational resource consumption of all computational tasks offloaded to server v cannot exceed the server's maximum computational capacity C. v (t). The controller decides whether to process and analyze the GoF transmitted in the current time slot; if o u,f,v (t) = 1 and So the computing device needs to calculate g u The original version is converted to its target processing ratio d u,f,g For point cloud frames, the processing and analysis time of the computation task, the allocated computing resources, the number of points in the frame, and the processing and analysis level all depend on the computational task. Therefore, the total computational latency... for in, User u selects a processing ratio of α u,f The computational task is designed to handle the computational latency required for processing frame group g. Furthermore, the computational resource consumption is... Among them, κ v This is a cost factor. The computing resources required for PCS user u to execute at maximum processing capacity are C′. v Given limited resources, to ensure carrying capacity in the future, resource consumption needs to be limited, and the average computational resource constraint over time is...

[0098] For computation functions f and f+1, there are sets of candidate paths on servers v and υ′, respectively. After function f processes the PCV stream, it determines the path based on the selected path. Upon reaching server υ′, function f+1 is executed. For the entire delivery chain of PCS user u, the selected path's transmission sequence is as follows: In the delivery chain, the virtual path from function f to function f+1 is mapped to the physical path from v to υ′. The binary decision variable y is set. u (t) represents the link mapping state. Meanwhile, the total bandwidth consumed by traffic along the transmission path cannot exceed the maximum transmission capacity of the physical path; the relevant constraint is... in, This refers to the computational resources consumed in wired or wireless transmission via physical link e during time slot t. Furthermore, the transmission delay is... Therefore, the total transmission delay is In addition, the transmission cost is in, This is the transmission cost coefficient from server v to v′. Set B′. eTo transmit the transmission capabilities originally possessed by the GoF, the time-dependent average transmission resource constraint is...

[0099] The edge network resource orchestration architecture in the holographic communication intelligent identification network provided in this invention maximizes system utility (combining user QoE utility and network satisfaction) through collaborative optimization of adaptive video processing, multi-task offloading, resource allocation, and transmission path selection. It models point cloud services (PCS) as multi-task dependency chains, dynamically adjusting processing ratios (such as compression and rendering levels) to balance computational latency and transmission bandwidth requirements, thus solving the holographic service orchestration challenge in resource-constrained environments.

[0100] Example 2

[0101] Building upon Example 1, Example 2 provides a point cloud service orchestration method based on graph attention networks. This method requires implementation based on the edge network resource orchestration architecture within the aforementioned holographic communication intelligent identification network. This example addresses resource contention and bandwidth constraints caused by uneven distribution of computing power among edge network servers, revealing that the local state of the servers can significantly impact the offloading and transmission performance of point cloud tasks. Based on this, a near-end policy optimization (GCNCO) framework incorporating a graph attention mechanism is designed. The policy network input of this framework uses an attention mechanism to dynamically aggregate the resource feature representations of neighboring nodes, and the policy update magnitude is controlled by the truncation ratio objective function of the PPO. The goal of this framework is to achieve global scheduling optimization for multiple PCS in resource-constrained environments.

[0102] like Figure 2 As shown, the point cloud service orchestration method based on graph attention networks includes the following steps:

[0103] S101 uses the point cloud service PCS to orchestrate neural networks and transform the original state of the edge network into a structured graph.

[0104] S102, a new feature vector is generated by fusing neighbor context information through a graph attention convolutional layer, and the new feature vector is then subjected to a nonlinear transformation through an activation function to obtain the transformed result.

[0105] S103, the transformed result is enhanced by a multi-head attention mechanism to obtain enhanced features, and the final point cloud service orchestration strategy is obtained by splicing and averaging.

[0106] In this embodiment, the GCNCO framework utilizes the improved GAT (i.e., GATv2) to evaluate link and computing resource status and formulate point cloud flow processing analysis and transmission optimization strategies. GAT is an advanced neural network designed for efficient processing of graph-structured data. In GAT, weights between different neighbors are adaptively assigned based on feature similarity and context, enhancing the model's information aggregation ability and robustness.

[0107] First, GAT transforms the original state of the edge network into a structured graph G' = (V', ε'), where Let ε represent the node set, and ε′ represent the edge set. The feature vectors associated with all nodes are represented as follows: Among them, h i Let F represent the feature vector of node i, and let F represent the dimension of the feature vector. Then, a graph attention convolutional layer (GAL) fuses neighbor context information to generate a new feature vector and performs a non-linear transformation through an activation function. Finally, a multi-head attention mechanism is implemented to enhance the model's ability to capture and integrate information. The multi-head attention mechanism captures different resource competition patterns in parallel, improving feature representation capabilities. Its outputs are concatenated to form a global context vector, which is then averaged. Furthermore, attention weights are constructed by sequentially performing linear transformations and concatenation on node features, resulting in a fixed ranking of attention weights among different query nodes. This limits the model's ability to capture multi-hop dependencies and interaction patterns in complex environments. To overcome this problem, GATv2 adjusts the computation order to achieve dynamic attention weights, thereby enhancing the model's expressive power.

[0108] In some embodiments, the point cloud service PCS orchestration neural network is trained with the aid of the PPO algorithm, which incorporates generalized advantage estimation (GAE) and pruning of surrogate target CSOs. PPO, as an advanced DRL algorithm based on an actor-critic framework, has been adopted for training effective and stable reinforcement learning models.

[0109] Specifically, the point cloud service PCS orchestration neural network is trained using the following steps:

[0110] Initialize the parameters of the Actor network and Critic network, and initialize the experience playback buffer.

[0111] In each loop, the agent interacts with the environment, constructing a graph representation of the network state under the currently arriving PCS request and sampling actions.

[0112] Under the premise of satisfying the constraints, the action is executed and feedback is obtained, and the graph structure observation trajectory is stored in the experience playback buffer.

[0113] Calculating the advantage function based on GAE A batch of samples is randomly drawn from the experience replay buffer, and the network parameters are updated by stochastic gradient descent until all point cloud service PCS orchestration is completed.

[0114] By repeating the above steps, an optimal policy model that can be used for the inference phase is obtained. For the orchestration process, inference is performed using a policy network with pre-trained parameters.

[0115] In other embodiments of the present invention, the architecture in Embodiment 1 above can also analyze the interaction between users and the network through a game theory model, and design a reasonable resource allocation mechanism so that all parties can obtain the best utility under limited resources.

[0116] In addition, fuzzy logic systems can be used to handle uncertainty and fuzziness, and dynamic resource allocation and task scheduling can be achieved through fuzzy reasoning on network status and user needs.

[0117] The point cloud service orchestration method based on graph attention networks provided in this invention modeles the edge network infrastructure (node ​​computing power, link bandwidth) as a graph structure and dynamically extracts resource competition relationships between nodes using a multi-head graph attention network (GATv2). It protects the dynamic attention weight calculation method, neighbor node feature aggregation method, and the mapping mechanism from the graph structure to the MDP state space, achieving a deep integration of network state awareness and intelligent decision-making. Furthermore, it combines Lyapunov optimization with Proximal Policy Optimization (PPO), using a quadratic Lyapunov function and drift-plus-penalty technique to transform long-term resource constraints into a single-slot queue stability problem. It protects the joint training mechanism based on GAE-based advantage function design, pruning the proxy objective function, and entropy reward enhancement to ensure the stability and convergence speed of policy updates. Finally, for the pipeline orchestration of holographic service task chains (such as encoding-decoding-rendering), it protects the modeling method for data dependencies between tasks (such as using task output as input for the next task) and the dynamic priority scheduling strategy. By combining user FoV changes and point density, the processing ratio is adaptively adjusted, and the balance between data volume and resource consumption is optimized through a non-linear complexity function.

[0118] In summary, this invention systematically solves the challenges of low-latency and high-reliability orchestration of existing holographic services in resource-constrained edge networks by employing dynamic graph-aware multi-target collaboration, task chain adaptive optimization, enhanced training stability, and cross-domain time-slot synchronization mechanisms. Furthermore, this invention designs an efficient edge network offloading and orchestration framework for holographic services and proposes a proximity strategy optimization algorithm based on graph attention-driven computation and network coordination. This ensures improved orchestration success rate and optimized system utility in high-PCS request scenarios, providing a feasible technical solution for immersive services in the 6G era.

[0119] Example 3

[0120] Based on Embodiment 1, Embodiment 3 provides a point cloud service orchestration device based on graph attention networks. This point cloud service orchestration device based on graph attention networks corresponds to the above-described point cloud service orchestration method based on graph attention networks. The device includes:

[0121] The first computing module uses the point cloud service PCS to orchestrate the neural network and convert the original state of the edge network into a structured graph.

[0122] The second calculation module fuses neighbor context information through graph attention convolutional layers to generate new feature vectors, and performs a non-linear transformation on the new feature vectors through an activation function to obtain the transformed result.

[0123] The third calculation module enhances the features of the transformed result through a multi-head attention mechanism to obtain enhanced features, and obtains the final point cloud service orchestration strategy through splicing and averaging.

[0124] For specific details, please refer to the description in the section on the orchestration method of point cloud services based on graph attention networks, which will not be repeated here.

[0125] Example 4

[0126] Embodiment 4 of the present invention provides an electronic device, including a memory and a processor, which communicate with each other. The memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute a point cloud service orchestration method based on a graph attention network. The method includes the following steps:

[0127] The original state of the edge network is transformed into a structured graph by orchestrating neural networks using the point cloud service PCS.

[0128] A new feature vector is generated by fusing neighbor context information through a graph attention convolutional layer, and the new feature vector is then subjected to a nonlinear transformation by an activation function to obtain the transformed result.

[0129] The transformed result is enhanced by a multi-head attention mechanism to obtain enhanced features, and the final point cloud service orchestration strategy is obtained by splicing and averaging.

[0130] Example 5

[0131] Embodiment 5 of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements a point cloud service orchestration method based on a graph attention network. The method includes the following steps:

[0132] The original state of the edge network is transformed into a structured graph by orchestrating neural networks using the point cloud service PCS.

[0133] A new feature vector is generated by fusing neighbor context information through a graph attention convolutional layer, and the new feature vector is then subjected to a nonlinear transformation by an activation function to obtain the transformed result.

[0134] The transformed result is enhanced by a multi-head attention mechanism to obtain enhanced features, and the final point cloud service orchestration strategy is obtained by splicing and averaging.

[0135] Example 6

[0136] Embodiment 6 of the present invention provides a computer program product, which includes a computer program and a point cloud service orchestration method based on a graph attention network when the computer program is executed by a processor. The method includes the following steps:

[0137] The original state of the edge network is transformed into a structured graph by orchestrating neural networks using the point cloud service PCS.

[0138] A new feature vector is generated by fusing neighbor context information through a graph attention convolutional layer, and the new feature vector is then subjected to a nonlinear transformation by an activation function to obtain the transformed result.

[0139] The transformed result is enhanced by a multi-head attention mechanism to obtain enhanced features, and the final point cloud service orchestration strategy is obtained by splicing and averaging.

[0140] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0141] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for method or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the description of the method embodiments. The method and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. 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 creative effort.

[0142] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An edge network resource orchestration system in a holographic communication intelligent identification network, characterized in that, The system includes: The infrastructure network layer is used to integrate computing and transmission resources based on the data processing and transmission capabilities of the physical infrastructure. The decision control layer is used to generate joint optimization strategies based on the real-time collected network status. The holographic service layer is used to aggregate different types of holographic service requests and trigger orchestration strategies based on the characteristics of the holographic service requests. The point cloud service model is used to decompose the received point cloud service into multiple sub-tasks to form a task chain; and to obtain the perception information of the point cloud service based on the task chain, point cloud frame group information, consumption information, latency, and processing level.

2. The system according to claim 1, characterized in that, The infrastructure network layer is a directed connected graph, which includes a set of servers characterized by computing resources and a set of physical links characterized by load bandwidth resources.

3. The system according to claim 1, characterized in that, The orchestration strategy triggered based on the characteristics of the holographic service request includes: The joint orchestration process of holographic services is modeled as a Markov decision process (MDP), and then processed using quadruples. It means that, among them, For state space, For the action space, For the transformation probability, This is the reward function.

4. A point cloud service orchestration method based on graph attention networks, characterized in that, The method is implemented based on the edge network resource orchestration system in the holographic communication intelligent identification network as described in any one of claims 1 to 3, and the method includes: The original state of the edge network is transformed into a structured graph by orchestrating the neural network using the point cloud service PCS; A new feature vector is generated by fusing neighbor context information through a graph attention convolutional layer, and the new feature vector is then subjected to a nonlinear transformation by an activation function to obtain the transformed result. The transformed result is enhanced by a multi-head attention mechanism to obtain enhanced features, and the final point cloud service orchestration strategy is obtained by splicing and averaging.

5. The method according to claim 4, characterized in that, The point cloud service PCS orchestration neural network is obtained through PPO algorithm training aided by generalized advantage estimation (GAE) and pruning surrogate target (CSO).

6. The method according to claim 5, characterized in that, The point cloud service PCS orchestration neural network is trained using the following steps: Initialize the parameters of the Actor network and Critic network, and initialize the experience replay buffer; In each loop, the agent interacts with the environment, constructing a graph representation of the network state under the currently arriving PCS request and sampling actions. Under the premise of satisfying the constraints, the action is executed and feedback is obtained, and the graph structure observation trajectory is stored in the experience playback buffer; The network parameters are updated using stochastic gradient descent based on a batch of samples randomly selected from the experience replay buffer based on the advantage function calculated by GAE, until all point cloud service PCS orchestration is completed.

7. A point cloud service orchestration device based on graph attention networks, characterized in that, The device is implemented based on the edge network resource orchestration system in the holographic communication intelligent identification network as described in any one of claims 1 to 3, and the device includes: The first computing module uses the point cloud service PCS to orchestrate the neural network and convert the original state of the edge network into a structured graph. The second calculation module fuses neighbor context information through graph attention convolutional layers to generate new feature vectors, and performs a non-linear transformation on the new feature vectors through an activation function to obtain the transformed result. The third calculation module enhances the features of the transformed result through a multi-head attention mechanism to obtain enhanced features, and obtains the final point cloud service orchestration strategy through splicing and averaging.

8. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the point cloud service orchestration method based on graph attention networks as described in any one of claims 4 to 6.

9. A computer-readable storage medium, characterized in that, It stores a computer program, which is loaded by a processor to execute the steps in the point cloud service orchestration method based on graph attention networks as described in any one of claims 4 to 6.

Images