A method for multi-dimensional resource coordination management in a vehicle network

Abstract

The application relates to a multi-dimensional resource cooperative management method in a vehicle networking network, and the method comprises the following steps: step 1, a clustering-based cooperative sensing network model in the vehicle networking network is constructed; step 2, on the basis of the cooperative sensing network model, a time-space joint value of a sensing information block is established and calculated; and step 3, steps such as radar mutual information MI and total time delay reached by VUE under different clustering modes are calculated. The superior effect of the application is that the method designs a clustering-based cooperative sensing network model in the vehicle networking network, provides a measurement mode of communication, sensing and calculation performance in the vehicle networking network, designs a VUE clustering mode and a sensing information block selection algorithm based on multi-agent reinforcement learning, reduces the communication time delay of the VUE uploaded sensing information block and improves the radar mutual information, and designs an improved interior point method calculation resource allocation algorithm, thereby reducing the calculation time delay of the cooperative sensing.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a method for multi-dimensional resource collaborative management in vehicle-to-everything (V2X) networks. Background Technology

[0002] Network systems enabled by Integrated Sensing and Communication (ISAC) technology offer advantages such as simplified structure, high interference suppression, and low resource consumption, making them a promising next-generation wireless communication network. In Vehicle Networks (VNETs), existing ISAC-supported designs alleviate the rapidly growing spectrum demands of VNETs and provide reliable sensing capabilities. However, in complex urban street scenarios, the sensing range and reliability of vehicles limit the development of fully autonomous driving. Furthermore, current research shows that the sensing range of ISAC-enabled Orthogonal Frequency Division Multiplexing (OFDM) radars is around 60 meters, and their sensing and communication capabilities are limited by the processing power of their onboard chips.

[0003] To overcome the shortcomings of separate perception, cooperative perception (also known as distributed perception) has recently attracted widespread attention from academia and industry. In addition to utilizing self-perceived information, cooperative perception technology collects perception information from nearby vehicles and roadside units through vehicle-to-everything (V2X) wireless communication technology, providing a broader global view of the road environment. To fully utilize the abundant computing and storage resources in VNETs, ​​academia has proposed fog-based vehicle networks (V2Ns) to effectively utilize these hierarchical resources and provide low-latency computational results for cooperative perception. Furthermore, international standardization organizations are currently accelerating the implementation of cooperative perception technology by formally defining its use cases and related information generation rules.

[0004] Despite the advantages mentioned above, cooperative perception still faces a key challenge: the simultaneous exchange of large volumes of sensing data can overwhelm existing vehicular networks with limited wireless spectrum and computing resources. Furthermore, most existing vehicles are designed for independent sensing, and their computing resources do not support cooperative perception, making it difficult to guarantee low latency requirements. Therefore, there is an urgent need for an effective cooperative perception method that can achieve a balance between the massive amount of uploaded data and limited wireless spectrum and computing resources in vehicular networks by supporting ISAC (Automatic Information Constraints).

[0005] In the prior art, Chinese invention patent application number CN202011504314.7 discloses a resource optimization method in a vehicle-to-everything (V2X) network based on non-orthogonal multiple access (NOMA). In a NOMA-assisted vehicle edge computing system, when processing vehicle tasks, the method minimizes the total energy consumption of the vehicle edge computing system. It determines the system's offloading and caching decisions, and the allocation of computing and caching resources to complete the resource optimization problem in the vehicle edge computing system. The method is characterized by the following steps: considering the random arrival of vehicle users and queue stability, a stochastic optimization problem is defined by jointly optimizing the computational offloading decision, content caching decision, and the allocation of computing and caching resources; a dynamic programming problem is proposed to solve this problem using Lyapunov optimization theory; the dynamic programming problem is decoupled into a computational offloading subproblem and a content caching subproblem by combining computational offloading, content caching, and resource allocation algorithms; the optimal solutions for the offloading decision and computing resource allocation are obtained by solving the computational offloading subproblem; and the optimal solutions for the caching decision and caching resource allocation are obtained by solving the content caching subproblem.

[0006] For example, Chinese invention patent application number CN201910178826.X discloses a dynamic resource optimization management system for a vehicle-to-everything (V2X) cloud-fog system based on SMDP. Its features include an integrated cloud-fog V2X architecture, comprising a central cloud, a fog service layer, and an in-vehicle cloud. The central cloud includes a server cluster that centralizes V2X application services, stores and analyzes massive amounts of data collected by terminal devices, analyzes V2X big data to obtain big data analysis results, manages traffic information, and provides services to in-vehicle users. The fog service layer extends computing and data analysis capabilities to the network edge by accessing fog devices, enabling some data in the V2X to be transferred locally for processing, providing services to in-vehicle users. The in-vehicle cloud includes in-vehicle nodes and user mobile terminals, dividing vehicles into different regional sets. Vehicles within the same regional set form a network through vehicle-to-vehicle communication, sharing computing, storage, and spectrum resources. After connecting to the network, vehicle terminals quickly obtain services from the fog service layer.

[0007] For example, Chinese invention patent application number CN201910178826.X discloses a dynamic resource optimization management system for a vehicle-to-everything (V2X) cloud-fog system based on SMDP. Its features include an integrated cloud-fog V2X architecture, comprising a central cloud, a fog service layer, and an in-vehicle cloud. The central cloud includes a server cluster that centralizes V2X application services, stores and analyzes massive amounts of data collected by terminal devices, analyzes V2X big data to obtain big data analysis results, manages traffic information, and provides services to in-vehicle users. The fog service layer extends computing and data analysis capabilities to the network edge by accessing fog devices, enabling some data in the V2X to be transferred locally for processing, providing services to in-vehicle users. The in-vehicle cloud includes in-vehicle nodes and user mobile terminals, dividing vehicles into different regional sets. Vehicles within the same regional set form a network through vehicle-to-vehicle communication, sharing computing, storage, and spectrum resources. After connecting to the network, vehicle terminals quickly obtain services from the fog service layer.

[0008] All of the above-mentioned invention patent applications do not consider the realization of collaborative perception between vehicles, and are difficult to meet the requirements of vehicle-to-everything (V2X) networks for perception capabilities and latency. Summary of the Invention

[0009] In view of this, the present invention provides a multi-dimensional resource collaborative management method in vehicle-to-everything (V2X) networks based on deep reinforcement learning, matching algorithms, and convex optimization theory, comprising the following specific steps:

[0010] Step 1: Construct a cluster-based collaborative perception network model in the vehicle-to-everything (V2X) network;

[0011] Step 2: Based on the collaborative sensing network model, establish and calculate the temporal-spatial joint value of the sensing information block;

[0012] Step 3: Calculate the radar mutual information (MI) and total delay achieved by VUE under different clustering modes;

[0013] Step 4: With the goal of maximizing user satisfaction, construct an optimization model for joint clustering, selection of perception information blocks, allocation of resource blocks (RBs), and optimal allocation of CPU cycle frequency.

[0014] Step 5: Establish a VUE clustering mode and perceptual information block selection model based on multi-agent deep reinforcement learning, and perform VUE clustering mode and perceptual information block selection.

[0015] Step 6: Train the reinforcement learning model and find the optimal clustering pattern and perceptual information block selection scheme for VUE;

[0016] Step 7: Perform joint radio and compute resource allocation for VUE;

[0017] Step 8: Input the network state of subsequent time slots into the model, repeat steps 6 to 8, and output the communication, perception and computing collaborative resource allocation scheme in the vehicle-to-everything network that maximizes long-term satisfaction while satisfying its time delay and radar mutual information constraints.

[0018] Furthermore, in step 1, a cluster-based collaborative perception network model is constructed in the vehicle-to-everything (V2X) network, including:

[0019] The collaborative sensing network model consists of one cloud server, M remote radio heads (RRHs), and N fog access points (F-APs). Multiple vehicle user equipment (VUEs) are configured to associate with the same fog node server or cloud server through RRHs or fog nodes, forming collaborative clusters and uploading their local sensing information blocks to the fog node server or cloud server for collaborative computation. This is referred to as a collaborative sensing task T. k ;

[0020] Set the connection selection for VUE k to x k，n If x k，0 =1, VUE k is associated with the cloud model through RRHs; conversely, if x k，n =1, n∈N, VUE k selects to access F-AP n.

[0021] Furthermore, in step 2, based on the collaborative sensing network model, the temporal-spatial joint value of the sensing information block is established and calculated, including:

[0022] Step 2.1: Divide the perception information blocks for each VUE in the network model:

[0023] Using quadtree data compression technology, the road area is divided into different two-dimensional regions, and the data contained in each two-dimensional region is compressed into independent sensing information blocks. The sensing information blocks around the VUE are obtained through its own wireless sensing capabilities and are called self-sensing information blocks. Within the VUE's expected sensing distance, the sensing information blocks where the wireless sensing capabilities of different vehicles of the VUE overlap are redundant blocks. The sensing information block is a square two-dimensional plane with a fixed side length, which contains vehicle and road information within the current plane. The self-sensing information block refers to the sensing information block within the sensing distance of the VUE through its own equipped ISAC integrated device. The expected sensing distance is the VUE front area with a length greater than the side length of the VUE self-sensing information block by a length of d.

[0024] Step 2.2: Constructing the temporal-spatial joint value of the perceived information block:

[0025] The temporal and spatial values ​​of the perceived information block are constructed as the basis for VUE to upload the perceived information block. The temporal and spatial values ​​refer to the degree of interest of the perceived information block in the VUE on the temporal and spatial scales: the closer the perceived information block is to the VUE and the fresher the generation time of the perceived information block, the greater the VUE's interest in it. The formula for the time value at time t is set as follows:

[0026]

[0027] Where, q k，b，0 >0 is the initial value of the perceived information block, τ dll It is the deadline for each block of perceived information, t. k，b，0 It is the time when the perceived information block is generated;

[0028] The formula for the spatial value of block b in VUE k is set as follows:

[0029]

[0030] Where, d k，b (t) represents the Euclidean distance between VUE k and the perceived information block b at time t, and θ k，b Let be the angle between the VUE k movement direction and the perceived information block b. The desired perception distance for VUE;

[0031] The formula for the joint time-space value of block information in VUE k is set as follows:

[0032]

[0033] Step 2.3: Calculate the joint temporal and spatial value of the perceived information block:

[0034] VUEs selectively upload the perception information block with the highest time-space joint value. Each VUE periodically selects and uploads the self-perception information block with the highest time-space joint value to its associated fog node server or cloud server. After receiving the uploaded perception information blocks from VUE members within the cluster, the cloud server and fog node F-AP readjust, integrate, infer, and map the collected perception information blocks to achieve collaborative perception among VUEs.

[0035] Furthermore, in step 3, the radar mutual information (MI) and total delay achievable by VUE under different clustering modes are calculated:

[0036] Step 3.1: Calculate the radar signal-to-interference-noise ratio (SINR) at VUE. RadBased on the achievable radar mutual information (MI), a perception model is established according to radar information estimation theory. Each VUE is configured to continuously transmit integrated OFDM waveforms via an onboard transmitter equipped with ISAC technology for radar detection. The radar signal received at VUE k is compared with the interference noise ratio (SINR). Rad for:

[0037]

[0038] Where, p k and p j For the transmission power of VUE k and other VUE j, here a k，s ∈{0,1} is the allocation vector for all k∈K and s∈S of resource block RB, G k，s (f) is g k，s The Fourier transform of (t) at time t, denoted as h, represents the channel gain of the radar receiver from another VUE j to VUE k. j，k，s (t); The radar mutual information MI that VUE k can realize within time slot t is defined as:

[0039]

[0040] Step 3.2: Calculate the total latency achieved by the VUE under different clustering modes. When the VUE selects the fog node clustering mode, the communication rate achieved is calculated using the following formula:

[0041]

[0042] Where W is the bandwidth size of the resource block, a k，s Assign vectors to binary resource blocks, p k For the transmit power of VUE k, p j To determine the transmit power of VUE j, which occupies the same resource block as VUE k, Let be the channel gain from VUEk to F-AP n on resource block s at time t. σ is the channel gain from other VUE j to F-AP n on resource block s at time t; 2 For noise power, the formula for the communication rate achieved when VUE selects the cloud server clustering mode is:

[0043]

[0044] Among them, g k，s It is the vector detected by the Minimum Mean Square Error (MMSE). It is the channel gain from VUE k to its associated RRHs M within time t, while This is the channel gain from VUE j to the associated RRHs of VUE k. Within a cluster, the communication delay of VUE k is the maximum upload delay among all VUE members, calculated using the following formula:

[0045]

[0046] Where, τ fh It is the fronthaul latency, measured by the actual network communication capabilities of the operator. I is the data size of each sensing information block, measured in bits. k，b For VUE k, the binary selection variable for perceptual information block b: e k，b =1 indicates that VUE k will upload the perception information block b; otherwise, e k，b =0 means that VUE k will not upload the perception information block b. VUE The computational delay for k is:

[0047]

[0048] Among them, f 0，max and f n，max These represent the maximum CPU cycle frequency of the cloud server and each fog node server, respectively. k，n Given the CPU cycle frequency allocated to VUE k, the total latency of VUE k is:

[0049] in, and These are local compression latency, communication latency, and task computation latency, respectively.

[0050] Furthermore, in step 4, with the goal of maximizing user satisfaction, an optimization model is constructed for joint clustering, perceptual information block selection, resource block (RB) allocation, and optimal CPU cycle frequency allocation:

[0051] The collaborative sensing optimization problem is formulated as follows:

[0052] P1:

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059] Among them, U k (t) is the satisfaction function, T is the total duration, and ε1 and ε2 are satisfaction weighting parameters to balance the impact of the spatiotemporal value of perceived information and the total task delay on satisfaction. It is the threshold for maximum latency tolerance, MI min The minimum radar mutual information tolerance threshold is given. Constraint C1 is the task latency requirement, determined by the maximum tolerable latency of the task. Constraint C2 indicates that each VUE should upload non-overlapping perception information to make reasonable use of computing resources. Constraint C3 indicates that each VUE can choose an offload node to enable reasonable clustering among VUEs. Constraint C4 indicates that each VUE selects a resource block RB for uploading, and the resource block RB can be reused by multiple VUEs in different clusters. Constraint C5 limits the CPU cycle frequency allocated by the fog node F-AP and the cloud server. Constraint C6 is the minimum perception performance threshold for VUEs with ISAC enabled.

[0060] Furthermore, in step 5, a VUE clustering pattern and perceptual information block selection model based on multi-agent reinforcement learning is established:

[0061] Step 5.1: Treat each VUE as an agent and construct an attention-assisted multi-agent DDPG algorithm model. The local state feature vector of agent k is defined as:

[0062]

[0063] in, For VUE k, the maximum tolerable delay, l k Let q′ be the current coordinate of VUE k. k Let {O′} be the time value of the perceived message of VUE at time t-1. n} n∈0∪N Let r be the remaining computing resources of fog node n and cloud server at time t-1. sat′ On a cloud server, the action of agent k is defined as the ratio of the global latency satisfaction of the previous action to the impact of the previous action.

[0064] a k =(n, e) k，b ),

[0065] Where n represents the sequence number of the fog node being connected;

[0066] Step 5.2: Construct the immediate reward function for each VUE. The reward function for each VUE k is defined as follows:

[0067]

[0068] Step 5.3: Construct the state features of the multi-agent Markov game model using the state feature vectors of the agents:

[0069] Let the Markov game model Γ for each agent be defined as:

[0070]

[0071] Where K is the number of agents. For state space, It is the action space, which is the system's instantaneous reward function determined by all reward agents, and γ is the exploration discount factor;

[0072] Step 5.4: Implement collaboration among agents using the Multi-Agent Development Program (DDPG) model, and then adjust the commenter network for each agent using a multilayer perceptual network with an attention mechanism. In the attention-based DDPG algorithm, each agent k has a commenter network responsible for finding a deterministic policy based on the local observation state, given random noise. The action a selected at time t k The formula for (t) is defined as:

[0073]

[0074] Where, θ k For the actor network parameters, s k (t) represents the local observation state of VUE k, μ k (s k (t); θ k ) is achieved by inputting the critic network parameter θ k and the local observation state s of VUE k k The deterministic policy obtained after (t) is such that each actor network adjusts the network parameters θ in the gradient direction. k To maximize the reward function, the formula is:

[0075]

[0076] in, It is a replay buffer used to store the experience tuples of all agents, Q. k (s, a) is the action-value function built by an attention-assisted commentator network;

[0077] Establish a multilayer perceptual network based on the observations of agents j and k, with attention weights α. k，j and attention value v k，j The calculation formulas are as follows:

[0078]

[0079] v k，j =h(V k e j ),

[0080] Among them, W q and W k Composition of e k and e j The bilinear mapping, h(·) is a ReLU function, V k This is the transformation matrix. The commentator network is set to update by minimizing the following loss function, as shown in the formula:

[0081]

[0082] Where, δ k These are the network parameters for the commentator, where (s, a, r, s′) is a quadruple consisting of state value, action value, immediate reward, and next state value, and s′ is the next state value. - (s′, a′) is the target action-value function inferred by the target attention-assisted critic network. It is the target action obtained from the network of target actors.

[0083] Furthermore, in step 6, a reinforcement learning model is trained, and the optimal clustering pattern and perceptual information block selection scheme for VUE are found:

[0084] Step 6.1: Initialize the actor network, attention mechanism critic network, target actor network, target critic network, and replay cache. and training batch size;

[0085] Step 6.2: Initialize state s in each training round, and observe the current state s of the agent in each time slot. k ;

[0086] Step 6.3: If the current round is in the pre-training phase, use the formula... Explore VUE clustering and message block selection, then return to step 6.2 until all time slots and agents have completed execution; otherwise, use formula a. k (t)=μ k (s k (t); θ k Select an action to perform VUE clustering and message block selection, and return to step 6.2 until all time slots and agents have completed their execution;

[0087] Step 6.4: Calculate the immediate reward r and the next state s', and then store the experience tuple (s, a, r, s') into the experience pool;

[0088] Step 6.5: For all agents, randomly select a batch of tuples from the replay buffer to update the actor network, the attention mechanism, and the commentator network respectively;

[0089] Step 6.6: Update the target's actors, attention mechanism, and commentator network using a soft update method, and repeat step 6.6 until all agents have completed their execution;

[0090] Furthermore, in step 7, joint radio and compute resource allocation is performed for the VUE:

[0091] Step 7.1: Perform optimal allocation of resource blocks (RBs) for VUEs based on an improved exchange matching algorithm. This involves providing an improved exchange matching algorithm to allocate resource blocks (RBs) for VUEs. For the initial matching, a preference list L is first constructed for each VUE and resource block (RB) in descending order. k and L s Then, the mismatched VUEs iteratively make requests to their optimal resource block RB. The resource block RB can accept its preferred VUEs and reject other VUEs until all VUEs are matched. The execution steps are as follows:

[0092] Step 7.1.1 Initialize the preference list L of VUE k and resource blocks RBs in descending order. k and L s ;

[0093] Step 7.1.2 Initialize the mismatched Vue collection

[0094] Step 7.1.3 When the VUE's preference list and the set of mismatched VUEs are not empty, for all VUEs in the set of mismatched VUEs, iteratively submit a matching request to the resource block RB that has not yet rejected it, in descending order of the product of the current resource block RB's achievable uplink delay and radar mutual information.

[0095] Step 7.1.4 For all resource blocks RB, if RB s receives a matching request from VUE k, then RB s maintains the match with VUE k and rejects all other matching requests; otherwise, resource block RBs accept the matching request from the VUE with the largest quotient of uplink delay and radar mutual information and reject the matching requests from other VUEs.

[0096] After step 7.1.5 completes the iteration, output the optimal resource block (RB) allocation result;

[0097] Step 7.2: Perform optimal allocation of computational resources for VUE based on the improved interior-point method:

[0098] Based on the optimal allocation result of RB obtained in step 7.1, further calculation of resource optimization allocation is performed. The formula for the resource optimization allocation problem is as follows:

[0099]

[0100] This paper presents an improved computational resource allocation algorithm based on the interior-point method to solve the problem. A penalty function is constructed as follows:

[0101]

[0102] Where f is a feasible solution, the execution steps are as follows:

[0103] Step 7.2.1 Initialize the penalty factor r, feasible solution f, and maximum number of iterations N. iter Maximum tolerance error ε0 and update coefficient b;

[0104] Step 7.2.2 Calculate for each VUE k

[0105] Step 7.2.3 If Obtain the unconstrained computational resource allocation result f under KKT conditions. k，n Otherwise, within the maximum number of iterations and the maximum tolerance error range, obtain the penalty function Ψ for the current r and f. i And calculate the remaining computing resources.

[0106] Step 7.2.4 For all VUE k, update the gradient value of VUE. and the results of the allocated computing resources

[0107] Step 7.2.5 Update Set r = cr, where c is the descent coefficient;

[0108] After the iteration in step 7.2.6 is completed, the optimal computing resource allocation result is output.

[0109] Furthermore, in step 8, the network state of subsequent time slots is input into the model respectively:

[0110] Repeat steps 6 to 8 to output the communication, perception and computing collaborative resource allocation scheme in the vehicle-to-everything (V2X) network that maximizes long-term satisfaction while satisfying its latency and radar mutual information constraints. The collaborative resource allocation scheme includes selecting the optimal resource blocks, perception information blocks, clusters and the CPU frequency of cloud servers or fog node servers.

[0111] The multi-dimensional resource collaborative management method in the vehicle-to-everything (V2X) network described in this invention has the following beneficial effects:

[0112] 1. The multi-dimensional resource collaborative management method in the vehicle-to-everything (V2X) network described in this invention designs a cluster-based collaborative perception network model in the V2X network and provides a way to measure the communication, perception, and computing performance in the V2X network.

[0113] 2. The multi-dimensional resource collaborative management method in the vehicle-to-everything (VUE) network described in this invention designs a VUE clustering mode and a perception information block selection algorithm based on multi-agent reinforcement learning, which greatly reduces the communication latency of VUE uploading perception information blocks and improves radar mutual information.

[0114] 3. The multi-dimensional resource collaborative management method in the vehicle-to-everything (V2X) network described in this invention designs an improved interior-point method for resource allocation, which greatly reduces the computational latency of collaborative perception.

[0115] 4. The multi-dimensional resource collaborative management method in the vehicle-to-everything (V2X) network described in this invention optimizes clustering, sensing information block selection, resource block (RB) allocation, and CPU cycle frequency in combination, so as to maximize the long-term overall satisfaction of collaborative sensing tasks while meeting the constraints of latency and radar mutual information. Attached Figure Description

[0116] Figure 1 This is a schematic diagram of the multi-dimensional resource collaborative management method in the vehicle-to-everything (V2X) network described in this invention.

[0117] Figure 2 This is a schematic diagram of the collaborative perception model of the vehicle-to-everything (V2X) network provided in this application.

[0118] Figure 3 This is a schematic diagram of the four-way tree division of the road area in the vehicle-to-everything (V2X) network provided in this application.

[0119] Figure 4 This is a schematic diagram of the collaborative perception task execution process provided in this application.

[0120] Figure 5 This is a schematic diagram of the Attention Multi-Agent (DDPG) framework for vehicle-to-everything (V2X) networks provided in this application. Detailed Implementation

[0121] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0122] like Figure 1-5 As shown, the multi-dimensional resource collaborative management method in the vehicle-to-everything (V2X) network of the present invention includes the following specific steps:

[0123] Step 1: Construct a cluster-based collaborative perception network model in the vehicle-to-everything (V2X) network:

[0124] like Figure 2As shown, the cooperative sensing network model consists of one cloud server, M remote radio heads (RRHs), and N fog access points (F-APs). Multiple vehicle user equipment (VUEs) are configured to associate with the same fog node server or cloud server through RRHs or fog nodes, forming cooperative clusters and uploading their local sensing information blocks to the fog node server or cloud server for collaborative computation. This is referred to as a cooperative sensing task T. k ;

[0125] Set the connection selection for VUE k to x k，n If x k，0 =1, VUE k is associated with the cloud model through RRHs; conversely, if x k，n =1, n∈N, VUE k chooses to access F-AP n;

[0126] Step 2, based on the collaborative sensing network model, establish and calculate the temporal-spatial joint value of the sensing information block, including:

[0127] Step 2.1: Divide the perception information blocks for each VUE in the network model:

[0128] like Figure 3 As shown, quadtree data compression technology is used to divide the road area into different two-dimensional regions, and the data contained in each two-dimensional region is compressed into independent sensing information blocks. The surrounding area of ​​the VUE is the sensing information block obtained by its own wireless sensing capability, which is called the self-sensing information block. Within the VUE's expected sensing distance, the sensing information blocks where the wireless sensing capabilities of different vehicles of the VUE overlap are redundant blocks. The sensing information block is a square two-dimensional plane with a fixed side length, which contains vehicle and road information in the current plane. The self-sensing information block refers to the sensing information block within the sensing distance of the VUE through its own equipped ISAC integrated device. The expected sensing distance is the VUE front area with a length greater than the side length of the VUE self-sensing information block by length d.

[0129] Step 2.2: Constructing the temporal-spatial joint value of the perceived information block:

[0130] The temporal and spatial values ​​of the perceived information block are constructed as the basis for VUE to upload the perceived information block. The temporal and spatial values ​​refer to the degree of interest of the perceived information block in the VUE on the temporal and spatial scales: the closer the perceived information block is to the VUE and the fresher the generation time of the perceived information block, the greater the VUE's interest in it. The formula for the time value at time t is set as follows:

[0131]

[0132] Where, q k，b，0 >0 is the initial value of the perceived information block, τ dll It is the deadline for each block of perceived information, t. k，b，0 It is the time when the perceived information block is generated;

[0133] The formula for the spatial value of block b in VUE k is set as follows:

[0134]

[0135] Where, d k，b (t) represents the Euclidean distance between VUE k and the perceived information block b at time t, and θ k，b Let be the angle between the VUE k movement direction and the perceived information block b. The desired perception distance for VUE;

[0136] The formula for the joint time-space value of block information in VUE k is set as follows:

[0137]

[0138] Step 2.3: Calculate the joint temporal and spatial value of the perceived information block:

[0139] VUE selectively uploads the perceptual information block with the highest combined time-space value, such as... Figure 4 As shown, each VUE periodically selects the self-perceived information block with the highest time-space joint value to upload to its associated fog node server or cloud server. After receiving the uploaded perception information blocks from VUE members within the cluster, the cloud server and fog node F-AP realize collaborative perception among VUEs by readjusting, integrating, inferring, and mapping the collected perception information blocks. Adjustment involves the cloud server and fog node server reordering the perception information blocks collected at different times; integration involves the cloud server and fog node server combining perception information blocks from different clusters; inference involves the cloud server and fog node server performing inference operations on the perception information blocks; and mapping involves the cloud server and fog node server projecting the inference operation results of different perception information blocks to the corresponding VUEs in the system.

[0140] Step 3: Calculate the radar mutual information (MI) and total delay achievable by VUE under different clustering modes:

[0141] Step 3.1: Calculate the radar signal-to-interference-noise ratio (SINR) at VUE. RadBased on the achievable radar mutual information (MI), a perception model is established according to radar information estimation theory. Each VUE is configured to continuously transmit integrated OFDM waveforms via an onboard transmitter equipped with ISAC technology for radar detection. The ratio of the radar signal received at VUE k to the interference noise (SINR) is calculated. Rad for:

[0142]

[0143] Where, p k and p j For the transmission power of VUE k and other VUE j, here a k，s ∈{0,1} is the allocation vector for all k∈K and s∈S of resource block RB, G k，s (f) is g k，s The Fourier transform of (t) at time t, denoted as h, represents the channel gain of the radar receiver from another VUE j to VUE k. j，k，s (t); The radar mutual information MI that VUE k can realize within time slot t is defined as:

[0144]

[0145] Step 3.2: Calculate the total latency achieved by the VUE under different clustering modes. When the VUE selects the fog node clustering mode, the communication rate achieved is calculated using the following formula:

[0146]

[0147] Where W is the bandwidth size of the resource block, a k，s Assign vectors to binary resource blocks, p k For the transmit power of VUE k, p j To determine the transmit power of VUE j, which occupies the same resource block as VUE k, Let be the channel gain from VUEk to F-AP n on resource block s at time t. σ is the channel gain from other VUE j to F-AP n on resource block s at time t; 2 For noise power, the formula for the communication rate achieved when VUE selects the cloud server clustering mode is:

[0148]

[0149] Among them, g k，s It is the vector detected by the Minimum Mean Square Error (MMSE). It is the channel gain from VUE k to its associated RRHs M within time t, while This is the channel gain from VUE j to the associated RRHs of VUE k. Within a cluster, the communication delay of VUE k is the maximum upload delay among all VUE members, calculated using the following formula:

[0150]

[0151] Where, τ fh It is the fronthaul latency, measured by the actual network communication capabilities of the operator. I is the data size of each sensing information block, measured in bits. k，b For VUE k, the binary selection variable for perceptual information block b: e k，b =1 indicates that VUE k will upload the perception information block b; otherwise, e k，b =0 indicates that VUE k will not upload the perception information block b, and the computation latency of VUE k is:

[0152]

[0153] Among them, f 0，max and f n，max These represent the maximum CPU cycle frequency of the cloud server and each fog node server, respectively. k，n Given the CPU cycle frequency allocated to VUE k, the total latency of VUE k is:

[0154]

[0155] in, and These are local compression latency, communication latency, and task computation latency, respectively.

[0156] Step 4: To maximize user satisfaction, construct an optimization model for joint clustering, perceptual information block selection, resource block (RB) allocation, and optimal CPU cycle frequency allocation.

[0157] The collaborative sensing optimization problem is formulated as follows:

[0158] P1

[0159]

[0160]

[0161]

[0162]

[0163]

[0164]

[0165] Among them, U k (t) is the satisfaction function, T is the total duration, and ε1 and ε2 are satisfaction weighting parameters to balance the impact of the spatiotemporal value of perceived information and the total task delay on satisfaction. It is the threshold for maximum latency tolerance, MI min The minimum radar mutual information tolerance threshold is the threshold. The satisfaction weight parameter is selected and formulated by the network operator according to the actual network conditions. Constraint C1 is the task latency requirement, which is determined by the maximum tolerable latency of the task. Constraint C2 indicates that each VUE should upload non-overlapping perception information to make reasonable use of computing resources. Constraint C3 indicates that each VUE can choose an offload node to make reasonable clustering among VUEs. Constraint C4 indicates that each VUE selects a resource block RB for uploading. The resource block RB can be reused by multiple VUEs in different clusters. Constraint C5 limits the CPU cycle frequency allocated by the fog node F-AP and the cloud server. Constraint C6 is the minimum perception performance threshold for VUEs with ISAC enabled.

[0166] Step 5: Establish a VUE clustering pattern and perceptual information block selection model based on multi-agent reinforcement learning:

[0167] Step 5.1: Treat each VUE as an agent and construct an attention-assisted multi-agent DDPG algorithm model. The local state feature vector of agent k is defined as:

[0168]

[0169] in, For VUE k, the maximum tolerable delay, l k Let q′ be the current coordinate of VUE k. k Let {O′} be the time value of the perceived message of VUE k at time t-1. n} n∈0∪N Let r be the remaining computing resources of fog node n and cloud server at time t-1. sat′ On a cloud server, the action of agent k is defined as the ratio of the global latency satisfaction of the previous action to the impact of the previous action.

[0170] a k =(n, e) k，b ),

[0171] Where n represents the sequence number of the fog node being connected;

[0172] Step 5.2: Construct the immediate reward function for each VUE. The reward function for each VUE k is defined as follows:

[0173]

[0174] Step 5.3: Construct the state features of the multi-agent Markov game model using the state feature vectors of the agents:

[0175] Let the Markov game model Γ for each agent be defined as:

[0176]

[0177] Where K is the number of agents. For state space, It is the action space, r K It is the system's instantaneous reward function determined by all reward agents, where γ is the exploration discount factor;

[0178] Step 5.4: Use the multi-agent DDPG model to realize cooperation between agents, and then use a multi-layer perceptual network with attention mechanism to adjust the critic network of each agent;

[0179] like Figure 5 As shown, in the Attention Multi-Agent (DDPG) algorithm, each agent k has a commentator network responsible for finding a deterministic policy based on the local observation state, given random noise. The action a selected at time t k The formula for (t) is defined as:

[0180]

[0181] Where, θ k For the actor network parameters, s k (t) represents the local observation state of VUE k, μ k (s k (t); θ k ) is achieved by inputting the critic network parameter θ k and the local observation state s of VUE k k The deterministic policy obtained after (t) is such that each actor network adjusts the network parameters θ in the gradient direction. k To maximize the reward function, the formula is:

[0182]

[0183] in, It is a replay buffer used to store the experience tuples of all agents, Q. k (s, a) is the action-value function built by an attention-assisted commentator network;

[0184] Establish a multilayer perceptual network based on the observations of agents j and k, with attention weights α. k，jand attention value v k，j The calculation formulas are as follows:

[0185]

[0186] v k，j =h(V k e j ),

[0187] Among them, W q and W k Composition of e k and e j The bilinear mapping, h(·) is a ReLU function, V k This is the transformation matrix. The commentator network is set to update by minimizing the following loss function, as shown in the formula:

[0188]

[0189] Where, δ k These are the network parameters for the commentator, where (s, a, r, s′) is a quadruple consisting of state value, action value, immediate reward, and next state value, and s′ is the next state value. - (s′, a′) is the target action-value function inferred by the target attention-assisted critic network. It is the target action obtained from the target actor network;

[0190] Step 6: Train the reinforcement learning model and find the optimal clustering pattern and perceptual information block selection scheme for VUE:

[0191] Step 6.1: Initialize the actor network, attention mechanism critic network, target actor network, target critic network, and replay cache. and training batch size;

[0192] Step 6.2: Initialize state s in each training round, and observe the current state s of the agent in each time slot. k ;

[0193] Step 6.3: If the current round is in the pre-training phase, use the formula... Explore VUE clustering and message block selection, then return to step 6.2 until all time slots and agents have completed execution; otherwise, use formula a. k (t)=μ k (s k (t); θ k Select an action to perform VUE clustering and message block selection, and return to step 6.2 until all time slots and agents have completed their execution;

[0194] Step 6.4: Calculate the immediate reward r and the next state s', and then store the experience tuple (s, a, r, s') into the experience pool;

[0195] Step 6.5: For all agents, randomly select a batch of tuples from the replay buffer to update the actor network, the attention mechanism, and the commentator network respectively;

[0196] Step 6.6: Update the target's actors, attention mechanism, and commentator network using a soft update method, and repeat step 6.6 until all agents have completed their execution;

[0197] Step 7: Perform joint radio and compute resource allocation for the VUE:

[0198] Step 7.1: Perform optimal allocation of resource blocks (RBs) for VUEs based on an improved exchange matching algorithm. This involves providing an improved exchange matching algorithm to allocate resource blocks (RBs) for VUEs. For the initial matching, a preference list L is first constructed for each VUE and resource block (RB) in descending order. k and L s Then, the mismatched VUEs iteratively make requests to their optimal resource block RB. The resource block RB can accept its preferred VUEs and reject other VUEs until all VUEs are matched. The execution steps are as follows:

[0199] Step 7.1.1 Initialize the preference list L of VUE k and resource block RBs in descending order. k and L s ;

[0200] Step 7.1.2 Initialize the mismatched Vue collection

[0201] Step 7.1.3 When the VUE's preference list and the set of mismatched VUEs are not empty, for all VUEs in the set of mismatched VUEs, iteratively submit a matching request to the resource block RB that has not yet rejected it, in descending order of the product of the current resource block RB's achievable uplink delay and radar mutual information.

[0202] Step 7.1.4 For all resource blocks RB, if RB s receives a matching request from VUE k, then RB s maintains the match with VUE k and rejects all other matching requests; otherwise, resource block RBs accept the matching request from the VUE with the largest quotient of uplink delay and radar mutual information and reject the matching requests from other VUEs.

[0203] After step 7.1.5 completes the iteration, output the optimal resource block (RB) allocation result;

[0204] Step 7.2: Perform optimal allocation of computational resources for VUE based on the improved interior-point method:

[0205] Based on the optimal allocation result of RB obtained in step 7.1, further calculation of resource optimization allocation is performed. The formula for the resource optimization allocation problem is as follows:

[0206]

[0207] This paper presents an improved computational resource allocation algorithm based on the interior-point method to solve the problem. A penalty function is constructed as follows:

[0208]

[0209] Where f is a feasible solution, the execution steps are as follows:

[0210] Step 7.2.1 Initialize the penalty factor r, feasible solution f, and maximum number of iterations N. iter Maximum tolerance error ε0 and update coefficient b;

[0211] Step 7.2.2 Calculate for each VUE k

[0212] Step 7.2.3 If Obtain the unconstrained computational resource allocation result f under KKT conditions. k，n Otherwise, within the maximum number of iterations and the maximum tolerance error range, obtain the penalty function Ψ for the current r and f. i And calculate the remaining computing resources.

[0213] Step 7.2.4 For all VUE k, update the gradient value of VUE. and the results of the allocated computing resources

[0214] Step 7.2.5 Update Set r = cr, where c is the descent coefficient;

[0215] After step 7.2.6 completes the iteration, output the optimal computational resource allocation result;

[0216] Step 8: Input the network state of subsequent time slots into the model respectively:

[0217] Repeat steps 6 to 8 to output the communication, perception and computing collaborative resource allocation scheme in the vehicle-to-everything (V2X) network that maximizes long-term satisfaction while satisfying its latency and radar mutual information constraints. The collaborative resource allocation scheme includes selecting the optimal resource blocks, perception information blocks, clusters and the CPU frequency of cloud servers or fog node servers.

[0218] By implementing the collaborative resource management scheme of communication, perception and computing fusion based on deep reinforcement learning, matching algorithm and convex optimization theory, which integrates joint clustering and sensing information block selection, a feasible solution is provided for the collaborative management of communication, perception and computing resources in vehicle networks.

[0219] This invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its concept and scope. All such changes and modifications fall within the scope of the invention as defined by the appended claims.

Claims

1. A method for multi-dimensional resource collaborative management in a vehicle-to-everything (V2X) network, comprising the following specific steps: Step 1: Construct a cluster-based collaborative perception network model in the vehicle-to-everything (V2X) network; Step 2: Based on the collaborative sensing network model, establish and calculate the joint temporal-spatial value of the sensing information block: Step 2.1: Divide the perception information blocks for each VUE in the network model: Using quadtree data compression technology, the road area is divided into different two-dimensional regions, and the data contained in each two-dimensional region is compressed into independent sensing information blocks. The surrounding area of ​​the VUE consists of sensing information blocks obtained through its own wireless sensing capabilities, denoted as self-sensing information blocks. Within the VUE's desired sensing distance, the overlapping sensing information blocks between different vehicles within the VUE are considered redundant blocks. Each sensing information block is a square two-dimensional plane with fixed side length, containing vehicle and road information within the current plane. Self-sensing information blocks refer to sensing information blocks within the sensing distance of the VUE's integrated ISAC device. The desired sensing distance is a length greater than the side length of the VUE's self-sensing information block. The front area of ​​VUE; Step 2.2: Constructing the temporal-spatial joint value of the perceived information block: The temporal and spatial value of the perceived information block are constructed as the basis for VUE to upload the perceived information block. Temporal and spatial value refer to the degree of VUE's interest in the perceived information block on a temporal and spatial scale: the closer the perceived information block is to the VUE and the more recent its generation time, the greater the VUE's interest in it. This is set... The formula for the time value of a moment is: ， in, It is the initial value of the perceived information block. It is the deadline for each block of perceived information. It is the time when the perceived information block is generated; Configure VUE Medium block The formula for spatial value is: ， in, for VUE Time and perceptual information blocks The Euclidean distance between them For VUE Movement direction and perception information block The angle between them The desired perception distance for VUE; Configure VUE The formula for the joint time-space value of information in medium blocks is: Step 2.3: Calculate the joint temporal and spatial value of the perceived information block: VUEs selectively upload the perception information block with the highest time-space joint value. Each VUE periodically selects and uploads the self-perception information block with the highest time-space joint value to its associated fog node server or cloud server. After receiving the uploaded perception information blocks from VUE members within the cluster, the cloud server and fog node F-AP readjust, integrate, infer, and map the collected perception information blocks to achieve collaborative perception between VUEs. Step 3: Calculate the radar mutual information (MI) and total delay achieved by VUE under different clustering modes; Step 4: With the goal of maximizing user satisfaction, construct an optimization model for joint clustering, selection of perception information blocks, allocation of resource blocks (RBs), and optimal allocation of CPU cycle frequency. Step 5: Establish a VUE clustering mode and perceptual information block selection model based on multi-agent deep reinforcement learning, and perform VUE clustering mode and perceptual information block selection. Step 6: Train the reinforcement learning model and find the optimal clustering pattern and perceptual information block selection scheme for VUE; Step 7: Perform joint radio and compute resource allocation for VUE; Step 8: Input the network state of subsequent time slots into the model, repeat steps 6 to 8, and output the communication, perception and computing collaborative resource allocation scheme in the vehicle-to-everything network that maximizes long-term satisfaction while satisfying its time delay and radar mutual information constraints.

2. The multi-dimensional resource collaborative management method in a vehicle-to-everything (V2X) network according to claim 1, characterized in that, Step 1, which involves constructing a cluster-based cooperative perception network model in the vehicle-to-everything (V2X) network, includes: The collaborative sensing network model is composed of A cloud server, Composed of several wireless remote radio frequency units, multiple vehicle user (VUE) units are configured to associate with the same fog node server or cloud server via RRH or fog node, forming a collaborative cluster and uploading their local sensing information blocks to the fog node server or cloud server for collaborative computation. This is called a collaborative sensing task. ; Configure VUE k connection selection as :if VUE k is associated with the cloud model through RRHs; conversely, if , VUE k selects to access F-AP .

3. The multi-dimensional resource collaborative management method in a vehicle-to-everything (V2X) network according to claim 1, characterized in that, Step 3 describes the calculation of the radar mutual information (MI) and total latency achievable by VUE under different clustering modes: Step 3.1: Calculate the radar signal to interference noise ratio at VUE. And achievable radar mutual information (MI), a perception model is established based on radar information estimation theory, and each VUE is set to be able to perform radar detection by continuously transmitting integrated OFDM waveforms through an onboard transmitter equipped with ISAC technology. The ratio of received radar signal to interference noise for: ， in, and For VUE Other VUE The transmission power, here It belongs to resource block RB and The assignment vector, yes The Fourier transform, in At that moment, from another VUE To VUE The channel gain of the radar receiver is denoted as In the time slot Inner VUE The achievable radar mutual information (MI) is defined as follows: ， Step 3.2: Calculate the total latency achieved by the VUE under different clustering modes. When the VUE selects the fog node clustering mode, the communication rate achieved is calculated using the following formula: ， in, This refers to the bandwidth size of the resource block. Assign vectors to binary resource blocks. For VUE The transmission power, For use with Vue Vue.js that uses the same resource blocks The transmission power, for Always in the resource block From VUE To F-AP Channel gain, for Always in the resource block From other Vue To F-AP Channel gain; For noise power, the formula for the communication rate achieved when VUE selects the cloud server clustering mode is: ， in, It is the vector detected by the Minimum Mean Square Error (MMSE). It is time Inside, from VUE To its associated RRHs The channel gain, and From VUE To VUE The associated channel gain of RRHs, in a cluster, VUE The communication latency is the maximum latency among all VUE members' upload latency, calculated using the following formula: ， in, This is the fronthaul latency, measured by the actual network communication capabilities of the operator. The data size for each perceptual information block is in bits. For VUE Perceptual information block Binary selection variables: VUE Upload the sensing information block ,on the contrary, This indicates VUE Will not upload perception information blocks VUE The computation delay is: ， in, and These are the maximum CPU cycle frequencies for the cloud server and each fog node server, respectively. To be assigned to VUE CPU cycle frequency, VUE The total latency is: ， in, and These are local compression latency, communication latency, and task computation latency, respectively.

4. The multi-dimensional resource collaborative management method in a vehicle-to-everything (V2X) network according to claim 3, characterized in that, Step 4 describes an optimization model that aims to maximize user satisfaction, involving joint clustering, selection of sensing information blocks, allocation of resource blocks (RBs), and optimal allocation of CPU cycle frequencies. The collaborative sensing optimization problem is described as follows: ， in, It is a satisfaction function. It is the total duration. and It is a satisfaction weighting parameter to balance perceived information. The impact of spatiotemporal value and total task delay on satisfaction. It is the threshold for maximum latency tolerance. It is the threshold for minimum radar mutual information tolerance, a constraint. It is the task latency requirement, determined by the task's maximum tolerable latency, and constrains it. This means that each VUE should upload non-overlapping perception information to make reasonable use of computing resources and constrain [the system / mechanism]. This means that each VUE can choose an offload node to allow VUEs to form reasonable clusters, constraining... This means that each VUE selects a resource block (RB) for uploading, and the resource block RB can be reused by multiple VUEs in different clusters, constraining... The CPU cycle frequency allocated to fog nodes (F-AP) and cloud servers was limited, constraining... It is the minimum perceived performance threshold for VUE with ISAC enabled.

5. The multi-dimensional resource collaborative management method in a vehicle-to-everything (V2X) network according to claim 4, characterized in that, Step 5 describes the establishment of a VUE clustering pattern and perceptual information block selection model based on multi-agent reinforcement learning: Step 5.1: Treat each VUE as an agent and construct an attention-assisted multi-agent DDPG algorithm model. The local state feature vector is defined as: ， in, Let k be the maximum tolerable delay for VUE. For VUE The current coordinates, for VUE Time The time value of the perceived message, Fog Node and cloud servers in time Remaining computing resources at that time On a cloud server, the action of agent k is defined as the ratio of the global latency satisfaction of the previous action to the impact of the previous action. ， in, Indicates the sequence number of the fog node being connected; Step 5.2: Construct the instant reward function for each VUE. The reward function is defined as: ， Step 5.3: Construct the state features of the multi-agent Markov game model using the state feature vectors of the agents: Define a Markov game model for each agent. Defined as: ， in, It is the number of intelligent agents. For state space, It is the action space. It is the system's instantaneous reward function determined by all reward agents. It is a discount factor for exploration; Step 5.4: Implement collaboration among agents using the multi-agent DDPG model, and then adjust the commenter network for each agent using a multi-layer perceptual network with an attention mechanism; each agent in the attention-based multi-agent DDPG algorithm... Each system has a network of critics responsible for finding deterministic strategies based on local observations, given random noise. In time The selected action The formula is defined as: ， in, For actor network parameters, It's VUE The local observation status, By inputting critic network parameters and VUE Local observation status The resulting deterministic policy involves each actor network adjusting its parameters along the gradient direction. To maximize the reward function, the formula is: ， in, It is a replay cache used to store the experience tuples of all agents. It is an action-value function built by an attention-assisted commentator network; Establish a multi-layer perception network based on intelligent agents. intelligent agent Observations, attention weights and attention value The calculation formulas are as follows: ， ， in, and constitute and bilinear mapping, It is a ReLU function. This is the transformation matrix. The commentator network is set to update by minimizing the following loss function, as shown in the formula: ， in, These are the network parameters of the critics. It consists of a state value, action value, immediate reward, and next state value group. The four-tuple formed It is the next state value. It is the target action value function inferred by a critic network assisted by target attention. It is the target action obtained from the network of target actors.

6. The multi-dimensional resource collaborative management method in a vehicle-to-everything (V2X) network according to claim 5, characterized in that, Step 6 describes training the reinforcement learning model and finding the optimal clustering pattern and perceptual information block selection scheme for VUE: Step 6.1: Initialize the actor network, attention mechanism critic network, target actor network, target critic network, and replay cache. and training batch size; Step 6.2: Initialize state s in each training round, and observe the current state of the agent in each time slot. ; Step 6.3: If the current round is in the pre-training phase, use the formula... Explore VUE clustering and message block selection, then return to step 6.2 until all time slots and agents have completed execution; otherwise, use the formula. Select an action to perform VUE clustering and message block selection, and return to step 6.2 until all time slots and agents have completed their execution; Step 6.4: Calculate the instant reward and the next state Then the experience tuple Stored in experience In the pool; Step 6.5: For all agents, randomly select a batch of tuples from the replay buffer to update the actor network, the attention mechanism, and the commentator network respectively; Step 6.6: Update the target's actor and attention mechanism commentator network using a soft update method, and repeat step 6.

6. Step 6.6 Continue until all agents have finished executing.

7. The multi-dimensional resource collaborative management method in a vehicle-to-everything (V2X) network according to claim 6, characterized in that, Step 7 describes the joint allocation of radio and compute resources for the VUE: Step 7.1: Perform optimal allocation of resource blocks (RBs) for VUEs based on an improved exchange matching algorithm. This involves providing an improved exchange matching algorithm to allocate resource blocks (RBs) to VUEs. For the initial matching, a preference list is first constructed for each VUE and resource block (RB) in descending order. and Then, the mismatched VUEs iteratively make requests to their optimal resource block RB. The resource block RB can accept its preferred VUEs and reject other VUEs until all VUEs are matched. The execution steps are as follows: Step 7.1.1 Initialize Vue in descending order and resource blocks RB Preference list and ; Step 7.1.2 Initialize the mismatched Vue collection ; Step 7.1.3 When the VUE's preference list and the set of mismatched VUEs are not empty, for all VUEs in the set of mismatched VUEs, iteratively make matching requests to the resource block RBs that have not yet rejected them, in descending order of the product of the current resource block RB's achievable uplink delay and radar mutual information. Step 7.1.4 For all resource blocks RB, if RB Received VUE If the matching request is received, then RB Keep with VUE The resource block RBs accepts the matching request of the VUE with the largest quotient of uplink latency and radar mutual information and rejects the matching requests of other VUEs. Step 7.1.5 After the iteration is complete, output the optimal resource block (RB) allocation result; Step 7.2: Perform optimal allocation of computational resources for VUE based on the improved interior-point method: Based on the optimal allocation result of RB obtained in step 7.1, further calculation of resource optimization allocation is performed. The formula for the resource optimization allocation problem is as follows: ， This paper presents an improved computational resource allocation algorithm based on the interior-point method to solve the problem. A penalty function is constructed as follows: ， in, This is a feasible solution, and the execution steps are as follows: Step 7.2.1 Initialize the penalty factor Feasible solution Maximum number of iterations Maximum tolerance error and update coefficient ; Step 7.2.2 Calculate for each VUE k ; Step 7.2.3 If Obtain the unconstrained computational resource allocation results under KKT conditions. Otherwise, within the maximum number of iterations and the maximum tolerance error range, obtain the current... and The penalty function below And calculate the remaining computing resources ; Step 7.2.4 For all Vue Update VUE's gradient values and the results of the allocated computing resources , Step 7.2.5 Update ,set up ,in The decreasing coefficient; Step 7.2.6 After the iteration is completed, output the optimal computing resource allocation result.

8. The multi-dimensional resource collaborative management method in a vehicle-to-everything (V2X) network according to claim 7, characterized in that, Step 8 involves inputting the network state of subsequent time slots into the model: Repeat steps 6 to 8 to output the communication, perception and computing collaborative resource allocation scheme in the vehicle-to-everything (V2X) network that maximizes long-term satisfaction while satisfying its latency and radar mutual information constraints. The collaborative resource allocation scheme includes selecting the optimal resource blocks, perception information blocks, clusters and the CPU frequency of cloud servers or fog node servers.

Images