Intelligent Slicing and Scheduling Method for Unmanned Cluster Network Resources

Abstract

The present invention discloses an intelligent slicing and scheduling method for unmanned cluster network resources. Through topology prediction and service clustering of non-backbone cluster networks, the types and quantities of backbone cluster resource slices are divided, and the time scales and operation sequences of the orchestration of each slice are determined. At a large time scale exceeding the time threshold, based on the joint prediction of the access relationship and service trend of the cluster network, the division and configuration between resource slices are implemented, and redundant resources are adaptively reserved in the slices to adapt to the fluctuation of service requirements and reduce the overhead of resource re-slicing. At a small time scale below the time threshold, based on service performance evaluation, in-slice resource allocation is implemented, which improves the operation efficiency of this slice while reducing the impact on the services of other slices. The present invention introduces the prediction of network access topology and mobile service requirements based on resource feasibility into slice management, making resource configuration more flexible and agile and adaptable to the changes in air cluster ad hoc networks.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technologies, and particularly to an intelligent slicing and scheduling method for unmanned cluster network resources. Background Art

[0002] The statements in this section only provide background information related to the present disclosure and may not constitute prior art.

[0003] Facing the dynamic topological features and time-varying communication requirements of multiple ad-hoc networks in the air, a networking architecture based on backbone clusters is constructed, and the communication access and resource scheduling of backbone clusters are focused on. Since the resources of the self-organizing network are limited when constructing group connections, there are limitations in the forwarding capacity of nodes and the bandwidth capacity of links. In order to ensure the quality of service (QoS) requirements between each group, it is necessary to reasonably allocate and reuse the limited resources to provide stable and reliable group communication. Therefore, under the advantages of network slicing technology in meeting differentiated service requirements, QoS guarantee, reliability, etc., it has become the main architecture in ad-hoc network services.

[0004] However, there are some challenges and limitations in the architecture of traditional network slice resource arrangement. First, if all services are carried on the same ad-hoc network, with the increase in the number of nodes and service types, the management and maintenance of traditional network slices become more complex for meeting different network transmission requirements. Second, when the service grouping access location or network topology changes, the deficiency of traditional network slices lies in their lack of dynamic flexibility, resulting in limited flexible configuration of network resources and difficulty in coping with rapidly changing requirements. In addition, due to the differences in the size and time scales of slice resource adjustment, the periodic reconfiguration of traditional network slices may lead to large interruptions or transition delays in the slice network.

[0005] Currently, the research on network slice reconfiguration mainly includes the following two types:

[0006] (1) Network Slicing Reconfiguration Based on Deep Reinforcement Learning under Variable Number of Service Function Chains (Reference: Tokuda K, Sato T, Oki E, "Network slice reconfiguration with deep reinforcement learning under variable number of service function chains", Computer networks, Volume 224, 2023, 109636, ISSN 1389-1286): This paper proposed the DNSR-DP model, a deep reinforcement learning model that combines virtual states / operations and online learning strategies. It can adapt to the dynamic changes in the number of service function chains, reduce the reconfiguration cost, and achieve stable learning. However, this model does not consider the impact of re-slicing on service interruption. The present invention further innovates by introducing resource reservation scheduling based on network topology and service requirements in resource allocation to reduce the impact of reconfiguration on service interruption and improve the flexibility and adaptability of resource allocation.

[0007] (2) Dynamic Slicing Reconfiguration for Virtualized 5G Networks Using ML Forecasting of Computing Capacity (Reference: Juan Sebastian Camargo, Estefanía Coronado, Wilson Ramirez, Daniel Camps, Sergi Sánchez Deutsch, Jordi Pérez-Romero, Angelos Antonopoulos, Oscar Trullols-Cruces, Sergio Gonzalez-Diaz, Borja Otura, Giovanni Rigazzi, "Dynamic slicing reconfiguration for virtualized 5G networks using ML forecasting of computing capacity", Computer Networks, Volume 236, 2023, 110001, ISSN 1389-1286): This paper proposed a machine learning (ML) model that can predict network slice traffic and dynamically adjust computing resources to meet the resource requirements of different slices and ensure service quality. Combining hysteresis rules, this model can accurately predict resource saturation and optimize resource allocation. However, this method does not consider the migration and configuration costs during the re-slicing process, which may lead to unnecessary waste of resources. Summary of the Invention

[0008] The object of the present invention is to provide an intelligent slicing and scheduling method for unmanned cluster network resources, aiming at the problems of limited network resource allocation and difficulty in quickly responding to demand changes under the current group task drive. A reconstruction scheme with small time scale and large time scale is introduced in slice management. By using the reconfiguration overhead function, the resource allocation strategy is optimized, allowing part of the resources of the current service to be reallocated as redundant resources for future service requirements. This strategy aims to achieve an effective balance between a moderate sacrifice of short-term service performance and a forward-looking reserve of long-term resource requirements, thus solving the above problems.

[0009] The technical solution of the present invention is as follows:

[0010] An intelligent slicing and scheduling method for unmanned cluster network resources, comprising:

[0011] Step S1: Dynamically optimize the reconfiguration threshold between slices by predicting the movement trajectories and service requirements of groups, and trigger cross-slice resource reconfiguration;

[0012] Step S2: Allocate adaptive resources between slices and make redundant reservations based on the prediction of group mobility and service requirements;

[0013] Step S3: Construct a non-cooperative resource game model and design a distributed learning algorithm to achieve the Nash equilibrium configuration of resource allocation;

[0014] Step S4: Dynamically update the small time scale threshold based on service performance, and adjust the resources within the slice to optimize the quality of service.

[0015] Further, the step S1 includes:

[0016] Step S11: Collect the movement trajectory data of all groups and extract features;

[0017] Step S12: Group the groups by a clustering method based on the Dirichlet process to identify clusters with similar movement patterns;

[0018] Step S13: Assign an initial cluster to each group, iteratively update the cluster assignment of each group, and use the Gibbs sampling method to generate samples from the multi-dimensional probability distribution;

[0019] Step S14: Estimate the transition kernel of the group at the current position;

[0020] Step S15: Predict the next access position of the group according to the transition kernel;

[0021] Step S16: Trigger slice reconfiguration when the matching degree between the predicted access topology and slice resources exceeds the large time scale trigger threshold.

[0022] Further, the movement trajectory data includes different locations visited in clusters and the timestamps of the visits, and the extracted features include the movement patterns and visit frequencies of the clusters.

[0023] Further, the step S15 includes:

[0024] Step S151: Calculate the transfer probability of the cluster from the current access location i to the location j ;

[0025] Step S152: Select the location corresponding to the maximum transfer probability j as the next access location.

[0026] Further, the transfer kernel is calculated using the following formula:

[0027]

[0028] where:

[0029] represents the transfer kernel of the cluster at the current location;

[0030] is the number of samples generated in Gibbs sampling; represents the th sample;

[0031] is the total number of clusters;

[0032] is the number of other clusters in the cluster c except the cluster ;

[0033] is the Dirac measure of the Dirichlet distribution;

[0034] is the current concentration parameter of the DP mixture model;

[0035] is the current base distribution.

[0036] Further, the transfer probability is expressed by the following formula:

[0037]

[0038] where:

[0039] represents the transfer probability of the cluster from the current location i to the location j ;

[0040] is the sub - clustering at the current position in time; To determine the next access position of the sub - clustering t the position that maximizes the transition probability needs to be found ; j ;

[0041] represents the conditional probability that the sub - clustering transfers from the current position i to the position j .

[0042] Furthermore, the step S2 includes:

[0043] Step S21: Construct a slice throughput model for evaluating transmission efficiency and calculate the expected throughput of the communication pair;

[0044] Step S22: Estimate the average throughput of each slice network by the Monte Carlo method;

[0045] Step S23: Estimate the expected average throughput of the slice and reserve redundant bandwidth and computing resources during resource allocation.

[0046] Furthermore, the step S3 includes:

[0047] Step S31: Define the strategy selection space for each slice, and require that the computing resource and bandwidth allocation satisfy the constraints of the slice resource game;

[0048] Step S32: Design a fully distributed learning algorithm to update the mixed strategy of the slice until it converges to the Nash equilibrium.

[0049] Furthermore, the constraints include:

[0050] a) The total computing resources for redeploying all slices cannot exceed the total remaining resources of the current physical network;

[0051] b) The total path delay within each slice is less than or equal to the delay constraint;

[0052] c) The throughput of each slice is greater than or equal to the minimum throughput limit;

[0053] d) The link bandwidth resources for redeploying all slices cannot exceed the total remaining bandwidth resources of the current physical network;

[0054] e) The number of VNFs of the slices deployed on each node cannot exceed one;

[0055] f) The number of VNFs deployed for each slice is greater than the number of VNF functional requirements;

[0056] g) Whether physical isolation between slices is required.

[0057] Further, the step S4 includes:

[0058] Step S41: Calculate the small time-scale trigger threshold by the fuzzy evaluation method;

[0059] Step S42: Under the condition of the small time-scale trigger threshold, optimize the resource allocation of the in-chip service flow through the in-chip optimization model, and adjust the link variables and node computing resources.

[0060] Compared with the existing technologies, the beneficial effects of the present invention are:

[0061] The present invention introduces a reconstruction scheme of small time-scale and large time-scale in slice management. By using the reconfiguration overhead function, it optimizes the resource allocation strategy, allowing the resources of a part of the current service to be reallocated as redundant resources for future service requirements. This strategy aims to achieve an effective balance between a moderate sacrifice of short-term service performance and a forward-looking reserve of long-term resource requirements. Description of the Drawings

[0062] Figure 1 It is a flowchart of the intelligent slice and scheduling method for unmanned cluster network resources provided by the present invention;

[0063] Figure 2 It is a schematic diagram of the scenario of the embodiment of the present invention;

[0064] Figure 3 It is a time-scale diagram. Detailed Embodiments

[0065] It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of additional identical elements in the process, method, article or device including the element.

[0066] First of all, it should be noted that the present invention intelligently clusters the UAV swarm service into corresponding network slices by dynamically analyzing business requirements. At each decision moment on a large time scale, resource reconfiguration between slices is performed to optimize the overall performance of the backbone network. In the small time scales included in the large time scale, service-oriented scheduling of intra-slice resources is carried out to meet the service transmission requirements. It is specifically divided into the following parts:

[0067] A1. In each decision cycle on a large time scale, detailed resource scheduling is carried out by embedding multiple sub-cycles of small time scales to achieve a rapid response to immediate business requirements. The trigger cycle lengths of the dual time scales are designed to be non-fixed and can be flexibly adjusted according to the dynamic changes in business requirements;

[0068] A2. By predicting the swarm movement trajectory, access topology, and business requirements through historical data and current trends, the large time scale trigger threshold based on resource utilization and service satisfaction rate is dynamically optimized to respond to the throughput fluctuations of time-varying business requirements, and the large time scale reconfiguration between slices is only triggered when the threshold is exceeded. On the premise of responding to changes in business requirements, the frequency of resource reconfiguration between slices is reduced. The setting of the variable large time scale trigger threshold distinguishes business types. That is: for services that are more sensitive to quality of service, the tolerance of their thresholds is lower. When the matching degree between the predicted business requirements and the slice services exceeds the threshold, the slice reconfiguration process will be triggered;

[0069] A3. In the process of adaptive resource allocation between slices, based on the prediction of future network access topology and business requirements, through redundant configuration of resources within the slice, the divided slices are made capable of adapting to changes in access topology and business requirements. That is: by moderately reducing the current network service quality, the flexibility of future resource scheduling and business adaptability of the slice are guaranteed, and the probability and cost of slice reconfiguration are reduced. Consider the impact of changes in the subsequent swarm access trajectory on the slice transmission rate. According to the dynamic network load and business requirements, a strategy for reserving redundant node computing resources and link bandwidth resources is formulated. This strategy ensures that when the swarm migrates to a new access point in the future at the cost of the current reconfiguration cost, the slice to which the access point belongs has sufficient resources to be used for scheduling, avoiding resource reconfiguration between slices;

[0070] A4. Construct a non - cooperative resource game model aimed at dynamically adjusting the resource allocation strategy. This model defines the strategy space available to each slice, that is, the range of reserved resource amounts. Apply a fully distributed learning algorithm that allows each slice to independently adjust its strategy based on local information, gradually achieving the convergence of the system to the Nash equilibrium state. Achieve the optimal resource allocation of each network slice in the Nash equilibrium state. This model considers the network utility and resource overhead of each slice to ensure the efficiency and fairness of resource allocation, thereby maximizing the total revenue of each slice. On a large time scale, when evaluating the total revenue of each slice, the revenue mainly comes from the throughput of the service slice. Use the Monte Carlo method to evaluate the connectivity of the slice network under specific node distributions and transmission ranges. By simulating the service flow paths in the clustered mobile mode, accurately calculate the average throughput of the slice network, providing data support for dynamically adjusting the service access points and resource allocation. The overhead includes reconfiguration delay and resource supply mismatch. The former is affected by the amount of resource adjustment and involves node configuration, virtual network function deployment, and routing reconstruction delay. A high degree of the latter mismatch will cause service interruption and additional delay costs;

[0071] A5. Based on the comprehensive evaluation of service performance such as communication delay, rate, and packet loss rate, dynamically update the resource scheduling threshold on a small time scale. This threshold is calculated by the fuzzy evaluation method, which considers the differences in service quality requirements for delay, bandwidth, and packet loss rate of different service types. Under the condition that the threshold is triggered, by adjusting and configuring the resources within the slice to make it as adaptable to the service requirements as possible, reduce the performance evaluation value below the threshold. By increasing the frequency of fine - tuning of resource scheduling, achieve more frequent reconstruction in a short time. Reconfigure the flow link and node computing resource allocation variables to control the in - slice reconfiguration cost and ensure that the cost remains at a low level.

[0072] The features and performance of the present invention will be further described in detail below in conjunction with the embodiments.

[0073] Embodiment 1

[0074] To facilitate the understanding of the present invention by ordinary technicians in the art, the following definitions are made for the technical terms involved in the present invention:

[0075] 1. Backbone network

[0076] In the backbone ad - hoc network composed of multiple manned and unmanned aircraft, the backbone network plays a crucial role as a communication hub connecting different service clusters. In this network architecture, the edge access nodes are responsible for the access of clustered communication, while the relay forwarding nodes are responsible for the forwarding and processing of data packets.

[0077] 2. Variable time scale

[0078] Based on the time-varying dynamic service requirements of each slice, the control center respectively performs slice resource partitioning and sub-slots in the dimension of large time slots and fine-tunes the resource scheduling within the slice. In each large time slot, the total bandwidth is divided into k mutually orthogonal frequency bands , and each slice network is assigned a unique frequency band. And in each sub-slot, physical resource blocks are allocated for the in-slice services.

[0079] 3. Service Slice

[0080] Within the coverage area of the backbone network, the clustering algorithm is first applied to classify all clusters by service. Communication services with strict requirements for reliability and latency are assigned to the slices with the highest priority; services that require high transmission rates and bandwidth are classified into the slices with ordinary priority; services without specific quality of service (QoS) requirements are assigned to the slices with low priority.

[0081] 4. Slice Reconfiguration Overhead

[0082] The slice reconfiguration overhead metrics include the delay of re-slicing and the degree of resource supply mismatch. The re-slicing delay is affected by the amount of slice resource adjustment and covers the delays of node configuration, virtual network function deployment, and service route reconstruction. The degree of resource supply mismatch includes the additional communication interruption overhead caused by the mismatch between resources and task requirements.

[0083] The present invention mainly includes two parts: First, the present invention jointly predicts the movement trajectory and service access trend of each cluster in view of the decrease in transmission rate and potential service interruption caused by cluster movement, and calculates the reserved required bandwidth and computing resources accordingly. This method aims to ensure transmission efficiency and service continuity by pre-configuring resources. Second, it involves dealing with the time-scale differences in resource allocation within and between slices. By implementing a reconfiguration strategy based on variable time scales, different triggering thresholds are set for different slices, and the in-slice allocation of resources is optimized within the slice resource scheduling strategy space on a large time scale to improve the efficient utilization of resources.

[0084] In this embodiment, specifically, please refer to Figure 1 , Figure 2 and Figure 3 , an intelligent slice and scheduling method for unmanned cluster network resources, including:

[0085] Step S1: Dynamically optimize the reconfiguration threshold between slices and trigger cross-slice resource reconfiguration by predicting the clustered movement trajectory and service requirements; that is, by jointly predicting the cluster access relationship and service trend, real-time optimize the reconfiguration threshold between slices to respond to the throughput fluctuations of time-varying service requirements, and trigger large-time-scale reconfiguration between slices when the threshold is exceeded to ensure the dynamic schedulability of network slice resources and respond to the real-time changes of service requirements.

[0086] Step S2: Allocate adaptive resources between slices and make redundant reservations based on the prediction of clustered mobility and service requirements; that is, perform adaptive resource allocation between slices, including dynamically adjusting spectrum and computing resources according to the priority of service slices. Based on the prediction of network access topology and service requirements, reserve redundant resources by moderately reducing the current service quality of services to ensure the abundance and flexibility of resources in future time periods to adapt to expected service requirement changes and clustered mobility.

[0087] Step S3: Construct a non-cooperative resource game model and design a distributed learning algorithm to achieve the Nash equilibrium configuration of resource allocation; that is, construct a non-cooperative resource game model, aiming to achieve the optimal resource allocation of each network slice in the Nash equilibrium state by dynamically adjusting the resource allocation strategy. This model considers the network utility and resource overhead of each slice to ensure the efficiency and fairness of resource allocation, thereby maximizing the total revenue of each slice.

[0088] Step S4: Dynamically update the small-time-scale threshold based on service performance and adjust the intra-slice resources to optimize the service quality; that is, based on the service performance evaluation of communication delay, rate, and packet loss rate, dynamically update the small-time-scale trigger period threshold, and finely adjust the slice resources in each micro-slot to optimize the data volume of service flows that meet the service quality standard, enhance the operation efficiency of this slice, and minimize the interference to the services of other slices at the same time.

[0089] In this embodiment, a clustered movement trajectory prediction model is established and the large-time-scale trigger threshold is dynamically updated, including the following steps (that is, the above-mentioned Step S1 includes the following steps):

[0090] Step S11: Collect the movement trajectory data of all clusters and extract features; the movement trajectory data includes different locations visited by the clusters and the timestamps of the visits, and the extracted features include the movement patterns and access frequencies of the clusters.

[0091] Step S12: Group the clusters by a clustering method based on the Dirichlet process to identify clusters with similar movement patterns; it should be noted that Step S12 specifically includes the following sub-steps: for each cluster and each sample , in the th iteration, the cluster The clustering assignment is based on the current concentration parameters of the DP mixture model and the base distribution for sampling. The concentration parameter controls the tightness of the clustering, while the base distribution provides prior knowledge of the cluster centers. The generation process of the clustered assignment samples is . denotes the Dirichlet process, and its sample value is the probability distribution of the partition. In this process, each cluster has a new clustering assignment;

[0092] Step S13: Assign an initial cluster to each cluster, iteratively update the clustering assignment of each cluster, and generate samples from the multi-dimensional probability distribution using the Gibbs sampling method;

[0093] Step S14: Estimate the transition kernel of the cluster at the current position; that is, the probability distribution of transitioning from the current position to the next position. The predicted transition kernel can be calculated by the following formula:

[0094]

[0095] where:

[0096] represents the transition kernel of the cluster at the current position;

[0097] is the number of samples generated in the Gibbs sampling; represents the th sample;

[0098] is the total number of clusters;

[0099] is the number of other clusters in cluster c except for cluster ;

[0100] is the Dirac measure of the Dirichlet distribution;

[0101] is the current concentration parameter of the DP mixture model;

[0102] is the current base distribution;

[0103] Step S15: Predict the next access position of the cluster according to the transition kernel; in this embodiment, specifically, the step S15 includes:

[0104] Step S151: Calculate the cluster from the current access positioni Transfer to position j The transition probability; that is, for each possible next position j, calculate the cluster The transition probability of transferring from the current position i to the position j can be expressed by the following formula:

[0105]

[0106] Where:

[0107] Represents the cluster Transfer from the current position i Transfer to position j The transition probability;

[0108] Is the cluster At time t The current position; to determine the next access position of the cluster The position that maximizes the transition probability needs to be found j ;

[0109] Represents the cluster Transfer from the current position i Transfer to position j The conditional probability.

[0110] Step S152: Select the position corresponding to the maximum transition probability j As the next access position; it should be noted that to determine the next access position of the cluster The position j that maximizes the transition probability needs to be found. The optimization objective is as follows:

[0111]

[0112] Where: Represents the next predicted position of the cluster ; The operation finds the position j that maximizes the internal sum; the internal summation is the accumulation of the transition probabilities from all possible current positions i to the position j; select the position with the maximum transition probability as the next predicted position of the cluster ; Represents the total access position;

[0113] Step S16: Trigger slice reconfiguration when the matching degree between the predicted access topology and the slice resources exceeds the large time-scale trigger threshold;

[0114] It should be noted that the cluster assignments and Dirichlet parameters are continuously updated through an iterative process. In each iteration, a Gibbs sampler is used to generate new cluster assignments for the clusters, and these assignments are updated as follows:

[0115]

[0116]

[0117] represents the cluster assignment in the k-th iteration; represents the cluster assignment in the (k + 1)-th iteration; is the number of members of sample b in cluster c in the k-th iteration, is the set of clusters, is the cluster of the trajectory length; represents the Dirichlet parameter in the (k + 1)-th iteration (i.e., the concentration parameter of the DP mixture model); represents the total Dirichlet parameter; represents the set of observed data of all clusters belonging to cluster c; The operation searches for the Dirichlet parameter that minimizes the internal sum;

[0118] Thus, the possible access trajectories for the next stage of the cluster are obtained , and the traffic within the prediction range of the backbone network is calculated throughput fluctuation , and the global resource utilization rate is comprehensively considered together with the service satisfaction rate . The triggering condition for large time-scale slice reconfiguration is:

[0119]

[0120] Among them, is the traffic a preset parameter.

[0121] In this embodiment, specifically, step S2 involves modeling the slice utility function and the reconfiguration overhead function, which specifically includes:

[0122] Step S21: Construct a slice throughput model for evaluating transmission efficiency , and calculate the expected throughput of the communication pair;

[0123] Step S22: Estimate the average throughput of each slice network through the Monte Carlo method;

[0124] Step S23: Estimate the expected average throughput of the slice, and reserve redundant bandwidth and computing resources during resource allocation;

[0125] It should be noted that the slice throughput model defines a temporary network throughput capacity for a communication pair , where belongs to the set of all possible node pairs and satisfies specific conditions . If the end-to-end throughput that a communication pair can expect is bits per second, then this communication mode has a throughput capacity of and a routing function . For a given source-destination pair , the resulting route consists of the set of edges included in the sequence , where ; ;

[0126] Assume that the reconfigured slice topology changes randomly with the movement of clustering. The expected throughput of a communication pair can be regarded as a random variable. Therefore, model the throughput capacity as a random variable and calculate its expected value , where . Since the path between two nodes is composed of a series of channel schedules, the throughput capacity is a concave function, reflecting the bottleneck effect, that is, the available throughput of the source-destination pair is limited by the node with the lowest bandwidth on the path. From to , the minimum number of available channels between two nodes along the path is represented as , represents the number of channels allocated on link e; the throughput capacity on its path satisfies: , where is the maximum calculation rate when all nodes are equal, is the total number of channels used;

[0127] Considering the ongoing traffic of all communication pairs within the slice, the connectivity of the network topology can be expressed as:

[0128]

[0129] where the load function , representing the degree to which a certain edge shares with other ongoing traffic, that is represents the degree to which a certain edge shares with other ongoing traffic; if each node is limited to transmitting on one channel, then Refers to in and the path availability between, the throughput capacity of the temporary network can be regarded as a direct measure of the connection probability;

[0130] To evaluate the network connectivity under a specific node distribution and transmission range, the Monte Carlo method is used to estimate the average throughput of each slice network. The expected slice average throughput expression can be obtained from the linear property of the expected value:

[0131]

[0132] where is the node location, is the specific set of node placements in the slice network; represents the expected value;

[0133] Furthermore, for slice a large time-scale delay overhead can be expressed as:

[0134]

[0135] represents the total traffic; interruption delay refers to the traffic request response failure time, due to changes in the traffic access trajectory or insufficient resources, nodes and links need to be reconfigured. can be expressed as:

[0136]

[0137]

[0138]

[0139]

[0140] where represents the slice in the traffic request for re-resource configuration waiting time, determined by the next large time reconfiguration time decided; represents the slice in the traffic time required to redeploy resources; represents the virtual network function information migration time; represents the slice in the traffic access node ; represents the slice nodes the change in computing resources before and after reconfiguration, represents the slice links the change in bandwidth resources before and after reconfiguration, is the unit deployment delay, 、 is the balance factor of resource deployment overhead; represents the slice nodes ; represents the slice total nodes; 、 、 respectively represent the occupied bandwidth, number of hops, and transmission delay of the path , is the balance coefficient; represents the slice total paths.

[0141] In this embodiment, specifically, the step S3 of establishing a service slice resource allocation model aims to optimize network resource allocation, and specifically includes:

[0142] Step S31: Define the policy selection space for each slice, requiring that the computing resource and bandwidth allocation satisfy the constraints of the slice resource game;

[0143] Step S32: Design a fully distributed learning algorithm to update the mixed strategy of the slice until it converges to the Nash equilibrium.

[0144] In this embodiment, specifically, the constraints include:

[0145] a) The total computing resources redeployed for all slices cannot exceed the total remaining resources of the current physical network;

[0146] b) The total delay of the paths within each slice is less than or equal to the delay constraint;

[0147] c) The throughput of each slice is greater than or equal to the minimum throughput limit;

[0148] d) The link bandwidth resources redeployed for all slices cannot exceed the total remaining bandwidth resources of the current physical network;

[0149] e) The number of VNFs of the slices deployed on each node cannot exceed one;

[0150] f) The number of VNFs deployed in each slice is greater than the number of VNF functional requirements;

[0151] g) Whether physical isolation between slices is required;

[0152] It should be noted that within the framework of non - cooperative games, each network slice independently formulates a resource scheduling strategy. Through Nash equilibrium analysis, it is ensured that each slice, while considering the strategies of other slices, selects an optimal strategy to maximize its own utility. The composition of the game specifically includes participants: slices , among which . Strategy selection space: , denotes the action of slice . denotes the computing resource allocation scheme of slice , denotes the bandwidth allocation scheme of slice ;

[0153] For each slice, the improvement in revenue mainly depends on two key factors: First, bandwidth and computing resource allocation are the core. More resource allocation can theoretically increase throughput and thus improve revenue. Second, effective management of service latency, including reducing resource deployment overhead and avoiding service interruptions caused by insufficient resources, is crucial for maintaining the total revenue. Therefore, slices must adopt strategies to reduce unnecessary latency and overhead to ensure maximum revenue.

[0154] Considering the mutual resource competition between slices, the revenue function of slice is expressed as:

[0155]

[0156] Among them, denotes the slice utility function, which depends on the throughput capacity of slice ; denotes the slice overhead function. , are balance factors;

[0157] The constraint conditions of the slice resource game are as follows:

[0158] 1) The total computing resources redeployed by all slices cannot exceed the total remaining resources of the current physical network:

[0159]

[0160] Among them, is the remaining computing capacity of node ; denotes the total number of slices; denotes the quantity of the r - th type of resource allocated on the i - th node; ​Represents the total resources.

[0161] 2) For each slice The total delay of the internal path is less than or equal to the delay constraint :

[0162]

[0163] Among them, Represents the internal path of the slice transmission delay; is the slice delay threshold.

[0164] 3) For each slice The throughput is greater than or equal to the minimum throughput limit :

[0165]

[0166] Among them, is the slice throughput threshold.

[0167] 4) The link bandwidth resources for redeployment of all slices cannot exceed the remaining total bandwidth resources of the current physical network:

[0168]

[0169] Among them, is the remaining bandwidth capacity of the link in the slice ; is the bandwidth resource allocated to the slice on a specific link; is the bandwidth resource allocated to the slice on the link ;

[0170] 5) The number of VNFs deployed for each slice on each node cannot exceed one:

[0171]

[0172] Among them, represents that the VNF function of the slice is deployed on the node ; represents the total functions deployed on the node ; represents the total functions deployed on all nodes;

[0173] 6) The number of VNFs deployed for each slice is greater than the number of VNF function requirements:

[0174]

[0175] Among them, is the slice requirement for the number of VNF deployments; represents the number of VNFs deployed for each slice deployed.

[0176] 7) Whether physical isolation between slices is required:

[0177]

[0178] In the formula represents the slice requiring physical isolation between slices;

[0179] Slices compete with each other for limited physical resources, aiming to obtain more resources to increase their own benefits while avoiding unnecessary congestion costs. Therefore, the slice game problem is solved by finding the Nash equilibrium. The revenue function is a convex function, and both parts of the strategy are continuous in the action space. It can be obtained that there exists a Nash equilibrium for this problem. When the game reaches the final state, for any arbitrary belonging to and satisfying:

[0180]

[0181] Among them, represents the slice utility value or resource allocation parameter under the service quality (QoS) constraint; represents the updated slice utility value or resource allocation parameter under the service quality (QoS) constraint.

[0182] The above Nash equilibrium is achieved by using a fully distributed learning algorithm. Each slice is based on its initial mixed strategy which is composed of the probability vector and satisfies the non-negativity and normalization conditions. The slice randomly selects an action according to this strategy and participates in the game. During this process, each slice obtains the corresponding utility based on the action it selects, and this utility value is determined based on the slice's own information;

[0183] Subsequently, the slice will adjust its mixed strategy based on the action it selects and the utility it obtains, following the update rule:

[0184]

[0185] Among them, represents the mixed strategy at time t; is the input value; is the revenue obtained by each slice at time t;

[0186] The update rule is as follows: If an action can improve the utility of a slice, the probability of this action being selected in the future will increase; otherwise, it will decrease. If the strategy reaches the pure strategy state (that is, there is at least one action whose probability of being selected approaches 1 and the probabilities of other actions approach 0) or the iteration reaches the preset upper limit T, the algorithm stops and converges stably to the Nash equilibrium of the game.

[0187] Among them, the distributed learning algorithm based on slice resource game is shown in Table 1:

[0188] Table 1 Distributed learning algorithm based on slice resource game

[0189]

[0190] In this embodiment, specifically, the step S4 includes:

[0191] Step S41: Calculate the small time-scale trigger threshold through the fuzzy evaluation method;

[0192] Step S42: Under the condition of the small time-scale trigger threshold, optimize the resource allocation of the in-chip service flow through the in-chip optimization model, and adjust the link variables and node computing resources;

[0193] Specifically, the small time-scale trigger condition is , is the preset confidence limit. Statistically calculated by the fuzzy evaluation method for each shard's service:

[0194]

[0195] Among them, refers to the number of services of slice within a small time scale. refers to the QoS ratio vector of the in-chip service , which is used to describe the differences in the requirements of different services for delay, bandwidth, and packet loss rate. refers to the fuzzy evaluation vector of the in-chip service . Among them, The value ranges of are all

[0196] To optimize the data volume of the service quality compliance service flow, the in-chip optimization model expression is:

[0197]

[0198] Wherein, is an indicator function. When then otherwise . and are the delay and bandwidth of the th service flow respectively. represents the service request mapping relationship within the slice . is the traffic bandwidth resource requirement for the request , and is the traffic computing resource requirement for the request . The bandwidth resource consumed by the flow with a unit flow rate is , and the computing resource is . Since the dynamic change of a single flow in the in-chip adjustment process is relatively small, it is only necessary to reconfigure the flow link variable and the node computing resource allocation variable for a single flow to ensure that the in-chip reconfiguration overhead remains at a low level.

[0199] It can be seen from the above embodiments that the intelligent slice and scheduling method for unmanned cluster network resources adopted by the present invention improves the air cluster ad hoc network communication performance. Through this method, the resource re-slice overhead is effectively reduced, making the resource configuration more flexible to adapt to the changes of the ad hoc network.

[0200] The above embodiments only represent the specific implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the protection scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the technical solution of the present application, several deformations and improvements can still be made, and these all belong to the protection scope of the present application.

[0201] This background technology section is provided to generally present the context of the present invention. The work of the currently named inventors, to the extent described in this background technology section, and aspects that are not prior art at the time of filing this application are neither expressly nor impliedly admitted as prior art to the present invention.

Claims

1. An intelligent slicing and scheduling method for unmanned cluster network resources, characterized in that Including: Step S1: Dynamically optimize the reconfiguration threshold between slices by predicting the clustered movement trajectory and service requirements, and trigger cross-slice resource reconfiguration; Step S2: Allocate adaptive resources between slices and perform redundant reservation based on the prediction of clustered mobility and service requirements; Step S3: Construct a non-cooperative resource game model and design a distributed learning algorithm to achieve the Nash equilibrium configuration of resource allocation; Step S4: Dynamically update the small time-scale threshold based on service performance, and adjust the resources within the slice to optimize the quality of service; The said Step S1 includes: Step S11: Collect the movement trajectory data of all clusters and extract features; Step S12: Group the clusters by the clustering method based on the Dirichlet process to identify clusters with similar movement patterns; Step S13: Assign an initial clustering to each cluster, iteratively update the clustering assignment of each cluster, and use the Gibbs sampling method to generate samples from the multi-dimensional probability distribution; Step S14: Estimate the transition kernel of the cluster at the current location; Step S15: Predict the next access location of the cluster according to the transition kernel; Step S16: Trigger slice reconfiguration when the matching degree between the predicted access topology and slice resources exceeds the large time-scale trigger threshold; The said Step S15 includes: Step S151: Calculate the transfer probability of the cluster moving from the current access location i to the location j ; Step S152: Select the position corresponding to the maximum transition probability j as the next access position; The transition kernel is calculated by the following formula: Where: Indicates clustering Transition kernel at the current position; is the number of samples generated in Gibbs sampling; represents the th sample; is the total number of groups; is the number of other sub - clusters in cluster c except the division sub - cluster ; that is, the number of other sub - clusters in cluster c other than the division sub - cluster is the Dirac measure of the Dirichlet distribution; is the current DP mixture model concentration parameter; is the current base distribution.

2. The intelligent slicing and scheduling method for unmanned cluster network resources according to claim 1, wherein The movement trajectory data includes different locations accessed by the cluster and the timestamps of access, and the extracted features include the movement pattern and access frequency of the cluster.

3. The intelligent slicing and scheduling method for unmanned cluster network resources according to claim 1, wherein The transition probability is expressed by the following formula: Where: Indicate clustering From the current position i Transfer to the position j Transition probability; is the clustering at the current position in time t ; To determine the next access position of the clustering it is necessary to find the position that maximizes the transition probability j ; Indicates clustering From the current position i Transfer to position j The conditional probability of 4. An intelligent slicing and scheduling method for unmanned cluster network resources according to claim 1, characterized in that, The said Step S2 includes: Step S21: Construct a slice throughput model for evaluating transmission efficiency and calculate the expected throughput of the communication pair; Step S22: Estimate the average throughput of each slice network by the Monte Carlo method; Step S23: Estimate the expected average throughput of the slice and reserve redundant bandwidth and computing resources during resource allocation.

5. An intelligent slicing and scheduling method for unmanned cluster network resources according to claim 1, characterized in that, The said Step S3 includes: Step S31: Define the strategy selection space of each slice, and require that the computing resource and bandwidth allocation satisfy the constraints of the slice resource game; Step S32: Design a fully distributed learning algorithm to update the mixed strategy of the slice until it converges to the Nash equilibrium.

6. The intelligent slicing and scheduling method for unmanned cluster network resources according to claim 5, wherein The said constraints include: a) The total computing resources redeployed by all slices cannot exceed the total remaining resources of the current physical network; b) The total path delay within each slice is less than or equal to the delay constraint; c) The throughput of each slice is greater than or equal to the minimum throughput limit; d) The link bandwidth resources redeployed by all slices cannot exceed the total remaining bandwidth resources of the current physical network; e) The number of VNFs of the slice deployed on each node cannot exceed one; f) The number of VNFs deployed in each slice is greater than the number of VNF functional requirements; g) Whether the slice requires physical isolation between slices.

7. The intelligent slicing and scheduling method for unmanned cluster network resources according to claim 1, characterized in that The said Step S4 includes: Step S41: Calculate the small time-scale trigger threshold by the fuzzy evaluation method; Step S42: Under the condition of the small time-scale trigger threshold, optimize the resource allocation of the in-slice service flow through the in-slice optimization model, and adjust the link variables and node computing resources.

Images