A network slice admission control method in a space-air-ground integrated network scenario

Abstract

The application discloses a network slice admission control method in a space-air-ground integrated network scene. The method is as follows: firstly, a slice admission control system model is established; then, a heterogeneous service priority model and a fairness model are designed; next, a slice admission control optimization problem of maximizing system service level and resource utilization is proposed; finally, the optimization problem is described as a Markov decision process and a deep reinforcement learning technology is used to solve the admission control decision. In the space-air-ground integrated network, the service target is heterogeneous and the resource is limited, the service priority and service discrimination problem are comprehensively considered, and the slice admission control decision is made to maximize the system service level and resource utilization, so that the system can guarantee the priority processing of the key business as much as possible when a large number of slice requests arrive, reduce the service discrimination problem caused by the priority level setting, and more efficiently utilize the space-air-ground network resources to serve the slice requests.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of space-air-ground integrated network, and particularly relates to a network slice admission control method in a space-air-ground integrated network scenario. BACKGROUND

[0002] When facing remote areas and harsh natural conditions, it is difficult for traditional ground networks to provide high-coverage, multi-scene and high-reliability services. In this context, the concept of space-air-ground integrated network is proposed and widely concerned, which deeply integrates space-based networks, air-based networks and ground-based networks, and uses satellite communication networks and air-based networks to provide valuable resources for expanding and supplementing ground networks. The diversity of service scenarios for space-air-ground integrated networks is increasing, requiring space-air-ground integrated networks to be customizable, flexible and support multi-service needs.

[0003] Network slicing is a resource allocation solution proposed in space-air-ground integrated networks to meet the increasing demand for diversified services, which transforms the traditional one-size-fits-all network into a set of logically independent networks, each customized to serve one type of service. Network slicing can provide customized network resource allocation for different service scenarios through end-to-end slicing on the shared physical infrastructure of space-air-ground integrated networks. Secondly, through network slicing, different business demand changes can be flexibly responded to, ensuring the continuity and efficiency of the business. At the same time, network slicing also separates the traffic of different services through logical function isolation, avoiding mutual interference between services.

[0004] Although network slicing as an enabling technology provides flexible customization capabilities for physical networks, in space-air-ground integrated networks, service demands are heterogeneous, and different users and application scenarios have different target requirements for the network. Moreover, various network slices with different needs share limited space-air-ground network resources such as satellite spectrum, and slice providers cannot always accept all network slice requests from tenants. How to make decisions through network slice admission control to comprehensively determine the acceptance or delayed rejection of slice requests has become a key problem that needs to be solved urgently. The solution to this problem is of great significance to promote the application of network slicing technology in space-air-ground integrated networks and other fields. SUMMARY

[0005] The purpose of the present application is to provide a network slice admission control method that can improve service level and resource utilization while considering service priority and fairness in complex heterogeneous space-air-ground integrated network scenarios.

[0006] The technical solution for achieving the purpose of the present application is: a network slice admission control method in a space-air-ground integrated network scenario, comprising the following steps:

[0007] Step 1, a slice admission control system model under the space-air-ground integrated network scene is established, and various resources required by the slice under the space-air-ground network are used to represent different slice types;

[0008] Step 2, a new priority representation model under the heterogeneous service scene is designed, which avoids the absolute correlation between priority definition and service, so as to more accurately ensure the communication demand of key business between slice requests pursuing heterogeneous service targets;

[0009] Step 3, a new fairness representation model under the heterogeneous service scene is designed, which comprehensively considers the guarantee of priority and fairness to reduce the degree of service discrimination;

[0010] Step 4, a slice admission control optimization problem for the space-air-ground integrated network is proposed;

[0011] Step 5, an admission control decision problem based on Markov decision process is proposed, and a deep reinforcement learning algorithm is designed to solve the problem, which maximizes the system service level and resource utilization rate on the basis of comprehensively considering priority and fairness.

[0012] Further, the slice admission control system model under the space-air-ground integrated network scene in step 1 uses various resources required by the slice under the space-air-ground network to represent different slice types, which is as follows:

[0013] It is assumed that the network slice tenant (ST) is expressed as a set K = {1, 2,..., k}, and the network slice provider (NSP) is expressed as another set N = {1, 2,..., n}. As a slice tenant, the service provider provides services for users by renting slices from the slice provider. After receiving the service request sent by the user, the slice tenant will maintain a service request queue, and then send a slice request to the slice provider. The tenant will weigh the default rate and request acceptance rate of each slice provider and then select the most suitable slice provider. When the user loses patience because the request sent to the tenant is not satisfied for a long time, the user will cancel the service request. As a slice provider, the infrastructure owner provides slices to its tenants, and they have different network infrastructures in the satellite layer, the air layer and the ground layer. After collecting the slice requests, the slice provider needs to consider the number of instances of each type of slice to be granted, and then allocate slice resources. The slice provider will publish its service default rate and cumulative request acceptance rate.

[0014] The present application considers three typical slices: high-throughput slice, low-latency slice and wide-coverage slice. The slice provider can support |S| kinds of slices, where S is the set of slice types. It is assumed that the slice tenant follows the parameter λ subPoisson probability initiates a slice request, and then selects a slice provider by probability P, which is determined by the service level indicator of the slice provider and the real-time queue length, which is defined as follows,

[0015]

[0016] where k n represents the real-time queue length, χ represents the service level, which is based on the default rate and request acceptance rate disclosed by the slice provider, and the calculation formula is as follows:

[0017] χ = w1·H n (t-1)-w2·φ n (t-1) (2)

[0018] where w1, w2 are two adjustable weight parameters, H n (t-1) and φ n (t-1) are the default rate and request acceptance rate disclosed by the NSP n at the last time, and the calculation formulas are given in equations 8 and 9 below. Time is divided into multiple slots t, which are fine enough to be used as basic units to measure the lifetime and waiting time of requests.

[0019] The slice provider establishes a queue Q s for each slice type, and will process these requests at the end of each time window. For each queue, use l s (t) to represent the real-time queue length. Use G = (g1(t), g2(t), g3(t)) T , A = (a1(t), a2(t), a3(t)) T , E = (e1(t), e2(t), e3(t)) T to represent the number of pre-authorized, accepted and exited requests in the queue of this slice provider at time t. For slice request i, use and to represent the slice lifetime, waiting time and patience time, respectively. At the end of each time window, some surviving requests will expire at this moment and release the resources they hold, while some requests will exit the queue because they have lost patience.

[0020] Further, the new priority representation model in step 2 described in the design of heterogeneous service scenario avoids the absolute correlation between priority definition and service, so as to more accurately guarantee the communication demand of key business between slice requests pursuing heterogeneous service goals, as follows:

[0021] The priority of a slice request is usually passed down by the priority of the service type that the slice corresponds to. To avoid the absolute relevance of priority indicator to service, the cumulative request acceptance rate (CRAR) and the weighted sum of the accumulated waiting time of the queue existing requests are used as priority measurement indicators.

[0022] For a queue maintained by a slice provider, the priority p s (t) is defined as follows:

[0023] p s (t) = a · η s (t) + β · W s (t) (3)

[0024] where a and β are two weight parameters that can be self-adjusted. η s (t) is the CRAR of the s-type slice at time t, and the calculation formula is as follows:

[0025]

[0026] where the numerator is the number of cumulative accepted requests, and the denominator is the sum of the cumulative accepted requests and the cumulative exited requests. W s (t) is the accumulated waiting time of all requests in the Q s queue at time t, and the calculation formula is as follows:

[0027] The priority indicates which type of request should be authorized first by the slice provider, and the priority needs to be recalculated and updated at each moment. An indicator function Θ(t) is introduced, and the priority vector P(t) = (p1(t), p2(t)...p s (t)) T As input, formula 6 gives the mathematical explanation of Θ(t). When Θ(t) = 0, it means that a priority order violation event has occurred, i.e., the CRAR level order of each service type does not match the priority order.

[0028]

[0029] where Δ|P(t)| is the vector that defines the service type index corresponding to each element in the vector P(t). And Δ|η(t)| defines the service type index corresponding to each element in the vector η(t).

[0030] Further, the new fairness representation model in the design of heterogeneous service scenarios described in step 3 comprehensively considers the priority and fairness guarantee to reduce the degree of service discrimination, as follows:

[0031] A part of fairness is embodied in the definition of priority, which takes into account the case of service discrimination. The accumulated waiting time W s (t) of a request in the queue

[0032] When the difference between the highest and lowest acceptance rate of requests of different services is higher than the overall acceptance rate, it is considered that unfair behavior has occurred. A indicator function Γ(t) is introduced, which uses η s (t) as input. The mathematical representation of Γ(t) is given in equation 7. When Γ(t) = 0, it represents the occurrence of an unfair event.

[0033]

[0034] Further, the step 4 proposes the slice admission control optimization problem for the space-air-ground integrated network, which is specifically as follows:

[0035] For an NSP, the total request acceptance rate H(t) is calculated by the following formula:

[0036]

[0037] For the above-mentioned violation rate φ n (t), it refers to the NSP after authorizing the request to join the waiting allocation of resources, but finally because of insufficient resources, it does not accept the request and allocate resources, which is defined as follows:

[0038]

[0039] Using o s (t) to represent the total amount of s-type slices that survive at the end of t, Z = (z1, z2, z3) T represents the total amount of spectrum, computing and storage resources owned by the slice provider, R s = (r1, r2, r3) T represents the amount of spectrum, computing and storage resources required by s-type slices. The resource utilization rate can be represented as follows:

[0040] The slice admission control optimization problem for the space-air-ground integrated network can be represented as:

[0041]

[0042] 0 ≤ g s (t) ≤ a s (t) ≤ l s (t) (13)

[0043]

[0044] r i <z i ,r i ∈R s (16)

[0045] Constraint (12) represents that the amount of resources used at any time cannot exceed the total amount of resources owned by the NSP, constraints 13, 14 represent that the number of requests authorized and accepted at a certain time should be within a reasonable range, constraint 15 represents that the number of slices granted at a certain time should be within a reasonable range, constraint 16 represents the resource constraint of the slice. Constraints 17 and 18 are introduced as soft constraints to emphasize the constraints of slice priority and fairness. Soft constraints are not necessary conditions, but conditions that should be met as much as possible if possible.

[0046] Further, step 5 describes the admission control decision problem based on Markov decision process, and an improved deep reinforcement learning algorithm is designed to solve the problem, which pursues the maximization of system service level and resource utilization on the basis of considering priority and fairness, as follows:

[0047] The decision related to admission control is regarded as a Markov model. Let K = (Sta, Act, Rew) represent the Markov decision process, where Sta represents the state space, Act represents the action space, and Rew represents the reward function. The environment of Markov decision is the slice admission control system in space-ground-integrated network, and the process is defined as follows:

[0048] · State: the state sta observed by the slice provider decision agent of the system t including, at the current time, the remaining available amount of each type of resource Z' = (z1(t), z2(t), z3(t)), the cumulative acceptance rate η s (t) of each type of service, the survival amount o s (t) of each type of slice at the current time, the cumulative waiting time W s (t) of each type of service queue, the queue length l s (t) and the proportion.

[0049] · Action: for each admission control decision window, the decision agent will take an n-element vector action, where the n elements respectively represent the acceptance percentage of each type of service request, action act t = (0.77, 0.8, 0.455), which means that for the requests in the three types of service queues, 77%, 80%, and 45.5% of the respective queue lengths are accepted.

[0050] • Reward: Whenever the decision agent chooses an action, a reward function evaluates the quality of the chosen action. The cumulative request acceptance rate, the default rate, and the resource utilization rate are used as the basic reward function, plus the consideration of the compliance of system decision to priority and fairness as punishment, the reward

[0051] The mathematical representation of the reward function is as follows:

[0052] rew t = H(t) - phi(t) + u(t) - pun1(t) - pun2(t) (19)

[0053] Where H(t), phi(t) and u(t) are the cumulative request acceptance rate, the default rate and the resource utilization rate, respectively, and pun1(t), pun2(t) are the punishment function, whose value is determined according to the value of the indicator function Theta(t), Gamma(t).

[0054]

[0055] Where p, is an adjustable parameter. The present application uses the double-delay deep deterministic policy gradient algorithm (TD3) in deep reinforcement learning to solve the optimal slice admission control decision in the integrated space-ground-terrestrial network.

[0056] Compared with the prior art, the present application has the following advantages: (1) Based on the priority and fairness recharacterization in the heterogeneous integrated space-ground-terrestrial network environment, the absolute correlation with services can be avoided, so that the priority processing of critical services can be guaranteed as much as possible, and the service discrimination problem caused by the priority level setting can be reduced. (2) The present application maximizes the service level and resource utilization rate to solve the network slice admission control problem under the condition of heterogeneous services and resource constraints, and proposes an admission control method based on deep reinforcement learning, which provides a technical foundation for the integration and development of network slices and integrated space-ground-terrestrial networks. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 The present application is a slice admission control system framework diagram under the integrated space-ground-terrestrial network scenario.

[0058] Figure 2 The present application is a network slice admission control method based on deep reinforcement learning. DETAILED DESCRIPTION

[0059] The present application will be further described in detail below with reference to the accompanying drawings.

[0060] The present application discloses a network slice admission control method under the integrated space-ground-terrestrial network scenario, comprising the following steps:

[0061] Step 1, establish a slice admission control system model under the space-air-ground integrated network scene, use various resources of the space-air-ground network required by the slice to represent different slice types, as follows:

[0062] In combination Figure 1 , set the network slice tenant (ST) as a set K = {1, 2, …, k}, and the network slice provider (NSP) as another set N = {1, 2, …, n}. As a slice tenant, the service provider provides services for users by renting slices from the slice provider. After receiving service requests sent by users, the slice tenant will maintain a service request queue, and then send a slice request to the slice provider. The tenant will weigh the default rate and request acceptance rate of each slice provider and then select the most suitable slice provider. When the user loses patience after the request sent to the tenant is not satisfied for a long time, the user will cancel the service request. As a slice provider, the infrastructure owner provides slices to its tenants, and they have different network infrastructures in the satellite layer, the air layer and the ground layer. After collecting slice requests, the slice provider needs to consider the number of instances of various slices to be granted, and then allocate slice resources. The slice provider will open and publish its service default rate and cumulative request acceptance rate.

[0063] The present application considers three typical slices: high throughput slice, low delay slice and wide coverage slice. The slice provider can support |S| kinds of slices, where S is the set of slice types. It is assumed that the slice tenant initiates a slice request following the Poisson probability with parameter λ sub , and then selects a slice provider through the probability P, which is determined by the service level indicator of the slice provider and the real-time queue length, and is defined as follows,

[0064]

[0065] Where k n represents the real-time queue length, χ represents the service level, which is based on the default rate and request acceptance rate disclosed by the slice provider, and the calculation formula is as follows:

[0066] χ = w1·H n (t-1)-w2·φ n (t-1) (23)

[0067] Where w1, w2 are two adjustable weight parameters, H n (t-1) and φ n (t-1) are the NSP nThe default rate and the acceptance rate at the last time instant are given in equations 8 and 9, respectively. Time is divided into slots t, which are fine enough to be used as the basic unit to measure the lifetime and the waiting time of a request.

[0068] The slice provider maintains a queue Q s for each slice type and processes the requests at the end of each time window. For each queue, we use l s (t) to denote the real-time queue length. We use G = (g1(t), g2(t), g3(t)) T , A = (a1(t), a2(t), a3(t)) T , E = (e1(t), e2(t), e3(t)) T to denote the number of pre-authorized, accepted and exited requests in the queue of the slice provider at time t, respectively. For a slice request i, we use and to denote the slice lifetime, the waiting time and the patience time, respectively. At the end of each time window, some alive requests will expire at this moment and release the resources they hold, while some requests will exit the queue because of losing patience.

[0069] Step 2, design a new priority representation model for heterogeneous service scenarios, avoid the absolute relevance of priority definition and service, in order to more accurately guarantee the communication demand of key business between slice requests pursuing heterogeneous service goals, as follows:

[0070] Usually, the priority of a slice request is passed by the priority of the service type corresponding to the slice. In order to avoid the absolute relevance of priority index and service, the cumulative request acceptance rate (CRAR) and the weighted sum of the cumulative waiting time of the existing requests in the queue are used as the priority measurement index.

[0071] For a queue maintained by a slice provider, the priority p s (t) of the s-type slice is defined as follows:

[0072] p s (t) = a · η s (t) + β · W s (t) (24)

[0073] Where a and β are two weight parameters that can be adjusted by themselves. η s (t) is the CRAR of the s-type slice at time t, and the calculation formula is as follows:

[0074]

[0075] Where, numerator is the cumulative accepted requests, denominator is the sum of cumulative accepted requests and cumulative exited requests. W s (t) is the priority of the request at time t s The cumulative waiting time of all requests in the queue, the calculation formula is as follows:

[0076] Priority indicates which type of request should be authorized first by the slice provider. The priority needs to be recalculated and updated at each time. A function Θ(t) is introduced, which uses the priority vector P(t) = (p1(t), p2(t)...p s (t)) T As input, formula 6 gives the mathematical explanation of Θ(t). When Θ(t) = 0, it means that the priority order violation event occurs, that is, the CRAR level order of each service type does not match the priority order.

[0077]

[0078] Where, Δ|P(t)| is the vector that defines the service type index corresponding to each element in the vector P(t). And Δ|η(t)| defines the service type index corresponding to each element in the vector η(t).

[0079] Step 3, design a new fairness representation model in heterogeneous service scenarios, comprehensively consider the priority and fairness protection, to reduce the degree of service discrimination, as follows:

[0080] In the above definition of priority, a part of fairness is reflected, considering the case of service discrimination. The cumulative waiting time of requests in the queue W s (t) is considered in the definition of priority to a certain extent, the weight parameter β can be adjusted according to the actual scene preference needs, so as to avoid the requests of low priority queue being ignored all the time.

[0081] When the difference between the highest value and the lowest value of the request acceptance rate of different services is higher than the overall request acceptance rate, it is considered that unfair behavior occurs at this time. A function Γ(t) is introduced, using η s (t) as input, formula 7 gives the mathematical representation of Γ(t). When Γ(t) = 0, it represents the occurrence of unfair events.

[0082]

[0083] Step 4, propose the slice admission control optimization problem for space-air-ground integrated network, as follows:

[0084] For an NSP, the total request acceptance rate H(t) is calculated as:

[0085]

[0086] For the above-mentioned default rate φ n (t), which refers to the NSP eventually not accepting the request and allocating resources due to insufficient resources after authorizing the request to join the waiting allocation of resources, is defined as follows:

[0087]

[0088] Use o s (t) to represent the total amount of s-type slices that survive at the end of time t, Z = (z1, z2, z3) T to represent the total amount of spectrum, computing and storage resources owned by the slice provider, R s = (r1, r2, r3) T to represent the amount of spectrum, computing and storage resources required by s-type slices. The resource utilization rate can be represented as follows:

[0089] The slice admission control optimization problem for space-air-ground integrated network can be represented as

[0090]

[0091] 0≤g s (t)≤a s (t)≤l s (t) (34)

[0092]

[0093] r i <z i , r i ∈R s (37)

[0094] Constraint (12) indicates that the amount of resources used at any time cannot exceed the total amount of resources owned by the NSP, constraints 13 and 14 indicate that the number of requests authorized and accepted at a certain time should be within a reasonable range, constraint 15 indicates that the number of slices granted to requests at a certain time should be within a reasonable range, and constraint 16 indicates the resource constraints of the slice. Constraints 17 and 18 are introduced as soft constraints to emphasize the constraints of slice priority and fairness. Soft constraints are not necessary conditions, but conditions that should be met as much as possible when possible.

[0095] Step 5, the admission control decision problem based on Markov decision process is proposed, and a deep reinforcement learning algorithm is designed to solve the problem. On the basis of considering priority and fairness, the system service level and resource utilization are maximized. Specifically as follows:

[0096] In combination Figure 2 , the admission control related decision is regarded as a Markov model. Let K = {Sta, Act, Rew) represent the Markov decision process, where Sta represents the state space, Act represents the action space, and Rew represents the reward function. The environment of Markov decision is the slice admission control system in space-ground-integrated network, and the process is defined as follows:

[0097] · State: the state sta observed by the slice provider decision agent of the system t Including, at the current time, the remaining available amount of each type of resource Z' = (z1(t), z2(t), z3(t)), the cumulative acceptance rate η s (t) of each type of service, the survival amount o s (t) of each type of slice at the current time, the cumulative waiting time W s (t) of each type of service queue, the queue length l s (t) and the proportion.

[0098] · Action: for each admission control decision window, the decision agent will take an n-element vector action, where the n elements represent the acceptance percentage of each type of service request, action act t = (0.77, 0.8, 0.455), which means that for the requests in the three types of service queues, 77%, 80%, and 45.5% of the respective queue lengths are accepted.

[0099] · Reward: each time the decision agent chooses an action, the reward function will evaluate the quality of the selected action. The cumulative request acceptance rate, the default rate and the resource utilization are used as the basic reward function, and the consideration of the system decision compliance to priority and fairness is added as punishment, reward

[0100] The mathematical expression of the reward function is as follows:

[0101] rew t = H(t) - φ(t) + u(t) - pun1(t) - pun2(t) (40)

[0102] Where H(t), φ(t) and u(t) are the cumulative request acceptance rate, the default rate and the resource utilization, respectively, and pun1(t), pun2(t) are the punishment functions, whose values are determined according to the values of the indicator functions Θ(t), Γ(t).

[0103]

[0104] where ρ, is a tunable parameter. The present application utilizes the Twin Delayed Deep Deterministic Policy Gradient algorithm (TD3) in deep reinforcement learning to solve the optimal slice admission control decision in space-air-ground integrated network.

[0105] The above describes the implementation process and advantages of the present application. Those skilled in the art should understand that various changes and improvements can be made to the present application without departing from the principles of the present application, and these changes and improvements all fall within the scope of the claimed present application.

Claims

1. A network slicing admission control method in an integrated air-space-ground network scenario, characterized in that, Comprising the following steps: Step 1, establish the slice admission control system model under the space-ground integration network scene, the model constructs a multi-level network slice supply architecture containing satellite layer, air layer and ground layer, supports dynamic rental of three typical slice types of high throughput, low delay and wide coverage, proposes an intelligent selection mechanism for slice providers based on service level indicators and real-time queue status, introduces request patience time and life cycle parameters, realizes a multi-class slice queue management mechanism on the slice provider side, and dynamically controls resource allocation and request scheduling through three state vectors of pre-authorization, acceptance and exit; Step 2, design a new priority representation model in heterogeneous service scenarios, abandon the traditional static service type priority, propose to take the weighted sum of cumulative request acceptance rate and queue cumulative waiting time as the dynamic priority index, through the definition of mathematical indicator function Θ(t), real-time monitor and judge whether the resource scheduling order is consistent with the calculated dynamic priority order; Θ(t) uses priority vector P(t) = (p1(t), p2(t)...p s (t)) T As input, when Θ(t) = 0, it indicates that a priority order violation event has occurred, that is, the CRAR level order of each service type does not coincide with the priority order. Wherein, Δ|P(t)| is a vector defining the service type index corresponding to each element in the vector P(t); and Δ|η(t)| defines the service type index corresponding to each element in the vector η(t); Step 3, design a new fairness representation model in a heterogeneous service scenario, integrate the cumulative waiting time of the queue into the original dynamic priority design, actively avoid long-term neglect of low-priority service requests by adjusting the weight parameter, and clearly define that when the difference between the highest and lowest request acceptance rates of different service types exceeds the overall acceptance rate of the system, a service discrimination event occurs; Step 4, propose a slice admission control optimization problem for space-ground integration networks, with the optimization objectives of maximizing the total request acceptance rate and optimizing resource utilization, while establishing the optimization problem under the multi-dimensional resource constraints composed of spectrum, computing and storage resources; Step 5, propose an admission control decision problem based on Markov decision process, dynamically determine the admission proportion of each type of slice request at each decision time, and clearly define the core elements of Markov decision process: state, action and reward; The state is the parameter that the slice provider can observe in the Markov decision process, including the remaining available amount of each type of resource, the cumulative acceptance rate of each type of service, the survival amount of each type of slice at the current time, the cumulative waiting time of each type of service queue, the length and proportion of each queue; The action is that for each admission control decision time t, the slice provider will calculate an n-element vector, where the n elements represent the acceptance percentage of each type of service request; The reward is that each time the slice provider selects an action, the quality of the selected action is evaluated, the cumulative request acceptance rate, the default rate and the resource utilization are used as the basic reward, and the compliance degree of priority and fairness is considered as the punishment, then the slice provider will optimize its action decision according to the reward, and solve the problem by improving the design of deep reinforcement learning algorithm TD3. 2.The method of claim 1, wherein, The slice admission control system model under the space-air-ground integrated network scenario is established as described in step 1. The model constructs a multi-level network slice supply architecture including satellite layer, air layer and ground layer, supports dynamic rental of three typical slice types, proposes an intelligent selection mechanism of slice provider based on service level indicators and real-time queue state, introduces request patience time and life cycle model, realizes multi-class slice queue management mechanism on the slice provider side, and dynamically controls resource allocation and request scheduling through three state vectors of pre-authorization, acceptance and exit. Specifically as follows: Set the network slice tenant (ST) as a set K = {1, 2,..., k}, and the network slice provider (NSP) as another set N = {1, 2,..., n}; as a slice tenant, the service provider provides services for users by renting slices from the slice provider; after receiving a service request sent by a user, the slice tenant will maintain a service request queue and then send a slice request to the slice provider; The tenant will weigh the default rate and request acceptance rate of each slice provider and then select the most suitable slice provider; when the user loses patience because the request sent to the tenant has not been met for a long time, the user will cancel the service request; as a slice provider, the infrastructure owner provides slices to its tenants, and they have different network infrastructures in the satellite layer, air layer and ground layer; After collecting slice requests, the slice provider needs to consider the number of instances of each type of slice to be granted and then allocate slice resources; The slice provider will publish its service default rate and cumulative request acceptance rate; The present application considers three typical slices: high throughput slice, low latency slice and wide coverage slice, and a slice provider can support |S| kinds of slices, wherein S is a set of slice types; it is assumed that a slice tenant initiates a slice request following a Poisson probability with a parameter λ sub , and then selects a slice provider through a probability P, which is determined by the service level index of the slice provider and the real-time waiting queue length, and is defined as follows, where k n The real-time queue length is represented by χ, and the service level is represented by χ, which is calculated based on the breach rate and request acceptance rate disclosed by the slice provider, and the formula is as follows: X = w1 · H n (t-1))-w2 · φ n (t-1) (3) where w1, w2 are two adjustable weight parameters, H n (t-1) and φ n (t-1) are NSP n The default rate and the request acceptance rate published at the last time, whose calculation formula is given in the formula 8, 9 of claim 5; the time is divided into multiple slots t, which are fine enough to be used as the basic unit for measuring the request lifetime and the waiting time; The slice provider builds a queue Q for each slice type s and processes the requests at the end of each time window; for each queue, we use l s (t) to denote the real-time queue length; we use G = (g1(t), g2(t), g3(t)) T , A = (a1(t), a2(t), a3(t)) T , E = (e1(t), e2(t), e3(t)) T to denote the number of pre-authorized, accepted and exited requests in the slice provider queue at time t; for a slice request i, we use the parameters and to denote the slice lifetime, the waiting time and the patience time; at the end of each time window, some alive requests will expire at this moment and release the resources they hold, i.e. are greater than the current time t, while some requests will exit the queue because they lose patience, i.e. . 3.The method of claim 1, wherein, Step 2 describes the design of a new priority representation model in a heterogeneous service scenario, which discards the traditional static service type priority and proposes to use the weighted sum of cumulative request acceptance rate (CRAR) and queue cumulative waiting time as a dynamic priority indicator. By defining a mathematical indicator function, it can monitor and determine whether the resource scheduling order is consistent with the calculated dynamic priority order in real time. Specifically as follows: The priority of a slice request is usually passed down from the service type priority corresponding to the slice; in order to avoid the absolute correlation between the priority indicator and the service, the weighted sum of cumulative request acceptance rate (CRAR) and queue cumulative waiting time of existing requests is used as a priority measurement indicator; For a slice provider-maintained queue, the priority p of the s-type slice s The definition of (t) is shown below: p s (t) = a · η s (t) + β · W s (t) (4) where a and β are two weight parameters that can be self-adjusted; η s (t) is the CRAR of s-type slice at time t, and the calculation formula is as follows: Wherein, the numerator is the cumulative accepted request quantity, and the denominator is the sum of the cumulative accepted request quantity and the cumulative exited request quantity; W s (t) is the Q at time t s The cumulative waiting time of all requests in the queue, and the calculation formula is as follows: The priority indicates which type of request should be authorized first by the slice provider, and the priority needs to be recalculated and updated at each moment; a mathematical indication function Θ(t) is introduced, using the priority vector P(t) = (p1(t), p2(t)...p s (t)) T As input, the formula 1 in the power requirement 1 gives a mathematical explanation of Θ(t); when Θ(t) = 0, it indicates that a priority order violation event has occurred, that is, the CRAR level order of each service type is inconsistent with the priority order. 4.The method of claim 1, wherein, Step 3 describes the design of a new fairness representation model in a heterogeneous service scenario, which integrates queue cumulative waiting time into the original dynamic priority design. By adjusting the weight parameter, it actively avoids long-term neglect of low-priority service requests, and clearly defines that when the difference between the highest and lowest request acceptance rates of different service types exceeds the overall acceptance rate of the system, a service discrimination event is determined to have occurred. Specifically as follows: The new priority definition of claim 3 has embodied a part of fairness, considering the case of service discrimination; the accumulated waiting time W of the request in the queue s (t) The definition of priority has taken into account to some extent, the weight parameter β can be adjusted according to the actual scene preference needs, so as to avoid the request of low priority queue being ignored all the time; When the difference between the highest and lowest values of request acceptance rate for different services is higher than the overall request acceptance rate, it is considered that unfair behavior has occurred; a indicator function Γ(t) is introduced, using η s (t) as input, and the mathematical representation of Γ(t) is given by equation 7; when Γ(t) = 0, it represents the occurrence of an unfair event.

5. The network slice admission control method in the space-air-ground integrated network scenario according to claim 1, characterized in that, Step 4 proposes a slice admission control optimization problem for space-air-ground integrated networks. The optimization objective is to maximize the total request acceptance rate and optimize resource utilization, while the optimization problem is established under the multi-dimensional resource constraints of spectrum, computing and storage resources. Specifically as follows: For an NSP, the total request acceptance rate H(t) is calculated as follows: For the default rate φ in claim 2 n (t) means that the NSP finally does not accept the request and allocate resources because of insufficient resources after the request to join the waiting for allocation of resources is authorized, and is defined as follows: Using o s (t) denotes the total amount of s-type slices that survive at the end of time t, Z = (z1, z2, z3) T denotes the total amount of spectrum, computing and storage resources owned by the slice provider, R s = (r1, r2, r3) T denotes the amount of spectrum, computing and storage resources needed by s-type slices; resource utilization can be expressed as follows: The slice admission control optimization problem for the space-air-ground integrated network can be expressed as: 0 < g s (t) < a s (t) < l s (t) (13) r i <2 i ，r i ∈R s (16) Constraint (12) represents that the amount of resources used at any time cannot exceed the total amount of resources owned by the NSP, constraints 13 and 14 represent that the number of authorized and accepted requests at a certain time should be within a reasonable range, constraint 15 represents that the number of slices granted at a certain time should be within a reasonable range, constraint 16 represents the resource constraint of the slice; constraints 17 and 18 are soft constraints introduced to emphasize the constraints of slice priority and fairness; the soft constraint is not a necessary condition, but a condition that should be met as much as possible when possible. 6.The method of claim 1, wherein, The admission control decision problem based on Markov decision process described in step 5 dynamically determines the admission ratio of each type of slice request at each decision time, and clearly defines the core elements of Markov decision process, including state, action and reward. A deep reinforcement learning algorithm TD3 is designed to solve the problem, and a priority experience replay mechanism and an adaptive action exploration noise are introduced, as follows: The decision related to admission control is regarded as a Markov model, and K=(Sta, Act, Rew) represents the Markov decision process, where Sta represents the state space, Act represents the action space, and Rew represents the reward function; The environment of Markov decision is the slice admission control system in the space-air-ground integrated network, the state is the parameter that the slice provider can observe in the Markov decision process, including the remaining available amount of each type of resource, the cumulative acceptance rate of each type of service, the survival amount of each type of slice at the current time, the cumulative waiting time of each type of service queue, and the queue length and proportion; The action is that for each admission control decision time t, the slice provider will calculate an n-element vector, where the n elements represent the acceptance percentage of each type of service request; The reward is that each time the slice provider selects an action, the quality of the selected action is evaluated, and the cumulative request acceptance rate, the default rate and the resource utilization rate are used as the basic reward, and the compliance degree of priority and fairness is considered as the punishment, then the slice provider will optimize its action decision according to the reward, the process elements are defined as follows: • state: the state sta observable by the slice provider of the system t including, at the current time, the remaining available amount of each type of resource Z' = (z1(t), z2(t), z3(t)), the cumulative acceptance rate of each type of service η s (t), the survival amount of each type of slice o s (t), the cumulative waiting time of each type of service queue W s (t), the length of each queue l s (t) and the proportion; • Action: For each admission control decision window, the slice provider takes an n-ary vector action, where n elements represent the acceptance percentage for each class of service request, action act t = (0.77, 0.8, 0.455), representing that for requests in the 3 classes of service queues, 77%, 80%, and 45.5% of the respective queue length are accepted, respectively; · Reward: Each time the slice provider selects an action, the reward function evaluates the quality of the selected action, and the cumulative request acceptance rate, the default rate and the resource utilization rate are used as the basic reward function, and the compliance of the system decision to priority and fairness is considered as the punishment, and the mathematical representation of the reward function is as follows: rew t = H(t) - φ(t) + u(t) - punl(t) - pun2(t) (19) Where H(t), φ(t) and u(t) are the cumulative request acceptance rate, the default rate and the resource utilization rate, pun1(t) and pun2(t) are the penalty functions, whose values are determined according to the values of the indicator functions Θ(t) and Γ(t), wherein p, is a parameter that can be adjusted; The application utilizes a double-delay deep deterministic policy gradient algorithm (TD3) in deep reinforcement learning to solve optimal slice admission control decisions in a space-air-ground integrated network; the TD3 algorithm observes the state of the system in real time through its own policy network, including resource reserves, queue state and historical performance indicators, and then outputs an action vector that decides the admission proportion of various slice requests; subsequently, a reward signal is fed back according to the action execution, which comprehensively considers the service acceptance rate, the default rate, the resource utilization rate and the degree of violation of priority and fairness; the TD3 algorithm continuously iteratively updates its decision strategy through the collaborative optimization of its own policy network and value network according to the reward signal, and finally learns an admission control strategy that can achieve high service level and resource utilization rate in the space-air-ground integrated network environment.

Images