A method and system for dynamic allocation of resources based on 5G slices

Abstract

The application discloses a kind of based on 5G slice resource dynamic allocation method and system, it is related to communication field, this method includes: acquisition multi-domain network operating parameter and service portrait information;Analysis is carried out to access network resource utilization, transmission link bandwidth, core network queue length and slice quota information, and the breach probability of slice service level agreement is output;Based on breach probability, the elastic interval of slice resource and cross slice borrowing matrix are calculated in digital twin model;Based on elastic interval and cross slice borrowing matrix, solve through security reinforcement learning algorithm, generate target resource allocation vector and action strategy;Target resource allocation vector and action strategy are issued to execution layer, control access network scheduling weight, transmission queue threshold and core network session anchor, realize slice resource dynamic allocation.The application effectively improves resource utilization, reduces cross-domain mismatch risk, and enhances service continuity and stability under complex business scenarios.

Description

Technical Field

[0001] This specification relates to the field of communications, and more specifically, to a method and system for dynamic resource allocation based on 5G slicing. Background Technology

[0002] With the gradual commercialization of 5G networks and the increasing diversification of business scenarios, the network resource requirements of enhanced mobile broadband (eMBB), massive machine-type communications (mMTC), and ultra-reliable low-latency communications (URLLC) differ significantly.

[0003] Existing slice resource allocation methods largely rely on static quotas or simple load balancing, failing to balance Service Level Agreement (SLA) compliance and overall resource utilization in situations involving sudden surges in traffic, cross-domain load fluctuations, and the coexistence of multiple services. On one hand, static quota methods can easily lead to breaches of critical slice services under high load scenarios, while resources may remain idle under low load scenarios. On the other hand, some machine learning-based prediction methods can only perform correlation analysis, lacking the ability to model the causal impact of resources, making it difficult to cope with dynamic changes under complex network conditions. Furthermore, existing resource scheduling algorithms lack end-to-end multi-domain collaborative control mechanisms, often optimizing only locally in the access network or core network, causing transmission links and core network anchor points to become bottlenecks, limiting overall performance improvement.

[0004] Therefore, there is an urgent need to propose a dynamic resource allocation method and system based on 5G slicing to at least solve some of the above problems. Summary of the Invention

[0005] The summary section introduces a series of simplified concepts, which will be further explained in detail in the detailed description section. The summary section of this invention is not intended to limit the key features and essential technical features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.

[0006] In a first aspect, the present invention proposes a method for dynamic resource allocation based on 5G slicing, the method comprising:

[0007] Collect multi-domain network operation parameters and service profile information, including access network resource utilization, transmission link bandwidth, core network queue length and slice quota information;

[0008] Based on the causal prediction model, the above access network resource utilization rate, above transmission link bandwidth, above core network queue length and above slice quota information are analyzed, and the probability of breach of slice service level agreement is output.

[0009] Based on the aforementioned default probability, the elasticity range of slice resources and the cross-slice borrowing matrix are calculated in the digital twin model;

[0010] In the near real-time policy controller, based on the above elastic interval and the above cross-slice borrowing matrix, a security reinforcement learning algorithm is used to solve the problem and generate the target resource allocation vector and action policy.

[0011] The aforementioned target resource allocation vector and action strategy are sent to the execution layer to control the access network scheduling weight, transmission queue threshold, and core network session anchor point, thereby achieving end-to-end dynamic allocation of slice resources.

[0012] In one feasible implementation, the above-mentioned analysis of the access network resource utilization, transmission link bandwidth, core network queue length, and slice quota information based on the causal prediction model, and the output of the probability of breach of slice service level agreement, includes:

[0013] Construct a causal relationship graph with the above-mentioned access network resource utilization rate, transmission link bandwidth, core network queue length and slice quota information as causal variables;

[0014] Intervention simulations were performed on the above slice quota information to obtain the distribution of service level agreement results under different resource allocations;

[0015] Based on the above result distribution, the probability that the above service level agreement does not meet the preset threshold is calculated, and this probability is used as the default probability of the above slice service level agreement. The construction of the above causal relationship graph, the above intervention simulation and the above probability calculation together constitute the implementation process of the above causal prediction model.

[0016] In one feasible implementation, a causal relationship graph is constructed using the aforementioned access network resource utilization rate, the aforementioned transmission link bandwidth, the aforementioned core network queue length, and the aforementioned slice quota information as causal variables, including:

[0017] The above access network resource utilization, transmission link bandwidth, core network queue length and slice quota information are normalized, and enhanced vectors are generated based on the service profile.

[0018] The causal structure learning algorithm is used to identify the directed causal relationships between the causal variables after the above normalization process, and a causal graph topology is formed by combining business prior constraints.

[0019] Based on structural equation modeling, the edges of the above causal graph topology are parametrically modeled to obtain edge weights that reflect the strength of the influence of each causal variable on the service level agreement, thereby completing the construction of the above causal relationship graph.

[0020] In one feasible implementation, the above-mentioned intervention simulation of the slice quota information to obtain the service level agreement result distribution under different resource allocations includes:

[0021] The above-mentioned slice quota information is set as multiple intervention values ​​to form a quota disturbance sequence covering the baseline, expansion and reduction.

[0022] Input the above quota perturbation sequence into the performance simulation module to obtain latency, packet loss rate and throughput samples under different intervention conditions;

[0023] Based on the aforementioned latency, packet loss rate, and throughput samples, the distribution of the service level agreement is fitted using a probability estimation module.

[0024] In one feasible implementation, the calculation of the elasticity range of slice resources and the cross-slice borrowing matrix in the digital twin model based on the aforementioned default probability includes:

[0025] In the digital twin model, the above-mentioned default probability is jointly input with historical network operation samples to obtain the performance boundary values ​​of each slice under different quotas. The above-mentioned digital twin model is a virtual network model for end-to-end performance simulation of access network, transmission network and core network.

[0026] Based on the above performance boundary values, the resource elasticity range for each slice is determined, wherein the above resource elasticity range includes the minimum guaranteed resource and the maximum allocatable resource;

[0027] Based on the correlation between the resource elasticity range and multiple slices, a cross-slice borrowing matrix is ​​generated, wherein the cross-slice borrowing matrix is ​​used to characterize the borrowable amount between slices under different load conditions.

[0028] In one feasible implementation, the near real-time policy controller, based on the aforementioned elastic interval and the aforementioned cross-slice borrowing matrix, generates a target resource allocation vector and action policy by solving a security reinforcement learning algorithm, including:

[0029] The aforementioned elastic interval is used as the feasible region constraint for resource allocation, and the aforementioned cross-slice borrowing matrix is ​​used as the inter-slice borrowing limit constraint.

[0030] Establish a Markov decision process with the above-mentioned resource utilization rate and the above-mentioned service level agreement achievement rate as reward functions;

[0031] During the training process, a security constraint function is introduced to give priority to slices that meet the above service level agreement latency threshold, thereby generating a target resource allocation vector and the above action strategy that meet the constraint conditions.

[0032] In one feasible implementation, the above method further includes:

[0033] In URLLC scenarios, select a security constraint function based on a barrier function to ensure that the latency does not exceed the hard threshold of the aforementioned service level agreement;

[0034] In the eMBB scenario, a security constraint function based on conditional risk value is selected to suppress extreme unfairness in bandwidth allocation;

[0035] In mMTC scenarios, an opportunity-based security constraint function is selected to ensure the overall reliability of large-scale access.

[0036] In hybrid scenarios, the aforementioned security constraint functions are dynamically switched based on the slice type and real-time load, thereby improving the adaptability and robustness of resource allocation.

[0037] In one feasible implementation, the aforementioned target resource allocation vector and action strategy are distributed to the execution layer to control the access network scheduling weight, transmission queue threshold, and core network session anchor point, thereby achieving end-to-end dynamic allocation of slice resources, including:

[0038] The aforementioned target resource allocation vector is mapped to the scheduling weight ratio of the access network, and the allocation of physical resource blocks is adjusted in real time on the access network side.

[0039] The above action strategy is converted into queue threshold control instructions for the transmission network, and the buffer allocation and discard strategy is dynamically adjusted in the above transmission nodes.

[0040] The above action strategy is synchronized to the session anchor selection module of the core network, and session migration is performed based on the load and path latency of the above session anchor, thereby completing the end-to-end dynamic allocation of slice resources.

[0041] In one feasible implementation, the specific steps taken by the session anchor selection module of the core network when performing the session migration include:

[0042] A comprehensive score is calculated based on the current load, path latency, and link packet loss rate of the aforementioned session anchor points.

[0043] When multiple candidate session anchors exist, a weighted evaluation is conducted using historical migration overhead and energy consumption metrics.

[0044] The optimal session anchor point is dynamically selected based on the comprehensive score mentioned above, and a buffer is allocated in advance during the session migration process to reduce handover latency, thereby improving the stability and continuity of end-to-end slice resource allocation.

[0045] Secondly, this invention proposes a dynamic resource allocation system based on 5G slicing, comprising:

[0046] The data acquisition unit is used to collect multi-domain network operation parameters and service profile information. The multi-domain network operation parameters include access network resource utilization, transmission link bandwidth, core network queue length, and slice quota information.

[0047] The analysis unit is used to analyze the above-mentioned access network resource utilization, transmission link bandwidth, core network queue length and slice quota information based on the causal prediction model, and output the probability of breach of slice service level agreement.

[0048] The calculation unit is used to calculate the elasticity range of slice resources and the cross-slice borrowing matrix in the digital twin model based on the above default probability;

[0049] The solution unit is used in the near real-time policy controller to solve the target resource allocation vector and action policy by using a security reinforcement learning algorithm based on the above elastic interval and the above cross-slice borrowing matrix.

[0050] The control unit is used to send the aforementioned target resource allocation vector and action strategy to the execution layer to control the access network scheduling weight, transmission queue threshold and core network session anchor point, thereby realizing end-to-end dynamic allocation of slice resources.

[0051] Thirdly, the present invention proposes an electronic device comprising: a memory and a processor, characterized in that the processor is configured to execute a computer program stored in the memory to implement the steps of the resource dynamic allocation method based on 5G slicing as described in any one of the first aspects.

[0052] Fourthly, the present invention proposes a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the 5G slicing-based dynamic resource allocation method as described in any one of the first aspects.

[0053] In summary, the 5G slicing-based dynamic resource allocation method proposed in this invention constructs a causal relationship graph through a causal prediction model and performs intervention simulations. This fundamentally characterizes the causal impact of slice quotas, access network utilization, transmission bandwidth, and core network queue length on SLA default probability, avoiding the bias of simple correlation prediction and improving the interpretability and reliability of the prediction results. A digital twin model is introduced, combining historical operating samples and default probabilities as inputs to accurately calculate the resource elasticity range and cross-slice borrowing matrix for each slice, enabling resource allocation to have adjustable boundaries and controllable mutual assistance mechanisms under different load scenarios. A security reinforcement learning algorithm is introduced into the near real-time policy controller, using resource utilization and SLA achievement rate as reward functions, and dynamically adjusting the policy search space in conjunction with security constraint functions to ensure that the latency of critical slices does not exceed the service threshold, achieving a balance between efficiency, fairness, and security. By distributing the target resource allocation vector and action policy to the execution layer, collaborative control is achieved in the access network, transmission network, and core network, realizing end-to-end dynamic scheduling. This effectively improves resource utilization, reduces cross-domain mismatch risk, and enhances service continuity and stability in complex business scenarios.

[0054] The present invention proposes a dynamic resource allocation method based on 5G slicing. Other advantages, objectives and features of the present invention will be partly apparent from the following description, and partly understood by those skilled in the art through research and practice of the present invention. Attached Figure Description

[0055] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0056] Figure 1 A flowchart illustrating a method for dynamic resource allocation based on 5G slicing, provided as an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of a resource dynamic allocation system based on 5G slicing provided in an embodiment of the present invention;

[0058] Figure 3 This is a structural schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0059] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The technical solutions of the embodiments of this invention will now be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them.

[0060] Please see Figure 1 This is a flowchart illustrating a method for dynamic resource allocation based on 5G slicing, provided by an embodiment of the present invention. Specifically, it may include:

[0061] S110. Collect multi-domain network operation parameters and service profile information, wherein the aforementioned multi-domain network operation parameters include access network resource utilization, transmission link bandwidth, core network queue length, and slice quota information.

[0062] S120. Based on the causal prediction model, analyze the above access network resource utilization rate, the above transmission link bandwidth, the above core network queue length and the above slice quota information, and output the probability of breach of slice service level agreement.

[0063] S130. Based on the above default probability, calculate the elasticity range of slice resources and the cross-slice borrowing matrix in the digital twin model;

[0064] S140. In the near real-time policy controller, based on the above elastic interval and the above cross-slice borrowing matrix, the target resource allocation vector and action policy are generated by solving the security reinforcement learning algorithm.

[0065] S150. The above target resource allocation vector and the above action strategy are sent to the execution layer to control the access network scheduling weight, transmission queue threshold and core network session anchor point, thereby realizing end-to-end dynamic allocation of slice resources.

[0066] For example, the system collects multi-domain network operation parameters in real time from three levels: access network, transmission network, and core network, and combines these parameters with service profile information to form an input dataset. Specifically, the multi-domain network operation parameters include access network resource utilization, transmission link bandwidth, core network queue length, and slice quota information. The service profile information includes user service type, traffic characteristics, and service sensitivity. This step ensures the completeness and timeliness of the data relied upon for subsequent modeling and decision-making.

[0067] The system analyzes the access network resource utilization, transmission link bandwidth, core network queue length, and slice quota information based on a causal prediction model. This model first constructs a causal relationship graph containing the aforementioned causal variables, and then simulates network operation under different quota adjustments through intervention operations, thereby obtaining the service level agreement (SLA) outcome distribution of slices under various resource allocation conditions. Based on this outcome distribution, the model further calculates the probability of SLA defaults for each slice under specific resource configurations, and uses this probability as input for subsequent resource allocation strategies.

[0068] The system inputs the aforementioned default probabilities into a digital twin model, which simultaneously considers historical operational data and simulation data to calculate the resource elasticity range and cross-slice borrowing matrix for each slice. The resource elasticity range is used to determine the minimum guaranteed resources and maximum allocable resources for each slice while meeting service demands. The cross-slice borrowing matrix describes the borrowable limits and constraints between different slices, thus providing boundary conditions and borrowing rules for dynamic resource adjustment.

[0069] The system inputs the aforementioned elasticity interval and cross-slice borrowing matrix into the near real-time policy controller and solves the problem using a security reinforcement learning algorithm. Specifically, the algorithm uses resource utilization and service level agreement (SLA) achievement rate as reward functions, the elasticity interval as a feasible region constraint, and the cross-slice borrowing matrix as a borrowing limit constraint. A security constraint function is introduced during training to ensure that the latency of critical slices does not exceed a preset threshold. Through this optimization process, the system ultimately generates a target resource allocation vector and action policy, ensuring fairness, stability, and service quality under dynamic network loads.

[0070] The system distributes the above-mentioned target resource allocation vector and the above-mentioned action policy to the execution layer, specifically including: mapping the target resource allocation vector to the scheduling weight ratio of the access network to adjust the allocation of physical resource blocks in real time; converting the action policy into a queue threshold control instruction for the transport network to dynamically adjust the buffer allocation and discard policy in the transport nodes; and synchronizing the action policy to the session anchor selection module of the core network to perform session migration according to the load and path delay of different anchors. Through the above multi-domain linkage control, the system realizes the dynamic allocation of end-to-end slice resources, which not only improves the resource utilization rate but also ensures the service level agreement achievement rate in diverse service scenarios.

[0071] In summary, the resource dynamic allocation method based on 5G slicing proposed by the present invention constructs a causal relationship graph through a causal prediction model and conducts intervention simulation, fundamentally characterizing the causal impact of slice quota, access network utilization rate, transmission bandwidth, and core network queue length on the SLA default probability, avoiding the deviation of pure correlation prediction, and enhancing the interpretability and reliability of the prediction results. By introducing a digital twin model and jointly inputting historical operation samples and default probabilities, the resource elasticity interval of each slice and the cross-slice borrowing matrix are accurately calculated, enabling the resource allocation to have adjustable boundaries and controllable mutual assistance mechanisms under different load scenarios. By introducing a safety reinforcement learning algorithm into the near-real-time policy controller, with the resource utilization rate and SLA achievement rate as the reward functions and combining with a safety constraint function to dynamically adjust the policy search space, ensuring that the delay of critical slices does not exceed the service threshold, the unity of efficiency, fairness, and safety is achieved. By distributing the target resource allocation vector and the action policy to the execution layer and performing collaborative control in the access network, transport network, and core network respectively, end-to-end dynamic scheduling is realized, effectively improving the resource utilization rate, reducing the cross-domain mismatch risk, and enhancing the service continuity and stability in complex service scenarios.

[0072] In a feasible implementation manner, analyzing the above-mentioned access network resource utilization rate, the above-mentioned transmission link bandwidth, the above-mentioned core network queue length, and the above-mentioned slice quota information based on the causal prediction model, and outputting the default probability of the slice service level agreement, including:

[0073] Constructing a causal relationship graph with the above-mentioned access network resource utilization rate, the above-mentioned transmission link bandwidth, the above-mentioned core network queue length, and the above-mentioned slice quota information as causal variables;

[0074] Performing intervention simulation on the above-mentioned slice quota information to obtain the service level agreement result distribution under different resource allocations;

[0075] Based on the above result distribution, the probability that the above service level agreement does not meet the preset threshold is calculated, and this probability is used as the default probability of the above slice service level agreement. The construction of the above causal relationship graph, the above intervention simulation and the above probability calculation together constitute the implementation process of the above causal prediction model.

[0076] For example, four types of parameters—access network resource utilization, transmission link bandwidth, core network queue length, and slice quota information—are used as causal variables, denoted as... ,in As an operable slice quota information node, the goal is to evaluate the SLA metrics when different quota values ​​are applied to it. The distribution changes and default probabilities. To eliminate the influence of dimensions, normalization and profile enhancement are performed to obtain the joint input:

[0077]

[0078] in, For the normalized first Network parameters, This is a semantically enhanced vector encoded from the business profile (business type, burstiness, latency / packet sensitivity, etc.). This refers to the joint features that enable entry into the causal learning and reasoning process.

[0079] In the causal structure learning stage, structural equation modeling (SEM) can be used to characterize the directed dependencies between variables, and the causal graph can be obtained by minimizing the following equation. Boundary rights :

[0080]

[0081] Where E represents taking the expected value of all expressions within the parentheses. express The set of parent nodes, and h The structural equations for intermediate variables and SLA outputs are respectively. The fitting loss (such as squared loss or log-likelihood). The sparsity regularization coefficient is used to control the sparsity of the edge set. Let be the parameter matrix for all directed edges.

[0082] After obtaining the structure, exogenous intervention is applied to the quota nodes during the intervention phase. And by "cutting off" its upstream incoming edges (graph cut), the intervened generation process is obtained:

[0083]

[0084] in, In addition to In addition to all factors affecting access network resource utilization, This indicates an impact on transmission link bandwidth. This indicates an impact on the core network queue length. Indicates except In addition to all random factors that directly affect the final SLA index, this formula represents the exogenous setting. conditions, The causal generation mechanism, This can be viewed as an external constraint (such as backbone bandwidth) or a semi-exogenous variable. To ignore the current network status, force the slice quota to be set to a specific value.

[0085] To obtain the distribution of SLA results under intervention, Monte Carlo sampling or digital twin-driven conditional simulation was used to obtain:

[0086]

[0087] Among them, the left-hand side is the intervention. The SLA probability distribution is given below; the right-hand side is derived from the structural equation and noise prior, and is numerically obtained through sampling / simulation.

[0088] The probability of default can be defined by comparing SLA metrics with the mean (e.g., in terms of achievement rate or time delay). Threshold and packet loss Boolean achievement variable of threshold):

[0089]

[0090] in, To achieve the SLA target, It is the first Next The generated SLA sample For the number of sampling or simulations, This is the indicator function. To provide the statistical confidence level, a binomial approximation confidence interval can be provided:

[0091]

[0092] in, For estimating the sample default rate, For standard normal quantiles, The significance level is indicated by .

[0093] If it is necessary to explicitly model the differences between different business profiles, the default probability can be decomposed conditionally and weighted by profile weights:

[0094]

[0095] in, For a set of image categories (such as URLLC / eMBB / mMTC, further subdivide the image). The conditional vibration term is determined by the proportion of images observed online or the strategy weights, and is derived from "in the image". Intervention simulations on subgroups were obtained.

[0096] In obtaining After obtaining the curve, it is possible to inversely deduce the risk tolerance of each slice within a given risk margin. The minimum guaranteed quota and the maximum allocable quota (used for subsequent flexibility) can be defined in their first order as follows:

[0097]

[0098] in, It is the upper limit for SLA breach tolerance set by the operations and maintenance side. and These serve as the feasible region boundaries for subsequent near-real-time optimization. To ensure model stability under network distribution drift, a sliding window adaptive re-evaluation of edge weights is employed, and relearning is triggered by the distribution divergence.

[0099]

[0100] in, The SLA distribution after intervention is estimated for adjacent time windows. This is the trigger threshold.

[0101] This embodiment uses structural equations and business priors to obtain an interpretable causal graph, and then... Within this framework, digital twins / simulations and sampling approximations are used to obtain... The resulting distribution ultimately yields the default probability through threshold comparison and statistical estimation. This default probability can be directly used as a risk metric input for the boundary calculation of the digital twin, or it can be linked with subsequent security reinforcement learning to constrain and weight the action space, ensuring that the strategy is both efficient and compliant in multi-load, mixed business scenarios.

[0102] In one feasible implementation, a causal relationship graph is constructed using the aforementioned access network resource utilization rate, the aforementioned transmission link bandwidth, the aforementioned core network queue length, and the aforementioned slice quota information as causal variables, including:

[0103] The above access network resource utilization, transmission link bandwidth, core network queue length and slice quota information are normalized, and enhanced vectors are generated based on the service profile.

[0104] The causal structure learning algorithm is used to identify the directed causal relationships between the causal variables after the above normalization process, and a causal graph topology is formed by combining business prior constraints.

[0105] Based on structural equation modeling, the edges of the above causal graph topology are parametrically modeled to obtain edge weights that reflect the strength of the influence of each causal variable on the service level agreement, thereby completing the construction of the above causal relationship graph.

[0106] For example, the access network resource utilization, transmission link bandwidth, core network queue length, and slice quota information are normalized to eliminate the impact of dimensional differences on structural identification. Enhanced vectors are generated based on service profiles, embedding semantic features such as user service type, traffic burstiness, and service sensitivity into the joint feature space of causal variables, enabling the causal relationship graph to reflect structural differences under different service scenarios. A causal structure learning algorithm is used to identify the directed causal relationships between the normalized causal variables, and a causal graph topology is formed by combining prior business constraints. Rules such as "the slice quota information has a unidirectional impact on the access network resource utilization," "the transmission link bandwidth has a positive constraint on end-to-end latency," and "the core network queue length is affected by both the access network resource utilization and the transmission link bandwidth" are used as prior constraints to limit the search space and directionality, ensuring that the obtained topology meets the physical rationality of network engineering.

[0107] After obtaining the aforementioned causal graph topology, the edges of the topology are parametrically modeled based on structural equation modeling. By minimizing the residuals between simulated samples and real-world replay samples on service level agreement (SLA) metrics, the edge weights reflecting the strength of each causal variable's influence on SLA are estimated. Sparse regularization and robust estimation are applied to the edge weights to suppress noise interference, and the edge weights corresponding to different business profiles are grouped and adapted to ensure the causal graph possesses stable interpretability and transferability under different loads and scenarios. Through these steps, the causal graph not only provides the directed dependency structure between variables but also quantifies the strength of each causal variable's influence on SLA with learnable edge weights, thus providing a structured and interpretable foundation for subsequent intervention simulations and default probability calculations.

[0108] In one feasible implementation, the above-mentioned intervention simulation of the slice quota information to obtain the service level agreement result distribution under different resource allocations includes:

[0109] The above-mentioned slice quota information is set as multiple intervention values ​​to form a quota disturbance sequence covering the baseline, expansion and reduction.

[0110] Input the above quota perturbation sequence into the performance simulation module to obtain latency, packet loss rate and throughput samples under different intervention conditions;

[0111] Based on the aforementioned latency, packet loss rate, and throughput samples, the distribution of the service level agreement is fitted using a probability estimation module.

[0112] For example, slice quota information is set as multiple intervention values, covering baseline, expansion, and reduction scenarios, thus forming a quota perturbation sequence. This perturbation sequence allows for a comprehensive observation of the impact of different quota adjustments on network performance under experimental conditions. Subsequently, the system inputs this quota perturbation sequence one by one into the performance simulation module, where an end-to-end simulation environment including the access network, transmission network, and core network is constructed. This environment is then run under different intervention conditions to generate corresponding latency, packet loss rate, and throughput samples.

[0113] This approach not only simulates performance under typical loads but also obtains performance sample sets under different random fluctuations through repeated simulations, better reflecting the network's true performance under uncertain conditions. Finally, the system inputs the obtained latency, packet loss rate, and throughput samples into the probability estimation module to statistically fit and model the distribution of performance indicators, thereby outputting the service level agreement (SLA) result distribution under different resource allocations. This distribution intuitively reflects the probability changes of SLA compliance or default under different quota conditions, providing solid data support for subsequent default probability calculations and resource elasticity range determination. This implementation method allows the impact of slice quota adjustments on SLAs to be reflected in a structured and quantitative form, improving the reliability and interpretability of the prediction results.

[0114] In one feasible implementation, the calculation of the elasticity range of slice resources and the cross-slice borrowing matrix in the digital twin model based on the aforementioned default probability includes:

[0115] In the digital twin model, the above-mentioned default probability is jointly input with historical network operation samples to obtain the performance boundary values ​​of each slice under different quotas. The above-mentioned digital twin model is a virtual network model for end-to-end performance simulation of access network, transmission network and core network.

[0116] Based on the above performance boundary values, the resource elasticity range for each slice is determined, wherein the above resource elasticity range includes the minimum guaranteed resource and the maximum allocatable resource;

[0117] Based on the correlation between the resource elasticity range and multiple slices, a cross-slice borrowing matrix is ​​generated, wherein the cross-slice borrowing matrix is ​​used to characterize the borrowable amount between slices under different load conditions.

[0118] For example, the above-mentioned slice quota information is set as multiple intervention values ​​covering three scenarios: baseline, expansion, and reduction, forming a quota perturbation sequence. The quota perturbation sequence includes equally spaced scan points and encrypted sampling points for high-load periods to ensure the ability to distinguish service level agreement boundary areas.

[0119] The above quota disturbance sequences are input one by one into the performance simulation module. The performance simulation module takes the end-to-end links of the access network, transmission network and core network as the object, derives the corresponding scheduling weight, queue threshold and session anchor configuration according to the given intervention value, and outputs end-to-end latency, packet loss rate and throughput samples under each intervention condition.

[0120] To enhance credibility, the performance simulation module generates sample clusters by replaying multiple random seeds and perturbing the service profile at each intervention value, thereby covering uncertainties such as wireless channel fluctuations, sudden service arrivals, and link jitter.

[0121] The system inputs the aforementioned latency samples, packet loss rate samples, and throughput samples into the probability estimation module for distribution fitting. The probability estimation module jointly models performance indicators under different intervention conditions based on a conditional mixture distribution and outputs the resulting distribution of the aforementioned service level agreement. During this process, the probability estimation module reweights anomalous long-tail samples to ensure that the distribution characterization remains robust and interpretable under extreme load or transient congestion conditions.

[0122] Through the above steps, the system not only obtains the distribution of service level agreement results under different resource allocation conditions, but also provides a statistical basis that can be directly invoked for subsequent default probability assessment and risk measurement, thereby providing quantitative support for subsequent resource elasticity range calculation and near real-time strategy optimization.

[0123] In one feasible implementation, the near real-time policy controller, based on the aforementioned elastic interval and the aforementioned cross-slice borrowing matrix, generates a target resource allocation vector and action policy by solving a security reinforcement learning algorithm, including:

[0124] The aforementioned elastic interval is used as the feasible region constraint for resource allocation, and the aforementioned cross-slice borrowing matrix is ​​used as the inter-slice borrowing limit constraint.

[0125] Establish a Markov decision process with the above-mentioned resource utilization rate and the above-mentioned service level agreement achievement rate as reward functions;

[0126] During the training process, a security constraint function is introduced to give priority to slices that meet the above service level agreement latency threshold, thereby generating a target resource allocation vector and the above action strategy that meet the constraint conditions.

[0127] In one feasible implementation, the above method further includes:

[0128] In URLLC scenarios, select a security constraint function based on a barrier function to ensure that the latency does not exceed the hard threshold of the aforementioned service level agreement;

[0129] In the eMBB scenario, a security constraint function based on conditional risk value is selected to suppress extreme unfairness in bandwidth allocation;

[0130] In mMTC scenarios, an opportunity-based security constraint function is selected to ensure the overall reliability of large-scale access.

[0131] In hybrid scenarios, the aforementioned security constraint functions are dynamically switched based on the slice type and real-time load, thereby improving the adaptability and robustness of resource allocation.

[0132] For example, end-to-end resource allocation can be abstracted as a constrained Markov decision process (CMDP) with security constraints. Let the time step be... System state, action, transition, and reward are defined as follows:

[0133]

[0134] For state space; For action space; For the transition probability; Rewards for joining a group; Discount factor; For constraint sets; For parameters random strategy, It represents the probability simplex.

[0135] For each slice in each cell / domain resource quota Apply a joint feasible region constraint to the elasticity interval and the borrowed matrix:

[0136]

[0137] For slices The minimum guaranteed and maximum allocatable resources (derived from the elastic range of digital twin computing). For a moment From slices Lend slices The limit; To borrow the upper limit of the far matrix across slices; This represents the total available resources for this cell / domain. For slices The basic quota.

[0138] To limit the risk of cumulative borrowing, aggregate constraints and dynamic repayment are introduced:

[0139]

[0140] This is the cumulative borrowing limit; To set a supply-side ratio ceiling and prevent excessive lending of individual components, Slicing under the current network conditions i The maximum amount of resources that can be allocated without causing serious breach of its own or other slice service level agreements.

[0141] Aggregate the multi-domain runtime states into:

[0142]

[0143] Parameter explanation: For the instantaneous PRB / power amplifier / dispatch utilization of each cell; Estimate the queue level and packet loss for each transmission node; For core network session anchor point load and path latency; Statistics were reached on historical borrowing balances and SLAs.

[0144] The joint action consists of three domain control vectors:

[0145]

[0146] Adjustments to the scheduling weights of each slice / cell (mapped to PRB weights). For transmission queue threshold / drop curve parameters; Select and migrate trigger instructions for session errors. Execute layer mapping functions. Will Configure changes that can be executed by the device in real time.

[0147] Use a combination of rewards:

[0148]

[0149] For the benefit of resource utilization; SLA achievement rate revenue (can be weighted by slice); For session migration overhead (state synchronization, route reconstruction); For borrowing costs (including repayment pressure); The cost of exceeding the boundary (such as prediction delays approaching the red line); The weights are adjustable.

[0150] 1) URLLC (Hard Delay): A delay-safe function based on the logarithmic barrier:

[0151]

[0152] A collection of URLLC slices; To predict end-to-end latency; This is the hard delay threshold; This is the obstacle coefficient. This term increases sharply when approaching the threshold, forcing the area away from the red line.

[0153] 2) eMBB (bandwidth fairness): CVaR constraint based on tail risk:

[0154]

[0155] Jain Fairness Index ( (eMBB slice throughput vector). For unfair losses; Confidence level; The upper limit for allowed tail unfairness; .

[0156] 3) mMTC (massive machine-type communications): Opportunity constraints:

[0157]

[0158] For the loss rate of cloud packaging; This is the acceptable upper limit; This represents the tolerance probability for default. In engineering practice, the probabilistic constraint can be transformed into a deterministic upper bound using the Chebyshev / Chernov or empirical quantile approximation.

[0159] 4) Dynamic switching / weighting of mixed scenes:

[0160]

[0161] These represent penalty / surrogate functions for URLLC barriers, eMBB tail risk, and mMTC opportunity constraints, respectively. The scene weights are adaptively updated based on real-time load and slice ratio. .

[0162] Rewrite the above safety items and feasible region as the Lagrange objective of CMDP:

[0163]

[0164] It includes all deterministic constraints (flexible range, borrowing limit, security constraint violation rate of proxying, etc.); For Lagrange multipliers; The step size is multiplied by the number of steps; [ ] It is a non-negative projection.

[0165] Employ a secure PPO-style policy update to control policy drift and incorporate security penalties:

[0166]

[0167] For advantage estimation (e.g., GAE); Here are the parameters for the previous round; KL represents the KL divergence between strategies. The upper bound of the stride limits the oscillation of the strategy update.

[0168] To tighten the action set in real time when a boundary violation is predicted, an action-level safety filter is introduced:

[0169]

[0170] The original policy output is used to project the actions onto a set that satisfies the feasible domain and security agent constraints through secondary projection. In engineering, it can be divided into three subspaces (RAN weight, queue threshold, and anchor migration) for dimensionality reduction and solution.

[0171] Autotightening is performed using the upper confidence bound of the delay / packet loss prediction error:

[0172]

[0173] Parameter explanation: For uncertainty estimation (such as Bayesian / deep integration variance); Let [the value be] the confidence coefficient. and Replace the upper bound with the entry. and To achieve conservative control.

[0174] Map policy outputs to device commands:

[0175] Weight table update Threshold issuance, MBbB anchor point migration. Mapping function. Responsible for unit conversion, sliding window interpolation, and debouncing control; receipts include PRB occupancy rate, queue level / ECN rate, and anchor load / delay, used for online updates of the value function, advantage estimation, and multipliers. It also triggers borrow rollback and threshold refinement when a deviation or risk increase is detected.

[0176] In one feasible implementation, the aforementioned target resource allocation vector and action strategy are distributed to the execution layer to control the access network scheduling weight, transmission queue threshold, and core network session anchor point, thereby achieving end-to-end dynamic allocation of slice resources, including:

[0177] The aforementioned target resource allocation vector is mapped to the scheduling weight ratio of the access network, and the allocation of physical resource blocks is adjusted in real time on the access network side.

[0178] The above action strategy is converted into queue threshold control instructions for the transmission network, and the buffer allocation and discard strategy is dynamically adjusted in the above transmission nodes.

[0179] The above action strategy is synchronized to the session anchor selection module of the core network, and session migration is performed based on the load and path latency of the above session anchor, thereby completing the end-to-end dynamic allocation of slice resources.

[0180] For example, based on the aforementioned target resource allocation vector, the access-side scheduling weight ratio is calculated for each cell and each slice. A one-to-one correspondence is established between the vector components and the physical resource block allocation. The physical resource block allocation is then adjusted in real time on the access network side through online updates of the scheduler weight table. To avoid sudden jitter, the update process employs a sliding window and progressive weight interpolation to ensure that weight changes between adjacent time slots are limited and do not introduce scheduling oscillations.

[0181] Based on the aforementioned action strategy, queue threshold control instructions are generated on the transmission side. The target queue occupancy rate, packet loss target, and congestion signal are mapped to minimum / maximum thresholds and drop curve parameters on each transmission node. Buffer allocation and drop strategies are dynamically adjusted within these transmission nodes. The drop strategy supports AQM / ECN linkage. When short-term link congestion is detected, explicit congestion marking is triggered first to suppress congestion propagation, rather than direct drop to protect critical slice flows.

[0182] The aforementioned action strategy is synchronized to the session anchor selection module of the core network to quickly assess the load and path latency of candidate session anchors. Under the premise of satisfying the service level agreement, a target anchor is selected and session migration is triggered. To reduce service interruption, the migration process adopts a "make-before-break" strategy. A mirror forwarding table is pre-established on the target anchor and its state is hot-synchronized. After the traffic switching conditions are met, a lossless switch is performed at the flow granularity, and the source anchor resources are reclaimed after the migration is complete.

[0183] After a coordinated deployment is completed, the execution layer returns receipts and real-time telemetry data by domain. The access side reports the updated PRB occupancy rate and scheduling achievement rate, the transmission side reports the queue level and ECN marking rate, and the core network side reports the load changes and end-to-end latency of the target session anchor point.

[0184] The near real-time policy controller compares the target resource allocation vector with the actual execution deviation. If a deviation or out-of-bounds risk is detected, it triggers fine-tuning or rollback based on the cross-slice borrowing matrix and refines and corrects the threshold and throttling parameters of the action policy.

[0185] Through this continuous process, the access network scheduling weight, transmission queue threshold, and core network session anchor point work together under a unified strategy, ensuring that the dynamic allocation of the aforementioned end-to-end slice resources remains stable, explainable, and auditable even under conditions of load bursts and link fluctuations.

[0186] In one feasible implementation, the specific steps taken by the session anchor selection module of the core network when performing the session migration include:

[0187] A comprehensive score is calculated based on the current load, path latency, and link packet loss rate of the aforementioned session anchor points.

[0188] When multiple candidate session anchors exist, a weighted evaluation is conducted using historical migration overhead and energy consumption metrics.

[0189] The optimal session anchor point is dynamically selected based on the comprehensive score mentioned above, and a buffer is allocated in advance during the session migration process to reduce handover latency, thereby improving the stability and continuity of end-to-end slice resource allocation.

[0190] For example, the core network's session anchor selection module first periodically samples the running state of candidate anchors to obtain three basic indicators for each candidate anchor: current load, path latency, and link packet loss rate. After normalizing these indicators, they are input into the comprehensive scorer.

[0191] The comprehensive scorer prioritizes low latency and high reliability, assigning higher weight to path latency as a sensitive factor. It also uses packet loss rate to reflect link stability and load to characterize processing capacity, thereby outputting a basic score that can truly represent end-to-end service quality.

[0192] When multiple candidate session anchors exist, the system introduces historical migration overhead and energy consumption metrics to weight and correct the basic score. On the one hand, historical migration overhead is used to penalize the state synchronization and session reconstruction costs caused by frequent switching, avoiding back-and-forth oscillations between anchors of similar quality. On the other hand, energy consumption metrics are used to prioritize the more energy-efficient anchor among multiple reachable paths, ensuring that session routing meets service level agreements while also considering operational energy efficiency goals. After obtaining the weighted comprehensive score, the system dynamically selects the current optimal session anchor and pre-allocates buffers and mirror forwarding tables on the target anchor before actually executing the session migration, achieving a lossless "build first, then switch" switchover. The critical states of the session to be migrated are hot-synchronized in the background, while an instantaneous buffer and congestion marking strategy are opened for the data stream to be switched to offset the short-term jitter at the moment of switching. When the path latency and packet loss rate of the target anchor are detected to have stabilized within a safe range, the system completes the direction switch at the flow granularity and reclaims the source anchor resources and temporary buffers after the switchover is completed. Through the above process, the selection and migration of session anchors not only reflects a sensitive response to real-time quality indicators, but also takes into account historical cost and energy efficiency constraints. Combined with buffering and state presetting before migration, the switching latency and packet loss peak are significantly reduced, thereby improving the stability and continuity of end-to-end slice resource dynamic allocation in complex load and multi-path environments.

[0193] like Figure 2 As shown, the present invention also provides a resource dynamic allocation system based on 5G slicing, comprising:

[0194] The acquisition unit 21 is used to acquire multi-domain network operation parameters and service profile information, wherein the aforementioned multi-domain network operation parameters include access network resource utilization, transmission link bandwidth, core network queue length, and slice quota information.

[0195] Analysis unit 22 is used to analyze the above access network resource utilization, the above transmission link bandwidth, the above core network queue length and the above slice quota information based on the causal prediction model, and output the probability of breach of slice service level agreement.

[0196] Calculation unit 23 is used to calculate the elasticity range of slice resources and the cross-slice borrowing matrix in the digital twin model based on the above default probability;

[0197] Solving unit 24 is used in the near real-time policy controller to solve the target resource allocation vector and action policy by using a security reinforcement learning algorithm based on the above elastic interval and the above cross-slice borrowing matrix.

[0198] The control unit 25 is used to send the above-mentioned target resource allocation vector and the above-mentioned action strategy to the execution layer to control the access network scheduling weight, transmission queue threshold and core network session anchor point, thereby realizing end-to-end dynamic allocation of slice resources.

[0199] The above system can also perform the following steps:

[0200] In one feasible implementation, the above-mentioned analysis of the access network resource utilization, transmission link bandwidth, core network queue length, and slice quota information based on the causal prediction model, and the output of the probability of breach of slice service level agreement, includes:

[0201] Construct a causal relationship graph with the above-mentioned access network resource utilization rate, transmission link bandwidth, core network queue length and slice quota information as causal variables;

[0202] Intervention simulations were performed on the above slice quota information to obtain the distribution of service level agreement results under different resource allocations;

[0203] Based on the above result distribution, the probability that the above service level agreement does not meet the preset threshold is calculated, and this probability is used as the default probability of the above slice service level agreement. The construction of the above causal relationship graph, the above intervention simulation and the above probability calculation together constitute the implementation process of the above causal prediction model.

[0204] In one feasible implementation, a causal relationship graph is constructed using the aforementioned access network resource utilization rate, the aforementioned transmission link bandwidth, the aforementioned core network queue length, and the aforementioned slice quota information as causal variables, including:

[0205] The above access network resource utilization, transmission link bandwidth, core network queue length and slice quota information are normalized, and enhanced vectors are generated based on the service profile.

[0206] The causal structure learning algorithm is used to identify the directed causal relationships between the causal variables after the above normalization process, and a causal graph topology is formed by combining business prior constraints.

[0207] Based on structural equation modeling, the edges of the above causal graph topology are parametrically modeled to obtain edge weights that reflect the strength of the influence of each causal variable on the service level agreement, thereby completing the construction of the above causal relationship graph.

[0208] In one feasible implementation, the above-mentioned intervention simulation of the slice quota information to obtain the service level agreement result distribution under different resource allocations includes:

[0209] The above-mentioned slice quota information is set as multiple intervention values ​​to form a quota disturbance sequence covering the baseline, expansion and reduction.

[0210] Input the above quota perturbation sequence into the performance simulation module to obtain latency, packet loss rate and throughput samples under different intervention conditions;

[0211] Based on the aforementioned latency, packet loss rate, and throughput samples, the distribution of the service level agreement is fitted using a probability estimation module.

[0212] In one feasible implementation, the calculation of the elasticity range of slice resources and the cross-slice borrowing matrix in the digital twin model based on the aforementioned default probability includes:

[0213] In the digital twin model, the above-mentioned default probability is jointly input with historical network operation samples to obtain the performance boundary values ​​of each slice under different quotas. The above-mentioned digital twin model is a virtual network model for end-to-end performance simulation of access network, transmission network and core network.

[0214] Based on the above performance boundary values, the resource elasticity range for each slice is determined, wherein the above resource elasticity range includes the minimum guaranteed resource and the maximum allocatable resource;

[0215] Based on the correlation between the resource elasticity range and multiple slices, a cross-slice borrowing matrix is ​​generated, wherein the cross-slice borrowing matrix is ​​used to characterize the borrowable amount between slices under different load conditions.

[0216] In one feasible implementation, the near real-time policy controller, based on the aforementioned elastic interval and the aforementioned cross-slice borrowing matrix, generates a target resource allocation vector and action policy by solving a security reinforcement learning algorithm, including:

[0217] The aforementioned elastic interval is used as the feasible region constraint for resource allocation, and the aforementioned cross-slice borrowing matrix is ​​used as the inter-slice borrowing limit constraint.

[0218] Establish a Markov decision process with the above-mentioned resource utilization rate and the above-mentioned service level agreement achievement rate as reward functions;

[0219] During the training process, a security constraint function is introduced to give priority to slices that meet the above service level agreement latency threshold, thereby generating a target resource allocation vector and the above action strategy that meet the constraint conditions.

[0220] In one feasible implementation, the above method further includes:

[0221] In URLLC scenarios, select a security constraint function based on a barrier function to ensure that the latency does not exceed the hard threshold of the aforementioned service level agreement;

[0222] In the eMBB scenario, a security constraint function based on conditional risk value is selected to suppress extreme unfairness in bandwidth allocation;

[0223] In mMTC scenarios, an opportunity-based security constraint function is selected to ensure the overall reliability of large-scale access.

[0224] In hybrid scenarios, the aforementioned security constraint functions are dynamically switched based on the slice type and real-time load, thereby improving the adaptability and robustness of resource allocation.

[0225] In one feasible implementation, the aforementioned target resource allocation vector and action strategy are distributed to the execution layer to control the access network scheduling weight, transmission queue threshold, and core network session anchor point, thereby achieving end-to-end dynamic allocation of slice resources, including:

[0226] The aforementioned target resource allocation vector is mapped to the scheduling weight ratio of the access network, and the allocation of physical resource blocks is adjusted in real time on the access network side.

[0227] The above action strategy is converted into queue threshold control instructions for the transmission network, and the buffer allocation and discard strategy is dynamically adjusted in the above transmission nodes.

[0228] The above action strategy is synchronized to the session anchor selection module of the core network, and session migration is performed based on the load and path latency of the above session anchor, thereby completing the end-to-end dynamic allocation of slice resources.

[0229] In one feasible implementation, the specific steps taken by the session anchor selection module of the core network when performing the session migration include:

[0230] A comprehensive score is calculated based on the current load, path latency, and link packet loss rate of the aforementioned session anchor points.

[0231] When multiple candidate session anchors exist, a weighted evaluation is conducted using historical migration overhead and energy consumption metrics.

[0232] The optimal session anchor point is dynamically selected based on the comprehensive score mentioned above, and a buffer is allocated in advance during the session migration process to reduce handover latency, thereby improving the stability and continuity of end-to-end slice resource allocation.

[0233] like Figure 3 As shown, the present invention also provides an electronic device 30, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor. When the processor 320 executes the computer program 311, it implements any step of the above-described resource dynamic allocation method based on 5G slicing.

[0234] The present invention also provides a computer program product comprising a computer program or computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer program or computer-executable instructions from the computer-readable storage medium and executes the computer program or computer-executable instructions, causing the electronic device to perform any step of the above-described 5G slicing-based dynamic resource allocation method of the present invention.

[0235] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1.A method for dynamic allocation of resources based on 5G slices, characterized in that, The method comprises the following steps: Collecting multi-domain network operation parameters and service portrait information, wherein the multi-domain network operation parameters include access network resource utilization, transmission link bandwidth, core network queue length and slice quota information; Analyzing the access network resource utilization, the transmission link bandwidth, the core network queue length and the slice quota information based on a causal prediction model, and outputting a default probability of a slice service level agreement; Calculating an elastic interval of slice resources and a cross-slice borrowing matrix in a digital twin model based on the default probability; In a near real-time policy controller, based on the elastic interval and the cross-slice borrowing matrix, a target resource allocation vector and an action policy are generated by solving a security reinforcement learning algorithm; The target resource allocation vector and the action policy are sent to the execution layer to control the access network scheduling weight, the transmission queue threshold and the core network session anchor, so as to realize the end-to-end dynamic allocation of slice resources. 2.The method of claim 1, wherein, The method comprises the following steps: A causal relationship graph is constructed with the access network resource utilization, the transmission link bandwidth, the core network queue length and the slice quota information as causal variables; An intervention simulation is performed on the slice quota information to obtain a service level agreement result distribution under different resource allocations; The probability that the service level agreement does not meet a preset threshold is calculated based on the result distribution, and this probability is taken as the default probability of the slice service level agreement, wherein the construction of the causal relationship graph, the intervention simulation and the probability calculation together constitute the implementation process of the causal prediction model. 3.The method of claim 2, wherein, A causal relationship graph is constructed with the access network resource utilization, the transmission link bandwidth, the core network queue length and the slice quota information as causal variables, comprising: The access network resource utilization, the transmission link bandwidth, the core network queue length and the slice quota information are normalized, and an enhanced vector is generated based on the service portrait; A causal structure learning algorithm is used to identify the directed causal relationship between each causal variable after normalization, and a causal graph topology is formed in combination with business prior constraints; Based on a structural equation model, the edges of the causal graph topology are parameterized modeled to obtain edge weights reflecting the influence intensity of each causal variable on the service level agreement, thereby completing the construction of the causal relationship graph. 4.The method of claim 2, wherein, The intervention simulation is performed on the slice quota information to obtain a service level agreement result distribution under different resource allocations, comprising: The slice quota information is set to a plurality of intervention values to form a quota disturbance sequence covering benchmarking, expansion and reduction; The quota disturbance sequence is input into a performance simulation module to obtain delay, packet loss rate and throughput samples under different intervention conditions; Based on the delay, the packet loss rate and the throughput samples, the result distribution of the service level agreement is fitted through a probability estimation module. 5.The method of claim 1, wherein, The method comprises the following steps: The default resource and the maximum allocable resource are determined based on the performance boundary value, wherein the resource elasticity interval comprises the minimum guaranteed resource and the maximum allocable resource; The cross-slice borrowing matrix is generated according to the correlation between the resource elasticity interval and the multi-slice, wherein the cross-slice borrowing matrix is used to represent the borrowing amount between slices under different load conditions. The method further comprises the following steps: 6.The method of claim 1, wherein, In the URLLC scenario, the safety constraint function based on the barrier function is selected to ensure that the delay does not exceed the hard threshold of the service level agreement; In the eMBB scenario, the safety constraint function based on the conditional value at risk is selected to suppress the extreme unfairness of bandwidth allocation; In the mMTC scenario, the safety constraint function based on the chance constraint is selected to ensure the overall reliability of large-scale access; In the mixed scenario, the safety constraint function is dynamically switched based on the slice type and real-time load, thereby improving the adaptability and robustness of resource allocation. 7.The method of claim 6, wherein, The target resource allocation vector and the action policy are sent to the execution layer to control the access network scheduling weight, the transmission queue threshold and the core network session anchor point, thereby realizing end-to-end dynamic allocation of slice resources, comprising: The target resource allocation vector is mapped to the scheduling weight proportion of the access network, and the physical resource block allocation is adjusted in real time on the access network side; The action policy is converted into a queue threshold control instruction of the transmission network, and the buffer allocation and discard strategy are dynamically adjusted in the transmission node; The action policy is synchronized to the session anchor point selection module of the core network, and the session migration is performed according to the load and path delay of the session anchor point, thereby completing the end-to-end dynamic allocation of slice resources. The specific steps of the session anchor point selection module of the core network when performing the session migration comprise: 8.The method of claim 1, wherein, The comprehensive score is calculated based on the current load, path delay and link packet loss rate of the session anchor point; When there are multiple candidate session anchor points, the historical migration overhead and energy consumption indicators are introduced for weighted evaluation. ​ ​ 9.The method of claim 8, wherein, ​ ​ ​ According to the comprehensive score, an optimal session anchor point is dynamically selected, and a buffer is allocated in advance in the session migration process to reduce switching delay, thereby improving stability and continuity of end-to-end slice resource allocation. 10.A system for dynamic allocation of resources based on 5G slices, the system comprising: Comprise: The acquisition unit is used for acquiring multi-domain network operation parameters and service portrait information, wherein the multi-domain network operation parameters include access network resource utilization, transmission link bandwidth, core network queue length and slice quota information; The analysis unit is used for analyzing the access network resource utilization, the transmission link bandwidth, the core network queue length and the slice quota information based on a causal prediction model, and outputting a violation probability of a slice service level agreement; The calculation unit is used for calculating an elastic interval of a slice resource and a cross-slice borrowing matrix in a digital twin model based on the violation probability; The solving unit is used for solving, in a near real-time policy controller, based on the elastic interval and the cross-slice borrowing matrix, through a safe reinforcement learning algorithm, to generate a target resource allocation vector and an action policy; The control unit is used for issuing the target resource allocation vector and the action policy to an execution layer to control access network scheduling weights, transmission queue thresholds and core network session anchor points, thereby realizing end-to-end dynamic allocation of slice resources.

Images