5G network end-to-end service quality dynamic optimization method

Abstract

The invention provides a 5G network end-to-end service quality dynamic optimization method, and belongs to the technical field of 5G network optimization, and the method comprises the steps: collecting multi-dimensional data in real time through a sensing module in a 5G network; taking the multi-dimensional data as training data, constructing a dynamic optimization model based on a near-end strategy optimization algorithm of meta-learning enhancement, and performing multi-task learning in a 5G network scene including normal load, high load and burst traffic; dynamically adjusting strategy network parameters of a near-end strategy optimization algorithm by utilizing a meta-learning mechanism, wherein the strategy network parameters comprise internal loop task specific adaptation and external loop meta-strategy updating; closed-loop integration of an algorithm and a 5G network architecture is realized through a cooperative controller, SLA compliance is verified regularly, and model reconfiguration is triggered during violation. The method has the advantages that the overall performance and reliability of network service are improved through an intelligent algorithm enhanced by meta-learning; real-time accurate perception and rapid dynamic response to the network state are realized, and the stability of the service quality under different network conditions is guaranteed.

Description

Technical Field

[0001] This invention belongs to the field of 5G network communication technology, specifically relating to a dynamic optimization method for end-to-end service quality of 5G networks. Background Technology

[0002] Currently, end-to-end quality of service (QoS) assurance in 5G networks primarily relies on the static QoS parameter configuration mechanism and network slicing technology defined by the 3GPP standard. This provides differentiated services for different service types through predefined 5QI levels and GBR parameters. Traditional methods employ rule-based empirical optimization or traditional control theories (such as PID control) for network resource management, which can meet basic service requirements in scenarios with relatively stable network conditions. Existing technologies mainly achieve QoS rule binding through the core network SMF component, execute flow control policies through the UPF, and utilize NWDAF for network data analysis, forming a complete QoS assurance system.

[0003] However, existing technologies have significant limitations: static QoS configuration cannot adapt to dynamic changes in network load, resulting in low resource allocation efficiency in high-load scenarios; traditional optimization methods have slow convergence speed, and multi-objective collaborative rapid optimization achieves poor results. Summary of the Invention

[0004] In view of the above, this invention addresses the shortcomings of existing technologies by providing a dynamic optimization method for end-to-end service quality in 5G networks. To solve the aforementioned technical problems, the technical solution adopted by this invention includes: Multidimensional data is collected in real time through sensing modules in the 5G network; Using multidimensional data as training data, a dynamic optimization model is constructed based on a meta-learning-enhanced proximal policy optimization algorithm. Multi-task learning is performed in 5G network scenarios including normal load, high load, and burst traffic. In the meta-learning stage, multi-task training is performed under preset network scenarios to obtain a meta-policy model that is universal across scenarios. In the online execution stage, policy parameters are quickly adjusted to adapt to new network states using a small number of real-time samples. The policy network parameters of the near-end policy optimization algorithm are dynamically adjusted using a meta-learning mechanism, including inner loop task-specific adaptation and outer loop meta-policy update, to optimize QoS parameters and network slice resource allocation. The algorithm is integrated into the 5G network architecture in a closed loop through a collaborative controller, which periodically verifies SLA compliance and triggers model reconfiguration when violations occur.

[0005] Furthermore, the multidimensional data includes: wireless access network load. The unit is a percentage, representing the utilization rate of wireless resource blocks; core network traffic. The unit is Gbps, representing the core network data transmission rate; the 5QI identifier for the service type. 5QI is the 5G QoS identifier, an integer used to distinguish service priorities; real-time latency. The unit is milliseconds, which represents the end-to-end transmission delay of data packets.

[0006] Furthermore, methods for constructing dynamic optimization models based on meta-learning-enhanced proximal policy optimization algorithms include: The meta-learning phase uses multidimensional data as training data, and the training objective is the loss function of the proximal policy optimization algorithm. It is defined as the weighted sum of pruning strategy loss, value loss, and entropy regularization: The strategy loss is based on the strategy ratio. ,in; The policy ratio is defined as the probability of the new policy. Probability of the old strategy The ratio of θ to θ measures the degree of policy change; θ represents the policy network parameters, indicating the weights of the neural network. Let be the expectation operator, representing the expected value at time step t; The advantage function measures the advantage of an action relative to the average level. For the clipping function, Limited to the range Internally, to prevent excessively large updates; To limit the update range for the trimming parameters; The state value function represents the state value at a given time. The expected cumulative reward; The target value function is typically obtained through time difference learning. Let entropy function represent the strategy. The randomness is used to encourage exploration; These are weighting coefficients used to balance different loss terms.

[0007] Furthermore, methods for dynamically adjusting the policy network parameters of the proximal policy optimization algorithm using meta-learning mechanisms include: The meta-learning phase involves multi-task training under various preset 5G network scenarios, sampling task sets. ,in Indicates the first Each task corresponds to a network state distribution; Adapting computational task-specific parameters through an inner loop. ,in; α represents the meta-policy parameters, indicating the shared policy network weights; α is the inner loop learning rate. For gradient operators, it means that with respect to parameters The partial derivatives; For the task The loss function of the near-end policy optimization algorithm on; Advantage function The generalized dominance estimate is calculated using the generalized dominance estimation method. ;in γ is the time difference error, representing the difference between the current reward and the value estimate; γ is the discount factor, which measures the importance of future rewards; λ is the GAE parameter, which controls the bias-variance tradeoff of advantage estimation. For the task The loss function of the meta-learning-enhanced proximal policy optimization algorithm; Update meta parameters via outer loop ,in The outer loop learning rate; ← represents the assignment operator, indicating that the updated value will be assigned to θ; meta-parameters. The update uses second-order gradient computation to capture shared structure across tasks, and trains the output meta-policy parameters. This enables meta-strategies to gain cross-scenario generalization capabilities.

[0008] Furthermore, methods for constructing dynamic optimization models based on meta-learning-enhanced proximal policy optimization algorithms also include: During the execution phase, status data is collected through the 5G network sensing module. Based on meta-policy parameters Input status To the policy network, output actions. ;in; Adjust the step size for 5QI level, with an integer range of -10 to +10; To ensure the step size for bit rate adjustment, the adjustment range is -20% to +20%; during the online adaptation phase, parameters are rapidly adjusted using a small number of real-time samples. action The rules are converted into QoS rules through the 5G core network SMF component and then sent to the UPF via the N4 interface for flow control. At the same time, they are integrated with the network slicing architecture to allocate resources for different slice types. Continuously monitor SLA metrics; the reward function adopts a multi-objective weighted form. Used to evaluate the effect of actions and trigger model reconfiguration; where; The reward value for time step t; This is the current delay, in milliseconds (ms). The maximum allowable delay for the SLA, in milliseconds (ms). Current throughput, in Mbps; Target throughput, in Mbps; For indicator functions, when The value is 1; α, β, and γ are the weighting coefficients for latency, throughput, and overload penalty, respectively.

[0009] Furthermore, the state data The metrics are collected through the NWDAF function, and the actions... The output is sent through the SMF component, and SMF will... Mapped to DSCP priority; and the flow control policy of UPF updated via N4 interface; Maintain independent meta-policy libraries for slice types such as eMBB and uRLLC, and trigger meta-learning to adapt quickly when slice resource contention is detected; Reward Function Integration with SLA: Reward Function Weights Dynamically adjust to match 5G SLA requirements and achieve closed-loop feedback through a collaborative controller.

[0010] Furthermore, methods for achieving closed-loop integration of algorithms and 5G network architecture through a collaborative controller, periodically verifying SLA compliance, and triggering model reconfiguration upon violation include: The collaborative controller is deployed on the 5G core network side and interacts with the SMF and NWDAF components to obtain network status data and service SLA indicators in real time. SLA compliance verification: Key performance indicators, including end-to-end latency, throughput, and packet loss rate, are periodically calculated within a set range and compared with SLA thresholds; if any indicator exceeds the threshold for three consecutive periods, it is considered an SLA violation. If an SLA violation is detected, the collaborative controller initiates the meta-learning rapid adaptation module to generate reconfiguration instructions based on the current network state. The reconfiguration actions include: The policy network parameters of the near-end policy optimization algorithm are dynamically adjusted, and task-specific parameters are updated through an inner loop. Optimize QoS rules, such as adjusting 5QI level or GBR value, and send them to UPF via SMF; If the SLA metrics do not improve after reconfiguration, the policy rollback mechanism will be triggered to restore the stable version policy. Closed-loop feedback: The reconfiguration effect is monitored by NWDAF and fed back to the collaborative controller to optimize subsequent decision-making logic.

[0011] Furthermore, the collaborative controller dynamically adjusts the meta-learning adaptation frequency based on network load; when the load rate... At that time, the adaptation frequency is increased to once per second; when the load rate At this time, the adaptation frequency is reduced to once every 10 seconds to balance computational overhead and response speed; Design differentiated SLA verification strategies for different service types, including eMBB and uRLLC; focus on monitoring latency metrics for uRLLC services and throughput metrics for eMBB services.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Improve the overall performance and reliability of network services through multi-dimensional data perception, meta-learning enhanced intelligent algorithms and closed-loop control mechanisms; achieve real-time and accurate perception of network status and rapid dynamic response, adapt to diverse scenario changes such as normal load, high load and sudden traffic, and ensure the stability of service quality under different network conditions.

[0013] 2. By deeply integrating meta-learning with proximal policy optimization algorithms, this application breaks through the limitations of traditional optimization methods in cross-scenario adaptation; the inner loop task-specific adaptation and the outer loop meta-policy update mechanism enable the model to extract shared knowledge from multi-task training, and quickly adjust policy parameters with a small number of samples when facing new business scenarios; improve optimization efficiency, and reduce the model's dependence on a large amount of historical data.

[0014] 3. Through SLA compliance verification and adaptive reconfiguration mechanism, a complete service quality assurance system has been built; by periodically monitoring key performance indicators and combining them with multi-objective reward function design, the system can proactively identify potential service quality risks, trigger policy adjustments or rollback operations in a timely manner, and improve the automation level and fault response capabilities of network operation and maintenance. Attached Figure Description

[0015] The present invention will now be described in further detail with reference to the accompanying drawings.

[0016] Figure 1 : A schematic diagram of the hierarchical closed-loop control process in this invention; Figure 2 : A schematic diagram of the sequential execution process in this invention. Detailed Implementation

[0017] To better understand the present invention, the content of the invention is further clearly illustrated below with reference to embodiments and accompanying drawings. However, the scope of protection of the present invention is not limited to the embodiments described below. Numerous specific details are set forth in the following description to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the present invention can be practiced without one or more of these details.

[0018] Example 1: See Figure 1-2 This embodiment of a 5G network end-to-end service quality dynamic optimization method includes: S1. Real-time collection of multi-dimensional data through sensing modules in the 5G network; real-time collection of multi-dimensional data enables the algorithm to accurately perceive changes in network status, avoiding the lag of traditional static configuration, and improving the accuracy and response speed of optimization.

[0019] S2. Using multidimensional data as training data, a dynamic optimization model is constructed based on the meta-learning-enhanced Proximal Policy Optimization (PPO) algorithm. Multi-task learning is performed in 5G network scenarios including normal load, high load, and burst traffic. The training objective is to minimize the PPO loss function. Pruning policy loss, value loss, and entropy regularization are combined to ensure the stability of policy updates. In the meta-learning phase, multi-task training is performed under preset network scenarios to obtain a cross-scenario general meta-policy model. In the online execution phase, policy parameters are quickly adjusted to adapt to new network states using a small number of real-time samples. The combination of meta-learning and the proximal policy optimization algorithm overcomes the shortcomings of traditional PPO algorithms, such as slow convergence and poor adaptability, enabling rapid adaptation in the dynamic environment of 5G networks, reducing reliance on large amounts of historical data, and improving model efficiency. Through multi-task learning, the model can maintain service quality in diverse network scenarios.

[0020] S3. The meta-learning mechanism is used to dynamically adjust the policy network parameters of the near-end policy optimization algorithm, including inner loop task-specific adaptation and outer loop meta-policy update, to optimize QoS parameters and network slice resource allocation. The meta-learning mechanism ensures the accuracy and efficiency of the dynamic adjustment of policy parameters. The inner loop adaptation realizes task-specific optimization, and the outer loop update improves cross-scenario consistency, optimizes resource utilization, and realizes the foundation for resource competition management and deployment between slices.

[0021] S4. By using a collaborative controller to achieve closed-loop integration of the algorithm and the 5G network architecture, the compliance of the Service Level Agreement (SLA) is verified regularly, and model reconfiguration is triggered when violations occur. Through regular verification and dynamic reconfiguration, continuous optimization of service quality is ensured, the need for manual intervention is reduced, and the system's adaptability and reliability are improved.

[0022] Step S11: The multidimensional data are status indicators collected from the 5G network, including: radio access network load. The unit is a percentage, representing the utilization rate of wireless resource blocks; core network traffic. The unit is Gbps, representing the core network data transmission rate; the 5QI identifier for the service type. 5QI is the 5G QoS identifier, an integer used to distinguish service priorities; it measures the end-to-end real-time latency of data packets from transmission to reception. The unit is milliseconds, which are collected through network probes or NWDAF function and are core verification parameters for SLA compliance.

[0023] Step S21: The method for constructing a dynamic optimization model based on the meta-learning-enhanced proximal policy optimization algorithm includes: In the meta-learning phase, multidimensional data is used as training data, and the training objective is the loss function of the proximal policy optimization algorithm. It is defined as a weighted sum of pruning policy loss, value loss, and entropy regularization, balancing policy update stability, value estimation accuracy, and exploration incentives. Through this weighted sum, the algorithm can simultaneously optimize policy performance (such as QoS parameter adjustment) and value estimation (such as network state prediction), avoiding overfitting to a single objective and improving the model's generalization ability in multiple 5G network scenarios. ; To prune the policy loss term, the policy update step size is limited to prevent training divergence. The pruning mechanism ensures the stability of QoS parameter adjustments (such as preventing sudden changes in 5QI level). As a value loss term, minimizing the state value estimation error improves the accuracy of network state prediction (such as delay estimation) and enhances prediction accuracy. Entropy regularization term increases policy randomness and avoids premature convergence; it promotes diversity in slice resource allocation, enhances the algorithm's adaptability to sudden traffic, and enables the model to quickly converge to the optimal policy through balancing policy improvement, value estimation, and exploration.

[0024] Among them, the strategy loss is based on the strategy ratio. ,in; The policy ratio is defined as the probability of the new policy. Probability of the old strategy The ratio measures the degree of policy change; a value close to 1 indicates small policy change, while a value far from 1 indicates large change; this is achieved through monitoring... The algorithm can dynamically control the step size of QoS parameter adjustment, avoiding the impact of drastic fluctuations on network stability; θ represents the policy network parameters, indicating the weights of the neural network; Let be the expectation operator, representing the expected value at time step t; The advantage function measures the advantage of an action relative to the average level, helping the algorithm identify the optimal action. For example, in high-load scenarios, it prioritizes the allocation of resources to high-priority services to improve the efficiency of slice resource utilization. For the clipping function, Limited to the range Internally, to prevent excessively large updates; To trim parameters, limit update magnitude, ensure policy updates are within a stable range, prevent 5G network service interruptions due to parameter mutations, and enhance system reliability; The state value function represents the state value at a given time. The expected cumulative reward; The target value function is typically obtained through time difference learning. Let entropy function represent the strategy. The randomness of the model is used to encourage exploration, which helps to discover new optimization strategies and improve model adaptability. These are weighting coefficients used to balance different loss terms; Step S31: The method for dynamically adjusting the policy network parameters of the proximal policy optimization algorithm using a meta-learning mechanism includes: The meta-learning phase involves multi-task training under various preset 5G network scenarios. Through multi-task training, the meta-learning model can learn shared knowledge across scenarios, improve the model's generalization ability, and avoid overfitting to a single scenario. (Sampling task sets are also included.) ,in Indicates the first There are 10 tasks, and each task corresponds to a network state distribution.

[0025] Adapting computational task-specific parameters through an inner loop. θ represents the task-specific parameters obtained by rapidly adjusting the meta-parameter θ based on the gradient of the loss function for task Ti. This process is based on PPO loss function optimization, and the inner loop adaptation enables the model to quickly customize policy parameters for each task. For example, in high-load tasks, resource allocation parameters are optimized first to improve response speed. Among these; α is the meta-policy parameter, representing the shared policy network weights; α is the inner loop learning rate, used to control the adjustment step size. For gradient operators, it means that with respect to parameters The partial derivatives; For the task The loss function of the near-end policy optimization algorithm.

[0026] Advantage function The GAE formula is calculated using the Generalized Advantage Estimation (GAE). ;in γ is the time difference error, representing the difference between the current reward and the value estimate; γ is the discount factor, which measures the importance of future rewards; λ is the GAE parameter, which controls the bias-variance tradeoff of advantage estimation. For the task The loss function of the meta-learning-enhanced proximal policy optimization algorithm is provided; GAE provides a more stable advantage estimate, helping the algorithm to accurately identify the optimal action (such as slice resource allocation decision) in multiple scenarios of 5G network and reduce policy oscillation.

[0027] Update meta parameters via outer loop ,in The outer loop learning rate; ← represents the assignment operator, meaning to assign the updated value to θ; meta-parameters. The update uses second-order gradient calculation, which involves taking the second derivative of the loss function (e.g., the Hessian matrix) to capture commonalities in shared structures across tasks, such as network load patterns, and trains the output meta-policy parameters. This enables meta-strategies to gain cross-scenario generalization capabilities.

[0028] During the outer loop meta-policy update, the Hessian matrix approximation is constructed by calculating the second derivative of the loss function with respect to the meta-parameter θ. The Hessian matrix approximation H(θ) is obtained through the second-order gradient operator. Calculate the curvature information of the loss function with respect to θ; update direction: The direction of the Hessian matrix is ​​updated by inverse matrix adjustment, and the diagonal approximation method is used to reduce computational complexity by calculating only the diagonal elements of the Hessian matrix. ; By combining second-order gradient calculation with real-time performance monitoring of 5G networks, the calculation frequency is dynamically adjusted: when the network load change rate exceeds the threshold (e.g., load rate change > 10% / second), full second-order gradient calculation is enabled; otherwise, a first-order approximation method is used. The dynamic adjustment mechanism optimizes the allocation of computing resources, ensures optimization quality during high load changes, and reduces overhead in a stable state. in; For the second-order gradient operator, it represents the mathematical operator for taking the second-order partial derivative with respect to the parameter θ; This indicates all tasks The loss function (e.g., under normal load, high load, and other 5G network scenarios) is summed. For the task The loss function for near-end strategy optimization is optimized with the goal of minimizing this loss to improve service quality. The outer loop learning rate controls the convergence speed of the meta-policy and trains the output meta-policy parameters. This enables meta-strategies to gain cross-scenario generalization capabilities; It is the identity matrix, used to maintain the stability of parameter updates and avoid matrix singularity problems; To adapt the task through the inner loop Specific parameters represent the results of fine-tuning the meta-parameter θ for a specific task.

[0029] Step S22: The method for constructing a dynamic optimization model based on the proximal policy optimization algorithm with meta-learning reinforcement also includes: During the execution phase, status data is collected through the 5G network sensing module. ;in; Wireless access network load factor, expressed as a percentage. Core network traffic, in Gbps; The current 5QI value, Real-time latency, in milliseconds (ms). Real-time throughput, in Mbps; Multi-dimensional data acquisition enables the algorithm to accurately perceive dynamic changes in the network (such as load fluctuations or service type switching), providing reliable input for subsequent optimization decisions and avoiding the lag of traditional static configuration; and Help identify network congestion. Supports business priority differentiation. and Directly related to SLA compliance; Based on meta-policy parameters Input status To the policy network, output actions. ;in; Adjust the step size for 5QI level, with an integer range of -10 to +10; To ensure the bit rate adjustment step size, the adjustment range is -20% to +20%; during the online adaptation phase, parameters are rapidly adjusted using a small number of real-time samples; through the meta-policy parameter θ, the algorithm can generalize to new network scenarios, reducing dependence on large amounts of training data and improving real-time performance. Action outputs directly optimize QoS parameters, for example, by increasing [something] when high latency is detected. To prioritize business operations or adjust strategies during traffic surges. Ensure bandwidth availability.

[0030] action The 5G core network SMF component converts the rules into QoS rules, which are then sent to the UPF (User Plane Function) via the N4 interface for flow control. The collaboration between SMF and UPF enables rapid deployment of policies, ensuring that optimization decisions take effect in a timely manner. At the same time, it integrates with the network slicing architecture to allocate resources for different slice types such as eMBB and uRLLC. Continuously monitor SLA metrics; the reward function adopts a multi-objective weighted form. Used to evaluate the effect of actions and trigger model reconfiguration; the multi-objective reward function balances latency, throughput, and load metrics to ensure that the optimization direction is consistent with SLA requirements; among which; The reward value for time step t; This is the current delay, in milliseconds (ms). The maximum allowable delay for the SLA, in milliseconds (ms). Current throughput, in Mbps; Target throughput, in Mbps; For indicator functions, when The time factor is set to 1; α, β, and γ are the weighting coefficients for latency, throughput, and overload penalty, respectively. The optimization focus is dynamically adjusted by adjusting the weighting coefficients.

[0031] Step S23: The state data The metrics are collected through the NWDAF (Network Data Analytics Function), a key network data analysis function of the 5G core network, which enables intelligent operation and maintenance and service optimization by collecting and analyzing network data in real time. The action The output is sent through the SMF component, and SMF will... Mapped to DSCP priority (e.g., 5QI=82 mapped to DSCP 46), DSCP (Differentiated Services Code Point) is a QoS classification standard; and the UPF flow control policy is updated through the N4 interface; Maintain independent meta-policy libraries for slice types such as eMBB and uRLLC, and trigger meta-learning to adapt quickly when slice resource contention is detected; Reward Function Integration with SLA: Reward Function Weights Dynamically adjust to match 5G SLA requirements and achieve closed-loop feedback through a collaborative controller.

[0032] Step S41: The methods for achieving closed-loop integration of the algorithm and the 5G network architecture through a collaborative controller, and for periodically verifying SLA compliance and triggering model reconfiguration include: The collaborative controller is deployed on the 5G core network side, interacts with SMF and NWDAF components, and obtains network status data and service SLA indicators in real time. Deploying on the core network side enables the controller to directly access key network components, ensuring low latency in data acquisition and command issuance. SLA compliance verification: Key performance indicators (KPIs), including end-to-end latency, throughput, and packet loss rate, are periodically calculated within a set range and compared with SLA thresholds. If any KPI exceeds the threshold for three consecutive periods, it is considered an SLA violation. Periodic verification provides continuous monitoring capabilities to ensure service quality consistently meets requirements. If latency exceeds the threshold for three consecutive periods (e.g., ...), the violation will be detected. The system identifies violations and triggers reconfiguration. Periodic detection results are more reliable than single detection, reducing false alarms and helping to maintain system stability.

[0033] If an SLA violation is detected, the collaborative controller activates the meta-learning rapid adaptation module. Utilizing an inner-loop adaptation and outer-loop update mechanism, it generates reconfiguration instructions based on the current network state to quickly adjust the strategy. The reconfiguration actions include: The policy network parameters of the near-end policy optimization algorithm are dynamically adjusted, and task-specific parameters are updated through an inner loop. Optimize QoS rules, such as adjusting 5QI level or GBR value, and send them to UPF via SMF; If the SLA metrics do not improve after reconfiguration, the policy rollback mechanism is triggered to restore the stable version of the policy. When new parameters cause performance degradation, rollback ensures rapid service recovery, reduces operational risks, prevents network degradation caused by optimization failures, and ensures system reliability. Closed-loop feedback: The reconfiguration effect is monitored by NWDAF and fed back to the collaborative controller to optimize subsequent decision-making logic. Closed-loop feedback enables continuous learning and optimization, allowing the algorithm to improve from historical experience.

[0034] Step S42: The collaborative controller dynamically adjusts the meta-learning adaptation frequency based on network load; when the load rate... At that time, the adaptation frequency is increased to once per second; when the load rate At this time, the adaptation frequency is reduced to once every 10 seconds to balance computational overhead and response speed; Design differentiated SLA verification strategies for different service types, including eMBB and uRLLC; focus on monitoring latency metrics for uRLLC services and throughput metrics for eMBB services.

[0035] Technical effects of this embodiment: 1. Through multi-dimensional data perception, meta-learning-enhanced intelligent algorithms, and closed-loop control mechanisms, the overall performance and reliability of network services are significantly improved. This application enables real-time and accurate perception of network status and rapid dynamic response, effectively adapting to diverse scenarios such as normal load, high load, and sudden traffic surges, ensuring the stability of service quality under different network conditions.

[0036] 2. By deeply integrating meta-learning with proximal policy optimization algorithms, this application breaks through the limitations of traditional optimization methods in cross-scenario adaptation; the inner loop task-specific adaptation and the outer loop meta-policy update mechanism enable the model to extract shared knowledge from multi-task training, and quickly adjust policy parameters with a small number of samples when facing new business scenarios; improve optimization efficiency, and reduce the model's dependence on a large amount of historical data.

[0037] 3. By combining service quality optimization with network slicing architecture, service assurance for differentiated business needs can be achieved. Through closed-loop integration of the collaborative controller, the system can intelligently coordinate the resources of the wireless access network and the core network, dynamically adjust the 5QI level and guarantee the bit rate according to the service priority, thereby optimizing the resource allocation efficiency between slices.

[0038] 4. Through SLA compliance verification and adaptive reconfiguration mechanism, a complete service quality assurance system has been built; by periodically monitoring key performance indicators and combining them with multi-objective reward function design, the system can proactively identify potential service quality risks, trigger policy adjustments or rollback operations in a timely manner, and improve the automation level and fault response capabilities of network operation and maintenance.

[0039] 5. By dynamically adjusting the meta-learning adaptation frequency based on network load, the system achieves reasonable control of computational overhead while ensuring optimization effect, and has an intelligent computing resource scheduling mechanism; it can meet the real-time requirements and operating costs of 5G networks, and is also feasible for large-scale deployment.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Any other modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention, as long as they do not depart from the spirit and scope of the technical solutions of the present invention, should be covered within the scope of the claims of the present invention.

Claims

1. A 5G network end-to-end quality of service dynamic optimization method, characterized in that, The method comprises the following steps: Real-time collection of multi-dimensional data by a perception module in a 5G network; Construction of a dynamic optimization model based on a meta-learning enhanced proximal policy optimization algorithm using multi-dimensional data as training data for multi-task learning in 5G network scenarios including normal load, high load and burst traffic; In the meta-learning stage, multi-task training is performed under preset network scenarios to obtain a meta-policy model applicable to different scenarios; in the online execution stage, the policy parameters are quickly adjusted to adapt to new network states through a small number of real-time samples; Dynamic adjustment of the policy network parameters of the proximal policy optimization algorithm using a meta-learning mechanism, including inner loop task-specific adaptation and outer loop meta-policy update, optimization of QoS parameters and network slice resource allocation; Closed-loop integration of the algorithm and the 5G network architecture through a collaborative controller, periodic verification of SLA compliance, and triggering of model reconfiguration when a violation occurs.

2. The 5G network end-to-end quality of service dynamic optimization method of claim 1, wherein, The multi-dimensional data comprises: radio access network load , expressed by percentage, indicating radio resource block utilization; core network traffic , expressed by Gbps, indicating core network data transmission rate; 5QI identifier of service type , 5QI is 5G QoS identifier, integer, used to distinguish service priority; real-time delay , expressed by millisecond, indicating end-to-end transmission delay of data packet.

3. The 5G network end-to-end quality of service dynamic optimization method of claim 2, wherein, The method for constructing a dynamic optimization model based on a meta-learning enhanced proximal policy optimization algorithm comprises: The meta-learning stage takes multi-dimensional data as training data, and the training target is a loss function of a proximal policy optimization algorithm defined as a weighted sum of a clipped policy loss, a value loss, and an entropy regularization: where the policy loss is based on the policy ratio where; is the policy ratio, defined as the ratio of the new policy probability to the old policy probability , which measures the degree of policy change; θ is the policy network parameter, representing the weights of the neural network; is the expected operator, representing the expected value at time step t; is the advantage function, measuring the advantage of an action relative to the average level; is the clipping function, limiting to the interval to prevent the update magnitude from being too large; is the clipping parameter, limiting the update magnitude; is the state value function, representing the expected cumulative reward in state ; is the target value function, usually obtained through temporal difference learning; is the entropy function, representing the randomness of the policy , used to encourage exploration; is the weight coefficient, used to balance different loss terms.

4. The 5G network end-to-end quality of service dynamic optimization method of claim 3, wherein, The method for dynamically adjusting the policy network parameters of the proximal policy optimization algorithm using a meta-learning mechanism comprises: The meta-learning phase involves multi-task training under various preset 5G network scenarios, sampling task sets. ,in Indicates the first Each task corresponds to a network state distribution; Adapting task-specific parameters through inner loop where; is the meta-strategy parameter representing shared policy network weights; a is the inner loop learning rate; is the gradient operator representing the partial derivative with respect to the parameter ; is the proximal policy optimization algorithm loss function on the task ; Advantage function The advantage function is calculated by generalized advantage estimation (GAE) as where is the time-difference error, representing the difference between the current reward and the value estimate; γ is the discount factor, measuring the importance of future rewards; λ is the GAE parameter, controlling the bias-variance trade-off of the advantage estimate; is the loss function of the meta-reinforcement learning enhanced proximal policy optimization algorithm on the task ; Updating meta-parameters by outer loop where is the outer loop learning rate; <- denotes the assignment operator, which means assigning the updated value to The update of makes the meta-policy to obtain the cross-scene generalization ability.

5. The 5G network end-to-end quality of service dynamic optimization method of claim 3, wherein, The method for constructing a dynamic optimization model based on a meta-learning enhanced proximal policy optimization algorithm further comprises: The execution phase collects state data through the 5G network perception module ; based on the meta-strategy parameters , input state to the policy network, output action ; wherein; is the 5QI level adjustment step, integer range -10 to +10; is the guaranteed bit rate adjustment step, adjustment range -20% to +20%; the online adaptation phase quickly adjusts the parameters through a small number of real-time samples; Actions Convert to QoS rules by 5G core network SMF component, issue to UPF through N4 interface to implement flow control, and integrate with network slice architecture to allocate resources for different slice types; The SLA indicators are continuously monitored, and the reward function adopts a multi-objective weighted form: for evaluating the effect of actions and triggering model reconfiguration; is the reward value for time step t; is the current latency, in ms; is the maximum latency allowed by the SLA, in ms; is the current throughput, in Mbps; is the target throughput, in Mbps; is the indicator function, which takes 1 when α, β, γ are the weight coefficients of latency, throughput, and overload penalty, respectively.

6. The 5G network end-to-end service quality dynamic optimization method of claim 5, wherein: The state data The indicators are collected by the NWDAF function, the actions The output is issued by the SMF component, which maps The output is issued by the SMF component, which maps Independent meta-policy libraries are maintained for eMBB, uRLLC and other slice types, and when slice resource competition is detected, meta-learning is triggered for quick adaptation; Reward function integration with SLA: reward function weights Dynamic adjustment to match 5G SLA requirements and closed-loop feedback through a coordination controller.

7. The 5G network end-to-end quality of service dynamic optimization method of claim 6, wherein, The method for closed-loop integration of the algorithm and the 5G network architecture through a collaborative controller, periodic verification of SLA compliance, and triggering of model reconfiguration comprises: The collaborative controller is deployed on the 5G core network side and interacts with SMF and NWDAF components to obtain real-time network state data and service SLA indicators; SLA compliance verification: periodically calculate key performance indicators within a set value range, including end-to-end latency, throughput and packet loss rate, and compare them with SLA thresholds; when any indicator exceeds the threshold for three consecutive periods, it is determined to be an SLA violation; If an SLA violation is detected, the collaborative controller starts the meta-learning quick adaptation module to generate a reconfiguration instruction based on the current network state; reconfiguration actions include: Dynamic adjustment of the policy network parameters of the proximal policy optimization algorithm through inner loop adaptation to update task-specific parameters; Optimization of QoS rules, such as adjustment of 5QI levels or GBR values, and delivery of the rules to UPF through SMF; If the SLA indicators do not improve after reconfiguration, trigger the policy rollback mechanism to restore to the stable version of the policy; Closed-loop feedback: the reconfiguration effect is monitored by NWDAF and fed back to the collaborative controller for optimization of subsequent decision logic.

8. The 5G network end-to-end quality of service dynamic optimization method of claim 7, wherein, The cooperative controller dynamically adjusts the meta-learning adaptation frequency based on network load; when the load rate is , the adaptation frequency is raised to 1 time per second; when the load rate , the adaptation frequency is reduced to 1 time per 10 seconds, to balance calculation overhead and response speed; Design differentiated SLA verification strategies for different business types including eMBB and uRLLC; uRLLC business focuses on monitoring latency indicators, and eMBB business focuses on monitoring throughput indicators.

Images