Network traffic regulation and control system and method, and storage medium

Abstract

The invention provides a network flow regulation and control system and method and a storage medium, and relates to the technical field of network communication. The network traffic regulation and control system comprises: an information collection unit, which is used for obtaining historical traffic data of all media plane network elements connected with a control plane network element and a topological relation of the media plane network elements; the AI model updating unit is used for updating the traffic regulation and control model of the control plane network element according to the historical traffic data and the topological relation, and the strategy generation unit is used for obtaining a traffic distribution strategy of the control plane network element; and the strategy execution unit is used for dynamically adjusting the media traffic of each media plane network element according to the traffic distribution strategy. According to the method, the network data is collected and analyzed in real time, and dynamic updating of the AI model is combined, so that dynamic regulation and control of the network traffic are realized, the method can quickly adapt to changes of network conditions, and the flexibility and efficiency of traffic regulation and control are improved.

Description

Technical Field

[0001] The present invention relates to the field of network communication technology, and in particular to a network traffic regulation system, method and storage medium. Background Art

[0002] With the development of 5G networks, 5G services are becoming more and more extensive. Among them, the 5G new communication service is a new type of real-time audio and video communication service derived from the IMS data channel technology. During the call, it establishes an IMS data channel for both parties to transmit a variety of different multimedia files, such as music, videos, pictures, documents, and applets. The 5G new communication service can also be superimposed with XR technology to derive more personalized applications, such as video XR effects, dynamic XR backgrounds, etc., bringing users a new communication experience. However, since the 5G new communication service can transmit a variety of different multimedia files, some of which require large network traffic to transmit, and some of which require small network traffic to transmit, how to reasonably allocate network traffic has become a technical problem that needs to be solved urgently.

[0003] In the related art, multimedia files are transmitted through media plane units. In order to ensure the transmission effect of multimedia files such as streaming media files, it is usually necessary to allocate fixed media plane unit traffic according to the priority of each service. For example, the higher the priority, the greater the allocated traffic. However, the fixed traffic allocation may cause some low-priority services in the media plane unit to be unable to obtain sufficient bandwidth resources for a long time, resulting in serious impact on their transmission performance, which in turn affects the transmission effect of low-priority multimedia files. Summary of the invention

[0004] The problem solved by the present invention is how to improve the traffic regulation effect of the new 5G communication service to improve the transmission effect of various multimedia files.

[0005] In order to solve the above problems, the present invention provides a network traffic control system, method and storage medium.

[0006] In a first aspect, the present invention provides a network traffic control system, comprising: An information collection unit, used to obtain historical traffic data of all media plane network elements connected to the control plane network element and the topological relationship of the media plane network elements; An AI model updating unit, configured to update a traffic control model of the control plane network element according to the historical traffic data and the topological relationship, wherein the traffic control model includes a deep reinforcement learning sub-model, a fusion sub-model, a graph neural network sub-model, and a Transformer sub-model; A policy generation unit, configured to input the current traffic data of the current time period into the updated traffic regulation model, perform time series prediction through the fusion sub-model to obtain the time series prediction result of the current traffic data; perform topology analysis through the graph neural network sub-model to obtain the graph structure of the media plane network elements; perform weight allocation through the Transformer sub-model to obtain the allocation weight policy of each media plane network element; and then perform reinforcement learning through the deep reinforcement learning sub-model according to the time series prediction result, the graph structure, and the allocation weight policy to obtain the traffic allocation policy of the control plane network elements. A policy execution unit, configured to dynamically adjust the media traffic of each media plane network element according to the traffic allocation policy.

[0007] Optionally, the information collection unit is specifically configured to: Obtain the operation status data of all the media plane network elements under a preset time duration from historical data according to a preset collection period, where the operation status data includes the occupancy of the CPU and the occupancy of the GPU; Determine the network bandwidth occupancy data of the media plane network elements according to the data traffic inflow and outflow information of the network interfaces of the media plane network elements in the historical data; Obtain the load data of the media plane network elements according to the signaling throughput data of all the media plane network elements in the historical data; Use the operation status data, the network bandwidth occupancy data, and the load data as the historical traffic data.

[0008] Optionally, the AI model update unit is specifically configured to: Input the historical traffic data and the topological relationship into the traffic regulation model for training; Verify the traffic allocation policy output by the traffic regulation model during training through a digital twin environment and A / B testing; When the traffic allocation policies output by the traffic regulation model during training all meet the test requirements of the digital twin environment and the A / B testing, stop training the traffic regulation model to obtain the updated traffic regulation model.

[0009] Optionally, the policy generation unit is specifically configured to: Obtain the operation status data, the network bandwidth occupancy data, and the load data of the media plane network elements within the current time period, and use the operation status data, the network bandwidth occupancy data, and the load data as the current traffic data to input into the updated traffic regulation model; Extract features from the current traffic data through the CNN network of the fusion sub-model to obtain the spatial features of the current traffic data; Extract features from the current traffic data according to the BiLSTM network of the fusion sub-model to obtain the time series of the current traffic data; Based on the spatial features and the time series, obtain the traffic change trend of the media plane network element within a preset future time period, and use the traffic change trend as the time series prediction result.

[0010] Optionally, the policy generation unit is specifically configured to: Input the topological relationship of the media plane network element within the current time period into the updated traffic regulation model; Through the graph neural network sub-model, construct the graph structure according to the topological relationship, where the nodes of the graph structure are the media plane network elements, and the edges of the graph structure are the connection relationships of the media plane network elements.

[0011] Optionally, the policy generation unit is specifically configured to: Obtain the operation status data, network bandwidth occupancy data, and load data of the media plane network element within the current time period, and use the operation status data, network bandwidth occupancy data, and load data as the current traffic data to input into the updated traffic regulation model; Through the Transformer sub-model, perform prediction according to the current traffic data to obtain the critical values of multiple preset metrics of each media plane network element within a preset future time period; Generate weight values for each preset metric according to the critical values to obtain the allocation weight policy of each media plane network element.

[0012] Optionally, the policy generation unit is specifically configured to: Perform reinforcement learning through the deep reinforcement learning sub-model according to the time series prediction result, the graph structure, and the allocation weight policy; Among them, according to the time series prediction result, the graph structure, and the allocation weight policy, generate an initial traffic allocation adjustment policy; Use the initial traffic allocation adjustment policy as an action, and use the digital twin environment as the environment, and calculate the reward value after the interaction between the action and the environment through the reward function; Judge whether the reinforcement learning of the initial traffic allocation adjustment policy is completed according to the reward value. When the reinforcement learning of the initial traffic allocation adjustment policy is completed, obtain the traffic allocation policy of the control plane network element.

[0013] Optionally, it further includes an early warning unit, and the early warning unit is configured to: Obtain the traffic change trend within the current time period according to the current traffic data of the current time period; Judge whether there is a sudden increase in traffic in the media plane network element according to the traffic change trend; When there is a sudden increase in traffic in the media plane network element, send an early warning message.

[0014] In a second aspect, the present invention provides a network traffic regulation method, including: Obtain the historical traffic data of all media plane network elements connected to the control plane network element and the topological relationship of the media plane network elements; Update the traffic regulation model of the control plane network element according to the historical traffic data and the topological relationship, and the traffic regulation model includes a deep reinforcement learning sub-model, a fusion sub-model, a graph neural network sub-model, and a Transformer sub-model; Input the current traffic data of the current time period into the updated traffic regulation model, perform time series prediction through the fusion sub-model to obtain the time series prediction result of the current traffic data; perform topological analysis through the graph neural network sub-model to obtain the graph structure of the media plane network element; perform weight allocation through the Transformer sub-model to obtain the allocation weight strategy of each media plane network element; and then perform reinforcement learning through the deep reinforcement learning sub-model according to the time series prediction result, the graph structure, and the allocation weight strategy to obtain the traffic allocation strategy of the control plane network element; Dynamically adjust the media traffic of each media plane network element according to the traffic allocation strategy.

[0015] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the above-mentioned network traffic regulation method is implemented.

[0016] The network traffic regulation system, method and storage medium of the present invention achieve the joint regulation of media plane network elements through an innovative multi-module cooperation mechanism, effectively solving the limitations of the fixed traffic allocation method in the prior art. The information collection unit collects historical traffic data (such as device CPU / GPU occupancy, short-term or instantaneous traffic, bandwidth occupancy, etc.) and the topological relationship of devices from media plane network elements at regular intervals, providing a detailed data basis for subsequent model training; enabling the system to comprehensively understand the historical state of the network and the connection relationship between devices, thus providing a basis for dynamic traffic allocation. The AI model update unit continuously updates the traffic regulation model through historical traffic data and topological relationship. The model includes a deep reinforcement learning sub-model, a fusion sub-model, a graph neural network sub-model and a Transformer sub-model. The cooperative effect of these sub-models enables the system to adapt to different network environments and conditions, thus improving the flexibility and efficiency of traffic allocation. The policy generation unit can predict the traffic trend in advance through the fusion sub-model for time series prediction, providing forward-looking guidance for traffic regulation; the graph neural network sub-model analyzes the topological relationship of devices, helping to reasonably allocate network resources; the Transformer sub-model performs weight allocation according to device performance and network conditions to ensure that critical services get sufficient bandwidth resources first. Moreover, the generation of these policies is based on dynamic network conditions and real-time data, also avoiding the unfairness problem in the traditional fixed allocation method. The policy execution unit dynamically adjusts the traffic of each media plane network element according to the traffic allocation policy obtained by the policy generation unit. By monitoring the network conditions in real time and quickly adjusting the traffic allocation, the policy execution unit can ensure the stability and efficiency of the network. For example, when the load of a certain media plane network element is too high, the system can transfer some of its traffic to other devices to achieve load balancing.

[0017] The technical solution of the present invention realizes the dynamic regulation of network traffic by collecting and analyzing network data in real time and combining with the dynamic update of the AI model. It can quickly adapt to the changes in network conditions, improving the flexibility and efficiency of traffic regulation. Through the weight allocation by the Transformer sub-model, the system can reasonably allocate bandwidth resources according to business priorities and device performance, avoiding the problem that low-priority services cannot obtain sufficient bandwidth for a long time and enhancing the fairness of traffic regulation. By using the graph neural network sub-model to analyze the graph structure, it realizes an in-depth understanding of the complex topological relationships between media plane network elements, thereby coordinating the traffic allocation of multiple network elements. This joint regulation method breaks the limitations of traditional single regulation. By comprehensively considering multiple factors such as timing, topology, and weight, the traffic allocation is more flexible and accurate, effectively avoiding the problem that low-priority services cannot obtain sufficient bandwidth for a long time. At the same time, the system can dynamically adjust the traffic allocation according to the real-time changes in network conditions and business requirements, adapting to the dynamic changes in network traffic demands and ensuring the transmission effect of various multimedia files. Meanwhile, the deep reinforcement learning sub-model can perform reinforcement learning based on the timing prediction results, graph structure, and weight allocation strategy to optimize the traffic allocation strategy, realizing the intelligence and automation of traffic regulation. By effectively managing network resources, it reduces the degradation of service quality caused by traffic congestion and enhances the user experience when using services such as high-definition video and XR.

[0018] In summary, through the synergistic effect of links such as information collection, model update, policy generation, and policy execution, the present invention realizes the intelligent regulation of 5G new communication service traffic, significantly enhancing the dynamics, flexibility, fairness, and resource utilization rate of traffic regulation, thus solving the problems in the prior art and achieving the goal of improving the traffic regulation effect of 5G new communication services. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a block diagram of the network traffic regulation system according to an embodiment of the present invention; Figure 2 It is a schematic diagram of the 5G new communication network architecture according to an embodiment of the present invention; Figure 3 It is a schematic diagram of the connection between the control plane and media plane devices according to an embodiment of the present invention; Figure 4 It is a flowchart of the network traffic regulation method according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following provides a detailed description of specific embodiments of the present invention with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments described herein. Instead, these embodiments are provided to more thoroughly and completely understand the present invention. It should be understood that the drawings and embodiments of the present invention are only for illustrative purposes and are not used to limit the scope of protection of the present invention.

[0021] It should be understood that the various steps described in the method embodiments of the present invention can be executed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this regard.

[0022] As used herein, the term "including" and its variants are open-ended, that is, "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "optional embodiments". The relevant definitions of other terms will be given in the following description. It should be noted that the concepts such as "first" and "second" mentioned in the present invention are only used to distinguish different devices, modules, or units, and are not used to limit the order of functions performed by these devices, modules, or units or their interdependent relationships.

[0023] It should be noted that the modifications of "one" and "plural" mentioned in the present invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise clearly specified in the context, it should be understood as "one or more".

[0024] The names of the messages or information exchanged between multiple devices in the embodiments of the present invention are only for illustrative purposes and are not used to limit the scope of these messages or information.

[0025] The 5G new communication service derives an emerging real-time audio and video communication service based on the IMS Data Channel (IP Multimedia Subsystem (IMS) is a framework that supports IP-based multimedia communication services, DC, data channel) technology. During a call, an IMS DC is established for both parties to transmit various multimedia files, including music, videos, pictures, documents, mini-programs, etc. Additionally, XR technology (Extended Reality) can be superimposed to achieve personalized applications such as video XR effects and dynamic XR backgrounds, as well as various personalized voice services such as intelligent customer service and real-time translation. Among them, XR technology is an environment of the integration of reality and virtuality and human-computer interaction generated by computer technology and wearable devices. The XR service type includes various technical forms, including Virtual Reality (VR), Augmented Reality (AR), Mixed Reality (MR), etc. Its service process involves transforming the existing IMS network and adding control plane and media plane network elements for 5G new communication to establish a data channel and basic service processes. However, since the 5G new communication service can transmit various multimedia files, especially streaming media files such as high-definition videos, which will occupy a large amount of network traffic, traffic control and scheduling of the network have become urgent problems to be solved.

[0026] In the prior art, the media plane unit mainly relies on a fixed traffic allocation method based on service priorities to transmit multimedia files. In this process, the media plane unit first determines the priorities of various services. For example, services with high real-time requirements such as streaming media files are determined as high-priority services, while services with relatively low real-time requirements such as ordinary pictures and documents are determined as low-priority services. Then, according to the pre-set priority rules, a fixed proportion of traffic is allocated to each service. The higher the priority, the larger the allocated traffic. Each media plane unit uses its own allocated fixed traffic channel to transmit the corresponding multimedia file. During the transmission process, regardless of how the network conditions change, the traffic allocation ratio remains relatively fixed and is difficult to flexibly adjust according to real-time requirements.

[0027] However, due to the limited traffic allocated to low-priority services, they may not be able to obtain sufficient bandwidth resources for a long time, resulting in a serious impact on their transmission performance, such as slow picture loading and document transmission interruption. On the other hand, network traffic demands are dynamically changing. The prior art cannot timely sense these changes and make corresponding adjustments, making it difficult to adapt to complex network environments, resulting in a serious impact on its transmission performance, and further affecting the transmission effect of low-priority multimedia files.

[0028] In view of the problems existing in the above related technologies, this embodiment provides a network traffic regulation system, method, and storage medium.

[0029] Combined with Figure 1 As shown, a network traffic regulation system provided by an embodiment of the present invention includes: An information collection unit, configured to obtain historical traffic data of all media plane network elements connected to the control plane network element and the topological relationship of the media plane network elements.

[0030] Specifically, the information collection unit can regularly collect historical traffic data from media plane network elements, including device CPU / GPU occupancy, short-term or instantaneous data traffic inflow / outflow, network bandwidth occupancy, etc. In addition, it is also responsible for collecting the connection relationships and geographical distribution of media plane network elements, constructing a topological relationship graph of the device cluster, and providing basic data support for subsequent traffic regulation.

[0031] Exemplarily, in a 5G new communication network, the information collection unit mainly collects data from the following media plane network element devices, such as media servers, video processing devices, media gateways, edge computing nodes, CDN nodes, and signaling processing devices. These devices usually exist in the form of rack-mounted servers, dedicated processing devices, or network nodes, and are deployed in data centers, communication machine rooms, or network edges. The information collection unit regularly obtains data such as the CPU / GPU (CPU, central processing unit, and GPU, Graphics Processing Unit) occupancy rate, traffic changes, and bandwidth usage of each device. For example, the traffic of a media server during a live broadcast can reach 10 Gbps (GigaBits Per Second, a unit representing data transmission rate, often used to describe the high speed of network communication), the average bandwidth occupancy of a CDN (Content Delivery Network) node is 5 Gbps, and the traffic of an edge computing node when processing local requests is 1.2 Gbps. These data provide comprehensive data support for the traffic regulation system and lay a foundation for realizing efficient network resource management and optimization.

[0032] An AI model update unit, configured to update the traffic regulation model of the control plane network element according to the historical traffic data and the topological relationship, where the traffic regulation model includes a deep reinforcement learning sub-model, a fusion sub-model, a graph neural network sub-model, and a Transformer sub-model.

[0033] Specifically, the AI model update unit trains and optimizes the deep reinforcement learning sub-model based on the collected historical traffic data and topological relationships to improve its decision-making ability in complex network environments. At the same time, the fusion sub-model is trained using historical data to enable more accurate time series prediction. The graph neural network sub-model is updated according to the topological relationships to better understand and process the connection relationships between network devices. The Transformer sub-model is also adjusted based on the new data to optimize its performance in weight allocation. Through this comprehensive model update method, it is ensured that the traffic regulation model can adapt to the changing network conditions.

[0034] The policy generation unit is used to input the current traffic data of the current time period into the updated traffic regulation model, perform time series prediction through the fusion sub-model to obtain the time series prediction result of the current traffic data; perform topological analysis through the graph neural network sub-model to obtain the graph structure of the media plane network element; perform weight allocation through the Transformer sub-model to obtain the allocation weight policy for each media plane network element; and then perform reinforcement learning through the deep reinforcement learning sub-model according to the time series prediction result, the graph structure, and the allocation weight policy to obtain the traffic allocation policy for the control plane network element.

[0035] Specifically, the policy generation unit first inputs the current traffic data into the traffic regulation model. The fusion sub-model in the traffic regulation model combines historical data for time series prediction and outputs the short-term and long-term trends of the current traffic. Then, the graph neural network sub-model analyzes the topological structure of the media plane network element to determine the positions and mutual relationships of each device in the network. Next, the Transformer sub-model assigns weights to each media plane network element according to the device performance indicators and network conditions, determining its priority in network traffic allocation. Finally, the deep reinforcement learning sub-model synthesizes the time series prediction result, the graph structure, and the weight allocation policy, and generates the optimal traffic allocation policy through the reinforcement learning algorithm to achieve network load balancing and efficient utilization of resources.

[0036] The policy execution unit is used to dynamically adjust the media traffic of each media plane network element according to the traffic allocation policy.

[0037] Specifically, after receiving the traffic allocation policy from the policy generation unit, the policy execution unit adjusts the traffic of each media plane network element in real time according to this policy. For example, when the load of a certain media plane network element is too high, the policy execution unit will transfer some of its traffic to other media plane network elements in the same province or adjacent regions to relieve the burden on this device. At the same time, it can also dynamically adjust the traffic allocation according to the real-time status of the network to ensure the stable operation of the network and a good experience for users. In addition, the policy execution unit will monitor the network status in real time so as to make rapid adjustments when necessary to cope with sudden traffic changes.

[0038] The network traffic regulation system of the present invention realizes the joint regulation of media plane network elements through an innovative multi-module cooperation mechanism, effectively solving the limitations of the fixed traffic allocation method in the prior art. The information collection unit collects historical traffic data (such as device CPU / GPU occupancy, short-term or instantaneous traffic, bandwidth occupancy, etc.) and the topological relationship of the devices from the media plane network elements at regular intervals, providing a detailed data basis for subsequent model training; enabling the system to comprehensively understand the historical state of the network and the connection relationship between devices, thereby providing a basis for dynamic traffic allocation. The AI model update unit continuously updates the traffic regulation model through historical traffic data and topological relationships. This model includes a deep reinforcement learning sub-model, a fusion sub-model, a graph neural network sub-model, and a Transformer sub-model. The collaborative effect of these sub-models enables the system to adapt to different network environments and conditions, thereby improving the flexibility and efficiency of traffic allocation. The policy generation unit can predict the traffic trend in advance through the fusion sub-model for traffic regulation, providing forward-looking guidance for traffic regulation; the graph neural network sub-model analyzes the topological relationship of the devices, which helps to reasonably allocate network resources; the Transformer sub-model assigns weights according to the device performance and network conditions to ensure that key services can obtain sufficient bandwidth resources first. Moreover, the generation of these policies is based on dynamic network conditions and real-time data, also avoiding the unfairness problem in the traditional fixed allocation method. The policy execution unit dynamically adjusts the traffic of each media plane network element according to the traffic allocation policy obtained by the policy generation unit. By monitoring the network status in real time and quickly adjusting the traffic allocation, the policy execution unit can ensure the stability and efficiency of the network. For example, when the load of a certain media plane network element is too high, the system can transfer some of its traffic to other devices to achieve load balancing.

[0039] The technical solution of the present invention realizes the dynamic regulation of network traffic by collecting and analyzing network data in real time and combining with the dynamic update of the AI model. It can quickly adapt to the changes in network conditions, improving the flexibility and efficiency of traffic regulation. Through the weight allocation by the Transformer sub-model, the system can reasonably allocate bandwidth resources according to service priorities and device performance, avoiding the problem that low-priority services cannot obtain sufficient bandwidth for a long time and enhancing the fairness of traffic regulation. By using the graph neural network sub-model to analyze the graph structure, an in-depth understanding of the complex topological relationships between media plane network elements is achieved, thereby coordinating the traffic allocation of multiple network elements. This joint regulation method breaks the limitations of traditional single regulation. By comprehensively considering multiple factors such as time series, topology, and weights, the traffic allocation is made more flexible and accurate, effectively avoiding the problem that low-priority services cannot obtain sufficient bandwidth for a long time. At the same time, the system can dynamically adjust the traffic allocation according to the real-time changes in network conditions and service requirements, adapting to the dynamic changes in network traffic demands and ensuring the transmission effect of various multimedia files. Meanwhile, the deep reinforcement learning sub-model can perform reinforcement learning based on the time series prediction results, graph structure, and weight allocation strategy to optimize the traffic allocation strategy, realizing the intelligentization and automation of traffic regulation. By effectively managing network resources, the degradation of service quality caused by traffic congestion is reduced, enhancing the user experience when using services such as high-definition video and XR.

[0040] In summary, through the collaborative action of links such as information collection, model update, policy generation, and policy execution, the present invention realizes the intelligent regulation of 5G new communication service traffic, significantly enhancing the dynamics, flexibility, fairness, and resource utilization rate of traffic regulation, thereby solving the problems in the prior art and achieving the goal of improving the traffic regulation effect of 5G new communication services.

[0041] Optionally, the information collection unit is specifically used for: Obtaining the operation status data of all the media plane network elements within a preset duration from historical data according to a preset collection period, where the operation status data includes the CPU occupancy and GPU occupancy; Determining the network bandwidth occupancy data of the media plane network elements according to the data traffic inflow and outflow information of the network interfaces of the media plane network elements in the historical data; Obtaining the load data of the media plane network elements according to the signaling throughput data of all the media plane network elements in the historical data; Taking the operation status data, the network bandwidth occupancy data, and the load data as the historical traffic data.

[0042] Specifically, the information collection unit obtains the operation status data of all media plane network elements within a preset duration from historical data according to a preset collection period. Among them, these operation status data cover the CPU occupancy and GPU occupancy of the media plane network elements. By analyzing these data, the usage of the computing resources of the device can be understood, providing an important basis for subsequent evaluation of the processing capacity and load status of the device. Based on the data traffic inflow and outflow information of the network interfaces of the media plane network elements in the historical data, the information collection unit determines the network bandwidth occupancy data of each media plane network element; the data traffic inflow and outflow information of the network interfaces reflects the scale and speed of data transmission between the media plane network elements and other parts of the network. By analyzing this information, the bandwidth resources occupied by each media plane network element in network communication can be accurately calculated, thus providing key network-level data support for traffic regulation.

[0043] In addition, the information collection unit also obtains the load data of each media plane network element according to the signaling throughput data of all media plane network elements in the historical data. The signaling throughput data reflects the working intensity and efficiency of the media plane network elements when processing signaling messages. By analyzing and processing these data, the load situation of the device in signaling processing can be obtained, further enriching the evaluation of the overall operation status of the media plane network elements.

[0044] Finally, the information collection unit integrates the collected operation status data, the determined network bandwidth occupancy data, and the obtained load data together as the historical traffic data. These historical traffic data provide a comprehensive, detailed, and multi-dimensional data basis for subsequent AI model updates and traffic regulation strategy generation, enabling the traffic regulation system to learn and optimize based on rich historical information, thus more accurately predicting and responding to changes in network traffic and achieving precise traffic regulation.

[0045] It should be noted that in an embodiment of the present invention, the operation status data collected by the information collection unit covers the CPU occupancy and GPU occupancy of the media plane network element. Here, the CPU and GPU specifically refer to the central processing unit (CPU) and the graphics processing unit (GPU) in the media plane network element device. Specifically, the CPU is the core processor in the media plane network element device. In the media plane network element, the CPU is mainly responsible for tasks such as processing control signaling, protocol processing of media streams, resource scheduling, and execution of business logic. For example, processing signaling messages from the control plane network element, executing forwarding and routing decisions for media streams, and managing resource allocation of the device. The CPU occupancy reflects the usage of computing resources of the device, including the busyness of processing tasks, resource utilization, etc. By monitoring the CPU occupancy, it is possible to evaluate whether the processing capacity of the device is sufficient and whether there is overloading or resource waste. The GPU is the graphics processing unit in the media plane network element device, which is specifically used to accelerate graphics and image processing tasks. In the media plane network element, the GPU is mainly used to process graphics and image tasks related to media streams, such as video encoding and decoding, image rendering, video special effect processing, XR (extended reality) effect generation, etc. For example, performing real-time encoding and decoding of high-definition videos, generating virtual backgrounds, and implementing special effect processing in videos. The GPU occupancy reflects the resource usage of the device in graphics and image processing. By monitoring the GPU occupancy, it is possible to evaluate the load status of the device when processing graphics tasks of media streams, ensure the reasonable utilization of GPU resources, and avoid delays or stuttering in media stream processing caused by GPU overload.

[0046] Furthermore, the data on CPU and GPU occupancy collected by the information collection unit specifically comes from the central processing unit (CPU) and the graphics processing unit (GPU) in the media plane network element device. These data comprehensively reflect the usage of computing resources of the media plane network element when processing media streams and executing business logic. By analyzing these data, it is possible to accurately evaluate the processing capacity and load status of the device, providing a detailed and multi-dimensional data basis for subsequent training of the traffic regulation model and formulation of strategies. Based on these data, the system can better perform resource allocation and load balancing, improve the utilization efficiency of network resources, and ensure the stable operation and good performance of the network.

[0047] In this optional embodiment, by obtaining the operating status data, network bandwidth occupancy data, and load data within a preset duration according to a preset collection period, the historical traffic data of the media plane network element can be comprehensively and systematically collected. These data cover various aspects of the operating status of the media plane network element, such as the CPU and GPU occupancy, the inflow and outflow information of data traffic on the network interface, and the signaling throughput data, providing a detailed and multi-dimensional data basis for subsequent traffic regulation model training and strategy formulation. Based on the operating status data, the computing resource usage of the media plane network element can be accurately understood, thereby evaluating its processing ability. By determining the network bandwidth occupancy data, the system can precisely master the bandwidth resources occupied by each media plane network element in network communication, providing key network-level data support for traffic regulation. By obtaining the load data, the evaluation of the overall operating status of the media plane network element is further enriched, helping to comprehensively understand the load situation of the device.

[0048] Moreover, using these historical traffic data as the input of the traffic regulation model provides rich information for the training and optimization of the model. Through the analysis and mining of historical traffic data, the traffic regulation model can more accurately predict and respond to changes in network traffic, achieving precise traffic regulation. Based on comprehensive and accurate historical traffic data, the system can better perform resource allocation and load balancing, improve the utilization efficiency of network resources, ensure the stable operation and good performance of the network. By effectively managing network resources, the degradation of service quality caused by traffic congestion is reduced, enhancing the user experience when using services such as high-definition video and XR.

[0049] In summary, the setting and function realization of the information collection unit can comprehensively and accurately collect the historical traffic data of the media plane network element, provide key data support for the traffic regulation system, help improve the effect of traffic regulation, and achieve the efficient utilization and optimized management of network resources.

[0050] Optionally, the AI model update unit is specifically used for: Inputting the historical traffic data and the topological relationship into the traffic regulation model for training; Verifying the traffic allocation strategy output by the traffic regulation model during training through the digital twin environment and A / B testing; When the traffic allocation strategies output by the traffic regulation model during training all meet the test requirements of the digital twin environment and the A / B testing, stop training the traffic regulation model to obtain the updated traffic regulation model.

[0051] Specifically, the historical traffic data collected by the information collection unit, including operation status data such as the CPU / GPU occupancy of devices, network bandwidth occupancy data, and load data, as well as the topological relationship information of media plane network elements, are input into the traffic regulation model for training. During the training process, the model continuously adjusts its own parameters based on this data to learn the laws of network traffic changes and the mutual influence relationships between media plane network elements. And digital twin environment and A / B testing are used to verify the traffic allocation strategy output during training; the digital twin environment can simulate a virtual environment highly consistent with the actual network environment. Testing the traffic allocation strategy in this environment can observe in advance the performance of the strategy under various possible network conditions, such as the effects under high load and burst traffic, so as to evaluate the feasibility and effectiveness of the strategy. At the same time, A / B testing is to select a part of the traffic or devices in the actual network environment for testing, and compare the test results with the situation without using this strategy, intuitively reflecting the improvement effect of the traffic allocation strategy on network performance, user experience, etc. By combining these two verification methods, the quality of the traffic allocation strategy can be evaluated more comprehensively and accurately. If the traffic allocation strategy output by the traffic regulation model during the training process meets the test requirements in both the digital twin environment and A / B testing, for example, can effectively achieve network load balancing, improve resource utilization rate, and ensure the transmission quality of key services, then the training of the traffic regulation model will be stopped, thus obtaining an updated traffic regulation model. This updated model contains the knowledge and laws learned from the previous training and can provide a more accurate and reliable basis for the generation of subsequent traffic allocation strategies.

[0052] Exemplarily, the AI (Artificial Intelligence) model update unit makes a judgment based on the resource usage conditions of each media plane network element collected and the trained large model on whether it is necessary to dynamically adjust and schedule nearby the media resources transmitted by some media plane network elements with too high resource utilization rate. If so, other media plane network elements in the same province or adjacent regions are selected for media resource transmission to reduce the load of individual media plane network elements. If the loads of media plane network elements in the same region are all high during some periods, the AI model update unit is used to schedule other adjacent region media plane network elements for load sharing.

[0053] Specifically, the control plane network element collects data such as the device CPU / GPU occupancy of the media plane network element, the short-term (taking 1 hour as an example, and the specific time can be adjusted by network operation and maintenance personnel) or instantaneous data traffic inflow / outflow, bandwidth occupancy, short-term / instantaneous signaling throughput, etc. through the AI model update unit. The operation and maintenance personnel set processing thresholds according to the network device conditions. Data exceeding the threshold is regarded as alarm data, and data within the threshold range is used as normal device operation data. Both types of data are sent to the AI model update unit for training to obtain an AI model that can automatically warn of signaling and traffic exceeding the threshold for the media plane network element. The control plane network element regularly sends requests to the media plane network element to obtain device condition data. When the data fed back by a certain media plane network element device obtained exceeds the threshold, the AI model update unit will immediately send device data acquisition requests to the media plane network elements connected to the periphery of the control plane network element in real time. Among the data of other surrounding media plane network elements collected, one or more media devices are selected according to influencing factors such as device CPU / GPU occupancy, delay, traffic, bandwidth occupancy, signaling throughput, etc. to share the signaling and media data traffic of the alarm device.

[0054] Moreover, for scenarios of sudden increase in instantaneous traffic, such as the appearance of popular stars in live shows, the appearance / winning of star players in sports events, etc., the operation and maintenance data of the control plane and media plane devices that have not been introduced into the AI model update unit can be collected for preliminary AI model training. During the actual operation of the network device, an alarm mechanism is set. When the signaling / bandwidth occupancy exceeds 2 times the daily data, an alarm notification is obtained. The AI model is mainly used for dynamic regulation, and the operation and maintenance personnel participate in model regulation. At the same time, a time period for predicting sudden increase in traffic is reserved for the AI model update unit. The AI model for sudden increase in traffic is enabled in advance during the reserved time period, or it is predicted in advance that the operation and maintenance personnel will pay attention to traffic changes synchronously and participate in AI model regulation.

[0055] In this alternative embodiment, by inputting rich historical traffic data and topological relationships into the model for training, the model can more comprehensively understand the operation of the network and the laws of traffic changes, thereby improving the accuracy of the traffic allocation strategy it generates. At the same time, the digital twin environment and A / B testing are used to strictly verify the strategy, ensuring that only strategies that meet the actual requirements and achieve the expected effects will be adopted, further enhancing the reliability of the model. Among them, the digital twin environment simulates various possible network conditions, enabling the model to be exposed to rich scenarios during the training process, thus enhancing its adaptability to different network environments. When the actual network environment changes, the model trained and verified in this way can make more rapid and effective adaptive adjustments, generate a traffic allocation strategy suitable for the current situation, and improve the practicality of the model in actual applications. The updated traffic control model can more accurately formulate a reasonable traffic allocation plan according to the actual operating state of the network and business requirements. This helps to optimize the allocation of network resources, improve the utilization efficiency of resources, avoid the situation where some devices are overloaded while others are idle, and thus enhance the performance and service quality of the entire network.

[0056] In addition, an accurate and reliable traffic allocation strategy helps to ensure the transmission quality of critical services and reduce problems such as service interruptions or delays caused by traffic congestion. For users, it means a smoother and more stable experience when using services such as high-definition video and XR. At the same time, reasonable resource allocation and effective traffic control are also conducive to maintaining the stable operation of the network, reducing the risk of system failures, and improving the reliability and stability of the entire network.

[0057] Optionally, the policy generation unit is specifically configured to: Obtain the operation state data, network bandwidth occupancy data, and load data of the media plane network element during the current time period, and use the operation state data, network bandwidth occupancy data, and load data as the current traffic data to input into the updated traffic control model; Extract features from the current traffic data through the CNN network of the fusion sub-model to obtain the spatial features of the current traffic data; Extract features from the current traffic data according to the BiLSTM network of the fusion sub-model to obtain the time series of the current traffic data; Obtain the traffic change trend of the media plane network element within a preset future time period according to the spatial features and the time series, and use the traffic change trend as the time series prediction result.

[0058] Specifically, the policy generation unit first obtains the operation status data, network bandwidth occupancy data, and load data of the media plane network elements within the current time period. Moreover, these data are consistent with the historical traffic data collected by the information collection unit in terms of type, including the CPU / GPU occupancy of devices, the network bandwidth occupancy determined by the data traffic inflow and outflow information of network interfaces, and the load conditions obtained from signaling throughput data, etc. Then, the obtained operation status data, network bandwidth occupancy data, and load data are integrated together and input into the updated traffic regulation model as the current traffic data. The fusion sub-model in the traffic regulation model includes a CNN network (Convolutional Neural Network) and a BiLSTM network (Bidirectional Long Short-Term Memory Neural Network). By using the CNN network to extract features from the current traffic data, the spatial features of the data can be mainly obtained. The convolution operation of the CNN can automatically learn the local spatial patterns and correlations in the data, such as the traffic distribution pattern in space between different devices, the pattern of resource occupancy, etc. By using the BiLSTM network to extract features from the current traffic data, the time series features of the data are obtained. The BiLSTM can perform bidirectional modeling on time series data, capturing the forward and backward dependencies and change trends of traffic data over time, such as the rising or falling trends of traffic at different time points, periodic changes, etc. After obtaining the spatial features and time series features of the current traffic data, the fusion sub-model combines these two features, further analyzes and predicts the traffic change trend of the media plane network elements within a preset future time period, and uses the prediction result as the time series prediction result, providing an important basis for formulating subsequent traffic allocation policies. By combining the features of the two dimensions of space and time, the prediction result can more comprehensively and accurately reflect the possible changes in future traffic, including which devices' traffic may increase, which may decrease, and the approximate amplitude and time range of traffic changes, etc.

[0059] In this optional embodiment, by using the CNN and BiLSTM networks in the fusion sub-model to extract the spatial features and time series features of the current traffic data respectively, various features and rules in the traffic data can be captured more comprehensively. The spatial features reflect the resource occupancy patterns and traffic distribution among different media plane network elements, and the time series features reflect the change trend of traffic over time. Combining these two features for traffic change trend prediction can greatly improve the prediction accuracy, enabling the system to understand the future traffic changes in advance and providing a more accurate basis for traffic regulation.

[0060] Moreover, based on the accurate prediction of the future traffic change trend of the media plane network element, the system can make corresponding traffic management decisions in advance. For example, if it is predicted that the traffic of a certain device will increase significantly in the future period and may exceed its processing capacity, then some traffic can be adjusted and allocated to other devices in advance to achieve dynamic traffic management. This dynamic adjustment ability helps to avoid network congestion and resource overload, ensuring the stable operation of the network. Accurately predicting the traffic change trend helps the system to allocate network resources more reasonably. By understanding in advance the changes in the traffic demands of each media plane network element, the resource allocation strategy can be dynamically adjusted according to the actual situation, making the resource allocation more refined and reasonable. This can not only improve the utilization efficiency of resources and avoid waste of resources, but also ensure that critical services and high-priority services can obtain sufficient resource support when needed, enhancing the quality and reliability of network services.

[0061] Optionally, the policy generation unit is specifically configured to: Input the topological relationship of the media plane network element in the current time period into the updated traffic regulation model; Through the graph neural network sub-model, construct the graph structure according to the topological relationship, where the nodes of the graph structure are the media plane network elements, and the edges of the graph structure are the connection relationships of the media plane network elements.

[0062] Specifically, the policy generation unit takes the topological relationship of the media plane network element in the current time period as the input. The topological relationship details the connection methods and network structures between each media plane network element, such as which devices have direct communication links and the hierarchical positions of the devices in the network. Through the graph neural network sub-model, the graph structure of the graph neural network is constructed according to the input topological relationship. In the graph structure, the media plane network elements are abstracted as the nodes of the graph, and the connection relationships between the media plane network elements are represented as the edges of the graph. Specifically, each node represents a specific media plane network element, and the attributes of the node can include information such as the device type, performance parameters, and current traffic load of the network element; the edge represents that there is a communication connection or data transmission path between two media plane network elements, and the weight of the edge can be associated with network parameters such as the connection bandwidth and delay, so as to reflect the tightness of the mutual connection and the communication ability between different devices.

[0063] In this alternative embodiment, the graph structure constructed by the graph neural network sub-model can accurately model the network topology relationship of the media plane network elements, enabling the traffic control system to intuitively understand the structural characteristics of the network and the mutual connection mode between devices, and integrating the network topology information into the traffic control decision-making process. Moreover, when formulating the traffic allocation strategy, it can fully consider the network positions of each media plane network element and its connection relationship with other devices, thereby making a more reasonable resource allocation decision. For example, for media plane network elements located on the critical path of the network or connected to more other devices, their resource requirements may be preferentially guaranteed to ensure smooth communication of the entire network. According to the graph structure, the system can adopt differentiated traffic control strategies for different regions or sub-networks in the network. For example, for relatively independent sub-networks in the topology structure, local traffic optimization can be performed according to the traffic conditions and topology relationships of the internal devices, while for core devices connecting multiple sub-networks, traffic control needs to be carried out from a global perspective to ensure the stable operation of the entire network.

[0064] By reasonably utilizing the graph structure of the graph neural network, the traffic control is made more in line with the actual topology of the network, thereby improving the performance and reliability of the entire network. Effective traffic control helps to avoid network congestion, reduce data transmission latency and packet loss rate, improve the utilization efficiency of network resources, ensure that the network can operate stably and efficiently, and improve the user experience.

[0065] Optionally, the policy generation unit is specifically configured to: Obtain the operation status data, the network bandwidth occupancy data, and the load data of the media plane network element during the current time period, and use the operation status data, the network bandwidth occupancy data, and the load data as the current traffic data to input into the updated traffic control model; Through the Transformer sub-model, perform prediction according to the current traffic data to obtain the critical values of multiple preset metrics of each media plane network element within a preset future time period; Generate the weight values of each preset metric according to the critical values to obtain the allocation weight policy of each media plane network element.

[0066] Specifically, the policy generation unit obtains the operation status data, network bandwidth occupancy data, and load data of the media plane network elements within the current time period. These data can reflect the operation status of the media plane network elements in real time, including the computing resource occupancy of the devices (such as the utilization rates of CPU and GPU), network resource occupancy (such as the inflow and outflow of data traffic), and signaling processing load, etc. Then, these data are integrated into the current traffic data and input into the traffic regulation model. The Transformer sub-model in the traffic regulation model is responsible for deeply analyzing and predicting the current traffic data. Through its powerful sequence processing ability and parallel computing advantages, the Transformer sub-model can capture the complex patterns and long-term dependencies in the data.

[0067] Specifically, the Transformer sub-model predicts the critical values of multiple preset metrics of each media plane network element within a preset time period in the future based on the current traffic data; these preset metrics include key performance indicators such as the resource utilization rate of the device, network bandwidth occupancy rate, and signaling processing delay; and the prediction of the critical values helps to identify potential resource bottlenecks or traffic congestion points in advance.

[0068] Among them, the critical value refers to a certain key threshold within the preset time period in the future, that is, the maximum value of the preset metric within the preset time period in the future, rather than simply referring to the maximum value or the value at the last moment. Specifically: the critical value is predicted based on the current traffic data and is used to represent a key threshold that a preset metric (such as device resource utilization rate, network bandwidth occupancy rate, signaling processing delay, etc.) of the media plane network element reaches within the preset time period in the future. The purpose of setting the critical value in this optional embodiment is to identify potential resource bottlenecks or traffic congestion points in advance so that the system can take measures to regulate the traffic in advance. The critical value is usually set based on the statistical analysis of historical data and network operation experience.

[0069] Exemplarily, if the historical data shows that there may be performance problems when the CPU utilization rate of a certain device exceeds 80%, then 80% is set as the critical value of the CPU utilization rate. The Transformer sub-model estimates whether a preset metric may reach or exceed this critical value at any moment within a certain time period in the future through the analysis and prediction of the current traffic data. If it is predicted that a preset metric may reach or exceed the critical value at any moment within the preset time period in the future, the system will consider that there is a risk for this metric and adjust the weight allocation policy accordingly to prioritize resource allocation to this media plane network element.

[0070] Another example, if the preset metric is network bandwidth occupancy rate and the critical value is set at 90%, and the Transformer sub-model predicts that at a certain future time, the bandwidth occupancy rate of a certain media plane network element will reach 90%, then 90% is the critical value. If the preset metric is the CPU utilization rate of a device, the critical value may be set at 80%. If it is predicted that at a certain future time, the CPU utilization rate of a certain device will reach 80%, this 80% is the critical value. In summary, in this alternative embodiment, the critical value is a key threshold used to determine whether a certain preset metric may reach or exceed the preset maximum critical value within a preset time period in the future, so that the system can perform traffic regulation in advance.

[0071] After obtaining the critical values of each preset metric, the policy generation unit generates weight values for each preset metric based on these critical values. The generation of weight values is based on the principle that key metrics or metrics close to the critical value will be assigned higher weights to reflect their importance and urgency in network traffic regulation. By comprehensively considering the weights of each metric, an allocation weight policy for each media plane network element is finally formed. This policy clarifies how much traffic resources each media plane network element should be preferentially allocated during traffic allocation, thereby providing a basis for subsequent dynamic traffic adjustment.

[0072] In this alternative embodiment, through in-depth analysis and prediction of the current traffic data by the Transformer sub-model, the critical values of the key performance indicators of each media plane network element can be accurately obtained. The weight values and allocation weight policies generated based on these critical values can achieve more precise resource allocation. For example, for a device whose resource utilization rate is about to reach the critical value, more traffic resources can be preferentially allocated to prevent resource exhaustion and ensure the stable operation of the network.

[0073] Moreover, this policy generation process can dynamically adjust the weight allocation according to real-time traffic data, making the traffic regulation adaptive. When the network condition changes, such as a sudden increase in the load of a certain media plane network element, the system can quickly identify and reallocate traffic by adjusting the weight policy, timely relieve network pressure and avoid congestion. This adaptive traffic regulation mechanism can effectively cope with various dynamic changes in the network, improving the flexibility and reliability of the network. Based on the prediction and weight generation mechanism of the Transformer sub-model, potential performance bottlenecks can be identified in advance and optimized through reasonable resource allocation. This helps to improve the performance of the entire network and reduce performance degradation caused by unreasonable resource allocation. By optimizing traffic allocation, data transmission latency can be reduced, network throughput can be increased, and the network experience of users can be improved, especially in application scenarios with high requirements for real-time performance and bandwidth, such as high-definition video conferencing, XR interaction, etc.

[0074] Finally, precise resource allocation and adaptive traffic regulation help reduce network overload and resource contention, thereby enhancing the reliability and stability of the system. In the face of sudden traffic or network anomalies, the system can adjust the traffic allocation strategy in a timely manner to ensure the continuity of critical services, reduce the risk of system failures, ensure the network remains stable, and improve the user network service experience.

[0075] Optionally, the policy generation unit is specifically configured to: Perform reinforcement learning through the deep reinforcement learning sub-model according to the time series prediction result, the graph structure, and the allocation weight policy; Wherein, an initial traffic allocation adjustment policy is generated according to the time series prediction result, the graph structure, and the allocation weight policy; Take the initial traffic allocation adjustment policy as an action, and take the digital twin environment as the environment, and calculate the reward value after the interaction between the action and the environment through the reward function; Judge whether the reinforcement learning of the initial traffic allocation adjustment policy is completed according to the reward value. When the reinforcement learning of the initial traffic allocation adjustment policy is completed, obtain the traffic allocation policy of the control plane network element.

[0076] Specifically, taking the time series prediction result, the graph structure, and the allocation weight policy as inputs, an initial traffic allocation adjustment policy is generated through the deep reinforcement learning sub-model to allocate traffic resources to each media plane network element. And taking the initial traffic allocation adjustment policy as an action and the digital twin environment as the reinforcement learning environment. The digital twin environment simulates the real network environment and traffic conditions, provides a platform for testing and optimizing the initial policy, and predicts the network state and performance indicators after traffic adjustment; the reward function is used to calculate the reward value after the interaction between the action and the environment. The reward function comprehensively considers the effects of the traffic allocation policy, such as network load balancing, resource utilization rate, and the transmission quality of critical services. A high reward value indicates a good policy, otherwise it needs to be optimized and adjusted. Judge whether the reinforcement learning is completed according to the reward value. If the reward value reaches the preset threshold or tends to be stable, it indicates that the optimization of the initial traffic allocation adjustment policy is completed, and the final traffic allocation policy is obtained.

[0077] In this optional embodiment, the deep reinforcement learning sub-model generates and optimizes the initial policy by combining time series prediction, graph structure, and weight strategy, improves the rationality of traffic allocation, enhances network performance. Simulating and evaluating the policy in the digital twin environment can also reduce the cost and risk of trial and error, improve the quality and reliability of decision-making, and ensure network stability. By continuously adjusting the policy through reinforcement learning, the system can adapt to the dynamic network environment, optimize traffic allocation, improve resource utilization and service performance. The deep reinforcement learning sub-model enhances the intelligent decision-making ability of the system, realizes automated and intelligent traffic regulation, reduces the burden on operation and maintenance personnel, and improves network management efficiency. The optimization and adjustment of reinforcement learning ensure the effectiveness and adaptability of the traffic allocation strategy, reduce congestion and overload, ensure service continuity, and enhance network stability. In addition, the optimized traffic allocation strategy guarantees the transmission quality of critical services and high-priority services, improves the user network service experience, especially in scenarios such as high-definition video and XR, reduces stuttering and latency, and improves satisfaction.

[0078] Optionally, it further includes a warning unit, and the warning unit is used to: Obtain the traffic change trend within the current time period according to the current traffic data of the current time period; Judge whether there is a sudden increase in traffic in the media plane network element according to the traffic change trend; When there is a sudden increase in traffic in the media plane network element, send a warning message.

[0079] Specifically, the warning unit obtains the current traffic data of the media plane network element in the current time period in real time, including operation status data, network bandwidth occupancy data, and load data, etc. By monitoring and analyzing these data in real time, the traffic change trend is calculated.

[0080] Exemplarily, time series analysis methods, such as moving average method, exponential smoothing method, etc., can be used to identify short-term fluctuations and long-term trends in traffic data. In addition, machine learning algorithms, such as linear regression, decision tree, etc., can be combined to model and predict traffic data to more accurately grasp the law of traffic change.

[0081] Based on the obtained traffic change trend, the warning unit determines whether there is a sudden increase in traffic in the media plane network element. The judgment basis for sudden traffic increase can include indicators such as the absolute increase in traffic, relative growth rate, growth rate, etc. For example, within a preset time window, the average growth rate of traffic can be calculated. If this rate exceeds the set threshold, it is determined that there is a sudden increase in traffic. The setting of the threshold can be based on the statistical analysis of historical data or dynamically adjusted according to the actual operating conditions of the network and business requirements. At the same time, algorithms for detecting traffic peaks, such as wavelet transform, sliding window extreme value detection, etc., can be combined to quickly identify the sudden increase in traffic. When it is determined that there is a sudden increase in traffic in the media plane network element, the warning unit promptly sends a warning message. The content of the warning message can include key information such as the device identifier of the sudden traffic increase, the scale of the sudden traffic increase, the time when the sudden increase occurred, and the possible scope of influence. The sending method of the warning message can be selected according to actual needs. For example, through means such as text messages, emails, instant messaging tools, and pop-up windows on the network management system interface, the warning message is promptly conveyed to network operation and maintenance personnel or relevant automated management systems. This can gain valuable time for subsequent traffic regulation and emergency handling, and take measures in advance to avoid network congestion and performance degradation.

[0082] In this optional embodiment, through the real-time monitoring and analysis of current traffic data, the warning unit can promptly detect the trend of sudden traffic increase and issue a warning before the problem occurs. This enables network operation and maintenance personnel or automated management systems to have sufficient time to take corresponding measures, such as adjusting traffic distribution strategies, increasing resource allocation, etc., thereby avoiding the occurrence of network congestion and performance problems in advance and improving the stability and reliability of the network. A sudden increase in traffic may lead to network congestion and affect the normal operation of services. Through early warning, measures such as traffic diversion and speed limiting can be taken in a timely manner to ensure the smooth operation of the network, improve the continuity and reliability of services, and reduce service interruptions or performance degradation caused by network problems. Moreover, early warning helps to reasonably arrange and allocate resources. Before the arrival of the traffic peak, additional bandwidth resources can be allocated in advance or traffic distribution strategies can be adjusted to ensure the stable operation of key services, improve the utilization efficiency of resources, and avoid resource waste. The warning message can also trigger the response of the automated management system, realizing the automated and intelligent management of the network, improving management efficiency, reducing the need for manual intervention, and alleviating the work burden of operation and maintenance personnel.

[0083] In a preferred embodiment of the present invention, in combination with Figure 2As shown in the figure, the main network elements in the 5G new communication network architecture and their interrelationships are presented. The control plane network element and the media plane network element are the core parts of the architecture. At the same time, other key network elements such as the business AS (Application Server, new communication application server) and the MPF (Media Processing Function, media capability platform) are also shown. The AI model update unit in the control plane is an embodiment of the network AI capability, which obtains real-time parameter information from the media plane through interface 5 for traffic control and optimization.

[0084] Specifically, in combination with Figure 2 introduce the main network elements and their functions. The NCP (5G New Communication Platform) complies with the 3GPP TS26.114 standard and supports the DCSF (DC Service Control Function) function; it is responsible for managing the establishment, maintenance, and release of the DC (Data Channel), triggering DC services; and provides a northbound interface for the business AS to call to achieve flexible customization and expansion of services. The UMF (Unified Media Function): complies with the 3GPP TS26.114 standard and supports the DCMF (DC Media Function) function; provides various DC service media capabilities, including but not limited to media replication, video stream synthesis, subtitle stream synthesis, background replacement, virtual portrait, etc.; supports establishing a DC channel with the terminal and transmitting application interaction information to achieve a rich media service experience. The business AS (Application Server, new communication application server) is used to provide various DC services, such as intelligent translation, XR calls, intelligent customer service, etc.; as the bearer entity of business logic, it works in coordination with the control plane network element and the media plane network element to provide convenient value-added services for users. The MPF (Media Processing Function, media capability platform) is a third-party platform that provides audio and video media enhancement processing capabilities; specific functions include speech-to-text conversion, translation, gesture recognition, etc., enriching the content and form of media services and enhancing the user experience. And, Figure 3Multiple interfaces (interface 1 to interface 16) are marked, and these interfaces define the communication paths and interaction methods between different network elements. For example: Interface 5 is the interface between the control plane and the media plane device, and the AI model update unit obtains real-time parameter information of the media plane through this interface. Interfaces such as interface 9 and interface 15 reflect the interaction relationships between the service AS and the control plane and the media plane, and are used for service triggering and data transmission. The control plane includes control plane network elements, which are responsible for signaling processing and the implementation of control functions; the AI model update unit is located in the control plane and realizes intelligent regulation of network traffic by analyzing the real-time data of the media plane network elements. The media plane includes media plane network elements, which are responsible for media stream processing and transmission; the media plane network elements serve media services, such as media stream forwarding, processing, and storage.

[0085] Among them, other network elements such as Figure 2 shown, such as ENUM / DNS, is used for domain name resolution and number query to help locate users and service servers; HSS / UDM is used for storing user subscription information and authentication data to support user identity authentication and service authorization; MMTel AS is used for providing control and management functions for traditional voice and multimedia telecommunications services; I / S-CSCF is used for responsible for the establishment and routing of the initial session, as well as user registration and authentication; P-CSCF / PCF is used as the entry point of the user plane and is responsible for session establishment and policy control; EPS / 5GS is used for providing connection and interoperability functions with the 4G / 5G core network to support user mobility and session management; UE (user equipment, which is a terminal device used by users, such as smartphones, tablets, etc., for accessing the 5G new communication network.

[0086] In this preferred embodiment, by showing the main network elements in the 5G new communication network architecture and their interrelationships, the application location and role of the network traffic control method that integrates the control plane network elements and AI technology in the present invention are reflected; the AI model update unit obtains real-time parameter information through the interface with the media plane device to realize intelligent regulation and optimization of network traffic.

[0087] In another preferred embodiment of the present invention, in combination with Figure 3As shown, the connection relationship between the control plane device and multiple media plane devices is presented, reflecting the network traffic control architecture in the present invention. Each control plane device includes an AI model update unit located inside the control plane device, which is responsible for collecting the real-time traffic data of the media plane devices, analyzing and making decisions, and generating traffic allocation strategies. Moreover, the two control plane devices are mutually primary and standby. One control plane device is the primary device, responsible for the main control and management functions; the other is the standby device, which takes over its functions when the primary device fails, ensuring the high availability and reliability of the system. Each control plane device is connected to multiple media plane devices; the media plane devices are responsible for the processing and transmission of media streams. Each control plane device is connected to multiple media plane devices to achieve centralized management and control of these media plane devices. There is also a connection between the two control plane devices for communication and data synchronization between the primary and standby devices. There is a direct connection between each media plane device and the control plane device for transmitting media streams and control signaling. The media plane devices provide real-time traffic data to the AI model update unit of the control plane device, including the operating status of the devices, network bandwidth occupancy, load conditions, etc.

[0088] In this preferred embodiment, the AI model update unit in the control plane device analyzes and predicts by periodically collecting the real-time traffic data of the connected media plane devices. The AI model update unit generates the optimal traffic allocation strategy based on the collected data using technologies such as deep reinforcement learning, fusion sub-models, graph neural networks, and Transformers. The AI model update unit dynamically adjusts the traffic allocation between the media plane devices according to the real-time traffic data to achieve network load balancing. When the traffic load of a certain media plane device is too high, the AI model update unit will transfer part of the traffic to other devices with lower load to ensure the stable operation of the network.

[0089] Exemplarily, the AI model update units of two control plane devices are set with the same timer duration, with time intersection (the start time of one timer is half a cycle earlier than that of the other timer), and they make decisions alternately, sharing decision data; the collected media resource parameters include but are not limited to data such as the CPU / GPU occupancy of the device, the short-term (taking 1 hour as an example, and the specific time can be adjusted by network operation and maintenance personnel) or instantaneous data traffic inflow / outflow, network bandwidth occupancy, short-term / instantaneous signaling throughput, etc. Data collection and model training are carried out during the trial commercialization stage of network devices. A digital twin environment is established to simulate a user scale of millions. Methods such as weighted round-robin / simulated annealing are used to compare with traditional algorithm baselines. The A / B test covers at least more than 10 pressure scenarios. In the formal stage, the time series prediction module is first launched as an early warning system, and the reinforcement learning controller is enabled in stages. Finally, through AI suggestions and the review of operation and maintenance personnel, the trained model is used for network device resource scheduling. During the operation process, a model deviation detection mechanism is established, a real-time visual heat map of each feature weight is made and displayed, and the policy rollback trigger condition is set for real-time monitoring to facilitate network operation and maintenance adjustment. When the instantaneous burst traffic is too large (2 times or more), the media plane device scheduling preferentially selects devices in the same province or the same major region where the resources are not fully occupied. If all resources are occupied, eligible devices in adjacent provinces / major regions are selected for load sharing.

[0090] This optional embodiment adopts a composite algorithm architecture. The main framework selects the Actor-Critic architecture based on deep reinforcement learning (DRL), such as DDPG / PPO / SAC, etc. The CNN-BiLSTM fusion structure is selected as the time series prediction component to handle the real-time traffic fluctuations in the network. The graph neural network is selected as the resource evaluation module to straighten out the topological relationship between the control plane and media plane devices. The lightweight Transformer is selected as the real-time decision-making module to handle instantaneous burst traffic. A topological relationship graph of device clusters is established from the spatial dimension, and a sliding window (30s / 1min / 5min / 15min / 30min / 1h) is established from the time dimension. Taking 10 layers as an example, the number of layers can be adjusted according to the actual network situation. The parameters are roughly divided into basic metrics, traffic characteristics, environmental parameters, etc. The types and weights can be adjusted according to the different network situations of operators. The weight dynamic allocation system is based on the meta-controller architecture of double-layer LSTM (Long Short-Term Memory). The critical thresholds of each metric are predicted through the first layer of LSTM, and the feature weight allocation strategy is generated through the second layer of GRU (Gated Recurrent Unit). The policy verification experiment is executed every hour, and the global parameters are optimized every week to adjust the parameter weights. For example: the basic metrics include CPU utilization rate (weight 0.15), GPU video memory occupancy (0.18), cross-node latency (0.22), etc.; the traffic characteristics include TCP retransmission rate (0.12), UDP jitter (0.08), QoE mapping parameter (0.15), etc.; the environmental parameters include temperature sensor data (0.05), power supply fluctuation (0.03), link optical attenuation (0.02), etc. Differentiable Neural Architecture Search (DNAS) is used to automatically screen features, and the importance of features is guided by the reward function, and the feature redundancy is controlled by the gradient penalty term, so as to adjust the dynamic feature weights. The objective function adopts the method of nesting multiple loss functions. The core objective layer adopts the min(1 - 99.99% SLA compliance rate × log(resource utilization rate)) function; the constraint condition layer is set is the bandwidth overrun penalty term, the threshold-triggered exponential penalty is used as the signaling storm suppression term, and the PUE²×TCO coefficient is the energy consumption sensitive term; the regularization layer uses the adjacent node resource difference regularization as the topology awareness constraint, and the scheduling policy mutation penalty is used as the time continuity constraint. The multi-objective Bayesian optimization is used to establish the Pareto surface, and a dynamic weight allocation strategy is established. Based on the context-aware weight of the current network state, the target priority is guided by short-term load prediction. The reward plasticity is generated by reverse engineering according to the actual SLA achievement, the adversarial samples are generated by simulating the network storm scenario, and the container is migrated through knowledge transfer across service scenarios. By performing the above operations, the dynamic scheduling of the media plane network device resources is realized through the AI model update unit of the control plane.

[0091] Figure 3 Shows the connection relationship between the control plane device and the media plane device in the present invention, reflecting the core role of the AI model update unit in the control plane device in network traffic control. Through the setting of the primary and standby control plane devices, the high availability of the system is ensured; through the intelligent analysis and decision-making of the AI model update unit, the dynamic regulation of the media plane device traffic is realized, the utilization of network resources is optimized, and the network performance and user experience are improved.

[0092] Combined with Figure 4 As shown, the present invention also provides a network traffic regulation method, including: Obtain the historical traffic data of all media plane network elements connected to the control plane network element and the topological relationship of the media plane network element; Update the traffic regulation model of the control plane network element according to the historical traffic data and the topological relationship, and the traffic regulation model includes a deep reinforcement learning sub-model, a fusion sub-model, a graph neural network sub-model and a Transformer sub-model; Input the current traffic data of the current time period into the updated traffic regulation model. Through the fusion sub-model, the time series prediction is carried out to obtain the time series prediction result of the current traffic data; through the graph neural network sub-model, the topological analysis is carried out to obtain the graph structure of the media plane network element; through the Transformer sub-model, the weight allocation is carried out to obtain the allocation weight strategy of each media plane network element; then through the deep reinforcement learning sub-model, the reinforcement learning is carried out according to the time series prediction result, the graph structure and the allocation weight strategy to obtain the traffic allocation strategy of the control plane network element; Dynamically adjust the media traffic of each media plane network element according to the traffic allocation strategy.

[0093] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned network traffic regulation method is realized.

[0094] A computer-readable storage medium provided by an embodiment of the present invention, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned network traffic regulation method is implemented.

[0095] Or, a non-volatile computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the processor performs the following operations: Obtain historical traffic data of all media plane network elements connected to the control plane network element and the topological relationship of the media plane network elements; Update the traffic regulation model of the control plane network element according to the historical traffic data and the topological relationship. The traffic regulation model includes a deep reinforcement learning sub-model, a fusion sub-model, a graph neural network sub-model, and a Transformer sub-model; Input the current traffic data of the current time period into the updated traffic regulation model. Perform time series prediction through the fusion sub-model to obtain the time series prediction result of the current traffic data; perform topological analysis through the graph neural network sub-model to obtain the graph structure of the media plane network element; perform weight allocation through the Transformer sub-model to obtain the allocation weight strategy of each media plane network element; then perform reinforcement learning through the deep reinforcement learning sub-model according to the time series prediction result, the graph structure, and the allocation weight strategy to obtain the traffic allocation strategy of the control plane network element; Dynamically adjust the media traffic of each media plane network element according to the traffic allocation strategy.

[0096] The computer-readable storage medium of the present invention has the same advantages as the above-mentioned network traffic regulation method compared with the prior art, and will not be elaborated here.

[0097] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will all fall within the protection scope of the present invention.

Claims

1. A network traffic regulation system, characterized in that, Including: An information collection unit, configured to obtain historical traffic data of all media plane network elements connected to a control plane network element and the topological relationship of the media plane network elements; An AI model update unit, configured to update a traffic regulation model of the control plane network element according to the historical traffic data and the topological relationship, where the traffic regulation model includes a deep reinforcement learning sub-model, a fusion sub-model, a graph neural network sub-model, and a Transformer sub-model; A policy generation unit, configured to input current traffic data of a current time period into the updated traffic regulation model, perform time series prediction through the fusion sub-model to obtain a time series prediction result of the current traffic data; perform topological analysis through the graph neural network sub-model to obtain a graph structure of the media plane network elements; perform weight allocation through the Transformer sub-model to obtain an allocation weight policy for each of the media plane network elements; Then, perform reinforcement learning through the deep reinforcement learning sub-model according to the time series prediction result, the graph structure, and the allocation weight policy to obtain a traffic allocation policy of the control plane network element; A policy execution unit, configured to dynamically adjust media traffic of each of the media plane network elements according to the traffic allocation policy.

2. The network traffic regulation system according to claim 1, wherein The information collection unit is specifically configured to: Obtain operation state data of all the media plane network elements within a preset duration from historical data according to a preset collection period, where the operation state data includes CPU occupancy and GPU occupancy; Determine network bandwidth occupancy data of the media plane network elements according to data traffic inflow and outflow information of network interfaces of the media plane network elements in the historical data; Obtain load data of the media plane network elements according to signaling throughput data of all the media plane network elements in the historical data; Use the operation state data, the network bandwidth occupancy data, and the load data as the historical traffic data.

3. The network traffic regulation system according to claim 1, characterized in that The AI model update unit is specifically configured to: Input the historical traffic data and the topological relationship into the traffic regulation model for training; Verify a traffic allocation policy output by the traffic regulation model during training through a digital twin environment and A / B testing; When the traffic allocation policies output by the traffic regulation model during training all meet the test requirements of the digital twin environment and the A / B testing, stop training the traffic regulation model to obtain the updated traffic regulation model.

4. The network traffic regulation system according to claim 2, wherein The policy generation unit is specifically configured to: Obtain the operation state data, the network bandwidth occupancy data, and the load data of the media plane network elements within the current time period, and use the operation state data, the network bandwidth occupancy data, and the load data as the current traffic data to input into the updated traffic regulation model; Extract features of the current traffic data through a CNN network of the fusion sub-model to obtain spatial features of the current traffic data; Extract features of the current traffic data according to a BiLSTM network of the fusion sub-model to obtain a time series of the current traffic data; Based on the spatial characteristics and the time series, obtain the traffic change trend of the media plane network element within a preset future time period, and use the traffic change trend as the time series prediction result.

5. The network traffic regulation system according to claim 2, wherein The policy generation unit is specifically configured to: Input the topological relationship of the media plane network element within the current time period into the updated traffic control model; Through the graph neural network sub-model, construct the graph structure according to the topological relationship, where the nodes of the graph structure are the media plane network elements, and the edges of the graph structure are the connection relationships of the media plane network elements.

6. The network traffic regulation system according to claim 2, wherein The policy generation unit is specifically configured to: Obtain the operation state data, network bandwidth occupancy data, and load data of the media plane network element within the current time period, and use the operation state data, network bandwidth occupancy data, and load data as the current traffic data to input into the updated traffic control model; Through the Transformer sub-model, perform prediction according to the current traffic data to obtain the critical values of multiple preset metrics of each media plane network element within a preset future time period; Generate the weight values of each preset metric according to the critical values to obtain the allocation weight policy of each media plane network element.

7. The network traffic regulation system according to claim 3, wherein The policy generation unit is specifically configured to: Through the deep reinforcement learning sub-model, perform reinforcement learning according to the time series prediction result, the graph structure, and the allocation weight policy; Among them, according to the time series prediction result, the graph structure, and the allocation weight policy, generate an initial traffic allocation adjustment policy; Use the initial traffic allocation adjustment policy as the action, and use the digital twin environment as the environment, and calculate the reward value after the interaction between the action and the environment through the reward function; Judge whether the reinforcement learning of the initial traffic allocation adjustment policy is completed according to the reward value. When the reinforcement learning of the initial traffic allocation adjustment policy is completed, obtain the traffic allocation policy of the control plane network element.

8. The network traffic regulation system according to claim 1, wherein It further includes a warning unit, and the warning unit is used for: Obtain the traffic change trend within the current time period according to the current traffic data of the current time period; Judge whether there is a sudden increase in traffic of the media plane network element according to the traffic change trend; When there is a sudden increase in traffic of the media plane network element, send a warning message.

9. A network traffic regulation method, characterized in that, It includes: Obtain the historical traffic data of all media plane network elements connected to the control plane network element and the topological relationship of the media plane network element; Update the traffic control model of the control plane network element according to the historical traffic data and the topological relationship. The traffic control model includes a deep reinforcement learning sub-model, a fusion sub-model, a graph neural network sub-model, and a Transformer sub-model; Input the current traffic data of the current time period into the updated traffic regulation model, perform time series prediction through the fusion sub-model to obtain the time series prediction result of the current traffic data; perform topological analysis through the graph neural network sub-model to obtain the graph structure of the media plane network element; perform weight allocation through the Transformer sub-model to obtain the allocation weight strategy for each media plane network element; then perform reinforcement learning through the deep reinforcement learning sub-model according to the time series prediction result, the graph structure, and the allocation weight strategy to obtain the traffic allocation strategy of the control plane network element; Dynamically adjust the media traffic of each media plane network element according to the traffic allocation strategy.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the storage medium, and when the computer program is executed by a processor, the network traffic regulation method described in claim 9 is implemented.

Images