A network optimization system

Abstract

The application discloses a network optimization system, comprising an endogenous AI module arranged in a network node and a comprehensive AI module arranged outside the network node; the endogenous AI module acquires node data of the corresponding network node, performs self-learning on the acquired node data through a pre-trained endogenous AI model, generates a local data set, transmits the local data set to the comprehensive AI module, and optimizes and adjusts the network node according to the feedback of the comprehensive AI module; the comprehensive AI module analyzes the acquired local data set through a pre-trained comprehensive AI model, generates an optimization strategy, and feeds back the optimization strategy to the endogenous AI module. The application scheme can avoid the problems of long data closed loop cycle, high data transmission cost and local optimal solution during network optimization, and improve the network optimization effect.

Description

Technical Field

[0001] This invention relates to the field of transmission and bearer technology, and in particular to a network optimization system. Background Technology

[0002] In the 5G era, AI models will adopt an add-on structure, integrating AI as an additional component into the existing network architecture. The RAN, bearer network, and core network will independently introduce AI based on their own needs.

[0003] While add-on AI applications can provide some network optimization capabilities, they often suffer from problems such as long data loop cycles, high data transmission costs, and local optima. Furthermore, the 6G core network introduces distributed micro-cloud units (SCUs), where base station service data flows in multiple directions, including to the core central node and other SCU nodes, and there is also data flow between SCU nodes. In this context, the local optima problems inherent in traditional add-on AI models may be further amplified, leading to poor performance of existing AI optimization methods. Summary of the Invention

[0004] The purpose of this invention is to provide a network optimization system that avoids the problems of long data loop cycles, high data transmission costs, and local optima during network optimization, thereby improving the network optimization effect.

[0005] To achieve the above objectives, embodiments of the present invention provide a network optimization system, the system comprising: an endogenous AI module configured inside a network node, and a comprehensive AI module configured outside the network node;

[0006] The endogenous AI module acquires node data of the corresponding network node, performs self-learning on the acquired node data through a pre-trained endogenous AI model, generates a local dataset, transmits the local dataset to the integrated AI module, and optimizes and adjusts the network node based on the feedback from the integrated AI module.

[0007] The integrated AI module analyzes the acquired local dataset using a pre-trained integrated AI model, generates optimization strategies, and feeds these optimization strategies back to the endogenous AI module.

[0008] Preferably, the endogenous AI module converts the local dataset into a preset protocol format and sends the converted local dataset to the endogenous integrated AI module;

[0009] The local dataset includes network optimization schemes, model performance evaluation results, and network device performance logs.

[0010] The network optimization scheme includes routing optimization strategies, dynamic resource allocation strategies, and fault prevention measures.

[0011] Preferably, the protocol format is the WebSocket protocol;

[0012] The endogenous AI module sends the network optimization scheme and model performance evaluation results to the integrated AI module through a connection established via the WebSocket protocol.

[0013] The intrinsic AI module transmits the network device performance logs to the integrated AI module via the uplink of the network node.

[0014] Preferably, the endogenous AI module obtains data from corresponding network nodes using different protocols through an intelligent data interface, and integrates the obtained data to generate node data in a unified format.

[0015] Preferably, the intelligent data interface sends configuration data and control commands through any one of the NetConf protocol, PCEP protocol, BGP-LS protocol, and Telemetry protocol;

[0016] The intelligent data interface uses the SFTP protocol to call the AI ​​model API interface and distribute AI model strategies.

[0017] The intelligent data interface uses the SFTP protocol for AI model loading and version control.

[0018] The intelligent data interface collects and reports large volumes of data via the SFTP protocol.

[0019] The intelligent data interface collects and reports the configuration processing results through the Netconf protocol.

[0020] Preferably, the endogenous AI module performs data preprocessing and data format conversion on the collected node data to obtain MCP protocol node data for learning the endogenous AI model.

[0021] Preferably, the endogenous AI module trains the AI ​​model using the acquired training data to obtain prediction results, and then compares the prediction results with the pre-labeled tags of the training data to optimize the AI ​​model and obtain the endogenous AI model.

[0022] Preferably, the network node includes a core network and / or a bearer network and / or a wireless network.

[0023] Preferably, when the network node includes a wireless access network, the node data acquired by the endogenous AI module includes base station data, node device data, terminal data, channel condition data, user behavior data, and base station traffic data;

[0024] The intrinsic AI module uses network performance metrics, including packet loss rate and throughput, as well as QoS metrics, including bandwidth, latency, and signal-to-noise ratio, as training labels to predict and analyze network performance and service quality, and to train the intrinsic AI model.

[0025] Preferably, when the network node includes a bearer network, the node data acquired by the endogenous AI module includes monitoring data of key equipment;

[0026] The key equipment includes routers and / or switches;

[0027] The monitoring data includes network topology information composed of network traffic, device status information, and link quality information, as well as network element information;

[0028] The endogenous AI module extracts key features from the node data to perform self-learning of the AI ​​model and outputs a network optimization scheme.

[0029] The key features include traffic patterns, link load, link latency, and packet loss rate.

[0030] Preferably, when the network node includes the core network, the node data acquired by the endogenous AI module includes user session data, network slice status, resource utilization, and fault records;

[0031] The endogenous AI module extracts features from the acquired training data, identifies key features that affect the core network performance, and uses network performance indicators and QoS indicators as labels to train the AI ​​model, thus obtaining the endogenous AI model.

[0032] The key characteristics include peak traffic periods, resource bottlenecks, and potential points of failure in the link.

[0033] Preferably, the integrated AI model is a deep neural network or a centralized endogenous AI model;

[0034] The integrated AI model includes the endogenous AI models of each network node.

[0035] Preferably, the integrated AI module uses an integrated AI model to perform comprehensive analysis on the acquired local dataset, identify cross-network patterns and development trends; assign weights to the collected data, prioritize and allocate resources to the optimization schemes of different network parts, call each sub-model or perform model pruning and other operations to adjust the local optimization strategies of network nodes, generate optimization strategies, and feed back the verified optimization strategies to the corresponding endogenous AI module.

[0036] Preferably, the endogenous AI module monitors the network performance indicators after the network nodes have been optimized and adjusted, and feeds back the monitored network performance indicators to the integrated AI module.

[0037] Preferably, after the endogenous AI module identifies an alarm event generated by a network node, it identifies the feature dataset of the fault through the configured endogenous AI model and performs self-repair of the alarm fault.

[0038] When self-repair fails, the feature dataset is uploaded to the integrated AI module.

[0039] Preferably, the endogenous AI module uses a configured endogenous AI model to identify the fault type based on the alarm information and generates an abnormal dataset according to time sorting;

[0040] The abnormal event chain is determined by a sliding window in the time dimension, and the feature sequence is obtained as the feature dataset. A self-repair scheme is generated based on the feature dataset.

[0041] Execute the self-healing scheme and determine whether the alarm event has been eliminated;

[0042] If the alarm event is not cleared, the self-repair is deemed to have failed.

[0043] Preferably, the integrated AI module uses the received feature dataset as input data for learning the integrated AI model to determine the fault log; and performs cluster analysis to trace the cause of the fault and output a fault repair plan.

[0044] The fault log includes the fault time, fault location, and fault cause.

[0045] Preferably, the system further includes a digital twin network;

[0046] The processor inputs the fault repair scheme output by the integrated AI model into the digital twin network;

[0047] The simulated network links constructed by the digital twin network are used for simulation verification; the fault repair plan after successful verification is fed back to each network node for execution.

[0048] Preferably, the digital twin network constructs a simulated network link by acquiring the resource configuration information of network nodes.

[0049] Preferably, the digital twin network obtains basic configuration information of network node devices through a management channel, obtains topology information related to the three-layer protocol through a control protocol channel, and obtains real-time performance data of the devices through a performance acquisition channel.

[0050] The device information includes interface information, device manufacturer information, network element type, and version signal;

[0051] The topology information includes link information, latency, interface bandwidth, and tags;

[0052] The real-time performance data includes bandwidth, latency, packet loss rate, and occupancy.

[0053] Preferably, the integrated AI module acquires data from each network node through an integrated AI model, performs correlation learning, splices and merges the feature information of each part to form a feature dataset of the link, and predicts the traffic of the link based on the feature dataset.

[0054] Compared to existing technologies, this invention discloses a network optimization system comprising an endogenous AI module configured inside network nodes and a comprehensive AI module configured outside the network nodes. The endogenous AI module acquires node data from the corresponding network nodes, performs self-learning on the acquired node data using a pre-trained endogenous AI model, generates a local dataset, transmits the local dataset to the comprehensive AI module, and optimizes and adjusts the network nodes based on feedback from the comprehensive AI module. The comprehensive AI module analyzes the acquired local dataset using a pre-trained comprehensive AI model, generates an optimization strategy, and feeds the optimization strategy back to the endogenous AI module. This solution avoids the problems of long data loop cycles, high data transmission costs, and local optima in network optimization, thereby improving network optimization performance. Attached Figure Description

[0055] Figure 1 This is a structural block diagram of the network optimization system provided in the embodiments of the present invention;

[0056] Figure 2 This is a schematic diagram of the network optimization system provided in an embodiment of the present invention;

[0057] Figure 3 This is another schematic diagram of the network optimization system provided in this embodiment of the invention;

[0058] Figure 4 This is a schematic diagram illustrating the principle of the network optimization system provided in this embodiment of the invention when performing network optimization;

[0059] Figure 5 This is a schematic diagram of the structure of a deep neural network provided in an embodiment of the present invention;

[0060] Figure 6 This is a schematic diagram of the structure of the centralized endogenous AI model provided in an embodiment of the present invention;

[0061] Figure 7 This is another schematic diagram of the network optimization system provided in this embodiment of the invention;

[0062] Figure 8 This is a schematic diagram of the framework of the digital twin network provided in an embodiment of the present invention. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] To address the aforementioned technical problems, this application proposes a network optimization system, which includes: an endogenous AI module configured inside a network node, and a comprehensive AI module configured outside the network node;

[0065] The endogenous AI module acquires node data of the corresponding network node, performs self-learning on the acquired node data through a pre-trained endogenous AI model, generates a local dataset, transmits the local dataset to the integrated AI module, and optimizes and adjusts the network node based on the feedback from the integrated AI module.

[0066] The integrated AI module analyzes the acquired local dataset using a pre-trained integrated AI model, generates optimization strategies, and feeds these optimization strategies back to the endogenous AI module.

[0067] In the specific implementation of this embodiment, please refer to Figure 1 This is a structural block diagram of a network optimization system provided in an embodiment of the present invention. The system includes an endogenous AI module configured inside a network node and a comprehensive AI module configured outside the network node.

[0068] It should be noted that the network nodes configured for the endogenous AI module provided in this application are various network nodes in the wireless communication network, which may specifically include network nodes such as wireless bearer network and core network.

[0069] It should be noted that, see Figure 2This is a schematic diagram illustrating the principle of the network optimization system provided in this embodiment of the invention. This application proposes a 6G end-to-end AI intrinsic collaborative large-scale model scheme, referring to a complete system used in 6G networks, including intrinsic AI small models in the wireless network, bearer network, and core network, as well as an external comprehensive AI model. Specifically, the AI ​​module loads a deep learning processor unit for algorithm acceleration and can include various AI chips, such as ARM chips, GPUs, FPGAs, and ASICs, while configuring a computing engine that supports most neural network architectures, including VGG, ResNet, YOLO, SSD, MobileNet, and FPN. The AI ​​module deploys and runs various AI models to achieve real-time inference acceleration, meeting the computing power requirements of various applications. The models undergo compression techniques such as distillation, pruning, and quantization, significantly reducing the hardware performance requirements of the model without substantially affecting the judgment results.

[0070] The endogenous AI design proposed in this application not only integrates the AI ​​module into the device architecture, providing the industry with a unified standard framework and facilitating end-to-end collaboration, but also provides a complete operating environment for the entire lifecycle of AI workflow, including data acquisition, data preprocessing, model training, model inference, and model evaluation, for 6G networks or devices themselves. Compared with traditional add-on AI, it achieves more efficient, real-time, and intelligent AI services.

[0071] The integrated AI model is equivalent to a central processing unit, which is a complex neural network structure. The endogenous AI mini-models of the wireless network, bearer network, and core network are derived from this integrated AI model through targeted model cutting or sub-model calling. Therefore, there is a certain correlation between the three AI mini-models, while they are also subsets of the integrated AI model. They can provide global optimization schemes for the integrated AI model and accept local optimization strategies from the integrated model.

[0072] In the target network environment, identify the network nodes that need optimization. These nodes can be devices with data processing and transmission capabilities, such as servers, routers, and switches. Equip each node with a built-in AI module, ensuring seamless integration between the module and the node's operating system and data acquisition interfaces to acquire various data during node operation in real time, such as CPU utilization, memory usage, network traffic, packet latency, and error rate.

[0073] The intrinsic AI module continuously monitors the operational status of network nodes, collecting various types of node data at a predetermined data collection frequency (e.g., once per second). It utilizes a pre-trained intrinsic AI model to learn from the pre-processed data. The model architecture can employ a recurrent neural network (RNN) or its variant, a long short-term memory network (LSTM), to capture the dynamic changes in node data over time. The model input is a sequence of node data within a time window, such as CPU usage and memory consumption data collected every 5 minutes over the past hour. After learning and analysis, the model generates a local dataset reflecting the current operational status, potential problems, and trends of the node. This dataset not only includes the statistical characteristics of the raw data, such as mean, standard deviation, maximum, and minimum values, but may also include indicators predicted by the model, such as whether CPU usage is likely to exceed a threshold within the next 10 minutes.

[0074] The generated local dataset is sent to the integrated AI module according to a predetermined data transmission protocol. Simultaneously, a thread or interface is set up to receive feedback, awaiting the optimization strategy returned by the integrated AI module. Upon receiving the optimization strategy, such as instructions to adjust resource allocation parameters for a service or change network configuration, the intrinsic AI module immediately initiates optimization adjustments to the network nodes. This may involve calling the node's operating system API or configuration management tools to implement specific adjustments, such as adjusting process priorities by modifying the "nice" value in a Linux system, thereby optimizing CPU resource allocation.

[0075] The integrated AI module continuously monitors data transmission requests from various endogenous AI modules on the central server. Upon receiving a local dataset, it categorizes and stores it according to network node identifiers. The datasets from different nodes are further integrated to construct a global data view, analyzing the correlations and mutual influences between different nodes. For example, if multiple nodes in a certain region simultaneously experience peak network traffic, it may indicate a source of network congestion or a concentration of business requests.

[0076] A pre-trained comprehensive AI model is used to perform in-depth analysis of the integrated global data. This comprehensive AI model can employ a graph neural network (GNN) architecture, treating network nodes as nodes in a graph and the connections between nodes as edges, thereby learning the overall network topology and traffic propagation patterns. Through model inference and computation, optimization strategies are generated for each network node. These strategies may include adjusting network routing tables to balance traffic, dynamically allocating server cluster resources, and starting or stopping certain non-critical services. The generation of optimization strategies must consider improving overall network performance while minimizing the impact on normal business operations, such as prioritizing the bandwidth requirements of critical services.

[0077] The generated optimization strategies are fed back to each intrinsic AI module according to their corresponding relationships. Simultaneously, the integrated AI module continuously collects feedback on network performance metrics, such as whether overall network latency has decreased and throughput has increased, and evaluates the effectiveness of the optimization strategies based on this feedback. If some strategies are found to be unsatisfactory, the integrated AI module can further adjust model parameters or retrain the model to continuously improve its optimization capabilities.

[0078] It's important to note that before the system's official launch, historical network data was used to train both the intrinsic AI model and the comprehensive AI model offline. Data on the normal operation of network nodes over a period of time, as well as data on instances of failures or performance bottlenecks, was collected to perform supervised learning on the model. The data was labeled with corresponding states (normal, abnormal, etc.) to enable the model to learn feature patterns under different operating conditions. Cross-validation and other techniques were employed to ensure the model's generalization ability and avoid overfitting. Simultaneously, network operations experts were invited to evaluate the model's training process and preliminary results. Based on expert feedback, the model structure and training parameters were adjusted, such as increasing the number of hidden layer neurons or adjusting the learning rate. Furthermore, during system operation, new network node data was continuously collected as incremental data. This incremental data was used periodically (e.g., weekly or monthly) to update the model online, adapting to dynamic changes in the network environment, such as the impact of new service launches or network equipment upgrades. During the update training process, incremental learning algorithms were used, adjusting parameters only for the parts of the model most affected by new data, avoiding retraining the entire model from scratch and improving training efficiency.

[0079] It should be noted that the endogenous AI mini-model is a partial model call or a model with deletions from the external AI comprehensive model.

[0080] For example, AI-inherent models are configured in the wireless network, bearer network, and core network.

[0081] Wireless Access Network (WLAN): The built-in AI model in the WLAN supports the collection of various network data through different protocols. It then analyzes and outputs AI data in a unified format through model self-training, forming proprietary AI dataset logs and global optimization strategy solutions. Simultaneously, it interacts with external integrated AI models, uploading AI data and global optimization solutions, downloading local optimization solutions, and then performing corresponding AI model inference and deployment, thus forming a complete lifecycle for WLAN AI.

[0082] Bearer Network: The intrinsic AI model in the bearer network supports the external integrated AI model and the digital twin network. By collecting physical information from devices in the network and pre-training the AI ​​model, it generates link logs for the bearer network, which are then uploaded to the external integrated AI model to assist in the construction and verification of the digital twin network. Simultaneously, it downloads and allocates link scheduling schemes to schedule service transmission links and manage overall network traffic.

[0083] Core Network: The endogenous AI model in the core network supports the collection of different types of network data using different protocols and transforms them into corresponding AI datasets. It also supports functions such as AI model inference deployment, log and scheme generation.

[0084] In addition to transmitting uplink and downlink information as usual, the three networks also transmit relevant information to the AI ​​model for data preprocessing, generating corresponding logs and global optimization strategies. Furthermore, there is an overall external integrated AI model that analyzes the AI ​​data after receiving the feature datasets and optimization strategies output by the three networks. Through the model adjustment and invocation described in section 1, a specific functional model is formed for training, generating optimization strategies and evaluating their effectiveness. At the same time, relevant verification is carried out in the digital twin network, ultimately providing guidance for local optimization of the three networks and end-to-end orchestration.

[0085] Through an endogenous AI module, AI capabilities are empowered throughout the 6G network and on terminals. All network nodes and terminals possess AI training and learning capabilities, achieving a high-performance, high-efficiency, generalizable, and universal 6G endogenous AI architecture. This enables network management, analysis and decision-making, coordination and scheduling functions, ensuring the transmission quality of the 6G network.

[0086] In another embodiment of the present invention, when transmitting a local dataset, the local dataset output by the endogenous AI model includes network optimization schemes, model performance evaluations, network device performance logs, etc.

[0087] The network optimization scheme includes routing optimization strategies, dynamic resource allocation strategies, and fault prevention measures.

[0088] Once the endogenous AI module completes self-learning of node data and generates a local dataset containing network optimization schemes, model performance evaluation results, and network device performance logs, it first sorts out these different types of data.

[0089] For routing optimization strategies in the network optimization scheme, information such as the source node, destination node, current routing path, and suggested modified routing path is compiled. The dynamic resource allocation strategy clarifies the current resource allocation (e.g., the proportion of CPU, memory, and bandwidth among various processes or services) and the adjustment direction and magnitude predicted by the learning model. Fault prevention measures detail potential fault points identified based on model analysis, possible fault types, and corresponding preventative steps. Model performance evaluation results cover quantitative indicators such as the model's accuracy, recall (e.g., the accuracy of predicting fault occurrence and the proportion of actual faults captured by the model), and mean squared error for predicting current node data, used to measure the effectiveness of the intrinsic AI model itself. Network device performance logs include device hardware status (e.g., temperature, fan speed), system operating parameters (e.g., power-on time, average system load), and network connection-related data (e.g., number of packets sent and received at ports, packet loss rate), organized in time series. Focusing on the network architecture level, the intrinsic AI model meticulously designs routing optimization strategies. Through detailed analysis of key elements such as network traffic flow and node load balancing, it cleverly plans the optimal path to ensure that data packets arrive at their destination with the fastest speed and lowest latency, effectively avoiding network congestion and greatly improving transmission efficiency. Simultaneously, in terms of resource management, it can keenly perceive the dynamics of resource supply and demand, dynamically allocating core resources such as CPU, memory, and bandwidth in real time to ensure sufficient support for critical business processes, eliminate resource idleness and waste, and maximize resource utilization. Furthermore, the intrinsic AI model also possesses a forward-looking fault prevention mindset. Based on deep learning of network device operating status and historical fault data, it identifies potential fault risk points in advance and tailors detailed and feasible fault prevention measures, such as early warning of abnormal temperatures in critical components and prediction of network interface speed attenuation, preventing problems before they occur. The intrinsic AI model also generates a comprehensive and detailed model performance evaluation report, which includes key quantitative indicators such as accuracy, recall, and F1 score, intuitively demonstrating its performance in network optimization tasks. The synchronously output network device performance log records the hardware health status of the device, including temperature, fan speed, power stability, and system operating parameters such as boot time, average load, and process resource usage. It also includes network connection-related details such as the number of data packets sent and received by the port, packet loss rate, and latency statistics, providing data for network operation and maintenance.

[0090] Based on the preset protocol format specifications, the integrated local dataset is converted.

[0091] The endogenous AI module converts the local dataset into a preset protocol format and sends the converted local dataset to the integrated AI module.

[0092] In another embodiment of the present invention, in the practical scenario of network optimization, the endogenous AI model, with its powerful computing capabilities and intelligent algorithms, deeply analyzes real-time network data and accurately outputs a series of highly targeted optimization solutions.

[0093] Data transmission between the device and the AI ​​mini-model is primarily achieved through the intelligent data interface described in section 1. AI mini-model data transmission mainly involves preprocessing the data, then encapsulating the output AI dataset into a format supported by the WebSocket protocol, and transmitting it via the WebSocket connection over the link. Interaction between the AI ​​mini-model and the external integrated AI model utilizes the WebSocket protocol to transmit data to the location of the external integrated AI model via the internet or a dedicated network for real-time data transmission. Simultaneously, by integrating the MCP protocol, a unified interface is provided for the external integrated AI model, supporting seamless integration between the external integrated AI model and external data sources.

[0094] To achieve efficient collaboration and rapid decision-making, the aforementioned optimization schemes, performance evaluation results, and network device performance logs were all carefully packaged into a specific format compatible with the WebSocket protocol. With the help of a pre-established stable WebSocket connection, this data can be transmitted to the integrated AI module in real time and without interruption.

[0095] Specifically, the network optimization scheme and model performance evaluation results are sent to the integrated AI model platform in real time via the established WebSocket connection for secondary confirmation of the analysis results, global analysis, and strategy adjustment.

[0096] The intrinsic AI module transmits the network device performance logs to the integrated AI module via the uplink of the network node.

[0097] The integrated AI module uses AI algorithms to reconfirm the optimization plan, taking into account the complex interactions of all nodes and links across the network. It makes precise adjustments to routing optimization and resource allocation strategies to ensure the scientific validity and feasibility of the plan. At the same time, network device performance logs will also take on a new role, laying a solid foundation for long-term monitoring, in-depth analysis, and accurate prediction, helping the operations and maintenance team to anticipate network risks and cope with various complex and ever-changing network conditions.

[0098] The endogenous AI model has excellent real-time response capabilities. In the face of sudden emergencies such as network traffic surges and equipment failures, it can adjust and optimize strategies in a lightning-fast manner. In the ever-changing network environment, it safeguards the efficient and stable operation of the core network with its intelligence and speed, and continuously empowers the network to upgrade and expand.

[0099] In another embodiment of the present invention, the endogenous AI module obtains data from the corresponding network nodes through intelligent data interfaces using different protocols, and integrates the obtained data to generate node data in a unified format.

[0100] The end-to-end endogenous AI big model solution also includes intelligent interfaces, which are the interfaces between AI models and devices. The intelligent interfaces include both functions and protocols.

[0101] Specifically, the functions of the intelligent interface are as follows:

[0102] It supports the ability to obtain information through general interfaces, including the issuance of existing configuration data and control commands, as well as the acquisition or reporting of process data such as configuration, alarms / events, performance, and status; it supports the acquisition or reporting of topology link information and status data; and it supports the issuance, acquisition, or reporting of tunnel configurations.

[0103] Supports calling AI model API interfaces and distributing AI model strategy or rule data;

[0104] It supports AI model loading and version control. AI model loading includes uploading, storing, and downloading different versions of AI model file packages, maintaining sessions, maintaining the security of upload / download channels, maintaining data security, and providing feedback on results.

[0105] Supports the collection and reporting of large volumes of data, including the second / millisecond performance and status required by AI agents;

[0106] It supports the collection and reporting of intelligent configuration processing results such as performance prediction results, quality defect detection results, fault detection analysis results, and equipment energy-saving status.

[0107] The protocol for the intelligent interface is as follows:

[0108] General interface capabilities are implemented using existing technical solutions. Traditional functions such as configuration data and control commands, topology link information and status data reporting, tunnel configuration distribution / acquisition and reporting are implemented using existing NetConf, PCEP, BGP-LS and Telemetry interface protocols.

[0109] AI model strategy or rule data is distributed via SFTP or other interfaces / protocols.

[0110] AI model loading and version control are performed via SFTP or other interfaces / protocols.

[0111] The second / millisecond level performance and state data required by AI models are reported or obtained through interfaces such as Telemetry.

[0112] The intelligent configuration processing results, such as performance prediction results, quality defect detection results, fault detection analysis results, and equipment energy-saving status, are reported through the Netconf protocol, and the interface definition adopts the XML / YANG model.

[0113] In another embodiment of the present invention, after the endogenous AI module acquires node data, it needs to perform data preprocessing and data format conversion on the node data. The endogenous AI module first needs to collect network data information, and through the model optimization module, it uses the Model Context Protocol (MCP) to convert data from multiple different protocols into node data in a unified format for learning the endogenous AI model.

[0114] In another embodiment of the present invention, the endogenous AI module pre-trains the AI ​​model to determine the endogenous AI model, specifically:

[0115] Training data is obtained from corresponding network nodes. These data sources are diverse, covering network traffic data (such as the number of data packets and bandwidth usage at different times), device performance data (such as CPU utilization and memory usage), and network connection status data (such as packet loss rate and latency). Data acquisition can be achieved through intelligent data interfaces, communicating with network nodes via different protocols.

[0116] After preprocessing the data, the preprocessed data is divided into a training set, a validation set, and a test set. For example, the training set accounts for approximately 70%-80% and is used for training the model; the validation set accounts for approximately 10%-15% and is used to adjust the hyperparameters of the model during training; and the test set accounts for approximately 10%-15% and is used to evaluate the performance of the trained model.

[0117] Choose an appropriate AI model architecture based on the nature of the task and the characteristics of the data. Common model architectures include: Neural network models: such as Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) and its variants (such as LSTM, GRU). RNNs and their variants may be more suitable for processing network traffic data with time-series characteristics; CNNs may be more effective for analyzing graphical network topologies. Decision tree models: such as Random Forest, Gradient Boosting Tree (GBDT), etc. These models perform well in processing structured data and feature selection. Support Vector Machines (SVM): suitable for binary or multi-class classification problems, and perform well when processing high-dimensional data.

[0118] Define a loss function to measure the difference between the model's predictions and the true labels.

[0119] The model parameters are updated using the selected optimization algorithm to minimize the loss function.

[0120] The training data is input into the selected AI model, and the model's prediction results are calculated through forward propagation. Then, the loss value between the prediction results and the true labels is calculated based on the loss function. Next, the model parameters are updated using an optimization algorithm. This process is repeated until the loss function converges or a preset number of training epochs are reached, resulting in the intrinsic AI model.

[0121] In another embodiment of the present invention, the network node includes a core network and / or a bearer network and / or a wireless network.

[0122] Built-in AI modules are deployed at core network nodes, which typically contain critical routing and switching equipment, core servers, etc. Data is collected via protocols such as SNMP (Simple Network Management Protocol) and NetFlow, utilizing the devices' built-in management interfaces or specially developed intelligent data acquisition plugins. For example, routing table information, port traffic data, and IP packet header information are collected from core routers to analyze network traffic flow and potential bottlenecks; CPU utilization, memory usage, and process running status at the operating system level are obtained from core servers to ensure the stable operation of core services. To meet the high reliability requirements of the core network, redundant links and backup acquisition mechanisms are set up for data acquisition. If the primary acquisition link fails, the backup link is immediately activated to ensure uninterrupted data transmission and to send real-time alerts to maintenance personnel.

[0123] For transmission equipment in the bearer network, such as optical transmission nodes and microwave transmission nodes, built-in AI modules are also deployed and equipped with adapted data acquisition components. Performance metrics of the transmission link, such as bit error rate, optical power, and dispersion, are extracted through the OTN (Optical Transport Network) overhead bytes of the optical transmission network. Signal strength, frequency offset, and modulation scheme information are collected using the network management interface of the microwave equipment to evaluate the transmission quality and stability of the bearer network. Considering the wide coverage of the bearer network, a distributed data acquisition strategy is adopted, decomposing the acquisition tasks to aggregation nodes in various regions, and then aggregating them to the central built-in AI module, reducing the risk of latency and data loss caused by long-distance transmission.

[0124] At wireless network nodes such as base stations, built-in AI modules connect to the base station equipment's internal APIs or open network management interfaces to collect key data such as wireless signal strength, channel quality, number of users accessing the network, and data rate. For example, by utilizing the Measurement Reporting (MR) function in LTE (Long Term Evolution) or 5G networks, real-time wireless experience data of users within the cell can be obtained, providing a basis for optimizing wireless network coverage and capacity. Since wireless networks are significantly affected by the environment, Geographic Information System (GIS) data is combined to correlate base station location, terrain, surrounding buildings, and other information with the collected wireless data, enabling more accurate analysis of issues such as signal attenuation and interference.

[0125] In yet another embodiment provided by the present invention, see Figure 3 This is another schematic diagram illustrating the principle of the network optimization system provided in this embodiment of the invention. Targeting the radio, transport, and core networks in 6G networks, it employs distributed collaborative learning, performing targeted training for different services to achieve global optimization of the external comprehensive AI model and local optimization of each network's smaller models, supporting various 6G service applications. The specific workflow of its AI model is shown in the figure, and the endogenous AI technology solution on the RAN side is also shown.

[0126] See Figure 4 This is a schematic diagram illustrating the principle of network optimization in the network optimization system provided in this embodiment of the invention. When the network node includes a wireless access network, in the wireless access network in the data uplink, the AI ​​module in the intrinsic AI transmission device first needs to collect network data information. Through the model optimization module, the Model Context Protocol (MCP) is used to convert various protocol data (base station, node device, terminal information, channel condition data, user behavior data, base station traffic data, etc.) for collection. The collected data undergoes preprocessing steps such as cleaning, standardization, and normalization to adapt to the model's input requirements. Then, the formed dataset is used to train the model, allowing the model to learn the patterns and rules in the data. In this process, network performance indicators (such as packet loss rate and throughput) and QoS indicators (such as bandwidth, latency, and signal-to-noise ratio) are used as the label set for model self-training, guiding the model to learn and perform self-training of the AI ​​mini-model, predicting and analyzing network performance and service quality. The output results include network optimization schemes, model performance evaluations, network device performance logs, etc. In addition to transmitting data to the bearer network, the wireless access network also encapsulates the dataset generated by the AI ​​model into a format suitable for transmission via the WebSocket protocol and sends it through a WebSocket connection. The network optimization scheme and model performance evaluation are sent to the integrated AI model via the Internet or a private network, while the network device performance logs are sent to the bearer network via the uplink.

[0127] In another embodiment of the present invention, when the network node includes a bearer network, the endogenous AI module is deployed at key network nodes, such as routers and switches, to monitor and collect network topology information and network element information, including network traffic, device status, and link quality, in real time. Using the Model Context Protocol (MCP), the AI ​​mini-model uniformly transforms the collected data from various protocols (such as BGP, OSPF, MPLS, etc.) for further processing. Preprocessing operations such as cleaning, denoising, standardization, and normalization are performed on the collected raw data and network device performance logs transmitted in the wireless access network to ensure data quality and meet the requirements of model input. Key features, including traffic patterns, link load, latency, and packet loss rate, are extracted from the preprocessed data to prepare input for model training. The AI ​​model is trained using the processed dataset. By learning patterns and rules in the data, the model can predict network performance trends and potential fault points. The model's prediction results are compared with actual network performance data to evaluate the model's accuracy and reliability, and a model optimization strategy is generated accordingly. Meanwhile, the AI ​​mini-model outputs network optimization solutions based on training results, such as routing optimization, load balancing, and fault prevention measures, to improve the overall performance and stability of the bearer network. It records network device performance logs, including device operating status, performance indicators, and any abnormal events, providing data support for subsequent analysis and troubleshooting. The network optimization solutions, model performance evaluation results, and network device performance logs generated by the AI ​​mini-model are encapsulated in a format suitable for WebSocket protocol transmission. Through a WebSocket connection, the optimization solutions and performance evaluation results are sent to the integrated AI model platform for more in-depth analysis and global optimization, while the fault and abnormal event chain is sent to the integrated AI model for storage. Simultaneously, datasets such as network device performance and fault and abnormal event chains are sent to the core network via the uplink. The AI ​​mini-model in the bearer network can respond to network changes in real time and quickly adjust optimization strategies to cope with sudden events such as network traffic surges and device failures.

[0128] In another embodiment of the present invention, when the network node includes a bearer network, it differs from the access network and bearer network described above. In the core network, the AI ​​mini-model's workflow primarily focuses on achieving high-speed data transmission, accurate user session management, and service quality optimization. The AI ​​mini-model first collects key network performance indicators through a model optimization module, including user session data, network slice status, resource utilization, and fault records. Using the Model Context Protocol (MCP), the AI ​​mini-model uniformly transforms and standardizes this data from different sources and protocols, and aggregates it with logs transmitted from the bearer network for further analysis. Next, the AI ​​mini-model performs in-depth analysis and feature extraction on the collected data, identifying key factors affecting core network performance, such as peak traffic periods, resource bottlenecks, and potential link failure points. Based on these features, the AI ​​mini-model uses machine learning algorithms for training, learning complex patterns and trends in network operation to predict future network behavior and potential problems. During training, the AI ​​mini-model uses network performance indicators and QoS indicators as labels to optimize the model and improve prediction accuracy. After training, the AI ​​mini-model outputs a series of optimization solutions, including routing optimization, dynamic resource allocation, and fault prevention measures, as well as model performance evaluation reports and network device performance logs. These optimization solutions and performance evaluation results are encapsulated in a format suitable for the WebSocket protocol and sent in real time to the integrated AI model platform via an established WebSocket connection for secondary confirmation of analysis results, global analysis, and strategy adjustment. Simultaneously, network device performance logs are also sent to the integrated AI model in the same way for subsequent monitoring, analysis, and prediction functions. The core network AI mini-model can respond to network changes in real time and quickly adjust optimization strategies to cope with sudden events such as network traffic surges and device failures, ensuring the efficient and stable operation of the core network.

[0129] In another embodiment of the present invention, the external integrated AI model has a large number of parameters, making it suitable for scenarios with large model scale, complex training datasets, and cross-domain data (wireless network, bearer network, core network), including scenarios containing user-related data. The AI ​​mini-models within each of the wireless, bearer network, and core network domains are limited in size and are mainly used for tasks with low latency requirements or low computational demands. Other complex tasks can be solved through collaborative work between different AI mini-models. When a small AI model struggles to solve a problem or is overloaded, it is uploaded to the external integrated AI model, where global resource orchestration is used to achieve the desired result.

[0130] The integrated AI model of the integrated AI module is either a deep neural network or a centralized endogenous AI model;

[0131] The integrated AI model includes the endogenous AI models of each network node.

[0132] See Figure 5This is a schematic diagram of the deep neural network structure provided in an embodiment of the present invention. The external comprehensive AI model composed of deep neural networks has hundreds of millions of layers and neurons. When dealing with different networks and different services, some neurons in the deep neural network are masked or simplified through pruning, fine-tuning, knowledge distillation, and specifying the input data format (prompt). This makes the individual comprehensive model a small AI model for different networks and different services, fulfilling the functional requirements of different small models, while improving the running speed of the overall model.

[0133] See Figure 6 This is a schematic diagram of the centralized endogenous AI model provided in this embodiment of the invention. The comprehensive AI model, formed by integrating multiple smaller models, is shown in the figure. Each square represents a neural network, and the specific neural network model can be an integration of commonly used AI models such as Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory Networks (LSTM), and Transformer structures. DG represents the decision generation module, which is responsible for implementing specific functions by calling designated smaller models for different network services. The specific method of calling smaller models is mainly achieved by adjusting the weight ratio of the smaller models to flexibly control the overall structure of the model.

[0134] In both of the above approaches, by downloading the composite neural network model (after cropping or adjusting the weights) into different networks, endogenous AI mini-models suitable for different networks are formed. The deep neural network approach may perform better when handling large-scale data and complex features, and is more effective when there are requirements for the overall network performance. The ensemble mini-model, on the other hand, may provide higher efficiency and accuracy for specific tasks and is more effective when flexibly scalable for business needs. Furthermore, if there are special requirements in both cases, a deep neural network can be integrated into the second approach to form a fusion-based composite model architecture.

[0135] In another embodiment of the present invention, the integrated AI model of the integrated AI module, as the central hub of the entire end-to-end endogenous AI solution, receives data transmitted from AI sub-models in the wireless access network, bearer network, and core network, including model optimization schemes, model performance evaluation results, and network device performance logs. It first performs comprehensive analysis on this data, utilizing big data analytics to mine massive amounts of data to identify cross-network patterns and trends. Next, the integrated AI model allocates weights based on the collected data, prioritizing and allocating resources for optimization schemes in different network parts to ensure optimal network operation. Subsequently, the integrated AI model calls upon various sub-models or performs model pruning to specifically adjust the optimization strategies for the wireless access network, bearer network, and core network, generating local optimization strategies. These strategies not only consider the performance improvement of individual network parts but also take into account the synergistic effect of the entire communication network to achieve end-to-end performance optimization. After the optimization strategies are generated, the integrated AI model further performs benefit evaluation, using simulation and predictive analysis to assess the potential network performance improvements and cost-effectiveness of implementing these strategies, ensuring the practical feasibility and return on investment of the optimization strategies. In addition, the integrated AI model is also responsible for digital twin network verification. It uses digital twin technology to create a virtual copy of the network, simulates and tests the optimization strategy, and ensures the effectiveness and security of the strategy before actual deployment.

[0136] Finally, the integrated AI model feeds back the validated local optimization strategies to the individual AI sub-models in each network component via the WebSocket solution, guiding them to make corresponding adjustments and optimizations. Simultaneously, the integrated AI model continuously monitors network performance to ensure the continued effectiveness of the optimization strategies and dynamically adjusts them based on changes in network conditions to address evolving network demands and challenges.

[0137] In another embodiment provided by the present invention, each endogenous AI module monitors the optimized network performance indicators of each network node and feeds them back to the integrated AI module, specifically:

[0138] The workflow for each network is as follows:

[0139] The core network's intrinsic AI model receives local optimization strategies from the integrated AI model and combines them with its own optimization schemes and performance evaluation results. It then applies these optimization strategies to the core network and adjusts them based on real-time core network data to achieve high-speed data transmission and optimized quality of service. Simultaneously, it collects key network performance indicators and fault records to provide real-time data support for the model. After implementation, the effectiveness of the optimization strategies is evaluated, and the results are fed back to the integrated AI model for subsequent global adjustments.

[0140] The inherent AI model of the bearer network receives local optimization strategies from the integrated AI model, as well as its own optimization schemes and performance evaluation results. Based on the specific conditions of the bearer network, it refines the local optimization strategies issued by the integrated model and implements them in the network, such as adjusting routing strategies and load balancing. Simultaneously, it collects network performance data and device status information for in-depth analysis to evaluate the effectiveness of the optimization strategies and feeds the results back to the integrated AI model.

[0141] Wireless Network: The built-in AI model of the wireless access network receives local optimization strategies from the comprehensive AI model via WebSocket. Based on the specific environment and conditions of the wireless access network, the optimization strategies are adapted and adjusted to ensure their effectiveness at the wireless access network level. The adjusted optimization strategies are applied to the wireless access network, including adjusting base station configuration and optimizing wireless resource allocation. Network performance after the implementation of the optimization strategies is monitored, and key performance indicators such as packet loss rate and throughput are collected. The effectiveness of the optimization strategies is evaluated, and the performance data and evaluation results are encapsulated in a format suitable for the WebSocket protocol and sent back to the comprehensive AI model.

[0142] The end-to-end AI large-scale model solution adopts a "centralized + distributed" collaborative architecture, supporting both local and global optimization. Each node and terminal in the 6G network can download the comprehensive AI model, make local optimizations and adjustments according to different business needs, and provide feedback on the optimization strategies.

[0143] By incorporating analytical and decision-making capabilities into each component of the end-to-end endogenous AI large model solution, and by analyzing and learning from feature datasets stored locally or uploaded by small AI models, timely coordination and scheduling of devices are achieved, ensuring the overall quality of the 6G network.

[0144] In another embodiment of the present invention, the end-to-end endogenous AI large model solution has fault management functionality. See also Figure 7 This is another schematic diagram of the network optimization system provided in this embodiment of the invention. It includes wireless, bearer, core nodes and devices, an endogenous AI model, and an external integrated AI model.

[0145] Wireless, bearer, core nodes and devices generate alarm content and upload the alarm information to their respective built-in AI mini-models.

[0146] The built-in AI module runs continuously on the network node side, tightly integrated with various monitoring systems of the network nodes. It comprehensively collects alarm information generated by the nodes through real-time polling and event-driven methods. This alarm information covers various possible fault scenarios, including network connection interruptions, excessive device temperature, abnormal port traffic, and protocol errors, originating from core network devices such as routers, switches, and servers, as well as various access devices. Using built-in intelligent analysis algorithms, the alarm information is initially screened and classified, removing duplicate or obviously false alarms and extracting the truly critical alarm events that require attention. For example, for alarms that occur multiple times in a short period with identical content, only the first occurrence is retained as a valid alarm; alarms caused by brief device restarts or temporary network jitter that quickly recover are marked as low priority or ignored.

[0147] Once a critical alarm event is confirmed, the intrinsic AI module quickly initiates the fault diagnosis process, conducting in-depth analysis based on a pre-configured intrinsic AI model. This model is trained on a large amount of historical fault data, simulated fault cases, and expert experience and knowledge, and can accurately extract feature datasets related to the current fault from massive amounts of alarm information and node operation data.

[0148] It should be noted that the extraction of the feature dataset involves multiple aspects of data. This includes not only the detailed information carried by the alarm event itself, such as alarm code, alarm level, and alarm occurrence time, but also key operational data of network nodes within a period before and after the alarm occurrence, such as CPU utilization, memory usage, network traffic flow, and port status changes of relevant devices. By organically integrating this multi-dimensional data, a complete feature dataset reflecting the fault state is formed, providing a strong basis for subsequent fault self-repair and accurate diagnosis.

[0149] Based on the identified fault feature dataset, the intrinsic AI module attempts to perform self-repair operations according to the built-in repair strategy library. This repair strategy library, also built upon extensive prior practical experience and simulation testing, outlines corresponding countermeasures for different types of fault scenarios.

[0150] For example, for alarms caused by network port congestion, the built-in AI module may automatically adjust the port's traffic control strategy, such as enabling traffic shaping technology to reasonably allocate incoming and outgoing traffic to the port according to pre-set bandwidth allocation rules, thereby alleviating congestion. If the alarm is found to be caused by excessive server process resource consumption, the module can try to use the process management tools provided by the operating system to appropriately reduce the priority of non-critical processes, freeing up more resources for critical business processes to ensure stable server operation. As another example, for alarms caused by excessive device temperature, the built-in AI module can remotely control the device's cooling fan speed to increase, or, if conditions permit, adjust the device's workload to reduce power consumption, thereby reducing temperature.

[0151] Shortly after the self-repair operation is completed (e.g., 1-2 minutes), the endogenous AI module closely monitors whether the fault has been effectively resolved. By polling the alarm status of network nodes again and comparing the changes in key operating parameters, it conducts a comprehensive evaluation of the self-repair results.

[0152] When self-repair fails, the feature dataset is uploaded to the integrated AI module.

[0153] By leveraging the built-in AI module for real-time monitoring and rapid self-healing at network nodes, a response can be initiated immediately upon a fault, significantly reducing the time interval between fault occurrence and the start of troubleshooting. Compared to traditional methods relying on manual fault diagnosis, response time can often be reduced from hours or even days to minutes, greatly minimizing the impact of faults on network services and ensuring the continuity of network services.

[0154] In yet another embodiment of the present invention, when the endogenous AI module performs self-repair, it specifically executes the following steps:

[0155] First, the fault type is determined by itself, and the alarm and other information are sorted by time to obtain the abnormal dataset.

[0156] Then, based on the sliding window along the time dimension, the corresponding abnormal event chain is obtained. The abnormal events are constructed into a feature sequence and used as the input dataset for the model to attempt self-repair. If self-repair is successful, the configuration scheme is applied to specific nodes and devices; if self-repair is not successful, the resulting feature dataset is uploaded to the comprehensive AI model via the WebSocket protocol for further diagnostic processing.

[0157] In another embodiment of the present invention, when the integrated AI module performs the repair process, specifically:

[0158] After receiving the feature dataset uploaded by the small model, it is used as the input dataset for training. Finally, the fault classification corresponding to the anomaly is determined. Based on the diagnostic module of the comprehensive model, logs containing information such as fault time, location, and cause are generated. Cluster analysis is performed to trace the cause of the fault, adjust the business transmission path, and provide fault repair suggestions and solutions.

[0159] In yet another embodiment of the present invention, the system further includes a digital twin network.

[0160] After generating corresponding fault analysis and handling solutions, the integrated AI model first applies the specific repair solutions to the digital twin network for simulation verification. If the fault is resolved in the digital twin network, the solution is distributed to the AI ​​mini-models of each network node to complete the configuration execution, thus achieving efficient intelligent fault handling.

[0161] The integrated AI model generates specific remediation plans and deploys them in a digital twin network. Following the instructions in the plan, it operates on the corresponding nodes and links in the digital twin network, simulating the execution process in a real network. For example, adjusting the routing table entries of virtual routers and changing the resource allocation ratio of virtual servers in the digital twin network is just like implementing it in a real network.

[0162] Initiate a simulation operation mechanism for the digital twin network to observe whether the fault has been effectively resolved according to the plan in a simulated environment. Monitoring indicators include network connectivity, the rationality of traffic distribution, and the performance recovery of equipment. Compare these indicators with the normal state before the fault occurred, and use quantitative evaluation metrics to determine the repair effectiveness.

[0163] After a period of simulation (set according to the complexity of the fault, such as 5 minutes for simple faults and 15 minutes for complex faults), a comprehensive evaluation of the fault repair in the digital twin network is conducted. If all monitoring indicators show that the fault has been properly resolved, the verification is passed, indicating that the repair plan is feasible. If the fault still exists or causes new problems, such as increased local network congestion or no improvement in equipment performance, the verification is deemed a failure, and the fault needs to be re-analyzed using the AI ​​model to optimize the repair plan.

[0164] In another embodiment of the present invention, the digital twin network constructs a simulated network link by acquiring the resource configuration information of network nodes.

[0165] Specifically, wireless, bearer, core nodes and devices generate alarm content, upload the alarm information to their respective AI mini-models, and simultaneously upload resources and configurations to the digital twin network to form a refined replication of the existing network situation.

[0166] Comprehensive network operation data is collected from various nodes of the actual network, including the core network, bearer network, and wireless network, using diverse data collection methods such as SNMP protocol for network device status collection, NetFlow for traffic information collection, and base station API for wireless parameters. This data covers various indicators such as device performance, connection status, and traffic load.

[0167] Using a dedicated data synchronization mechanism, these real-time data are transmitted to the digital twin network platform to ensure that the digital twin can accurately reflect the actual network status in real time. The data transmission frequency is set according to the dynamic characteristics of the network. For example, traffic data that changes frequently is synchronized once per second, and basic device performance data is synchronized once every 10 seconds.

[0168] In a digital twin network platform, a digital twin model of the network is constructed using modeling techniques based on collected data. For the network topology, graph theory methods are used to construct a connection graph of nodes and links to accurately simulate the architecture layout of the actual network. For the device operation mechanism, device function models are established, such as the routing and forwarding model of routers and the resource scheduling model of servers, so that the behavior of digital twin devices is highly consistent with that of actual devices.

[0169] By calibrating parameters, the performance parameters and initial states of actual network devices are mapped to the digital twin model, achieving seamless integration between the two in terms of initial conditions and operational characteristics, making the digital twin network a "mirror" of the actual network.

[0170] In another embodiment of the present invention, multi-dimensional network data, including time, space, and perception data, from nodes, terminals, and AI platforms in a 6G network is collected using different protocols. Simultaneously, the integrated AI model can write the data into a digital twin network, enabling parallel and real-time data collection across a large-scale network. This achieves a refined "replication" of the entire process of network elements, topology, services, and data from the real network. (See also...) Figure 8 This is a schematic diagram of the framework of the digital twin network provided in an embodiment of the present invention.

[0171] Digital twin networks obtain basic configuration information of network node devices through management channels (NETCONF / SNMP);

[0172] Topology information related to Layer 3 protocols is obtained through the Control Protocol Channel (BGP-LS);

[0173] Real-time performance data of the device is obtained through the performance acquisition channel (Telemetry / SNMP);

[0174] The device information includes interface information, device manufacturer information, network element type, and version signal;

[0175] The topology information includes link information, latency, interface bandwidth, and tags;

[0176] The real-time performance data includes bandwidth, latency, packet loss rate, and occupancy.

[0177] It also has corresponding storage functions. The platform can classify data according to different business types to form a queryable dataset that can serve 6G services. Specifically, it should include the time (year, month, day), space, physical entity information (port, link, etc.), operation status information (device status, performance data, etc.), configuration-related information and data flow information (source / destination IP, protocol, etc.) mentioned above.

[0178] In another embodiment of the present invention, the integrated AI model scheme has traffic prediction function. In the future 6G network, with the addition of SCU, the information flow of the base station will no longer go to a single core, but can go to different nodes. According to the data collection and storage function proposed in 3, the AI ​​model can extract data from each node and terminal in the 6G network, perform correlation learning, splice and fuse the feature information of each part to form a feature dataset of the link, and complete the traffic prediction of the link.

[0179] The feature dataset is used as input data for AI traffic prediction models (RNN, LSTM, etc.) to learn the traffic situation of each path at different times, and to make traffic predictions for different functions in different time ranges.

[0180] Short-term forecasting: Predicts traffic over several minutes to several hours, suitable for scenarios requiring real-time responses to network conditions, such as network management and load balancing.

[0181] Mid-term forecast: Covers traffic from a few hours to a few days, suitable for routine network maintenance and upgrade plans.

[0182] Long-term forecasting: Predicts traffic trends for the next few days, weeks, or even months, suitable for long-term network planning and expansion.

[0183] The AI ​​model can extract logs from various nodes and terminals in a 6G network. Based on traffic prediction in 5G, it assesses and classifies the transmission quality of each link, including real-time status, connectivity, congestion, and latency. Specific functions should include:

[0184] Visualized Energy Consumption Interface: The proposed solution supports energy consumption visualization, which can intuitively display the real-time energy consumption and energy efficiency data of each link in the entire network, form a network-wide energy consumption map, and trace the historical trends of the entire network, helping customers to fully understand the overall energy consumption of the network.

[0185] Intelligent multi-service transmission control: In multi-service transmission control, the MTL (Multi-task Learning) neural network is used to comprehensively judge and classify the optional paths of each service, and finally to carry out refined management and intelligent control of service transmission paths to ensure the overall network quality.

[0186] Energy-saving solution evaluation: Through digital twin networks, various network energy-saving and line optimization solutions can be explored without affecting the normal use of the network. The optimization quality and risks of the solutions can be evaluated as a whole, so as to select the relatively optimal solution and backup solutions.

[0187] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A network optimization system, characterized in that, The system includes: an endogenous AI module configured inside the network node, and a comprehensive AI module configured outside the network node; The endogenous AI module acquires node data of the corresponding network node, performs self-learning on the acquired node data through a pre-trained endogenous AI model, generates a local dataset, transmits the local dataset to the integrated AI module, and optimizes and adjusts the network node based on the feedback from the integrated AI module. The integrated AI module analyzes the acquired local dataset using a pre-trained integrated AI model, generates optimization strategies, and feeds these optimization strategies back to the endogenous AI module.

2. The network optimization system as described in claim 1, characterized in that, The endogenous AI module converts the local dataset into a preset protocol format and sends the converted local dataset to the endogenous integrated AI module. The local dataset includes network optimization schemes, model performance evaluation results, and network device performance logs. The network optimization scheme includes routing optimization strategies, dynamic resource allocation strategies, and fault prevention measures.

3. The network optimization system as described in claim 2, characterized in that, The specific protocol format is the WebSocket protocol; The endogenous AI module sends the network optimization scheme and model performance evaluation results to the integrated AI module through a connection established via the WebSocket protocol. The intrinsic AI module transmits the network device performance logs to the integrated AI module via the uplink of the network node.

4. The network optimization system as described in claim 1, characterized in that, The intrinsic AI module obtains data from corresponding network nodes through intelligent data interfaces using different protocols, and integrates the obtained data to generate node data in a unified format.

5. The network optimization system as described in claim 4, characterized in that, The intelligent data interface sends configuration data and control commands through any one of the NetConf protocol, PCEP protocol, BGP-LS protocol and Telemetry protocol. The intelligent data interface uses the SFTP protocol to call the AI ​​model API interface and distribute AI model strategies. The intelligent data interface uses the SFTP protocol for AI model loading and version control. The intelligent data interface collects and reports large volumes of data via the SFTP protocol. The intelligent data interface collects and reports the configuration processing results through the Netconf protocol.

6. The network optimization system as described in claim 1, characterized in that, The endogenous AI module performs data preprocessing and data format conversion on the collected node data to obtain MCP protocol node data for learning the endogenous AI model.

7. The network optimization system as described in claim 1, characterized in that, The endogenous AI module trains the AI ​​model using the acquired training data, obtains prediction results, and compares the prediction results with the pre-labeled tags in the training data to optimize the AI ​​model and obtain the endogenous AI model.

8. The network optimization system as described in claim 1, characterized in that, The network nodes include the core network and / or the bearer network and / or the wireless network.

9. The network optimization system as described in claim 1, characterized in that, When the network node includes a wireless access network, the node data acquired by the endogenous AI module includes base station data, node device data, terminal data, channel condition data, user behavior data, and base station traffic data. The intrinsic AI module uses network performance metrics, including packet loss rate and throughput, as well as QoS metrics, including bandwidth, latency, and signal-to-noise ratio, as training labels to predict and analyze network performance and service quality, and to train the intrinsic AI model.

10. The network optimization system as described in claim 1, characterized in that, When the network node includes a bearer network, the node data acquired by the endogenous AI module includes monitoring data of key equipment; The key equipment includes routers and / or switches; The monitoring data includes network topology information composed of network traffic, device status information, and link quality information, as well as network element information; The endogenous AI module extracts key features from the node data to perform self-learning of the AI ​​model and outputs a network optimization scheme. The key features include traffic patterns, link load, link latency, and packet loss rate.

11. The network optimization system as described in claim 1, characterized in that, When the network node includes the core network, the node data acquired by the endogenous AI module includes user session data, network slice status, resource utilization, and fault records. The endogenous AI module extracts features from the acquired training data, identifies key features that affect the core network performance, and uses network performance indicators and QoS indicators as labels to train the AI ​​model, thus obtaining the endogenous AI model. The key characteristics include peak traffic periods, resource bottlenecks, and potential points of failure in the link.

12. The network optimization system as described in claim 1, characterized in that, The integrated AI model is a deep neural network or a centralized endogenous AI model; The integrated AI model includes the endogenous AI models of each network node.

13. The network optimization system as described in claim 1, characterized in that, The integrated AI module uses an integrated AI model to comprehensively analyze the acquired local dataset, identify cross-network patterns and development trends, assign weights to the collected data, prioritize and allocate resources for optimization schemes of different network parts, call various sub-models or perform model pruning and other operations to adjust the local optimization strategies of network nodes, generate optimization strategies, and feed back the verified optimization strategies to the corresponding endogenous AI module.

14. The network optimization system as described in claim 1, characterized in that, The endogenous AI module monitors the network performance indicators after the network nodes have been optimized and adjusted, and feeds back the monitored network performance indicators to the integrated AI module.

15. The network optimization system as described in claim 1, characterized in that, After the intrinsic AI module identifies the alarm event generated by the network node, it identifies the feature dataset of the fault through the configured intrinsic AI model and performs self-repair of the alarm fault. When self-repair fails, the feature dataset is uploaded to the integrated AI module.

16. The network optimization system as described in claim 1, characterized in that, The intrinsic AI module uses a configured intrinsic AI model to identify fault types based on alarm information and generates an abnormal dataset by sorting by time. The abnormal event chain is determined by a sliding window in the time dimension, and the feature sequence is obtained as the feature dataset. A self-repair scheme is generated based on the feature dataset. Execute the self-healing scheme and determine whether the alarm event has been eliminated; If the alarm event is not cleared, the self-repair is deemed to have failed.

17. The network optimization system as described in claim 15, characterized in that, The integrated AI module uses the received feature dataset as input data for learning the integrated AI model to identify fault logs; it also performs cluster analysis to trace the cause of the fault and outputs a fault repair plan. The fault log includes the fault time, fault location, and fault cause.

18. The network optimization system as described in claim 17, characterized in that, The system also includes a digital twin network; The processor inputs the fault repair scheme output by the integrated AI model into the digital twin network; The simulated network links constructed by the digital twin network are used for simulation verification; The verified fault repair plan will be fed back to each network node for execution.

19. The network optimization system as described in claim 18, characterized in that, The digital twin network constructs simulated network links by acquiring resource configuration information of network nodes.

20. The network optimization system as described in claim 19, characterized in that, The digital twin network obtains basic configuration information of network node devices through the management channel, obtains topology information related to the three-layer protocol through the control protocol channel, and obtains real-time performance data of the devices through the performance acquisition channel. The device information includes interface information, device manufacturer information, network element type, and version signal; The topology information includes link information, latency, interface bandwidth, and tags; The real-time performance data includes bandwidth, latency, packet loss rate, and occupancy.

21. The network optimization system as described in claim 1, characterized in that, The integrated AI module acquires data from each network node through an integrated AI model, performs correlation learning, splices and merges the feature information of each part to form a feature dataset of the link, and predicts the traffic of the link based on the feature dataset.

Images