Wireless network artificial intelligence model management method, apparatus, medium, and product

Abstract

Embodiments of the present application provide a wireless network artificial intelligence model management method, device, medium and product. The method comprises: in the case that an artificial intelligence (AI) model has been deployed at a base station and a terminal, collecting index data of the base station; evaluating whether the AI model has positive gain on the wireless network based on the collected index data; in the case that the AI model has positive gain on the wireless network, saving the index data corresponding to the AI model; in the case that the AI model does not have positive gain on the wireless network, retraining the AI model, and deploying the retrained AI model to the base station and the terminal according to a required application scenario. The scheme of the present application can solve the difference between offline simulation and network application, the limitation of channel estimation, and the isolation of base station AI model deployment by collecting index data to evaluate whether the AI model brings positive gain to the network.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a method, apparatus, medium, and product for managing artificial intelligence models in wireless networks. Background Technology

[0002] Currently, there is a lack of unified standards for the transmission of artificial intelligence (AI) model metrics and parameters between base stations and terminals, and there are no clear evaluation thresholds to measure the positive or negative gain that AI models bring to the network. This leads to a series of challenges for AI-based wireless networks in practical applications.

[0003] In one existing technology, current AI model simulation uses offline channel data. In live network applications, different AI neural network models have different memory requirements for terminals. When switching AI models, base stations must consider the UE's chip capabilities and the list of deployed AI models. Currently, there is a lack of corresponding standards and mechanisms to support this collaboration, resulting in discrepancies between offline simulation and live network applications. In another existing technology, the AI ​​model may exhibit local optima, while the global network's optimality is uncertain, leading to inconsistencies between local and global optima and limitations in channel estimation. Yet another existing technology deploys AI training models independently on each base station, requiring separate resource allocation for each station. There is a lack of a common management platform to integrate the network element AI capabilities of the current area, and a lack of information sharing mechanisms between base stations, resulting in the isolation of base station AI model deployment. Summary of the Invention

[0004] At least one embodiment of this application provides a wireless network artificial intelligence model management method, apparatus, medium, and product to address the problems of discrepancies between offline simulation and live network applications, limitations of channel estimation, and isolation of base station AI model deployment in the prior art.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows:

[0006] In a first aspect, embodiments of this application provide a wireless network artificial intelligence model management method, applied to an Operation and Maintenance Center (OMC), comprising:

[0007] With AI models already deployed on the base station and terminal, collect the indicator data of the base station;

[0008] Based on the collected indicator data, evaluate whether the AI ​​model has a positive gain for the wireless network;

[0009] If the AI ​​model has a positive gain for the wireless network, save the index data corresponding to the AI ​​model;

[0010] If the AI ​​model does not provide a positive gain for the wireless network, the AI ​​model is retrained, and the retrained AI model is deployed to the base station and terminal according to the required application scenario.

[0011] Optionally, if an artificial intelligence (AI) model has been deployed on the base station and the terminal, the method further includes the following step before collecting the indicator data of the base station:

[0012] Receive target cell information reported by the base station using a preset interface;

[0013] Based on the target information, determine whether the base station has deployed the AI ​​model.

[0014] Optionally, the target information element includes:

[0015] Terminal chip performance parameters;

[0016] List of AI models supported by the terminal;

[0017] AI model parameters and metrics have been deployed on the terminal.

[0018] Optionally, after determining whether the base station has deployed an artificial intelligence (AI) model based on the target information, the method further includes:

[0019] In the absence of AI models deployed on base stations and terminals, preset indicator triggering conditions can be set using the task orchestration module.

[0020] When the base station meets the preset indicator triggering condition, the network information corresponding to the preset indicator triggering condition in the base station is collected using the preset interface;

[0021] Based on the network information, suitable target AI models for deployment are selected, and it is determined whether the current computing power resources of the base station meet the deployment requirements of the target AI models.

[0022] If the deployment requirements are met, a data collection task is issued through the preset interface; the data collection task is used to instruct the collection of training data required by the target AI model.

[0023] After preprocessing the target data acquired by the data acquisition task, the processed data is sent to the AI ​​model resource pool for model training to determine the first model index parameters after training.

[0024] The first model index parameters are deployed and verified, and the first target parameters after successful deployment and verification are sent to the base station.

[0025] Optionally, based on the collected indicator data, evaluating whether the AI ​​model has a positive gain for the wireless network includes:

[0026] Based on the performance indicators in the collected indicator data, evaluate whether the performance indicators meet the preset threshold conditions corresponding to the performance indicators.

[0027] If each parameter in the performance index meets the corresponding preset threshold condition, then the AI ​​model is determined to have a positive gain for the wireless network.

[0028] If at least one of the performance metrics fails to meet the corresponding preset threshold condition, then the AI ​​model is determined to have no positive gain for the wireless network.

[0029] Optionally, the preset interface is used to transmit at least one of the following information:

[0030] Terminal chip capability parameters;

[0031] AI model coefficients, AI model weights, and a list of AI models supported by the base station and / or terminal;

[0032] AI model evaluation metrics;

[0033] AI Task ID;

[0034] AI task strategy.

[0035] Optionally, the OMC includes an evaluation center module, an AI model resource pool module, a task orchestration module, and a deployment verification module, and the method includes performing at least one of the following functions:

[0036] Using the evaluation center module, performance indicator thresholds, model evaluation indicator thresholds, and business indicator thresholds are set for different business scenarios that support AI models;

[0037] The AI ​​model resource pool module is used to train different AI models and neural networks, and to centrally process the training tasks of base station and terminal AI models.

[0038] The task orchestration module is used to store the AI ​​model capability list of each base station, maintain the AI ​​tasks of each base station, and periodically inspect the deployment data of the AI ​​models of each base station.

[0039] Using the deployment verification module, the trained model is evaluated for computing power based on the base stations to be deployed, and the model weights are pruned.

[0040] Secondly, this application provides a wireless network artificial intelligence model management device, applied in an Operation and Maintenance Center (OMC), comprising:

[0041] The first processing module is used to collect the indicator data of the base station when an artificial intelligence (AI) model has been deployed in the base station and the terminal.

[0042] The second processing module is used to evaluate whether the AI ​​model has a positive gain for the wireless network based on the collected indicator data.

[0043] The third processing module is used to save the index data corresponding to the AI ​​model when the AI ​​model has a positive gain for the wireless network.

[0044] The fourth processing module is used to retrain the AI ​​model when the AI ​​model does not have a positive gain for the wireless network, and to deploy the retrained AI model to the base station and the terminal.

[0045] Thirdly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0046] Fourthly, this application provides a computer program product including computer instructions that, when executed by a processor, implement the steps of the method described in the first aspect.

[0047] Compared with existing technologies, the wireless network artificial intelligence model management method, apparatus, medium, and product provided in this application, when an artificial intelligence (AI) model has been deployed on the base station and terminal, collects base station indicator data in real time and evaluates whether the AI ​​model has a positive benefit to the wireless network based on the collected data. If the model performs poorly, it is fed back to the simulation environment for optimization. When the model performs poorly, it is retrained using more diverse datasets and global optimization algorithms. The retrained AI model is then deployed to the base station and terminal and uniformly managed and updated through OMC. The AI ​​model is continuously optimized according to changes in the network environment and user needs to maintain its optimal performance. Through the above solutions, the differences between offline simulation and live network applications, the limitations of channel estimation, and the isolation of base station AI model deployment can be effectively solved. Attached Figure Description

[0048] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0049] Figure 1 A flowchart illustrating the wireless network artificial intelligence model management method provided in this application embodiment;

[0050] Figure 2 An architecture diagram of the Operation and Maintenance Center (OMC) provided for embodiments of this application;

[0051] Figure 3 A method flowchart of a first specific embodiment provided in this application;

[0052] Figure 4 A flowchart illustrating the method of a second specific embodiment provided in this application;

[0053] Figure 5 This is a structural diagram of the wireless network artificial intelligence model management device provided in an embodiment of this application. Detailed Implementation

[0054] The terms "first," "second," etc., used in this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, without limiting the number of objects; for example, the first object can be one or more. Furthermore, "or" in this application indicates at least one of the connected objects. For example, "A or B" covers three scenarios: Scenario 1: including A but not B; Scenario 2: including B but not A; Scenario 3: including both A and B. The character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0055] The term "instruction" in this application can be either a direct instruction (or explicit instruction) or an indirect instruction (or implicit instruction). A direct instruction can be understood as one in which the sender explicitly informs the receiver of specific information, the operation to be performed, or the requested result, etc., in the instruction sent. An indirect instruction can be understood as one in which the receiver determines the corresponding information based on the instruction sent by the sender, or makes a judgment and determines the operation to be performed or the requested result, etc., based on the judgment result.

[0056] As described in the background section, current AI model simulations mostly utilize offline channel data, which eliminates the need to consider the AI ​​capabilities of the UE (User Equipment) chip during the simulation phase. However, in live network applications, different AI neural network models have varying memory requirements, requiring base stations to consider the UE's chip capabilities and the list of deployed AI models when switching AI models. This necessitates signaling collaboration between the UE and the base station, but currently lacks corresponding standards and mechanisms to support this collaboration, resulting in discrepancies between offline simulation and live network applications. In the pursuit of low overhead and high accuracy, channel estimation can easily lead AI models into local optima. This means that while good results may be achieved in link-level simulations, it's uncertain whether this optimization brings positive gains to the entire network from a global network perspective. This inconsistency between local and global optima limits the effectiveness of channel estimation. Currently, AI training models are typically deployed independently on each base station, requiring separate resource allocation for each base station to support their operation. Furthermore, there is a lack of a common management platform to integrate the AI ​​capabilities of network elements within a given area, as well as information sharing mechanisms between base stations. This isolation not only reduces resource utilization efficiency but also hinders the widespread application and further optimization of AI technology in wireless networks, resulting in the problem of isolated deployment of AI models in base stations. To address at least one of the above problems, embodiments of this application provide a method, apparatus, medium, and product for managing artificial intelligence models in wireless networks, which can reduce or avoid the occurrence of the above situations, improve network stability, and prevent unsuitable AI models from causing losses to the wireless network.

[0057] Please refer to Figure 1 This application provides a wireless network artificial intelligence model management method, applied to an Operation and Maintenance Center (OMC), comprising:

[0058] Step 11: If the base station and terminal have already deployed an artificial intelligence (AI) model, collect the base station's performance data. This performance data includes, but is not limited to, model inference results and wireless network performance metrics.

[0059] In this embodiment, when an artificial intelligence (AI) model has been deployed on the base station and the terminal, the base station's indicator data is collected using the preset interface of the OMC. The preset interface of this application can be selected as the AI ​​capability interface. In this application, an intelligent task orchestration module is used to set a timed collection task. The timed task can be triggered based on a preset time detection or a specific event. The timed collection task collects the base station's indicator data through the preset interface data and comprehensively collects the indicator data of the AI ​​model deployed on the base station through the AI ​​capability interface.

[0060] Model inference result evaluation metrics are used to assess the performance and effectiveness of AI models. These metrics include, but are not limited to, NMSE (Normalized Mean Squared Error, which measures the difference between model predictions and actual values), Squared Generalized Cosine Similarity (SGCS), and FLOPS (Floating-Point Operations, which measures the computational complexity of the model). These metrics allow for a quantitative evaluation of AI model performance, enabling the selection of the optimal model for deployment.

[0061] Performance metrics for wireless networks are primarily used to measure the performance of wireless communication networks. These metrics include, but are not limited to, throughput (representing the amount of data successfully transmitted by the network per unit time), CQI (Channel Quality Indicator, used to reflect the transmission quality of the wireless channel), and RSRP (Reference Signal Received Power, used to assess the strength of the wireless signal). Performance metrics can help to comprehensively understand the network's operating status and performance level.

[0062] For example, taking the CSI compression model as an example, the AI ​​model needs to collect data such as its NMSE, FLOPS, SGCS, as well as the antenna configuration and network performance indicators of the base station and the terminal.

[0063] Step 12: Based on the collected indicator data, evaluate whether the AI ​​model has a positive gain for the wireless network;

[0064] Step 13: If the AI ​​model has a positive gain for the wireless network, save the index data corresponding to the AI ​​model;

[0065] Step 14: If the AI ​​model does not provide a positive gain for the wireless network, retrain the AI ​​model and deploy the retrained AI model to the base station and terminal according to the required application scenario.

[0066] In this embodiment, after the AI ​​model has been deployed on the base station and terminal, the base station's performance data is collected. This step involves collecting the model inference results of the AI ​​model and the performance metrics of the wireless network. The model inference results may include NMSE (Normalized Mean Square Error), FLOPS (Floating-Point Operations), etc., to evaluate the model's accuracy and efficiency. The wireless network's performance metrics may cover throughput, CQI (Channel Quality Indicator), RSRP (Reference Signal Received Power), etc., to comprehensively reflect the network's actual performance. Based on the collected performance data, the AI ​​model is evaluated to determine whether it has a positive impact on the wireless network. By comparing the performance metrics of the deployed model, such as throughput compared to a corresponding preset threshold, it is determined whether the throughput has increased. Similarly, by using similar methods, it is determined whether latency has decreased, CQI has improved, etc., to determine whether the current AI model has a positive impact on the wireless network. At the same time, the stability and accuracy of the model inference results also need to be considered to ensure the reliability of the model in practical applications.

[0067] If the AI ​​model provides a positive gain for the wireless network, the corresponding metric data should be saved. Positive gain refers to whether there is a global and local improvement in network performance metrics before and after model deployment. If there is only a local improvement and a global decline, it may have a negative gain. Saving the metric data not only provides valuable reference for subsequent model optimization but also facilitates the rapid replication and promotion of successful models in similar application scenarios. The saved metric data should include model parameters, model inference results, and corresponding wireless network performance metrics to ensure information integrity and traceability.

[0068] Conversely, if the AI ​​model does not provide a positive gain for wireless networks, it needs to be retrained. Retraining may involve adjusting the model structure, optimizing hyperparameters, or increasing training data to improve model performance. During retraining, the specific requirements and limitations of the wireless network should be fully considered to ensure model compatibility with the network. After retraining, the optimized AI model needs to be redeployed to base stations and terminals according to their application scenarios, and a new round of evaluation and optimization should be conducted.

[0069] Through the above steps, this application enables continuous evaluation and optimization of AI models in wireless networks, ensuring that the model consistently brings positive benefits to the network, improving network stability, and preventing unsuitable AI models from causing losses to the wireless network.

[0070] Optionally, step 12 above includes:

[0071] Based on the performance indicators in the collected indicator data, evaluate whether the performance indicators meet the preset threshold conditions corresponding to the performance indicators.

[0072] If each parameter in the performance index meets the corresponding preset threshold condition, then the AI ​​model is determined to have a positive gain for the wireless network.

[0073] If at least one of the performance metrics fails to meet the corresponding preset threshold condition, then the AI ​​model is determined to have no positive gain for the wireless network.

[0074] In this embodiment of the application, when evaluating whether the AI ​​model has a positive gain for the wireless network based on the collected performance index data, it is first necessary to define the collected performance indicators. These indicators may include throughput, CQI (Channel Quality Indicator), RSRP (Reference Signal Received Power), etc., which can comprehensively reflect the actual performance of the wireless network. At the same time, corresponding preset threshold conditions should be set for these performance indicators. These threshold conditions are determined based on network operation requirements, industry standards, or historical experience.

[0075] Next, the actual collected performance metrics data are compared one by one with the corresponding preset threshold conditions. This step is crucial for evaluating whether the AI ​​model has a positive benefit to the wireless network. Specifically: if each parameter in the performance metrics meets the corresponding preset threshold conditions, that is, the actual performance data reaches or exceeds the expected standard, then it can be determined that the AI ​​model has a positive benefit to the wireless network. This means that the introduction of the AI ​​model improves network performance, such as increased throughput, improved channel quality, and enhanced signal reception power, thereby improving the overall network performance and user experience.

[0076] If at least one parameter in the performance metrics fails to meet its corresponding preset threshold, meaning the actual performance data does not meet the expected standard, then it can be determined that the AI ​​model does not provide a positive benefit to the wireless network. In this case, the AI ​​model may have failed to effectively improve network performance, and may even have had a negative impact on the network. Further optimization or adjustment of the AI ​​model is then necessary to improve its performance.

[0077] This application accurately assesses whether an AI model provides a positive benefit to a wireless network by comparing actual collected performance data with preset threshold conditions. This method is not only scientifically effective but also provides strong support for network operation and optimization.

[0078] Optionally, refer to Figure 2 As shown, the OMC includes an evaluation center module, an AI model resource pool module, a task orchestration module, and a deployment and verification module. The method includes performing at least one of the following functions:

[0079] Using the evaluation center module, performance indicator thresholds, model evaluation indicator thresholds, and business indicator thresholds are set for different business scenarios that support AI models;

[0080] The AI ​​model resource pool module is used to train different AI models and neural networks, and to centrally process the training tasks of base station and terminal AI models.

[0081] The task orchestration module is used to store the AI ​​model capability list of each base station, maintain the AI ​​tasks of each base station, and periodically inspect the deployment data of the AI ​​models of each base station.

[0082] Using the deployment verification module, the trained model is evaluated for computing power based on the base stations to be deployed, and the model weights are pruned.

[0083] In this embodiment, the core of the wireless network AI model management method lies in achieving centralized training, scheduled evaluation, and intelligent deployment of AI models through the functional collaboration of multiple key modules. The evaluation center module sets performance indicator thresholds, model evaluation indicator thresholds, and service indicator thresholds for different business scenarios supporting AI models. Performance indicators such as throughput, CQI, and RSRP directly reflect the actual performance of the network; model evaluation indicators such as NMSE, SGCS, and FLOPS reflect the accuracy and efficiency of the model; and service indicator thresholds such as packet loss rate and latency are key considerations for service quality. Through the setting of these thresholds, the evaluation center can comprehensively and accurately evaluate the performance of AI models in the network.

[0084] The AI ​​model resource pool is the core venue for model training and optimization. It is responsible for training different AI models and neural networks, and classifies models according to scenarios, such as channel estimation, localization enhancement, and beam management. Through centralized processing, the AI ​​model resource pool improves the efficiency and quality of model training, providing networks with a rich and diverse selection of models.

[0085] The task orchestration module is responsible for storing the capability list of AI models for each base station and maintaining the AI ​​tasks for each base station. These tasks include the collection, distribution, and switching of model indicator data, which are crucial for ensuring the effective operation of AI models in the network. Simultaneously, the task orchestration module also periodically checks the deployment data of AI models at each base station to ensure the stability and reliability of the models.

[0086] The deployment and validation module performs computational power evaluation and model weight pruning on the trained model to ensure it can adapt to different base stations and terminals. This step is the last line of defense before model deployment, ensuring the model's feasibility and effectiveness in real-world networks.

[0087] Through the synergistic function of the key modules mentioned above, this wireless network AI model management method enables centralized training of AI models and periodic evaluation of the gains of deployed AI models on the existing network. When a model exhibits negative gain, it can be retrained or switched in a timely manner to prevent network performance degradation. Simultaneously, the timed inspection function of task orchestration enables intelligent deployment of AI models; when network problems occur, corresponding AI models can be deployed to address them. This method not only improves the network's intelligence level but also significantly enhances its performance and stability.

[0088] Optionally, the OMC also includes a data center module, which acts as a central hub for collecting user-level AI data. This data is extensive and in-depth, including but not limited to user chip specifications, lists of supported AI models, and information on deployed AI models. This data provides a solid foundation for subsequent model training and evaluation.

[0089] Furthermore, the preset interface is used to transmit at least one of the following information:

[0090] Terminal chip capability parameters;

[0091] AI model coefficients, AI model weights, and a list of AI models supported by the base station and / or terminal;

[0092] AI model evaluation metrics;

[0093] AI Task ID;

[0094] AI task strategy.

[0095] In this embodiment of the application, the preset interface can be Figure 2 The AI ​​capability interface shown is responsible for transmitting various metrics, tasks, and strategies related to the AI ​​model in the wireless network AI model management method. Specifically, the AI ​​capability interface may contain the following data:

[0096] Chip capability parameters: These data reflect the AI ​​processing capabilities of different chips, including key indicators such as computing power and power consumption. By transmitting chip capability parameters, the system can understand the hardware foundation of each base station and terminal, thereby selecting appropriate AI models for deployment.

[0097] AI model coefficients and weights: These are the core data of the AI ​​model, determining its behavior and performance. By transmitting these parameters, the system can remotely update and configure the AI ​​model to meet different network environments and business needs.

[0098] Supported Model List: This list details the types and versions of AI models supported by the device or system. By transmitting this list, the system can understand the AI ​​capabilities of each base station and terminal, thereby enabling model matching and deployment decisions.

[0099] AI model evaluation metrics: These data are used to evaluate the performance of AI models, such as accuracy, recall, and F1 score. By transmitting these AI model evaluation metrics, the system can monitor and evaluate the performance of AI models in real-world networks, allowing for timely adjustments and optimizations.

[0100] AI Task ID: Each AI task has a unique identifier (ID) used to distinguish and track different tasks. By transmitting the AI ​​Task ID, the system can schedule and manage AI tasks.

[0101] Task strategies: These strategies define key information such as the execution method, timing, and priority of AI tasks. By transmitting task strategies, the system can ensure the efficient and orderly execution of AI tasks within the network.

[0102] By transmitting this data through AI capability interfaces, the system can remotely manage, update, and configure AI models, thereby improving the network's intelligence and performance. Simultaneously, this also facilitates centralized training, evaluation, and optimization of AI models, contributing to the rapid development and application of AI technology in wireless networks.

[0103] Optionally, if an artificial intelligence (AI) model has been deployed on the base station and the terminal, the method further includes the following step before collecting the indicator data of the base station:

[0104] Receive target cell information reported by the base station using a preset interface;

[0105] Based on the target information, determine whether the base station has deployed the AI ​​model.

[0106] In this embodiment, when an artificial intelligence (AI) model has been deployed on both the base station and the terminal, before collecting the base station's indicator data, target information cell information reported by the base station using a preset interface is received. Here, the target information cell information is obtained through interaction between the base station and the terminal via Radio Resource Control (RRC) signaling. For example, the base station has a terminal (UE) capability query function. The base station sends a capability query request to the terminal via RRC signaling. This request includes a capability query request for target information cell information, used to query the terminal's target information cell information.

[0107] Correspondingly, upon receiving a capability query request, the terminal parses the request and prepares the corresponding capability information for reporting. The terminal itself possesses UE capability reporting capabilities. Based on the base station's capability query request, the terminal encapsulates its own target information element information within the target information element information of the RRC signaling. Subsequently, the terminal reports this capability information to the base station via RRC signaling. In this way, the base station and the terminal can efficiently transmit capability information related to the AI ​​model, providing strong support for intelligent network management and optimization. Simultaneously, this also demonstrates the flexibility and scalability of 5G networks in supporting new services and applications.

[0108] After receiving the capability reports from the terminal, the base station will further process this information and report it to the OMC (Operation and Maintenance Center) through the AI ​​capability interface, so as to make subsequent network management, optimization and AI model deployment decisions.

[0109] OMC first receives target cell information reported from the base station. After parsing the target cell information, it matches the parsed target cell information with its own maintained AI model database. The database stores detailed information about various AI models, including model name, version, parameters, etc.

[0110] Through the matching process, the OMC can determine whether a base station has deployed a specific AI model. If the AI ​​model information in the target cell matches a record in the database, it can be confirmed that the base station has deployed the AI ​​model. Conversely, if no match is found, it can be assumed that the base station has not yet deployed the model.

[0111] Based on the determined results, the OMC can then perform further actions. For example, if the base station has already deployed the required AI model, the OMC may continue to monitor the model's performance metrics; if the base station has not yet deployed it, the OMC may trigger the model deployment process or issue an alert to the network administrator so that necessary measures can be taken in a timely manner.

[0112] This application utilizes OMC to analyze target information and match it with its own maintained AI model database, enabling accurate determination of whether a base station has deployed the AI ​​model. This process is of great significance for achieving intelligent network management and optimization.

[0113] Optionally, the target information element includes:

[0114] Terminal chip performance parameters;

[0115] List of AI models supported by the terminal;

[0116] AI model parameters and metrics have been deployed on the terminal.

[0117] In this embodiment, the terminal chip performance metrics are used to describe the chip's processing power, power consumption, and AI-related computing performance, such as operations per second (OPS) and computing performance per watt (OPS / W). In actual communication processes, these metrics can help the base station or network side understand the terminal's hardware capabilities, thereby making more reasonable resource allocation and scheduling decisions.

[0118] List of AI Models Supported by the Terminal: This list includes the types and versions of AI models that the terminal can support and run. With the continuous development of AI technology, more and more AI models are being applied to terminal devices, such as smartphones and smart wearable devices. Through this list, the network side can understand the range of AI capabilities of the terminal and then push suitable AI services or applications.

[0119] AI model parameters and metrics deployed on the terminal: These parameters and metrics describe the specific information of the AI ​​models already deployed on the terminal, including the model name, version, computational complexity (such as FLOPs), memory usage (such as MACs), etc.

[0120] Target information is crucial for the network side because it helps understand the AI ​​model currently being used by the terminal, enabling more effective network management and optimization. For example, based on the computational needs of the deployed AI model, the network side can dynamically adjust the terminal's transmission power or data rate to ensure the stability and efficiency of AI services.

[0121] The target information in this application covers the performance indicators of the terminal chip, the list of supported AI models, and the parameters and indicators of the deployed AI models, providing technical support for optimizing base station AI models.

[0122] Optionally, after determining whether the base station has deployed an artificial intelligence (AI) model based on the target information, the method further includes:

[0123] In the absence of AI models deployed on base stations and terminals, preset indicator triggering conditions can be set using the task orchestration module.

[0124] When the base station meets the preset indicator triggering condition, the network information corresponding to the preset indicator triggering condition in the base station is collected using the preset interface;

[0125] Based on the network information, suitable target AI models for deployment are selected, and it is determined whether the current computing power resources of the base station meet the deployment requirements of the target AI models.

[0126] If the deployment requirements are met, a data collection task is issued through the preset interface; the data collection task is used to instruct the collection of training data required by the target AI model.

[0127] After preprocessing the target data acquired by the data acquisition task, the processed data is sent to the AI ​​model resource pool for model training to determine the first model index parameters after training.

[0128] The first model index parameters are deployed and verified, and the first target parameters after successful deployment and verification are sent to the base station.

[0129] In this embodiment, after determining whether the base station has deployed an AI model based on the target information, and if it is determined that neither the base station nor the terminal has deployed an AI model, firstly, preset indicator trigger conditions are set: using the task orchestration module, a series of preset indicator trigger conditions are set according to network conditions, service requirements, or specific scenarios. These conditions may include base station load, data transmission rate, user request frequency, etc., to determine whether the base station needs to deploy an AI model.

[0130] Secondly, network information is collected: when a base station meets preset indicator trigger conditions, relevant network information is automatically collected from the base station through a preset interface. This information may include the current load status of the base station, user behavior data, network transmission quality, etc., providing a basis for subsequent selection of suitable AI models for deployment.

[0131] Next, target AI models are selected and computing resources are evaluated: Based on the collected network information and the model information in the AI ​​model resource pool, suitable target AI models for deployment are selected. Simultaneously, it is assessed whether the current computing resources of the base station meet the deployment requirements of the target AI model. This step requires comprehensive consideration of factors such as the model's computational complexity, memory usage, and the base station's hardware performance.

[0132] Then, a data collection task is issued: if the base station's computing resources meet the deployment requirements, a data collection task is issued through a preset interface. This task instructs the base station or related systems to collect the training data required by the target AI model, providing data support for subsequent model training.

[0133] Subsequently, data preprocessing and model training occur: the target data acquired through the data collection task is preprocessed, including data cleaning, format conversion, and normalization. Then, the processed data is sent to the AI ​​model resource pool for model training. Through the training process, the first set of model metrics parameters are determined; these parameters reflect the model's performance and effectiveness.

[0134] Finally, deployment verification and target parameter distribution: The first model metric parameters are deployed and verified to ensure the stability and reliability of the model in a real-world environment. If deployment verification is successful, the first target parameters are distributed to the base station. These parameters may include model weights, biases, hyperparameters, etc., and are used for the actual deployment and operation of the AI ​​model on the base station.

[0135] Optionally, the base station can send the first model indicator parameters to the terminal, allowing the terminal to synchronously update the current AI model on the terminal side.

[0136] By following the steps above, AI models can be selected, trained, deployed, and validated even without their deployment on base stations and terminals, thereby meeting business and network optimization needs in specific scenarios.

[0137] It should be noted that the step of retraining the AI ​​model when it is deployed on the base station and terminal, and when the AI ​​model does not provide a positive gain for the wireless network, includes:

[0138] In the AI ​​model resource pool, suitable target AI models are selected for deployment based on the application scenarios required by the base station and terminal. It is then determined whether the base station's current computing power resources meet the deployment requirements of the target AI model. If the deployment requirements are met, a data acquisition task is issued through a preset interface. This data acquisition task instructs the collection of training data required by the target AI model. After preprocessing the target data acquired by the data acquisition task, the processed data is sent to the AI ​​model resource pool for model training to determine the second model indicator parameters after training. The second model indicator parameters are then deployed and verified, and the successfully verified second target parameters are sent to the base station. The base station uses the second model indicator parameters to update its current AI model and continues to monitor newly deployed AI models on the base station, initiating a new round of evaluation and optimization cycles.

[0139] Optionally, the base station can send the second model indicator parameters to the terminal, allowing the terminal to synchronously update the current AI model on the terminal side.

[0140] To further understand the above method, this application provides two specific embodiments.

[0141] In the first specific embodiment, the modules processed sequentially in the OMC are: task orchestration module, AI model resource pool, deployment verification module, and data center module. The specific steps are as follows: The base station reports its current performance and service data through the AI ​​capability interface, requesting the task orchestration center to verify whether the current base station can deploy the relevant AI model. If not, no further processing is performed; if so, a data collection task is issued through the AI ​​capability interface using task orchestration. The data collected by the base station is preprocessed by the data center and then enters the AI ​​model resource pool for model training. After deployment verification, the trained model parameters are issued to the base station in the form of a policy through the AI ​​capability interface. The base station and the terminal then deploy the relevant model.

[0142] Specifically, refer to Figure 2 and Figure 3 As shown, the existing network base station terminals do not have AI models deployed (taking the CSI compressed model as an example):

[0143] (1) Task orchestration settings for wireless and performance metrics triggering mechanisms: The task orchestration center first sets a series of triggering mechanisms for wireless and performance metrics. These metrics include, but are not limited to, RSRP (Reference Signal Received Power), SINR (Signal-to-Interference-plus-Noise Ratio), latency, and throughput. When these metrics are below or above the set thresholds, the base station will report the current network metrics, the list of supported AI models, and the remaining computing resources through the AI ​​capability interface.

[0144] For example, when the base station's air interface resource load is too high, and the number of RRC (Radio Resource Control) connections and the utilization rate of PRB (Physical Resource Block) both exceed the CSI (Channel State Information) compression model trigger threshold set by the task orchestration center (such as RRC connection count > 100 or PRB utilization rate > 60%), the base station will automatically report these network indicators, as well as the list of supported AI models (such as TransNet, CsiNet, etc.) and the current memory status.

[0145] (2) Selecting suitable AI models for deployment and assessing computing resources: After receiving the data indicators reported by the base station, the task orchestration center will select suitable AI models for deployment based on this information. At the same time, it will also assess whether the current computing resources of the base station meet the deployment requirements of the corresponding AI model.

[0146] If the base station's computing resources meet the requirements, the task orchestration center will issue a data collection task through the AI ​​capability interface to collect the data needed for AI model training. If the base station's computing resources do not meet the requirements, the current deployment request will be rejected.

[0147] For example, if the current base station's computing resources only support deployment of a model like "TransNet", the task orchestration center will send the parameters of the "TransNet model" (such as NMSE, flops, etc.) to the base station side through the AI ​​capability interface.

[0148] (3) Data Preprocessing and Model Training: The collected data undergoes preprocessing such as cleaning and merging in the data center to remove redundant and invalid data, and is merged or decomposed according to relevant algorithm requirements. After preprocessing, the data is pushed to the AI ​​model resource pool for model training. After training, the relevant model metrics (such as neural network parameters: neuron weights, network structure, etc.) are sent to the deployment and verification center for subsequent verification.

[0149] For example, the CSI matrix of the current channel is collected and sent to the data center for preprocessing, and then pushed to the AI ​​model resource pool for training. After training is completed, the model metrics are sent to the deployment and verification center.

[0150] (4) Deployment Verification and Model Distribution: The deployment verification center will verify the parameters of the trained model and remove objective functions with small weight factors to reduce the overhead of the model. At the same time, it will also check the computing power requirements of the model to ensure that the model can be successfully deployed on the base station.

[0151] If the model meets all requirements, the deployment and verification center will distribute the model to the base station in the form of a policy through the AI ​​capability interface. After receiving the model, the base station will deploy the relevant model and interact with the terminal through RRC signaling to exchange model parameters and indicators, thus entering the live network application stage.

[0152] For example, a deployment verification center verifies the network structure of the TransNet model and performs appropriate network pruning based on the memory availability of the base station or terminal. The pruned model is then distributed to the base station for deployment via an AI capability interface.

[0153] In the second specific embodiment, the modules processed sequentially in the OMC are: task orchestration module, AI capability interface, evaluation center module, AI model resource pool, and deployment verification module. The specific steps are as follows: Task orchestration periodically collects AI data through the AI ​​capability interface; the base station reports relevant collected datasets and data center performance service data to the evaluation center; the network performance status after deploying the AI ​​model is evaluated, and the gain brought to the network by the AI ​​model is given based on corresponding indicator thresholds, and the results are fed back to task orchestration; based on the feedback results, task orchestration archives positive gains and sends negative gain AI models for retraining; after deployment verification, the retrained AI model is sent to the base station in the form of a policy through the AI ​​capability interface.

[0154] Specifically, refer to Figure 2 and Figure 4 As shown, the AI ​​model is deployed on the existing network base station terminal (taking the CSI compression model as an example):

[0155] (1) The task orchestration module sets up timed tasks and collects base station data of currently deployed AI models through the AI ​​capability interface. The inference results of the current model include, but are not limited to: NMSE, FLOPS, SGCS, number of transmit and receive antennas between base station and terminal, and wireless network indicators including, but not limited to, throughput, CQI, RANK, RSRP, SINR, packet loss rate, latency, etc.

[0156] For example: The task orchestration collected the following CSI compression model parameters deployed on the current terminal side: NMSE, FLOPS, SGCS, base station transceiver antenna count 2T4R, terminal receiving antenna count: 1, current network CQI: 12, RSRP: -105db.

[0157] (2) After the collected data is integrated by the data center module, it enters the evaluation center module. The evaluation center module evaluates whether the current network performance triggers each preset threshold, such as: RSRP is lower than XXX and throughput is lower than XXX. If the threshold requirements are met (a multi-objective optimization function is set, and each sub-objective is the corresponding performance service threshold; the optimization algorithm is not described here), then the current AI model generates a positive gain for the network. The task orchestration storage records the model inference results and the corresponding network performance indicators. Subsequent base stations with similar application scenarios can carry out intelligent deployment.

[0158] For example, the evaluation threshold set by the evaluation center for the CSI compression model is: CQI>10 and RSRP>-110db. It can be seen that the current CSI compression model meets the evaluation center threshold, and it is determined that the current CSI compression model produces a positive gain to the network. Then, the current model inference results (neuron weights, network structure, NMSE, RSRP, CQI, etc.) are recorded.

[0159] (3) If the relevant threshold requirements are not met, the model will be re-entered into the AI ​​model resource pool for model switching and retraining. The retrained model parameters will be deployed and verified by the module.

[0160] (4) Deploy and verify the model parameters, remove objective functions with small weight factors, reduce the overhead of the model, check the model computing power requirements, and if the requirements are met, issue them to the base station in the form of a policy through the AI ​​capability interface. The base station deploys the relevant model and interacts with the terminal through RRC signaling to exchange model parameter indicators, and then enters the live network application.

[0161] Compared with existing technical solutions, this application demonstrates significant technical advantages in AI model management and wireless network optimization:

[0162] (1) Completeness of AI Model Management Process: This application proposes a complete AI model management process, including model evaluation, policy distribution, inspection mechanism, and model switching and retraining. The establishment of this process makes the application of AI models in wireless networks more standardized and systematic, effectively avoiding the continuous negative gains caused by unsuitable models. In contrast, existing technologies may lack such a complete management process, resulting in more arbitrary model selection, deployment, and optimization processes, making it difficult to guarantee stable improvement in network performance.

[0163] (2) This application proposes a new southbound interface for uplink and downlink transmission of AI data and transmission of management metrics for AI tasks. The southbound interface refers to the data interface used to manage base stations, enabling centralized monitoring and statistics of information reported by base stations. This interface design not only enriches the data transmission capabilities of wireless networks but also provides strong support for the flexible application of AI models in wireless networks. Existing interfaces may not fully meet the needs of AI data and management metric transmission, limiting the optimization effect of AI models in wireless networks.

[0164] (3) By introducing an AI model management process, this application ensures that the most suitable AI model is always deployed in the wireless network, thereby effectively improving network stability and reliability. For operators, this means fewer network failures and higher user satisfaction. The automated AI model management process reduces the need for manual intervention and lowers the operator's operating costs. At the same time, by accurately predicting and selecting the optimal model, unnecessary resource waste can also be avoided.

[0165] The various methods of the embodiments of this application have been described above. Apparatus for implementing the above methods will now be provided.

[0166] Please refer to Figure 5 This application also provides a wireless network artificial intelligence model management device, applied to an Operation and Maintenance Center (OMC), comprising:

[0167] The first processing module 51 is used to collect the indicator data of the base station when an artificial intelligence (AI) model has been deployed in the base station and the terminal.

[0168] The second processing module 52 is used to evaluate whether the AI ​​model has a positive gain for the wireless network based on the collected index data.

[0169] The third processing module 53 is used to save the index data corresponding to the AI ​​model when the AI ​​model has a positive gain for the wireless network.

[0170] The fourth processing module 54 is used to retrain the AI ​​model when the AI ​​model does not have a positive gain for the wireless network, and to deploy the retrained AI model to the base station and the terminal.

[0171] Optionally, the device further includes:

[0172] The receiving module is used to receive target cell information reported by the base station using a preset interface;

[0173] The first determining module is used to determine whether the base station has deployed the AI ​​model based on the target information.

[0174] Optionally, the target information element includes:

[0175] Terminal chip performance parameters;

[0176] List of AI models supported by the terminal;

[0177] AI model parameters and metrics have been deployed on the terminal.

[0178] Optionally, the device further includes:

[0179] The fifth processing module is used to set preset indicator trigger conditions using the task orchestration module when no artificial intelligence (AI) model is deployed on the base station and terminal.

[0180] The sixth processing module is used to collect network information in the base station corresponding to the preset indicator triggering condition using the preset interface when the base station meets the preset indicator triggering condition.

[0181] The second determining module is used to filter suitable target AI models for deployment based on the network information and determine whether the current computing power resources of the base station meet the deployment requirements of the target AI model.

[0182] The third determining module is used to issue a data collection task through the preset interface if the deployment requirements are met; the data collection task is used to instruct the collection of training data required by the target AI model.

[0183] The fourth determining module is used to preprocess the target data acquired by the data acquisition task, send the processed data to the AI ​​model resource pool for model training, and determine the first model index parameters after training.

[0184] The seventh processing module is used to deploy and verify the first model index parameters, and send the first target parameters after successful deployment and verification to the base station.

[0185] Optionally, the second processing module 52 described above includes:

[0186] The first processing unit is used to evaluate whether the performance indicators in the collected indicator data meet the preset threshold conditions corresponding to the performance indicators.

[0187] The second processing unit is used to determine that the AI ​​model has a positive gain for the wireless network if each parameter in the performance index meets the corresponding preset threshold condition.

[0188] The third processing unit is used to determine that the AI ​​model does not have a positive gain for the wireless network if at least one of the performance indicators does not meet the corresponding preset threshold condition.

[0189] Optionally, the preset interface is used to transmit at least one of the following information:

[0190] Terminal chip capability parameters;

[0191] AI model coefficients, AI model weights, and a list of AI models supported by the base station and / or terminal;

[0192] AI model evaluation metrics;

[0193] AI Task ID;

[0194] AI task strategy.

[0195] Optionally, the OMC includes an evaluation center module, an AI model resource pool module, a task orchestration module, and a deployment verification module, and the device includes the ability to perform at least one of the following functions:

[0196] Using the evaluation center module, performance indicator thresholds, model evaluation indicator thresholds, and business indicator thresholds are set for different business scenarios that support AI models;

[0197] The AI ​​model resource pool module is used to train different AI models and neural networks, and to centrally process the training tasks of base station and terminal AI models.

[0198] The task orchestration module is used to store the AI ​​model capability list of each base station, maintain the AI ​​tasks of each base station, and periodically inspect the deployment data of the AI ​​models of each base station.

[0199] Using the deployment verification module, the trained model is evaluated for computing power based on the base stations to be deployed, and the model weights are pruned.

[0200] It should be noted that the device in this embodiment corresponds to the method applied to the OMC side described above. The implementation methods in the above embodiments are all applicable to the embodiments of this device and can achieve the same technical effect. The device provided in this application embodiment can implement all the method steps implemented in the above method embodiments and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiments and the beneficial effects will not be described in detail.

[0201] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the above-described wireless network artificial intelligence model management method embodiments and achieves the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0202] This application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they implement the various processes of the above-described wireless network artificial intelligence model management method embodiment and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0203] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0204] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0205] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A wireless network artificial intelligence model management method, characterized in that, Applications in Operations and Maintenance Centers (OMCs) include: With AI models already deployed on the base station and terminal, collect the indicator data of the base station; Based on the collected indicator data, evaluate whether the AI ​​model has a positive gain for the wireless network; If the AI ​​model has a positive gain for the wireless network, save the index data corresponding to the AI ​​model; If the AI ​​model does not provide a positive gain for the wireless network, the AI ​​model is retrained, and the retrained AI model is deployed to the base station and terminal according to the required application scenario.

2. The method of claim 1, wherein, When an artificial intelligence (AI) model has been deployed on the base station and the terminal, before collecting the indicator data of the base station, the method further includes: Receive target cell information reported by the base station using a preset interface; Based on the target information, determine whether the base station has deployed the AI ​​model.

3. The method of claim 2, wherein, The target information element includes: Terminal chip performance parameters; List of AI models supported by the terminal; AI model parameters and metrics have been deployed on the terminal.

4. The method of claim 2, wherein, After determining whether the base station has deployed an artificial intelligence (AI) model based on the target information, the method further includes: In the absence of AI models deployed on base stations and terminals, preset indicator triggering conditions can be set using the task orchestration module. When the base station meets the preset indicator triggering condition, the network information corresponding to the preset indicator triggering condition in the base station is collected using the preset interface; Based on the network information, suitable target AI models for deployment are selected, and it is determined whether the current computing power resources of the base station meet the deployment requirements of the target AI models. If the deployment requirements are met, a data collection task is issued through the preset interface; the data collection task is used to instruct the collection of training data required by the target AI model. After preprocessing the target data acquired by the data acquisition task, the processed data is sent to the AI ​​model resource pool for model training to determine the first model index parameters after training. The first model index parameters are deployed and verified, and the first target parameters after successful deployment and verification are sent to the base station.

5. The method according to claim 1, characterized in that, Based on the collected indicator data, the evaluation of whether the AI ​​model has a positive gain for the wireless network includes: Based on the performance indicators in the collected indicator data, evaluate whether the performance indicators meet the preset threshold conditions corresponding to the performance indicators. If each parameter in the performance index meets the corresponding preset threshold condition, then the AI ​​model is determined to have a positive gain for the wireless network. If at least one of the performance metrics fails to meet the corresponding preset threshold condition, then the AI ​​model is determined to have no positive gain for the wireless network.

6. The method according to claim 2, characterized in that, The preset interface is used to transmit at least one of the following information: Terminal chip capability parameters; AI model coefficients, AI model weights, and a list of AI models supported by the base station and / or terminal; AI model evaluation metrics; AI Task ID; AI task strategy.

7. The method according to claim 1, characterized in that, The OMC includes an evaluation center module, an AI model resource pool module, a task orchestration module, and a deployment and verification module. The method includes performing at least one of the following functions: Using the evaluation center module, performance indicator thresholds, model evaluation indicator thresholds, and business indicator thresholds are set for different business scenarios that support AI models; The AI ​​model resource pool module is used to train different AI models and neural networks, and to centrally process the training tasks of base station and terminal AI models. The task orchestration module is used to store the AI ​​model capability list of each base station, maintain the AI ​​tasks of each base station, and periodically inspect the deployment data of the AI ​​models of each base station. Using the deployment verification module, the trained model is evaluated for computing power based on the base stations to be deployed, and the model weights are pruned.

8. A wireless network artificial intelligence model management device, characterized in that, Applications in Operations and Maintenance Centers (OMCs) include: The first processing module is used to collect the indicator data of the base station when an artificial intelligence (AI) model has been deployed in the base station and the terminal. The second processing module is used to evaluate whether the AI ​​model has a positive gain for the wireless network based on the collected indicator data. The third processing module is used to save the index data corresponding to the AI ​​model when the AI ​​model has a positive gain for the wireless network. The fourth processing module is used to retrain the AI ​​model when the AI ​​model does not have a positive gain for the wireless network, and to deploy the retrained AI model to the base station and the terminal.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 7.

Images