Model generation method, apparatus, and storage medium

Abstract

The present application relates to the technical field of communications, and discloses a model generation method, an apparatus, and a storage medium. The method is applied to a Near-RT RIC and comprises: receiving AI model requirement information reported by a user equipment, and generating an AI global model parameter on the basis of the AI model requirement information; sending the AI global model parameter to the user equipment, wherein the AI global model parameter is used for the user equipment to generate an AI model training parameter after using a model training strategy to train the AI global model parameter; receiving the AI model training parameter sent by the user equipment, and generating an AI model file of a target AI model on the basis of the AI model training parameter and the AI global model parameter; and sending model management information of the target AI model and the AI model file of the target AI model to a Non-RT RIC.

Description

A model generation method, apparatus and storage medium

[0001] Cross-reference to related applications

[0002] This application claims priority to Chinese Patent Application No. 202411455474.5, filed on October 17, 2024, entitled "A Model Generation Method, Apparatus and Storage Medium", the entire contents of which are incorporated herein by reference. Technical Field

[0003] This application relates to the field of communication technology, and in particular to a model generation method, apparatus and storage medium. Background Technology

[0004] With the rapid development of mobile communication technology, research on the integration of Artificial Intelligence (AI) models and Radio Access Networks (RANs) has gradually moved from theory to practice. This research aims to combine the powerful representation and feature extraction capabilities of AI models with wireless communication scenarios to achieve time-frequency domain resource optimization and cost reduction. This research direction is of great significance for improving the overall performance of wireless communication systems.

[0005] However, with the increasing demand for AI services from user equipment (UE), customized AI models have gained widespread attention. In the future, communication and AI integration scenarios, more and more users will be provided with more professional and personalized AI services to meet their diverse performance requirements. Summary of the Invention

[0006] This application provides a model generation method, apparatus, and storage medium to meet users' customized AI service needs.

[0007] Firstly, a model generation method is provided, applied to a near real-time wireless access network intelligent controller, the method comprising:

[0008] Receive AI model requirement information reported by user devices, and generate global AI model parameters based on the AI ​​model requirement information;

[0009] Send the AI ​​global model parameters to the user equipment; wherein, the AI ​​global model parameters are used by the user equipment to generate AI model training parameters after training the AI ​​global model parameters using a model training strategy;

[0010] Receive the AI ​​model training parameters sent by the user equipment, and generate the AI ​​model file of the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters;

[0011] The model management information of the target AI model and the AI ​​model file are sent to the Non-Real-Time Access Network Intelligent Management Controller; wherein, the model management information is used by the Non-RT RIC to manage the target AI model.

[0012] In some embodiments, receiving the AI ​​model training parameters sent by the user equipment and generating an AI model file for the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters includes:

[0013] In the presence of multiple user devices, the AI ​​model training parameters sent by each of the multiple user devices are received, and a model aggregation operation is performed on the gradients of each AI model training parameter to obtain the aggregated AI model training parameters; wherein, the AI ​​model training parameters are local model parameters or local model gradients.

[0014] Based on the aggregated AI model training parameters and the AI ​​global model parameters, the current target AI global model parameters are generated;

[0015] If the target AI model generated based on the target AI global model parameters is found to meet the model training termination condition, then the AI ​​model file of the target AI model is generated based on the target global model parameters.

[0016] In some embodiments, the model management information is stored in a model management list built by the Non-RT RIC, including at least the model application scenario, model revenue, model cost, model file storage location of the target AI model, and the cell identifier associated with the target AI model. The model management list is built based on the model identifiers and / or model functions of multiple AI models, and the AI ​​model file is stored on the cloud server associated with the Non-RT RIC.

[0017] In some embodiments, after generating the target AI model parameters based on the AI ​​model training parameters and the AI ​​global model parameters, the method further includes:

[0018] Send a model deployment request to the Non-RT RIC; wherein the model deployment request carries AI model requirement information sent by the user equipment, which is used by the Non-RT RIC to determine K target AI models that match the AI ​​model requirement information from the model management list, where K is an integer greater than 0;

[0019] Receive model deployment strategies corresponding to each of the K target AI models from the Non-RT RIC; wherein, the model deployment strategies corresponding to each of the K target AI models are generated at least based on the model management information corresponding to each of the K target AI models;

[0020] Based on the storage location of the model files corresponding to the K target AI models, download the AI ​​model files corresponding to the K target AI models from the cloud server;

[0021] The model deployment strategies and AI model files corresponding to the K target AI models are encapsulated to obtain model deployment strategy signaling, and the model deployment strategy signaling is sent to the user equipment; wherein, the model deployment strategy signaling is used to instruct the user equipment to deploy models based on the model deployment strategies and AI model files corresponding to the K target AI models.

[0022] In some embodiments, the method further includes:

[0023] In response to a detected model processing trigger event of the target AI model, a model processing instruction for the target AI model is sent to the user equipment;

[0024] Wherein, if the model processing trigger event is a model update, the model processing instruction includes at least a model update mode and a model update method. The model update mode is used to instruct the user device to update the target AI model using an offline update mode or an online update mode, and the model update method is used to instruct the user device to update the target AI model using a local model update or a global model update. If the model processing trigger event is a model uninstallation, the model processing instruction includes at least a model uninstallation method. The model uninstallation method is used to instruct the user device to uninstall the target AI model using a soft uninstallation method or a hard uninstallation method.

[0025] In some embodiments, the AI ​​model requirements information includes at least:

[0026] The identification information of the user equipment;

[0027] The local data format indication configured in the user equipment;

[0028] Model application scenario indication;

[0029] The privacy protection level of the user device;

[0030] Model performance gain requirement indication;

[0031] Model service quality requirements indication;

[0032] The computing power indication of the user equipment;

[0033] The user equipment requests a timeliness indication of the model.

[0034] In this embodiment, the system first receives AI model requirement information reported by the user device, generates global AI model parameters based on the AI ​​model requirement information, and sends the global AI model parameters to the user device. Then, the user device uses a model training strategy to train the global AI model parameters to generate AI model training parameters, enabling model training to be completed on the user device, effectively protecting user privacy. Finally, after receiving the AI ​​model training parameters sent by the user device, the system generates an AI model file of the target AI model based on the AI ​​model training parameters and the global AI model parameters, and sends the model management information of the target AI model and the AI ​​model file to the Non-RT RIC, facilitating the Non-RT RIC to manage the target AI model, thereby meeting the user's customized AI service needs and improving the user experience.

[0035] Secondly, a model generation method is provided for application in a user device, the method comprising:

[0036] Report AI model requirements to Near-RT RIC;

[0037] Receive AI global model parameters from the Near-RT RIC; wherein the AI ​​global model parameters are generated by the Near-RT RIC based on the AI ​​model requirement information;

[0038] The AI ​​global model parameters are trained using a model training strategy to obtain the AI ​​model training parameters;

[0039] The AI ​​model training parameters are sent to the Near-RT RIC; wherein, the AI ​​model training parameters are used by the Near-RT RIC to generate an AI model file of the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters.

[0040] In some embodiments, the method further includes:

[0041] The model deployment request is sent to the Non-RT RIC via the Near-RT RIC; wherein the model deployment request is used by the Non-RT RIC to determine K target AI models that match the AI ​​model requirement information from the model management list, the model management list is constructed based on the model identifiers and / or model functions of multiple AI models, and K is an integer greater than 0;

[0042] Receive model deployment signaling from the Near-RT RIC; wherein the model deployment signaling is encapsulated based on the model deployment strategy and AI model file corresponding to each of the K target AI models, and the model deployment strategy is generated by the Non-RT RIC based on the model management information corresponding to each of the K target AI models recorded in the model management list;

[0043] The model deployment operation is performed based on the model deployment strategy and AI model file corresponding to each of the K target AI models.

[0044] In some embodiments, after performing the model deployment operation based on the model deployment strategy and AI model file corresponding to each of the K target AI models, the method further includes:

[0045] The K target AI models are used to infer the t-th data sample in the sample set to obtain K inference results for the t-th data sample; where t is an integer greater than 0.

[0046] The target inference result of the t-th data sample is determined based on the K inference results, and the ensemble weights of the K target AI models are updated based on the target inference result and the loss functions of the K target AI models.

[0047] In some embodiments, the method further includes:

[0048] Receive model processing instructions for the target AI model sent by the Near-RT RIC; wherein, the model processing instructions are sent by the Near-RT RIC when it detects a model processing trigger event for the target AI model;

[0049] Perform model processing operations on the target AI model according to the model processing instructions;

[0050] Wherein, if the model processing trigger event is a model update, the model processing instruction includes at least a model update mode and a model update method. The model update mode is used to instruct the user device to update the target AI model using an offline update mode or an online update mode, and the model update method is used to instruct the user device to update the target AI model using a local model update or a global model update. If the model processing trigger event is a model uninstallation, the model processing instruction includes at least a model uninstallation method. The model uninstallation method is used to instruct the user device to uninstall the target AI model using a soft uninstallation method or a hard uninstallation method.

[0051] In some embodiments, the AI ​​model requirements information includes at least:

[0052] The identification information of the user equipment;

[0053] The local data format indication configured in the user equipment;

[0054] Model application scenario indication;

[0055] The privacy protection level of the user device;

[0056] Model performance gain requirement indication;

[0057] Model service quality requirements indication;

[0058] The computing power indication of the user equipment;

[0059] The user equipment requests a timeliness indication of the model.

[0060] In this embodiment, the user equipment can report AI model requirement information to the Near-RT RIC, which facilitates the Near-RT RIC to generate suitable global AI model parameters based on the AI ​​model requirement information. After receiving the global AI model parameters, the user equipment can adopt a model training strategy to train the global AI model parameters locally to obtain AI model training parameters, thus protecting the user equipment's data privacy. Then, the user equipment sends the AI ​​model training parameters to the Near-RT RIC. Subsequently, the Near-RT RIC generates the AI ​​model file of the target AI model based on the AI ​​model training parameters and the global AI model parameters, thereby meeting the user's customized AI service needs and improving the user experience.

[0061] Thirdly, a real-time wireless access network intelligent controller is provided, comprising:

[0062] The receiving module is used to receive AI model requirement information reported by the user device and generate global AI model parameters based on the AI ​​model requirement information.

[0063] The first sending module is used to send the AI ​​global model parameters to the user equipment; wherein, the AI ​​global model parameters are used by the user equipment to generate AI model training parameters after training the AI ​​global model parameters using a model training strategy;

[0064] The generation module is used to receive the AI ​​model training parameters sent by the user equipment, and generate the AI ​​model file of the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters;

[0065] The second sending module is used to send the model management information of the target AI model and the AI ​​model file to the Non-RT RIC; wherein, the model management information is used by the Non-RT RIC to manage the target AI model.

[0066] Fourthly, a user equipment is provided, comprising:

[0067] The first transmitting module is used to report AI model requirement information to the Near-RT RIC intelligent controller of the near real-time wireless access network.

[0068] The receiving module is used to receive AI global model parameters from the Near-RT RIC; wherein the AI ​​global model parameters are generated based on the AI ​​model requirement information;

[0069] The training module is used to train the global AI model parameters using a model training strategy to obtain the AI ​​model training parameters.

[0070] The second sending module is used to send the AI ​​model training parameters to the Near-RT RIC; wherein, the AI ​​model training parameters are used by the Near-RT RIC to generate an AI model file of the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters.

[0071] Fifthly, a communication device is provided, comprising:

[0072] A memory for storing computer programs; a processor for executing the computer programs stored in the memory to implement the method described in either the first aspect or the second aspect.

[0073] A sixth aspect provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of either the first aspect or the second aspect. Attached Figure Description

[0074] Figure 1 is a schematic diagram of the application scenarios applicable to the embodiments of this application;

[0075] Figure 2 is a flowchart of a model generation method provided in an embodiment of this application;

[0076] Figure 3 is a flowchart of a model deployment provided in an embodiment of this application;

[0077] Figure 4 is a flowchart of another model generation method provided in an embodiment of this application;

[0078] Figure 5 is a signaling interaction diagram of model generation and model deployment provided in an embodiment of this application;

[0079] Figure 6 is a schematic diagram of a Near-RT RIC provided in an embodiment of this application;

[0080] Figure 7 is a schematic diagram of the structure of a user equipment provided in an embodiment of this application;

[0081] Figure 8 is a schematic diagram of the structure of a communication device provided in an embodiment of this application. Detailed Implementation

[0082] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0083] The terms "first" and "second" in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the term "comprising" and any variations thereof are intended to cover non-exclusive protection. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. The term "multiple" in this application can mean at least two, for example, two, three, or more, and the embodiments of this application do not impose limitations.

[0084] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These embodiments should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description. It should be noted that in the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solutions of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0085] To better understand the embodiments of this application, the technical terms involved in the embodiments of this application will be explained below.

[0086] (1) The Near-RT Intelligent Controller (Near-RT RIC) is an important functional unit defined by the Open Radio Access Network (O-RAN) Alliance for intelligent radio access. For example, it can collect base station data and generate strategies using artificial intelligence algorithms. It is a near-real-time radio resource control device for base stations at the millisecond level. It has functions such as radio resource management, intelligent control, low-latency control, and open architecture.

[0087] (2) The non-real-time wireless access network intelligent controller is a functional unit used to realize the intelligent control of the non-real-time wireless access network. For example, it is responsible for the analysis and processing of cross-domain network-level, multi-dimensional, and ultra-large-scale data, and supports policy management and control at the second level or above. It has functions such as service and intent policy management, wireless network analysis, and artificial intelligence (AI) model training.

[0088] (3) Lifecycle Management (LCM) is an important concept proposed in the current 3GPP standardization research. As a comprehensive management solution covering the entire process of AI models, LCM includes data collection, model training, model inference, model management (e.g., selecting model activation / deactivation, switching, rollback, monitoring, updating, and uninstallation based on model functions or model identifiers), and model transfer. In the concept of LCM, model generation and model deployment are two important steps. Model generation refers to generating a series of AI models according to the configuration for performing model training and model verification processes; model deployment refers to deploying the generated models locally or in the cloud for performing operations such as model inference; and the importance of AI model lifecycle management lies in its ability to ensure the quality and efficiency of AI models throughout their entire lifecycle, reducing potential usage costs.

[0089] (4) The Physical Downlink Shared Channel (PDSCH) can be used to transmit user data, transmit control information, etc.

[0090] (5) The Cell-Radio Network Temporary Identifier (C-RNTI) is a dynamic identifier assigned to user equipment by the base station. The C-RNTI uniquely identifies a user equipment under the cell air interface, and the C-RNTI is only valid for user equipment in the connected state.

[0091] (6) The Random Access Radio Network Temporary Identifier (RA-RNTI) is used to identify the user equipment performing the random access procedure, ensuring that each UE is unique when attempting to access the network, thereby avoiding conflicts.

[0092] The following is a brief introduction to the application scenarios to which the technical solutions of the embodiments of this application are applicable. It should be noted that the application scenarios described below are only for illustrating the embodiments of this application and are not intended to limit the scope. In specific implementation, the technical solutions provided by the embodiments of this application can be flexibly applied according to actual needs.

[0093] Figure 1 is a schematic diagram of an application scenario applicable to the embodiments of this application. As shown in the figure, the application scenario mainly includes: user equipment (1,2,3…,N) and base station 11. Among them, user equipment (1,2,3…,N) and base station 11 can exchange information through a communication network. The communication network can adopt communication methods including wireless communication and wired communication.

[0094] In some scenarios, user equipment (1,2,3…,N) can access the network and communicate with base station 11 through cellular mobile communication technology, which may include 5th generation mobile network (5G) technology, 6G technology, etc.

[0095] In some scenarios, user equipment (1,2,3…,N) can access the network and communicate with base station 11 through short-range wireless communication. This short-range wireless communication method may include Wireless Fidelity (Wi-Fi) technology.

[0096] User equipment (1,2,3…,N) is a device that can provide voice and / or data connectivity to a user, including: handheld user equipment with wireless connectivity, vehicle-mounted terminal equipment, etc. For example, user equipment (1,2,3…,N) includes, but is not limited to: mobile phones, tablets, laptops, PDAs, mobile internet devices (MIDs), wearable devices, virtual reality (VR) devices, augmented reality (AR) devices, wireless terminal equipment in industrial control, wireless terminal equipment in autonomous driving, wireless terminal equipment in smart grids, wireless terminal equipment in transportation safety, wireless terminal equipment in smart cities, or wireless terminal equipment in smart homes, etc.

[0097] Base station 11 includes a centralized unit (CU), a distributed unit (DU), etc. The CU may include a user plane (CU-UP) responsible for processing user data, and a control plane (CU-CP) responsible for interface management, system information management, etc. Furthermore, base station 11 may deploy a Near-RT RIC, which can be connected to the CU via an E2 interface and is responsible for controlling radio resources within the coverage area of ​​base station 11. Base station 11 may also deploy a Non-RT RIC, which can be specifically deployed on the radio network management system. It should be noted that a base station can deploy one or more Near-RT RICs; this embodiment does not impose any limitations.

[0098] In this embodiment, based on the scenario shown in Figure 1 above, terminals (1,2,3…,N) can report their respective AI model requirements to the Near-RT RIC. Then, the Near-RT RIC generates AI models that meet the needs of each terminal (1,2,3…,N) based on the AI ​​model requirements, providing more professional and customized AI services to different users to meet their diverse performance requirements. Furthermore, the AI ​​model files of each AI model and related model management information can be sent to the Non-RT RIC for unified management by the Non-RT RIC.

[0099] To further illustrate the technical solutions provided in the embodiments of this application, a detailed description is provided below in conjunction with the accompanying drawings and specific implementation methods. Although the embodiments of this application provide method operation steps as shown in the following embodiments or drawings, the method may include more or fewer operation steps based on conventional or non-inventive methods. In steps where there is no logically necessary causal relationship, the execution order of these steps is not limited to the execution order provided in the embodiments of this application. In actual processing or when the device executes the method, it may be executed in the order shown in the embodiments or drawings, or in combination.

[0100] Figure 2 is a flowchart of a model generation method provided in an embodiment of this application. This process can be executed by Near-RT RIC to meet users' customized AI service needs. As shown in Figure 2, the process may include the following steps:

[0101] 201: Receive AI model requirement information reported by the user device, and generate global AI model parameters based on the AI ​​model requirement information.

[0102] In some embodiments, the user equipment (user equipment 1 as shown in FIG1) can carry AI model requirement information in uplink control information (UCI) and send it to the base station (base station 11 as shown in FIG1), and the base station 11 then uploads it to Near-RT RIC through the E2 interface.

[0103] In other embodiments, the uplink control information also includes a cyclic redundancy check (CRC) scrambled using C-RNTI or RA-RNTI, so that the base station can effectively identify different user equipment and protect the security of data transmission.

[0104] In some embodiments, the AI ​​model requirement information may include at least the user device's identification information, the user device's configured local data format, the model application scenario, the user device's privacy protection level, the model performance gain requirement, the model service quality requirement, the user device's computing power, and the timeliness of the user device's request to obtain the model. It may also include other customized requirements. Table 1 illustrates an example table of AI model requirement information provided in an embodiment of this application.

[0105] Table 1: Example of AI Model Requirements Information

[0106] In other embodiments, when multiple user devices are present, Near-RT RIC is configured with multiple near real-time RIC platform extended applications (xAPPs) for data interaction and model processing interaction with different user devices. Furthermore, different xAPPs can set up interfaces to exchange information, or be uniformly controlled by Near-RT RIC, so as to receive AI model requirement information reported by multiple user devices and generate a shared global AI model parameter based on their respective AI model requirement information.

[0107] 202: Send the above AI global model parameters to the user equipment.

[0108] The AI ​​global model parameters are used by the user device to generate AI model training parameters after training the AI ​​global model parameters using a model training strategy.

[0109] In some embodiments, the AI ​​global model parameters can be determined by Near-RT RIC based on the local data format, model application scenario, model performance gain requirements, model service quality requirements, computing power, and current network transmission capacity and network congestion in the AI ​​model requirement information, in order to adapt to the AI ​​service needs of user devices.

[0110] In other embodiments, sending the aforementioned AI global model parameters to the user device can also be achieved by first initializing the AI ​​global model parameters by generating random numbers using a fixed random seed. Then, based on the local data format, model application scenario, model performance gain requirements, model service quality requirements, computing power, and current network transmission capacity and network congestion status in the AI ​​model requirement information, AI global model parameters adapted to the user device are generated, laying a stable and predictable foundation for subsequent model training and application.

[0111] In other embodiments, before sending the aforementioned AI global model parameters to the user equipment, the Near-RT RIC can be configured with a list of user equipment nodes to determine those participating in model generation. For example, the Near-RT RIC can record the user equipment identifiers from the AI ​​model requirement information reported by all user equipment in the user equipment node list according to a set period, and send it to the base station through the E2 interface so that the base station can broadcast the user equipment node list to all user equipment. For example, the user equipment node list can be broadcast to all user equipment with a System Information Block 1 (SIB1) message to ensure that the user equipment can access the system correctly, participate in the subsequent model training, and have better model generation and model training interactions with the Near-RT RIC.

[0112] In some embodiments, sending the aforementioned AI global model parameters to the user equipment may involve sending the AI ​​global model parameters to the base station, and then the base station sending downlink control information carrying the AI ​​global model parameters to the user equipment.

[0113] Furthermore, after obtaining the global AI model parameters, the user device uses a model training strategy to train the global AI model parameters, generating AI model training parameters, which may specifically include:

[0114] The user device prepares local data samples for model training, and starts model training with these local data samples, a specific model training strategy, and training hyperparameters. During the training process, the global AI model parameters are continuously updated and iterated according to the gradient of the loss function, thereby generating the corresponding AI model training parameters.

[0115] In some embodiments, the model training strategy can be configured by the user equipment itself, or it can be configured by the base station to the user equipment through at least one of system messages, Radio Resource Control (RRC) signaling, Media Access Control element (MAC CE) signaling, and downlink control signaling.

[0116] Furthermore, the model training strategy includes at least an optimization algorithm and a loss function; the optimization algorithm can be a gradient descent algorithm (e.g., at least one of stochastic gradient descent and batch gradient descent); the loss function includes at least one of mean squared error loss function, cross-entropy loss function, Hinge loss function, Huber loss function, and Focal loss function.

[0117] Furthermore, the model training strategy can be reasonably configured based on factors such as user behavior patterns, network traffic distribution, signal strength changes, user equipment computing power, model application scenarios, model performance gain requirements, model parameters and complexity within the cell covered by the base station.

[0118] In some embodiments, the training hyperparameters may include learning rate, batch size, weight of training set, validation set, test set, number of training rounds, learning rate decay rate, etc.

[0119] Similarly, when there are multiple user devices, the base station can send shared global AI model parameters to each user device. This allows each user device to perform its own model training process based on its own configured local data samples, model training strategies, and training hyperparameters. This federated learning mechanism enables distributed model training to ensure the data privacy of each user device.

[0120] 203: Receive AI model training parameters sent by the user equipment, and generate the AI ​​model file of the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters.

[0121] In this step, receiving the AI ​​model training parameters sent by the user equipment can be achieved by the user equipment uploading the AI ​​model training parameters to the base station through the user plane, and the base station then forwarding them to the Near-RT RIC through the E2 interface.

[0122] In this step, the AI ​​model training parameters sent by the user device are received, and based on the AI ​​model training parameters and the AI ​​global model parameters, the current target AI global model parameters are generated. It is then determined whether the target AI model constructed by the target AI global model parameters meets the model training termination condition. If yes, the AI ​​model file of the target AI model is generated based on the target AI global model parameters. If no, the current target AI global model parameters are sent to the user device for a new round of model training until the target AI model generated by the final target AI global model parameters meets the model training termination condition.

[0123] In some embodiments, Near-RT RIC can obtain the AI ​​model training parameters through a subscription mechanism. For example, firstly, Near-RT RIC's xAPP registers in the system so that the base station can recognize its identity and permissions. After successful registration, xAPP sends a model subscription request to the base station. After receiving the model subscription request, the base station confirms and verifies it. If the subscription is successful, Near-RT RIC can obtain the AI ​​model training parameters uploaded by the user device through the E2 interface.

[0124] In some embodiments, the AI ​​model training parameters are local model parameters or local model gradients generated by the user device during the training process.

[0125] In other embodiments, when multiple user devices exist, generating the AI ​​model file of the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters may include the following process:

[0126] First, in the presence of multiple user devices, the system receives AI model training parameters sent by each user device and performs a model aggregation operation on each AI model training parameter to obtain aggregated AI model training parameters, thereby accelerating the subsequent model training speed. Then, based on the aggregated AI model training parameters and the AI ​​global model parameters from the previous round, the system generates the current target AI global model parameters. If the target AI model generated by the target AI global model parameters is detected to meet the model training termination condition, the system generates the AI ​​model file of the target AI model based on the target AI global model parameters of this round, so that it can be sent to the Non-RT RIC for unified model management.

[0127] In some embodiments, when the model aggregation operation is an average aggregation operation, the specific implementation process includes: summing up each local model parameter or local model gradient according to a set vector level, taking the average value, and then combining it with the AI ​​global model parameters of the previous round to generate the target AI global model parameters for the current round.

[0128] In other embodiments, when the model aggregation operation is a weighted aggregation operation, the specific implementation process includes: according to a set vector level, the aggregation weights of each local model parameter or local model gradient are summed, and then combined with the AI ​​global model parameters of the previous round to generate the target AI global model parameters of the current round. The aggregation weights are related to the configuration of the target AI model, the model application scenario, etc.

[0129] In some embodiments, Near-RT RIC can execute a model detection mechanism to detect the performance metrics of the target AI model in real time (these performance metrics can be configured on the core network side or the network management side). When it is detected that the performance metrics of the target AI model have reached the model training termination condition, it can complete the model generation process without interacting with the user device for model training. Furthermore, the aforementioned model training termination condition may include at least one of the following:

[0130] (1) Accuracy performance indicators termination conditions, for example, the loss function value of the target AI model is less than or equal to the loss threshold, and the accuracy is greater than or equal to the accuracy threshold, as detected by the validation set.

[0131] (2) Termination conditions for communication overhead indicators, such as the network congestion level of the target AI model being greater than a certain threshold, the channel quality in the channel state information being lower than a certain threshold, the throughput being higher than a certain threshold, the bit error rate or block error rate being higher than a certain threshold, and the end-to-end transmission delay being higher than a certain threshold.

[0132] (3) Termination conditions for computational overhead indicators, such as detecting that the number of model parameters of the target model is higher than a certain threshold, the computational complexity is higher than a certain threshold, or the CPU / GPU computing resources occupied are higher than a certain threshold.

[0133] It should be noted that the model training termination conditions listed above are only examples. There may be other model training termination conditions, such as the total number of training rounds of the target AI model reaching a certain threshold. These conditions can be flexibly set according to actual needs, and the embodiments of this application do not impose any restrictions on them.

[0134] 204: Send the model management information and AI model file of the target AI model to the Non-RT RIC. The model management information is used by the Non-RT RIC to manage the target AI model.

[0135] In this step, the model management information and AI model file can be sent to the Non-RT RIC via the A1 interface connected to the Non-RT RIC, so that the Non-RT RIC can effectively manage the target AI model.

[0136] In some embodiments, the model management information is stored in a model management list built by Non-RT RIC, which includes at least the model application scenario, model benefits, model costs, model file storage location, and cell identifier associated with the target AI model.

[0137] Furthermore, the model management list is constructed based on the model identifiers and / or model functions of multiple AI models. It is used to store model management information for different AI models, enabling unified management of all models generated by Near-RT RIC. Table 2 below illustrates an example of a model management list constructed based on model functions, and Table 3 illustrates an example of a model management list constructed based on model identifiers.

[0138] Table 2: Model Management List Based on Model Functionality

[0139] Table 3: Model Management List Based on Model Functionality

[0140] It should be noted that Tables 2 and 3 above are just examples. Other management information may also be included in the model management list, which can be recorded according to actual needs. This application embodiment does not impose any restrictions here.

[0141] In some embodiments, the AI ​​model files can be stored using a shared storage architecture, on a cloud server associated with the Non-RT RIC (e.g., on OAM or NWDAF), thereby simplifying the access process for upper-layer applications to the AI ​​model files. Furthermore, the file format of the AI ​​model files can be a resolvable binary executable file format, conforming to the current specifications for saving and loading AI model weights in deep learning frameworks.

[0142] In some embodiments, when a user device requests model deployment, the Near-RT RIC can interact with the Non-RT RIC and the user device to complete the model deployment process. Figure 3 illustrates a flowchart of a model deployment method provided in an embodiment of this application. The process includes the following steps:

[0143] 301: Send a model deployment request to the Non-RT RIC.

[0144] In this step, the model deployment request carries the AI ​​model requirement information sent by the user device. This information is used by the Non-RT RIC to determine K target AI models that match the AI ​​model requirement information from the model management list (such as Table 2 or Table 3 above). K is an integer greater than 0. For example, the Non-RT RIC determines K target AI models that match the model application scenario, model performance gain requirements, service quality requirements, computing power, etc. from the model management list.

[0145] In some embodiments, the value of K can be configured by the core network side and the network management side, or it can be carried in the AI ​​model requirement information reported by the user equipment.

[0146] 302: Receive the model deployment strategies corresponding to each of the K target AI models from the Non-RT RIC.

[0147] The deployment strategies for each of the K target AI models are generated at least based on the model management information corresponding to each of the K target AI models.

[0148] In some embodiments, the model deployment strategy may include the model identifier or model function of each of the K target AI models, the model file storage location, the model deployment method, the model deployment validity period, the model deployment reference information, computing power, storage capacity, and model inference strategy.

[0149] Furthermore, the model deployment method may include at least one of joint deployment and independent deployment; the model deployment validity period refers to how long the user device can use the model, and once it expires, new AI model requirement information needs to be uploaded again; the model deployment reference information may include recommended configurations for data collection, model compression, model transmission, model monitoring, etc.; the computing power refers to how much computing resources the user device needs to reserve for running the above model; the storage capacity refers to how much storage space the user device needs to reserve for storing the above model locally; and the model inference strategy may include configuration information related to the model inference process.

[0150] 303: Based on the storage location of the model files corresponding to the K target AI models, download the AI ​​model files corresponding to the K target AI models from the cloud server.

[0151] In this step, Near-RT RIC downloads the corresponding AI model files for each of the K target AI models from the cloud server, based on the model file storage location included in the model deployment strategy for each of the K target AI models.

[0152] 304: The model deployment strategy and AI model file corresponding to each of the K target AI models are encapsulated to obtain the model deployment strategy signaling, and the model deployment strategy signaling is sent to the user equipment.

[0153] In this step, sending the model deployment strategy signaling to the user equipment can be done by Near-RT RIC sending the encapsulated model deployment strategy signaling to the base station. The base station then sends the signaling to the base station via the control plane in the form of downlink control information. Furthermore, the downlink control information includes a CRC scrambled by C-RNTI or RA-RNTI.

[0154] Furthermore, the model deployment strategy signaling is used to instruct user equipment to deploy models based on the model deployment strategies and AI model files corresponding to the K target AI models. For example, based on their respective model deployment strategies, user equipment stores each AI model file in local space. After preparing the corresponding computing resources, it completes local model identification and various configurations related to model inference, etc., according to the deployment method of independent deployment or joint deployment.

[0155] In some embodiments, Near-RT RIC can also generate a model deployment list based on the model deployment strategies of the K target AI models. This model deployment list can be used to record information such as user device identification information, model identification or model function, model deployment method, model deployment validity period, number of local models, updated or uninstalled models, and user device feedback, which is convenient for subsequent review and management.

[0156] In other embodiments, the user equipment may be configured with a model monitoring mechanism during the model inference phase, and promptly feed back model monitoring information to the Near-RT RIC, so that the Near-RT RIC can decide whether to update or uninstall the model, etc. Specifically, it may include the following process:

[0157] In response to a detected model processing trigger event of the target AI model, a model processing instruction for the target AI model is sent to the user device. If the model processing trigger event is to update the model, the model processing instruction includes at least the model update mode and the model update method, and may also include model identifier or model function, model update algorithm, etc. The model update mode is used to instruct the user device to update the target AI model using offline update mode or online update mode, and the model update method is used to instruct the user device to update the target AI model using local model update or global model update. If the model processing trigger event is to uninstall the model, the model processing instruction includes at least the model uninstallation method, and may also include uninstallation confirmation information, model identifier or model function, etc. The model uninstallation method is used to instruct the user device to uninstall the target AI model using soft uninstallation or hard uninstallation.

[0158] Furthermore, the offline update mode indicates that only the target AI model stored locally on the user device is updated, and no update is performed on target AI models that have the same model identifier or model function and have been deployed on other nodes (user devices). The online update mode indicates that not only the target AI model stored locally on the user device is updated, but also the target AI models that have the same model identifier or model function and have been deployed on other nodes (user devices) are updated. The local model update indicates that the model parameters of a specific part / module / component of the target AI model are updated, and the global model update indicates that all model parameters of the target AI model are updated.

[0159] Furthermore, the hard unloading method indicates that after receiving a model processing instruction (specifically, a model unloading instruction), the user device immediately stops using the corresponding target AI model and deletes the corresponding AI model file in the local storage space. The soft unloading method indicates that after receiving a model unloading instruction, the user device completes all operations related to the target AI model before stopping using the target AI model and deleting the corresponding AI model file in the local storage space.

[0160] In some embodiments, the model processing triggering event may be determined by the Near-RT RIC based on model monitoring information fed back by the user equipment, indicating that a model update triggering threshold or a model offload triggering threshold has been exceeded, and triggering model processing instructions for one or more target models. Further, the model monitoring information may be reported by the user equipment to the base station via uplink control information or MAC CE, and then forwarded by the base station to the Near-RT RIC; the model update triggering threshold and the model offload triggering threshold are associated with specific configurations / conditions of the user equipment and the network, and also with auxiliary conditions (e.g., scenario, site, channel, etc.).

[0161] In other embodiments, the model processing triggering event may also be that Near-RT RIC triggers model processing instructions for one or more models according to a preset time period.

[0162] In some embodiments, the above-mentioned model processing instructions may be carried by downlink control information, MAC CE, or RRC reconfiguration signaling.

[0163] In this embodiment, the system first receives AI model requirement information reported by the user device, generates global AI model parameters based on the AI ​​model requirement information, and sends the global AI model parameters to the user device. Then, the user device uses a model training strategy to train the global AI model parameters to generate AI model training parameters, enabling model training to be completed on the user device, effectively protecting user privacy. Finally, after receiving the AI ​​model training parameters sent by the user device, the system generates an AI model file of the target AI model based on the AI ​​model training parameters and the global AI model parameters, and sends the model management information of the target AI model and the AI ​​model file to the Non-RT RIC, facilitating the Non-RT RIC to manage the target AI model, thereby meeting the user's customized AI service needs and improving the user experience.

[0164] Figure 4 is a flowchart of another model generation method provided in an embodiment of this application. This process can be executed by a user device (as shown in Figure 1) to meet the user's customized AI service needs. As shown in Figure 4, the process may include the following steps:

[0165] 401: Report AI model requirements to the Near-RT RIC.

[0166] In this step, the explanations of AI model requirements and AI global model parameters can be found in the description of section 201 in Figure 2 above, and will not be repeated here.

[0167] 402: Received AI global model parameters from Near-RT RIC.

[0168] The AI ​​global model parameters are generated by Near-RT RIC based on AI model requirement information. For details, please refer to the relevant description in Figure 201, which will not be repeated here.

[0169] 403: The AI ​​global model parameters are trained using a model training strategy to obtain the AI ​​model training parameters.

[0170] In this step, the user device can prepare local data samples for model training, and start model training with these local data samples, a specific model training strategy, and training hyperparameters. During the training process, the AI ​​global model parameters will be continuously updated and iterated according to the gradient of the loss function, thereby generating the corresponding AI model training parameters.

[0171] In some embodiments, the model training strategy can be configured by the user equipment itself, or it can be configured by the base station to the user equipment through at least one of system messages, Radio Resource Control (RRC) signaling, Media Access Control element (MAC CE) signaling, and downlink control signaling.

[0172] Furthermore, the model training strategy includes at least an optimization algorithm and a loss function; the optimization algorithm can be a gradient descent algorithm (e.g., at least one of stochastic gradient descent and batch gradient descent); the loss function includes at least one of mean squared error loss function, cross-entropy loss function, Hinge loss function, Huber loss function, and Focal loss function.

[0173] Furthermore, the model training strategy can be reasonably configured based on factors such as user behavior patterns, network traffic distribution, signal strength changes, user equipment computing power, model application scenarios, model performance gain requirements, model parameters and complexity within the cell covered by the base station.

[0174] In some embodiments, the training hyperparameters may include learning rate, batch size, weight of training set, validation set, test set, number of training rounds, learning rate decay rate, etc.

[0175] Similarly, when there are multiple user devices, the base station can send shared global AI model parameters to each user device. This allows each user device to perform its own model training process based on its own configured local data samples, model training strategies, and training hyperparameters. This federated learning mechanism enables distributed model training to ensure the data privacy of each user device.

[0176] 404: Sending AI model training parameters to Near-RT RIC.

[0177] The AI ​​model training parameters are used by Near-RT RIC to generate the AI ​​model file of the target AI model based on the AI ​​model training parameters and the global AI model parameters.

[0178] In this step, the user equipment can upload the AI ​​model training parameters to the base station through the user plane, and the base station will then forward them to the Near-RT RIC through the E2 interface.

[0179] In other embodiments, when a user device requests model deployment, the following process may be included:

[0180] First, a model deployment request is sent to the Non-RT RIC (Non-Real-Time Access Network Intelligent Management Controller) via the Near-RT RIC. This model deployment request is used by the Non-RT RIC to determine K target AI models that match the AI ​​model requirement information in the model deployment request from the model management list. This model management list is constructed based on the model identifiers and / or model functions of multiple AI models, where K is an integer greater than 0. For details, please refer to the relevant description in Figure 3, which will not be repeated here.

[0181] Secondly, it receives model deployment signaling from Near-RT RIC. The model deployment signaling is encapsulated by Near-RT RIC based on the model deployment strategy and AI model file corresponding to each of the K target AI models. The model deployment strategy is generated by Non-RT RIC based on the model management information corresponding to each of the K target AI models recorded in the model management list. For details, please refer to the relevant description in Figure 3, which will not be repeated here.

[0182] Then, model deployment operations are performed based on the model deployment strategies and AI model files corresponding to the K target AI models.

[0183] In other embodiments, after the user device completes the model deployment operation, it can continuously perform model inference on the deployed K target AI models, thereby adapting to the user device's local data format, gradually improving model inference performance, and enhancing the level of AI model customization services. Specifically, this may include the following processes:

[0184] First, K target AI models are used to infer the t-th data sample in the sample set, resulting in K inference results for the t-th data sample; where t is an integer greater than 0.

[0185] In some embodiments, before inferring the t-th data sample in the sample set using K target AI models, the user device may initialize the ensemble weights of the K target AI models and configure information such as the frequency of updating the ensemble weights, the algorithm used for ensemble learning, and hyperparameter settings based on their respective model inference strategies.

[0186] Then, the target inference result for the t-th data sample is determined based on the K inference results, and the ensemble weights of the K target AI models are updated based on the target inference result and the loss functions of the K target AI models.

[0187] In some embodiments, the updated integration weights satisfy the following expression:

[0188] in, Let w be the ensemble weight for the k-th target AI model out of K target AI models on the t-th data sample, where ε is a preset hyperparameter. k Let l(w) be the model parameters of the k-th target AI model. k ;x t ,y t Let be the loss function of the k-th target AI model on the t-th data sample.

[0189] In other embodiments, when a user equipment receives a model processing instruction for a target AI model sent by Near-RT RIC, it can perform model processing operations on the target AI model according to the model processing instruction. The model processing instruction is sent by Near-RT RIC when it detects a model processing trigger event of the target AI model (which may be one or more). The explanation of the model processing trigger event can be found in the relevant description above, and will not be repeated here.

[0190] Furthermore, if the model processing trigger event is to update the model, the model processing instruction shall at least include the model update mode and the model update method, and may also include the model identifier or model function, model update algorithm, etc. The model update mode is used to instruct the user device to update the target AI model using offline update mode or online update mode, and the model update method is used to instruct the user device to update the target AI model using local model update or global model update. If the model processing trigger event is to uninstall the model, the model processing instruction shall at least include the model uninstallation method, and may also include uninstallation confirmation information, model identifier or model function, etc. The model uninstallation method is used to instruct the user device to uninstall the target AI model using soft uninstallation or hard uninstallation.

[0191] In other embodiments, after the user equipment completes the relevant operations according to the model processing instructions, it can also send feedback information to Near-RT RIC indicating that the model update or model uninstallation is complete, so that Near-RT RIC can update the above-mentioned model deployment list accordingly after receiving the feedback information, such as deleting the relevant information of the target AI model that has been uninstalled.

[0192] In this embodiment, the user equipment can report AI model requirement information to the Near-RT RIC, which facilitates the Near-RT RIC to generate suitable global AI model parameters based on the AI ​​model requirement information. After receiving the global AI model parameters, the user equipment can adopt a model training strategy to train the global AI model parameters locally to obtain AI model training parameters, thus protecting the user equipment's data privacy. Then, the user equipment sends the AI ​​model training parameters to the Near-RT RIC. Subsequently, the Near-RT RIC generates the AI ​​model file of the target AI model based on the AI ​​model training parameters and the global AI model parameters, thereby meeting the user's customized AI service needs and improving the user experience.

[0193] Based on the methods shown in Figures 2 and 4 above, Figure 5 exemplarily illustrates a signaling interaction diagram for model generation and model deployment provided by an embodiment of this application. As shown in Figure 5, the steps include:

[0194] 501: User equipment reports AI model requirement information to Near-RT RIC via base station.

[0195] 502: Near-RT RIC generates global AI model parameters based on the above AI model requirement information.

[0196] 503: Near-RT RIC sends the above-mentioned AI global model parameters to the user equipment via the base station.

[0197] 504: The user equipment uses a model training strategy to train the above-mentioned global AI model parameters and generates AI model training parameters.

[0198] 505: User equipment reports AI model training parameters to Near-RT RIC via base station.

[0199] 506: Near-RT RIC generates the AI ​​model file of the target AI model based on the above AI model training parameters and AI global model parameters.

[0200] 507: Near-RT RIC reports the AI ​​model file and model management information of the target AI model to Non-RT RIC.

[0201] 508: Non-RT RIC adds the model management information of the target AI model to the model management list and stores the AI ​​model file of the target AI model to the cloud server.

[0202] 509: When Non-RT RIC receives a model deployment request from Near-RT RIC, it determines K target AI models that match the AI ​​model requirement information sent by the user device in the model deployment request from the model management list.

[0203] 510: Non-RT RIC sends the model deployment strategies corresponding to each of the K target AI models to Near-RT RIC.

[0204] 511: Near-RT RIC downloads the respective AI model files from the cloud server based on the storage location of the model files of the K target AI models in the model deployment strategy.

[0205] 512: Near-RT RIC encapsulates the model deployment strategies and AI model files corresponding to each of the K target AI models to obtain the model deployment strategy signaling.

[0206] 513: Near-RT RIC sends the above model deployment strategy signaling to user equipment via the base station.

[0207] 514: The user equipment performs model deployment operations and model inference operations based on the model deployment strategies and AI model files corresponding to the K target AI models in the above model deployment strategy signaling.

[0208] 515: User equipment feeds back model monitoring information to Near-RT RIC in real time via base stations.

[0209] 516: If Near-RT RIC triggers a model processing trigger event based on the above model monitoring information, it sends a model processing instruction to the user equipment via the base station.

[0210] 517: The user equipment executes the corresponding model processing operation according to the above model processing instructions.

[0211] Based on the same technical concept, this application also provides a structural schematic diagram of a Near-RT RIC, which can implement the model generation method flow of the Near-RT RIC side described in this application.

[0212] Figure 6 is a schematic diagram of a Near-RT RIC structure provided in an embodiment of this application. The structure includes a receiving model 601, a first transmitting module 602, a generating module 603, and a second transmitting module 604.

[0213] The receiving module 601 is used to receive AI model requirement information reported by the user device and generate AI global model parameters based on the AI ​​model requirement information.

[0214] The first sending module 602 is used to send the AI ​​global model parameters to the user equipment; wherein, the AI ​​global model parameters are used by the user equipment to generate AI model training parameters after training the AI ​​global model parameters using a model training strategy.

[0215] The generation module 603 is used to receive the AI ​​model training parameters sent by the user equipment, and generate the AI ​​model file of the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters.

[0216] The second sending module 604 is used to send the model management information of the target AI model and the AI ​​model file to the Non-RT RIC; wherein, the model management information is used by the Non-RT RIC to manage the target AI model.

[0217] Based on the same technical concept, this application also provides a schematic diagram of the structure of a user equipment, which can implement the model generation method flow on the user equipment side described above in this application.

[0218] Figure 7 is a schematic diagram of the structure of a user equipment provided in an embodiment of this application. The structure includes a first transmitting module 701, a receiving module 702, a training module 703, and a second transmitting module 704.

[0219] The first transmitting module 701 is used to report AI model requirement information to the Near-RT RIC intelligent controller.

[0220] The receiving module 702 is used to receive AI global model parameters from the Near-RT RIC; wherein the AI ​​global model parameters are generated based on the AI ​​model requirement information;

[0221] The training module 703 is used to train the global AI model parameters using a model training strategy to obtain the AI ​​model training parameters.

[0222] The second sending module 704 is used to send the AI ​​model training parameters to the Near-RT RIC; wherein, the AI ​​model training parameters are used by the Near-RT RIC to generate an AI model file of the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters.

[0223] It should be noted that the apparatus provided in this application embodiment can implement all the method steps in the above method embodiment and achieve the same technical effect. Therefore, the parts and beneficial effects that are the same as those in the method embodiment will not be described in detail here.

[0224] Based on the same technical concept, this application also provides a communication device that can realize the functions of the aforementioned user equipment or Near-RT RIC.

[0225] Figure 8 is a schematic diagram of the structure of a communication device provided in an embodiment of this application.

[0226] At least one processor 801 and a memory 802 connected to at least one processor 801 are included. In this embodiment, the specific connection medium between the processor 801 and the memory 802 is not limited. Figure 8 illustrates an example where the processor 801 and the memory 802 are connected via a bus 800. The bus 800 is represented by a thick line in Figure 8. The connection methods between other components are for illustrative purposes only and are not intended to be limiting. The bus 800 can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in Figure 8, but this does not indicate that there is only one bus or one type of bus. Alternatively, the processor 801 can also be called a controller; the name is not limited.

[0227] In this embodiment, the memory 802 stores instructions executable by at least one processor 801. By executing the instructions stored in the memory 802, the at least one processor 801 can perform a model generation method as described above. The processor 801 can implement the functions of each module in the device shown in FIG6 or FIG7.

[0228] The processor 801 is the control center of the device. It can connect to various parts of the control device through various interfaces and lines. By running or executing instructions stored in memory 802 and calling data stored in memory 802, the processor can perform various functions and process data, thereby monitoring the device as a whole.

[0229] In this embodiment, processor 801 may include one or more processing units. Processor 801 may integrate an application processor and a modem processor. The application processor mainly handles the operating system, user interface, and applications, while the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into processor 801. In some embodiments, processor 801 and memory 802 may be implemented on the same chip; in other embodiments, they may be implemented on separate chips.

[0230] The processor 801 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit, field-programmable gate array or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of a model generation method disclosed in the embodiments of this application can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0231] Memory 802, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory 802 may include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic storage, magnetic disk, optical disk, etc. Memory 802 can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. In the embodiments of this application, memory 802 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0232] By designing and programming the processor 801, the code corresponding to a model generation method described in the foregoing embodiments can be embedded into the chip, enabling the chip to execute a model generation method of the embodiment shown in Figure 2 or Figure 4 during runtime. How to design and program the processor 801 is a technique well-known to those skilled in the art and will not be elaborated upon here.

[0233] It should be noted that the communication device provided in this application embodiment can implement all the method steps implemented in the above method embodiment and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0234] Based on the same technical concept, embodiments of this application provide a computer storage medium, which includes computer program code. When the computer program code is run on a computer, it causes the computer to execute any of the model generation methods described above. Since the principle by which the computer storage medium solves the problem is similar to that of a model generation method, the implementation of the computer storage medium can be referred to the implementation of the method, and repeated details will not be elaborated further.

[0235] In specific implementation, computer storage media can include: Universal Serial Bus Flash Drive (USB), portable hard drive, Read-Only Memory (ROM), Random Access Memory (RAM), magnetic disk or optical disk, and other storage media that can store program code.

[0236] Based on the same technical concept, this application also provides a computer program product, which includes computer program code. When the computer program code is run on a computer, it causes the computer to execute any of the model generation methods described above. Since the principle by which the above computer program product solves the problem is similar to that of a model generation method, the implementation of the above computer program product can refer to the implementation of the method, and repeated details will not be described again.

[0237] Computer program products may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0238] The methods in this application can be implemented, in whole or in part, by software, hardware, firmware, or any combination thereof. When implemented in software, they can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed on a computer, the processes or functions described in this application are performed, in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, network equipment, user equipment, core network equipment, OAM, or other programmable devices.

[0239] The computer program or instructions may be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer program or instructions may be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium may be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; or an optical medium, such as a digital video optical disc; or a semiconductor medium, such as a solid-state drive. The computer-readable storage medium may be a volatile or non-volatile storage medium, or may include both volatile and non-volatile types of storage media.

[0240] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0241] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more flowchart illustrations and / or one or more block diagrams.

[0242] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction device that implements the functions specified in one or more flowcharts and / or one or more block diagrams.

[0243] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0244] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations of the present invention fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.

Claims

1. A model generation method characterized by comprising: The method, applied to Near-RT RIC (Near-Real-Time Radio Access Network) intelligent controllers, includes: Receive AI model requirement information reported by user devices, and generate global AI model parameters based on the AI ​​model requirement information; Send the AI ​​global model parameters to the user equipment; wherein, the AI ​​global model parameters are used by the user equipment to generate AI model training parameters after training the AI ​​global model parameters using a model training strategy; Receive the AI ​​model training parameters sent by the user equipment, and generate the AI ​​model file of the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters; The model management information of the target AI model and the AI ​​model file are sent to the Non-RT RIC (Non-Real-Time Access Network Intelligent Management Controller); wherein the model management information is used by the Non-RT RIC to manage the target AI model.

2. The method of claim 1, wherein, The step of receiving the AI ​​model training parameters sent by the user equipment and generating an AI model file for the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters includes: In the presence of multiple user devices, the AI ​​model training parameters sent by each of the multiple user devices are received, and a model aggregation operation is performed on the gradients of each AI model training parameter to obtain the aggregated AI model training parameters; wherein, the AI ​​model training parameters are local model parameters or local model gradients. Based on the aggregated AI model training parameters and the AI ​​global model parameters, the current target AI global model parameters are generated; If the target AI model generated based on the target AI global model parameters is found to meet the model training termination condition, then the AI ​​model file of the target AI model is generated based on the target global model parameters.

3. The method of claim 1, wherein, The model management information is stored in the model management list built by the Non-RT RIC, and includes at least the model application scenario, model benefits, model costs, model file storage location of the target AI model, and the cell identifier associated with the target AI model. The model management list is built based on the model identifiers and / or model functions of multiple AI models, and the AI ​​model file is stored on the cloud server associated with the Non-RT RIC.

4. The method of claim 3, wherein, After generating the target AI model parameters based on the AI ​​model training parameters and the AI ​​global model parameters, the process further includes: Send a model deployment request to the Non-RT RIC; wherein the model deployment request carries AI model requirement information sent by the user equipment, which is used by the Non-RT RIC to determine K target AI models that match the AI ​​model requirement information from the model management list, where K is an integer greater than 0; Receive model deployment strategies corresponding to each of the K target AI models from the Non-RT RIC; wherein, the model deployment strategies corresponding to each of the K target AI models are generated at least based on the model management information corresponding to each of the K target AI models; Based on the storage location of the model files corresponding to the K target AI models, download the AI ​​model files corresponding to the K target AI models from the cloud server; The model deployment strategies and AI model files corresponding to the K target AI models are encapsulated to obtain model deployment strategy signaling, and the model deployment strategy signaling is sent to the user equipment; wherein, the model deployment strategy signaling is used to instruct the user equipment to deploy models based on the model deployment strategies and AI model files corresponding to the K target AI models.

5. The method according to any one of claims 1 to 4, characterized in that, The method further includes: In response to a detected model processing trigger event of the target AI model, a model processing instruction for the target AI model is sent to the user equipment; Wherein, if the model processing trigger event is a model update, the model processing instruction includes at least a model update mode and a model update method. The model update mode is used to instruct the user device to update the target AI model using an offline update mode or an online update mode, and the model update method is used to instruct the user device to update the target AI model using a local model update or a global model update. If the model processing trigger event is a model uninstallation, the model processing instruction includes at least a model uninstallation method. The model uninstallation method is used to instruct the user device to uninstall the target AI model using a soft uninstallation method or a hard uninstallation method.

6. The method according to any one of claims 1 to 4, wherein The AI ​​model requirements information includes at least: The identification information of the user equipment; The local data format indication configured in the user equipment; Model application scenario indication; The privacy protection level of the user device; Model performance gain requirement indication; Model service quality requirements indication; The computing power indication of the user equipment; The user equipment requests a timeliness indication of the model.

7. A model generation method characterized by comprising: Applied to user equipment, the method includes: Report artificial intelligence (AI) model requirements to the Near-RT RIC (Near-RT Intelligent Controller); Receive AI global model parameters from the Near-RT RIC; wherein the AI ​​global model parameters are generated by the Near-RT RIC based on the AI ​​model requirement information; The AI ​​global model parameters are trained using a model training strategy to obtain the AI ​​model training parameters; The AI ​​model training parameters are sent to the Near-RT RIC; wherein the AI ​​model training parameters are used by the Near-RT RIC to generate an AI model file of the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters.

8. The method of claim 7, wherein, The method further includes: The model deployment request is sent to the Non-RT RIC via the Near-RT RIC; wherein the model deployment request is used by the Non-RT RIC to determine K target AI models that match the AI ​​model requirement information from the model management list, the model management list is constructed based on the model identifiers and / or model functions of multiple AI models, and K is an integer greater than 0; Receive model deployment signaling from the Near-RT RIC; wherein the model deployment signaling is encapsulated based on the model deployment strategy and AI model file corresponding to each of the K target AI models, and the model deployment strategy is generated by the Non-RT RIC based on the model management information corresponding to each of the K target AI models recorded in the model management list; The model deployment operation is performed based on the model deployment strategy and AI model file corresponding to each of the K target AI models.

9. The method of claim 8, wherein, After performing the model deployment operation based on the model deployment strategies and AI model files corresponding to the K target AI models, the process further includes: The K target AI models are used to infer the t-th data sample in the sample set to obtain K inference results for the t-th data sample; where t is an integer greater than 0. The target inference result of the t-th data sample is determined based on the K inference results, and the ensemble weights of the K target AI models are updated based on the target inference result and the loss functions of the K target AI models.

10. The method of claim 7, wherein, The method further includes: Receive model processing instructions for the target AI model sent by the Near-RT RIC; wherein, the model processing instructions are sent by the Near-RT RIC when it detects a model processing trigger event for the target AI model; Perform model processing operations on the target AI model according to the model processing instructions; Wherein, if the model processing trigger event is a model update, the model processing instruction includes at least a model update mode and a model update method. The model update mode is used to instruct the user device to update the target AI model using an offline update mode or an online update mode, and the model update method is used to instruct the user device to update the target AI model using a local model update or a global model update. If the model processing trigger event is a model uninstallation, the model processing instruction includes at least a model uninstallation method. The model uninstallation method is used to instruct the user device to uninstall the target AI model using a soft uninstallation method or a hard uninstallation method.

11. The method of claim 10, wherein, The AI ​​model requirements information includes at least: The identification information of the user equipment; The local data format indication configured in the user equipment; Model application scenario indication; The privacy protection level of the user device; Model performance gain requirement indication; Model service quality requirements indication; The computing power indication of the user equipment; The user equipment requests a timeliness indication of the model.

12. A real-time radio access network intelligent controller, characterized by, include: The receiving module is used to receive AI model requirement information reported by the user device and generate global AI model parameters based on the AI ​​model requirement information. The first sending module is used to send the AI ​​global model parameters to the user equipment; wherein, the AI ​​global model parameters are used by the user equipment to generate AI model training parameters after training the AI ​​global model parameters using a model training strategy; The generation module is used to receive the AI ​​model training parameters sent by the user equipment, and generate the AI ​​model file of the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters; The second sending module is used to send the model management information of the target AI model and the AI ​​model file to the Non-RT RIC (Non-Real-Time Access Network Intelligent Management Controller); wherein the model management information is used by the Non-RT RIC to manage the target AI model.

13. A user equipment, comprising: include: The first transmitting module is used to report AI model requirement information to the Near-RT RIC intelligent controller of the near real-time wireless access network. The receiving module is used to receive AI global model parameters from the Near-RT RIC; wherein the AI ​​global model parameters are generated based on the AI ​​model requirement information; The training module is used to train the global AI model parameters using a model training strategy to obtain the AI ​​model training parameters. The second sending module is used to send the AI ​​model training parameters to the Near-RT RIC; wherein, the AI ​​model training parameters are used by the Near-RT RIC to generate an AI model file of the target AI model based on the AI ​​model training parameters and the AI ​​global model parameters.

14. A communications device, characterized by include: Memory, used to store computer programs; A processor, when executing a computer program stored in the memory, implements the method of any one of claims 1-6, or implements the method of any one of claims 7-11.

15. 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 method of any one of claims 1-6, or the method of any one of claims 7-11.

Images