Model fine tuning method and device, program product and electronic equipment

Abstract

The invention provides a model fine tuning method and device, a program product and electronic equipment, and is applied to a model fine tuning system, the system comprises a terminal node, a base station node and a core network node, and the method comprises the following steps: the base station node receives a model fine tuning request sent by at least one terminal node, sending a model fine tuning request signaling of a base station level to the core network node; the base station node receives a model fine tuning strategy signaling sent by the core network node and a registered pre-training model; the pre-training model is determined according to a model fine tuning request signaling of a base station level; the base station node generates a model fine tuning configuration signaling according to the model fine tuning strategy signaling, issues the model fine tuning configuration signaling to all the terminal nodes in the model fine tuning terminal node group, and transmits the first model to all the terminal nodes, all the terminal nodes execute local model fine tuning on the first model according to the model fine tuning configuration information contained in the model fine tuning configuration signaling, and a corresponding first local fine tuning model is obtained.

Description

Technical Field

[0001] The present disclosure relates to the field of computer and communication technology, and in particular to a model fine-tuning method, device, program product, and electronic device. Background Art

[0002] Model fine-tuning involves retraining a pre-trained model using task-specific data to better adapt it to new tasks or datasets. Model fine-tuning requires establishing a communication link between participating nodes to communicate forward inference results, backward gradients, model parameters, and other business flows. This significantly saves time and computing resources because only the pre-trained model needs to be fine-tuned, eliminating the need to train from scratch.

[0003] In related technologies, how to improve the training efficiency of model fine-tuning has become an urgent problem to be solved. Summary of the Invention

[0004] The present disclosure provides a model fine-tuning method, a model fine-tuning device, a computer program product, and an electronic device to improve the training efficiency of model fine-tuning, reduce time costs, and lower system resource overhead.

[0005] According to a first aspect of the present disclosure, a model fine-tuning method is provided, which is applied to a model fine-tuning system, wherein the model fine-tuning system includes a terminal node, a base station node, and a core network node. The method includes:

[0006] The base station node receives a model fine-tuning request sent by at least one terminal node, and sends a base station-level model fine-tuning request signaling to the core network node;

[0007] The base station node receives the model fine-tuning strategy signaling and the registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling;

[0008] The base station node generates a model fine-tuning configuration signaling according to the model fine-tuning strategy signaling, and sends the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, and transmits the first model to all the terminal nodes, so that all the terminal nodes perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, and obtain the corresponding first local fine-tuning model; wherein, the first model is obtained by model processing of the registered pre-trained model.

[0009] In a possible implementation, after the base station node receives the model fine-tuning strategy signaling and the registered pre-trained model sent by the core network node, the method further includes:

[0010] The base station node selects a number of terminal nodes from the covered cells and / or cell groups based on the user scheduling policy in the model fine-tuning policy signaling to form the model fine-tuning terminal node group.

[0011] In one possible implementation, the method further includes:

[0012] The base station node performs federal model fine-tuning on the first local fine-tuning model based on the local model fine-tuning result uploaded by the terminal node in the model fine-tuning terminal node group and according to the federal model fine-tuning algorithm indicated in the model fine-tuning strategy signaling to obtain a second model;

[0013] The local model fine-tuning result is determined based on the first local fine-tuning model.

[0014] In one possible implementation, the method further includes:

[0015] The base station node sends the second model to all terminal nodes in the model fine-tuning terminal node group, so that all terminal nodes perform local model fine-tuning on the second model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling to obtain the corresponding second local fine-tuning model.

[0016] In one possible implementation, the method further includes:

[0017] The base station node receives a model fine-tuning termination request signaling sent by a terminal node in the model fine-tuning terminal node group, and deletes the terminal node from the model fine-tuning terminal node group according to the model fine-tuning termination request signaling;

[0018] If the base station node determines that there is no terminal node in the model fine-tuning terminal node group, it stops performing model fine-tuning on the registered pre-trained model.

[0019] In one possible implementation, the method further includes:

[0020] If the base station node receives the model fine-tuning termination signaling sent by the core network node, it stops performing model fine-tuning on the registered pre-trained model.

[0021] In a possible implementation, the model fine-tuning strategy signaling includes at least two of a model fine-tuning algorithm indication, a synchronization fine-tuning indication, a synchronization fine-tuning time interval indication, a user scheduling strategy indication, a model processing strategy indication, a model deployment mode indication, a privacy protection level indication, a service quality indication, and a federated model fine-tuning algorithm indication;

[0022] Among them, the model fine-tuning algorithm indication is used to instruct the terminal node to use a first preset algorithm to perform local model fine-tuning on the first model; the synchronous fine-tuning indication is used to instruct the base station node to control the terminal node to adopt synchronous fine-tuning or asynchronous fine-tuning to perform model fine-tuning method; the synchronous fine-tuning time interval indication is used to indicate the uniform time interval length set under synchronous fine-tuning; the user scheduling policy indication is used to instruct the base station node to select at least one terminal node to perform local model fine-tuning on the first model; the model processing strategy indication is used to indicate the model processing rules performed by the base station node on the registered pre-trained model; the model deployment method indication is used to indicate the rules for deploying the registered pre-trained model on the terminal node; the privacy protection level indication is used to indicate the protection level of the registered pre-trained model; the service quality indication is used to indicate the service quality performance requirements of the base station node for terminal node model fine-tuning; the federal model fine-tuning algorithm indication is used to indicate the second preset algorithm used by the base station node when aggregating the local model fine-tuning results corresponding to all terminal nodes.

[0023] According to a second aspect of the present disclosure, a model fine-tuning method is provided, which is applied to a model fine-tuning system, wherein the model fine-tuning system includes a terminal node, a base station node, and a core network node. The method includes:

[0024] The terminal node sends a model fine-tuning request signaling to the base station node and reports model fine-tuning related information related to the terminal node, so that the base station node sends a base station-level model fine-tuning request signaling to the core network node and receives the model fine-tuning policy signaling and the registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling;

[0025] The terminal node receives the model fine-tuning configuration signaling and the first model sent by the base station node; wherein the model fine-tuning configuration signaling is generated according to the model fine-tuning strategy signaling; the first model is obtained by model processing the registered pre-trained model;

[0026] The terminal node collects a local fine-tuning data set, performs local model fine-tuning based on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, obtains a first local fine-tuning model, and reports the local model fine-tuning result determined according to the first local fine-tuning model to the base station node.

[0027] In a possible implementation, the terminal node-related model fine-tuning related information includes at least one of the terminal node data type, terminal node model processing capability, terminal node model storage capability, model fine-tuning application scenario, service quality performance requirements, model monitoring related information, and model fine-tuning time limit requirements.

[0028] In a possible implementation manner, the model fine-tuning request signaling is carried via uplink control information, a media access control layer control unit, or a signaling radio bearer.

[0029] In a possible implementation, the local model fine-tuning result is at least one of a local fine-tuning model gradient, a local fine-tuning model loss value, or a local fine-tuning model parameter updated after model fine-tuning.

[0030] In a possible implementation manner, the local model fine-tuning result is reported to the base station node via uplink control information of the control plane or a data radio bearer of the user plane.

[0031] According to a third aspect of the present disclosure, a model fine-tuning device is provided, which is applied to a model fine-tuning system. The model fine-tuning system includes a terminal node, a base station node, and a core network node. The device includes:

[0032] A first processing unit is configured to receive, by the base station node, a model fine-tuning request sent by at least one of the terminal nodes, and send base station-level model fine-tuning request signaling to the core network node;

[0033] A second processing unit is configured to receive, by the base station node, a model fine-tuning strategy signaling and a registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling;

[0034] A model fine-tuning unit is used for the base station node to generate a model fine-tuning configuration signaling according to the model fine-tuning strategy signaling, and send the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, and transmit the first model to all the terminal nodes, so that all the terminal nodes perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, and obtain the corresponding first local fine-tuning model; wherein, the first model is obtained by model processing of the registered pre-trained model.

[0035] In a possible implementation manner, the second processing unit is further configured to:

[0036] The base station node selects a number of terminal nodes from the covered cells and / or cell groups based on the user scheduling policy in the model fine-tuning policy signaling to form the model fine-tuning terminal node group.

[0037] In a possible implementation, the model fine-tuning unit is further configured to:

[0038] The base station node performs federal model fine-tuning on the first local fine-tuning model based on the local model fine-tuning result uploaded by the terminal node in the model fine-tuning terminal node group and according to the federal model fine-tuning algorithm indicated in the model fine-tuning strategy signaling to obtain a second model;

[0039] The local model fine-tuning result is determined based on the first local fine-tuning model.

[0040] In a possible implementation, the model fine-tuning unit is further configured to:

[0041] The base station node sends the second model to all terminal nodes in the model fine-tuning terminal node group, so that all terminal nodes perform local model fine-tuning on the second model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling to obtain the corresponding second local fine-tuning model.

[0042] In a possible implementation, the model fine-tuning unit is further configured to:

[0043] The base station node receives a model fine-tuning termination request signaling sent by a terminal node in the model fine-tuning terminal node group, and deletes the terminal node from the model fine-tuning terminal node group according to the model fine-tuning termination request signaling;

[0044] If the base station node determines that there is no terminal node in the model fine-tuning terminal node group, it stops performing model fine-tuning on the registered pre-trained model.

[0045] In a possible implementation, the model fine-tuning unit is further configured to:

[0046] If the base station node receives the model fine-tuning termination signaling sent by the core network node, it stops performing model fine-tuning on the registered pre-trained model.

[0047] In a possible implementation, the model fine-tuning strategy signaling includes at least two of a model fine-tuning algorithm indication, a synchronization fine-tuning indication, a synchronization fine-tuning time interval indication, a user scheduling strategy indication, a model processing strategy indication, a model deployment mode indication, a privacy protection level indication, a service quality indication, and a federated model fine-tuning algorithm indication;

[0048] Among them, the model fine-tuning algorithm indication is used to instruct the terminal node to use a first preset algorithm to perform local model fine-tuning on the first model; the synchronous fine-tuning indication is used to instruct the base station node to control the terminal node to adopt synchronous fine-tuning or asynchronous fine-tuning to perform model fine-tuning method; the synchronous fine-tuning time interval indication is used to indicate the uniform time interval length set under synchronous fine-tuning; the user scheduling policy indication is used to instruct the base station node to select at least one terminal node to perform local model fine-tuning on the first model; the model processing strategy indication is used to indicate the model processing rules performed by the base station node on the registered pre-trained model; the model deployment method indication is used to indicate the rules for deploying the registered pre-trained model on the terminal node; the privacy protection level indication is used to indicate the protection level of the registered pre-trained model; the service quality indication is used to indicate the service quality performance requirements of the base station node for terminal node model fine-tuning; the federal model fine-tuning algorithm indication is used to indicate the second preset algorithm used by the base station node when aggregating the local model fine-tuning results corresponding to all terminal nodes.

[0049] According to a fourth aspect of the present disclosure, a model fine-tuning device is provided, which is applied to a model fine-tuning system. The model fine-tuning system includes a terminal node, a base station node, and a core network node. The device includes:

[0050] A sending unit, configured for the terminal node to send a model fine-tuning request signaling to the base station node, and to report model fine-tuning related information related to the terminal node, so that the base station node sends a base station-level model fine-tuning request signaling to the core network node, and receives a model fine-tuning policy signaling and a registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling;

[0051] A receiving unit, configured for the terminal node to receive a model fine-tuning configuration signaling and a first model sent by the base station node; wherein the model fine-tuning configuration signaling is generated according to the model fine-tuning strategy signaling; and the first model is obtained by model processing the registered pre-trained model;

[0052] A model fine-tuning unit is used for the terminal node to collect a local fine-tuning data set, perform local model fine-tuning based on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, obtain a first local fine-tuning model, and report the local model fine-tuning result determined according to the first local fine-tuning model to the base station node.

[0053] In a possible implementation, the terminal node-related model fine-tuning related information includes at least one of the terminal node data type, terminal node model processing capability, terminal node model storage capability, model fine-tuning application scenario, service quality performance requirements, model monitoring related information, and model fine-tuning time limit requirements.

[0054] In a possible implementation manner, the model fine-tuning request signaling is carried via uplink control information, a media access control layer control unit, or a signaling radio bearer.

[0055] In a possible implementation, the local model fine-tuning result is at least one of a local fine-tuning model gradient, a local fine-tuning model loss value, or a local fine-tuning model parameter updated after model fine-tuning.

[0056] In a possible implementation manner, the local model fine-tuning result is reported to the base station node via uplink control information of the control plane or a data radio bearer of the user plane.

[0057] According to a fifth aspect of the present disclosure, a model fine-tuning system is provided, comprising: a terminal node, a base station node, and a core network node, wherein:

[0058] The terminal node sends a model fine-tuning request signaling to the base station node, and reports model fine-tuning related information related to the terminal node;

[0059] The base station node receives a model fine-tuning request sent by at least one terminal node, and sends a base station-level model fine-tuning request signaling to the core network node;

[0060] The base station node receives the model fine-tuning strategy signaling and the registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling;

[0061] The base station node generates a model fine-tuning configuration signaling according to the model fine-tuning strategy signaling, and sends the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, and transmits the first model to all the terminal nodes;

[0062] The terminal node collects a local fine-tuning data set, performs local model fine-tuning based on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, obtains a first local fine-tuning model, and reports the local model fine-tuning result determined according to the first local fine-tuning model to the base station node.

[0063] In one possible implementation, the base station node performs federal model fine-tuning on the first local fine-tuning model based on the local model fine-tuning results uploaded by the terminal nodes in the model fine-tuning terminal node group and in accordance with the federal model fine-tuning algorithm indicated in the model fine-tuning strategy signaling to obtain a second model.

[0064] In one possible implementation, the base station node sends the second model to all terminal nodes in the model fine-tuning terminal node group, so that all terminal nodes perform local model fine-tuning on the second model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling to obtain the corresponding second local fine-tuning model.

[0065] In a possible implementation, the terminal node-related model fine-tuning related information includes at least one of the terminal node data type, terminal node model processing capability, terminal node model storage capability, model fine-tuning application scenario, service quality performance requirements, model monitoring related information, and model fine-tuning time limit requirements.

[0066] In a possible implementation manner, the model fine-tuning request signaling is carried via uplink control information, a media access control layer control unit, or a signaling radio bearer.

[0067] In a possible implementation, the local model fine-tuning result is at least one of a local fine-tuning model gradient, a local fine-tuning model loss value, or a local fine-tuning model parameter updated after model fine-tuning.

[0068] In a possible implementation manner, the local model fine-tuning result is reported to the base station node via uplink control information of the control plane or a data radio bearer of the user plane.

[0069] According to a sixth aspect of the present disclosure, a computer program product is provided, comprising a computer program, wherein when the computer program is executed by a processor, the method of the first and second aspects and possible implementation methods thereof are implemented.

[0070] According to the seventh aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the methods of the above-mentioned first and second aspects and their possible implementation methods by executing the executable instructions.

[0071] The technical solution disclosed in this disclosure has the following beneficial effects:

[0072] In an embodiment of the present disclosure, since the model fine-tuning system includes terminal nodes, base station nodes, and core network nodes, the base station node can receive a model fine-tuning request sent by at least one terminal node and send a base station-level model fine-tuning request signaling to the core network node; and the base station node can receive a model fine-tuning policy signaling and a registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling; thus, the base station node generates a model fine-tuning configuration signaling according to the model fine-tuning policy signaling, and sends the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, transmitting the first model to all terminal nodes so that all terminal nodes perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, thereby obtaining a corresponding first local fine-tuning model; wherein the first model is obtained by model processing the registered pre-trained model. It can be seen that in the embodiment of the present disclosure, based on the collaboration of the base station node, the terminal node, and the core network node, it is possible to quickly implement fine-tuning of the artificial intelligence model, improve the fine-tuning efficiency of the artificial intelligence model, reduce time cost, and reduce system resource overhead.

[0073] Other features and advantages of the present disclosure will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present disclosure. The purposes and other advantages of the present disclosure can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings required for use in the embodiments of the present disclosure. Obviously, the drawings introduced below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0075] Figure 1 A schematic diagram showing a model fine-tuning system according to this exemplary embodiment is shown;

[0076] Figure 2 A flowchart showing a model fine-tuning method in this exemplary embodiment is shown;

[0077] Figure 3 A flowchart showing a model fine-tuning method in this exemplary embodiment is shown;

[0078] Figure 4 An interactive schematic diagram illustrating a model fine-tuning method in this exemplary embodiment;

[0079] Figure 5A schematic structural diagram of a model fine-tuning device in this exemplary embodiment is shown;

[0080] Figure 6 A schematic structural diagram of a model fine-tuning device in this exemplary embodiment is shown;

[0081] Figure 7 A schematic structural diagram of an electronic device in this exemplary embodiment is shown. DETAILED DESCRIPTION

[0082] In order to make the purpose, technical solutions and advantages of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure. In the absence of conflict, the embodiments in the present disclosure and the features in the embodiments can be arbitrarily combined with each other. In addition, although the logical order is shown in the flow chart, in some cases, the steps shown or described can be performed in an order different from that here.

[0083] The terms "comprises" and "comprising" and any variations thereof in the specification and claims of this disclosure are intended to cover non-exclusive protection. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements, but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.

[0084] In the embodiments of the present disclosure, one or more, "many" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: the existence of A alone, the existence of A and B at the same time, and the existence of B alone, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following (items)" or similar expressions refers to any combination of these items, including any combination of single items (items) or plural items (items). For example, at least one item (item) of a, b or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, c can be single or multiple.

[0085] It should be noted that the terms "first", "second", "third", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order, sequence, size and priority. For example, the first terminal node and the second terminal node in the embodiment of the present disclosure are only used to distinguish different terminal nodes. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0086] The following describes exemplary embodiments of the present disclosure in conjunction with the accompanying drawings. The accompanying drawings are schematic diagrams of the present disclosure and are not necessarily drawn to scale. Some of the block diagrams shown in the accompanying drawings may be functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, or in hardware modules or integrated circuits, or in networks, processors or microcontrollers. The embodiments can be implemented in various forms and should not be construed as being limited to the examples set forth herein. The features, structures or characteristics described in the present disclosure can be combined in one or more embodiments in any suitable manner. In the description below, many specific details are provided to provide a full description of the embodiments of the present disclosure. However, those skilled in the art will appreciate that one or more specific details may be omitted when implementing the technical solutions of the present disclosure, or that other methods, components, devices, steps, etc. may be used to replace one or more specific details.

[0087] It should be noted that in the embodiments of this disclosure, certain software, components, models, and other existing industry solutions may be mentioned. These should be considered exemplary and intended solely to illustrate the feasibility of implementing the technical solutions of this disclosure. This does not imply that the applicant has or will necessarily use such solutions. The use, training, and dissemination of model data in this technical solution comply with relevant national laws and regulations.

[0088] Life Cycle Management (LCM) is a key concept currently being proposed in 3GPP standardization research. As a full-process management solution for artificial intelligence (AI) models, LCM encompasses data collection, model training, model inference, model management, and model transmission. Model management includes functionality / model ID selection, model activation / deactivation, switching, rollback, monitoring, updates, and uninstallation. The importance of AI model lifecycle management lies in ensuring the quality and efficiency of AI models throughout their entire lifecycle, potentially reducing operational costs.

[0089] The 3GPP Radio Access Network (RAN) Working Group has conducted standardization research on key air interface AI technologies, combining the powerful representation and feature extraction capabilities of AI models with wireless communication scenarios to optimize time-frequency resources and reduce overhead. When implementing edge intelligence in wireless networks, when the AI ​​model's deployment environment deviates from its training environment, the resulting data distribution mismatch can lead to reduced model performance. Applying model fine-tuning techniques can significantly improve this issue.

[0090] Model fine-tuning technology involves retraining a pre-trained model using task-specific data to better adapt it to new tasks or datasets. Model fine-tuning requires establishing a communication link between participating nodes to process business flows such as forward reasoning results, backward gradients, and model parameters. This significantly saves time and computing resources, as only the pre-trained model needs to be fine-tuned, eliminating the need to train from scratch. Furthermore, model fine-tuning can improve the generalization and flexibility of AI models for new tasks, enhancing the application prospects of AI algorithms in network edge scenarios.

[0091] Furthermore, according to white papers published by organizations such as 3GPP, ITU-T, and CCSA, 5G-A / 6G networks must deliver higher system performance and lower latency, integrating AI technologies to enable more intelligent application scenarios. With the increasing number of user terminals at the network edge and the growing demand for AI services, there is a pressing need to leverage the cloud's powerful computing power to support the network edge and enable the coordinated deployment of large language models in the cloud and smaller models at the edge to meet the requirements of future network intelligence.

[0092] In view of one or more of the above problems, an exemplary embodiment of the present disclosure provides a model fine-tuning method. Since the model fine-tuning system includes terminal nodes, base station nodes and core network nodes, the base station node can receive a model fine-tuning request sent by at least one terminal node and send a base station-level model fine-tuning request signaling to the core network node; and the base station node can receive a model fine-tuning policy signaling and a registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling; in this way, the base station node generates a model fine-tuning configuration signaling according to the model fine-tuning policy signaling, and sends the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, and transmits the first model to all terminal nodes, so that all terminal nodes perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, and obtain the corresponding first local fine-tuning model; wherein the first model is obtained by model processing of the registered pre-trained model. It can be seen that in the embodiment of the present disclosure, based on the collaboration of base station nodes, terminal nodes and core network nodes, it is possible to quickly fine-tune the artificial intelligence model, improve the efficiency of fine-tuning the artificial intelligence model, reduce time costs, and reduce system resource overhead.

[0093] To better understand the technical solutions provided by the embodiments of the present disclosure, the following briefly introduces the system architecture applicable to the technical solutions provided by the embodiments of the present disclosure. It should be noted that the system architecture introduced below is only used to illustrate the embodiments of the present disclosure and is not intended to limit the scope of the present disclosure. In specific implementations, the technical solutions provided by the embodiments of the present disclosure can be flexibly applied according to actual needs.

[0094] See also Figure 1 As shown, Figure 1 A schematic diagram of a model fine-tuning system in this exemplary embodiment is shown. The model fine-tuning system includes a terminal node 101, a base station node 102, and a core network node 103. The terminal node 101 is deployed in a user equipment (UE), the base station node 102 is deployed in an access network centralized unit (CU) or an access network distributed unit (DU), and the core network node is deployed in the core network's network management system, which has an orchestration and management (OAM) or a third-party server (Over-the-top Server, OTT Server) or a network data analytics function (NWDAF).

[0095] In an embodiment of the present disclosure, the terminal node 101 is configured to send a model fine-tuning request signaling to the base station node 102, and report the model fine-tuning related information related to the terminal node. And, according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, local model fine-tuning is performed on the first model to obtain the corresponding first local fine-tuning model; wherein the first model is obtained by model processing of the registered pre-trained model. It should be noted that the first model can be understood as a pre-trained model that has been processed by the model or a model group consisting of multiple pre-trained models that have been processed by the model. The base station node 102 is configured to receive a model fine-tuning request sent by at least one terminal node 101, and send a base station-level model fine-tuning request signaling to the core network node 103, and, according to the model fine-tuning policy signaling, generate a model fine-tuning configuration signaling, and send the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group. The core network node 103 is configured to respond to the base station-level model fine-tuning request, select a pre-trained model that meets the requirements from the model storage module to form a pre-trained model / model group according to the base station-level model fine-tuning related information carried in the model fine-tuning request signaling, perform model identification and model registration, generate a model fine-tuning strategy signaling and send it to the base station node 102, and at the same time, the core network node sends the registered pre-trained model / model group to the base station node 102.

[0096] In the disclosed embodiment, terminal node 101 includes a model fine-tuning module, a model monitoring module, and a data acquisition module. The model fine-tuning module is used to fine-tune the model delivered by base station node 102. The model monitoring module is used to monitor the fine-tuning status, i.e., the training status, of the model delivered by base station node 102. When the results on the local validation set meet the model fine-tuning termination criteria, the model fine-tuning process is terminated and a model fine-tuning termination request signaling is sent to the base station. The data acquisition module is used to collect online or offline data to form a local fine-tuning dataset.

[0097] In an embodiment of the present disclosure, the base station node 102 includes a model fine-tuning module, a model management module, a model storage module, etc., wherein the model fine-tuning module is used to perform model processing on the pre-trained model / model group issued by the core network node 103, the model management module is used to manage the model, and the model storage module is used to store each received model.

[0098] In the disclosed embodiment, the core network node 103 includes a policy formulation module, a pre-training module, a model storage module, etc. Among them, the policy formulation module is used to formulate a model fine-tuning strategy, the pre-training module is used to train the artificial intelligence model, and the model storage module stores pre-trained models to form a pre-trained model / model group.

[0099] In an embodiment of the present disclosure, the base station node 102 receives a model fine-tuning request sent by at least one terminal node 101, and sends a base station-level model fine-tuning request signaling to the core network node; the base station node 102 receives a model fine-tuning policy signaling and a registered pre-trained model sent by the core network node 103; the pre-trained model is determined according to the base station-level model fine-tuning request signaling; the base station node 102 generates a model fine-tuning configuration signaling according to the model fine-tuning policy signaling, and sends the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, and transmits the first model to all terminal nodes, so that all terminal nodes 101 perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, and obtain the corresponding first local fine-tuning model; wherein the first model is obtained by model processing of the registered pre-trained model.

[0100] To further illustrate the technical solutions provided by the embodiments of the present disclosure, this is described in detail below in conjunction with the accompanying drawings and specific implementation methods. Although the embodiments of the present disclosure provide the method operation steps as shown in the following embodiments or drawings, more or fewer operation steps may be included in the method based on routine or no creative labor. In steps where there is no necessary causal relationship logically, the execution order of these steps is not limited to the execution order provided by the embodiments of the present disclosure. In the actual processing process or when the device is executed, the method may be executed in the order of the methods shown in the embodiments or drawings or in parallel.

[0101] See Figure 2 , Figure 2 FIG is a flow chart of a model fine-tuning method in an embodiment of the present disclosure. Figure 1 The base station node 102 in the embodiment of the present invention performs the following specific implementation process:

[0102] Step 201: A base station node receives a model fine-tuning request sent by at least one terminal node, and sends a base station-level model fine-tuning request signaling to a core network node.

[0103] Step 202: The base station node receives the model fine-tuning strategy signaling and the registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling.

[0104] Step 203: The base station node generates a model fine-tuning configuration signaling according to the model fine-tuning strategy signaling, and sends the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, and transmits the first model to all terminal nodes, so that all terminal nodes perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, and obtain the corresponding first local fine-tuning model; wherein the first model is obtained by model processing of the registered pre-trained model.

[0105] In an embodiment of the present disclosure, since the model fine-tuning system includes terminal nodes, base station nodes, and core network nodes, the base station node can receive a model fine-tuning request sent by at least one terminal node and send a base station-level model fine-tuning request signaling to the core network node; and the base station node can receive a model fine-tuning policy signaling and a registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling; thus, the base station node generates a model fine-tuning configuration signaling according to the model fine-tuning policy signaling, and sends the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, transmitting the first model to all terminal nodes so that all terminal nodes perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, thereby obtaining a corresponding first local fine-tuning model; wherein the first model is obtained by model processing the registered pre-trained model. It can be seen that in the embodiment of the present disclosure, based on the collaboration of the base station node, the terminal node, and the core network node, it is possible to quickly implement fine-tuning of the artificial intelligence model, improve the fine-tuning efficiency of the artificial intelligence model, reduce time cost, and reduce system resource overhead.

[0106] See Figure 3 , Figure 3 FIG is a flow chart of a model fine-tuning method in an embodiment of the present disclosure. Figure 1 The terminal node 101 in the embodiment of the present invention is executed, and the specific implementation process of the method is as follows:

[0107] Step 301: The terminal node sends a model fine-tuning request signaling to the base station node and reports the model fine-tuning related information related to the terminal node, so that the base station node sends a base station-level model fine-tuning request signaling to the core network node and receives the model fine-tuning strategy signaling and the registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling;

[0108] Step 302: The terminal node receives a model fine-tuning configuration signaling and a first model sent by the base station node; wherein the model fine-tuning configuration signaling is generated according to the model fine-tuning strategy signaling; and the first model is obtained by model processing a registered pre-trained model.

[0109] Step 303: The terminal node collects a local fine-tuning data set, performs local model fine-tuning based on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, obtains a first local fine-tuning model, and reports the local model fine-tuning result determined according to the first local fine-tuning model to the base station node.

[0110] In the embodiment of the present disclosure, since the model fine-tuning system includes terminal nodes, base station nodes and core network nodes, artificial intelligence model fine-tuning can be achieved based on the collaboration of base station nodes, terminal nodes and core network nodes. In this way, compared with the method in which a single terminal device alone achieves artificial intelligence model fine-tuning in the related technology, artificial intelligence model fine-tuning can be achieved quickly, the fine-tuning efficiency of the artificial intelligence model can be improved, the time cost can be reduced, and the system resource overhead can be reduced.

[0111] The following is an introduction to the model fine-tuning method provided by the embodiment of the present disclosure using a specific example. Figure 4 shown.

[0112] Step 401: The terminal node sends a model fine-tuning request signaling to the base station node, and reports model fine-tuning related information related to the terminal node.

[0113] In the disclosed embodiments, each terminal node in the model fine-tuning system can send a model fine-tuning request signaling to a base station node and report model fine-tuning related information related to the terminal node. In this way, the base station node can receive a model fine-tuning request sent by at least one terminal node and the model fine-tuning related information reported by the terminal node.

[0114] In an optional embodiment, the terminal node-related model fine-tuning related information includes at least one of the terminal node data type, the terminal node model processing capability, the terminal node model storage capability, the model fine-tuning application scenario, the QoS performance requirements, the model monitoring related information, and the model fine-tuning time limit requirements. In other words, the terminal node-related model fine-tuning related information includes the terminal node data type, the terminal node model processing capability, and the terminal node model storage capability, or the terminal node-related model fine-tuning related information includes the terminal node model processing capability, the terminal node model storage capability, and the model fine-tuning time limit requirements, which are not limited in the embodiments of the present disclosure.

[0115] In an optional implementation manner, the model fine-tuning request signaling is carried by uplink control information (Uplink Control Information, UCI).

[0116] In an optional implementation manner, the model fine-tuning request signaling is carried by a media access control layer control element (MAC Control Element, MAC CE).

[0117] In an optional implementation manner, the model fine-tuning request signaling is carried via a signaling radio bearer (SRB).

[0118] Step 402: The base station node receives and responds to the model fine-tuning request from one or more terminal nodes, and sends a base station-level model fine-tuning request signaling to the core network node.

[0119] In an embodiment of the present disclosure, a base station node may receive and respond to model fine-tuning requests from one or more terminal nodes, and send base station-level model fine-tuning request signaling to a core network node.

[0120] In an optional implementation, the base station-level model fine-tuning request signaling may be forwarded based on a user plane function (UPF) network element on the user plane.

[0121] In an optional implementation, the base station-level model fine-tuning request signaling may be forwarded by an Access and Mobility Management Function (AMF) network element of the control plane.

[0122] In the embodiment of the present disclosure, the base station-level model fine-tuning request signaling carries base station-level model fine-tuning related information.

[0123] In an optional embodiment, the model fine-tuning related information at the base station level includes the base station identifier, the model fine-tuning deployment range, the model fine-tuning application scenario, the network load information of each cell / cell group, the computing and storage capacity of each user / user group, the QoS requirement list of each user / user group, and at least one of the model fine-tuning time limit requirements of each user / user group.

[0124] Step 403: The core network node responds to the model fine-tuning request, selects a pre-trained model that meets the requirements from the model storage module according to the base station-level model fine-tuning related information carried in the model fine-tuning request signaling to form a pre-trained model / model group, and performs model identification and model registration.

[0125] In an embodiment of the present disclosure, the core network node responds to a base station-level model fine-tuning request, selects a pre-trained model and / or a model group formed based on the pre-trained model that meets the preset conditions from the model storage module according to the base station-level model fine-tuning related information carried in the model fine-tuning request signaling, and performs model identification and model registration on the pre-trained model / model group formed based on the pre-trained model that meets the preset conditions.

[0126] In an optional embodiment, the preset condition is, for example, that the quantitative score reaches a preset threshold or that the quantitative score is topM. Among them, the quantitative score in the embodiment of the present disclosure can be determined based on the compliance of each pre-trained model with the sub-information in the model fine-tuning related information at the base station level. Optionally, the quantitative score can be determined by the compliance of each pre-trained model with the sub-information in the model fine-tuning related information at the base station level, or it can be comprehensively determined based on the compliance of multiple pre-trained models with the sub-information in the model fine-tuning related information at the base station level.

[0127] For example, the compliance of the application scenarios in the model fine-tuning related information at the base station level with each pre-trained model is quantitatively scored to obtain a first quantitative score set; the compliance of the QoS requirements in the model fine-tuning related information at the base station level with each pre-trained model is quantitatively scored to obtain a second quantitative score set; the compliance of the deployment scope in the model fine-tuning related information at the base station level with each pre-trained model is quantitatively scored to obtain a third quantitative score set.

[0128] Optionally, the core network node can select any quantization score set from the first quantization score set, the second quantization score set, and the third quantization score set, and based on the sorting method, select the top 5 pre-trained models with the highest sorting scores to form a pre-trained model group, or select the pre-trained model with the highest quantization score as the pre-trained model sent to the base station node.

[0129] Optionally, the core network node can also comprehensively determine the comprehensive quantitative score of each pre-trained model based on the first quantitative score set, the second quantitative score set, and the third quantitative score set, thereby obtaining a comprehensive quantitative score set, and then based on the sorting method, select the top 7 pre-trained models with the highest comprehensive quantitative scores from the comprehensive quantitative score set to form a pre-trained model group, or select the pre-trained model with the highest comprehensive quantitative score as the pre-trained model sent to the base station node.

[0130] In one possible implementation, the pre-trained model in the model storage module may be at least one large language model, and the generation method of the large language model may be a typical generation method based on existing generative artificial intelligence models, such as generative adversarial networks (GANs), variational autoencoders (VAEs), diffusion models, and transformer models.

[0131] In one possible embodiment, the process of performing model identification on a pre-trained model that meets preset conditions and / or a model group formed based on a pre-trained model includes identifying the model structure, model parameters, model storage format, model storage space size, model input and output methods, model performance, etc. of the selected pre-trained model.

[0132] In one possible implementation, the process of model registration for a pre-trained model that meets preset conditions and / or a model group formed based on a pre-trained model includes indexing the identified pre-trained model and forming a model identification list, so that the registration of the model identification list can be completed at the core network node.

[0133] In a possible implementation, the model storage module of the core network node may be deployed in a distributed manner on the access network side, or in a centralized manner on the core network side.

[0134] Step 404: The core network node generates a model fine-tuning strategy signaling and sends it to the base station node. At the same time, the core network node sends the registered pre-trained model / model group to the base station node.

[0135] In an embodiment of the present disclosure, the core network node can generate model fine-tuning strategy signaling and send it to the base station node. At the same time, the core network node can send the registered pre-trained model and / or model group to the base station node.

[0136] In one possible implementation, the core network node sends the model file corresponding to the pre-trained model / model group to the CU-UP plane of the base station node through the UPF network element from the E2 interface, and the model fine-tuning strategy signaling is sent to the base station node through the user plane UPF network element or the control plane AMF network element.

[0137] In a possible implementation, the model fine-tuning strategy signaling includes at least two of a model fine-tuning algorithm indication, a synchronization fine-tuning indication, a synchronization fine-tuning time interval indication, a user scheduling strategy indication, a model processing strategy indication, a model deployment mode indication, a privacy protection level indication, a QoS indication, and a federated model fine-tuning algorithm indication;

[0138] Among them, the model fine-tuning algorithm indication is used to instruct the terminal node to use the first preset algorithm to perform local model fine-tuning on the first model; the synchronous fine-tuning indication is used to instruct the base station node to control the terminal node to adopt synchronous fine-tuning or asynchronous fine-tuning to perform the model fine-tuning method; the synchronous fine-tuning time interval indication is used to indicate the uniform time interval length set under synchronous fine-tuning; the user scheduling policy indication is used to instruct the base station node to select at least one terminal node to perform local model fine-tuning on the first model; the model processing strategy indication is used to indicate the model processing rules performed by the base station node on the registered pre-trained model; the model deployment method indication is used to indicate the rules for deploying the registered pre-trained model on the terminal node; the privacy protection level indication is used to indicate the protection level of the registered pre-trained model; the service quality indication is used to indicate the QoS performance requirements of the base station node for the terminal node model fine-tuning; the federal model fine-tuning algorithm indication is used to indicate the second preset algorithm used by the base station node when aggregating the local model fine-tuning results corresponding to all terminal nodes.

[0139] For example, see Table 1, which is a schematic table of model fine-tuning strategy signaling in an embodiment of the present disclosure.

[0140]

[0141] Table 1

[0142] Step 405: The base station node processes the registered pre-trained model / model group based on the model processing strategy in the model fine-tuning strategy signaling to form a pre-trained model / model group that has undergone model processing.

[0143] In the disclosed embodiments, the base station node processes the registered pre-trained model / model group based on the model processing strategy in the model fine-tuning strategy signaling to form a pre-trained model / model group that has undergone model processing. For ease of description, this is referred to as the first model below. That is, the first model in the disclosed embodiments can be understood as a pre-trained model that has undergone model processing, or as a model group composed of pre-trained models that have undergone model processing.

[0144] In the embodiments of the present disclosure, the base station node may process the registered pre-trained model / model group in the following ways, but not limited to:

[0145] Method 1.

[0146] The base station node modifies and customizes the model modules, parameters, etc. Optionally, the concept of a small neural network adapter can be introduced, that is, the registered pre-trained model / model group is processed by updating only the adapter model parameters instead of all pre-trained model / model group parameters to obtain the first model.

[0147] This approach can meet the requirements of modifying and customizing the registered pre-trained model / model group while reducing the computational and storage overhead incurred during model fine-tuning.

[0148] Method 2.

[0149] The base station node can process the registered pre-trained model / model group by adding a low-rank matrix operation module to the convolution layer or fully connected layer of the registered pre-trained model / model group to replace the original matrix operation module to obtain the first model.

[0150] This approach can reduce the complexity of matrix operations, thereby improving the training efficiency of registered pre-trained models / model groups.

[0151] It should be noted that in the embodiment of the present disclosure, the registered pre-trained model / model group can also be processed in combination with the aforementioned method one and method two to obtain the first model. Of course, the registered pre-trained model / model group can also be processed based on other methods, which is not limited in the embodiment of the present disclosure.

[0152] In a possible implementation, the base station node may store the model file corresponding to the first model / model group in at least one of a centralized unit (CU), a distributed unit (DU), or a database close to the base station node.

[0153] Step 406: The base station node selects several terminal nodes from the covered cells / cell groups based on the user scheduling policy in the model fine-tuning policy signaling to form a model fine-tuning terminal node group.

[0154] In an embodiment of the present disclosure, the base station node may select several terminal nodes from the covered cells and / or cell groups based on the user scheduling policy in the model fine-tuning policy signaling to form a model fine-tuning terminal node group.

[0155] In a possible implementation, the base station node may perform user scheduling in a list sorting manner, wherein parameters of the list sorting may be determined based on actual conditions and are not limited in the embodiments of the present disclosure.

[0156] In one possible implementation, the base station node may determine the channel state information obtained by each terminal node based on a channel state information reference signal (CSI-RS), sort the channel states in descending order, and select the top K terminal nodes to form a model fine-tuning terminal node group, where K is a positive integer.

[0157] In one possible implementation, the base station node may determine the time limit requirements for model fine-tuning for each terminal node, sort the specific durations corresponding to the time limit requirements in ascending order, and select the top N terminal nodes to form a model fine-tuning terminal node group. N depends on the number of terminal nodes in the cell / cell group that report model fine-tuning request signaling. N should be less than or equal to the total number of terminal nodes of this type, and N is a positive integer.

[0158] In a possible implementation, the base station node may transmit user scheduling notification signaling via a downlink broadcast channel, and broadcast relevant information of the model fine-tuning terminal node group to all terminal nodes in the covered cell / cell group.

[0159] Optionally, the user scheduling notification signaling is carried by System Information Block Type 1 (SIB1).

[0160] In one possible implementation, the model fine-tuning terminal node group is autonomously controlled by the base station node, supporting the real-time addition of new terminal nodes or the deletion of existing terminal nodes. In other words, in this embodiment of the present disclosure, the base station node can perform operations such as adding and deleting terminal nodes included in the model fine-tuning terminal node group.

[0161] Step 407: The base station node generates a model fine-tuning configuration signaling based on the model fine-tuning strategy signaling, and sends it to all terminal nodes in the model fine-tuning terminal node group, and transmits the pre-trained model / model group that has undergone model processing to all terminal nodes.

[0162] In the disclosed embodiment, the base station node generates model fine-tuning configuration signaling based on the model fine-tuning strategy signaling. Optionally, the model fine-tuning configuration signaling can be carried by at least one control signaling of SRB, downlink control information (DCI), MAC CE, or by a user plane data radio bearer (DRB).

[0163] In one possible implementation, the model fine-tuning configuration signaling includes information such as a model fine-tuning algorithm indication, a synchronization fine-tuning indication, a synchronization fine-tuning time interval indication, a model deployment method indication, a privacy protection level indication, a QoS indication, and the like, and can reuse the corresponding fields in the model fine-tuning policy signaling, and also includes a terminal node identification indication, a frequency domain resource allocation indication, a time domain resource allocation indication, a model alignment method, and at least two of the model fine-tuning configuration parameters.

[0164] Among them, the terminal node identifier indicates a specific terminal node in the model fine-tuning terminal node group; the time domain resource allocation indicates the time domain resources allocated for the pre-trained model / model group file transfer; the frequency domain resource allocation indicates the frequency domain resources allocated for the pre-trained model / model group file transfer; the model alignment method indicates the model processing method that the terminal node needs to perform on the local model fine-tuning results when aggregating the local model fine-tuning results at the base station node to maintain alignment in terms of model size, model dimension, etc.; the model fine-tuning configuration parameters indicate that the base station node needs to configure configuration parameters related to model deployment and application scenarios to facilitate the terminal node to perform model fine-tuning based on this model.

[0165] For example, see Table 2, which is a schematic table of a design embodiment of a model fine-tuning configuration DCI signaling in an embodiment of the present disclosure.

[0166]

[0167]

[0168] Table 2

[0169] In an embodiment of the present disclosure, the base station node may send the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, and transmit the first model to all terminal nodes.

[0170] In a possible implementation manner, the base station node transmits the first model in at least one of the following manners:

[0171] The base station node transmits data from the CU to the DU via the F1 interface, and then transmits data from the DU to the terminal node via the Uu interface;

[0172] The base station node transmits data from the DU to the terminal node via the Uu interface;

[0173] The base station node transmits the information from the base station database to the DU through a specific interface, and then transmits it from the DU to the terminal node through the Uu interface.

[0174] Step 408: All terminal nodes in the model fine-tuning terminal node group collect local fine-tuning data sets, perform local model fine-tuning based on the pre-trained model / model group that has been processed by the model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, obtain the local fine-tuning model, and report the local model fine-tuning results to the base station node.

[0175] In an embodiment of the present disclosure, after all terminal nodes in the model fine-tuning terminal node group receive the first model and model fine-tuning configuration signaling, they can collect a local fine-tuning data set, perform local model fine-tuning based on the first model / model group according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, obtain the corresponding first local fine-tuning model, and report the model fine-tuning result determined according to the first local fine-tuning model to the base station node.

[0176] In one possible implementation, the terminal node performs model fine-tuning based on the model fine-tuning algorithm indicated in the model fine-tuning configuration signaling. Optionally, the model fine-tuning algorithm may be, for example, a gradient-based model fine-tuning algorithm. In this manner, the terminal node may use a gradient descent algorithm based on a locally collected fine-tuning dataset to update the obtained pre-trained model / model group that has undergone model processing, i.e., the corresponding first local fine-tuning model, and obtain a gradient value for the updated local fine-tuning model.

[0177] In one possible implementation, the local fine-tuning dataset can be collected through online or offline data collection. The online fine-tuning dataset is a dataset collected in real time by the terminal node via a data collection module. This dataset is relatively small and real-time. The offline fine-tuning dataset is a dataset collected non-real time by the terminal node. This dataset is relatively large and spans a long timeframe. The terminal node can determine whether to use online or offline data collection for the local fine-tuning dataset based on information indicated in the model fine-tuning configuration signaling. This is not limited in the present embodiments.

[0178] In a possible implementation, the local model fine-tuning result may be at least one of a local fine-tuning model gradient, a local fine-tuning model loss value, and a local fine-tuning model parameter updated after model fine-tuning.

[0179] In one possible implementation, the local model fine-tuning result may be reported to the base station node via at least one of the UCI of the control plane or the DRB of the user plane. Optionally, the terminal node selects the time interval for uploading the local model fine-tuning result based on the time interval indication in the model fine-tuning configuration signaling.

[0180] Step 409: All terminal nodes in the model fine-tuning terminal node group monitor the local model fine-tuning process in real time through the model monitoring module. If the results of the local fine-tuning model on the local validation set meet the model fine-tuning termination criteria, the terminal node terminates the model fine-tuning process and sends a model fine-tuning termination request signaling to the base station. The base station responds to the termination request and removes the terminal node from the model fine-tuning terminal node group.

[0181] In the disclosed embodiments, all terminal nodes in the model fine-tuning terminal node group can monitor the local model fine-tuning process in real time through their respective model monitoring modules. Thus, if the results of the local fine-tuning model on the local validation set meet the model fine-tuning termination conditions, the terminal node terminates the model fine-tuning process and sends a model fine-tuning termination request signaling to the base station. The base station then responds to the termination request and removes the terminal node from the model fine-tuning terminal node group.

[0182] In one possible implementation, the model fine-tuning termination condition includes at least one of the following:

[0183] The loss value on the local validation set is lower than the first threshold, or the loss value of T consecutive rounds is lower than the second threshold;

[0184] The accuracy on the local validation set is higher than the third threshold;

[0185] Other evaluation metrics on the local validation set exceed the corresponding fourth threshold. Examples of other evaluation metrics include spectral efficiency, throughput, latency, and energy consumption. In other words, if spectral efficiency on the local validation set exceeds the corresponding fourth threshold, throughput on the local validation set exceeds the corresponding fourth threshold, latency on the local validation set exceeds the corresponding fourth threshold, and energy consumption on the local validation set exceeds the corresponding fourth threshold, these can all serve as model fine-tuning termination conditions.

[0186] Obviously, in a possible implementation method, the model fine-tuning termination condition can be understood as the parameter value corresponding to the evaluation index on the local verification set exceeds the corresponding threshold, then it is determined that the model fine-tuning termination condition is met. The evaluation index is, for example, accuracy, spectrum efficiency, throughput, latency, energy consumption, etc.

[0187] It should be noted that the specific values ​​of the aforementioned "first threshold", "second threshold", "third threshold" and "fourth threshold" can be the same or different, and can all be based on the actual implementation of the corresponding color child, which is not limited in the embodiments of the present disclosure.

[0188] In a possible implementation, the model fine-tuning termination request signaling includes at least one of the following information: terminal node identification information, a termination request flag, model monitoring information, and local fine-tuning log information.

[0189] In a possible implementation, the model fine-tuning termination request signaling may be carried by UCI, MAC CE, or SRB.

[0190] Step 410: The base station node performs federated model fine-tuning based on the local model fine-tuning result uploaded by the terminal node and in accordance with the federated model fine-tuning algorithm indicated in the model fine-tuning strategy signaling.

[0191] In an embodiment of the present disclosure, the base station node performs federal model fine-tuning on the first local fine-tuning model based on the local model fine-tuning results uploaded by the terminal nodes in the model fine-tuning terminal node group, and obtains a second model according to the federal model fine-tuning algorithm indicated in the model fine-tuning policy signaling; wherein the local model fine-tuning results are determined based on the first local fine-tuning model.

[0192] In one possible implementation, the federated model fine-tuning algorithm may be implemented as follows: the base station node weights and sums the local fine-tuning model gradient values ​​uploaded by each terminal node in the model fine-tuning terminal node group to obtain an aggregated federated fine-tuning model weight value. Based on the aggregated federated fine-tuning model weight value and the pre-trained model / model group of the current round, a federated fine-tuned pre-trained model / model group is obtained, i.e., the second model. Each model in the pre-trained model group corresponds to a federated fine-tuning model weight value.

[0193] In one possible implementation, the federated fine-tuning model weight value can be set as the ratio of the sample size of the fine-tuning dataset of each terminal node participating in this round of model fine-tuning to the sum of the sample sizes of the fine-tuning datasets of all terminal nodes that upload local model fine-tuning results in this round.

[0194] Step 411: The base station node sends the pre-trained model / model group that has undergone federal fine-tuning to all terminal nodes in the model fine-tuning terminal node group, and repeats steps 409 to 410 until the model fine-tuning terminal node group is empty, or the core network node sends a model fine-tuning termination signaling to terminate this model fine-tuning process.

[0195] In an embodiment of the present disclosure, the base station node sends the second model to all terminal nodes in the model fine-tuning terminal node group, so that all terminal nodes perform local model fine-tuning on the second model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, and obtain the corresponding second local fine-tuning model.

[0196] In an embodiment of the present disclosure, the terminal node may also transmit a local model fine-tuning result corresponding to the second local fine-tuning model to the base station node. In this way, the base station node may further perform federated model fine-tuning on the second local fine-tuning model based on the latest local model fine-tuning result uploaded by the terminal nodes in the model fine-tuning terminal node group and in accordance with the federated model fine-tuning algorithm indicated in the model fine-tuning policy signaling, thereby obtaining a third model. The third model is then distributed to all terminal nodes in the model fine-tuning terminal node group, so that all terminal nodes perform local model fine-tuning on the third model according to the model fine-tuning configuration information included in the model fine-tuning configuration signaling, thereby obtaining the corresponding third local fine-tuning model.

[0197] That is, in the disclosed embodiment, the base station node can continuously determine a new model after executing federated model fine-tuning, and then send this model to the terminal node for model fine-tuning. This means that the federated model fine-tuning process is continuously repeated, thereby reducing the fine-tuning process at the terminal node, speeding up the model fine-tuning process, improving the efficiency of model fine-tuning, reducing time costs, and lowering system resource overhead. In other words, in the disclosed embodiment, steps 409-411 can be repeatedly executed, i.e., model fine-tuning is iteratively executed, thereby achieving the purpose of improving model fine-tuning efficiency.

[0198] In one possible implementation, the base station node receives a model fine-tuning termination request signaling sent by a terminal node in the model fine-tuning terminal node group, and deletes the terminal node from the model fine-tuning terminal node group based on the model fine-tuning termination request signaling; if the base station node determines that there is no terminal node in the model fine-tuning terminal node group, it stops performing model fine-tuning on the registered pre-trained model.

[0199] In a possible implementation, if the base station node receives a model fine-tuning termination signaling sent by the core network node, it stops performing model fine-tuning on the registered pre-trained model.

[0200] In a possible implementation, the model fine-tuning termination signaling may be transmitted via AMF control plane signaling.

[0201] In a possible implementation, for any terminal node in the model fine-tuning terminal node group, after it terminates model fine-tuning and exits the model fine-tuning terminal node group, the terminal node will deploy the local fine-tuning model at this time to perform a local model inference process.

[0202] It can be seen that in the embodiment of the present disclosure, a federated model fine-tuning method based on federated learning is designed to improve the convergence speed and operation efficiency of model fine-tuning while effectively protecting user privacy. That is, by continuously fine-tuning the model sent by the base station node to all terminal nodes in the terminal node group, and cyclically regulating the local model fine-tuning results uploaded by the terminal node, the fine-tuning efficiency of the model can be accelerated, thereby improving the fine-tuning efficiency of the model, reducing time cost, and reducing system resource overhead.

[0203] In one possible implementation, the terminal node triggers a model fine-tuning event in real time based on the local model inference result, or the base station node periodically triggers the model fine-tuning event, so that the terminal node sends a model fine-tuning request signaling to the base station node and reports model fine-tuning related information. In this way, steps 401 to 411 can be re-executed to continuously fine-tune the model.

[0204] The exemplary embodiment of the present disclosure also provides a model fine-tuning device. Figure 5 As shown, the model fine-tuning device 500 includes the following program modules:

[0205] The first processing unit 501 is configured to receive, by the base station node, a model fine-tuning request sent by at least one terminal node, and send base station-level model fine-tuning request signaling to the core network node;

[0206] The second processing unit 502 is configured to receive, by the base station node, a model fine-tuning strategy signaling and a registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling;

[0207] The model fine-tuning unit 503 is used for the base station node to generate a model fine-tuning configuration signaling according to the model fine-tuning strategy signaling, and send the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, and transmit the first model to all the terminal nodes, so that all the terminal nodes perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, and obtain the corresponding first local fine-tuning model; wherein, the first model is obtained by model processing of the registered pre-trained model.

[0208] In a possible implementation manner, the second processing unit 502 is further configured to:

[0209] The base station node selects a number of terminal nodes from the covered cells and / or cell groups based on the user scheduling policy in the model fine-tuning policy signaling to form the model fine-tuning terminal node group.

[0210] In a possible implementation, the model fine-tuning unit 503 is further configured to:

[0211] The base station node performs federal model fine-tuning on the first local fine-tuning model based on the local model fine-tuning result uploaded by the terminal node in the model fine-tuning terminal node group and according to the federal model fine-tuning algorithm indicated in the model fine-tuning strategy signaling to obtain a second model;

[0212] The local model fine-tuning result is determined based on the first local fine-tuning model.

[0213] In a possible implementation, the model fine-tuning unit 503 is further configured to:

[0214] The base station node sends the second model to all terminal nodes in the model fine-tuning terminal node group, so that all terminal nodes perform local model fine-tuning on the second model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling to obtain the corresponding second local fine-tuning model.

[0215] In a possible implementation, the model fine-tuning unit 503 is further configured to:

[0216] The base station node receives a model fine-tuning termination request signaling sent by a terminal node in the model fine-tuning terminal node group, and deletes the terminal node from the model fine-tuning terminal node group according to the model fine-tuning termination request signaling;

[0217] If the base station node determines that there is no terminal node in the model fine-tuning terminal node group, it stops performing model fine-tuning on the registered pre-trained model.

[0218] In a possible implementation, the model fine-tuning unit 503 is further configured to:

[0219] If the base station node receives the model fine-tuning termination signaling sent by the core network node, it stops performing model fine-tuning on the registered pre-trained model.

[0220] In a possible implementation, the model fine-tuning strategy signaling includes at least two of a model fine-tuning algorithm indication, a synchronization fine-tuning indication, a synchronization fine-tuning time interval indication, a user scheduling strategy indication, a model processing strategy indication, a model deployment mode indication, a privacy protection level indication, a service quality indication, and a federated model fine-tuning algorithm indication;

[0221] Among them, the model fine-tuning algorithm indication is used to instruct the terminal node to use a first preset algorithm to perform local model fine-tuning on the first model; the synchronous fine-tuning indication is used to instruct the base station node to control the terminal node to adopt synchronous fine-tuning or asynchronous fine-tuning to perform model fine-tuning method; the synchronous fine-tuning time interval indication is used to indicate the uniform time interval length set under synchronous fine-tuning; the user scheduling policy indication is used to instruct the base station node to select at least one terminal node to perform local model fine-tuning on the first model; the model processing strategy indication is used to indicate the model processing rules performed by the base station node on the registered pre-trained model; the model deployment method indication is used to indicate the rules for deploying the registered pre-trained model on the terminal node; the privacy protection level indication is used to indicate the protection level of the registered pre-trained model; the service quality indication is used to indicate the service quality performance requirements of the base station node for terminal node model fine-tuning; the federal model fine-tuning algorithm indication is used to indicate the second preset algorithm used by the base station node when aggregating the local model fine-tuning results corresponding to all terminal nodes.

[0222] The exemplary embodiment of the present disclosure also provides a model fine-tuning device. Figure 6 As shown, the model fine-tuning device 600 includes the following program modules:

[0223] A sending unit 601 is configured for the terminal node to send a model fine-tuning request signaling to the base station node, and to report model fine-tuning related information related to the terminal node, so that the base station node sends a base station-level model fine-tuning request signaling to the core network node, and receives a model fine-tuning policy signaling and a registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling;

[0224] A receiving unit 602 is configured to receive, at the terminal node, a model fine-tuning configuration signaling and a first model sent by the base station node; wherein the model fine-tuning configuration signaling is generated according to the model fine-tuning strategy signaling; and the first model is obtained by model processing the registered pre-trained model;

[0225] The model fine-tuning unit 603 is used for the terminal node to collect a local fine-tuning data set, perform local model fine-tuning based on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, obtain a first local fine-tuning model, and report the local model fine-tuning result determined according to the first local fine-tuning model to the base station node.

[0226] In a possible implementation, the terminal node-related model fine-tuning related information includes at least one of the terminal node data type, terminal node model processing capability, terminal node model storage capability, model fine-tuning application scenario, service quality performance requirements, model monitoring related information, and model fine-tuning time limit requirements.

[0227] In a possible implementation manner, the model fine-tuning request signaling is carried via uplink control information, a media access control layer control unit, or a signaling radio bearer.

[0228] In a possible implementation, the local model fine-tuning result is at least one of a local fine-tuning model gradient, a local fine-tuning model loss value, or a local fine-tuning model parameter updated after model fine-tuning.

[0229] In a possible implementation manner, the local model fine-tuning result is reported to the base station node via uplink control information of the control plane or a data radio bearer of the user plane.

[0230] The specific details of each part of the above-mentioned device have been described in detail in the implementation method part. The undisclosed details can be found in the implementation method part, so they will not be repeated here.

[0231] It should be noted that although several modules or units of the device for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to the exemplary embodiments of the present disclosure, the features and functions of two or more modules or units described above can be concretized in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into multiple modules or units to be concretized.

[0232] The exemplary embodiment of the present disclosure also provides a computer program product. The computer program product includes a computer program, which implements the above-mentioned Figure 2-4 The model fine-tuning method shown.

[0233] In one embodiment, a computer program product may be a tangible product containing a computer program, such as a computer-readable storage medium storing the computer program. The computer-readable storage medium may be a storage medium based on electrical, magnetic, optical, electromagnetic, infrared, or other signals, including but not limited to random access memory (RAM), read-only memory (ROM), magnetic tape, floppy disk, flash memory (Flash), mechanical hard disk drive (HDD), solid-state drive (SSD), and the like. Exemplarily, the computer program product may be implemented as a non-volatile storage medium storing the computer program, such as a read-only memory, NAND flash memory, and the like.

[0234] In one embodiment, the computer program product may be an intangible product containing a computer program. For example, the computer program product may be implemented as a virtual digital product, such as a digital file such as an executable file or installation package storing the computer program.

[0235] The code of the computer program can be written in one or more programming languages. Programming languages ​​include C, Java, C++, etc. The program code can be executed entirely on the user computing device, partially on the user computing device, or as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device via any type of network, such as a local area network (LAN), a wide area network (WAN), etc., or can be connected to an external computing device (e.g., via an Internet connection provided by a carrier).

[0236] Computer programs can be carried or transmitted via electrical, magnetic, optical, electromagnetic, infrared, or other signals. Electronic devices can convert signals carrying computer programs into digital signals, thereby running the computer programs. When a computer program is run on an electronic device, its code is used to enable the electronic device to execute (more specifically, enable the processor of the electronic device to execute) the method steps of various exemplary embodiments of the present disclosure, such as the above-mentioned model fine-tuning method, the above-mentioned Figure 2-Figure 4 The model fine-tuning method shown in Figure 2 The illustrated model fine-tuning method includes the following steps: Step 201: The base station node receives a model fine-tuning request sent by at least one terminal node and sends base station-level model fine-tuning request signaling to the core network node. Step 202: The base station node receives model fine-tuning policy signaling and a registered pre-trained model sent by the core network node; the pre-trained model is determined based on the base station-level model fine-tuning request signaling. Step 203: The base station node generates model fine-tuning configuration signaling based on the model fine-tuning policy signaling, and sends the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, transmitting a first model to all terminal nodes so that all terminal nodes perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, thereby obtaining a corresponding first locally fine-tuned model; wherein the first model is obtained by model processing the registered pre-trained model.

[0237] By implementing the above method steps through a computer program, a model fine-tuning request sent by at least one terminal node can be received, and a base station-level model fine-tuning request signaling can be sent to the core network node; and the base station node can receive the model fine-tuning policy signaling and the registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling; in this way, the base station node generates a model fine-tuning configuration signaling according to the model fine-tuning policy signaling, and sends the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, and transmits the first model to all terminal nodes, so that all terminal nodes respectively perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, and obtain a corresponding first local fine-tuning model; wherein the first model is obtained by model processing the registered pre-trained model. It can be seen that in the embodiment of the present disclosure, based on the collaboration of the base station node, the terminal node and the core network node, the fine-tuning of the artificial intelligence model can be quickly achieved, the fine-tuning efficiency of the artificial intelligence model can be improved, the time cost can be reduced, and the system resource overhead can be reduced.

[0238] The exemplary embodiments of the present disclosure further provide an electronic device, which may include a processor and a memory. The memory stores executable instructions for the processor, such as a computer program. The processor executes the executable instructions to perform the method steps of various exemplary embodiments of the present disclosure.

[0239] Reference below Figure 7 , the electronic device is exemplarily described in the form of a general-purpose computing device. It should be understood that Figure 7 The electronic device 700 shown is merely an example and should not limit the functions and scope of use of the embodiments of the present disclosure.

[0240] like Figure 7 As shown, the electronic device 700 may include: a processor 710 , a memory 720 , a bus 730 , an I / O (input / output) interface 740 , and a network adapter 770 .

[0241] The memory 720 may include volatile memory, such as RAM 721 and cache unit 722, and may also include non-volatile memory, such as ROM 723. The memory 720 may also include one or more program modules 724. Such program modules 724 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. For example, the program modules 724 may include the modules in the aforementioned devices.

[0242] The processor 710 may include one or more processing units, for example: the processor 710 may include an AP (Application Processor), a modem processor, a GPU (Graphics Processing Unit), an ISP (Image Signal Processor), a controller, an encoder, a decoder, a DSP (Digital Signal Processor), a baseband processor and / or an NPU (Neural-Network Processing Unit), etc.

[0243] The processor 710 can be used to execute the executable instructions stored in the memory 720, as described above. Figure 2-Figure 4 The model fine-tuning method shown in Figure 2The illustrated model fine-tuning method includes the following steps: Step 201: The base station node receives a model fine-tuning request sent by at least one terminal node and sends base station-level model fine-tuning request signaling to the core network node. Step 202: The base station node receives model fine-tuning policy signaling and a registered pre-trained model sent by the core network node; the pre-trained model is determined based on the base station-level model fine-tuning request signaling. Step 203: The base station node generates model fine-tuning configuration signaling based on the model fine-tuning policy signaling, and sends the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, transmitting a first model to all terminal nodes so that all terminal nodes perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, thereby obtaining a corresponding first locally fine-tuned model; wherein the first model is obtained by model processing the registered pre-trained model.

[0244] By executing the above method steps through the processor 710, a model fine-tuning request sent by at least one terminal node can be received, and a base station-level model fine-tuning request signaling can be sent to the core network node; and the base station node can receive the model fine-tuning policy signaling and the registered pre-trained model sent by the core network node; the pre-trained model is determined according to the base station-level model fine-tuning request signaling; in this way, the base station node generates a model fine-tuning configuration signaling according to the model fine-tuning policy signaling, and sends the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, and transmits the first model to all terminal nodes, so that all terminal nodes perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, and obtain a corresponding first local fine-tuning model; wherein the first model is obtained by model processing the registered pre-trained model. It can be seen that in the embodiment of the present disclosure, based on the collaboration of the base station node, the terminal node and the core network node, the fine-tuning of the artificial intelligence model can be quickly achieved, the fine-tuning efficiency of the artificial intelligence model can be improved, the time cost can be reduced, and the system resource overhead can be reduced.

[0245] The bus 730 is used to connect different components of the electronic device 700 and may include a data bus, an address bus, and a control bus.

[0246] The electronic device 700 can communicate with one or more external devices 800 (eg, a keyboard, a mouse, an external controller, etc.) through the I / O interface 740 .

[0247] The electronic device 700 can communicate with one or more networks via a network adapter 770. For example, the network adapter 770 can provide mobile communication solutions such as 3G / 4G / 7G, or wireless communication solutions such as wireless LAN, Bluetooth, and near-field communication. The network adapter 750 can communicate with other modules of the electronic device 700 via the bus 730.

[0248] although Figure 7 Not shown, other hardware and / or software modules may also be provided in the electronic device 700, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0249] As can be seen from the above, the technical solutions of the present disclosure can be implemented as methods, devices, systems, computer program products, storage media, electronic devices, etc. Those skilled in the art will understand that various aspects of the present disclosure can be specifically implemented in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or an implementation that combines hardware and software aspects, such as "circuit", "module" or "system".

[0250] It should be understood that the present disclosure is not limited to the specific method steps or structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from the scope thereof. Those skilled in the art will easily think of other embodiments based on the specific embodiments provided by the present disclosure. Therefore, the specific embodiments provided by the present disclosure are merely exemplary, and the scope and spirit of the present disclosure are indicated by the claims, which should cover any variations, uses or adaptive changes of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary technical means in the field of the present technology that are not disclosed in the present disclosure.

Claims

1. A model fine-tuning method, characterized in that: Applied to a model fine-tuning system, the model fine-tuning system includes a terminal node, a base station node and a core network node, and the method includes: The base station node receives a model fine-tuning request sent by at least one of the terminal nodes, and sends a base station-level model fine-tuning request signaling to the core network node; The base station node receives the model fine-tuning strategy signaling and the registered pre-trained model sent by the core network node; the pre-trained model is determined according to the model fine-tuning request signaling at the base station level; The base station node generates a model fine-tuning configuration signaling according to the model fine-tuning strategy signaling, and sends the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, and transmits the first model to all the terminal nodes, so that all the terminal nodes perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, and obtain the corresponding first local fine-tuning model; wherein the first model is obtained by model processing the registered pre-trained model.

2. The method according to claim 1, characterized in that The method further comprises: The base station node performs federal model fine-tuning on the first local fine-tuning model based on the local model fine-tuning result uploaded by the terminal node in the model fine-tuning terminal node group and according to the federal model fine-tuning algorithm indicated in the model fine-tuning strategy signaling to obtain a second model; The local model fine-tuning result is determined based on the first local fine-tuning model.

3. The method according to claim 1 or 2, characterized in that: After the base station node receives the model fine-tuning strategy signaling and the registered pre-trained model sent by the core network node, the method further includes: The base station node selects a number of terminal nodes from the covered cells and / or cell groups based on the user scheduling policy in the model fine-tuning policy signaling to form the model fine-tuning terminal node group.

4. The method according to claim 2, characterized in that: The method further comprises: The base station node sends the second model to all terminal nodes in the model fine-tuning terminal node group, so that all terminal nodes perform local model fine-tuning on the second model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling to obtain the corresponding second local fine-tuning model.

5. The method according to claim 4, characterized in that The method further comprises: The base station node receives a model fine-tuning termination request signaling sent by a terminal node in the model fine-tuning terminal node group, and deletes the terminal node from the model fine-tuning terminal node group according to the model fine-tuning termination request signaling; If the base station node determines that there is no terminal node in the model fine-tuning terminal node group, it stops performing model fine-tuning on the registered pre-trained model.

6. The method according to claim 4, characterized in that The method further comprises: If the base station node receives the model fine-tuning termination signaling sent by the core network node, it stops performing model fine-tuning on the registered pre-trained model.

7. The method according to claim 1 or 2, characterized in that: The model fine-tuning strategy signaling includes at least two of a model fine-tuning algorithm indication, a synchronization fine-tuning indication, a time interval indication for synchronization fine-tuning, a user scheduling strategy indication, a model processing strategy indication, a model deployment mode indication, a privacy protection level indication, a service quality indication, and a federated model fine-tuning algorithm indication; Among them, the model fine-tuning algorithm indication is used to indicate the first preset algorithm used by the terminal node to perform local model fine-tuning on the first model; the synchronous fine-tuning indication is used to indicate the base station node to control the terminal node to adopt synchronous fine-tuning or asynchronous fine-tuning to perform the model fine-tuning method; the synchronous fine-tuning time interval indication is used to indicate the uniform time interval length set under synchronous fine-tuning; the user scheduling policy indication is used to indicate the rule for the base station node to select at least one terminal node to perform local model fine-tuning on the first model; the model processing strategy indication is used to indicate the model processing rules performed by the base station node on the registered pre-trained model; the model deployment mode indication is used to indicate the rules for deploying the registered pre-trained model on the terminal node; the privacy protection level indication is used to indicate the protection level of the registered pre-trained model; the service quality indication is used to indicate the service quality performance requirements of the base station node for terminal node model fine-tuning; the federal model fine-tuning algorithm indication is used to indicate the second preset algorithm used by the base station node when aggregating the local model fine-tuning results corresponding to all terminal nodes.

8. A model fine-tuning device, characterized in that: Applied to a model fine-tuning system, the model fine-tuning system includes a terminal node, a base station node and a core network node, and the device includes: A first processing unit, configured for the base station node to receive a model fine-tuning request sent by at least one of the terminal nodes, and to send a base station-level model fine-tuning request signaling to the core network node; A second processing unit is used for the base station node to receive the model fine-tuning strategy signaling and the registered pre-trained model sent by the core network node; the pre-trained model is determined according to the model fine-tuning request signaling at the base station level; A model fine-tuning unit is used for the base station node to generate a model fine-tuning configuration signaling according to the model fine-tuning strategy signaling, and send the model fine-tuning configuration signaling to all terminal nodes in the model fine-tuning terminal node group, and transmit the first model to all the terminal nodes, so that all the terminal nodes perform local model fine-tuning on the first model according to the model fine-tuning configuration information contained in the model fine-tuning configuration signaling, and obtain the corresponding first local fine-tuning model; wherein the first model is obtained by model processing the registered pre-trained model.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: include: processor; A memory, configured to store executable instructions of the processor; The processor is configured to perform the method of any one of claims 1 to 7 by executing the executable instructions.

Images