Online model fine-tuning method and apparatus based on edge intelligence, and computer device

Abstract

The present application relates to an online model fine-tuning method and apparatus based on edge intelligence, and a computer device. The method comprises: receiving an online model fine-tuning request uploaded by at least one terminal node to which an edge node belongs, and on the basis of the at least one terminal node, generating an online model fine-tuning node group; on the basis of a model fine-tuning configuration file list, broadcasting model fine-tuning feature signaling to each terminal node in the online model fine-tuning node group; acquiring subscription request signaling sent by any terminal node; and transmitting a subscribed model parameter configuration file to said terminal node, such that said terminal node performs online model fine-tuning on a pre-trained model on the basis of the subscribed model parameter configuration file, so as to obtain an updated model.

Description

Methods, apparatus, and computer equipment for online model fine-tuning based on edge intelligence

[0001] Related applications

[0002] This application claims priority to Chinese patent application No. 2024119955338, filed on December 31, 2024, entitled "Method, Apparatus and Computer Equipment for Online Model Fine-tuning Based on Edge Intelligence", the entire contents of which are incorporated herein by reference. Technical Field

[0003] This application relates to the field of wireless communication technology, and in particular to an online model fine-tuning method, apparatus, computer device, computer-readable storage medium, and computer program product based on edge intelligence. Background Technology

[0004] With the continuous improvement of user device computing power and the increasing demand for massive data processing, network intelligence is expanding from cloud servers to the network edge, making edge intelligence a key driving force for the application of future sixth-generation mobile communication standards. Edge intelligence combines artificial intelligence models with edge computing technology, relying on an edge computing architecture to distribute data processing, model inference, and some training tasks across network edge nodes and terminal nodes, rather than relying on traditional centralized cloud server computing. By performing intelligent computing at edge nodes close to the data source, edge intelligence reduces the latency of data transmission to the cloud, thereby improving response speed and privacy protection capabilities. Edge intelligence is widely used in smart homes, autonomous driving, industrial automation, and other fields, enabling functions such as real-time monitoring, anomaly detection, and predictive maintenance.

[0005] When implementing edge intelligence in wireless networks, discrepancies arise between the deployment environment and training environment of an AI model. The data distribution learned by the model in the training environment differs from the actual data distribution in the deployment environment, leading to performance degradation. Therefore, model configuration is necessary based on existing pre-trained models—that is, adjusting model parameters using new data to adapt the model to new environments or tasks. However, with the increasing number of user terminals accessing the network edge and the growing demand for AI services, model configuration faces multiple challenges, including limited computing and communication resources and data security concerns, resulting in low efficiency in edge intelligence model configuration. Summary of the Invention

[0006] Firstly, this application provides an online model fine-tuning method based on edge intelligence, applied to the edge nodes of a model fine-tuning system, which further includes terminal nodes and network nodes; including:

[0007] Receive online model fine-tuning requests uploaded by at least one terminal node to which the edge node belongs, and generate an online model fine-tuning point group based on at least one terminal node;

[0008] According to the model fine-tuning configuration file list, model fine-tuning feature signaling is broadcast to each of the terminal nodes in the online model fine-tuning point group; the model fine-tuning configuration file list includes at least one model fine-tuning feature information generated based on the model fine-tuning requirement information of each of the terminal nodes.

[0009] Obtain a subscription request signaling sent by any of the terminal nodes; the subscription request signaling is sent by any of the terminal nodes after determining the model parameter configuration file to be subscribed based on the node running information in the current time period and the model fine-tuning feature information carried by the model fine-tuning feature signaling.

[0010] The subscribed model parameter configuration file is transmitted to any of the terminal nodes, so that any of the terminal nodes can perform online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file to obtain an updated model.

[0011] In one embodiment, after obtaining the updated model, the method further includes:

[0012] Obtain online model monitoring information sent by any of the terminal nodes; the online model monitoring information includes the model verification results of any of the terminal nodes performing online verification of the updated model;

[0013] When the online model monitoring information indicates that the updated model meets the fine-tuning requirements, any of the terminal nodes is removed from the online model fine-tuning point group to obtain the updated online model fine-tuning point group.

[0014] In one embodiment, the model fine-tuning feature information includes at least one of the following: model parameter configuration file identifier, computational complexity, storage complexity, implemented function, task type, expected performance gain, fine-tuning latency, and security level; after obtaining the updated online model fine-tuning point set, the method further includes:

[0015] Based on the online model monitoring information, the subscribed model parameter configuration file is updated to obtain the updated model parameter configuration file, and the model fine-tuning feature information is updated to obtain the updated model fine-tuning feature information.

[0016] In one embodiment, after obtaining the updated model fine-tuning feature information, the method further includes:

[0017] Returning to the step of broadcasting model fine-tuning feature signaling to each terminal node in the online model fine-tuning point group according to the model fine-tuning configuration file list, until the model fine-tuning termination condition is met;

[0018] The model fine-tuning termination condition includes all terminal nodes either moving out of the online model fine-tuning point group or receiving an online model fine-tuning termination signal from the network node.

[0019] In one embodiment, prior to receiving an online model fine-tuning request uploaded by at least one terminal node to which the edge node belongs, the method further includes:

[0020] Based on the model fine-tuning requirement information of each terminal node to which the edge node belongs, generate the model fine-tuning requirement information of the cell group.

[0021] The model fine-tuning requirements of the cell group are uploaded to the network node, so that the network node can generate a corresponding model fine-tuning configuration file list based on the pre-trained model according to the model fine-tuning requirements of the cell group.

[0022] Obtain the list of model fine-tuning configuration files sent by the network node.

[0023] In one embodiment, the fine-tuning requirement conditions include at least one of the following: the online model validation loss value or convergence speed is less than a first preset threshold; the online model validation accuracy is greater than a second preset threshold; the change in the number of consecutive T rounds is less than a third preset threshold; or, the online model validation latency, throughput, or spectral efficiency is less than a fourth preset threshold.

[0024] In one embodiment, the model parameter configuration file includes improvements to at least one aspect of the pre-trained model, including its structure, functionality, number of model parameters, and connection method.

[0025] In one embodiment, the pre-trained model includes a large language model and a neural network model; wherein the large language model includes a converter model, a generative pre-trained converter model, and a decoding-based enhancement model; and the neural network model includes a convolutional neural network, a recurrent neural network, a fully connected network, and a generative adversarial network.

[0026] In one embodiment, the model fine-tuning feature signaling is either attached to the downlink control information or transmitted within the downlink control information field.

[0027] Secondly, this application also provides an online model fine-tuning system based on edge intelligence, the system being used to implement the steps of the above method; the system includes network nodes, edge nodes, and terminal nodes;

[0028] The network node is used to generate a corresponding model fine-tuning configuration file list based on the cell group model fine-tuning requirement information uploaded by the edge node, on the basis of the pre-trained model; the cell group model fine-tuning requirement information is generated based on the model fine-tuning requirement information of each terminal node to which the edge node belongs.

[0029] The edge node is used to obtain the list of model fine-tuning configuration files sent by the network node;

[0030] The terminal node is used to send an online model fine-tuning request to the edge node;

[0031] The edge node is configured to generate an online model fine-tuning point group based on at least one of the terminal nodes, and to broadcast model fine-tuning feature signaling to each of the terminal nodes in the online model fine-tuning point group based on the model fine-tuning configuration file list.

[0032] The terminal node is used to determine the subscribed model parameter configuration file based on the node operation information in the current time period and the model fine-tuning feature information carried by the model fine-tuning feature signaling, and send a subscription request signaling to the edge node.

[0033] The edge node is used to transmit the subscribed model parameter configuration file to the terminal node;

[0034] The terminal node is used to perform online model fine-tuning of the pre-trained model based on the subscribed model parameter configuration file to obtain an updated model.

[0035] Thirdly, this application also provides an online model fine-tuning device based on edge intelligence, applied to the edge nodes of a model fine-tuning system, wherein the model fine-tuning system further includes terminal nodes and network nodes; the device includes:

[0036] The receiving module is used to receive online model fine-tuning requests uploaded by at least one terminal node to which the edge node belongs, and to generate an online model fine-tuning point group based on at least one terminal node;

[0037] The broadcast module is used to broadcast model fine-tuning feature signaling to each of the terminal nodes in the online model fine-tuning point group according to the model fine-tuning configuration file list; the model fine-tuning configuration file list includes at least one model fine-tuning feature information generated according to the model fine-tuning requirement information of each of the terminal nodes.

[0038] The acquisition module is used to acquire a subscription request signaling sent by any of the terminal nodes; the subscription request signaling is sent by any of the terminal nodes after determining the model parameter configuration file to be subscribed based on the node running information in the current time period and the model fine-tuning feature information carried by the model fine-tuning feature signaling.

[0039] The transmission module is used to transmit the subscribed model parameter configuration file to any of the terminal nodes, so that any of the terminal nodes can perform online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file to obtain an updated model.

[0040] Fourthly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0041] Fifthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0042] Sixthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0043] Details of one or more embodiments of this application are set forth in the following drawings and description. Other features, objects, and advantages of this application will become apparent from the specification, drawings, and claims. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 shows the application environment of an online model fine-tuning method based on edge intelligence in one embodiment;

[0046] Figure 2 is a flowchart illustrating an online model fine-tuning method based on edge intelligence in one embodiment;

[0047] Figure 3 is a signaling interaction logic diagram of an online model fine-tuning method based on edge intelligence in one embodiment;

[0048] Figure 4 is a flowchart illustrating an online model fine-tuning method based on edge intelligence in another embodiment;

[0049] Figure 5 is a structural block diagram of an online model fine-tuning device based on edge intelligence in one embodiment;

[0050] Figure 6 is an internal structure diagram of a computer device in one embodiment. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] The edge intelligence-based online model fine-tuning method provided in this application embodiment can be applied to the application environment shown in Figure 1. The edge intelligence-based online model fine-tuning system includes a network node 100, an edge node 102, and a terminal node 104.

[0053] Network node 100 can manage and generate model parameter configuration files, maintain model versions, and configure signaling. It also centrally stores various pre-trained models and model parameter configuration files for edge nodes 102 and terminal nodes 104 to access. Network node 100 optimizes resource allocation and task scheduling by monitoring the overall network operation status. For example, network node 100 possesses orchestration and management (OAM) functions, network data analytics (NWDAF) functions, over-the-top (OTT) server functions, and network functions for network management.

[0054] Edge node 102 serves as an intermediate layer connecting user equipment (terminal node 104) and the core network (network node 100). It shares the computing tasks of the core network and provides low-latency services to the terminal nodes. Edge node 102 can be a node deployed close to the data source or user equipment, capable of performing data processing, storage, and some artificial intelligence computing tasks. For example, edge node 102 may include base stations, access network units, edge servers, industrial edge gateways, or edge computing devices.

[0055] In one embodiment, edge node 102 may include a Centralized Unit (CU) and a Distributed Unit (DU). Specifically, the Centralized Unit is an important component of the radio access network, primarily responsible for higher-level protocol processing and resource management, and can be used for model transmission, signaling configuration, model updates, etc. The Centralized Unit can be further divided into two parts: a Control Plane Centralized Unit (CU-CP) and a User Plane Centralized Unit (CU-UP). The Control Plane Centralized Unit (CU-CP) is responsible for processing control signaling, such as user equipment access management, handover management, and connection maintenance, and centrally manages and coordinates multiple Distributed Units (DUs). The User Plane Centralized Unit (CU-UP) is responsible for forwarding and processing user data, including packet data transmission, providing data plane processing, and assisting the Distributed Units (DUs) in completing efficient data transmission tasks. Specifically, the Distributed Unit is mainly responsible for low-level physical layer and real-time task processing, executing specific tasks assigned by the Centralized Unit, and can be used for data collection, model transmission, etc.

[0056] In this system, terminal node 104 is a user equipment (UE). Terminal node 104 can be used for data collection, such as collecting environmental and task-related data (e.g., sensor data, user behavior data, etc.). Terminal node 104 can also be used for model subscription and local model fine-tuning, that is, adjusting the parameters of the pre-trained model locally according to the received model parameter configuration file to generate a model adapted to the current environment or task. Terminal node 104 can also perform performance monitoring, monitoring the model's running status in real time, including accuracy, latency, and stability, and feeding back the monitoring information to edge node 102.

[0057] As can be seen, terminal node 104 may include a model fine-tuning function module, a model monitoring function module, a data acquisition function module, etc.; edge node 102 may include a model fine-tuning function module, a model management function module, a model storage function module, etc.; network node 100 may include a model management module, a model storage module, etc.

[0058] In the specific implementation, network node 100 is used to generate a corresponding model fine-tuning configuration file list based on the cell group model fine-tuning requirement information uploaded by edge node 102 and the pre-trained model; the cell group model fine-tuning requirement information is generated based on the model fine-tuning requirement information of each terminal node 104 to which edge node 102 belongs.

[0059] Edge node 102 is used to obtain the list of model fine-tuning configuration files sent by network node 100;

[0060] Terminal node 104 is used to send online model fine-tuning requests to edge node 102;

[0061] Edge node 102 is used to generate an online model fine-tuning point group based on at least one terminal node 104, and to broadcast model fine-tuning feature signaling to each terminal node 104 in the online model fine-tuning point group based on the model fine-tuning configuration file list.

[0062] Terminal node 104 is used to determine the subscribed model parameter configuration file based on the node running information in the current time period and the model fine-tuning feature information carried by the model fine-tuning feature signaling, and send a subscription request signaling to edge node 102.

[0063] Edge node 102 is used to transmit the subscribed model parameter configuration file to terminal node 104;

[0064] Terminal node 104 is used to perform online model fine-tuning of the pre-trained model based on the subscribed model parameter configuration file to obtain the updated model.

[0065] In an exemplary embodiment, as shown in Figure 2, an online model fine-tuning method based on edge intelligence is provided. Taking the application of this method to edge node 102 in Figure 1 as an example, the method includes:

[0066] Step S202: Receive online model fine-tuning requests uploaded by at least one terminal node to which the edge node belongs, and generate an online model fine-tuning point group based on at least one terminal node.

[0067] The online model fine-tuning request can be a request sent by the terminal node to the edge node during operation, expressing the terminal node's current need for fine-tuning the artificial intelligence model. The online model fine-tuning request may include the terminal node's current resource requirements (such as computing power, bandwidth, etc.), real-time requirements, and the type of artificial intelligence task to be processed (such as natural language processing, image classification, object detection, etc.). In specific implementation, the terminal node can report the online model fine-tuning request to the edge node at any time within each time period, and participate in the online model fine-tuning process of the current or next time period after the edge node responds to the request.

[0068] Optionally, online model fine-tuning requests can be sent in the form of Downlink Control Information (DCI), Media Access Control (MAC) Control Element (CE), or Radio Resource Control (RRC) signaling. DCI can be signaling used between the user equipment and the base station to transmit control information; MAC Control Element (CE) can be control signaling used between the user equipment and the base station for status reporting and command transmission; and RRC signaling can be signaling located in the Radio Resource Control layer responsible for configuring and managing the radio link. Through these signaling messages, online model fine-tuning requests can be embedded into different layers of the protocol stack to ensure efficient transmission in different scenarios.

[0069] Optionally, the triggering method for the online model fine-tuning request includes at least one of the following: the terminal node determines whether online model fine-tuning needs to be performed based on model monitoring information, and triggers an online model fine-tuning event, sending an online model fine-tuning request to the edge node; the network node or edge node periodically sends model fine-tuning instructions to the terminal node so that the terminal node generates an online model fine-tuning request.

[0070] In practice, edge nodes can record the terminal nodes that send online model fine-tuning requests within the current time period, identify these terminal nodes as all the terminal nodes participating in online model fine-tuning within the current time period, and form an online model fine-tuning point group based on these terminal nodes.

[0071] In practice, the online model fine-tuning point group can include terminal nodes newly added to the online model fine-tuning process within the current time period, as well as terminal nodes that have joined the online model fine-tuning process in previous time periods and have not exited. The online model fine-tuning point group can record the information of each type of terminal node separately, or it can record the information of all terminal nodes jointly without distinguishing between the two types.

[0072] Optionally, the online model fine-tuning point group should at least record the terminal node identifier (ID), the cell / cell group to which the terminal node belongs, the timestamp of the terminal node joining the online model fine-tuning, and whether it participates in joint fine-tuning for each terminal node. Here, the cell to which the terminal node belongs can refer to an area in the wireless network loaded by a single base station, and the cell group can refer to a set of multiple adjacent or functionally related cells; whether it participates in joint fine-tuning can indicate that multiple terminal nodes can collaboratively complete the model fine-tuning task and share model update information or fine-tuning results.

[0073] In practice, the online model fine-tuning point group allows terminal nodes to join and leave the group at any time, and updates the information of the online model fine-tuning point group after each terminal node joins or leaves the group.

[0074] Step S204: Based on the model fine-tuning configuration file list, broadcast the model fine-tuning feature signaling to each terminal node in the online model fine-tuning point group.

[0075] In practice, edge nodes can select model fine-tuning feature information from the model fine-tuning configuration file list, generate model fine-tuning feature signaling based on this information, and broadcast this signaling to each terminal node in the online model fine-tuning point group. This signaling carries the model fine-tuning feature information. The model fine-tuning configuration file list can include multiple model parameter configuration files and their corresponding model fine-tuning feature information. Specifically, the model fine-tuning configuration file list can be generated by network nodes based on the model fine-tuning requirements of terminal nodes, using a pre-trained model, and then distributed to the edge nodes.

[0076] The model parameter configuration file can include parameter configuration information for the pre-trained model, which can be used to guide online model fine-tuning. Each model parameter configuration file can correspond to unique model fine-tuning feature information.

[0077] Among them, model fine-tuning feature information can be used to describe the feature information in the model parameter configuration file related to performing model fine-tuning.

[0078] The model fine-tuning configuration file list includes at least one model fine-tuning feature information generated based on the model fine-tuning requirements of each terminal node. Optionally, the model fine-tuning requirements information may include computing power configuration information relevant to the corresponding terminal node that can be used for model fine-tuning, storage space capacity information that can be used to store the model, service quality performance requirements for model fine-tuning, latency tolerance, application scenarios, data types or data formats, and other historical data related to model fine-tuning.

[0079] In one embodiment, model fine-tuning feature information can be stored in the model fine-tuning configuration file list of the edge node according to the storage method of hash table. The file identifier ID of each model parameter configuration file is used as the key. A mapping function is designed to map the key to the storage address (value) of the corresponding model fine-tuning feature information. Thus, the storage address of the unique model fine-tuning feature information is determined according to each file identifier ID, and a one-to-one correspondence between multiple key-value pairs is obtained to ensure efficient storage and retrieval.

[0080] In one embodiment, the model fine-tuning feature information includes at least one of the following: the identifier of the model parameter configuration file, computational complexity, storage complexity, implemented functions, task type, expected performance gain, fine-tuning latency, and security level.

[0081] The model parameter configuration file is identified by a unique ID to distinguish it from other configuration files. The computational complexity represents the computational resources required for this configuration file, including the number of CPU / GPU processors, cycles, and floating-point operations needed for model fine-tuning. The storage complexity represents the storage resources required, including disk space usage, storage format, and whether compression is supported. The functionality characterizes the input and output types of the configuration file, such as the data type of the input data, the data format of the output data, and whether multimodal data input is used. The task type indicates the type of task the configuration file can perform, including natural language processing, image classification, and sentiment analysis. The model parameter configuration file is used for various applications, including class analysis, time series prediction, and target detection. The expected performance gain characterizes the anticipated performance gains of the model parameter configuration file in terms of Quality of Service (QoS), energy efficiency, and system performance, and is related to the self-evaluation results of network nodes. Fine-tuning latency characterizes the processing time of the model parameter configuration file when performing fine-tuning tasks, and is generally related to the type of task, such as the time required to process one data sample or a batch of data samples. The security level indicates the confidentiality level of the model parameter configuration file. If its model structure or design method exhibits significant autonomy and innovation, and has a high confidentiality level, its distribution to end nodes will be restricted to or only distributed to security-compliant end nodes to protect data privacy.

[0082] For example, the model fine-tuning feature information can be shown in Table 1.

[0083] Table 1

[0084] In practice, edge nodes determine the corresponding model fine-tuning application scenario based on the online model fine-tuning requests sent by terminal nodes. For different application scenarios, the edge nodes select different model fine-tuning feature information, thus generating different model fine-tuning feature signaling. Model fine-tuning application scenarios can be specific usage occasions or task requirements of artificial intelligence models in different practical applications. Since different scenarios have different requirements for model functionality, performance, and resource consumption, edge nodes need to select the most suitable model fine-tuning feature information based on these scenario characteristics. For example, terminal nodes may need different types of model fine-tuning algorithms for different application tasks; for instance, speech recognition, user behavior prediction, object detection, path planning, device anomaly detection, and predictive maintenance all correspond to different types of model fine-tuning algorithms. Furthermore, the computing power, storage space, and bandwidth resources of terminal nodes limit the complexity of the model and the fine-tuning method; for example, high-performance terminal nodes can run complex fine-tuning algorithms, while low-performance terminal nodes can run lightweight models. Therefore, for these different model fine-tuning application scenarios, edge nodes can select different model fine-tuning feature information, thereby broadcasting different model fine-tuning feature signaling.

[0085] In one embodiment, the edge node generates model fine-tuning feature signaling based on model fine-tuning feature information. This model fine-tuning feature signaling may include model fine-tuning feature information and may also implicitly or explicitly include a list of model fine-tuning configuration files. Implicitly, the model fine-tuning feature information already includes the identifier ID of the model parameter configuration file, eliminating the need for the edge node to transmit it separately; explicitly, the edge node transmits the list of model fine-tuning configuration files to the terminal node only once.

[0086] In specific implementation, edge nodes broadcast model fine-tuning feature signaling to each terminal node in the online model fine-tuning point group. Optionally, the model fine-tuning feature signaling can be transmitted by the control plane centralized unit in the edge node via downlink control information (DCI) or system information broadcast (SIB), or it can be sent by the user plane centralized unit in the edge node via data radio bearer (DRB). The model fine-tuning feature signaling can be transmitted after the downlink control information (DCI) or it can be transmitted within the downlink control information (DCI) field.

[0087] Step S206: Obtain the subscription request signaling sent by any terminal node.

[0088] The subscription request signaling is sent by any terminal node after determining the model parameter configuration file to subscribe to, based on the node operation information in the current time period and the model fine-tuning feature information carried in the model fine-tuning feature signaling. The subscription request signaling is a type of signaling sent by a terminal node to an edge node to request a specific model parameter configuration file.

[0089] The node operation information includes at least one of performance information, environmental information, and resource information. The terminal node can determine the model parameter configuration file to subscribe to based on the matching between the node operation information for the current time period and the model fine-tuning feature information.

[0090] The model parameter configuration file contains the parameter configuration information for the pre-trained model. It can include configurations optimized for specific tasks or terminal node requirements. The model parameter configuration file can define parameter adjustment ranges, optimization objectives, training strategies, etc.

[0091] In one embodiment, the model parameter configuration file includes improvements to the structure, function, number of model parameters, connection methods, etc. of the pre-trained model, which can change some modules of the original backbone model or add new modules on it.

[0092] In one embodiment, the pre-trained model can be a large language model or a smaller general neural network model, such as a convolutional neural network (CNN), a recurrent neural network (RNN), a fully connected network (MLP), a generative adversarial network (GAN), etc.

[0093] In practice, the terminal node selects one or more model parameter configuration files based on performance, environmental, and resource information for the current time period. Optionally, the granularity of the time period can be a time slot, half-frame, frame, or a longer time interval, which is uniformly configured by the edge node and communicated to the terminal node.

[0094] Optionally, performance information can refer to the device's own performance requirements, which include requirements for performance indicators related to model fine-tuning such as accuracy, convergence speed, energy consumption, latency, throughput, spectral efficiency, and quality of service (QoS); environmental information can refer to environmental conditions, which include wireless channel conditions, data arrival conditions, and service processing conditions; and resource information can refer to available resources, which may include computing resources, communication resources, and storage resources available to the local device.

[0095] In one embodiment, the method by which a terminal node selects a model parameter configuration file to subscribe to may include: the terminal node determining, based on performance information, environmental information, and resource information in the current time period, the application scenario requiring model fine-tuning, performance requirement thresholds, etc.; according to the model fine-tuning configuration file list, selecting the model fine-tuning feature information corresponding to each model parameter configuration file that matches its own application scenario, comparing it with its own performance requirement threshold, obtaining a score for each model parameter configuration file, and not scoring model parameter configuration files that do not match its own application scenario; selecting the N configuration files with the highest scores from all model parameter configuration files as the model parameter configuration files to be subscribed to by this terminal node, forming a model parameter configuration file subscription list. Optionally, the value of N is configured by the terminal node itself, or it can be configured by the edge node through downlink control information or the media access control layer control unit.

[0096] In one embodiment, the model parameter configuration file subscription list can correspond to the model fine-tuning configuration file list issued by the edge node, and be arranged in the same order to maintain alignment at the edge node and the terminal node. A 1-bit flag can be added after the file identifier ID in the model fine-tuning configuration file list to obtain the model parameter configuration file subscription list. The flag position of the subscribed model parameter configuration file can be set to 1 in the model parameter configuration file subscription list, and the flag positions of the other modules can be set to 0.

[0097] In one embodiment, the subscription request signaling may include a model parameter configuration file subscription list related to the terminal node, as well as information such as timestamp, subscription duration, subscription method, and supported storage formats.

[0098] Optionally, subscription request signaling can be carried by control plane uplink control information (UCI) signaling or user plane data radio bearer (DRB).

[0099] Step S208: The subscribed model parameter configuration file is transmitted to any terminal node so that the terminal node can perform online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file to obtain the updated model.

[0100] In practice, edge nodes select the model parameter configuration file they want to subscribe to and transmit it to the corresponding terminal node based on the subscription request signaling reported by the terminal node. The model parameter configuration file may include parameter configuration information of the pre-trained model, which can be used to guide how to fine-tune and configure the pre-trained model.

[0101] Optionally, edge nodes can transmit subscriptions based on model parameter profiles and use downlink control information (DCI) to indicate allocated time-frequency domain transmission resources.

[0102] In specific implementations, the method for transmitting the subscribed model parameter configuration file may include: the edge node transmitting the model file of the model parameter configuration file from the centralized unit (CU) to which the terminal node belongs via the F1 interface to the distributed unit (DU), and then transmitting it to the terminal node via the Uu interface, where the F1 interface is the communication interface connecting the distributed unit (DU) and the centralized unit (CU), and the Uu interface is the wireless interface between the terminal device and the base station; or, the edge node transmitting the model file of the model parameter configuration file from a centralized unit (CU) to which the terminal node currently belongs via the Xn interface, then transmitting the model file from the centralized unit (CU) to the distributed unit (DU) via the F1 interface, and then transmitting it to the terminal node via the Uu interface, where Xn is the interface used to connect different base stations; or, the edge node transmitting the model file of the model parameter configuration file from the distributed unit (DU) to the terminal node via the Uu interface.

[0103] In one embodiment, edge nodes can transmit customized fine-tuning modes based on the storage format of the model parameter configuration file. The model parameter configuration file can be stored in an extensible format such as .h5, .pb, or .onnx, or in a serialized storage format such as .pt, .json, or .pkl. Different terminal nodes may support different deep learning frameworks, thus supporting different storage formats. Therefore, edge nodes may store the same model parameter configuration file in different storage formats and send model files with customized fine-tuning modes in different storage formats based on different terminal nodes.

[0104] In one embodiment, edge nodes can participate in the module subscription decision-making process. For example, based on factors such as the global quality of service (QoS) requirements of the cell / cell group, global energy consumption requirements, computing power scheduling capabilities, model parameter configuration file utilization, stability requirements, and generalization requirements, the edge nodes can adjust the module subscription results of the terminal nodes and ensure that the module subscription requests of the terminal nodes are met to the greatest extent possible.

[0105] In practice, the terminal node performs model fine-tuning using the online fine-tuning dataset based on the subscribed model parameter configuration file transmitted by the edge node, obtaining the updated model. The model fine-tuning algorithm can employ gradient descent, determining the update direction of the model by using the gradient values ​​of the model parameters in the configuration file on each data sample or batch of data samples, and updating the model parameters of the pre-trained model by setting the update step size.

[0106] In the aforementioned online model fine-tuning method based on edge intelligence, the following steps are taken: receiving an online model fine-tuning request uploaded by at least one terminal node to which the edge node belongs, and generating an online model fine-tuning adjustment point group based on at least one terminal node; broadcasting model fine-tuning feature signaling to each terminal node in the online model fine-tuning adjustment point group according to a model fine-tuning configuration file list, wherein the model fine-tuning configuration file list includes at least one model fine-tuning feature information generated based on the model fine-tuning requirement information of each terminal node; obtaining a subscription request signaling sent by any terminal node; wherein the subscription request signaling is sent by any terminal node after determining the subscribed model parameter configuration file based on the node running information in the current time period and the model fine-tuning feature information carried in the model fine-tuning feature signaling; and transmitting the subscribed model parameter configuration file to any terminal node so that any terminal node can perform online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file to obtain an updated model. By receiving online model fine-tuning requests from terminal nodes, an online model fine-tuning point group is dynamically generated. Model fine-tuning feature signaling is broadcast according to the model fine-tuning configuration file list. Through the signaling interaction process between edge nodes and terminal nodes, it ensures that model configuration can be efficiently optimized according to actual needs. Terminal nodes can autonomously and flexibly select suitable model parameter configuration files based on their own performance, environment, and resource conditions, and perform online model fine-tuning of the pre-trained model. This achieves flexible adaptation to time-varying environments, enhances the generalization of artificial intelligence models at the network edge, and provides customized artificial intelligence services to edge users based on general pre-trained models. This reduces data transmission volume and communication latency while improving model configuration efficiency and real-time performance.

[0107] In another embodiment, after obtaining the updated model, the method further includes: acquiring online model monitoring information sent by any terminal node; the online model monitoring information includes the model verification results of any terminal node performing online verification of the updated model; and if the online model monitoring information indicates that the updated model meets the fine-tuning requirements, removing any terminal node from the online model fine-tuning point group to obtain the updated online model fine-tuning point group.

[0108] In the implementation, terminal nodes collect online data in real time and divide the collected data into online fine-tuning datasets and online validation datasets. The division ratio is determined by the terminal nodes themselves or configured by the edge nodes. Based on the subscribed model parameter configuration files transmitted from the edge nodes, the terminal nodes perform model fine-tuning using the online fine-tuning dataset to obtain the updated model. The model fine-tuning algorithm can employ gradient descent. The direction of model update is determined by the gradient values ​​of the model parameters in the model parameter configuration file on each data sample or batch of data samples. The update step size is set to update the model parameters of the pre-trained model. The terminal nodes perform online validation of the updated model. Based on the updated model parameters and the validation results obtained for each data sample in the online validation dataset, the validation results are compared with the label values ​​to obtain the online validation performance on that data sample. Finally, the average value is calculated over all online validation data samples to obtain the online validation performance based on the online validation dataset, thus obtaining the model validation result.

[0109] In one embodiment, if the terminal node subscribes to multiple model parameter configuration files, it uses the multiple updated model parameter configuration files to calculate the average of the validation results on the online validation dataset or uses a specific model ensemble algorithm to obtain the online validation performance and obtain the model validation result. The model ensemble algorithm is determined autonomously by the terminal node.

[0110] In practice, the terminal node executes an online model monitoring mechanism to monitor online validation performance in real time and obtain model validation results. These results may include at least one of the following: loss value, gradient value, convergence speed, accuracy, latency, throughput, and spectral efficiency. They may also include feedback on model usage, descriptions of model monitoring metrics, and online model fine-tuning logs. Based on these validation results, the terminal node can generate online model monitoring information and upload it to the edge nodes.

[0111] The online model monitoring information may include the model verification results mentioned above, namely the feedback information generated after the terminal node verifies the updated model, including model performance, adaptability, and task completion status, which can be used to evaluate the effectiveness of the updated model and model parameter configuration file.

[0112] Optionally, terminal nodes can report online model monitoring information to edge nodes via uplink control information (UCI) signaling or user plane data radio bearer (DRB). The reporting period can be set by the edge node or terminal node, but it must be ensured that the reporting is completed within the current time period.

[0113] In one embodiment, the online model monitoring information corresponds to the model parameter configuration file. If the terminal node subscribes to multiple model parameter configuration files, it reports the online model monitoring information corresponding to multiple modules respectively. For example, a list is formed according to the order of the identifier IDs of the model parameter configuration files, i.e., [ID of module 1: online model monitoring information of module 1, ..., ID of module M: ​​online model monitoring information of module M].

[0114] In practice, since online model monitoring information can reflect the model fine-tuning process and model configuration results, if the online model monitoring information indicates that the updated model meets the fine-tuning requirements, it can be said that the model configuration has been successful and meets the requirements of the fine-tuning task, thereby triggering the online model fine-tuning exit process of the terminal node.

[0115] Optionally, the fine-tuning requirements may include at least one of the following conditions: the online model validation loss value or convergence speed is less than a first preset threshold; the online model validation accuracy is greater than a second preset threshold, or the change over consecutive T rounds is less than a third preset threshold; the online model validation latency, throughput, or spectral efficiency is less than a fourth preset threshold. The first, second, third, and fourth preset thresholds, as well as the T value, can be configured by the edge node system or by the terminal node itself.

[0116] In one embodiment, the terminal node triggers an online model fine-tuning exit process by sending an online model fine-tuning exit request to the edge node and exiting the online model fine-tuning adjustment point group. The edge node responds to the online model fine-tuning exit request by ceasing to send model parameter configuration files and model fine-tuning feature information to the terminal node in the next time period and updating the online model fine-tuning adjustment point group.

[0117] In one embodiment, the online model fine-tuning exit process for any terminal node can be triggered by an edge node based on the online model monitoring information reported by the terminal node, and an online model fine-tuning exit signaling message can be sent to the corresponding terminal node. Upon confirmation from the terminal node, the edge node removes the terminal node from the online model fine-tuning adjustment point group, ceases sending model parameter configuration file model fine-tuning feature information to that terminal node in the next time period, and updates the online model fine-tuning adjustment point group.

[0118] The technical solution of this embodiment, by introducing online model monitoring information, can verify in real time whether the updated model in the terminal node meets the fine-tuning requirements, realizing dynamic performance evaluation during the fine-tuning process and enhancing adaptability to complex scenarios and task requirements. When a terminal node verifies that its updated model meets the fine-tuning requirements, it is removed from the online model fine-tuning point group, which can reduce the resource allocation for the already completed fine-tuning points and allocate more resources to other nodes that need fine-tuning. Dynamically removing nodes that have completed tasks avoids redundancy in system resource allocation. Timely removal of the node group after the terminal node completes fine-tuning and meets the requirements can accelerate the iteration efficiency of the fine-tuning task. The dynamic addition and removal mechanism of terminal nodes adapts to the dynamic changes in the states of network nodes and terminal nodes in the edge intelligent system and enhances real-time response capabilities.

[0119] In another embodiment, after the updated online model fine-tuning point group, the method further includes: updating the subscribed model parameter configuration file based on the online model monitoring information to obtain an updated model parameter configuration file, and updating the model fine-tuning feature information to obtain updated model fine-tuning feature information.

[0120] The model fine-tuning feature information includes at least one of the following: model parameter configuration file identifier, computational complexity, storage complexity, implemented functions, task type, expected performance gain, fine-tuning latency, and security level.

[0121] Edge nodes can improve existing model parameter configuration files based on the shortcomings of the model reported in the monitoring information. For example, updating the subscribed model parameter configuration file may include adjusting the range of model parameters to better suit specific scenarios or task requirements.

[0122] In practice, edge nodes can use the gradient values ​​from the online model monitoring information to perform a weighted summation of the online model validation gradient values ​​reported by terminal nodes in all online model fine-tuning point groups, and then use gradient descent to update the model parameters in the subscribed model parameter configuration file. Alternatively, edge nodes can combine the online model validation latency information reported by terminal nodes in all online model fine-tuning point groups with the subscribed model parameter configuration file, and increase or decrease the number of model parameters in the subscribed model parameter configuration file according to a certain proportion, thereby reducing the transmission latency of online model validation.

[0123] Accordingly, if the model parameter configuration file is updated, the corresponding model fine-tuning feature information is also updated.

[0124] The technical solution of this embodiment dynamically updates the model parameter configuration file and its feature information by acquiring online monitoring information in real time. This enables rapid optimization of module performance based on the latest feedback from terminal nodes, ensuring that the model always adapts to actual application scenarios. By updating the subscribed fine-tuning modules, unnecessary repeated debugging and performance degradation are avoided, ensuring that the model optimization process is more accurate and efficient. By updating feature information, edge nodes can more accurately allocate fine-tuning modules to terminal nodes that meet their needs, enabling the edge intelligent system to quickly respond to changes in terminal node requirements and adapt to complex scenarios involving multiple tasks and environments.

[0125] In another embodiment, after obtaining the updated model fine-tuning feature information, the method further includes: returning to the step of broadcasting model fine-tuning feature signaling to each terminal node in the online model fine-tuning adjustment point group according to the model fine-tuning configuration file list, until the model fine-tuning termination condition is met; wherein, the model fine-tuning termination condition includes each terminal node being removed from the online model fine-tuning adjustment point group or receiving an online model fine-tuning termination signaling issued by the network node.

[0126] Optionally, the online model fine-tuning termination signaling issued by the network node can be forwarded to the edge node through a dedicated control plane interface. The edge node can then issue the online model fine-tuning termination signaling to all terminal nodes, using downlink control information, media access control layer control elements, or radio resource control signaling to transmit the signaling.

[0127] In practice, the fine-tuning process iterates continuously until the model fine-tuning termination condition is met. Once a terminal node confirms through online model verification that its model meets the fine-tuning requirements, it will be removed from the fine-tuning point group one by one. Once the online model fine-tuning point group is empty, it means that the fine-tuning tasks of all terminal nodes have been completed, and the process terminates. Alternatively, the network node can proactively issue a termination signal based on global needs (such as system resource reallocation, task priority adjustment, etc.) to forcibly end the fine-tuning process.

[0128] During the iterative process of fine-tuning, after obtaining the updated model fine-tuning configuration file list, the process returns to the step of broadcasting model fine-tuning feature signaling to each terminal node in the online model fine-tuning point group based on the model fine-tuning configuration file list. This allows for dynamic updates of the model fine-tuning feature signaling, ensuring that each terminal node obtains the model parameter configuration file most suitable for its current state in each iteration. Through multiple iterations, model performance is gradually optimized to ultimately meet the task requirements.

[0129] The technical solution of this embodiment optimizes the model fine-tuning process through dynamic iteration, ensuring that the terminal node can continuously receive the most suitable model parameter configuration file and gradually meet the task requirements; when the fine-tuning termination condition is met, the process ends naturally, which can efficiently complete the model fine-tuning task in dynamic environments and multi-task scenarios, while optimizing resource allocation.

[0130] In another embodiment, before receiving an online model fine-tuning request uploaded by at least one terminal node, the method further includes: generating model fine-tuning requirement information for a cell group based on the model fine-tuning requirement information of each terminal node to which the edge node belongs; uploading the model fine-tuning requirement information of the cell group to the network node, so that the network node generates a corresponding model fine-tuning configuration file list based on the pre-trained model according to the model fine-tuning requirement information of the cell group; and obtaining the model fine-tuning configuration file list sent by the network node.

[0131] The model fine-tuning requirements information may include computing power configuration information for the corresponding terminal node that can be used for model fine-tuning, storage space capacity information that can be used to store the model, service quality performance requirements for model fine-tuning, latency tolerance, application scenarios, data types or data formats, and other historical data related to model fine-tuning.

[0132] The model fine-tuning requirements for cell groups can include model fine-tuning requirements for individual cells and for the entire cell group. A cell can refer to an area in a wireless network that is loaded by a single base station, while a cell group can refer to a collection of multiple adjacent or functionally related cells. In addition to the model fine-tuning requirements for each terminal node, the model fine-tuning requirements for cell groups also include requirements for the number of models deployed within the cell / cell group, stability requirements, and generalization requirements.

[0133] In one embodiment, the model parameter configuration file includes improvements to the structure, functionality, number of model parameters, and connection methods of the pre-trained model. This can modify some modules of the original backbone model or add new modules. Optionally, the model parameter configuration file can be implemented using techniques such as Low-Rank Adaptive (LoRA), adapter, prefix tuning, and prompt tuning.

[0134] In one embodiment, the pre-trained model can be a large language model, such as a Transformer model, a generative pre-trained Transformer model, or a decoder-enhanced BERT model, or a smaller general neural network model, such as a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), a Fully Connected Network (MLP), or a Generative Adversarial Network (GAN).

[0135] In practical implementation, network nodes can generate corresponding model parameter configuration files based on the model fine-tuning requirements of cell groups, and generate corresponding model fine-tuning feature information based on the model fine-tuning-related feature information in the model parameter configuration files. Finally, a list of model fine-tuning configuration files is generated based on the model parameter configuration files and the model fine-tuning feature information. For example, the way network nodes generate corresponding model parameter configuration files based on the pre-trained model can be through low-rank adaptive (LoRA) technology. Specifically, the weight matrix of the pre-trained model can be decomposed into two matrices with fewer parameters, and the number of parameters in the model parameter configuration files can be controlled by adjusting the rank of the matrices, forming a series of model parameter configuration files to match the performance requirements of different terminal nodes.

[0136] Optionally, network nodes can transmit model parameter configuration files to edge nodes via user plane data radio bearer (DRB) and transmit corresponding configuration information via dedicated control signaling.

[0137] In one embodiment, the edge node sorts the model parameter configuration files based on the model parameter configuration files transmitted by the network node according to at least one of the following: module parameter size, model structure complexity, customized design method used, stability, generalization, etc., and records the identifier ID information of each module in order to generate a model fine-tuning configuration file list.

[0138] The technical solution of this embodiment, before the terminal node issues a fine-tuning request, involves edge nodes and network nodes collaborating to generate a fine-tuning module, reducing response latency; aggregating the demand information of the block group avoids terminal nodes transmitting their own demands, saving communication resources; edge nodes are only responsible for distribution, while network nodes centrally complete complex model optimization tasks, improving the overall utilization efficiency of system resources; model parameter configuration files are generated according to the block group requirements, capable of adapting to multiple terminal nodes simultaneously, reducing redundant calculations; and by dynamically maintaining the list of model fine-tuning configuration files, it supports different changes in the needs of terminal nodes, adapting to complex multi-task scenarios. In summary, through the collaboration of edge nodes and network nodes, a list of model fine-tuning configuration files adapted to the block group requirements is generated in advance, enabling a rapid response to subsequent online model fine-tuning requests. This optimizes resource allocation and improves the efficiency and adaptability of model fine-tuning, making it suitable for handling complex scenarios involving multiple terminals and multiple tasks.

[0139] To facilitate understanding by those skilled in the art, Figure 3 provides an exemplary signaling interaction logic diagram of an online model fine-tuning method based on edge intelligence.

[0140] The edge intelligence-based online model fine-tuning method of this application is applicable to edge intelligence of communication standards such as 5G and 6G. Based on a general pre-trained model, it provides customized artificial intelligence services to edge users. Within each time period, the edge device autonomously decides to deploy and execute the model fine-tuning algorithm according to its own performance requirements, environmental conditions, and available resources, thereby achieving flexible adaptation to time-varying environments and enhancing the generalization of the artificial intelligence model at the network edge. By proposing a cloud-edge-device collaborative model fine-tuning configuration method and signaling interaction process, and proposing a model parameter configuration file generation and deployment strategy, a signaling method for characterizing model fine-tuning features is designed, applicable to various typical wireless artificial intelligence use cases. This method provides a solution for edge devices to perform online model fine-tuning, providing low-latency, customized model fine-tuning services to downstream edge devices based on cloud-based pre-trained models, and effectively addressing data distribution deviations caused by time-varying environments in online scenarios. This paper presents a customized model fine-tuning service for edge devices by designing an online model fine-tuning configuration method and signaling interaction mechanism. It also designs a model parameter configuration file generation and deployment strategy, defines a signaling method conforming to the 3GPP standard for characterizing model fine-tuning features, and provides the corresponding implementation process. Furthermore, it designs an online model fine-tuning implementation method, allowing terminal nodes to autonomously decide on the deployment and execution of model fine-tuning algorithms within each time period, thereby improving the adaptability of edge devices to time-varying environments while effectively protecting user privacy.

[0141] In another embodiment, as shown in Figure 4, an online model fine-tuning method based on edge intelligence is provided. Taking the application of this method to edge node 102 in Figure 1 as an example, the method includes the following steps:

[0142] S402, based on the model fine-tuning requirement information of each terminal node to which the edge node belongs, generates the model fine-tuning requirement information of the cell group.

[0143] S404. Upload the model fine-tuning requirements of the cell group to the network node so that the network node can generate a corresponding model fine-tuning configuration file list based on the pre-trained model according to the model fine-tuning requirements of the cell group.

[0144] S406, retrieve the list of model fine-tuning configuration files sent by the network nodes.

[0145] S408, receive an online model fine-tuning request uploaded by at least one terminal node to which the edge node belongs, and generate an online model fine-tuning point group based on at least one terminal node.

[0146] S410 broadcasts model fine-tuning feature signaling to each terminal node in the online model fine-tuning point group according to the model fine-tuning configuration file list.

[0147] The model fine-tuning configuration file list includes at least one model fine-tuning feature information generated based on the model fine-tuning requirements of each terminal node.

[0148] S412, obtain the subscription request signaling sent by any terminal node.

[0149] Among them, the subscription request signaling is sent by any terminal node after determining the model parameter configuration file to be subscribed based on the node operation information in the current time period and the model fine-tuning feature information carried by the model fine-tuning feature signaling.

[0150] S414 transmits the subscribed model parameter configuration file to any terminal node, so that any terminal node can perform online model fine-tuning based on the subscribed model parameter configuration file to obtain the updated model.

[0151] S416: Obtain online model monitoring information sent by any terminal node.

[0152] The online model monitoring information includes the model verification results of any terminal node performing online verification of the updated model.

[0153] S418, if the updated model, as represented by the online model monitoring information, meets the fine-tuning requirements, remove any terminal node from the online model fine-tuning point group to obtain the updated online model fine-tuning point group.

[0154] S420 updates the subscribed model parameter configuration file based on online model monitoring information to obtain the updated model parameter configuration file, and updates the model fine-tuning feature information to obtain the updated model fine-tuning feature information.

[0155] The model fine-tuning feature information includes at least one of the following: model parameter configuration file identifier, computational complexity, storage complexity, implemented functions, task type, expected performance gain, fine-tuning latency, and security level.

[0156] S422, determine whether the model fine-tuning termination condition is met. If it is met, end the process; otherwise, return to step S410.

[0157] The termination conditions for model fine-tuning include all terminal nodes either moving out of the online model fine-tuning point group or receiving an online model fine-tuning termination signal from the network node.

[0158] It should be noted that the specific limitations of the above steps can be found in the specific limitations of an online model fine-tuning method based on edge intelligence described above.

[0159] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0160] Based on the same inventive concept, this application also provides an edge-intelligence-based online model fine-tuning device for implementing the above-described edge-intelligence-based online model fine-tuning method. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations in one or more embodiments of the edge-intelligence-based online model fine-tuning device provided below can be found in the limitations of the edge-intelligence-based online model fine-tuning method described above, and will not be repeated here.

[0161] In an exemplary embodiment, as shown in FIG5, an online model fine-tuning device based on edge intelligence is provided, applied to the edge node of a model fine-tuning system. The model fine-tuning system further includes terminal nodes and network nodes; including:

[0162] The receiving module 510 is used to receive online model fine-tuning requests uploaded by at least one terminal node to which the edge node belongs, and to generate an online model fine-tuning point group based on at least one terminal node.

[0163] The broadcast module 520 is used to broadcast model fine-tuning feature signaling to each of the terminal nodes in the online model fine-tuning point group according to the model fine-tuning configuration file list; the model fine-tuning configuration file list includes at least one model fine-tuning feature information generated according to the model fine-tuning requirement information of each of the terminal nodes.

[0164] The acquisition module 530 is used to acquire a subscription request signaling sent by any of the terminal nodes; the subscription request signaling is sent by any of the terminal nodes after determining the model parameter configuration file to be subscribed based on the node running information in the current time period and the model fine-tuning feature information carried by the model fine-tuning feature signaling.

[0165] The transmission module 540 is used to transmit the subscribed model parameter configuration file to any of the terminal nodes, so that any of the terminal nodes can perform online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file to obtain an updated model.

[0166] In one embodiment, the edge intelligence-based online model fine-tuning device further includes a removal module, which is specifically used to acquire online model monitoring information sent by any of the terminal nodes; the online model monitoring information includes the model verification result of any of the terminal nodes performing online verification of the updated model; when the online model monitoring information indicates that the updated model meets the fine-tuning requirements, any of the terminal nodes is removed from the online model fine-tuning point group to obtain the updated online model fine-tuning point group.

[0167] In one embodiment, the model fine-tuning feature information includes at least one of the following: the identifier of the model parameter configuration file, computational complexity, storage complexity, implemented function, task type, expected performance gain, fine-tuning latency, and security level. The edge intelligence-based online model fine-tuning device further includes an update module, which is specifically used to update the subscribed model parameter configuration file according to the online model monitoring information to obtain the updated model parameter configuration file, and to update the model fine-tuning feature information to obtain the updated model fine-tuning feature information.

[0168] In one embodiment, the edge intelligence-based online model fine-tuning device further includes a return module, which is specifically used to return to the step of broadcasting model fine-tuning feature signaling to each of the terminal nodes in the online model fine-tuning adjustment point group according to the model fine-tuning configuration file list, until the model fine-tuning termination condition is met; wherein, the model fine-tuning termination condition includes each of the terminal nodes being removed from the online model fine-tuning adjustment point group or receiving an online model fine-tuning termination signaling issued by the network node.

[0169] In one embodiment, the acquisition module 530 is specifically used to generate model fine-tuning requirement information for cell groups based on the model fine-tuning requirement information of each terminal node to which the edge node belongs; upload the model fine-tuning requirement information of the cell groups to the network node, so that the network node generates a corresponding model fine-tuning configuration file list based on the pre-trained model according to the model fine-tuning requirement information of the cell groups; and acquire the model fine-tuning configuration file list sent by the network node.

[0170] The modules in the aforementioned edge intelligence-based online model fine-tuning device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0171] In an exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram is shown in Figure 6. The computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the computer device stores model data. The I / O interfaces of the computer device are used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an online model fine-tuning method based on edge intelligence.

[0172] Those skilled in the art will understand that the structure shown in Figure 6 is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0173] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0174] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0175] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0176] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0177] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0178] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0179] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. An online model fine-tuning method based on edge intelligence, applied to edge nodes of a model fine-tuning system, wherein the model fine-tuning system further includes terminal nodes and network nodes; the method includes: Receive online model fine-tuning requests uploaded by at least one terminal node to which the edge node belongs, and generate an online model fine-tuning point group based on at least one terminal node; According to the model fine-tuning configuration file list, model fine-tuning feature signaling is broadcast to each of the terminal nodes in the online model fine-tuning point group; the model fine-tuning configuration file list includes at least one model fine-tuning feature information generated based on the model fine-tuning requirement information of each of the terminal nodes. Obtain a subscription request signaling sent by any of the terminal nodes; the subscription request signaling is sent by any of the terminal nodes after determining the model parameter configuration file to be subscribed based on the node running information in the current time period and the model fine-tuning feature information carried by the model fine-tuning feature signaling. The subscribed model parameter configuration file is transmitted to any of the terminal nodes, so that any of the terminal nodes can perform online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file to obtain an updated model.

2. The method of claim 1, wherein after obtaining the updated model, the method further comprises: Obtain online model monitoring information sent by any of the terminal nodes; The online model monitoring information includes the model verification results of any of the terminal nodes performing online verification of the updated model. When the online model monitoring information indicates that the updated model meets the fine-tuning requirements, any of the terminal nodes is removed from the online model fine-tuning point group to obtain the updated online model fine-tuning point group.

3. The method according to claim 2, wherein the model fine-tuning feature information includes at least one of the following: model parameter configuration file identifier, computational complexity, storage complexity, implemented function, task type, expected performance gain, fine-tuning latency, and security level; after obtaining the updated online model fine-tuning point set, the method further includes: Based on the online model monitoring information, the subscribed model parameter configuration file is updated to obtain the updated model parameter configuration file, and the model fine-tuning feature information is updated to obtain the updated model fine-tuning feature information.

4. The method according to claim 3, wherein after obtaining the updated model fine-tuning feature information, the method further comprises: Returning to the step of broadcasting model fine-tuning feature signaling to each terminal node in the online model fine-tuning point group according to the model fine-tuning configuration file list, until the model fine-tuning termination condition is met; The model fine-tuning termination condition includes all terminal nodes either moving out of the online model fine-tuning point group or receiving an online model fine-tuning termination signal from the network node.

5. The method according to claim 2, wherein the fine-tuning requirement conditions include at least one of the following conditions: the online model validation loss value or convergence speed is less than a first preset threshold; the online model validation accuracy is greater than a second preset threshold; the change in consecutive T rounds is less than a third preset threshold; or, the online model validation latency, throughput or spectral efficiency is less than a fourth preset threshold.

6. The method of claim 1, wherein before receiving the online model fine-tuning request uploaded by at least one terminal node to which the edge node belongs, the method further comprises: Based on the model fine-tuning requirement information of each terminal node to which the edge node belongs, generate the model fine-tuning requirement information of the cell group. The model fine-tuning requirements of the cell group are uploaded to the network node, so that the network node can generate a corresponding model fine-tuning configuration file list based on the pre-trained model according to the model fine-tuning requirements of the cell group. Obtain the list of model fine-tuning configuration files sent by the network node.

7. The method according to claim 1, wherein the model parameter configuration file includes improvements in at least one aspect of the structure, function, number of model parameters, and connection method of the pre-trained model.

8. The method according to claim 1, wherein the pre-trained model includes a large language model and a neural network model; wherein the large language model includes a converter model, a generative pre-trained converter model, and a decoding-based enhancement model; and the neural network model includes a convolutional neural network, a recurrent neural network, a fully connected network, and a generative adversarial network.

9. The method according to claim 1, wherein the model fine-tuning feature signaling is transmitted after the downlink control information or is transmitted within the downlink control information field.

10. An online model fine-tuning system based on edge intelligence, wherein the system is used to perform the steps of the online model fine-tuning method based on edge intelligence as described in any one of claims 1 to 9; the system includes network nodes, edge nodes, and terminal nodes; The network node is used to generate a corresponding model fine-tuning configuration file list based on the model fine-tuning requirements information of the cell group uploaded by the edge node; The model fine-tuning requirement information for the cell group is generated based on the model fine-tuning requirement information of each terminal node to which the edge node belongs. The edge node is used to obtain the list of model fine-tuning configuration files sent by the network node; The terminal node is used to send an online model fine-tuning request to the edge node; The edge node is configured to generate an online model fine-tuning point group based on at least one of the terminal nodes, and to broadcast model fine-tuning feature signaling to each of the terminal nodes in the online model fine-tuning point group based on the model fine-tuning configuration file list. The terminal node is used to determine the subscribed model parameter configuration file based on the node operation information in the current time period and the model fine-tuning feature information carried by the model fine-tuning feature signaling, and send a subscription request signaling to the edge node. The edge node is used to transmit the subscribed model parameter configuration file to the terminal node; The terminal node is used to perform online model fine-tuning of the pre-trained model based on the subscribed model parameter configuration file to obtain an updated model.

11. An online model fine-tuning device based on edge intelligence, applied to the edge nodes of a model fine-tuning system, the model fine-tuning system further comprising terminal nodes and network nodes; the device comprising: The receiving module is used to receive online model fine-tuning requests uploaded by at least one terminal node to which the edge node belongs, and to generate an online model fine-tuning point group based on at least one terminal node; The broadcast module is used to broadcast model fine-tuning feature signaling to each terminal node in the online model fine-tuning point group according to the model fine-tuning configuration file list; The model fine-tuning configuration file list includes at least one model fine-tuning feature information generated based on the model fine-tuning requirements of each terminal node; The acquisition module is used to acquire subscription request signaling sent by any of the terminal nodes; The subscription request signaling is sent by any of the terminal nodes after determining the model parameter configuration file to be subscribed based on the node operation information in the current time period and the model fine-tuning feature information carried by the model fine-tuning feature signaling. The transmission module is used to transmit the subscribed model parameter configuration file to any of the terminal nodes, so that any of the terminal nodes can perform online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file to obtain an updated model.

12. A computer device comprising a memory and a processor, the memory storing a computer program, wherein the processor executes the computer program to implement the steps of the method according to any one of claims 1 to 9.

13. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method according to any one of claims 1 to 9.

14. A computer program product comprising a computer program, wherein the computer program, when executed by a processor, implements the steps of the method according to any one of claims 1 to 9.

Images