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

Abstract

The application relates to an online model fine-tuning method, device and computer equipment based on edge intelligence. The method comprises the following steps: receiving an online model fine-tuning request uploaded by at least one terminal node to which an edge node belongs, and generating an online model fine-tuning node group according to the at least one terminal node; broadcasting model fine-tuning feature signaling to each terminal node in the online model fine-tuning node group according to a model fine-tuning configuration file list; obtaining a subscription request signaling sent by any terminal node; and transmitting a subscribed model parameter configuration file to any terminal node, so that any terminal node performs online model fine-tuning on a pre-training model based on the subscribed model parameter configuration file, and obtains an updated model. The method can improve the efficiency of model configuration in edge intelligence.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to an online model fine-tuning method and device based on edge intelligence, a computer device, a computer readable storage medium and a computer program product. BACKGROUND

[0002] With the continuous improvement of user equipment computing power and the increasing demand for massive data processing, network intelligence is expanding from cloud servers to network edges, making edge intelligence a key driving factor for the application of the sixth generation of mobile communication standards. Edge intelligence combines artificial intelligence models with edge computing technology and relies on edge computing architecture to distribute data processing, model inference and part of the training tasks on the edge nodes and terminal nodes of the network, rather than relying on traditional centralized cloud server computing. Edge intelligence reduces the delay of data transmission to the cloud by performing intelligent computing on edge nodes close to data sources, thereby improving response speed and privacy protection capabilities. Edge intelligence is widely used in smart homes, autonomous driving, industrial automation and other fields, and can realize real-time monitoring, anomaly detection, predictive maintenance and other functions.

[0003] In the implementation of edge intelligence of wireless networks, when there is a certain deviation between the deployment environment of the artificial intelligence model and its training environment, the data distribution learned by the artificial intelligence model in the training environment is inconsistent with the actual data distribution in the deployment environment, which will cause the performance of the artificial intelligence model to decline. Therefore, model configuration needs to be performed on the basis of the existing pre-trained model, that is, the model is adjusted in parameters using new data so that the model can adapt to the new environment or task. However, with more and more user terminals accessing the network edge and people's increasing demand for artificial intelligence services, model configuration faces multiple challenges such as limited computing and communication resources and data security, resulting in low efficiency of model configuration in edge intelligence. SUMMARY

[0004] Therefore, it is necessary to provide an online model fine-tuning method, device, computer device, computer readable storage medium and computer program product based on edge intelligence, which can improve the efficiency of model configuration in edge intelligence.

[0005] In a first aspect, the present application provides an online model fine-tuning method based on edge intelligence, applied to an edge node of a model fine-tuning system, the model fine-tuning system further comprising a terminal node and a network node; comprising:

[0006] 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 node group according to at least one terminal node;

[0007] According to a model fine-tuning configuration file list, broadcast model fine-tuning feature signaling to each of the terminal nodes in the online model fine-tuning node group; the model fine-tuning configuration file list includes at least one model fine-tuning feature information generated according to model fine-tuning requirement information of each of the terminal nodes;

[0008] Obtain subscription request signaling sent by any of the terminal nodes; the subscription request signaling is sent by any of the terminal nodes in a case where model parameter configuration files subscribed are determined according to node running information in a current time period and model fine-tuning feature information carried by the model fine-tuning feature signaling;

[0009] Transmit the subscribed model parameter configuration files to any of the terminal nodes, so that any of the terminal nodes performs online model fine-tuning on a pre-trained model based on the subscribed model parameter configuration files to obtain an updated model.

[0010] In one of the embodiments, after the updated model is obtained, the method further includes:

[0011] Obtain online model monitoring information sent by any of the terminal nodes; the online model monitoring information includes a model verification result of online verification of the updated model by any of the terminal nodes;

[0012] In a case where the online model monitoring information indicates that the updated model meets fine-tuning requirement conditions, remove any of the terminal nodes from the online model fine-tuning node group to obtain an updated online model fine-tuning node group.

[0013] In one of the embodiments, the model fine-tuning feature information includes at least one of an identifier of a model parameter configuration file, a calculation complexity, a storage complexity, an implementation function, a task type, an expected performance gain, a fine-tuning time delay, and a security level; after the updated online model fine-tuning node group is obtained, the method further includes:

[0014] According to the online model monitoring information, update the subscribed model parameter configuration files to obtain updated model parameter configuration files, and update the model fine-tuning feature information to obtain updated model fine-tuning feature information.

[0015] In one of the embodiments, after the updated model fine-tuning feature information is obtained, the method further includes:

[0016] Return to the step of broadcasting model fine-tuning feature signaling to each of the terminal nodes in the online model fine-tuning node group according to the model fine-tuning configuration file list until a model fine-tuning termination condition is met;

[0017] The model fine-tuning termination condition comprises that each of the terminal nodes moves out of the online model fine-tuning node group or receives online model fine-tuning termination signaling sent by the network node.

[0018] In one of the embodiments, before the online model fine-tuning request is uploaded to at least one terminal node to which the edge node belongs, the method further comprises:

[0019] According to the model fine-tuning requirement information of each terminal node to which the edge node belongs, model fine-tuning requirement information of a cell group is generated.

[0020] The model fine-tuning requirement information of the cell group is uploaded to a network node, so that the network node generates a corresponding model fine-tuning configuration file list on the basis of a pre-trained model according to the model fine-tuning requirement information of the cell group.

[0021] The model fine-tuning configuration file list sent by the network node is acquired.

[0022] In a second aspect, the present application further provides an online model fine-tuning system based on edge intelligence, which is used to implement the steps of the above method; the system comprises a network node, an edge node and a terminal node.

[0023] The network node is configured to generate a corresponding model fine-tuning configuration file list on the basis of a pre-trained model according to model fine-tuning requirement information of a cell group uploaded by the edge node; the model fine-tuning requirement information of the cell group is generated according to model fine-tuning requirement information of each terminal node to which the edge node belongs.

[0024] The edge node is configured to acquire the model fine-tuning configuration file list sent by the network node.

[0025] The terminal node is configured to send an online model fine-tuning request to the edge node.

[0026] The edge node is configured to generate an online model fine-tuning node group according to at least one terminal node, and broadcast model fine-tuning feature signaling to each of the terminal nodes in the online model fine-tuning node group according to the model fine-tuning configuration file list.

[0027] The terminal node is configured to determine a subscribed model parameter configuration file according to node running information in a current time period and model fine-tuning feature information carried by the model fine-tuning feature signaling, and send a subscription request signaling to the edge node.

[0028] The edge node is configured to transmit the subscribed model parameter configuration file to the terminal node.

[0029] The terminal node is configured to perform online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file, to obtain an updated model.

[0030] In a third aspect, the present application further provides an online model fine-tuning device based on edge intelligence, applied to an edge node of a model fine-tuning system, the model fine-tuning system further comprising a terminal node and a network node; the device comprises:

[0031] The receiving module is configured to 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 node group according to the at least one terminal node;

[0032] The broadcasting module is configured to broadcast model fine-tuning feature signaling to each terminal node in the online model fine-tuning node group according to a model fine-tuning configuration file list; the model fine-tuning configuration file list comprises at least one model fine-tuning feature information generated according to model fine-tuning requirement information of each terminal node;

[0033] The obtaining module is configured to obtain a subscription request signaling sent by any terminal node; the subscription request signaling is sent by any terminal node in a case where the terminal node determines a subscribed model parameter configuration file according to node running information in a current time period and model fine-tuning feature information carried by the model fine-tuning feature signaling;

[0034] The transmission module is configured to transmit the subscribed model parameter configuration file to any terminal node, so that any terminal node performs online model fine-tuning on a pre-trained model based on the subscribed model parameter configuration file, to obtain an updated model.

[0035] In a fourth aspect, the present application further provides a computer device comprising a memory and a processor, the memory storing a computer program, and the processor realizing steps of the above method when executing the computer program.

[0036] In a fifth aspect, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program realizes steps of the above method when executed by a processor.

[0037] In a sixth aspect, the present application further provides a computer program product comprising a computer program, and the computer program realizes steps of the above method when executed by a processor.

[0038] The online model fine-tuning method, device, computer equipment, computer readable storage medium and computer program product based on edge intelligence 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 node group according to the at least one terminal node; according to a model fine-tuning configuration file list, broadcast model fine-tuning feature signaling to each terminal node in the online model fine-tuning node group, wherein the model fine-tuning configuration file list includes at least one model fine-tuning feature information generated according to model fine-tuning requirement information of each terminal node; obtain a subscription request signaling sent by any terminal node; wherein the subscription request signaling is sent by any terminal node in a case where the model parameter configuration file subscribed is determined according to node running information in a current time period and model fine-tuning feature information carried by the model fine-tuning feature signaling; and transmit the subscribed model parameter configuration file to any terminal node, so that any terminal node performs online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file, and obtains an updated model. Through receiving the online model fine-tuning request of the terminal node, dynamically generating the online model fine-tuning node group, and broadcasting the model fine-tuning feature signaling according to the model fine-tuning configuration file list, through the signaling interaction process between the edge node and the terminal node, it is ensured that the model configuration can be efficiently and specifically optimized according to the actual demand, the terminal node autonomously and flexibly selects the adaptive model parameter configuration file according to the performance, environment and resource status of the terminal node, and performs online model fine-tuning on the pre-trained model, realizes flexible adaptation to the time-varying environment, enhances the generalization of the artificial intelligence model at the network edge side, and thus provides customized artificial intelligence services for edge users based on the general pre-trained model, reduces the data transmission amount and communication delay, and improves the model configuration efficiency and real-time performance. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the description of the embodiments of the present application or the related art will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other related drawings can be obtained by those skilled in the art without creating any inventive labor.

[0040] Figure 1 An application environment diagram of an online model fine-tuning method based on edge intelligence in an embodiment;

[0041] Figure 2 A flowchart of an online model fine-tuning method based on edge intelligence in an embodiment;

[0042] Figure 3 A signaling interaction logic diagram of an online model fine-tuning method based on edge intelligence in an embodiment;

[0043] Figure 4 This is a flowchart illustrating an online model fine-tuning method based on edge intelligence in another embodiment;

[0044] Figure 5 This is a structural block diagram of an online model fine-tuning device based on edge intelligence in one embodiment;

[0045] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0046] 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.

[0047] The online model fine-tuning method based on edge intelligence provided in this application can be applied to, for example... Figure 1 The application environment shown. This edge intelligence-based online model fine-tuning system includes network node 100, edge node 102, and terminal node 104.

[0048] Network node 100 manages and generates model parameter configuration files, maintains model versions, and configures 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 monitors the overall network operation status to optimize resource allocation and task scheduling. 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.

[0049] 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.

[0050] In one of the embodiments, the edge node 102 can include a centralized unit (CU) and a distributed unit (DU). Specifically, the centralized unit is an important component in the radio access network, mainly responsible for higher level protocol processing and resource management, which can be used for model transmission, signaling configuration, model update, etc. The centralized unit can be further divided into two parts: control plane centralized unit (CU-CP) and user plane centralized unit (CU-UP). Among them, the control plane centralized unit (CU-CP) can be responsible for the processing of control signaling, such as user equipment access management, handover management and connection maintenance, centralized management and coordination of multiple distributed units (DU). The user plane centralized unit (CU-UP) can be responsible for the forwarding and processing of user data, including the transmission of packet data, providing data plane processing, assisting the distributed unit (DU) to complete efficient data transmission tasks. Specifically, the distributed unit is mainly responsible for low-level physical layer and real-time task processing, and executes specific tasks allocated by the centralized unit, which can be used for data collection, model transmission, etc.

[0051] Among them, the terminal node 104 is a user equipment (UE), which can be used for data collection, such as collecting environment and task related data (such as sensor data, user behavior data, etc.). The terminal node 104 can also be used for model subscription and local model fine-tuning, that is, adjusting the pre-trained model locally according to the received model parameter configuration file to generate a model suitable for the current environment or task; the terminal node 104 can also perform performance monitoring, real-time monitoring of the running state of the model, including accuracy, delay and stability, etc., and feed back the monitoring information to the edge node 102.

[0052] It can be seen that the terminal node 104 can include a model fine-tuning function module, a model monitoring function module, a data acquisition function module, etc.; the edge node 102 can include a model fine-tuning function module, a model management function module, a model storage function module, etc.; the network node 100 can include a model management module, a model storage module, etc.

[0053] In a specific implementation, the network node 100 is configured to generate 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 uploaded by the edge node 102; the model fine-tuning requirement information of the cell group is generated according to the model fine-tuning requirement information of each terminal node 104 to which the edge node 102 belongs;

[0054] The edge node 102 is configured to obtain the model fine-tuning configuration file list sent by the network node 100;

[0055] The terminal node 104 is configured to send an online model fine-tuning request to the edge node 102.

[0056] The edge node 102 is configured to generate an online model fine-tuning node group according to at least one terminal node 104, and broadcast model fine-tuning feature signaling to each terminal node 104 in the online model fine-tuning node group according to a model fine-tuning configuration file list.

[0057] The terminal node 104 is configured to determine a subscribed model parameter configuration file according to node running information in a current time period and model fine-tuning feature information carried by the model fine-tuning feature signaling, and send a subscription request signaling to the edge node 102.

[0058] The edge node 102 is configured to transmit the subscribed model parameter configuration file to the terminal node 104.

[0059] The terminal node 104 is configured to perform online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file, to obtain an updated model.

[0060] In an exemplary embodiment, as shown in Figure 2 , an edge intelligence-based online model fine-tuning method is provided, which is applied to the edge node 102 in Figure 1 for example, and includes the following steps.

[0061] In step S202, an online model fine-tuning request uploaded by at least one terminal node to which the edge node belongs is received, and an online model fine-tuning node group is generated according to the at least one terminal node.

[0062] The online model fine-tuning request can be a request sent by the terminal node to the edge node during running, to express the current demand of the terminal node for artificial intelligence model fine-tuning. The online model fine-tuning request can include the current resource demand (such as computing power, bandwidth, etc.), real-time demand, and current artificial intelligence task type (such as natural language processing, image classification, target detection, etc.) of the terminal node. In a specific implementation, the terminal node can report the online model fine-tuning request to the edge node at any time in each time period, and after the edge node responds to the request, participate in the online model fine-tuning process in the current time period or the next time period.

[0063] Optionally, the online model fine-tuning request can be sent in the form of Downlink Control Information (DCI), Media Access Control Control Element (MAC CE), or Radio Resource Control Signaling (RRC). The Downlink Control Information can be a signaling between the user equipment and the base station for transmitting control information, the Media Access Control Control Element can be a control signaling between the user equipment and the base station for state reporting and instruction transmission, and the Radio Resource Control Signaling can be a signaling in the Radio Resource Control layer responsible for configuring and managing the wireless link. Through these signalings, the online model fine-tuning request can be embedded in different layers of the protocol stack to ensure efficient transmission in different scenarios.

[0064] Optionally, the triggering mode of the online model fine-tuning request includes at least one of the following: the terminal node determines whether to perform online model fine-tuning based on model monitoring information, triggers an online model fine-tuning event, and sends an online model fine-tuning request to the edge node; the network node or the edge node periodically issues a model fine-tuning instruction to the terminal node to make the terminal node generate an online model fine-tuning request.

[0065] In a specific implementation, the edge node can record the terminal nodes that send online model fine-tuning requests in the current time period, determine these terminal nodes as all terminal nodes participating in online model fine-tuning in the current time period, and form an online model fine-tuning node group according to these terminal nodes.

[0066] In a specific implementation, the online model fine-tuning node group can include terminal nodes that newly join the online model fine-tuning process in the current time period and terminal nodes that have joined the online model fine-tuning process in the previous time period and have not exited. The online model fine-tuning node group can record the information of each type of terminal node according to the two types of terminal nodes, or can jointly record all terminal node information without distinguishing between the two types.

[0067] Optionally, the online model fine-tuning node group records at least the terminal node identifier (ID) of each terminal node, the cell / cell group to which the terminal node belongs, the timestamp of the terminal node joining the online model fine-tuning, and whether the terminal node participates in joint fine-tuning. The cell can refer to an area served by a base station in the wireless network, and the cell group can refer to a set composed of multiple adjacent or functionally related cells. Whether the terminal node participates in joint fine-tuning can mean that multiple terminal nodes can cooperatively complete the model fine-tuning task and can share model update information or fine-tuning results.

[0068] In a specific implementation, the online model fine-tuning node group supports the terminal nodes to join and leave the group at any time, and the information of the online model fine-tuning node group is updated after each terminal node joining and leaving event occurs.

[0069] In step S204, according to the model fine-tuning configuration file list, the model fine-tuning feature signaling is broadcasted to each terminal node in the online model fine-tuning node group.

[0070] In a specific implementation, the edge node can filter the model fine-tuning feature information from the model fine-tuning configuration file list according to the model fine-tuning configuration file list, generate the model fine-tuning feature signaling based on the model fine-tuning feature information, and broadcast the model fine-tuning feature signaling to each terminal node in the online model fine-tuning node group, where the model fine-tuning feature signaling carries the model fine-tuning feature information. The model fine-tuning configuration file list can include a plurality of model parameter configuration files and corresponding model fine-tuning feature information. In a specific implementation, the model fine-tuning configuration file list can be generated by the network node based on the model fine-tuning requirement information of the terminal node and on the basis of the pre-trained model, and then distributed to the edge node.

[0071] The model parameter configuration file can include parameter configuration information of the pre-trained model, which can be used to guide the online model fine-tuning of the pre-trained model. Each model parameter configuration file can correspond to unique model fine-tuning feature information.

[0072] The model fine-tuning feature information can be used to describe the feature information related to the execution of the model fine-tuning in the model parameter configuration file.

[0073] 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 terminal node. Optionally, the model fine-tuning requirement information can include the computing power configuration information of the corresponding terminal node that can be used for model fine-tuning, the storage space capability information that can be used for storing the model, the quality of service performance requirement of the model fine-tuning, the time delay tolerance, the application scenario, the data type or data format, and other historical data related to the model fine-tuning.

[0074] In one embodiment, the model fine-tuning feature information can be stored in the model fine-tuning configuration file list of the edge node in the form of a hash table, taking the file identifier ID of each model parameter configuration file as the key (key), designing a mapping function to map the key to the storage address (value) of the corresponding model fine-tuning feature information, so as to determine the storage address of the unique model fine-tuning feature information according to each file identifier ID, obtain a one-to-one correspondence between a plurality of key-value pairs, and ensure the efficiency of storage and retrieval.

[0075] In one of the embodiments, the model fine-tuning feature information includes at least one of the following: an identification of the model parameter configuration file, a computational complexity, a storage complexity, an implementation function, a task type, an expected performance gain, a fine-tuning delay, and a security level.

[0076] The identification of the model parameter configuration file can be a unique ID of the model parameter configuration file, used to distinguish between various model parameter configuration files. The computational complexity represents the size of the computing resources required by the model parameter configuration file, including the number of CPU / GPU processors, the number of cycles, the amount of floating-point operations, and the like required to perform model fine-tuning. The storage complexity represents the size of the storage resources required by the model parameter configuration file, including the number of disk space occupied, the storage format, and whether compression processing is supported. The implementation function represents the input and output types of the model parameter configuration file, such as the data type of the input data, the data format of the output data, and whether multi-modal data input is supported. The task type represents the type of tasks that can be performed by the model parameter configuration file, including natural language processing, image classification, sentiment classification, time series prediction, object detection, and the like. The expected performance gain represents the expected performance gain of the model parameter configuration file in terms of quality of service (QoS), energy efficiency, system performance, and the like, which is related to the self-evaluation results of the network node. The fine-tuning delay represents the processing time of the model parameter configuration file when performing fine-tuning tasks, which is generally related to the type of tasks performed, such as the time required to process one data sample, the time required to process a batch of data samples, and the like. The security level represents the confidentiality level of the model parameter configuration file. If the model structure or design method of the model parameter configuration file has significant autonomy and innovation, it has a high security level, and to protect the security of data privacy, it will be limited to distribution to terminal nodes or only to distribution to terminal nodes that comply with security requirements.

[0077] For example, the model fine-tuning feature information can be as shown in Table 1.

[0078] Table 1

[0079]

[0080] In a specific implementation, the edge node can determine the corresponding model fine-tuning application scenario according to the online model fine-tuning request sent by the terminal node. For different model fine-tuning application scenarios, the model fine-tuning feature information filtered by the edge node is different, and therefore the generated model fine-tuning feature signaling is different. The model fine-tuning application scenario can be the specific use occasion or task requirement of the artificial intelligence model in different actual applications. Since there are differences in the requirements of different scenarios for the functions, performance, resource consumption, etc. of the model, the edge node needs to filter the most suitable model fine-tuning feature information according to these scenario characteristics. For example, the terminal node may need different types of model fine-tuning algorithms for different application tasks, such as speech recognition, user behavior prediction, target detection, path planning, device anomaly detection, and predictive maintenance, which correspond to different types of model fine-tuning algorithms. For another example, the computing power, storage space, bandwidth, etc. of the terminal node will limit the complexity and fine-tuning method of the model, for example, a high-performance terminal node can run a complex fine-tuning algorithm, and a low-performance terminal node can run a lightweight model. Therefore, for these different model fine-tuning application scenarios, the edge node can filter different model fine-tuning feature information, thereby broadcasting different model fine-tuning feature signaling.

[0081] In one of the embodiments, the edge node generates model fine-tuning feature signaling based on the model fine-tuning feature information. The model fine-tuning feature signaling can contain the model fine-tuning feature information and can also implicitly or explicitly include a model fine-tuning configuration file list. Implicitly, the model fine-tuning feature information already contains the identification ID of the model parameter configuration file, and the edge node does not need to transmit it separately. Explicitly, the edge node transmits the model fine-tuning configuration file list to the terminal node separately once.

[0082] In a specific implementation, the edge node broadcasts the model fine-tuning feature signaling to each terminal node in the online model fine-tuning node group. Optionally, the model fine-tuning feature signaling can be transmitted by the control plane centralized unit in the edge node through downlink control information (DCI) or system information broadcast (SIB), or can be sent by the user plane centralized unit in the edge node through a data radio bearer (DRB) to the user plane centralized unit. The model fine-tuning feature signaling can be transmitted after the downlink control information (DCI) or can be contained in the downlink control information (DCI) field.

[0083] In step S206, the subscription request signaling sent by any terminal node is obtained.

[0084] The subscription request signaling is sent by any terminal node after determining the subscribed model parameter configuration file according to 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 subscription request signaling is a signaling sent by the terminal node to the edge node, which is used to request a specific model parameter configuration file.

[0085] The node running information includes at least one of performance information, environment information, and resource information. The terminal node determines the model parameter configuration file to be subscribed according to matching of the node running information of the current time period and the model fine-tuning feature information.

[0086] The model parameter configuration file includes parameter configuration information of the pre-trained model. The model parameter configuration file can include configurations optimized for specific tasks or requirements of the terminal node. The model parameter configuration file can define parameter adjustment ranges, optimization objectives, training strategies, and the like.

[0087] In one embodiment, the model parameter configuration file includes improvements in structure, function, model parameter quantity, connection mode, and the like for the pre-trained model, which can change part of the original backbone model or add new modules based on the original backbone model.

[0088] In one embodiment, the pre-trained model can be a large language model, or a general neural network model with a smaller scale, such as a convolutional neural network CNN, a recurrent neural network RNN, a fully connected network MLP, a generative adversarial network GAN, and the like.

[0089] In a specific implementation, the terminal node selects one or more model parameter configuration files according to the performance information, environment information, and resource information in the current time period. Optionally, the granularity of the time period can be a time slot, a half frame, a frame, or a longer time interval, which is uniformly configured and notified to the terminal node by the edge node.

[0090] Optionally, the performance information can refer to self-performance requirements, including requirements for model fine-tuning-related performance indicators such as precision, convergence speed, energy consumption, latency, throughput, spectral efficiency, quality of service (QoS), and the like; the environment information can refer to environmental conditions, including wireless channel conditions, data arrival conditions, service processing conditions, and the like; and the resource information can refer to available resources, which can include computing resources, communication resources, storage resources, and the like that can be used by the local device.

[0091] In one of the embodiments, the way the terminal node selects the subscribed model parameter configuration file can include: the terminal node determines the application scenario in which the terminal node needs to perform model fine-tuning, the performance requirement threshold, and the like according to the performance information, the environment information, and the resource information in the current time period; selects the model fine-tuning feature information corresponding to each model parameter configuration file that matches the application scenario of the terminal node according to the model fine-tuning configuration file list, compares the model fine-tuning feature information with the performance requirement threshold of the terminal node, and obtains the score of each model parameter configuration file. The model parameter configuration file that does not match the application scenario of the terminal node is not scored. The terminal node selects the N model parameter configuration files with the highest scores from all the model parameter configuration files as the model parameter configuration files to be subscribed by the terminal node, and forms a model parameter configuration file subscription list. Optionally, the value of N is configured by the terminal node itself, or can be configured by the edge node through downlink control information or a medium access control layer control unit.

[0092] In one of the embodiments, the model parameter configuration file subscription list can correspond to the model fine-tuning configuration file list issued by the edge node, and the same arrangement order is maintained to keep alignment at the edge node and the terminal node. A 1-bit flag bit can be added after each file identification ID in the model fine-tuning configuration file list to obtain the model parameter configuration file subscription list. The flag bit of the subscribed model parameter configuration file in the model parameter configuration file subscription list is set to 1, and the flag bits of the remaining model parameter configuration files are set to 0.

[0093] In one of the embodiments, the subscription request signaling can include the model parameter configuration file subscription list related to the terminal node, and further include information such as a timestamp, a subscription duration, a subscription mode, and a supported storage format.

[0094] Optionally, the subscription request signaling can be carried through control plane uplink control information (UCI) signaling or user plane data radio bearer (DRB).

[0095] Step S208: transmitting the subscribed model parameter configuration file to any terminal node, so that any terminal node performs online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file to obtain an updated model.

[0096] In a specific implementation, the edge node transmits the subscribed model parameter configuration file to the corresponding terminal node based on the subscription request signaling reported by the terminal node. The model parameter configuration file can include parameter configuration information of the pre-trained model, which can be used to guide how to perform model fine-tuning and model configuration on the pre-trained model.

[0097] Optionally, the edge node can transmit the subscribed model parameter configuration file based on a radio bearer (DRB), and use downlink control information (DCI) to indicate the allocated time-frequency domain transmission resource.

[0098] In a specific implementation, the method for transmitting the model parameter configuration file of the subscription can include: the edge node transmitting the model file of the model parameter configuration file to the distributed unit (DU) through an F1 interface from the centralized unit (CU) to which the terminal node belongs, and transmitting the model file to the terminal node through a Uu interface, where the F1 interface is a communication interface connecting the distributed unit (DU) and the centralized unit (CU), and the Uu interface is a 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 the centralized unit (CU) to which other non-terminal node belongs to the centralized unit (CU) to which the terminal node currently belongs through an Xn interface, and then transmitting the model file to the distributed unit (DU) through an F1 interface from the centralized unit (CU), and transmitting the model file to the terminal node through a Uu interface, where the Xn is an interface for connecting different base stations; or the edge node transmitting the model file of the model parameter configuration file to the terminal node through a Uu interface from the distributed unit (DU).

[0099] In one of the embodiments, the edge node can transmit the customized fine-tuning mode based on the storage format of the model parameter configuration file. The storage format of the model parameter configuration file can adopt an extensible format such as.h5,.pb,.onnx format, or a serialized storage format such as.pt,.json,.pkl. Different terminal nodes can support different deep learning frameworks, thereby supporting different storage formats. Therefore, the edge node can save different storage formats of the same model parameter configuration file, and transmit the model file of the customized fine-tuning mode in different storage formats based on different terminal nodes.

[0100] In one of the embodiments, the edge node can participate in the decision process of the module subscription, for example, based on the global quality of service (QoS) requirement, the global energy consumption requirement, the computing power scheduling capability, the model parameter configuration file utilization rate, the stability requirement, the generalization requirement, and other factors of the current cell / cell group, to adjust the module subscription result of the terminal node, and to ensure the compliance with the initial module subscription request of the terminal node to the greatest extent.

[0101] In a specific implementation, the terminal node performs model fine-tuning using an online fine-tuning dataset based on the subscribed model parameter configuration file transmitted by the edge node, and obtains an updated model. The model fine-tuning algorithm can adopt a gradient descent algorithm, determines the update direction of the model through the gradient value of the model parameter of the model parameter configuration file on each data sample or a batch of data samples, and updates the model parameters of the pre-trained model by setting an update step.

[0102] In the online model fine-tuning method based on edge intelligence, the online model fine-tuning request uploaded by at least one terminal node to which the edge node belongs is received, and an online model fine-tuning node group is generated according to the at least one terminal node; according to the model fine-tuning configuration file list, the model fine-tuning feature signaling is broadcasted to each terminal node in the online model fine-tuning node group, wherein the model fine-tuning configuration file list includes at least one model fine-tuning feature information generated according to the model fine-tuning demand information of each terminal node; the subscription request signaling sent by any terminal node is obtained; wherein the subscription request signaling is sent by any terminal node in the case that the model parameter configuration file to be subscribed is determined according to the node running information in the current time period and the model fine-tuning feature information carried in the model fine-tuning feature signaling; the subscribed model parameter configuration file is transmitted to any terminal node, so that any terminal node performs online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file to obtain an updated model. Through the online model fine-tuning request of the terminal node, the online model fine-tuning node group is dynamically generated, and the model fine-tuning feature signaling is broadcasted according to the model fine-tuning configuration file list. Through the signaling interaction process between the edge node and the terminal node, it is ensured that the model configuration can be efficiently and specifically optimized according to the actual demand. The terminal node autonomously and flexibly selects the adaptive model parameter configuration file according to the performance, environment and resource status of the terminal node, and performs online model fine-tuning on the pre-trained model. The flexible adaptation to the time-varying environment is realized, the generalization of the artificial intelligence model at the network edge side is enhanced, and thus the customized artificial intelligence service is provided for the edge user based on the general pre-trained model, while the data transmission amount and the communication delay are reduced, and the model configuration efficiency and the real-time performance are improved.

[0103] In another embodiment, after obtaining the updated model, the method further includes: obtaining online model monitoring information sent by any terminal node; the online model monitoring information includes a model verification result of any terminal node performing online verification on the updated model; in the case that the online model monitoring information indicates that the updated model meets the fine-tuning demand condition, any terminal node is removed from the online model fine-tuning node group to obtain an updated online model fine-tuning node group.

[0104] In a specific implementation, the terminal node collects online data in real time, and divides the collected data into an online fine-tuning data set and an online verification data set, with the division ratio being determined autonomously by the terminal node or configured by the edge node. Based on the subscribed model parameter configuration file transmitted by the edge node, the terminal node performs model fine-tuning using the online fine-tuning data set to obtain an updated model. The model fine-tuning algorithm can use a gradient descent algorithm to determine the update direction of the model by the gradient value of the model parameters in the model parameter configuration file on each data sample or a batch of data samples, and update the model parameters of the pre-trained model by setting the update step size. The terminal node performs online verification on the updated model, compares the verification result obtained based on the model parameters of the updated model and each data sample in the online verification data set with the label value to obtain the online verification performance on the data sample, and finally obtains the online verification performance based on the online verification data set by averaging all online verification data samples to obtain the model verification result.

[0105] In one of the embodiments, if the terminal node subscribes to multiple model parameter configuration files, the multiple updated model parameter configuration files are used to obtain the model verification result based on the verification result on the online verification data set, with the verification result being averaged or obtained by a specific model integration algorithm to obtain the online verification performance, and the model integration algorithm being determined autonomously by the terminal node.

[0106] In a specific implementation, the terminal node performs an online model monitoring mechanism to monitor the online verification performance in real time and obtain the model verification result. The model verification result can include at least one of the loss value, gradient value, convergence speed, precision, latency, throughput, and spectral efficiency of the online model verification, and can also include model usage feedback, model monitoring indicator description, online model fine-tuning log, and other information. The terminal node can generate online model monitoring information according to these model verification results and upload the information to the edge node.

[0107] The online model monitoring information can include the above-described model verification result, i.e., the feedback information generated by the terminal node after verifying the updated model, including the model performance, adaptability, and task completion status, which can be used to evaluate the effectiveness of the updated model and the model parameter configuration file.

[0108] Optionally, the terminal node can report the online model monitoring information to the edge node through uplink control information (UCI) signaling or user plane data radio bearer (DRB), and the reporting period can be set by the edge node or the terminal node, but it is necessary to ensure that the reporting is completed within the current time period.

[0109] In one of the embodiments, the online model monitoring information corresponds to the model parameter configuration file, and if the terminal node subscribes to multiple model parameter configuration files, it reports the online model monitoring information corresponding to multiple modules respectively, such as forming a list in the order of the identification ID of the model parameter configuration file, that is, [the ID of module 1: the online model monitoring information of module 1, …, the ID of module M: the online model monitoring information of module M].

[0110] In the specific implementation, since the online model monitoring information can reflect the model fine-tuning process and the model configuration result, if the online model monitoring information represents that the updated model meets the fine-tuning requirement condition, it can be indicated that the model configuration has been successful and meets the requirement of the fine-tuning task, thereby triggering the online model fine-tuning exit process of the terminal node.

[0111] Optionally, the fine-tuning requirement condition can include at least one of the following conditions: the online model verification loss value or the convergence speed is less than a certain preset threshold; the online model verification accuracy is greater than a certain preset threshold, or the change condition of continuous T rounds is less than a certain preset threshold; the online model verification delay, throughput or spectral efficiency is less than a certain preset threshold. The above-mentioned preset threshold and T value can be configured by the edge node system or configured by the terminal node itself.

[0112] In one of the embodiments, the terminal node triggers the online model fine-tuning exit process, sends an online model fine-tuning exit request to the edge node, and exits the online model fine-tuning node group. The edge node responds to the online model fine-tuning exit request, no longer sends the model parameter configuration file model fine-tuning feature information to the terminal node in the next time period, and updates the online model fine-tuning node group.

[0113] In one of the embodiments, the online model fine-tuning exit process of any terminal node can be triggered by the edge node according to the online model monitoring information reported by the terminal node, and the online model fine-tuning exit signaling is issued to the corresponding terminal node. In the case of feedback confirmation information of the terminal node, the edge node removes it from the online model fine-tuning node group, no longer sends the model parameter configuration file model fine-tuning feature information to the terminal node in the next time period, and updates the online model fine-tuning node group.

[0114] The technical scheme of the embodiment can verify whether the updated model in the terminal node meets the fine-tuning requirement condition in real time through the introduction of online model monitoring information, realize dynamic performance evaluation in the fine-tuning process, and enhance the adaptability to complex scenes and task requirements; when the updated model of a terminal node meets the fine-tuning requirement, the terminal node is removed from the online model fine-tuning node group, the resource allocation to the fine-tuned node can be reduced, more resources can be allocated to other nodes that need fine-tuning, and the dynamically removed node that has completed the task can avoid the redundancy of system resource allocation; the node group is removed in time after the terminal node completes fine-tuning and meets the requirement, which can accelerate the iteration efficiency of the fine-tuning task, and the dynamic joining and removing mechanism of the terminal node adapts to the characteristics of the dynamic change of the network node and the terminal node state in the edge intelligent system, and enhances the real-time response capability.

[0115] In another embodiment, after updating the online model fine-tuning node group, the method further includes: updating the subscribed model parameter configuration file according to 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.

[0116] The model fine-tuning feature information includes at least one of the following: an identifier of the model parameter configuration file, a calculation complexity, a storage complexity, an implementation function, a task type, an expected performance gain, a fine-tuning delay, and a security level.

[0117] The edge node can improve the existing model parameter configuration file according to the feedback of the model deficiencies in the monitoring information, for example, updating the subscribed model parameter configuration file can include adjusting the model parameter range to make it more suitable for specific scenes or task requirements.

[0118] In a specific implementation, the edge node can weight-sum all the online model verification gradient values reported by the terminal nodes in the online model fine-tuning node group according to the gradient value information in the online model monitoring information according to a certain weight, and update the model parameters contained in the subscribed model parameter configuration file by using the gradient descent method. Alternatively, the edge node can increase or decrease the number of model parameters contained in the subscribed model parameter configuration file according to a certain proportion by comprehensively considering the online model verification delay information reported by the terminal nodes in all online model fine-tuning node groups, so as to reduce the transmission delay of the online model verification.

[0119] Correspondingly, if the model parameter configuration file is updated, the corresponding model fine-tuning feature information is also updated.

[0120] The technical scheme of the embodiment can dynamically update the model parameter configuration file and the feature information thereof by acquiring online monitoring information in real time, can quickly optimize the module performance according to the latest feedback of the terminal node, and can make the model always adapt to the actual application scene; the fine-tuning module of the subscription is updated, unnecessary repeated debugging and performance degradation are avoided, and the optimization process of the model is ensured to be more accurate and efficient; the feature information is updated, the edge node can more accurately allocate the fine-tuning module to the terminal node meeting the demand, and the edge intelligent system can quickly respond to the change of the demand of the terminal node and adapt to the complex scene of multiple tasks and multiple environments.

[0121] In another embodiment, after obtaining the updated model fine-tuning feature information, the method further comprises: returning to the step of broadcasting the model fine-tuning feature signaling to each terminal node in the online model fine-tuning node 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 comprises that each terminal node is removed from the online model fine-tuning node group or that the online model fine-tuning termination signaling is received from the network node.

[0122] Optionally, the online model fine-tuning termination signaling sent by the network node can be forwarded to the edge node through a control plane dedicated interface. The edge node can issue the online model fine-tuning termination signaling to all terminal nodes, and can use the downlink control information, the medium access control layer control element, or the radio resource control signaling to transmit the online model fine-tuning termination signaling.

[0123] In a specific implementation, the iteration of the fine-tuning process continues until the model fine-tuning termination condition is met. When the terminal node confirms that the model meets the fine-tuning demand through online model verification, the terminal node is removed from the fine-tuning node group one by one. Once the online model fine-tuning node group is empty, it means that the fine-tuning task of all terminal nodes is completed, and the process is terminated at this time. Alternatively, the network node can actively issue a termination signaling according to the global demand (such as system resource reallocation, task priority adjustment, etc.), and forcibly end the fine-tuning process.

[0124] In the iteration process of the fine-tuning process, after obtaining the updated model fine-tuning configuration file list, the step of returning to broadcasting the model fine-tuning feature signaling to each terminal node in the online model fine-tuning node group according to the model fine-tuning configuration file list can dynamically update the model fine-tuning feature signaling, and ensure that the terminal node can obtain the model parameter configuration file most suitable for the current state in each iteration. The model performance is gradually optimized through multiple iterations, and finally the task demand is met.

[0125] The technical scheme of the embodiment realizes optimization of the model fine-tuning process in a dynamic iteration manner, ensures that the terminal node can continuously receive the most suitable model parameter configuration file and gradually meet the task requirements, and when the fine-tuning termination condition is met, the process naturally ends, thereby efficiently completing the model fine-tuning task in a dynamic environment and a multi-task scenario, and optimizing resource allocation.

[0126] In another embodiment, before receiving the online model fine-tuning request uploaded by the at least one terminal node, the method further comprises: generating model fine-tuning requirement information of a cell group according to 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.

[0127] The model fine-tuning requirement information can include computing power configuration information, storage space capability information, quality of service performance requirements, time delay tolerance, application scenarios, data types or data formats, and other historical data related to model fine-tuning of the corresponding terminal node.

[0128] The model fine-tuning requirement information of the cell group can include model fine-tuning requirement information of a cell and model fine-tuning requirement information of a cell group. The cell can refer to an area served by a base station in a wireless network, and the cell group can refer to a collection of multiple adjacent or functionally related cells. In addition to the model fine-tuning requirement information of each terminal node, the model fine-tuning requirement information of the cell group also includes information such as the number of models deployed within the cell / cell group, stability requirements, and generalization requirements.

[0129] In one embodiment, the model parameter configuration file includes improvements in structure, function, model parameter quantity, connection mode, etc. of the pre-trained model, which can change part of the original backbone model or add new modules based on the original backbone model. Optionally, the model parameter configuration file can be implemented using technologies such as low-rank adaptation (LoRA), adapter, prefix tuning, prompt tuning, etc.

[0130] In one embodiment, the pre-trained model can be a large language model such as a Transformer model, a generative pre-trained Transformer model, a decoding enhancement-based BERT model, etc. or a smaller general-purpose 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.

[0131] In a specific implementation, the network node can fine-tune the demand information according to the model of the cell group, generate a corresponding model parameter configuration file based on the pre-trained model, and generate corresponding model fine-tuning feature information according to the feature information related to model fine-tuning in the model parameter configuration file, and generate a model fine-tuning configuration file list based on the model parameter configuration file and the model fine-tuning feature information. Illustratively, the network node can generate a corresponding model parameter configuration file based on the pre-trained model by using the low-rank adaptive (LoRA) technology to generate the model parameter configuration file. Specifically, the weight matrix of the pre-trained model can be decomposed into two matrices with fewer parameters by low-rank decomposition, and the parameter size of the model parameter configuration file can be controlled by adjusting the rank of the matrix to form a series of model parameter configuration files for matching the performance requirements of different terminal nodes.

[0132] Optionally, the network node can transmit the model parameter configuration file to the edge node through a user plane data radio bearer (DRB) and transmit the corresponding configuration information through dedicated control signaling.

[0133] In one embodiment, the edge node sorts the model parameter configuration file based on at least one of the model parameter size, the model structure complexity, the customized design method used, the stability, and the generalization according to the model parameter configuration file transmitted by the network node, records the identification ID information of each module in order, and generates a model fine-tuning configuration file list.

[0134] The technical solution of the present embodiment reduces the response delay by generating fine-tuning modules in cooperation with the edge node and the network node before the terminal node sends a fine-tuning request; saves communication resources by aggregating the demand information of the cell group to avoid separate transmission of demand by the terminal node; improves the overall utilization efficiency of system resources by having the edge node only responsible for distribution and the network node concentrating on complex model optimization tasks; reduces repeated calculations by generating model parameter configuration files according to the demand of the cell group, which can simultaneously adapt to multiple terminal nodes; and supports different demand changes of the terminal node by dynamically maintaining the model fine-tuning configuration file list, thereby adapting to complex multi-task scenarios. In summary, through the cooperation of the edge node and the network node, the model fine-tuning configuration file list that adapts to the demand of the cell group is generated in advance, so that subsequent online model fine-tuning requests can be quickly responded to, which not only optimizes resource allocation but also improves the efficiency and adaptability of model fine-tuning, and is suitable for processing complex scenarios with multiple terminals and multiple tasks.

[0135] For the convenience of those skilled in the art, Figure 3 An exemplary signaling interaction logic diagram of an online model fine-tuning method based on edge intelligence is provided.

[0136] The online model fine-tuning method based on edge intelligence provided by the embodiments of the present application can face edge intelligence of communication standards such as the fifth generation mobile communication standard and the sixth generation mobile communication standard, provide customized artificial intelligence services for edge users based on a general pre-training model, and in each time period, the edge device autonomously decides to deploy and execute a model fine-tuning algorithm according to the performance requirements, environmental conditions and available resources of the edge device in the current time period, thereby realizing flexible adaptation to time-varying environments and enhancing the generalization of artificial intelligence models at the network edge. By proposing a cloud-edge-end collaborative model fine-tuning configuration method and a signaling interaction process, a model parameter configuration file generation and deployment strategy is proposed, and a signaling for representing model fine-tuning feature information is designed, which is suitable for various typical wireless artificial intelligence use case scenarios. The method provides a solution for edge devices to perform online model fine-tuning, provides low-latency, customized model fine-tuning services for downstream edge devices based on cloud pre-training models, and effectively deals with the data distribution deviation problem caused by time-varying environments in online scenarios. By designing an online model fine-tuning configuration method and a signaling interaction mechanism, customized model fine-tuning services are provided for edge devices; a model parameter configuration file generation and deployment strategy is designed, a signaling conforming to the Third Generation Partnership Project Plan (3GPP) standard for representing model fine-tuning feature information is defined, and the corresponding implementation process is given; an online model fine-tuning implementation method is designed, in which the terminal node autonomously decides to deploy and execute the model fine-tuning algorithm in each time period, and the adaptability of the edge device to time-varying environments is improved under the premise of effectively protecting user privacy.

[0137] In another embodiment, as shown in Figure 4 , an online model fine-tuning method based on edge intelligence is provided, which is applied to the edge node 102 in Figure 1 for example, and includes the following steps:

[0138] S402, generating model fine-tuning requirement information of a cell group according to model fine-tuning requirement information of each terminal node to which the edge node belongs.

[0139] S404, uploading the model fine-tuning requirement information of the cell group to a network node, so that the network node generates a corresponding model fine-tuning configuration file list based on the model fine-tuning requirement information of the cell group and the pre-training model.

[0140] S406, obtaining the model fine-tuning configuration file list sent by the network node.

[0141] S408, 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 node group according to the at least one terminal node.

[0142] S410, according to the model fine-tuning configuration file list, broadcast model fine-tuning feature signaling to each terminal node in the online model fine-tuning node group.

[0143] 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 terminal node.

[0144] S412, obtain the subscription request signaling sent by any terminal node.

[0145] The subscription request signaling is sent by any terminal node in the case of determining the subscribed model parameter configuration file according to the node running information in the current time period and the model fine-tuning feature information carried by the model fine-tuning feature signaling.

[0146] S414, transmit the subscribed model parameter configuration file to any terminal node, so that any terminal node performs online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file to obtain an updated model.

[0147] S416, obtain online model monitoring information sent by any terminal node.

[0148] The online model monitoring information includes model verification results of any terminal node on the updated model.

[0149] S418, in the case that the online model monitoring information represents that the updated model meets the fine-tuning requirement condition, remove any terminal node from the online model fine-tuning node group to obtain an updated online model fine-tuning node group.

[0150] S420, according to the online model monitoring information, update the subscribed model parameter configuration file to obtain an updated model parameter configuration file, and update the model fine-tuning feature information to obtain updated model fine-tuning feature information.

[0151] The model fine-tuning feature information includes at least one of the following: identification of the model parameter configuration file, calculation complexity, storage complexity, implementation function, task type, expected performance gain, fine-tuning delay, security level.

[0152] S422, determine whether the model fine-tuning termination condition is met, if yes, end the process, if not, return to step S410.

[0153] The model fine-tuning termination condition includes that each terminal node is removed from the online model fine-tuning node group or that the online model fine-tuning termination signaling is received from the network node.

[0154] It should be noted that the specific definitions of the above steps can refer to the specific definitions of the online model fine-tuning method based on edge intelligence described above.

[0155] It should be understood that, although the steps in the flowcharts involved in the embodiments described above are displayed in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowcharts involved in the embodiments described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be alternately executed with at least part of other steps or steps or stages in other steps.

[0156] Based on the same inventive concept, the embodiments of the present application also provide an edge intelligence-based online model fine-tuning device for implementing the above-mentioned edge intelligence-based online model fine-tuning method. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific definitions in one or more edge intelligence-based online model fine-tuning device embodiments provided below can refer to the definitions of the edge intelligence-based online model fine-tuning method described above, which will not be repeated here.

[0157] In one exemplary embodiment, as shown in Figure 5 An edge intelligence-based online model fine-tuning device is provided, applied to an edge node of a model fine-tuning system, the model fine-tuning system further comprising a terminal node and a network node; comprising:

[0158] The receiving module 510 is configured to 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 node group according to the at least one terminal node.

[0159] The broadcasting module 520 is configured to broadcast model fine-tuning feature signaling to each terminal node in the online model fine-tuning node group according to a model fine-tuning configuration file list; the model fine-tuning configuration file list comprises at least one model fine-tuning feature information generated according to model fine-tuning requirement information of each terminal node.

[0160] The obtaining module 530 is configured to obtain a subscription request signaling sent by any terminal node; the subscription request signaling is sent by any terminal node in a case where the model parameter configuration file to be subscribed is determined according to node running information in a current time period and model fine-tuning feature information carried by the model fine-tuning feature signaling.

[0161] The transmission module 540 is configured to transmit the subscribed model parameter configuration file to any of the terminal nodes, so that any of the terminal nodes performs online model fine-tuning on the pre-trained model based on the subscribed model parameter configuration file to obtain an updated model.

[0162] In one of the embodiments, the edge intelligence-based online model fine-tuning apparatus further includes a removal module. The removal module is specifically configured to acquire online model monitoring information sent by any of the terminal nodes. The online model monitoring information includes a model verification result of online verification of the updated model by any of the terminal nodes. In a case where the online model monitoring information indicates that the updated model meets fine-tuning requirement conditions, any of the terminal nodes is removed from the online model fine-tuning node group, and an updated online model fine-tuning node group is obtained.

[0163] In one of the embodiments, the model fine-tuning feature information includes at least one of an identifier of a model parameter configuration file, a computational complexity, a storage complexity, an implementation function, a task type, an expected performance gain, a fine-tuning delay, and a security level. The edge intelligence-based online model fine-tuning apparatus further includes an update module. The update module is specifically configured to update the subscribed model parameter configuration file according to the online model monitoring information to obtain an updated model parameter configuration file, and update the model fine-tuning feature information to obtain updated model fine-tuning feature information.

[0164] In one of the embodiments, the edge intelligence-based online model fine-tuning apparatus further includes a return module. The return module is specifically configured to return to the step of broadcasting model fine-tuning feature signaling to each of the terminal nodes in the online model fine-tuning node group according to the model fine-tuning configuration file list until a model fine-tuning termination condition is met. The model fine-tuning termination condition includes that each of the terminal nodes is removed from the online model fine-tuning node group or that online model fine-tuning termination signaling issued by a network node is received.

[0165] In one of the embodiments, the acquisition module 530 is specifically configured to generate model fine-tuning requirement information of a cell group according to model fine-tuning requirement information of each of the terminal nodes to which an edge node belongs, upload the model fine-tuning requirement information of the cell group to a network node, so that the network node generates a corresponding model fine-tuning configuration file list on the basis of a pre-trained model according to the model fine-tuning requirement information of the cell group, and acquires the model fine-tuning configuration file list sent by the network node.

[0166] The various modules in the online model fine-tuning device based on edge intelligence can be implemented by software, hardware, and combinations thereof, in whole or in part. The various modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a memory in a computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to the various modules.

[0167] In an example embodiment, a computer device, which can be a server, is provided, and an internal structure diagram of the computer device can be as shown in Figure 6 The computer device includes a processor, a memory, an input / output interface (I / O), and a communication interface. The processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store model data. The input / output interface of the computer device is configured to exchange information between the processor and external devices. The communication interface of the computer device is configured to communicate with terminals outside through a network connection. The computer program is executed by the processor to implement an online model fine-tuning method based on edge intelligence.

[0168] Those skilled in the art can understand that Figure 6 The structure shown in the above

[0169] In an example embodiment, a computer device is provided, including a memory and a processor, and the memory stores a computer program. The processor executes the computer program to implement the steps in the above method embodiments.

[0170] In an embodiment, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0171] In an embodiment, a computer program product is provided, and the computer program product includes a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0172] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0173] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments of each method. In the embodiments provided in the present application, any reference to memory, database or other medium 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 storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive 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. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., without being limited thereto.

[0174] Any technical features in the above embodiments can be combined, and for the sake of brevity, not all possible combinations are described above, however, any combination of these technical features is deemed to be within the scope of the present application.

[0175] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be pointed out that, for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. An edge intelligence based online model fine-tuning method, characterized in that, The method is applied to an edge node of a model fine-tuning system, the model fine-tuning system further comprising a terminal node and a network node; the method comprising: 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 node group according to the at least one terminal node; broadcasting model fine-tuning feature signaling to each terminal node in the online model fine-tuning node group according to a model fine-tuning configuration file list; the model fine-tuning configuration file list comprising at least one model fine-tuning feature information generated according to model fine-tuning requirement information of each terminal node; obtaining a subscription request signaling sent by any terminal node; the subscription request signaling being sent by any terminal node in a case where the terminal node determines to subscribe to a model parameter configuration file according to node running information in a current time period and model fine-tuning feature information carried by the model fine-tuning feature signaling; transmitting the subscribed model parameter configuration file to any terminal node, so that any terminal node performs online model fine-tuning on a pre-trained model based on the subscribed model parameter configuration file to obtain an updated model.

2. The method of claim 1, wherein, After the updated model is obtained, the method further comprises: obtaining online model monitoring information sent by any terminal node; the online model monitoring information comprising a model verification result of online verification of the updated model by any terminal node; in a case where the online model monitoring information indicates that the updated model meets fine-tuning requirement conditions, removing any terminal node from the online model fine-tuning node group to obtain an updated online model fine-tuning node group.

3. The method of claim 2, wherein, The model fine-tuning feature information comprises at least one of an identifier, a computational complexity, a storage complexity, an implementation function, a task type, an expected performance gain, a fine-tuning time delay, and a security level of a model parameter configuration file; After the updated online model fine-tuning node group is obtained, the method further comprises: updating the subscribed model parameter configuration file according to 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.

4. The method of claim 3, wherein, After the updated model fine-tuning feature information is obtained, 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 node group according to the model fine-tuning configuration file list until a model fine-tuning termination condition is met; wherein the model fine-tuning termination condition comprises that each terminal node is removed from the online model fine-tuning node group or that online model fine-tuning termination signaling is received from the network node.

5. The method of claim 1, wherein, Before the online model fine-tuning request uploaded by at least one terminal node to which the edge node belongs is received, the method further comprises: generating model fine-tuning requirement information of a cell group according to model fine-tuning requirement information of each terminal node to which the edge node belongs; uploading model fine-tuning requirement information of the cell group to a network node, so that the network node generates a corresponding model fine-tuning configuration file list based on a pre-trained model according to the model fine-tuning requirement information of the cell group; obtaining the model fine-tuning configuration file list sent by the network node.

6. An edge intelligence based online model fine-tuning system, characterized in that, The system is used to execute the steps of the edge intelligence-based online model fine-tuning method in any one of claims 1 to 5; the system comprises a network node, an edge node and a terminal node; The network node is configured to generate a corresponding model fine-tuning configuration file list based on a pre-trained model according to model fine-tuning requirement information of a cell group uploaded by the edge node; The model fine-tuning requirement information of the cell group is generated according to model fine-tuning requirement information of each terminal node to which the edge node belongs; The edge node is configured to obtain the model fine-tuning configuration file list sent by the network node; The terminal node is configured to send an online model fine-tuning request to the edge node; The edge node is configured to generate an online model fine-tuning node group according to at least one terminal node, and broadcast model fine-tuning feature signaling to each terminal node in the online model fine-tuning node group according to the model fine-tuning configuration file list; The terminal node is configured to determine a subscribed model parameter configuration file according to node running information in a current time period and 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 configured to transmit the subscribed model parameter configuration file to the terminal node; The terminal node is configured to perform online model fine-tuning on a pre-trained model based on the subscribed model parameter configuration file, to obtain an updated model.

7. An edge intelligence based online model fine-tuning apparatus, characterized by, An edge node applied to a model fine-tuning system, the model fine-tuning system further comprising a terminal node and a network node; the device comprises: A receiving module configured to 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 node group according to at least one terminal node; A broadcasting module configured to broadcast model fine-tuning feature signaling to each terminal node in the online model fine-tuning node group according to a model fine-tuning configuration file list; the model fine-tuning configuration file list comprises at least one model fine-tuning feature information generated according to model fine-tuning requirement information of each terminal node; An obtaining module configured to obtain a subscription request signaling sent by any terminal node; the subscription request signaling is sent by any terminal node in a case where the terminal node determines a subscribed model parameter configuration file according to node running information in a current time period and model fine-tuning feature information carried by the model fine-tuning feature signaling; A transmitting module configured to transmit the subscribed model parameter configuration file to any terminal node, so that any terminal node performs online model fine-tuning on a pre-trained model based on the subscribed model parameter configuration file, to obtain an updated model. 8.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-7. The processor executes the computer program to implement the steps of the method in any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, which when executed by a processor, implements the steps of the method of any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, The computer program, which when executed by a processor, implements the steps of the method of any one of claims 1 to 6.

Images