Model management method and device based on hierarchical federal learning and computer equipment

Abstract

The invention relates to a model management method and device based on hierarchical federal learning and computer equipment. The method comprises: receiving a first configuration message sent by a base station, the first configuration message comprising user scheduling result indication information based on hierarchical federal learning, the user scheduling result indication information being obtained by the base station based on user feature information of each terminal device; the user feature information comprises at least one of user data feature information and user channel state information. Reporting information is sent to the network, the network stores the hierarchical federal learning model and the issuing strategy generated by the base station, the reporting information at least comprises at least one of the terminal capability, the reasoning task type and the model performance monitoring feedback information, and the reporting information is used for the network to generate a second configuration message; the second configuration message is model reasoning strategy issuing information or model updating strategy issuing information based on hierarchical federated learning. And obtaining a model reasoning result or a model updating result based on the second configuration message.

Description

Technical Field

[0001] The present application relates to the technical field of federated learning, and in particular to a model management method, apparatus and computer equipment based on hierarchical federated learning. Background Art

[0002] With the rapid development of 3GPP radio access network (RAN) technology, the powerful representation and feature extraction capabilities of AI models can be combined with wireless communication scenarios to achieve time-frequency domain resource optimization and cost reduction. Among them, federated learning, as a distributed framework, can effectively improve the training efficiency of AI models, while reducing transmission overhead and protecting the privacy and security of local data. Therefore, how to realize the application of federated learning in 5G-A / 6G networks is a technical problem that needs to be solved urgently. Summary of the invention

[0003] Based on this, it is necessary to provide a model management method, device, computer equipment, computer-readable storage medium and computer program product based on hierarchical federated learning, which can utilize the cloud-edge-end collaborative network architecture to improve the coverage of the algorithm and enrich the applicable scenarios of the generated federated learning model, in response to the above-mentioned technical problems.

[0004] In a first aspect, the present application provides a model management method based on hierarchical federated learning, which is applied to a terminal device in a mobile communication system, wherein the mobile communication system also includes a network, wherein the network includes a network node and a plurality of base stations, each of the base stations corresponds to a plurality of terminal devices, wherein the terminal device is a distributed node for performing hierarchical federated learning, and the base station is a central node for performing hierarchical federated learning; the method includes:

[0005] receiving a first configuration message sent by the base station, where the first configuration message includes user scheduling result indication information based on hierarchical federated learning, where the user scheduling result indication information is obtained by the base station based on user feature information of each of the terminal devices, where the user feature information includes at least one of user data feature information and user channel state information;

[0006] Sending reporting information to the network, the network storing the hierarchical federated learning model and the delivery strategy generated by the base station, the reporting information including at least one of terminal capabilities, reasoning task types, and model performance monitoring feedback information, the reporting information is used by the network to generate a second configuration message, and the second configuration message is model reasoning strategy delivery information or model update strategy delivery information based on hierarchical federated learning;

[0007] The second configuration message is received, and a model inference result or a model update result is obtained based on the second configuration message.

[0008] In one of the embodiments, the terminal capability includes one or more of terminal computing capability, terminal channel status, and reasoning characteristics, the terminal computing capability is used to characterize the complexity of the model supported by the terminal device, and the terminal channel status characterizes the uplink channel status of the terminal device.

[0009] In one embodiment, sending the reporting information to the network includes:

[0010] Based on the target transmission mode, reporting information is sent to the base station so that the base station sends reporting information to the network node, and the target transmission mode includes at least one of system message, RRC signaling, MAC CE signaling, DCI signaling, and NAS message.

[0011] In one of the embodiments, the model performance monitoring feedback information includes one or more of a system performance index, an accuracy index, a complexity index, and a generalization performance index.

[0012] In one embodiment, the method further comprises:

[0013] If it is determined that the model performance monitoring feedback information and the preset indicator threshold meet the preset update conditions, a model update request is sent to the base station so that the network responds to the model update request, and the model update request includes one or more of the model to be updated, the function identifier to be updated, and the reference information.

[0014] In one embodiment, the receiving the second configuration message includes:

[0015] Receive a second configuration message sent by the base station based on a target transmission mode, where the second configuration message is sent by a network node to the base station, and the target transmission mode includes at least one of a system message, an RRC signaling, a MAC CE signaling, a DCI signaling, and a NAS message.

[0016] In one embodiment, the second configuration message is model update strategy delivery information, and the model update strategy delivery information includes the model to be updated and the update strategy; and obtaining the model inference result or the model update result based on the second configuration message includes:

[0017] The model to be updated is updated based on the update strategy to obtain a model update result, wherein the updated model includes the model to be updated / the model group to be updated, and the update strategy includes one or more of a local update strategy, a global update strategy, a model replacement strategy and a model deletion strategy.

[0018] In one embodiment, the process of performing the update process includes at least one of the following:

[0019] Update some modules of the update model to obtain the model update result;

[0020] Based on the target update algorithm, the model parameters of each of the models to be updated are updated to obtain a model update result, wherein the target update algorithm includes one or more of a momentum gradient descent algorithm, a stochastic gradient descent algorithm, and a mini-batch gradient descent algorithm;

[0021] Based on the new model parameters, the model parameters of each model to be updated are updated to obtain a model update result;

[0022] Delete the model parameters and model content / function identifiers of the updated model to obtain the model update result.

[0023] In one embodiment, the second configuration message is model reasoning strategy delivery information, and the model reasoning strategy delivery information includes a model identifier of the model to be reasoned, a model file, and identification information of a terminal device to perform model reasoning; and obtaining a model reasoning result or a model update result based on the second configuration message includes:

[0024] Based on the model identification of the model to be inferred, the model file and the identification information of the terminal device on which the model inference is to be executed, model inference in an online mode or model inference in an offline mode is performed to obtain a model inference result.

[0025] In a second aspect, an embodiment of the present application provides a model management method based on hierarchical federated learning, which is applied to a network node in a mobile communication system, wherein the network includes a network node and a plurality of base stations, each of the base stations corresponds to a plurality of terminal devices, and the method includes:

[0026] Receiving reporting information sent by the terminal device through the base station, the reporting information including at least one of terminal capability, reasoning task type, and model performance monitoring feedback information;

[0027] A second configuration message is generated based on the reported information, and the second configuration message is sent to the terminal device through the base station; the second configuration message is a model reasoning strategy based on hierarchical federated learning or a model update strategy, and the second configuration message is used to enable the terminal device to obtain a model reasoning result or a model update result.

[0028] In one of the embodiments, the terminal capability includes one or more of terminal computing capability, terminal channel status, and reasoning characteristics, the terminal computing capability is used to characterize the complexity of the model supported by the terminal device, and the terminal channel status characterizes the uplink channel status on the terminal device side.

[0029] In one of the embodiments, the reported information includes the terminal capability, and the receiving the reported information sent by the terminal device through the base station includes:

[0030] Based on the target transmission mode, the terminal capabilities reported by the terminal device through the base station are received, and the target transmission mode includes at least one of system messages, RRC signaling, MAC CE signaling, DCI signaling, and NAS messages.

[0031] In one embodiment, the method further comprises:

[0032] Receive the model inference request signaling sent by the terminal device through the base station, and return a response message corresponding to the inference request, so that the terminal device sends reporting information based on the response message.

[0033] In one embodiment, the second configuration message includes model reasoning strategy delivery information or model update strategy delivery information based on hierarchical federated learning; and generating the second configuration message based on the reported information includes:

[0034] Based on the reported information, determining at least one model to be inferred whose target dimension satisfies a preset matching condition among the multiple first models;

[0035] Based on the model identifier of the model to be inferred, the model file, and the identification information of the terminal device on which the model inference is to be executed, model inference strategy delivery information based on hierarchical federated learning is generated.

[0036] In one of the embodiments, determining at least one model to be inferred whose target dimension satisfies a preset matching condition among multiple first models based on the reported information includes:

[0037] Based on the reported information, a preset number of models to be inferred whose model structures, model complexity and model accuracy meet preset matching conditions are determined among the multiple first models.

[0038] In one of the embodiments, sending the second configuration message to the terminal device by the base station includes:

[0039] The second configuration message is sent to the base station through the user plane interface, so that the base station sends the second configuration message to the terminal device through the target transmission mode, and the target transmission mode includes at least one of system message, RRC signaling, MAC CE signaling, DCI signaling, and NAS message.

[0040] In one of the embodiments, the model performance monitoring feedback information includes one or more of a system performance index, an accuracy index, a complexity index, and a generalization performance index.

[0041] In one embodiment, the reported information includes model performance monitoring feedback information, and the generating the second configuration message based on the reported information includes:

[0042] Receive model performance monitoring feedback;

[0043] Receiving a model update request, the model update request including a to-be-updated model / to-be-updated function identifier and reference information, the model update request being generated when the model performance monitoring feedback information of the terminal device and a preset indicator threshold satisfy a preset update condition;

[0044] Respond to the model update request and generate a second configuration message.

[0045] In one embodiment, the reported information includes model performance monitoring feedback information, and the generating the second configuration message based on the reported information includes:

[0046] Receive model performance monitoring feedback;

[0047] Based on a preset time period, a model update event is triggered; or, if the model performance monitoring feedback information and the preset indicator threshold meet the preset update condition, an update event is triggered; the preset time period is associated with the terminal capability and the auxiliary information;

[0048] Based on the triggered update event, a second configuration message is generated.

[0049] In one embodiment, the second configuration message includes information about issuing a model reasoning strategy based on hierarchical federated learning; and generating the second configuration message includes:

[0050] determining at least one model to be updated;

[0051] Based on the model to be updated, the identification information of the terminal device on which the model update is to be executed, and the update strategy, model reasoning strategy delivery information based on hierarchical federated learning is generated.

[0052] In one of the embodiments, the update strategy includes one or more of a local update strategy, a global update strategy, a model replacement strategy, and a model deletion strategy.

[0053] In one embodiment, the method further comprises:

[0054] Receiving model data sent by each of the base stations, the model data including a model file and a model content / function identifier;

[0055] The first models are determined based on the model data, and are arranged according to evaluation indicators of the first models to obtain a first model list.

[0056] In one embodiment, the method further comprises:

[0057] An identification request and a model capability sent by a base station are received, wherein the identification request includes a model identification request and a function identification request, and the model capability is used to indicate the models / supported functions supported by the base station.

[0058] In one embodiment, the method further comprises:

[0059] An identification request is sent to each of the base stations so that each of the base stations reports a model capability, where the model capability is used to indicate a model / supported function supported by the base station.

[0060] In a third aspect, an embodiment of the present application provides a model management method based on hierarchical federated learning, which is applied to a base station in a mobile communication system, wherein the mobile communication system includes a network and multiple base stations, each of the base stations corresponds to multiple terminal devices, the terminal devices are distributed nodes for performing hierarchical federated learning, and the base stations are central nodes for performing hierarchical federated learning; the method includes:

[0061] Based on the user characteristic information of each terminal device, obtain the user scheduling result indication information corresponding to each terminal device, and send a first configuration message to each terminal device, wherein the user characteristic information includes at least one of user data characteristic information and user channel state information, and the first configuration message includes the user scheduling result indication information;

[0062] Receiving reporting information sent by the terminal device, the reporting information including at least one of terminal capability, reasoning task type, and model performance monitoring feedback information;

[0063] The reported information is sent to the network node so that the network node obtains a second configuration message; the second configuration message is a model reasoning strategy issuance information or a model update strategy issuance information based on hierarchical federated learning, and the second configuration message is used to enable the terminal device to obtain a model reasoning result or a model update result.

[0064] In one embodiment, the method further comprises:

[0065] Model data is sent to the network node, where the model data includes a model file and a model content / function identifier, so that the network node determines each first model based on each model data and obtains a first model list.

[0066] In one embodiment, the method further comprises:

[0067] An identification request and a model capability are sent to the network node, wherein the identification request includes a model identification request and a function identification request, and the model capability is used to indicate the model / supported function supported by the base station.

[0068] In one embodiment, the method further comprises:

[0069] An identification request sent by the network node is received, and a model capability is reported to the network node based on the identification request, where the model capability is used to indicate the model / supported function supported by the base station.

[0070] In one embodiment, the method further comprises:

[0071] Based on the model training results sent by each of the terminal devices, a trained global model is obtained.

[0072] In one embodiment, the terminal device is a terminal device of a scheduling user, and obtaining a trained global model based on the model training results sent by each of the terminal devices includes:

[0073] For the i-th round of training, based on the training results sent by each of the terminal devices in the i-1-th round, obtain the model parameters of the global model of the i-th round;

[0074] When the preset training completion conditions are not met, the model parameters of the global model of the i-th round are broadcast to each of the terminal devices, so that each of the terminal devices performs training based on the model parameters of the global model of the i-th round, until the base station determines that the preset training completion conditions are met, thereby obtaining a trained global model.

[0075] In one embodiment, the preset training completion conditions include one or more of the number of training rounds reaching a preset upper limit, the loss function of the global model being less than or equal to a preset value, and the difference between the loss function of the global model in two adjacent rounds being less than or equal to a preset value.

[0076] In one embodiment, the method further comprises:

[0077] Obtaining user scheduling result indication information corresponding to each of the terminals based on user characteristic information of each of the terminal devices, wherein the user characteristic information includes at least one of user data characteristic information and user channel state information;

[0078] A first configuration message is sent to each of the terminal devices respectively, wherein the first configuration message includes the user scheduling result indication information.

[0079] In one embodiment, obtaining the user scheduling result indication information corresponding to each terminal based on the user characteristic information of each terminal device includes:

[0080] Get the number of pre-configured scheduled users;

[0081] Among all users within the coverage of the base station, based on at least one of the user data characteristic information and user channel state information of each terminal device, the terminal devices with the scheduled number of users are screened out, and the user scheduling result indication information is determined as the scheduled indication information.

[0082] In a fourth aspect, an embodiment of the present application provides a model reasoning device based on hierarchical federated learning, which is applied to a terminal device in a mobile communication system, wherein the mobile communication system also includes a network, wherein the network includes a network node and a plurality of base stations, each of the base stations corresponds to a plurality of terminal devices, wherein the terminal device is a distributed node for performing hierarchical federated learning, and the base station is a central node for performing hierarchical federated learning; the device includes:

[0083] A first receiving module is configured to receive a first configuration message sent by the base station, where the first configuration message includes user scheduling result indication information based on hierarchical federated learning, where the user scheduling result indication information is obtained by the base station based on user feature information of each of the terminal devices, where the user feature information includes at least one of user data feature information and user channel state information;

[0084] A first sending module, configured to send reporting information to the network, wherein the network stores the hierarchical federated learning model and the issuing strategy generated by the base station, wherein the reporting information includes at least one of terminal capabilities, reasoning task types, and model performance monitoring feedback information, and the reporting information is used by the network to generate a second configuration message, wherein the second configuration message is model reasoning strategy issuing information or model update strategy issuing information based on hierarchical federated learning;

[0085] The second receiving module is used to receive the second configuration message and obtain a model inference result or a model update result based on the second configuration message.

[0086] In a fifth aspect, an embodiment of the present application provides a model management device based on hierarchical federated learning, which is applied to a network node in a mobile communication system, wherein the network includes a network node and multiple base stations, each of the base stations corresponds to multiple terminal devices, and the device includes:

[0087] A third receiving module is used to receive the reporting information sent by the terminal device through the base station, where the reporting information includes at least one of terminal capability, reasoning task type, and model performance monitoring feedback information;

[0088] The first generation module is used to generate a second configuration message based on the reported information, and send the second configuration message to the terminal device through the base station; the second configuration message is a model reasoning strategy based on hierarchical federated learning or a model update strategy, and the second configuration message is used to enable the terminal device to obtain a model reasoning result or a model update result.

[0089] In a sixth aspect, an embodiment of the present application provides a model management device based on hierarchical federated learning, which is applied to a base station in a mobile communication system, wherein the mobile communication system includes a network and multiple base stations, each of the base stations corresponds to multiple terminal devices, the terminal devices are distributed nodes for performing hierarchical federated learning, and the base stations are central nodes for performing hierarchical federated learning; the device includes:

[0090] A second sending module is used to obtain user scheduling result indication information corresponding to each terminal device based on the user characteristic information of each terminal device, and send a first configuration message to each terminal device, wherein the user characteristic information includes at least one of user data characteristic information and user channel state information, and the first configuration message includes the user scheduling result indication information;

[0091] A fourth receiving module, configured to receive reporting information sent by the terminal device, wherein the reporting information includes at least one of terminal capability, reasoning task type, and model performance monitoring feedback information;

[0092] The third sending module is used to send the reported information to the network node so that the network node obtains a second configuration message; the second configuration message is a model reasoning strategy based on hierarchical federated learning or a model update strategy. The second configuration message is used to enable the terminal device to obtain the model reasoning result or the model update result.

[0093] In a seventh aspect, the present application further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps described in the above embodiments when executing the computer program.

[0094] In an eighth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and the computer program implements the steps described in the above embodiments when executed by a processor.

[0095] In a ninth aspect, the present application also provides a computer program product, including a computer program, which implements the steps described in the above embodiments when executed by a processor.

[0096] The above-mentioned model management method, device, and computer equipment based on hierarchical federated learning, wherein the method includes: receiving a first configuration message sent by a base station, the first configuration message includes user scheduling result indication information based on hierarchical federated learning, the user scheduling result indication information is obtained by the base station based on user feature information of each terminal device, and the user feature information includes at least one of user data feature information and user channel state information; sending reporting information to the network, the network stores the hierarchical federated learning model and the issuing strategy generated by the base station, the reporting information includes at least one of the terminal capability, the reasoning task type, and the model performance monitoring feedback information, and the reporting information is used for the network to generate a second configuration message, the second configuration message is the model reasoning strategy issuing information or the model update strategy issuing information based on hierarchical federated learning; receiving the second configuration message, and obtaining the model reasoning result or the model update result based on the second configuration message. By adopting this method, data interaction between the base station and the terminal can be realized, the management efficiency and generalization performance in the wireless intelligent federated learning model can be improved, the additional consumption of communication resources and the computing overhead can be reduced, the resources to be allocated can be flexibly configured, and the terminal devices for federated learning can be flexibly scheduled, and the applicable scenarios of the model can be further expanded. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0098] Figure 1 is an application environment diagram of a model management method based on hierarchical federated learning in one embodiment;

[0099] Figure 2 A flowchart of a model management method based on hierarchical federated learning in one embodiment;

[0100] Figure 3 A schematic diagram of a flow chart of steps for obtaining a model update result in one embodiment;

[0101] Figure 4 A schematic diagram of a sending step in an embodiment;

[0102] Figure 5 A schematic diagram of a flow chart of a step of generating a second configuration message in an embodiment;

[0103] Figure 6 A schematic diagram of a flow chart of a step of generating a second configuration message in an embodiment;

[0104] Figure 7 A schematic diagram of a flow chart of a step of generating a second configuration message in an embodiment;

[0105] Figure 8 A schematic diagram of a flow chart of steps for obtaining a model list in one embodiment;

[0106] Fig. 9 A schematic diagram of a process of reporting information transmission steps in an embodiment;

[0107] Fig.10 is a flowchart of a training step in an embodiment;

[0108] Fig.11a is a flow chart of a model management method based on hierarchical federated learning in another embodiment;

[0109] Fig.11b A schematic diagram of a curve showing the positioning accuracy of a solution in an embodiment;

[0110] Fig.11c A schematic diagram of a positioning error change trend of a federated learning algorithm in one embodiment;

[0111] Fig.11d Schematic diagram of a CDF curve of positioning error in one embodiment;

[0112] Fig.12 is a structural block diagram of a model reasoning device based on hierarchical federated learning in one embodiment;

[0113] Fig.13 is a structural block diagram of a model reasoning device based on hierarchical federated learning in one embodiment;

[0114] Fig.14 is a structural block diagram of a model reasoning device based on hierarchical federated learning in one embodiment;

[0115] Fig.15 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0116] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0117] The model management method based on hierarchical federated learning provided in the embodiment of the present application can be applied to Figure 1In the mobile communication system in the application environment shown. Among them, the mobile communication system includes a network, the network includes a network node and a plurality of base stations, each base station corresponds to a plurality of terminal devices, each terminal device is a distributed node for performing hierarchical federated learning, and each base station is a central node for performing hierarchical federated learning; the network node can communicate with each base station, and the base station can communicate with the terminal device corresponding to the base station. The mobile communication system can be a distributed network architecture, the UE (User Equipment, user equipment / terminal equipment) can be a terminal node, a terminal device or a local client, the base station can be an edge node or a server side, specifically, for example, it can be a CU / BBU under a gNB / CU-DU separation architecture, and the network node can be a cloud, specifically, for example, it can be LMF / OAM (Orchestration and Management, orchestration and management) / network data analysis function NWDAF (NW side). Among them, the terminal device can be, but is not limited to, various personal computers, laptops, smart phones, tablets, etc.

[0118] In an exemplary embodiment, Figure 2 As shown in the figure, a model management method based on hierarchical federated learning is provided. Figure 1 The method is described as follows: a terminal device in a mobile communication system in the example, the mobile communication system also includes a network, the network includes a network node and a plurality of base stations, each base station corresponds to a plurality of terminal devices, the terminal device is a distributed node for performing hierarchical federated learning, and the base station is a central node for performing hierarchical federated learning; wherein the method includes:

[0119] Step 202: Receive a first configuration message sent by a base station.

[0120] The first configuration message includes user scheduling result indication information based on hierarchical federated learning, and the user scheduling result indication information is obtained by the base station based on user characteristic information of each terminal device, and the user characteristic information includes at least one of user data characteristic information and user channel state information. The user data characteristic information may be characteristic information of a local update gradient of the terminal device, and the user channel state information may be information obtained based on user CSI-RS (Channel State Information Reference Signal, Channel State Information Reference Signal) channel estimation feedback.

[0121] Specifically, the base station can determine all terminal devices within the coverage of the base station, and screen among all the terminal devices based on hierarchical federated learning to determine the selected terminal devices, and obtain a first configuration message based on the user scheduling indication information. The base station can send the first configuration message to each selected terminal device, and the user scheduling result indication information contained in the first configuration message can indicate that the terminal device is selected by the base station.

[0122] In one example, the specific process of scheduling by the base station may include: the base station determines the first number of terminal devices that currently need to be scheduled (used), and screens all terminal devices within the coverage of the base station based on the user feature information of each terminal device, extracts the first number of terminal devices, and determines that the user scheduling result indication information of the first number of terminal devices is the first indication information, and the first indication information can represent that the terminal device is selected by the base station; the user scheduling indication information of other terminal devices can be empty or the second indication information, and the second indication information is used to represent that the terminal device is not selected by the base station. The present disclosure does not limit the specific form of the first indication information and the second indication information, and those skilled in the art can make specific determinations based on actual application scenarios.

[0123] Step 204: Send reporting information to the network.

[0124] Among them, the network stores the hierarchical federated learning model and delivery strategy generated by the base station, and the reported information includes at least one of the terminal capabilities, the reasoning task type, and the model performance monitoring feedback information. The reported information is used for the network to generate a second configuration message, and the second configuration message is the model reasoning strategy delivery information or the model update strategy delivery information based on hierarchical federated learning.

[0125] Specifically, the terminal device can send reporting information to the network. After receiving the reporting information, the network can generate a second configuration message based on the reporting information. Based on this, the network can send the second configuration message to each terminal device. In the model reasoning process, the network sends a second configuration message to each terminal, and the second configuration message may include model reasoning strategy delivery information based on hierarchical federated learning. In the model update process, the network sends a second configuration message to each terminal device, and the second configuration message may include model update strategy delivery information based on hierarchical federated learning.

[0126] Step 206: Receive a second configuration message, and obtain a model inference result or a model update result based on the second configuration message.

[0127] Specifically, the terminal device can receive the second configuration message, and perform model reasoning based on the model reasoning strategy in the second configuration message to obtain the model reasoning result; the terminal device can also send information based on the model update strategy in the second configuration message to update the model and obtain the model update result.

[0128] In the above-mentioned model management method based on hierarchical federated learning, a first configuration message sent by a base station is received, the first configuration message includes user scheduling result indication information based on hierarchical federated learning, the user scheduling result indication information is obtained by the base station based on user feature information of each terminal device, and the user feature information includes at least one of user data feature information and user channel state information. Reporting information is sent to the network, and the network stores the hierarchical federated learning model and the delivery strategy generated by the base station. The reporting information includes at least one of the terminal capability, the reasoning task type, and the model performance monitoring feedback information. The reporting information is used for the network to generate a second configuration message, and the second configuration message is the model reasoning strategy delivery information or the model update strategy delivery information based on hierarchical federated learning. The second configuration message is received, and the model reasoning result or the model update result is obtained based on the second configuration message. By adopting this method, data interaction between the base station and the terminal can be realized, the management efficiency and generalization performance in the wireless intelligent federated learning model can be improved, the additional consumption of communication resources and the computing overhead can be reduced, the resources to be allocated can be flexibly configured, and the terminal devices for federated learning can be flexibly scheduled, and the applicable scenarios of the model can be further expanded.

[0129] In one of the embodiments, the terminal capability includes one or more of terminal computing capability, terminal channel status, and reasoning characteristics. The terminal computing capability is used to characterize the complexity of the model supported by the terminal device, and the terminal channel status characterizes the uplink channel status of the terminal device.

[0130] Specifically, the terminal computing capability is an indicator that characterizes the strength of the local computing capability of the terminal device. The terminal computing capability can be used to determine the model complexity supported by the terminal device. The terminal computing capability of the terminal device can be positively correlated with the model complexity supported by the terminal device, that is, the stronger the terminal computing capability, the higher the model complexity that can be supported; the terminal computing capability may include the main frequency, number of threads, number of cores of the terminal device CPU, the bit width, memory size, etc. of the GPU; the terminal channel state is used to characterize the uplink channel state (CSI); the inference feature can be an intelligent feature used to characterize the local data of the terminal device, for example, the inference feature may include the number of data set features, number of samples, dimension, data type, learning rate, data batch and other hyperparameters; the data type is used to characterize whether it is a time series; the terminal computing capability and inference feature can be optional items for the terminal device to send reporting information to the network, and the terminal device can determine whether the reporting information includes the terminal computing capability and inference feature based on the needs of the actual application scenario; or it can be determined based on the configuration of the network side whether the reporting information includes the terminal computing capability and inference feature.

[0131] Optionally, the reported information may include terminal capabilities, and may also include reasoning task types, auxiliary information, etc.; the reasoning task type may characterize the type of model reasoning task, that is, the model type or the type of model reasoning result, which may include one or more of classification tasks, prediction tasks, and clustering tasks, and the corresponding models used for model reasoning may include one or more of classification models, prediction models, and clustering models.

[0132] In this embodiment, the terminal capabilities carried in the reported information may include multiple types, and the specific content of the terminal capabilities may be determined based on the actual needs of the terminal device, thereby ensuring the flexibility of the information represented by the terminal capabilities.

[0133] In one embodiment, the specific implementation process of the step of "sending reporting information to the network" may include:

[0134] Based on the target transmission mode, the reporting information is sent to the base station, so that the base station sends the reporting information to the network node.

[0135] The target transmission mode includes at least one of system message, RRC signaling, MAC CE signaling, DCI signaling, and NAS message.

[0136] Specifically, the reporting information sent by the terminal device to the network can be terminal device capability reporting signaling. The terminal device can report the terminal device capability reporting signaling to the base station (edge ​​node) through at least one transmission method including system messages, RRC (Radio Resource Control) signaling, MAC CE signaling (MAC Control Element), DCI signaling (Downlink Control Information), and NAS (Non Access Stratum) messages, that is, report to the edge node through at least one transmission method including system messages, RRC signaling, MACCE signaling, DCI signaling, and NAS messages. After receiving the reporting information, the edge node can send the reporting information to the network node corresponding to the edge node. For example, the edge node can send the reporting information to the network node through the NG-C interface.

[0137] In this embodiment, the terminal device can transmit data to the base station through multiple transmission modes, thereby ensuring the diversity and reliability of data transmission.

[0138] In one embodiment, the model performance monitoring feedback information includes one or more of a system performance index, an accuracy index, a complexity index, and a generalization performance index.

[0139] Specifically, the model performance monitoring feedback information may include multiple indicators. System performance indicators may include one or more of throughput, block error rate, and latency. Accuracy indicators may include one or more of the accuracy of the model output, MSE / NMSE / cosine similarity between the model input and output, etc. Complexity indicators may include model size and computational complexity. The model size refers to the number of parameters contained in the model, and the computational complexity may be FLOPs (Floating Point Operations Per second). The model performance monitoring feedback information may also include generalized performance indicators, and the specific content may be determined based on the actual application scenario.

[0140] Optionally, after the UE performs model inference, it can monitor the model performance in real time, obtain the model performance monitoring result, and obtain the model performance monitoring feedback information based on the result. The UE can feedback the model performance monitoring feedback information to the edge node corresponding to the terminal device, and the edge node can send the model performance monitoring feedback information to the network node (NW side) corresponding to the edge node after receiving the model performance monitoring feedback information.

[0141] In this embodiment, by monitoring the performance indicators during the model reasoning process, the model performance can be monitored in real time and adjusted in time to ensure the timeliness of the model.

[0142] In one embodiment, the model management method based on hierarchical federated learning further includes:

[0143] If it is determined that the model performance monitoring feedback information and the preset indicator threshold meet the preset update condition, a model update request is sent to the base station so that the network responds to the model update request.

[0144] Among them, the model update request includes one or more of the model to be updated, the function identifier to be updated, and the reference information; the preset update condition is used to determine whether to trigger a model update event; if it is determined that the preset update condition is met, the update event is triggered; if it is determined that the preset update condition is not met at present, the update event will not be triggered.

[0145] Specifically, the terminal device can obtain model performance monitoring feedback information based on the model performance monitoring results, for example, the model performance monitoring results can be used as model performance monitoring feedback information; the terminal device can determine the model performance monitoring results as reporting information, and send the reporting information to the base station to which the terminal device belongs, and the base station can send the reporting information to the network node corresponding to the base station.

[0146] After the base station sends the reporting information to the network node corresponding to the base station, a model update process may be triggered. In one example, the triggering process of the model update may be actively triggered by the terminal device. The terminal device may compare the indicator in the model performance monitoring feedback information with the preset indicator threshold corresponding to the indicator. If it is determined that there is an indicator that meets the preset update condition, the terminal device may trigger a model update event, that is, the terminal device may generate a model update request and send the model update request to the base station corresponding to the terminal device; after receiving the model update request, the base station may report the model update request to the network node, and the network node may respond to the model update request. The response process may include: the network node determines at least one model to be updated for model update to be executed, generates a model update policy issuance information based on the at least one model to be updated, that is, obtains a second configuration message, the network node may send the model update policy issuance information to the terminal device to be updated through the base station, and the terminal device may perform the model update based on the model update policy issuance information to obtain the model update result. The specific update process will be described in detail in the following embodiments.

[0147] In another example, the triggering process of the model update can be actively triggered by the network side. After receiving the reported information, the network node can trigger the model update event based on the target trigger condition. The specific process of the target trigger condition may include periodic model update and non-periodic model update. One possible implementation scenario may be that the network node triggers the update event of the model to be updated / the model group to be updated according to a preset time period, and generates a second configuration message, wherein the preset time period may be associated with the terminal capability, or may be associated with the change of auxiliary information, for example, the auxiliary information may be a scene, a site, a channel, or a data sampling speed, etc. Another possible triggering method may be that the network node triggers the model update event based on the received model performance monitoring feedback information, specifically, the network node may compare the indicator contained in the model performance monitoring feedback information with the preset indicator threshold corresponding to the indicator, and if it is determined that the indicator and the preset indicator threshold meet the preset update condition, the model update event is triggered.

[0148] Optionally, the specific process of the network node / UE determining whether the preset update condition is met may be that, when the indicator included in the model performance monitoring feedback information is delay, if the network node / UE determines that the delay is greater than or equal to a preset delay threshold corresponding to the delay, the network node / UE may determine that the preset update condition is currently met, triggering a model update event; when the indicator included in the model performance monitoring feedback information is model accuracy, if the network node / UE determines that the model accuracy is less than or equal to the preset delay threshold corresponding to the delay, the network node / UE may determine that the preset update condition is currently met, triggering a model update event; the network node / UE may determine whether the preset update condition is met based on the meaning of the indicator and the relationship between the indicator and the preset indicator threshold corresponding to the indicator.

[0149] In one example, the terminal device may report model performance monitoring feedback information to the base station to which the terminal device belongs. After receiving the model performance monitoring feedback information, the base station may send the model performance monitoring feedback information to the network node (NW side) corresponding to the base station. The model update process is triggered by the terminal device side or the network side.

[0150] In one example, a network node may configure a preset indicator threshold through a target transmission mode, and the target transmission mode includes at least one of a system message, RRC signaling, MAC CE signaling, DCI signaling, and a NAS message; the network node may send an indicator threshold configuration message to a terminal device through a base station, and the indicator threshold configuration message includes at least a specific value of the indicator threshold and an indicator identifier; the indicator threshold configuration message may be at least one of a system message, RRC signaling, MAC CE signaling, DCI signaling, and a NAS message.

[0151] Optionally, the model update request sent by the terminal device specifically includes the model to be updated, the function identifier to be updated, and reference information. The reference information may specifically be the model update reference information fed back by the terminal device, such as one or more of gradient reference information, loss function reference information, and logs.

[0152] Optionally, the base station to which the terminal device belongs may be within the coverage of the base station. The base station serves as the central node of hierarchical federated learning, and each terminal device within the coverage of the base station serves as a distributed node of hierarchical federated learning. The network node corresponding to the base station may be a network node for managing each base station.

[0153] In this embodiment, the terminal device / network node can actively determine whether a model update process is needed based on the monitored model indicator information, thereby improving the control of the terminal device / network node over the model update process and being able to update the model in a timely manner when the model does not meet the requirements.

[0154] In one embodiment, the specific implementation process of the step of "receiving the second configuration message" may include:

[0155] A second configuration message sent by the base station based on the target transmission mode is received.

[0156] The second configuration message is sent by the network node to the base station, and the target transmission mode includes at least one of a system message, an RRC signaling, a MAC CE signaling, a DCI signaling, and a NAS message.

[0157] Specifically, after receiving the reporting information sent by the terminal device, the network node can process the reporting information to obtain a second configuration message and send the second configuration message to the base station. After receiving the second configuration message, the base station can send the second configuration message to each terminal device corresponding to the base station.

[0158] In one example, a network node generates a second configuration message based on the reported information received from the terminal device, and the network node may send the second configuration message to the base station to which the terminal device belongs, so that the base station sends the second configuration message to the terminal device; in another example, the network node may send the second configuration message to each base station managed by the network node, so that each base station may send the second configuration message to each terminal device within the range of the base station; in another example, the network node may screen each base station managed by the network node, and send the second configuration message to the screened base stations, and each base station may schedule the terminal device based on the method described in the above embodiment to obtain the scheduled terminal device, and send the second configuration message to each scheduled terminal device, wherein each base station may also send the second configuration message to the terminal device corresponding to the identifier of the terminal device carried in the second configuration message.

[0159] In this embodiment, data transmission can be achieved between network nodes, base stations and terminal devices through multiple transmission modes, ensuring the reliability, diversity and flexibility of data transmission.

[0160] In one embodiment, the second configuration message is model update policy delivery information, and the model update policy delivery information includes the model to be updated and the update policy. Specifically, in the model update process, the second configuration message generated by the network node may include the model update policy delivery information; in the model reasoning process summary, the second configuration message generated by the network node may include the model reasoning policy delivery information.

[0161] In one example, the second configuration message is the model update policy delivery information; the process of the network node generating the second configuration message may be that the network node determines at least one model to be updated based on the reported information, encapsulates the model to be updated, the user identifier of the terminal device to be updated, and the update policy to obtain the model update policy delivery information; the network node may send the model update policy delivery information to the base station through the user plane interface, and the base station may send the second configuration message to the terminal device corresponding to the user identifier based on the user identifier in the received model update policy delivery information. The user plane interface may be an NG-U interface. The model to be updated determined by the network node may be a model to be updated or a group of models to be updated.

[0162] Accordingly, the specific implementation process of the step of “obtaining a model inference result or a model update result based on the second configuration message” may include:

[0163] The model to be updated is updated based on the update strategy to obtain a model update result.

[0164] Among them, the update model includes the model to be updated / model group to be updated, the update strategy includes one or more of the local update strategy, the global update strategy, the model replacement strategy and the model deletion strategy, and the update strategy can be the model update method used by the terminal device when updating the model.

[0165] Specifically, after receiving the second configuration message, the terminal device can obtain the model update policy delivery information carried in the second configuration message, and update the model to be updated based on the update policy in the model update policy delivery information to obtain a model update result, which can be the model to be updated after the update process is executed. In other words, the terminal device currently receiving the second configuration message is a terminal device selected by the base station based on the user identifier of the model to be updated carried in the second configuration message, and the terminal device is a terminal device to be updated. The terminal device can execute the model update process based on the user device to be updated based on the model to be updated / model group to be updated in the model update policy delivery information and the update policy indicated by the delivery information to obtain the model update result, and the obtained model update result can be the updated model / model group obtained by the terminal device after executing the model update method selected by the network node. The model update policy delivery information can be a model update policy delivery signaling.

[0166] In this embodiment, the base station may send model update policy delivery information to a specific terminal device based on an instruction of a network node, so that the selected terminal device performs model update processing.

[0167] In one embodiment, if Figure 3 As shown, the specific implementation process of the step "performing update processing" may include at least one of the following:

[0168] Step 302, update some modules of the model to be updated to obtain a model update result.

[0169] Among them, some models can be specific modules or modules pre-selected by network nodes.

[0170] Specifically, when the update strategy includes a local update strategy, the terminal device can determine the model to be updated / model group to be updated based on the information sent by the model update strategy, and determine some modules in the model to be updated / model group to be updated, and update some modules in the model to be updated / model group to be updated based on the collected data to obtain an updated model / model group; the terminal device can also update the model parameters of some modules in the model to be updated / model group to be updated to obtain an updated model / model group, and determine the updated model / model group as the model update result.

[0171] Step 304: based on the target update algorithm, update the model parameters of each model to be updated to obtain a model update result.

[0172] Among them, the target update algorithm includes one or more of a momentum gradient descent algorithm, a stochastic gradient descent algorithm, and a mini-batch gradient descent algorithm.

[0173] Specifically, the terminal device can send information from the model update strategy to determine the model to be updated / the model group to be updated, and determine the global update strategy and the target update algorithm contained in the update strategy, and update the collected data and all model parameters of the model to be updated / the model group to be updated through the target update algorithm to obtain the updated model / model group. Based on this, the terminal device can determine that the updated model / model group is the model update result. The target update algorithm contained in the update strategy includes at least one of: momentum gradient descent algorithm, stochastic gradient descent algorithm, and small batch gradient descent algorithm.

[0174] Step 306: Update the model parameters of each model to be updated based on the new model parameters to obtain a model update result.

[0175] Specifically, the terminal device can send information based on the model update policy to determine the model to be updated / the model group to be updated, and determine the model replacement policy included in the update policy; the terminal device can determine new model parameters based on the model replacement policy, and replace the model parameters in the model to be updated / the model group to be updated based on the new model parameters, and replace the model parameters of the model to be updated / the model group to be updated with the new model parameters to obtain an updated model / model group. Based on this, the terminal device can determine that the updated model / model group is the model update result.

[0176] Step 308, delete the model parameters and model content / function identifiers of the updated model to obtain the model update result.

[0177] Specifically, if the terminal device determines that the update strategy is a model deletion strategy, the terminal device can determine the model to be updated / model group to be updated, delete the model parameters of the model to be updated / model group to be updated, and delete the model content / function identifier corresponding to the model to be updated / model group to be updated to obtain the model update result.

[0178] It should be noted that the present disclosure does not limit the execution order and execution scope of the above steps 302, 304, 306 and 308. Those skilled in the art can determine the specific execution steps based on the specific content of the update strategy included in the second configuration message.

[0179] In this embodiment, the network node can configure the specific process of executing the model update on the terminal device by transmitting the second configuration message, thereby ensuring the degree of information synchronization between the network node and the terminal device.

[0180] In one embodiment, the second configuration message is model reasoning strategy delivery information, and the model reasoning strategy delivery information includes the model identifier of the model to be reasoned, the model file, and the identification information of the terminal device to perform model reasoning. Specifically, during the model reasoning process, the second configuration message generated by the network node may include the model reasoning strategy delivery information.

[0181] In one example, the second configuration message is model inference strategy delivery information; the process of the network node generating the second configuration message may be that the terminal device to be executed model inference can send a model inference request signaling to the network node through the base station, the network node can respond to the model inference request signaling, generate a model inference response message, and send the model inference response message to the terminal device to be executed model inference through the base station; the terminal device receiving the model inference response message can send reporting information to the base station (edge ​​node) to which the terminal device belongs, and the base station can send the reporting information to the network node through the NG-C interface. After receiving the reporting information, the network node can determine, based on the reporting information, at least one model to be inferred whose target dimension meets the preset matching conditions among multiple first models stored in the network node, and based on the model identifier of the model to be inferred, the model file and the identification information of the terminal device to be executed model inference, generate model inference strategy delivery information based on hierarchical federated learning, that is, generate the second configuration message, and transmit it to the edge node through the NG-U interface, and transmit it to the terminal device to be executed model inference by the edge node.

[0182] Among them, the process of the network node determining the model to be inferred / the group of models to be inferred may be that the multiple models stored locally by the network node may be multiple first models, among which the network node may be based on the reported information and the pre-configured preset number, by maximizing the model accuracy, minimizing the model complexity, or other dimensions for matching, to obtain the preset number of matching models to be inferred / the group of models to be inferred, which will be described in detail in the following embodiments and will not be repeated here. The first model may be a model that has completed model identification, or a model that has completed function identification.

[0183] Accordingly, the specific implementation process of the step of “obtaining a model inference result or a model update result based on the second configuration message” may include:

[0184] Based on the model identification of the model to be inferred, the model file and the identification information of the terminal device on which the model inference is to be executed, the model inference in the online mode or the model inference in the offline mode is performed to obtain the model inference result.

[0185] The model identifier of the model to be inferred may be an identifier of a list of models to be inferred, for example, a list of models to be inferred / a list of models to be inferred.

[0186] Specifically, the network node can encapsulate the model to be inferred / the group of models to be inferred, the model files of each model to be inferred, and the identification information of the terminal device to be executed model inference, and obtain the model inference strategy delivery information. The network node can transmit the model inference strategy delivery information as the second configuration message to the base station through the NG-U interface. The base station can transmit the second configuration message to each terminal device to be executed model inference based on the identification of the terminal device to be executed model inference in the second configuration message. After the terminal device receives the second configuration message, it can

[0187] In one example, the base station may transmit the second configuration message to each terminal device to be executed model reasoning through a target transmission mode, and the base station may transmit the second configuration message to the terminal device to be executed model reasoning through at least one message transmission mode of system message, RRC signaling, MAC CE signaling, DCI signaling, and NAS message. The form of model transmission may include one or more of the following: sending in a private format, sending in a public format, and the user device knows the model structure, and sending in a public format, and the user device does not know the model structure.

[0188] Based on this, the terminal device can perform model reasoning based on the information sent by the model reasoning strategy to obtain the model reasoning result. For application scenarios where the delay requirement is greater than the first delay threshold, the terminal device can adopt an online model reasoning strategy. After receiving the information sent by the model reasoning strategy, the terminal device collects real-time data, and performs real-time model reasoning based on the collected real-time data and the second configuration message to obtain online model reasoning results; for application scenarios where the delay requirement is less than the second delay threshold, the terminal device can obtain data stored in advance locally on the terminal device, and perform model reasoning based on the data and the second configuration message to obtain offline model reasoning results.

[0189] In this embodiment, the terminal device can perform real-time model reasoning online based on the configuration requirements of the network nodes or perform offline model reasoning based on pre-collected data, thereby expanding the scenarios in which model reasoning is applicable.

[0190] In an exemplary embodiment, Figure 4 As shown in the figure, a model management method based on hierarchical federated learning is provided for application Figure 1The network node in the mobile communication system includes a network node and multiple base stations, each base station corresponds to multiple terminal devices, and the terminal device corresponding to each base station can include multiple terminal devices within the communication coverage of each base station. The method includes:

[0191] Step 402: Receive reporting information sent by the terminal device through the base station.

[0192] The reported information includes at least one of terminal capabilities, reasoning task types, and model performance monitoring feedback information.

[0193] Specifically, the terminal device may send reporting information to the base station to which the terminal device belongs, and the base station may send the received reporting information to the network node.

[0194] Step 404: Generate a second configuration message based on the reported information, and send the second configuration message to the terminal device through the base station.

[0195] Among them, the second configuration message is the model reasoning strategy delivery information or the model update strategy delivery information based on hierarchical federated learning, and the second configuration message is used to enable the terminal device to obtain the model reasoning result or the model update result.

[0196] Specifically, based on this, the network node can send the second configuration message to each terminal device. During the model reasoning process, the network node sends a second configuration message to each terminal device, and the second configuration message may include model reasoning strategy delivery information based on hierarchical federated learning. During the model update process, the network node sends a second configuration message to each terminal device, and the second configuration message may include model update strategy delivery information based on hierarchical federated learning. The terminal device can perform model reasoning based on the model reasoning strategy delivery information in the received second configuration message to obtain a model reasoning result; the terminal device can also update the model based on the model update strategy delivery information in the second configuration message to obtain a model update result.

[0197] In this embodiment, data interaction between base stations and terminals can be realized, the management efficiency and generalization performance in the wireless intelligent federated learning model can be improved, the additional consumption of communication resources and computing overhead can be reduced, the resources to be allocated can be flexibly configured, and the terminal devices for federated learning can be flexibly scheduled, further expanding the applicable scenarios of the model.

[0198] In one of the embodiments, the terminal capability includes one or more of terminal computing capability, terminal channel status, and reasoning characteristics. The terminal computing capability is used to characterize the complexity of the model supported by the terminal device, and the terminal channel status characterizes the uplink channel status on the terminal device side.

[0199] Specifically, the terminal computing capability is an indicator that characterizes the strength of the local computing capability of the terminal device. The terminal computing capability can be used to determine the model complexity supported by the terminal device. The terminal computing capability of the terminal device can be positively correlated with the model complexity supported by the terminal device, that is, the stronger the terminal computing capability, the higher the model complexity that can be supported; the terminal computing capability may include the main frequency, number of threads, number of cores of the terminal device CPU, the bit width, memory size, etc. of the GPU; the terminal channel state is used to characterize the uplink channel state (CSI); the inference feature can be an intelligent feature used to characterize the local data of the terminal device, for example, the inference feature may include the number of data set features, number of samples, dimension, data type, learning rate, data batch and other hyperparameters; the data type is used to characterize whether it is a time series; the terminal computing capability and inference feature can be optional items for the terminal device to send reporting information to the network, and the terminal device can determine whether the reporting information includes the terminal computing capability and inference feature based on the needs of the actual application scenario; or it can be determined based on the configuration of the network side whether the reporting information includes the terminal computing capability and inference feature.

[0200] Optionally, the reported information may include terminal capabilities, and may also include reasoning task types, auxiliary information, etc.; the reasoning task type may characterize the type of model reasoning task, that is, the model type or the type of model reasoning result, which may include one or more of classification tasks, prediction tasks, and clustering tasks, and the corresponding models used for model reasoning may include one or more of classification models, prediction models, and clustering models.

[0201] In this embodiment, the terminal capabilities carried in the reported information may include multiple types, and the specific content of the terminal capabilities may be determined based on the actual needs of the terminal device, thereby ensuring the flexibility of the information represented by the terminal capabilities.

[0202] In one embodiment, the reported information includes terminal capabilities, and the specific implementation process of the step of "receiving the reported information sent by the terminal device through the base station" may include:

[0203] Based on the target transmission mode, the terminal capabilities reported by the terminal device through the base station are received.

[0204] The target transmission mode includes at least one of system message, RRC signaling, MAC CE signaling, DCI signaling, and NAS message.

[0205] Specifically, the terminal device may report the terminal device capability reporting signaling to the base station (edge ​​node) through at least one transmission method among system messages, RRC signaling, MAC CE signaling, DCI signaling, and NAS messages, that is, report the reporting information to the edge node through at least one transmission method among system messages, RRC signaling, MAC CE signaling, DCI signaling, and NAS messages. After the edge node receives the reporting information, the edge node may send the reporting information to the network node corresponding to the edge node. For example, the edge node may send the reporting information to the network node through the NG-C interface.

[0206] In this embodiment, the terminal device can transmit data to the base station through multiple transmission modes, thereby ensuring the diversity and reliability of data transmission.

[0207] In one embodiment, the model management method based on hierarchical federated learning further includes:

[0208] Receive the model inference request signaling sent by the terminal device through the base station, and return a response message corresponding to the inference request, so that the terminal device sends reporting information based on the response message; the response message can be a model inference response message.

[0209] Specifically, during the model reasoning process, the terminal device may be a terminal device to perform model reasoning. The terminal device may generate a model reasoning request signaling and send the model reasoning request signaling to the base station to which the terminal device belongs. After receiving the model reasoning request signaling, the base station may send the model reasoning request signaling to the network node. The network node may respond to the model reasoning request signaling, generate a model reasoning response message, and return the model reasoning response message to the terminal device to perform model reasoning through the base station. The terminal device that receives the model reasoning response message may send reporting information to the base station (edge ​​node) to which the terminal device belongs. The base station may send the reporting information to the network node through the NG-C interface. The network node may receive the reporting information and obtain a second configuration message.

[0210] In this embodiment, the terminal device can actively trigger the model reasoning process and promptly receive relevant information of the model reasoning sent by the network node, thereby improving the efficiency of the model reasoning.

[0211] In one embodiment, if Figure 5 As shown, the second configuration message includes model reasoning strategy delivery information or model update strategy delivery information based on hierarchical federated learning. The specific implementation process of the step "generating the second configuration message based on the reported information" can be:

[0212] Step 502: Based on the reported information, determine at least one model to be inferred whose target dimension meets a preset matching condition among multiple first models.

[0213] The multiple models stored locally in the network node may be multiple first models, and the first model may be a model that has completed model identification, or a model that has completed function identification. The target dimension may include model structure dimension, model complexity dimension, model accuracy, etc.; the preset matching condition may be the least complex, the most complex, the most accurate, the lowest accurate, or the most moderate, etc., which may be determined based on the actual application scenario.

[0214] Specifically, the network node can screen multiple first models based on the reported information sent by the received terminal device and the pre-configured preset number of models to be inferred, and extract the model that best matches the terminal capabilities carried in the reported information in terms of model structure, model complexity, accuracy, etc. as the model to be inferred / group of models to be inferred, or select a preset number of models as the model to be inferred / group of models to be inferred by compromise.

[0215] Step 504: Generate model reasoning strategy delivery information based on hierarchical federated learning based on the model identifier of the model to be inferred, the model file, and the identification information of the terminal device on which the model reasoning is to be executed.

[0216] Specifically, the network node can encapsulate the model identifier of the model to be inferred / the group of models to be inferred selected in the above steps, the model file of each model to be inferred, and the identification information of the terminal device to perform model inference, obtain the encapsulation result, and determine that the encapsulation result is the model inference strategy issuance information, and the model inference strategy issuance information can also be the model inference strategy issuance signaling. The network node can transmit the model inference strategy issuance information to the base station through NG-U, and the base station can transmit the model inference strategy issuance information to each terminal device to perform model inference.

[0217] In this embodiment, the network node can determine the model to be inferred based on the reported information, and synchronize data with the base station and the terminal device to be inferred in a timely manner to ensure the timeliness of the data and improve the efficiency of model reasoning.

[0218] In one embodiment, the specific implementation process of the step of "determining, based on the reported information, at least one model to be inferred whose target dimension satisfies a preset matching condition among multiple first models" may include:

[0219] Based on the reported information, a preset number of models to be inferred whose model structures, model complexity and model accuracy meet preset matching conditions are determined among the multiple first models.

[0220] Specifically, the network node can screen multiple first models based on the reported information sent by the received terminal device and the pre-configured preset number of models to be inferred, and extract the model that best matches the terminal capabilities carried in the reported information in terms of model structure, model complexity, accuracy, etc. as the model to be inferred / group of models to be inferred, or select a preset number of models as the model to be inferred / group of models to be inferred by compromise.

[0221] In this embodiment, the network node can screen the models to be inferred from multiple dimensions based on the reported information, thereby ensuring the comprehensiveness and flexibility of the inference model screening.

[0222] In one embodiment, the specific implementation process of the step of "sending the second configuration message to the terminal device through the base station" may include:

[0223] The second configuration message is sent to the base station through the user plane interface, so that the base station sends the second configuration message to the terminal device through the target transmission mode.

[0224] The target transmission mode includes at least one of a system message, an RRC signaling, a MAC CE signaling, a DCI signaling, and a NAS message. The user plane interface may be an NG-U interface.

[0225] Specifically, after receiving the reporting information sent by the terminal device, the network node can process based on the reporting information to obtain the second configuration message, and send the second configuration message to the base station. After receiving the second configuration message, the base station can send the second configuration message to each terminal device corresponding to the base station. The network node can transmit the second configuration message to each base station through the NG-U interface, and the base station can transmit the second configuration message to multiple terminal devices corresponding to the base station in the form of a target transmission mode.

[0226] Optionally, the base station may transmit the second configuration message to the terminal device in at least one message transmission mode of system message, RRC signaling, MAC CE signaling, DCI signaling, and NAS message.

[0227] In one example, a network node generates a second configuration message based on the reported information received from the terminal device, and the network node may send the second configuration message to the base station to which the terminal device belongs, so that the base station sends the second configuration message to the terminal device; in another example, the network node may send the second configuration message to each base station managed by the network node, so that each base station may send the second configuration message to each terminal device within the range of the base station; in another example, the network node may screen each base station managed by the network node, and send the second configuration message to the screened base stations, and each base station may schedule the terminal device based on the method described in the above embodiment to obtain the scheduled terminal device, and send the second configuration message to each scheduled terminal device, wherein each base station may also send the second configuration message to the terminal device corresponding to the identifier of the terminal device carried in the second configuration message.

[0228] In this embodiment, data transmission can be achieved between network nodes, base stations and terminal devices through multiple transmission modes, ensuring the reliability, diversity and flexibility of data transmission.

[0229] In one embodiment, the model performance monitoring feedback information includes one or more of a system performance index, an accuracy index, a complexity index, and a generalization performance index.

[0230] Specifically, the model performance monitoring feedback information may include multiple indicators. System performance indicators may include one or more of throughput, block error rate, and latency. Accuracy indicators may include one or more of the accuracy of the model output, MSE / NMSE / cosine similarity between the model input and output, etc. Complexity indicators may include model size and computational complexity. The model size refers to the number of parameters contained in the model, and the computational complexity may be FLOPs. The model performance monitoring feedback information may also include generalization performance indicators, and the specific content may be determined based on the actual application scenario.

[0231] Optionally, after the UE performs model inference, it can monitor the model performance in real time, obtain the model performance monitoring result, and obtain the model performance monitoring feedback information based on the result. The UE can feedback the model performance monitoring feedback information to the edge node corresponding to the terminal device, and the edge node can send the model performance monitoring feedback information to the network node (NW side) corresponding to the edge node after receiving the model performance monitoring feedback information.

[0232] In this embodiment, by monitoring the performance indicators during the model reasoning process, the model performance can be monitored in real time and adjusted in time to ensure the timeliness of the model.

[0233] In one of the embodiments, the reported information includes model performance monitoring feedback information.

[0234] Accordingly, if Figure 6 As shown, the specific implementation process of the step "generating a second configuration message based on the reported information" may include:

[0235] Step 602: Receive model performance monitoring feedback information.

[0236] Specifically, the terminal device may report the model performance monitoring feedback information to the base station to which the terminal device belongs. After receiving the model performance monitoring feedback information, the base station may send the model performance monitoring feedback information to the network node (NW side) corresponding to the base station. The network node may receive the model performance monitoring feedback information sent by the base station.

[0237] Step 604: Receive a model update request, where the model update request includes an identifier of the model to be updated / a function to be updated and reference information.

[0238] The model update request is generated when the terminal device model performance monitoring feedback information and the preset indicator threshold meet the preset update conditions.

[0239] Specifically, the terminal device can compare the indicator contained in the model performance monitoring feedback information with the preset indicator threshold corresponding to the indicator. If it is determined that the indicator and the preset indicator threshold meet the preset update conditions, a model update event is triggered and a model update request is generated. The terminal device can send the model update request to the network node through the base station.

[0240] The specific process of the UE determining whether the preset update condition is met may be: when the indicator included in the model performance monitoring feedback information is the delay, if the UE determines that the delay is greater than or equal to the preset delay threshold corresponding to the delay, the UE may determine that the preset update condition is currently met and trigger a model update event; when the indicator included in the model performance monitoring feedback information is the model accuracy, if the UE determines that the model accuracy is less than or equal to the preset delay threshold corresponding to the delay, the UE may determine that the preset update condition is currently met and trigger a model update event; the UE may determine whether the preset update condition is met based on the meaning of the indicator and the relationship between the indicator and the preset indicator threshold corresponding to the indicator.

[0241] Step 606: respond to the model update request and generate a second configuration message.

[0242] Specifically, the network node can respond to the model update request, and the response process may include: the network node determines at least one model to be updated for model update, generates model update policy information based on the at least one model to be updated, that is, obtains a second configuration message, and the network node can send the model update policy information to the terminal device to be updated through the base station, and the terminal device can perform model update based on the model update policy information to obtain a model update result.

[0243] In this embodiment, the terminal device / network node can actively determine whether a model update process is needed based on the monitored model indicator information, thereby improving the control of the terminal device / network node over the model update process and being able to update the model in a timely manner when the model does not meet the requirements.

[0244] In one of the embodiments, the reported information includes model performance monitoring feedback information.

[0245] Accordingly, if Figure 7 As shown, the specific implementation process of the step "generating a second configuration message based on the reported information" may include:

[0246] Step 702: Receive model performance monitoring feedback information.

[0247] Specifically, the terminal device can obtain model performance monitoring feedback information based on the model performance monitoring results, for example, the model performance monitoring results can be used as model performance monitoring feedback information; the terminal device can determine the model performance monitoring results as reporting information, and send the reporting information to the base station to which the terminal device belongs, and the base station can send the reporting information to the network node corresponding to the base station.

[0248] Step 704: trigger a model update event based on a preset time period, or trigger an update event if the model performance monitoring feedback information and the preset indicator threshold meet the preset update condition.

[0249] The preset time period is associated with the terminal capability or the auxiliary information.

[0250] Specifically, the triggering process of the model update can be actively triggered by the network side. After receiving the reported information, the network node can trigger the model update event based on the target trigger condition. The specific process of the target trigger condition may include periodic model update and non-periodic model update; one possible implementation scenario may be that the network node triggers the update event of the model to be updated / the model group to be updated according to a preset time period, and generates a second configuration message, wherein the preset time period may be associated with the terminal capability, or may be associated with the change of auxiliary information, for example, the auxiliary information may be a scene, a site, a channel, or a data sampling speed, etc.; another possible triggering method may be that the network node triggers the model update event based on the received model performance monitoring feedback information, specifically, the network node may compare the indicator contained in the model performance monitoring feedback information with the preset indicator threshold corresponding to the indicator, and if it is determined that the indicator and the preset indicator threshold meet the preset update condition, the model update event is triggered.

[0251] Optionally, the specific process of the network node determining whether the preset update condition is met may be that, when the indicator included in the model performance monitoring feedback information is delay, if the network node determines that the delay is greater than or equal to a preset delay threshold corresponding to the delay, then the network node may determine that the preset update condition is currently met, triggering a model update event; when the indicator included in the model performance monitoring feedback information is model accuracy, if the network node determines that the model accuracy is less than or equal to the preset delay threshold corresponding to the delay, then the network node may determine that the preset update condition is currently met, triggering a model update event; the network node may determine whether the preset update condition is met based on the meaning of the indicator and the relationship between the indicator and the preset indicator threshold corresponding to the indicator.

[0252] Step 706: Generate a second configuration message based on the triggered update event.

[0253] Specifically, the network node determines at least one model to be updated for model update to be executed, generates model update policy information based on the at least one model to be updated, that is, obtains a second configuration message, and the network node can send the model update policy information to the terminal device to be updated through the base station, and the terminal device can execute the model update based on the model update policy information to obtain the model update result.

[0254] In this embodiment, the terminal device / network node can actively determine whether a model update process is needed based on the monitored model indicator information, thereby improving the control of the terminal device / network node over the model update process and being able to update the model in a timely manner when the model does not meet the requirements.

[0255] In one of the embodiments, the second configuration message includes model reasoning strategy delivery information based on hierarchical federated learning.

[0256] Accordingly, the specific implementation process of the step “generating a second configuration message” may include:

[0257] At least one model to be updated is determined. Based on the model to be updated, identification information of a terminal device on which the model update is to be executed, and an update strategy, model reasoning strategy delivery information based on hierarchical federated learning is generated.

[0258] Specifically, the second configuration message is the model reasoning strategy delivery information; the process of the network node generating the second configuration message may be that the terminal device to be executed model reasoning can send a model reasoning request signaling to the network node through the base station, the network node can respond to the model reasoning request signaling, generate a model reasoning response message, and send the model reasoning response message to the terminal device to be executed model reasoning through the base station; the terminal device receiving the model reasoning response message can send a reporting information to the base station (edge ​​node) to which the terminal device belongs, and the base station can send the reporting information to the network node through the NG-C interface. After receiving the reporting information, the network node can determine, based on the reporting information, at least one model to be inferred whose target dimension meets the preset matching conditions among the multiple first models stored in the network node, and based on the model identifier of the model to be inferred, the model file and the identification information of the terminal device to be executed model reasoning, generate the model reasoning strategy delivery information based on hierarchical federated learning, that is, generate the second configuration message, and transmit it to the edge node through the NG-U interface, and transmit it to the terminal device to be executed model reasoning by the edge node.

[0259] In one example, the network node can encapsulate the model identifier of the model to be inferred, the model file, and the identification information of the terminal device on which the model inference is to be executed to obtain a encapsulation result, and determine that the encapsulation result is information issued as a model inference strategy based on hierarchical federated learning.

[0260] Among them, the process of the network node determining the model to be inferred / the group of models to be inferred may be that the multiple models locally stored in the network node may be multiple first models, among which the network node may be based on the reported information and the pre-configured preset number, and match them by maximizing the model accuracy, minimizing the model complexity, or other dimensions to obtain a matching preset number of models to be inferred / the group of models to be inferred.

[0261] In this embodiment, the network node can select a suitable model to be inferred during the model inference process to perform model inference processing, thereby ensuring the efficiency of model inference.

[0262] In one embodiment, the update strategy includes one or more of a local update strategy, a global update strategy, a model replacement strategy, and a model deletion strategy.

[0263] Specifically, the specific meaning of the update strategy and the specific process of the terminal device performing model update processing based on the update strategy to obtain the model update result have been described in detail in the corresponding embodiments of the above steps 302 to 308, and will not be repeated here.

[0264] In one embodiment, if Figure 8 As shown, the model management method based on hierarchical federated learning also includes:

[0265] Step 802: Receive model data sent by each base station.

[0266] The model data includes model files and model content / function identifiers.

[0267] Specifically, the base station can determine the identified federated learning model indicated by the base station, and send the model data of the identified federated learning model to the network node through the NG-U interface. The network node can receive the model data sent by each base station respectively. The model data represents each federated learning model that the base station has identified. The model data includes model files, model content or function identifiers.

[0268] Step 804: determine each first model based on each model data, and arrange them according to the evaluation index of each first model to obtain a first model list.

[0269] The evaluation index may include one or more of accuracy, convergence speed, model complexity or other evaluation indexes.

[0270] Specifically, the network node obtains multiple identified models, namely, first models, based on the model files in each model data. The network node can sort each first model according to the evaluation index of each first model to obtain a first model list, which can be a model list or a function list.

[0271] In this embodiment, the network node can efficiently store the functions / models that have been identified by each base station, thereby improving the storage efficiency of the model.

[0272] In one embodiment, the model management method based on hierarchical federated learning further includes:

[0273] An identification request and a model capability sent by a base station are received, where the identification request includes a model identification request and a function identification request, and the model capability is used to indicate the models / supported functions supported by the base station.

[0274] Specifically, after the training of the federated learning model is completed, an identification event can be triggered. The identification event can be a model identification event or a function identification event. The identification event can be triggered by a base station. The specific identification process can be: the base station can send an identification request to the network node through the NG-C interface, and the base station reports the model capability to the network node. The model capability reported by the base station is associated with a specific configuration / specific condition. The model capability reported by the base station to the network node can also be associated with an auxiliary condition. The specific configuration / specific condition can be a specific configuration / specific condition under the capabilities of each terminal in the area where the base station is located. The area where the base station is located can be the cell / sector / cell group where the base station is located; the auxiliary conditions can be but are not limited to scenarios, sites, channels, data sets, etc.

[0275] In one example, the model capabilities reported by the base station may include the model type, the input and output descriptions of the model, the AI ​​(Artificial Intelligence) related features involved in the model, the accuracy that can be achieved on the standard data set, the convergence speed, the model complexity, the application scenarios, etc. The interaction of model capabilities between the base station and the network node can be carried through user plane signaling. For the models reported by all base stations, the network node sorts and establishes a model / function list according to accuracy, convergence speed, model complexity or other evaluation indicators, and assigns a model / function ID to each base station. The network node sends the model / function ID indication information to the corresponding base station through the NG-U interface. The model / function ID is unique on the network side, and the format of the ID includes at least one of the following:

[0276] Two parts of bits are connected in series, including: an identification bit of the cell / sector / cell group to which it belongs, and a bit used to distinguish different models / functions in the same cell / sector / cell group;

[0277] Other bit encoding methods used to distinguish different models / functions.

[0278] Based on this, the base station can determine the identified federated learning model indicated by the base station, and send the model data of the identified federated learning model to the network node through the NG-U interface. The network node can receive the model data sent by each base station respectively. The model data represents each federated learning model that the base station has identified. The model data includes model files, model content or function identifiers.

[0279] In this embodiment, the base station can actively trigger model recognition to improve the flexibility of model recognition.

[0280] In one embodiment, the model management method based on hierarchical federated learning further includes:

[0281] An identification request is sent to each base station so that each base station reports model capabilities, where the model capabilities are used to indicate the models / supported functions supported by the base station.

[0282] Specifically, after the training of the federated learning model is completed, an identification event can be triggered. The identification event can be a model identification event or a function identification event. The identification event can be triggered by a network node. The specific identification process can be: the network node can send an identification request to the base station through the NG-C interface. After the base station receives the identification request, it can report the model capability to the network node in response to the model identification request. The model capability reported by the base station is associated with a specific configuration / specific condition. The model capability reported by the base station to the network node can also be associated with an auxiliary condition. The specific configuration / specific condition can be a specific configuration / specific condition under the capabilities of each terminal in the area where the base station is located. The area where the base station is located can be the cell / sector / cell group where the base station is located; the auxiliary conditions can be but are not limited to scenarios, sites, channels, data sets, etc.

[0283] The base station reports the model capability to the network node. The model capability reported by the base station is associated with a specific configuration / specific condition. The model capability reported by the base station to the network node may also be associated with an auxiliary condition. The specific configuration / specific condition may be a specific configuration / specific condition under the capabilities of each terminal in the area where the base station is located. The area where the base station is located may be the cell / sector / cell group where the base station is located. The auxiliary conditions may be but are not limited to scenarios, sites, channels, data sets, etc.

[0284] In one example, the model capabilities reported by the base station may include the model type, the input and output descriptions of the model, the AI-related features involved in the model, the accuracy that can be achieved on the standard data set, the convergence speed, the model complexity, the application scenarios, etc. The interaction of model capabilities between the base station and the network node can be carried through user plane signaling. For the models reported by all base stations, the network node sorts and establishes a model / function list according to accuracy, convergence speed, model complexity or other evaluation indicators, and assigns a model / function ID to each base station. The network node sends the model / function ID indication information to the corresponding base station through the NG-U interface. The model / function ID is unique on the network side, and the format of the ID includes at least one of the following:

[0285] Two parts of bits are connected in series, including: an identification bit of the cell / sector / cell group to which it belongs, and a bit used to distinguish different models / functions in the same cell / sector / cell group;

[0286] Other bit encoding methods used to distinguish different models / functions.

[0287] Based on this, the base station can determine the identified federated learning model indicated by the base station, and send the model data of the identified federated learning model to the network node through the NG-U interface. The network node can receive the model data sent by each base station respectively. The model data represents each federated learning model that the base station has identified. The model data includes model files, model content or function identifiers.

[0288] In this embodiment, the network node can actively trigger model recognition to improve the flexibility of model recognition.

[0289] In an exemplary embodiment, Fig. 9 As shown in the figure, a model management method based on hierarchical federated learning is provided for application Figure 1 A base station in a mobile communication system, the mobile communication system includes a network and multiple base stations, each base station corresponds to multiple terminal devices, the terminal devices are distributed nodes for performing hierarchical federated learning, and the base stations are central nodes for performing hierarchical federated learning. The method includes:

[0290] Step 902: obtaining user scheduling result indication information corresponding to each terminal device based on user characteristic information of each terminal device, and sending a first configuration message to each terminal device.

[0291] The user characteristic information includes at least one of user data characteristic information and user channel state information, and the first configuration message includes user scheduling result indication information.

[0292] Specifically, the base station can determine all terminal devices within the coverage of the base station, and screen among all the terminal devices based on hierarchical federated learning to determine the selected terminal devices, and obtain a first configuration message based on the user scheduling indication information. The base station can send the first configuration message to each selected terminal device, and the user scheduling result indication information contained in the first configuration message can indicate that the terminal device is selected by the base station.

[0293] In one example, the specific process of scheduling by the base station may include: the base station determines the first number of terminal devices that currently need to be scheduled (used), and screens all terminal devices within the coverage of the base station based on the user feature information of each terminal device, extracts the first number of terminal devices, and determines that the user scheduling result indication information of the first number of terminal devices is the first indication information, and the first indication information can represent that the terminal device is selected by the base station; the user scheduling indication information of other terminal devices can be empty or the second indication information, and the second indication information is used to represent that the terminal device is not selected by the base station. The present disclosure does not limit the specific form of the first indication information and the second indication information, and those skilled in the art can make specific determinations based on actual application scenarios.

[0294] Step 904: receiving reporting information sent by the terminal device.

[0295] The reported information includes at least one of terminal capabilities, reasoning task types, and model performance monitoring feedback information.

[0296] Step 906: Send the reporting information to the network node, so that the network node obtains the second configuration message.

[0297] Among them, the second configuration message is the model reasoning strategy delivery information or the model update strategy delivery information based on hierarchical federated learning, and the second configuration message is used to enable the terminal device to obtain the model reasoning result or the model update result.

[0298] Specifically, the terminal device can send reporting information to the base station, and the base station can send the received reporting information to the network node. After receiving the reporting information, the network node can generate a second configuration message based on the reporting information. Based on this, the network node can send the second configuration message to the base station, and the base station can send the received second configuration message to each terminal device. In the model reasoning process, the base station sends a second configuration message to each terminal device, and the second configuration message can include model reasoning strategy delivery information based on hierarchical federated learning. In the model update process, the network node sends a second configuration message to each terminal device, and the second configuration message can include model update strategy delivery information based on hierarchical federated learning. The terminal device can receive the second configuration message, and perform model reasoning based on the model reasoning strategy delivery information in the second configuration message to obtain a model reasoning result; the terminal device can also update the model based on the model update strategy delivery information in the second configuration message to obtain a model update result.

[0299] In this embodiment, data interaction between base stations and terminals can be realized, the management efficiency and generalization performance in the wireless intelligent federated learning model can be improved, the additional consumption of communication resources and computing overhead can be reduced, the resources to be allocated can be flexibly configured, and the terminal devices for federated learning can be flexibly scheduled, further expanding the applicable scenarios of the model.

[0300] In one embodiment, the model management method based on hierarchical federated learning further includes:

[0301] Model data is sent to a network node, where the model data includes a model file and a model content / function identifier, so that the network node determines each first model based on each model data and obtains a first model list.

[0302] Specifically, the base station can determine the identified federated learning model indicated by the base station, and send the model data of the identified federated learning model to the network node through the NG-U interface. The network node can receive the model data sent by each base station respectively. The model data represents each federated learning model that the base station has identified. The model data includes model files, model content or function identifiers.

[0303] The network node obtains multiple identified models, i.e., first models, based on the model files in each model data. The network node can sort each first model according to the evaluation index of each first model to obtain a first model list, which can be a model list or a function list. The evaluation index can include one or more of accuracy, convergence speed, model complexity or other evaluation indexes.

[0304] In this embodiment, the network node can efficiently store the functions / models that have been identified by each base station, thereby improving the storage efficiency of the model.

[0305] In one embodiment, the model management method based on hierarchical federated learning further includes:

[0306] An identification request and a model capability are sent to a network node, wherein the identification request includes a model identification request and a function identification request, and the model capability is used to indicate the models / supported functions supported by the base station.

[0307] Specifically, after the training of the federated learning model is completed, an identification event can be triggered. The identification event can be a model identification event or a function identification event. The identification event can be triggered by a base station. The specific identification process can be: the base station can send an identification request to the network node through the NG-C interface, and the base station reports the model capability to the network node. The model capability reported by the base station is associated with a specific configuration / specific condition. The model capability reported by the base station to the network node can also be associated with an auxiliary condition. The specific configuration / specific condition can be a specific configuration / specific condition under the capabilities of each terminal in the area where the base station is located. The area where the base station is located can be the cell / sector / cell group where the base station is located; the auxiliary conditions can be but are not limited to scenarios, sites, channels, data sets, etc.

[0308] In one example, the model capabilities reported by the base station may include the model type, the input and output descriptions of the model, the AI-related features involved in the model, the accuracy that can be achieved on the standard data set, the convergence speed, the model complexity, the application scenarios, etc. The interaction of model capabilities between the base station and the network node can be carried through user plane signaling. For the models reported by all base stations, the network node sorts and establishes a model / function list according to accuracy, convergence speed, model complexity or other evaluation indicators, and assigns a model / function ID to each base station. The network node sends the model / function ID indication information to the corresponding base station through the NG-U interface. The model / function ID is unique on the network side, and the format of the ID includes at least one of the following:

[0309] Two parts of bits are connected in series, including: an identification bit of the cell / sector / cell group to which it belongs, and a bit used to distinguish different models / functions in the same cell / sector / cell group;

[0310] Other bit encoding methods used to distinguish different models / functions.

[0311] Based on this, the base station can determine the identified federated learning model indicated by the base station, and send the model data of the identified federated learning model to the network node through the NG-U interface. The network node can receive the model data sent by each base station respectively. The model data represents each federated learning model that the base station has identified. The model data includes model files, model content or function identifiers.

[0312] In this embodiment, the base station can actively trigger model recognition to improve the flexibility of model recognition.

[0313] In one embodiment, the model management method based on hierarchical federated learning further includes:

[0314] Receive an identification request sent by a network node, and report model capabilities to the network node based on the identification request, where the model capabilities are used to indicate the models / supported functions supported by the base station.

[0315] Specifically, after the training of the federated learning model is completed, an identification event can be triggered. The identification event can be a model identification event or a function identification event. The identification event can be triggered by a network node. The specific identification process can be: the network node can send an identification request to the base station through the NG-C interface. After the base station receives the identification request, it can report the model capability to the network node in response to the model identification request. The model capability reported by the base station is associated with a specific configuration / specific condition. The model capability reported by the base station to the network node can also be associated with an auxiliary condition. The specific configuration / specific condition can be a specific configuration / specific condition under the capabilities of each terminal in the area where the base station is located. The area where the base station is located can be the cell / sector / cell group where the base station is located; the auxiliary conditions can be but are not limited to scenarios, sites, channels, data sets, etc.

[0316] The base station reports the model capability to the network node. The model capability reported by the base station is associated with a specific configuration / specific condition. The model capability reported by the base station to the network node may also be associated with an auxiliary condition. The specific configuration / specific condition may be a specific configuration / specific condition under the capabilities of each terminal in the area where the base station is located. The area where the base station is located may be the cell / sector / cell group where the base station is located. The auxiliary conditions may be but are not limited to scenarios, sites, channels, data sets, etc.

[0317] In one example, the model capabilities reported by the base station may include the model type, the input and output descriptions of the model, the AI-related features involved in the model, the accuracy that can be achieved on the standard data set, the convergence speed, the model complexity, the application scenarios, etc. The interaction of model capabilities between the base station and the network node can be carried through user plane signaling. For the models reported by all base stations, the network node sorts and establishes a model / function list according to accuracy, convergence speed, model complexity or other evaluation indicators, and assigns a model / function ID to each base station. The network node sends the model / function ID indication information to the corresponding base station through the NG-U interface. The model / function ID is unique on the network side, and the format of the ID includes at least one of the following:

[0318] Two parts of bits are connected in series, including: an identification bit of the cell / sector / cell group to which it belongs, and a bit used to distinguish different models / functions in the same cell / sector / cell group;

[0319] Other bit encoding methods used to distinguish different models / functions.

[0320] Based on this, the base station can determine the identified federated learning model indicated by the base station, and send the model data of the identified federated learning model to the network node through the NG-U interface. The network node can receive the model data sent by each base station respectively. The model data represents each federated learning model that the base station has identified. The model data includes model files, model content or function identifiers.

[0321] In this embodiment, the network node can actively trigger model recognition to improve the flexibility of model recognition.

[0322] In one embodiment, the model management method based on hierarchical federated learning further includes:

[0323] Based on the model training results sent by each terminal device, a trained global model is obtained.

[0324] Specifically, the base station can obtain a trained global model based on the model training results received from each terminal device.

[0325] In one of the embodiments, the terminal device is a terminal device of a scheduling user.

[0326] Accordingly, if Fig.10 As shown, the specific implementation process of the step of "obtaining a trained global model based on the model training results sent by each terminal device" may include:

[0327] Step 1002, for the i-th round of training, based on the training results sent by each terminal device in the i-1-th round, obtain the model parameters of the global model of the i-th round.

[0328] Step 1004, when the preset training completion conditions are not met, broadcast the model parameters of the global model of the i-th round to each terminal device, so that each terminal device performs training based on the model parameters of the global model of the i-th round, until the base station determines that the preset training completion conditions are met, and a trained global model is obtained.

[0329] Among them, the training results sent by each terminal device can be a local update gradient. The base station can perform model aggregation based on the local update gradients sent by each terminal device to obtain updated global model parameters. When the base station determines that the preset training completion conditions are not met at present, the base station can broadcast the updated global model parameters to the terminal device corresponding to the base station through the physical downlink shared channel (PDSCH), so that the terminal device that receives the global model parameters can perform training until a trained global model is obtained when the preset training completion conditions are met.

[0330] In one example, the network may include M cells / sectors / cell groups (with the same cell ID), each cell / sector / cell group is provided with an edge node (base station) and N user equipments (terminal devices), and there is a network node in the network. For each cell / sector / cell group, the following steps are performed independently:

[0331] Step 1: The user equipment and edge node start model training. The model training process includes at least one or more of periodic model training and non-periodic model training:

[0332] Periodic model training: The model training process is started between the user equipment and the edge node according to the preset time period value, where the time period value is associated with the specific configuration / specific conditions under the capabilities of each UE in the cell / sector / cell group, and the time period value can also be associated with auxiliary conditions, which can include scenarios, sites, channels, data sampling speeds, typical parameters or other requirements. Optionally, the network node can configure the time period value through at least one of system messages, RRC signaling, MAC CE signaling, DCI signaling, and NAS messages.

[0333] Aperiodic model training: The model training process is started between the user equipment and the edge node in an aperiodic manner (triggered by the UE side or the NW side).

[0334] Model training is triggered by at least one of the following methods:

[0335] The UE side triggers, and the UE side feeds back the model training request signaling to the edge node for one or more AI-related features. The edge node responds to the request and starts model training.

[0336] The NW side triggers, and the NW side sends a model training request signaling to the user device through the edge node for one or more AI-related features. The user device responds to the request and starts model training.

[0337] Step 2: The model training is performed for T rounds. During the i-th round of model training, 0≤i≤T, the user device trains the local model and obtains the local model update result.

[0338] In step 3, the edge node can determine the set of users that need to be scheduled in this round of model training based on the user scheduling method. Only the scheduled user equipment needs to perform the local model training process in step 102. Specifically, the edge node pre-sets the number of users to be scheduled in this round of model training, and determines the set of users to be scheduled from all users in the cell / sector / cell group within its coverage area based on the channel state information fed back by the user CSI-RS channel estimation, the characteristic information between the user's local update gradient, or a random selection of a part of the users, and broadcasts the ID set of the users to be scheduled to all users in the cell / sector / cell group through the downlink physical broadcast channel PBCH.

[0339] In step 4, the scheduled user equipment uploads the local model update result (training result) to the edge node through the physical uplink shared channel (PUSCH, Physical Uplink Shared Channel).

[0340] In step 5, the edge node performs model aggregation based on the local update gradients in the training results uploaded by all scheduled users to obtain the model parameters of the updated global model.

[0341] In step 6, the edge node broadcasts the updated model parameters of the global model to all user equipments through a physical downlink shared channel (PDSCH).

[0342] If it is determined that the preset training completion condition is not met at present, steps 1 to 6 are repeated until the global model converges, that is, the preset training completion condition is met. The preset training completion condition may be a convergence condition, and determining whether the model meets the preset convergence condition may include at least one of the following:

[0343] The number of training rounds reaches the preset upper limit, the loss function of the global model on the validation set is less than or equal to the preset value, and the difference between the loss function of the global model on the validation set in two adjacent rounds is less than or equal to the preset value.

[0344] Specifically, in the same cell / sector / cell group, one or more models can be trained, and the training modes include at least one of the following: parallel mode: the training of multiple models is simultaneous or overlapping in time; serial mode: the training of multiple models is sequential in time and without overlap.

[0345] In one embodiment, the preset training completion conditions include one or more of the number of training rounds reaching a preset upper limit, the loss function of the global model being less than or equal to a preset value, and the difference between the loss function of the global model in two adjacent rounds being less than or equal to a preset value.

[0346] Specifically, the preset training completion condition may be a convergence condition, and determining whether the model satisfies the preset convergence condition may include at least one of the following: the number of training rounds reaches a preset upper limit, the loss function of the global model on the validation set is less than or equal to a preset value, and the difference between the loss function of the global model on the validation set in two adjacent rounds is less than or equal to a preset value.

[0347] In one embodiment, the model management method based on hierarchical federated learning further includes:

[0348] Based on the user characteristic information of each terminal device, the user scheduling result indication information corresponding to each terminal device is obtained, and the user characteristic information includes at least one of user data characteristic information and user channel state information. A first configuration message is sent to each terminal device, and the first configuration message includes the user scheduling result indication information.

[0349] In one embodiment, the specific implementation process of the step of "obtaining user scheduling result indication information corresponding to each terminal device based on user characteristic information of each terminal device" may include:

[0350] Obtain the pre-configured number of scheduled users. Among all users in the coverage of the base station, based on at least one of the user data feature information and user channel state information of each terminal device, screen out the terminal devices with the scheduled number of users, and determine the user scheduling result indication information as the scheduled indication information.

[0351] Specifically, the base station can determine all terminal devices within the coverage of the base station, and screen among all the terminal devices based on hierarchical federated learning to determine the selected terminal devices, and obtain a first configuration message based on the user scheduling indication information. The base station can send the first configuration message to each selected terminal device, and the user scheduling result indication information contained in the first configuration message can indicate that the terminal device is selected by the base station.

[0352] In one example, the specific process of scheduling by the base station may include: the base station determines the first number of terminal devices that currently need to be scheduled (used), and screens all terminal devices within the coverage of the base station based on the user feature information of each terminal device, extracts the first number of terminal devices, and determines that the user scheduling result indication information of the first number of terminal devices is the first indication information, and the first indication information can represent that the terminal device is selected by the base station; the user scheduling indication information of other terminal devices can be empty or the second indication information, and the second indication information is used to represent that the terminal device is not selected by the base station. The present disclosure does not limit the specific form of the first indication information and the second indication information, and those skilled in the art can make specific determinations based on actual application scenarios.

[0353] In this embodiment, the base station can perform user scheduling to ensure the reliability of terminal devices performing model training and improve the efficiency of model training.

[0354] The following describes in detail the specific implementation process of the model management method based on hierarchical federated learning in conjunction with a specific embodiment. Before introducing the specific implementation process, the professional terms involved in this embodiment are explained first:

[0355] 6G: 6th Generation Mobile Communication Technology

[0356] IoT: Internet of Things

[0357] LCM: Life Cycle Management

[0358] LMF: Location Management Function

[0359] SINR: Signal to Interference plus Noise Ratio, signal interference noise ratio

[0360] SGD: Stochastic Gradient Descent, stochastic gradient descent

[0361] SSB: Synchronization Signal Block, synchronization signal block

[0362] RSRP: Reference Signal Receiving Power, reference signal receiving power

[0363] KPI: Key Performance Indicators

[0364] MSE / NMSE: (Normalized) Mean Squared error, (normalized) mean square error

[0365] The method provided in this embodiment mainly realizes lifecycle management based on hierarchical federated learning, and provides corresponding signaling interaction mechanisms and configuration information, which can improve transmission reliability and reduce interruption probability as the goal, and can also improve the convergence speed of the federated learning model by scheduling user equipment. The method provided in this embodiment is a lifecycle management method that supports flexible user scheduling methods, which can improve the management efficiency and generalization performance of the intelligent air interface federated learning model, reduce communication and computing overhead, effectively combat the impact of data heterogeneity, and improve the convergence speed and accuracy performance of model training. Considering strategies to reduce energy consumption, the cost of deploying federated learning algorithms in communication systems is reasonably reduced, and considering model generalization performance, the federated learning model can be well adapted to new application scenarios or deployed on newly added network nodes. The method provided in this embodiment is based on the fact that the convergence of the federated learning model is closely related to the data distribution of each user device, and can effectively perform user scheduling and accelerate the convergence speed of the federated learning model.

[0366] The lifecycle management (LCM) involved in the method provided in this embodiment is an important concept proposed in the current 3GPP standardization research based on the wireless AI framework. As a full-process management solution covering AI models, LCM includes data collection, model training, model reasoning, model management (based on functionality / model ID selection, activation / deactivation, model switching / fallback, model monitoring, model update), model transmission, etc. Therefore, the method provided in this embodiment is designed and optimized under the condition of meeting the lifecycle management framework defined by 3GPP.

[0367] Specifically, the distributed network architecture in the method provided in this embodiment includes network nodes, base stations and terminal devices. The network node can be a central node / cloud, for example, LMF / OAM / NWDAF (NW side); the base station can be an edge node / server side, for example, CU / BBU under the gNB / CU-DU separation architecture; the terminal device can be a terminal node / user equipment / local client, for example, UE (UE side). Fig.11aAs shown, the lifecycle management part based on hierarchical federated learning in the method provided in this embodiment at least includes a model training process based on hierarchical federated learning, a model identification process based on hierarchical federated learning, a model reasoning process based on hierarchical federated learning, and a model update process based on hierarchical federated learning. Specifically, initialization, initializing model parameters for all user devices; model training based on hierarchical federated learning, for a total of M cells / sectors / cell groups in the network, in each cell / sector / cell group, a user device or edge node triggers the start of the model training process, any scheduled user device updates the local model parameters, and transmits the model parameters to the edge node through the air interface, the edge node performs model aggregation to update the global model parameters, and through several rounds of iterative updates until the global model parameters converge, M clustered federated learning models are formed; model / function identification based on hierarchical federated learning, the network node performs model / function identification on the federated learning models that have been trained on all cells / sectors / cell groups, and all edge nodes connect the identified federated learning models to the NG interface. The port is sent to the network node for model storage. The network node assigns model / function IDs to all identified models for model deployment / inference / update and other operations. Based on model inference of hierarchical federated learning, the user equipment to perform model inference sends a model inference request signaling to the NW side to report UE capabilities and information related to the inference task. The network node determines the matching inference model / model group based on the information reported by the UE side, generates a signaling containing the inference model / model group configuration information, and transmits it to the user equipment to perform model inference. The user equipment to perform model inference performs model inference and obtains the inference result, and monitors the model performance in real time. Based on model update of hierarchical federated learning, the model update process is triggered by the UE side or the NW side, and the network node executes the model update process.

[0368] The following describes the model training process based on hierarchical federated learning, which includes:

[0369] The network may include M cells / sectors / cell groups (with the same cell ID), each cell / sector / cell group is provided with an edge node (base station) and N user equipments (terminal devices), and there is a network node in the network. For each cell / sector / cell group, the following steps are performed independently:

[0370] Step 1: The user equipment and edge node start model training. The model training process includes at least one or more of periodic model training and non-periodic model training:

[0371] Periodic model training: The model training process is started between the user equipment and the edge node according to the preset time period value, where the time period value is associated with the specific configuration / specific conditions under the capabilities of each UE in the cell / sector / cell group, and the time period value can also be associated with auxiliary conditions, which can include scenarios, sites, channels, data sampling speeds, typical parameters or other requirements. Optionally, the network node can configure the time period value through at least one of system messages, RRC signaling, MAC CE signaling, DCI signaling, and NAS messages.

[0372] Aperiodic model training: The model training process is started between the user equipment and the edge node in an aperiodic manner (triggered by the UE side or the NW side).

[0373] Model training is triggered by at least one of the following methods:

[0374] The UE side triggers, and the UE side feeds back the model training request signaling to the edge node for one or more AI-related features. The edge node responds to the request and starts model training.

[0375] The NW side triggers, and the NW side sends a model training request signaling to the user device through the edge node for one or more AI-related features. The user device responds to the request and starts model training.

[0376] Step 2: The model training is performed for T rounds. During the i-th round of model training, 0≤i≤T, the user device trains the local model and obtains the local model update result.

[0377] Specifically, the process of “the user device trains the local model and obtains the local model update result” may be that the user device may be a terminal device, the local model update result (training result) may be a local update gradient, and the terminal device may obtain the local update gradient based on local data through a stochastic gradient descent (SGD) algorithm. For example, the local update gradient g may be obtained through the following formula: i (w i,t ):

[0378]

[0379] Among them, w i,t is the local model parameter of the i-th user, is a data batch randomly sampled from the user’s local data set. w i,t The gradient of the loss function at the data point (x,y).

[0380] In step 3, the edge node can determine the set of users that need to be scheduled in this round of model training based on the user scheduling method. Only the scheduled user equipment needs to perform the local model training process in step 102. Specifically, the edge node pre-sets the number of users to be scheduled in this round of model training, and determines the set of users to be scheduled from all users in the cell / sector / cell group within its coverage area based on the channel state information fed back by the user CSI-RS channel estimation, the characteristic information between the user's local update gradient, or a random selection of a part of the users, and broadcasts the ID set of the users to be scheduled to all users in the cell / sector / cell group through the downlink physical broadcast channel PBCH.

[0381] The number of scheduled users pre-configured by the base station is less than or equal to the number of all users in the RRC activated state in the cell / sector / cell group.

[0382] Specifically, the base station can define the utility function of each terminal device as the intra-cell benefit minus the inter-cell benefit, where the intra-cell benefit is the average of the cosine similarities between the local update gradient of the user and the local update gradient of other users in the same cell / sector / cell group. The local update gradient of each user device can be fed back to the base station through the uplink data channel. For example, the local update gradient of the user device can be compressed to obtain a compression result, and the compression result is fed back to the base station through the uplink data channel to reduce transmission overhead; the inter-cell benefit is the average of the cosine similarities between the local update gradient of the user and the local update gradient of other users not in the same cell / sector / cell group.

[0383] The base station can also define the optimization problem as maximizing the sum of weighted utility functions of all users in the cell / sector / cell group, where the weight is the proportion of the data volume of each user. The base station can perform user scheduling based on the proportional fairness algorithm to determine each user equipment to be scheduled. For example, the real utility of each user equipment can be defined as the ratio of the utility function value of the current training round to the time average utility function value of all past training rounds. The first K users with the largest real utility are selected from all users in the cell / sector / cell group as the users scheduled in this training round, that is, the user scheduling result indication information corresponding to each scheduled user equipment is determined to be scheduled, where K is specific to the cell / sector / cell group, configured by the NW side, and can be included in the system message or NAS message.

[0384] In step 4, the scheduled user equipment uploads the local model update result (training result) to the edge node through the physical uplink shared channel (PUSCH, Physical Uplink Shared Channel).

[0385] Specifically, the physical uplink shared channel may be a flat Rayleigh fading channel, and the signal to interference plus noise ratio (SINR) may be calculated by the following formula:

[0386]

[0387] Among them, P i，t is the transmission power of the ith user (user equipment), h i，t Used to characterize the fading of small-scale flat Rayleigh channels, d i is the distance from the user equipment to the base station, Φ o，i represents the set of users that interfere with the i-th user, α represents the path loss coefficient, σ 2 represents the additive white Gaussian noise power.

[0388] In step 5, the edge node performs model aggregation based on the local update gradients in the training results uploaded by all scheduled users to obtain the model parameters of the updated global model.

[0389] Specifically, the base station updates the global model parameters based on a weighted summation method, where the weight is the amount of data of each scheduled user equipment, as shown in the following formula.

[0390]

[0391] in, represents the global model parameter of the mth cell / sector / cell group, η represents the learning rate, β i represents the ratio of the data size of the ith user to the data size of all users, λ i，t Indicates whether the i-th user successfully transmits the local update gradient in the t-th round of model training. If the transmission is successful, λ i，t =1, otherwise λ i，t =0.

[0392] In step 6, the edge node broadcasts the updated model parameters of the global model to all user equipments through a physical downlink shared channel (PDSCH).

[0393] If it is determined that the preset training completion condition is not met at present, steps 1 to 6 are repeated until the global model converges, that is, the preset training completion condition is met. The preset training completion condition may be a convergence condition, and determining whether the model meets the preset convergence condition may include at least one of the following:

[0394] The number of training rounds reaches the preset upper limit, the loss function of the global model on the validation set is less than or equal to the preset value, and the difference between the loss function of the global model on the validation set in two adjacent rounds is less than or equal to the preset value.

[0395] Specifically, in the same cell / sector / cell group, one or more models can be trained, and the training modes include at least one of the following: parallel mode: the training of multiple models is simultaneous or overlapping in time; serial mode: the training of multiple models is sequential in time and without overlap.

[0396] The following describes the model identification / function identification process based on hierarchical federated learning, which includes:

[0397] Step a: The network node performs model identification / function identification on the trained federated learning models on all cells / sectors / cell groups. The triggering method of the identification process includes at least one of the following: edge node triggering or network node triggering. The edge node triggering includes: the edge node sends a model / function identification request to the network node through the NG-C interface and reports the model capability. The model capability is the capability of the model on the edge node, which is used to indicate the model / function supported by the edge node; the network node triggering includes: the network node initiates a model / function identification request to the edge node through the NG-C interface, and the edge node responds to the request and reports the model capability. The model capability is the capability of the model on the edge node, which is used to indicate the model / function supported by the capability of the model on the edge node.

[0398] In step b, all edge nodes send the indicated federated learning model files to the network nodes through the NG-U interface according to the model / function ID for model deployment / inference / update / storage.

[0399] The following describes a specific implementation example in the model reasoning process based on hierarchical federated learning. The process by which a network node obtains information sent down by a model reasoning strategy may be that the network node may encapsulate the reasoning model / model group list and model file information in sequence in the order of “model / function ID of model 1, model file of model 1, model / function ID of model 2, model file of model 2,…” to obtain an encapsulation result, and determine that the encapsulation result is the model reasoning strategy sent down information. The encapsulation format of the encapsulation result is an IP message, and the format of the model reasoning strategy sent down information may be a binary executable file format specific to the user device.

[0400] In one embodiment, a 5G NR indoor positioning algorithm based on hierarchical federated learning is proposed. The user terminal is mainly responsible for updating the local model parameters, the base station is responsible for collecting local model parameter information, using the aggregation algorithm to aggregate the local parameters to update the global model, and then sending the new model parameters to the local terminal. The network node is responsible for sending the positioning strategy and resource control.

[0401] As an efficient positioning method for indoor scenes, AI technology can be applied to indoor industrial Internet scenes to improve the problem of inaccurate positioning of traditional methods in indoor heavy NLOS scenes. The distributed federated learning training architecture is relatively advanced. The advantages of the proposed algorithm include: high user privacy protection, collaborative positioning to increase positioning accuracy, reduced communication overhead, and faster training convergence.

[0402] As shown in Table 1, the positioning errors of different positioning schemes are recorded when the positioning accuracy is 50%, 67%, 80% and 90% percentiles:

[0403] Table 1

[0404] Positioning solution 50％ 67％ 80％ 90％ Traditional methods 11.89 13.62 14.78 16.36 AI+TOA 0.37 0.46 0.57 0.69 AI+DL-TDOA 0.38 0.49 0.59 0.73 AI+CIR 0.27 0.36 0.43 0.54 AI+RSRP+TOA 0.22 0.29 0.36 0.43 AI+RSRP+DL-TDOA 0.19 0.25 0.31 0.38

[0405] like Fig.11b As shown in Table 1, it can be the CDF curve of positioning accuracy of the AI ​​solution (CDF of positioning accuracy), which is the CDF corresponding to the positioning accuracy of the AI-based method. Fig.11b It can be seen from the results listed in that the AI-based method can significantly improve positioning accuracy. When the traditional positioning algorithm is used for UE coordinate prediction, the positioning error exceeds 10 meters at the 90% CDF percentile. The AI-based method can reduce the positioning error to within 1 meter. When multiple measurement information is combined as the input of the AI ​​model, the positioning accuracy can be further improved. Simulation results show that when RSRP is used to obtain features, the positioning error of 90% of users can be reduced to within 0.38 meters.

[0406] In the simulation setting, when 10 clients / base stations are online, the personalized federated learning algorithm FedProx and the federated average algorithm FedAvg are used as comparison algorithms. In addition, the error change trend of the Per-FedAvg algorithm is compared with that of the centralized training. The positioning error change trend is shown in Figure 2. Fig.11c As shown. Fig.11c It can be seen that compared with the centralized training algorithm, the method of using federated learning for distributed training has better convergence.

[0407] like Fig.11d As shown in the figure, using the CDF of the positioning error for comparison, the 90% percentile of the Per-FedAvg positioning error corresponds to an error of 0.22m, and there is no outstanding value from the distribution function. In the case of centralized training, the 90% percentile of the positioning error reached after 300 iterations corresponds to a value of 0.54m. Compared with the case of centralized training, distributed training significantly reduces the positioning error.

[0408] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages 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 to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0409] Based on the same inventive concept, the embodiment of the present application also provides a model reasoning device based on hierarchical federated learning for implementing the model management method based on hierarchical federated learning involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more embodiments of the model reasoning device based on hierarchical federated learning provided below can refer to the limitations of the model management method based on hierarchical federated learning above, and will not be repeated here.

[0410] In an exemplary embodiment, Fig.12 As shown, a model reasoning device 1200 based on hierarchical federated learning is provided, which is applied to a terminal device in a mobile communication system, wherein the mobile communication system also includes a network, wherein the network includes a network node and a plurality of base stations, wherein each of the base stations corresponds to a plurality of terminal devices, wherein the terminal devices are distributed nodes for performing hierarchical federated learning, and the base stations are central nodes for performing hierarchical federated learning. The device includes:

[0411] The first receiving module 1202 is used to receive a first configuration message sent by the base station, wherein the first configuration message includes user scheduling result indication information based on hierarchical federated learning, and the user scheduling result indication information is obtained by the base station based on user characteristic information of each of the terminal devices, and the user characteristic information includes at least one of user data characteristic information and user channel state information.

[0412] The first sending module 1204 is used to send reporting information to the network, and the network stores the hierarchical federated learning model and the delivery strategy generated by the base station. The reporting information includes at least one of the terminal capabilities, the reasoning task type, and the model performance monitoring feedback information. The reporting information is used for the network to generate a second configuration message, and the second configuration message is a model reasoning strategy delivery information or a model update strategy delivery information based on hierarchical federated learning.

[0413] The second receiving module 1206 is used to receive the second configuration message, and obtain a model inference result or a model update result based on the second configuration message.

[0414] In one of the embodiments, the terminal capability includes one or more of terminal computing capability, terminal channel status, and reasoning characteristics, the terminal computing capability is used to characterize the complexity of the model supported by the terminal device, and the terminal channel status characterizes the uplink channel status of the terminal device.

[0415] In one embodiment, the first sending module is specifically used for:

[0416] Based on the target transmission mode, reporting information is sent to the base station so that the base station sends reporting information to the network node, and the target transmission mode includes at least one of system message, RRC signaling, MAC CE signaling, DCI signaling, and NAS message.

[0417] In one of the embodiments, the model performance monitoring feedback information includes one or more of a system performance index, an accuracy index, a complexity index, and a generalization performance index.

[0418] In one embodiment, the device further comprises:

[0419] The first determination module is specifically used for: if it is determined that the model performance monitoring feedback information and the preset indicator threshold meet the preset update conditions, then sending a model update request to the base station so that the network responds to the model update request, and the model update request includes one or more of the model to be updated, the function identifier to be updated, and the reference information.

[0420] In one embodiment, the second receiving module is specifically configured to:

[0421] Receive a second configuration message sent by the base station based on a target transmission mode, where the second configuration message is sent by a network node to the base station, and the target transmission mode includes at least one of a system message, an RRC signaling, a MAC CE signaling, a DCI signaling, and a NAS message.

[0422] In one embodiment, the second configuration message is model update strategy delivery information, and the model update strategy delivery information includes the model to be updated and the update strategy. The second receiving module is specifically configured to:

[0423] The model to be updated is updated based on the update strategy to obtain a model update result, wherein the updated model includes the model to be updated / the model group to be updated, and the update strategy includes one or more of a local update strategy, a global update strategy, a model replacement strategy and a model deletion strategy.

[0424] In one embodiment, the second receiving module is specifically used for at least one of the following:

[0425] Some modules of the update model are updated to obtain the model update result.

[0426] Based on the target update algorithm, the model parameters of each model to be updated are updated to obtain a model update result. The target update algorithm includes one or more of a momentum gradient descent algorithm, a stochastic gradient descent algorithm, and a small batch gradient descent algorithm.

[0427] The model parameters of each of the models to be updated are updated based on the new model parameters to obtain a model update result.

[0428] Delete the model parameters and model content / function identifiers of the updated model to obtain the model update result.

[0429] In one embodiment, the second configuration message is model reasoning strategy delivery information, and the model reasoning strategy delivery information includes a model identifier of the model to be reasoned, a model file, and identification information of a terminal device to perform model reasoning. The second receiving module is specifically configured to:

[0430] Based on the model identification of the model to be inferred, the model file and the identification information of the terminal device on which the model inference is to be executed, model inference in an online mode or model inference in an offline mode is performed to obtain a model inference result.

[0431] In an exemplary embodiment, Fig.13 As shown, a model reasoning device 1300 based on hierarchical federated learning is provided, which is applied to a network node in a mobile communication system, wherein the network includes a network node and a plurality of base stations, each of which corresponds to a plurality of terminal devices, and the device includes:

[0432] The third receiving module 1302 is used to receive the reporting information sent by the terminal device through the base station, and the reporting information includes at least one of terminal capabilities, reasoning task types, and model performance monitoring feedback information.

[0433] The first generation module 1304 is configured to generate a second configuration message based on the reported information, and send the second configuration message to the terminal device through the base station. The second configuration message is model reasoning strategy delivery information or model update strategy delivery information based on hierarchical federated learning, and the second configuration message is used to enable the terminal device to obtain a model reasoning result or a model update result.

[0434] In one of the embodiments, the terminal capability includes one or more of terminal computing capability, terminal channel status, and reasoning characteristics, the terminal computing capability is used to characterize the complexity of the model supported by the terminal device, and the terminal channel status characterizes the uplink channel status on the terminal device side.

[0435] In one embodiment, the third receiving module is specifically used for:

[0436] Based on the target transmission mode, the terminal capabilities reported by the terminal device through the base station are received, and the target transmission mode includes at least one of system messages, RRC signaling, MAC CE signaling, DCI signaling, and NAS messages.

[0437] In one embodiment, the device further comprises:

[0438] The fourth receiving module is used to receive the model inference request signaling sent by the terminal device through the base station, and return a response message corresponding to the inference request, so that the terminal device sends reporting information based on the response message.

[0439] In one embodiment, the second configuration message includes model reasoning strategy delivery information or model update strategy delivery information based on hierarchical federated learning. The first generation module is specifically used to:

[0440] Based on the reported information, at least one model to be inferred whose target dimension satisfies a preset matching condition is determined among multiple first models.

[0441] Based on the model identifier of the model to be inferred, the model file, and the identification information of the terminal device on which the model inference is to be executed, model inference strategy delivery information based on hierarchical federated learning is generated.

[0442] In one of the embodiments, the first generation module is specifically used to: based on the reported information, determine a preset number of models to be inferred from multiple first models whose model structures, model complexity and model accuracy meet preset matching conditions.

[0443] In one of the embodiments, sending the second configuration message to the terminal device by the base station includes:

[0444] The second configuration message is sent to the base station through the user plane interface, so that the base station sends the second configuration message to the terminal device through the target transmission mode, and the target transmission mode includes at least one of system message, RRC signaling, MAC CE signaling, DCI signaling, and NAS message.

[0445] In one of the embodiments, the model performance monitoring feedback information includes one or more of a system performance index, an accuracy index, a complexity index, and a generalization performance index.

[0446] In one of the embodiments, the reported information includes model performance monitoring feedback information, and the first generating module is specifically used to receive model performance monitoring feedback information.

[0447] A model update request is received, wherein the model update request includes an identifier of a model to be updated / a function to be updated and reference information, and the model update request is generated when the model performance monitoring feedback information of the terminal device and a preset indicator threshold meet a preset update condition.

[0448] Respond to the model update request and generate a second configuration message.

[0449] In one of the embodiments, the reported information includes model performance monitoring feedback information, and the first generating module is specifically used to receive model performance monitoring feedback information.

[0450] Based on a preset time period, a model update event is triggered. Alternatively, if the model performance monitoring feedback information and a preset indicator threshold meet a preset update condition, an update event is triggered. The preset time period is associated with the terminal capability and the auxiliary information.

[0451] Based on the triggered update event, a second configuration message is generated.

[0452] In one embodiment, the second configuration message includes information on issuing a model reasoning strategy based on hierarchical federated learning. The first generation module is specifically used to: determine at least one model to be updated.

[0453] Based on the model to be updated, the identification information of the terminal device on which the model update is to be executed, and the update strategy, model reasoning strategy delivery information based on hierarchical federated learning is generated.

[0454] In one of the embodiments, the update strategy includes one or more of a local update strategy, a global update strategy, a model replacement strategy, and a model deletion strategy.

[0455] In one embodiment, the device further comprises:

[0456] The fifth receiving module is used to receive the model data sent by each base station, and the model data includes a model file and a model content / function identifier.

[0457] An arrangement module is used to determine each of the first models based on each of the model data, and arrange them according to the evaluation index of each of the first models to obtain a first model list.

[0458] In one embodiment, the device further comprises:

[0459] The sixth receiving module is used to receive an identification request and a model capability sent by a base station, wherein the identification request includes a model identification request and a function identification request, and the model capability is used to indicate the model / supported function supported by the base station.

[0460] In one embodiment, the device further comprises:

[0461] The second sending module is used to send an identification request to each base station so that each base station reports a model capability, and the model capability is used to indicate the model / supported function supported by the base station.

[0462] In an exemplary embodiment, Fig.14 As shown, a model reasoning device 1400 based on hierarchical federated learning is provided, which is applied to a base station in a mobile communication system, and the device includes:

[0463] The third sending module 1402 is used to obtain user scheduling result indication information corresponding to each terminal device based on the user characteristic information of each terminal device, and send a first configuration message to each terminal device, wherein the user characteristic information includes at least one of user data characteristic information and user channel state information, and the first configuration message includes the user scheduling result indication information;

[0464] The seventh receiving module 1404 is used to receive the reporting information sent by the terminal device, where the reporting information includes at least one of the terminal capability, the reasoning task type, and the model performance monitoring feedback information;

[0465] The fourth sending module 1406 is used to send the reported information to the network node so that the network node obtains a second configuration message; the second configuration message is a model reasoning strategy issuance information or a model update strategy issuance information based on hierarchical federated learning, and the second configuration message is used to enable the terminal device to obtain the model reasoning result or the model update result.

[0466] In one embodiment, the device further comprises:

[0467] The fifth sending module is used to send model data to the network node, where the model data includes a model file and a model content / function identifier, so that the network node determines each first model based on each model data and obtains a first model list.

[0468] In one embodiment, the device further comprises:

[0469] The sixth sending module is used to send an identification request and a model capability to the network node, wherein the identification request includes a model identification request and a function identification request, and the model capability is used to indicate the model / supported function supported by the base station.

[0470] In one embodiment, the device further comprises:

[0471] An eighth receiving module is used to receive an identification request sent by the network node, and report model capabilities to the network node based on the identification request, where the model capabilities are used to indicate the models / supported functions supported by the base station.

[0472] In one embodiment, the device further comprises:

[0473] The seventh sending module is used to obtain a trained global model based on the model training results sent by each of the terminal devices.

[0474] In one embodiment, the seventh sending module is specifically used to: for the i-th round of training, based on the training results sent by each of the terminal devices in the i-1th round, obtain the model parameters of the global model of the i-th round;

[0475] When the preset training completion conditions are not met, the model parameters of the global model of the i-th round are broadcast to each of the terminal devices, so that each of the terminal devices performs training based on the model parameters of the global model of the i-th round, until the base station determines that the preset training completion conditions are met, thereby obtaining a trained global model.

[0476] In one embodiment, the preset training completion conditions include one or more of the number of training rounds reaching a preset upper limit, the loss function of the global model being less than or equal to a preset value, and the difference between the loss function of the global model in two adjacent rounds being less than or equal to a preset value.

[0477] In one embodiment, the device further comprises:

[0478] An indication information determination module, configured to obtain user scheduling result indication information corresponding to each of the terminals based on user characteristic information of each of the terminal devices, wherein the user characteristic information includes at least one of user data characteristic information and user channel state information;

[0479] A first configuration message is sent to each of the terminal devices respectively, wherein the first configuration message includes the user scheduling result indication information.

[0480] In one of the embodiments, the indication information determination module is specifically used to obtain a pre-configured number of scheduled users;

[0481] Among all users within the coverage of the base station, based on at least one of the user data characteristic information and user channel state information of each terminal device, the terminal devices with the scheduled number of users are screened out, and the user scheduling result indication information is determined as the scheduled indication information.

[0482] Each module in the above-mentioned model reasoning device based on hierarchical federated learning can be implemented in whole or in part by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0483] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Fig.15 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, 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. Among them, the processor of the computer device is used 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 the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data for model lifecycle management. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a model management method based on hierarchical federated learning is implemented.

[0484] Those skilled in the art will understand that Fig.15 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0485] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps described in the above embodiment when executing the computer program.

[0486] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps described in the above embodiment are implemented.

[0487] In one embodiment, a computer program product is provided, including a computer program, which implements the steps described in the above embodiment when executed by a processor.

[0488] It should be noted that the user information (including but not limited to terminal device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0489] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and 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 embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0490] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0491] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations 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 shall be subject to the attached claims.

Claims

1. A model management method based on hierarchical federated learning, characterized in that: A terminal device applied to a mobile communication system, wherein the mobile communication system further comprises a network, wherein the network comprises a network node and a plurality of base stations, each of the base stations corresponds to a plurality of terminal devices, wherein the terminal devices are distributed nodes for performing hierarchical federated learning, and the base stations are central nodes for performing hierarchical federated learning; and the method comprises: receiving a first configuration message sent by the base station, where the first configuration message includes user scheduling result indication information based on hierarchical federated learning, where the user scheduling result indication information is obtained by the base station based on user feature information of each of the terminal devices, where the user feature information includes at least one of user data feature information and user channel state information; Sending reporting information to the network, the network storing the hierarchical federated learning model and the delivery strategy generated by the base station, the reporting information including at least one of terminal capabilities, reasoning task types, and model performance monitoring feedback information, the reporting information is used by the network to generate a second configuration message, and the second configuration message is model reasoning strategy delivery information or model update strategy delivery information based on hierarchical federated learning; The second configuration message is received, and a model inference result or a model update result is obtained based on the second configuration message.

2. The method according to claim 1, characterized in that The terminal capability includes one or more of terminal computing capability, terminal channel status, and reasoning characteristics. The terminal computing capability is used to characterize the complexity of the model supported by the terminal device, and the terminal channel status characterizes the uplink channel status of the terminal device.

3. The method according to claim 1, characterized in that The sending the reporting information to the network includes: Based on the target transmission mode, reporting information is sent to the base station so that the base station sends reporting information to the network node, and the target transmission mode includes at least one of system message, RRC signaling, MAC CE signaling, DCI signaling, and NAS message.

4. The method according to claim 1, characterized in that: The model performance monitoring feedback information includes one or more of a system performance index, an accuracy index, a complexity index, and a generalization performance index.

5. The method according to claim 4, characterized in that The method further comprises: If it is determined that the model performance monitoring feedback information and the preset indicator threshold meet the preset update conditions, a model update request is sent to the base station so that the network responds to the model update request, and the model update request includes one or more of the model to be updated, the function identifier to be updated, and the reference information.

6. The method according to claim 1, characterized in that The receiving the second configuration message includes: Receive a second configuration message sent by the base station based on a target transmission mode, where the second configuration message is sent by a network node to the base station, and the target transmission mode includes at least one of a system message, an RRC signaling, a MAC CE signaling, a DCI signaling, and a NAS message.

7. The method according to claim 1, characterized in that The second configuration message is model update strategy delivery information, and the model update strategy delivery information includes a model to be updated and an update strategy; and obtaining a model inference result or a model update result based on the second configuration message includes: The model to be updated is updated based on the update strategy to obtain a model update result, wherein the updated model includes the model to be updated / the model group to be updated, and the update strategy includes one or more of a local update strategy, a global update strategy, a model replacement strategy and a model deletion strategy.

8. The method according to claim 7, characterized in that The process of performing the update process includes at least one of the following: Update some modules of the update model to obtain the model update result; Based on the target update algorithm, the model parameters of each of the models to be updated are updated to obtain a model update result, wherein the target update algorithm includes one or more of a momentum gradient descent algorithm, a stochastic gradient descent algorithm, and a mini-batch gradient descent algorithm; Based on the new model parameters, the model parameters of each model to be updated are updated to obtain a model update result; Delete the model parameters and model content / function identifiers of the updated model to obtain the model update result.

9. The method according to claim 1, characterized in that: The second configuration message is model reasoning strategy delivery information, and the model reasoning strategy delivery information includes a model identifier of the model to be reasoned, a model file, and identification information of a terminal device on which model reasoning is to be performed; The obtaining a model inference result or a model update result based on the second configuration message includes: Based on the model identification of the model to be inferred, the model file and the identification information of the terminal device on which the model inference is to be executed, model inference in an online mode or model inference in an offline mode is performed to obtain a model inference result.

10. A model management method based on hierarchical federated learning, characterized in that: A network node applied to a mobile communication system, the network comprising a network node and a plurality of base stations, each of the base stations corresponding to a plurality of terminal devices, the method comprising: Receiving reporting information sent by the terminal device through the base station, the reporting information including at least one of terminal capability, reasoning task type, and model performance monitoring feedback information; A second configuration message is generated based on the reported information, and the second configuration message is sent to the terminal device through the base station; the second configuration message is a model reasoning strategy based on hierarchical federated learning or a model update strategy, and the second configuration message is used to enable the terminal device to obtain a model reasoning result or a model update result.

11. The method according to claim 10, characterized in that The terminal capability includes one or more of terminal computing capability, terminal channel status, and reasoning characteristics. The terminal computing capability is used to characterize the complexity of the model supported by the terminal device, and the terminal channel status characterizes the uplink channel status on the terminal device side.

12. The method according to claim 10, characterized in that The reporting information includes the terminal capability, and the receiving the reporting information sent by the terminal device through the base station includes: Based on the target transmission mode, the terminal capabilities reported by the terminal device through the base station are received, and the target transmission mode includes at least one of system messages, RRC signaling, MAC CE signaling, DCI signaling, and NAS messages.

13. The method according to claim 10, characterized in that The method further comprises: Receive the model inference request signaling sent by the terminal device through the base station, and return a response message corresponding to the inference request, so that the terminal device sends reporting information based on the response message.

14. The method according to claim 10, characterized in that The second configuration message includes model reasoning strategy delivery information or model update strategy delivery information based on hierarchical federated learning; and the generating of the second configuration message based on the reported information includes: Based on the reported information, determining at least one model to be inferred whose target dimension satisfies a preset matching condition among the multiple first models; Based on the model identifier of the model to be inferred, the model file, and the identification information of the terminal device on which the model inference is to be executed, model inference strategy delivery information based on hierarchical federated learning is generated.

15. The method according to claim 11, characterized in that The step of determining, based on the reported information, at least one model to be inferred whose target dimension satisfies a preset matching condition among the multiple first models includes: Based on the reported information, a preset number of models to be inferred whose model structures, model complexity and model accuracy meet preset matching conditions are determined among the multiple first models.

16. The method according to claim 14, characterized in that The sending the second configuration message to the terminal device by the base station includes: The second configuration message is sent to the base station through the user plane interface, so that the base station sends the second configuration message to the terminal device through the target transmission mode, and the target transmission mode includes at least one of system message, RRC signaling, MAC CE signaling, DCI signaling, and NAS message.

17. The method according to claim 14, characterized in that The model performance monitoring feedback information includes one or more of a system performance index, an accuracy index, a complexity index, and a generalization performance index.

18. The method according to claim 14, characterized in that The reported information includes model performance monitoring feedback information, and the generating a second configuration message based on the reported information includes: Receive model performance monitoring feedback; Receive a model update request, the model update request including a to-be-updated model / to-be-updated function identifier and reference information, the model update request being generated when the model performance monitoring feedback information of the terminal device and a preset indicator threshold satisfy a preset update condition; Respond to the model update request and generate a second configuration message.

19. The method according to claim 14, characterized in that The reported information includes model performance monitoring feedback information, and the generating a second configuration message based on the reported information includes: Receive model performance monitoring feedback; Based on a preset time period, a model update event is triggered; or, if the model performance monitoring feedback information and the preset indicator threshold meet the preset update condition, an update event is triggered; the preset time period is associated with the terminal capability and the auxiliary information; Based on the triggered update event, a second configuration message is generated.

20. The method according to claim 18 or 19, characterized in that: The second configuration message includes model reasoning strategy delivery information based on hierarchical federated learning; The generating the second configuration message comprises: determining at least one model to be updated; Based on the model to be updated, the identification information of the terminal device on which the model update is to be executed, and the update strategy, model reasoning strategy delivery information based on hierarchical federated learning is generated.

21. The method according to claim 20, characterized in that The update strategy includes one or more of a local update strategy, a global update strategy, a model replacement strategy, and a model deletion strategy.

22. The method according to claim 18, characterized in that The method further comprises: Receiving model data sent by each of the base stations, the model data including a model file and a model content / function identifier; The first models are determined based on the model data, and are arranged according to evaluation indicators of the first models to obtain a first model list.

23. The method according to claim 22, characterized in that The method further comprises: An identification request and a model capability sent by a base station are received, wherein the identification request includes a model identification request and a function identification request, and the model capability is used to indicate the models / supported functions supported by the base station.

24. The method according to claim 22, characterized in that The method further comprises: An identification request is sent to each of the base stations so that each of the base stations reports a model capability, where the model capability is used to indicate a model / supported function supported by the base station.

25. A model management method based on hierarchical federated learning, characterized in that: A base station applied to a mobile communication system, the mobile communication system comprising a network and a plurality of base stations, each of the base stations corresponding to a plurality of terminal devices, the terminal devices being distributed nodes for performing hierarchical federated learning, and the base stations being central nodes for performing hierarchical federated learning; the method comprising: Based on the user characteristic information of each terminal device, obtain the user scheduling result indication information corresponding to each terminal device, and send a first configuration message to each terminal device, wherein the user characteristic information includes at least one of user data characteristic information and user channel state information, and the first configuration message includes the user scheduling result indication information; Receiving reporting information sent by the terminal device, the reporting information including at least one of terminal capability, reasoning task type, and model performance monitoring feedback information; The reported information is sent to the network node so that the network node obtains a second configuration message; the second configuration message is a model reasoning strategy issuance information or a model update strategy issuance information based on hierarchical federated learning, and the second configuration message is used to enable the terminal device to obtain a model reasoning result or a model update result.

26. The method according to claim 25, characterized in that The method further comprises: Model data is sent to the network node, where the model data includes a model file and a model content / function identifier, so that the network node determines each first model based on each model data and obtains a first model list.

27. The method according to claim 25, characterized in that The method further comprises: An identification request and a model capability are sent to the network node, wherein the identification request includes a model identification request and a function identification request, and the model capability is used to indicate the model / supported function supported by the base station.

28. The method according to claim 25, characterized in that The method further comprises: An identification request sent by the network node is received, and a model capability is reported to the network node based on the identification request, where the model capability is used to indicate the model / supported function supported by the base station.

29. The method according to claim 25, characterized in that The method further comprises: Based on the model training results sent by each of the terminal devices, a trained global model is obtained.

30. The method according to claim 29, characterized in that The terminal device is a terminal device of a scheduling user, and obtaining a trained global model based on the model training results sent by each of the terminal devices includes: For the i-th round of training, based on the training results sent by each of the terminal devices in the i-1-th round, obtain the model parameters of the global model of the i-th round; When the preset training completion conditions are not met, the model parameters of the global model of the i-th round are broadcast to each of the terminal devices, so that each of the terminal devices performs training based on the model parameters of the global model of the i-th round, until the base station determines that the preset training completion conditions are met, thereby obtaining a trained global model.

31. The method according to claim 30, characterized in that The preset training completion conditions include one or more of the number of training rounds reaching a preset upper limit, the loss function of the global model being less than or equal to a preset value, and the difference between the loss function of the global model in two adjacent rounds being less than or equal to a preset value.

32. The method according to claim 25, characterized in that The method further comprises: Obtaining user scheduling result indication information corresponding to each of the terminals based on user characteristic information of each of the terminal devices, wherein the user characteristic information includes at least one of user data characteristic information and user channel state information; A first configuration message is sent to each of the terminal devices respectively, wherein the first configuration message includes the user scheduling result indication information.

33. The method according to claim 32, characterized in that The obtaining the user scheduling result indication information corresponding to each of the terminals based on the user characteristic information of each of the terminal devices includes: Get the number of pre-configured scheduled users; Among all users within the coverage of the base station, based on at least one of the user data characteristic information and user channel state information of each terminal device, the terminal devices with the scheduled number of users are screened out, and the user scheduling result indication information is determined as the scheduled indication information.

34. A model reasoning device based on hierarchical federated learning, characterized in that: A terminal device applied to a mobile communication system, wherein the mobile communication system further comprises a network, wherein the network comprises a network node and a plurality of base stations, wherein each of the base stations corresponds to a plurality of terminal devices, wherein the terminal devices are distributed nodes for performing hierarchical federated learning, and the base stations are central nodes for performing hierarchical federated learning; the device comprises: A first receiving module is configured to receive a first configuration message sent by the base station, where the first configuration message includes user scheduling result indication information based on hierarchical federated learning, where the user scheduling result indication information is obtained by the base station based on user feature information of each of the terminal devices, where the user feature information includes at least one of user data feature information and user channel state information; A first sending module, configured to send reporting information to the network, wherein the network stores the hierarchical federated learning model and the issuing strategy generated by the base station, wherein the reporting information includes at least one of terminal capabilities, reasoning task types, and model performance monitoring feedback information, and the reporting information is used by the network to generate a second configuration message, wherein the second configuration message is model reasoning strategy issuing information or model update strategy issuing information based on hierarchical federated learning; The second receiving module is used to receive the second configuration message and obtain a model inference result or a model update result based on the second configuration message.

35. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 33 are implemented.

36. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 33 are implemented.

37. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 33 are implemented.

Images