First access network equipment and terminal equipment

Abstract

The application discloses a first access network device and a terminal device wherein the first access network device comprises a first transceiver; and a first processor coupled to the first transceiver; the first processor is configured to: receive, via the first transceiver, a first message sent by the terminal device, the first message comprising model training capability information, the first message comprising at least one of: a terminal capability information message, a terminal capability response message; based on the first message, model training strategy indication information is determined, and the model training strategy indication information is used for indicating the terminal equipment to upload training data or upload a locally updated model; sending a media access control element (MAC CE) to the terminal device via the first transceiver, wherein the MAC CE comprises model training strategy indication information; and receiving model data sent by the terminal equipment through the first transceiver, wherein the model data is data which is determined based on the model training strategy indication information and is related to the update of the first model.

Description

Technical Field

[0001] This application relates to, but is not limited to, the field of mobile communication technologies, and particularly relates to a first access network device and a terminal device. Background Art

[0002] With the development of the intelligence of terminal devices, artificial intelligence (AI) has witnessed an explosive development. More and more terminal data is sent to the server to form its training dataset. However, the increase in the amount of data in the dataset will not only affect the communication quality of other terminal devices, but also reduce the accuracy of the finally trained model. Summary of the Invention

[0003] In view of this, embodiments of this application at least provide a first access network device and a terminal device.

[0004] The technical solution of the embodiments of this application is implemented as follows:

[0005] Embodiments of this application provide a first access network device, which includes a first transceiver; and

[0006] a first processor coupled to the first transceiver; the first processor is configured to:

[0007] receive, via the first transceiver, a first message sent by a terminal device, the first message including model training capability information, and the first message including at least one of the following: a terminal capability information message, a terminal capability response message;

[0008] determine, based on the first message, model training policy indication information for indicating that the terminal device uploads training data or uploads a locally updated model;

[0009] send, via the first transceiver, a media access control control element (MAC CE) to the terminal device, the MAC CE including the model training policy indication information;

[0010] receive, via the first transceiver, model data sent by the terminal device, the model data being data determined based on the model training policy indication information and related to the update of a first model.

[0011] In some embodiments, the model training capability information includes at least one of the following:

[0012] the size of the local training data of the terminal device;

[0013] the computing frequency of the terminal device;

[0014] the central processing unit (CPU) cycles of the terminal device.

[0015] In some embodiments, the first message further includes at least one of the following of the terminal device: Reference Signal Received Power (RSRP), Reference Signal Received Quality (RSRQ).

[0016] In some embodiments, the first processor is configured to:

[0017] Receive, via the first transceiver, the model training requirement information sent by the core network device;

[0018] Determine the model training strategy indication information based on the model training requirement information and the first message.

[0019] In some embodiments, the model training requirement information includes at least one of the following: device list, model training time delay information of the model training task, model accuracy information of the model training task;

[0020] Wherein, the device list includes at least one terminal device participating in the model training task;

[0021] The model training task includes multiple iterative trainings, and the training of the first model is the Nth iterative training in the multiple iterative trainings.

[0022] In some embodiments, the MAC CE further includes at least one of the following:

[0023] First power information, which is used to indicate the transmit power when the terminal device uploads training data;

[0024] Second power information, which is used to indicate the transmit power when the terminal device uploads the model;

[0025] Time information, which is used to indicate the application time of the model training strategy;

[0026] Time indication information, which is used to indicate the representation method adopted by the time information;

[0027] The first model.

[0028] In some embodiments, the time information includes at least one of the following:

[0029] Start time and end time;

[0030] Duration.

[0031] In some embodiments, the time indication information is used to indicate the representation method adopted by the start time and the end time, and the representation methods of the start time and the end time include one of the following:

[0032] Absolute timestamp;

[0033] A time offset relative to a first time, where the first time is the time when the terminal device receives the model training policy indication information.

[0034] In some embodiments, the first processor is configured to:

[0035] Send a second message to the terminal device via the first transceiver, where the second message is used to trigger the sending of the first message, and the second message includes one of the following: a terminal capability query message, a terminal capability request message.

[0036] In some embodiments, the second message includes at least one of the following:

[0037] Model training capability request information;

[0038] A first threshold, where the first threshold is used to control the sending of the first message based on the size of the training data of the terminal device;

[0039] A first model.

[0040] In some embodiments, the first processor is configured to:

[0041] When the model training policy instructs the terminal device to upload training data, receive the training data uploaded by the terminal device via the first transceiver;

[0042] Update the first model based on the training data to obtain a second model.

[0043] In some embodiments, the first processor is configured to:

[0044] When the model training policy instructs the terminal device to upload the locally updated model, receive the second model uploaded by the terminal device via the first transceiver;

[0045] The second model is obtained by the terminal device updating the first model based on the training data.

[0046] In some embodiments, the first processor is configured to:

[0047] Receive the local models sent by each of one or more second access network devices via the first transceiver;

[0048] Determine a global model based on the local models sent by each of one or more second access network devices and the local model of the first access network device, where the local model of the first access network device is determined based on model data;

[0049] Send the global model to the terminal device via the first transceiver.

[0050] In some embodiments, the terminal device is served jointly by multiple distributed access network devices, and the multiple distributed access network devices include a first access network device and a third access network device;

[0051] The first processor is configured to:

[0052] Receive, via the first transceiver, model training configuration information sent by the core network device, where the model training configuration information indicates that the first access network device participates in the Nth iterative training of the model training task;

[0053] Wherein, the distributed access network device participating in the Mth iterative training of the model training task is the first access network device or the third access network device.

[0054] This application provides a terminal device, which includes a second transceiver; and

[0055] A second processor, which is coupled to the second transceiver; the second processor is configured to:

[0056] Send, via the second transceiver, a first message to the first access network device, where the first message includes model training capability information, and the first message includes at least one of the following: a terminal capability information message, a terminal capability response message;

[0057] Receive, via the second transceiver, a MAC CE sent by the first access network device, where the MAC CE includes model training policy indication information, and the model training indication information is determined based on the first message, and the model training policy indication information is used to instruct the terminal device to upload training data or upload the locally updated model;

[0058] Determine model data based on the model training policy indication information, where the model data is related to the update of the first model;

[0059] Send the model data to the first access network device via the second transceiver.

[0060] In some embodiments, the second processor is configured to:

[0061] Receive, via the second transceiver, a second message sent by the first access network device, where the second message is used to trigger the sending of the first message, and the second message includes one of the following: a terminal capability query message, a terminal capability request message.

[0062] In some embodiments, the second processor is configured to:

[0063] Update the first model based on the training data to obtain a second model;

[0064] Send the second model to the first access network device via the second transceiver.

[0065] In some embodiments, the second processor is configured to:

[0066] Training data sent to the first access network device via the second transceiver.

[0067] This application provides a wireless communication method, which is applied to a first access network device. The method includes:

[0068] Receiving a first message sent by a terminal device, the first message includes model training capability information, and the first message includes at least one of the following: a terminal capability information message, a terminal capability response message;

[0069] Based on the first message, determining model training policy indication information, where the model training policy indication information is used to instruct the terminal device to upload training data or upload a locally updated model;

[0070] Sending a Media Access Control Control Element (MAC CE) to the terminal device, where the MAC CE includes the model training policy indication information;

[0071] Receiving model data sent by the terminal device, where the model data is determined based on the model training policy indication information and is related to the update of the first model.

[0072] This application provides a wireless communication method, which is applied to a terminal device. The method includes:

[0073] Sending a first message to the first access network device, the first message includes model training capability information, and the first message includes at least one of the following: a terminal capability information message, a terminal capability response message (;

[0074] Receiving a MAC CE sent by the first access network device, the MAC CE includes model training policy indication information, the model training indication information is determined based on the first message, and the model training policy indication information is used to instruct the terminal device to upload training data or upload a locally updated model;

[0075] Based on the model training policy indication information, determining model data, where the model data is related to the update of the first model;

[0076] Sending the model data to the first access network device.

[0077] This application also provides a storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned wireless communication method is implemented.

[0078] In the embodiments of the present application, based on the model training capability information of the terminal device, model training strategy indication information is determined to control the terminal device to upload training data or the locally updated model. In this way, it is not necessary for all terminal devices to upload training data, reducing the amount of data uploaded to the first access network device and solving the problem in the prior art of the impact on the communication quality of other terminal devices due to the upload of a large amount of training data. On the other hand, the model training strategy indication information in the present application is determined based on the model training capability information of the terminal device. In this way, it is possible to instruct the terminal device with stronger computing power to upload the locally updated model and instruct the terminal device with weaker computing power to upload training data. By jointly training the model by the terminal device and the first access network device, the training duration of the model trained alone by the first access network device is reduced, the efficiency of model training is improved, the problem of delay in the model training task in the prior art is solved, and thus the computing accuracy of the model training task is improved. Description of the Drawings

[0079] Figure 1 It is an optional flowchart of the wireless communication method provided by the embodiments of the present application;

[0080] Figure 2 It is an optional flowchart of the wireless communication method provided by the embodiments of the present application;

[0081] Figure 3 It is an optional flowchart of the wireless communication method provided by the embodiments of the present application;

[0082] Figure 4 It is an optional schematic diagram of cell-free MIMO provided by the embodiments of the present application;

[0083] Figure 5 It is an optional schematic diagram of the wireless communication system provided by the embodiments of the present application;

[0084] Figure 6 It is an optional flowchart of the wireless communication method provided by the embodiments of the present application;

[0085] Figure 7 It is an optional flowchart of the wireless communication method provided by the embodiments of the present application;

[0086] Figure 8 It is an optional flowchart of the wireless communication method provided by the embodiments of the present application;

[0087] Figure 9 It is an optional flowchart of the wireless communication method provided by the embodiments of the present application;

[0088] Figure 10An optional schematic diagram of the MAC CE provided by an embodiment of the present application;

[0089] Figure 11 A schematic diagram of the hardware entity of a device provided by an embodiment of the present application. Detailed implementation manners

[0090] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following will further describe the specific technical solutions of the application in detail with reference to the accompanying drawings in the embodiments of the present application. The following embodiments are used to illustrate the present application but are not intended to limit the scope of the present application.

[0091] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments. However, it can be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.

[0092] In the following description, the terms "first / second / third" are only used to distinguish different objects, and do not represent a specific order for the objects, and there is no limitation on the order. It can be understood that "first / second / third" can be interchanged with a specific order or sequence when allowed, so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0093] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0094] In a first aspect, a wireless communication method provided by an embodiment of the present application is applied to a first access network device, such as Figure 1 shown, and includes:

[0095] S101. Receive a first message sent by a terminal device, where the first message includes model training capability information, and the first message includes at least one of the following: a terminal capability information message, a terminal capability response message.

[0096] S102. Determine model training policy indication information based on the first message, where the model training policy indication information is used to instruct the terminal device to upload training data or upload a locally updated model.

[0097] S103. Send a media access control control element (MAC CE) to the terminal device, where the MAC CE includes the model training policy indication information.

[0098] S104. Receive the model data sent by the terminal device. The model data is data determined based on the model training policy indication information and related to the update of the first model.

[0099] In a second aspect, a wireless communication method provided by an embodiment of the present application is applied to a terminal device. As Figure 2 shown, it includes:

[0100] S201. Send a first message to the first access network device. The first message includes model training capability information, and the first message includes at least one of the following: terminal capability information message, terminal capability response message.

[0101] S202. Receive the MAC CE sent by the first access network device. The MAC CE includes model training policy indication information. The model training indication information is determined based on the first message, and the model training policy indication information is used to instruct the terminal device to upload training data or upload the locally updated model.

[0102] S203. Determine model data based on the model training policy indication information. The model data is related to the update of the first model.

[0103] S204. Send the model data to the first access network device.

[0104] In a third aspect, a wireless communication method provided by an embodiment of the present application is applied to a wireless communication system including a first access network device and a terminal device. As Figure 3 shown, it includes:

[0105] S301. The terminal device sends a first message to the first access network device. The first message includes model training capability information, and the first message includes at least one of the following: terminal capability information message, terminal capability response message;

[0106] S302. The first access network device determines model training policy indication information based on the first message. The model training policy indication information is used to instruct the terminal device to upload training data or upload the locally updated model;

[0107] S303. The first access network device sends the MAC CE to the terminal device. The MAC CE includes model training policy indication information.

[0108] S304. The terminal device determines model data based on the model training policy indication information included in the MAC CE. The model data is related to the update of the first model.

[0109] S305. The terminal device sends the model data to the first access network device.

[0110] Next, for Figure 1 , Figure 2 orFigure 3 The described wireless communication method will be described.

[0111] Here, the terminal capability information message refers to the message in which the terminal device reports its hardware and software capabilities to the network. The terminal capability response message is the terminal capability response message sent by the terminal to the network after receiving the terminal capability query message sent by the network. This terminal capability response message is used to reply to the network with the terminal's capability information. Among them, both the terminal capability information message and the terminal capability response message are Radio Resource Control (RRC) messages.

[0112] In some embodiments, the first access network device may be a base station, or any device such as a management node used for communicating with the terminal device.

[0113] In some embodiments, the terminal device may be a mobile phone, or any device such as an Internet of Things device that is connected to the communication network and realizes functions such as data transmission.

[0114] In some embodiments, the first model may be a model to be trained, and the first model may be pre-stored in the model base of the first access network device. Among them, when the model training task indicates that the current is the first iteration, the model to be trained is the initial model; when the model training task indicates that the current is not the first iteration, the model to be trained is the model generated after the previous iteration.

[0115] In some embodiments, the communication interaction between the first access network device and the terminal device can be realized through the forwarding plane in the first access network device.

[0116] In some embodiments, the model training strategy indication information may be determined by the Distributed Computing Unit (DCU) in the computing plane of the first access network device.

[0117] In some embodiments, the model training capability information includes at least one of the following: the size of the local training data of the terminal device; the computing frequency of the terminal device; the Central Processing Unit (CPU) cycles of the terminal device.

[0118] Here, the computing frequency of the terminal device refers to the clock frequency at which the Central Processing Unit (CPU) of the terminal device operates, that is, the clock oscillation frequency of the CPU. The CPU cycles of the terminal device refer to the time required for the terminal device to calculate the local training data once. The size of the local training data of the terminal device refers to the size of the data set used by the terminal device for training the model locally.

[0119] It can be understood that based on the computing frequency of the terminal device, the CPU cycles of the terminal device, and the size of the local training data, the computing ability of the terminal device to perform model training based on the size of the local training data can be determined. In the case where the computing ability of the terminal device is relatively good, performing model training through the terminal device can not only relieve the operating pressure of the first access network device, but also improve the efficiency of the first access network device for model training and reduce the latency of model training. In the case where the computing ability of the terminal device is relatively weak, if the terminal device performs model training, it will take a long time to wait. In this scenario, the local training data can be uploaded to the first access network device, and the first access network device can perform model training, which can improve the efficiency of model training and reduce the latency of model training.

[0120] In some embodiments, the first access network device may obtain model training ability information through a terminal capability information message and / or a terminal capability response message.

[0121] In one example, the terminal device sends a terminal capability information message to the first access network device. The terminal capability information message includes an indicator for indicating that the terminal device reports model training ability, and the indicator contains at least one of the following parameter information: computing frequency, CPU cycles, size of local training data.

[0122] In one example, the terminal device sends a terminal capability response message to the first access network device. The terminal capability information message includes an indicator for indicating that the terminal device reports model training ability, and the indicator contains at least one of the following parameter information: computing frequency, CPU cycles, size of local training data.

[0123] In one example, the terminal device sends a terminal capability response message and a terminal capability information message to the first access network device. Among them, the terminal capability information message contains at least one of the following parameter information: computing frequency, CPU cycles, and the terminal capability response message contains the size of local training data.

[0124] In some embodiments, the first message further includes at least one of the following of the terminal device: reference signal received power, reference signal received quality.

[0125] Here, the Reference Signal Receiving Power (RSRP) is one of the key parameters representing the wireless signal strength and the physical layer measurement requirements in the LTE network, which is the average value of the signal power received on all resource particles carrying reference signals within a certain symbol. The Reference Signal Receiving Quality (RSRQ) is an index for measuring the signal quality by the ratio of the reference signal receiving power to the reference signal receiving quality.

[0126] In some embodiments, when the terminal device sends a terminal capability response message and / or a terminal capability information message to the first access network device, the first message further includes the RSRP and / or RSRQ of the terminal device, where the RSRP and / or RSRQ are used to characterize the wireless environment information of the terminal device. Since the wireless environment where the terminal device is located may be different when the terminal device sends the model training capability to the first access network device multiple times, and the model training capabilities of the terminal device in different wireless environments may be different, the RSRP and / or RSRQ can be used to associate the model training capability of the terminal device with the wireless environment when the terminal device sends the model training capability.

[0127] In some embodiments, after receiving the first message, the terminal device parses the first message to obtain the model training strategy indication information in the first message, and determines to upload training data or upload the locally updated model based on the model strategy indication information.

[0128] In one example, the model training strategy indication information can be identified by 0 and 1. When the model training strategy indication information is set to 1, it indicates that the terminal device uploads training data. In this scenario, the model data is the training data uploaded by the terminal device; when the model training strategy indication information is set to 0, it indicates that the terminal device uploads the locally updated model. In this scenario, the model data is the locally updated model uploaded by the terminal device, or the model data is the model parameters of the locally updated model.

[0129] In one example, the model training strategy indication information can be identified by 0 and 1. When the model training strategy indication information is set to 1, it indicates that the terminal device uploads training data. In this scenario, when the model training strategy indication information is set to 0, it indicates that the terminal device uploads training data. In this scenario, the model data is the training data uploaded by the terminal device; when the model training strategy indication information is set to 1, it indicates that the terminal device uploads the locally updated model. In this scenario, the model data is the locally updated model uploaded by the terminal device, or the model data is the model parameters of the locally updated model.

[0130] In the embodiments of the present application, based on the model training capability information of the terminal device, the model training strategy indication information is determined to control the terminal device to upload training data or the locally updated model. In this way, it is not necessary for all terminal devices to upload training data, reducing the amount of data uploaded to the first access network device and solving the problem in the prior art that the communication quality of other terminal devices is affected due to the upload of a large amount of training data. On the other hand, the model training strategy indication information in the present application is determined based on the model training capability information of the terminal device. In this way, it is possible to instruct the terminal device with stronger computing power to upload the locally updated model and instruct the terminal device with weaker computing power to upload training data. In this way, through the joint training of the model by the terminal device and the first access network device, the training duration of the model trained by the first access network device alone is reduced, the efficiency of model training is improved, and the delay problem of the model training task in the prior art is solved, thereby improving the computing accuracy of the model training task.

[0131] In some embodiments, determining the model training strategy indication information based on the first message in S102 includes:

[0132] S1021. Receive the model training requirement information sent by the core network device.

[0133] Here, the core network device refers to the device in the core network that provides service support for the terminal device. The core network is an important part of the communication network, responsible for managing data, sorting data, and determining the data transmission path.

[0134] In some embodiments, it may be the service manager in the core network that sends the model training requirement information to the first access network device.

[0135] In some embodiments, the model training requirement information may be determined by the service manager based on the model training description information sent by the application controller / application server in the core network. Among them, the model training description information may include at least one of the following: model training task, model training routine, and service function.

[0136] Here, the model training task is a core process of artificial intelligence. By using a set of known data (usually called training data) to teach or train a model, enabling it to learn and identify patterns or features in the data, and these patterns or features can then be used to predict or classify new, unseen data. The model training task includes multiple iterative trainings of the model. The model training routine refers to all the steps executed for one iterative training; the service function refers to the logical processing capabilities required to complete one model training routine, such as computing power, algorithms, data, etc.

[0137] In some embodiments, the model training requirement information includes at least one of the following: a device list, model training latency information of a model training task, and model accuracy information of the model training task;

[0138] Wherein, the device list includes at least one terminal device participating in the model training task;

[0139] The model training task includes multiple iterative trainings, and the training of the first model is the Nth iterative training in the multiple iterative trainings.

[0140] S1022. Determine model training policy indication information based on the model training requirement information and the first message.

[0141] In some embodiments, based on the model training requirement information and the first message, the type of the terminal device can be determined; and model policy indication information corresponding to the type of the terminal device can be determined.

[0142] In an example, the types of the terminal device include a first type and a second type. The model training policy indication information corresponding to the first type is to upload training data to the first access network device, and the model training policy indication information corresponding to the second type is to upload the locally updated model to the first access network device.

[0143] In some embodiments, the model training requirement and the first message are input into a target policy model, and the output of the target policy model is obtained; based on the output of the target policy model, the model training policy indication information of the terminal device is determined. Wherein, the target policy model can be any algorithm and / or relationship for determining the model training policy indication information of the terminal device.

[0144] In the embodiments of the present application, by simultaneously considering the model training latency information and the model accuracy information of the model training task, the model training policy indication information is determined. In this way, from the aspect of the computing requirements of the model training task, the first access network device can provide corresponding resource guarantees such as computing, storage, and communication for the model training task.

[0145] In some embodiments, the MAC CE further includes at least one of the following: first power information for indicating the transmission power of the terminal device when uploading training data; second power information for indicating the transmission power of the terminal device when uploading the model; time information for indicating the application time of the model training policy; time indication information for indicating the representation method adopted by the time information; the first model.

[0146] In some embodiments, the first power information and / or the second power information can be determined based on the identification information.

[0147] In one example, at least one transmission power is preset in the first access network device, and at least one transmission power is identified, so that the identification information corresponding to the transmission power when the terminal device uploads training data can be determined as the first power information.

[0148] In one example, at least one transmission power is preset in the first access network device, and at least one transmission power is identified, so that the identification information corresponding to the transmission power when the terminal device uploads the locally updated model can be determined as the second power information.

[0149] In other embodiments, the transmission power when the terminal device uploads training data is determined as the first power information, and the transmission power when the terminal device uploads the locally updated model is determined as the second power information.

[0150] In some embodiments, the time information includes at least one of the following: start time and end time; duration.

[0151] Here, the start time and end time refer to the start time and end time when the terminal device sends model data to the first access network device.

[0152] In some embodiments, the terminal device uploads model data based on the duration, and / or the start time and end time.

[0153] In one example, taking the time information including the start time and end time as an illustration, after the terminal device receives the start time and end time corresponding to the model training policy indication information, the terminal device uploads model data based on the start time and ends the upload of model data based on the end time.

[0154] In one example, taking the time information including the duration as an illustration, when the terminal device receives the duration, it completes the upload of model data within the duration.

[0155] In some embodiments, when the MAC CE does not include power information (the first power information and / or the second power information), the terminal device determines the power information based on the duration, or the start time and end time.

[0156] In one example, the terminal device obtains the target duration for uploading model data based on the start time and end time, and determines the power information when the terminal device uploads model data to the first access network device based on the target duration.

[0157] In one example, the terminal device determines the transmission rate corresponding to the duration, and determines the power information when the terminal device uploads model data to the first access network device based on the transmission rate.

[0158] In some other embodiments, when the terminal device fails to upload model data based on the time information, the terminal device stops uploading the model data.

[0159] In one example, it is described that the time information includes a start time and an end time. After the terminal device reaches the end time and has not completed the upload of the model data, the terminal device stops uploading the model data.

[0160] In one example, it is described that the time information includes a duration. If the terminal device fails to upload the model data within the duration, the terminal device stops uploading the model data.

[0161] In some embodiments, when the terminal device fails to upload model data based on the time information, the terminal device sends a failure message to the first access network device, and the failure message is used to indicate that the terminal device has not completed the upload of the model data.

[0162] In some other embodiments, when the terminal device fails to upload model data based on the time information, the terminal device sends a failure message to the first access network device, and the failure message includes information about the model data that has not been uploaded. For example, the size of the model data that has not been uploaded, or the proportion of the model data that has not been uploaded to all the model data.

[0163] In some embodiments, when the MAC CE includes a start time, an end time, and a duration, the duration can be determined based on the start time and the end time, or can be set arbitrarily.

[0164] In some embodiments, when the duration is set arbitrarily, after receiving the MAC CE, the terminal device first determines the target duration corresponding to the start time and the end time, then determines the larger value between the target duration and the duration as the transmission duration, and uploads the model data to the first access network device based on the transmission duration.

[0165] In some embodiments, when the duration is set arbitrarily, after receiving the MAC CE, the terminal device first determines the target duration corresponding to the start time and the end time, then determines the smaller value between the target duration and the duration as the transmission duration, and uploads the model data to the first access network device based on the transmission duration.

[0166] In some embodiments, the time indication information is used to indicate the representation methods of the start time and the end time, and the representation methods of the start time and the end time include one of the following: an absolute timestamp; a time offset relative to a first time, where the first time is the time when the terminal device receives the model training policy indication information.

[0167] In some embodiments, the time indication information may be identified by 0 and 1. For example, when the time indication information is set to 1, the representation used to indicate the start time and the end time is an absolute timestamp; when the time indication information is set to 0, the representation used to indicate the start time and the end time is a time offset relative to the first time. Alternatively, when the time indication information is set to 0, the representation used to indicate the start time and the end time is an absolute timestamp; when the time indication information is set to 1, the representation used to indicate the start time and the end time is a time offset relative to the first time.

[0168] In one example, the absolute timestamp may be Coordinated Universal Time (UTC), that is, the start time and the end time are UTC times. For example, the start time for the terminal device to upload model data is 14:00:00 on January 16, 2025, and the end time for the terminal device to upload model data is 14:02:00 on January 16, 2025.

[0169] In one example, the start time for the terminal device to upload model data is 0:01:00, and the end time for the terminal device to upload model data is 0:03:00.

[0170] In some embodiments, when the representation of the start time and the end time is an absolute timestamp, the terminal device starts uploading model data when the start time arrives and completes the upload of the model data before the end time arrives.

[0171] In some other embodiments, when the representation of the start time and the end time is an absolute timestamp, the terminal device starts uploading model data when the start time arrives. If the model data has not been completely uploaded when the end time arrives, the terminal device stops uploading the model data.

[0172] In some embodiments, when the representation of the start time and the end time is a time offset relative to the first time, the terminal device starts uploading model data when it determines that it has reached the time offset represented by the start time relative to the first time; and completes the upload of the model data when it determines that it has reached the time offset represented by the end time relative to the first time.

[0173] In some other embodiments, when the representation of the start time and the end time is a time offset relative to the first time, the terminal device starts uploading model data when it determines that it has reached the time offset represented by the start time relative to the first time; and if the model data has not been completely uploaded when it determines that it has reached the time offset represented by the end time relative to the first time, it stops uploading the model data.

[0174] In some embodiments, the first processor is configured to:

[0175] Send a second message to the terminal device via the first transceiver, where the second message is used to trigger the sending of the first message, and the second message includes one of the following: a terminal capability query message, a terminal capability request message.

[0176] Correspondingly, based on Figure 2 the wireless communication method, the wireless communication method provided by the embodiments of the present application further includes: receiving a second message sent by the first access network device, where the second message is used to trigger the sending of the first message, and the second message includes one of the following: a terminal capability query message, a terminal capability request message.

[0177] Here, the terminal capability query message refers to a message or instruction used by the network layer to query the hardware or software capabilities of the terminal device. The terminal capability request message is a message sent by the network to the terminal device, used to request the terminal device to report its hardware and software capability information. Among them, both the terminal capability query message and the terminal capability request message are RRC messages.

[0178] In some embodiments, the first access network device sends a second message to the terminal device, and the second message includes an indicator. If the value of the indicator is set to 1, it means that the first access network device requests the terminal device to report model training capability information.

[0179] In one example, the first access network device sends a terminal capability request message to the terminal device, and the terminal capability request message includes an indicator. If the value of the indicator is set to 1, it means that the first access network device requests the terminal device to report model training capability information.

[0180] In one example, the first access network device sends a terminal capability query message to the terminal device, and the terminal capability query message includes an indicator. If the value of the indicator is set to 1, it means that the first access network device requests the terminal device to report model training capability information.

[0181] In some other embodiments, the second message includes a model training type indication, used to indicate that the model training is federated learning. At this time, the second message is used to indicate that the first access network device requests the terminal device to report model training capability information in a federated learning environment.

[0182] In one example, the terminal capability query message includes a model training type indication, used to indicate that the model training is federated learning. At this time, the terminal capability query message is used to indicate that the first access network device requests the terminal device to report model training capability information in a federated learning environment.

[0183] In some embodiments, the second message includes at least one of the following: model training ability request information; a first threshold, which is used to control the sending of the first message based on the size of the training data of the terminal device; a first model.

[0184] In some embodiments, when the training data of the terminal device is greater than or equal to the first threshold, the terminal device needs to send model training ability information to the first access network device.

[0185] In some embodiments, when the training data of the terminal device is less than or equal to the first threshold, the terminal device updates the first model locally.

[0186] In the embodiments of the present application, a data transmission format corresponding to the model training policy indication information sent by the first access network device to the terminal device is proposed. In this way, the first access network device can generate corresponding transmission data based on this data transmission format, so that the terminal device can obtain the model training policy indication information based on this transmission data.

[0187] In some embodiments, based on Figure 1 the wireless communication method, the wireless communication method provided by the embodiments of the present application further includes: when the model training policy instructs the terminal device to upload training data, receiving the training data uploaded by the terminal device; updating the first model based on the training data to obtain a second model.

[0188] Correspondingly, based on Figure 2 the wireless communication method, the wireless communication method provided by the embodiments of the present application further includes: when the model training policy instructs the terminal device to upload training data, sending the training data to the first access device.

[0189] In some embodiments, the terminal device may upload the training data to the DCU in the first access network device through the wireless access unit in the first access network device.

[0190] In some embodiments, the DCU in the first access network device updates the first model based on the training data to obtain a second model.

[0191] In some embodiments, when the model training policies corresponding to all the terminal devices participating in the model training task under the first access network device instruct the terminal devices to upload training data, the first model is updated based on the training data uploaded by all the terminal devices to obtain a second model.

[0192] In some embodiments, when the model training strategies respectively corresponding to all the terminal devices participating in the model training task under the first access network device instruct the terminal devices to upload training data, the first model is updated based on the training data uploaded by each terminal device to obtain a third model; an aggregation operation is performed on all the third models to obtain a second model.

[0193] In the embodiments of the present application, when the first access network device receives the training data sent by the terminal device, the first model can be updated, which can enable the first access network device to not need to wait for all the terminal devices to upload training data, reduce the waiting time of the first access network device, thereby reducing the latency of the model training task and improving the accuracy of the second model.

[0194] In some embodiments, based on Figure 1 the wireless communication method, the wireless communication method provided in the embodiments of the present application further includes: when the model training strategy instructs the terminal device to upload the locally updated model, receiving the second model uploaded by the terminal device; the second model is obtained by the terminal device updating the first model based on the training data.

[0195] Correspondingly, based on Figure 2 the wireless communication method, the wireless communication method provided in the embodiments of the present application further includes: updating the first model based on the training data to obtain a second model; sending the second model to the first access network device.

[0196] In some embodiments, the terminal device may upload the second model to the DCU in the first access network device through the wireless access unit in the first access network device.

[0197] In some embodiments, the terminal device includes a DCU, and the second model may be obtained by the DCU in the terminal device updating the first model based on the training data.

[0198] In the embodiments of the present application, the terminal device can update the first model to obtain a second model, and then send the second model to the first access network device, reducing the amount of data sent by the terminal device to the first access network, not only reducing the impact on the communication quality of other terminal devices, but also reducing the latency of the model training task and improving the accuracy of the second model.

[0199] In some embodiments, based on Figure 1The wireless communication method provided by the embodiments of the present application further includes: receiving local models sent by each of one or more second access network devices; determining a global model based on the local models sent by each of the one or more second access network devices and the local model of the first access network device, where the local model of the first access network device is determined based on model data; and sending the global model to the terminal device.

[0200] In some embodiments, the local model may be generated based on a second model.

[0201] In one example, taking the first access network device as an illustration, the DCU in the first access network device may receive training data uploaded by each terminal device. Thus, when the model training strategies respectively corresponding to all the terminal devices participating in the model training task under the first access network device instruct the terminal devices to upload training data, the DCU in the first access network device may receive the training data respectively sent by all the terminal devices participating in the model training task, and based on all the training data, train a first model to obtain a second model, and determine the second model as the local model.

[0202] In one example, when the model training strategies respectively corresponding to all the terminal devices participating in the model training task under the first access network device instruct the terminal devices to upload training data, the DCU in the first access network device may first train the first model for the training data uploaded by each terminal device to obtain a third model corresponding to the terminal device, so that the same number of third models as the terminal devices participating in the model training task can be obtained. Then, obtain the model parameters respectively corresponding to all the third models, and perform weighted average and aggregation processing on all the model parameters to obtain a second model, and determine the second model as the local model.

[0203] In one example, taking the first access network device as an illustration, the DCU in the first access network device may receive the second model uploaded by each terminal device. Thus, when the model training strategies respectively corresponding to all the terminal devices participating in the model training task under the first access network device instruct the terminal devices to upload the locally updated models, the DCU in the first access network device may receive the second models respectively sent by all the terminal devices participating in the model training task. The DCU in the first access network device first obtains the model parameters of each second model, and then performs weighted average and aggregation processing on all the model parameters to obtain the local model.

[0204] In one example, taking the first access network device as an illustration, in the case where the model training policies corresponding to a part of the terminal devices participating in the model training task under the first access network device respectively instruct the terminal devices to upload training data, and the model training policies corresponding to another part of the terminal devices respectively instruct the terminal devices to upload the locally updated models, a local model is determined based on all the second models uploaded by a part of the terminal devices and the second models corresponding to another part of the terminal devices determined by the first access network device. Among them, for the method of determining the second model, refer to the description in the case where the model training policies corresponding to all the above-mentioned terminal devices respectively instruct the terminal devices to upload training data.

[0205] In some embodiments, based on the local models of all the second access network devices and the local model of the first access network device, the model parameters corresponding to all the received local models are determined; the model parameters corresponding to all the local models are subjected to weighted average and aggregation processing to obtain a global model. Among them, for the manner of obtaining the local model of the second access network device, refer to the method of determining the local model of the above-mentioned first access network device.

[0206] In some embodiments, when the DCU in the first access network device is a routine server, the first access network device is used to determine the global model. When the DCU in the first access network device is not a routine server, the first access network device sends the local model to the access network device corresponding to the routine server, and the global model is determined by this access network device. Among them, the routine server can be determined based on the resource manager in the access network device.

[0207] In the embodiments of the present application, based on the model training ability information of the terminal device, the model training policy indication information is determined to control the terminal device to upload training data or upload the locally updated model, reducing the amount of data uploaded to the access network device, thereby improving the accuracy of the obtained local model.

[0208] In some embodiments, the terminal device is jointly served by multiple distributed access network devices, and the multiple distributed access network devices include a first access network device and a third access network device;

[0209] Based on Figure 1 the wireless communication method, the wireless communication method provided by the embodiments of the present application further includes receiving model training configuration information sent by the core network device, where the model training configuration information instructs the first access network device to participate in the Nth iteration training of the model training task;

[0210] Among them, the distributed access network device participating in the Mth iteration training of the model training task is the first access network device or the third access network device.

[0211] In some embodiments, when performing a model training task, for each iterative training of the model training task, the access network device that jointly performs the iterative training with the terminal device can be different.

[0212] In one example, the first access network device participates in the Nth iterative training of the model training task, and the first access network device can perform the Mth iterative training of the model training task.

[0213] In one example, the first access network device participates in the Nth iterative training of the model training task, and the third access network device can perform the Mth iterative training of the model training task.

[0214] In a cell-free multiple-input multiple-output (MIMO) system, a terminal device can be jointly served by multiple distributed access network devices. Therefore, for the access network device participating in each iterative training of the model training task, it can be determined by the core network from multiple distributed access network devices serving the terminal device. The access network device for each iterative training is independent. The access network devices for different iterative trainings can be the same distributed access network device or different distributed access network devices.

[0215] After the core network device determines the access network device participating in the iterative training, it sends model training requirement information to the access network device. Based on the received model training requirement information, the access network device determines to participate in the iterative training and can send a second message to the terminal device.

[0216] In the embodiments of the present application, the terminal device is jointly served by multiple distributed access network devices. In this way, through the joint cooperation of multiple access network devices, the inter-cell interference in the traditional cell-centered wireless communication network can be eliminated, the hard segmentation between cells can be eliminated, making the network a cell-free network, allowing the terminal device to seamlessly switch between different access points, obtaining a higher macro-diversity gain from distributed antennas, reducing the communication distance between the terminal device and the network, thereby improving the transmission rate of the data uploaded to the first access network device, further reducing the latency of the model training task, and improving the accuracy of the second model.

[0217] Next, the wireless communication method provided by the embodiments of the present application will be described.

[0218] With the development of the intelligence of terminal devices, AI has witnessed an explosive growth. More and more terminal data is sent to the server to form its training dataset. However, terminal data involves personal privacy or confidentiality agreements. Therefore, uploading terminal data to the server may lead to the leakage of personal privacy. To protect personal privacy, Federated Learning (FL) is introduced.

[0219] Federated Learning is a distributed model training method. In this method, terminal devices use their local data to train the model, update the model parameters (for example, update the gradients or weights of the model), and then send the updated model parameters to the server. The server aggregates the received model parameters to generate a global model, thus ensuring the global consistency of the model. In this way, through Federated Learning, local training data does not need to be sent to the server. Federated Learning mainly includes the following steps:

[0220] 1) The server randomly selects multiple terminal devices or selects multiple terminal devices with typical characteristics and sends the initial model to the terminal devices.

[0221] 2) The terminal devices train the initial model based on the local training data to obtain an updated model.

[0222] 3) The terminal devices send the updated model parameters (or the updated model itself) to the server.

[0223] 4) The terminal devices perform weighted averaging and aggregation on the received updated model parameters to obtain a new model. This new model can be used for the next round of model training or for model inference.

[0224] In the prior art, the communication process between terminal devices is synchronous, that is, in each model training iteration, each terminal device needs to obtain the same initial model. On the other hand, the computing process only occurs between the terminal devices and the centralized processing unit (server), that is, only the terminal devices and the centralized processing unit in the network can perceive the model training, while other network devices do not perceive the model training.

[0225] Furthermore, when deploying Federated Learning in a wireless network, the existing architecture is difficult to support the optimization of model training, because:

[0226] 1) In Federated Learning, all model training parameters need to be transmitted through the wireless link. Therefore, the quality of the wireless link will affect the quality of model training. However, existing wireless resource allocation technologies usually only consider communication efficiency and do not consider the efficiency of model training, such as model accuracy, etc.

[0227] 2) Due to the limited wireless bandwidth, the base station usually selects terminal devices with better wireless link quality for local model training, rather than those with limited wireless link quality, resulting in insufficient generalization performance of the model.

[0228] To make the global model converge, the terminal devices participating in federated learning need to frequently upload the updated model parameters through the wireless communication link during the training process. Since the existing network was not designed to consider how to efficiently provide AI capabilities and services, there are the following pain points in deploying federated learning on the existing network:

[0229] 1) For federated learning, the more terminal devices participate, the better the performance of the final global model (e.g., the higher the model accuracy). However, the frequent transmission of high-dimensional model parameters between a large number of terminal devices and the server will result in huge communication overhead, making it difficult for the network to guarantee the quality of service requirements of other services.

[0230] 2) Due to the inherent instability of the wireless channel, the simultaneous upload of model parameters by a large number of terminal devices in different cells will seriously interfere with the wireless channels of terminal devices in other cells, thus reducing the communication quality of other user devices.

[0231] 3) For federated learning, in each round of model training, the server must receive the model parameters sent by all terminal devices before it can perform model parameter aggregation to carry out the next round of model training iteration. Therefore, the latency of each round of model training in federated learning depends on the terminal device with the longest model training latency. Due to the heterogeneity of terminal devices (i.e., different computing capabilities), the model training latency of the entire federated learning system will increase sharply due to the overly long local model training latency of some terminal devices, resulting in the "straggler effect".

[0232] 4) The asymmetry of the uplink and downlink transmission rates in the wireless network, that is, the uplink transmission latency is much higher than the downlink transmission latency, will further exacerbate the transmission differences between terminal devices, thus increasing the model training latency of the entire federated learning.

[0233] 5) The complexity of maximizing the convergence speed of federated learning: on the one hand, minimizing the latency of federated learning model training iterations will result in a loss of model accuracy; on the other hand, maximizing model accuracy requires more model training iterations, resulting in greater latency.

[0234] Traditional wireless communication networks are a cell - centered wireless network architecture. Different from traditional wireless networks, the cell - free MIMO system is a terminal - centered wireless network architecture. By deploying a large number of distributed access points and introducing cooperation among the access points to fully eliminate inter - cell interference, it breaks the cell boundaries of traditional cellular networks and provides a new network coverage and user service mode. The cell - free MIMO re - defines the cell concept in the traditional cellular network from the perspective of the terminal device, eliminates the hard segmentation between cells, makes the network a cell - free area, allows the terminal device to seamlessly switch between different access points, and obtains a high macro - diversity gain from the distributed antennas. In addition, since the access points are deployed close to the terminal device, the communication distance between the terminal device and the network is reduced, thereby improving the spectral efficiency, reducing the data transmission delay, and improving the energy efficiency.

[0235] The core idea of cell - free MIMO is to adopt a flexible cell concept at the terminal device granularity, that is, the network dynamically selects the access points around the terminal device to serve the terminal device according to the service requirements, location, and mobility of the terminal device. From the perspective of the terminal device, as the terminal moves, the associated access points change dynamically; from the perspective of the network, each access point is dynamically associated with multiple terminals.

[0236] When terminals reuse the same time - frequency resources, distributed cooperative transmission needs to flexibly support the scalability of the terminal device scale and the cooperative access point scale. There are the following three ways of distributed cooperative transmission, as Figure 4 shown:

[0237] 1) Scalable full - dynamic cooperative clustering (Dynamic Cooperative Clustering, DCC): The access point (Access Point, AP) sends the received terminal device, i.e., user equipment (User Equipment, UE) signal to the CPU, and the CPU realizes dynamic cooperative transceiver. The CPU is usually deployed in the cloud.

[0238] 2) Fully distributed cooperation: The AP locally realizes the coherent transmission and coherent reception of multiple terminal device signals, and the CPU realizes the distribution of downlink data and the combination of uplink data.

[0239] 3) Dynamic cooperation based on fixed clustering: The AP sends the terminal device signal to the edge distributed unit (Edge Distributed Unit, EDU), and the EDU realizes the relevant transmission and relevant reception of multiple terminal device signals. A user - centered distributed unit (User Centric Distributed Unit, UCDU) is deployed in the cloud, and the UCDU realizes the processing of user data (for example, the distribution of downlink data and the combination of uplink data).

[0240] Based on this, the present application applies federated learning to cell-free MIMO and proposes a wireless communication system. This wireless communication system simultaneously considers the model training iteration delay and model accuracy, jointly designs communication and computing, enabling the network to provide corresponding resource guarantees for computing, storage, communication, etc. according to the quality of service requirements of federated learning (specific delay, jitter, packet loss rate, reliability, etc.), thereby improving the Service Level Agreement (SLA).

[0241] First, introduce the wireless communication system as Figure 5 shown. This wireless communication system mainly includes the following three-plane functions:

[0242] I. Control Plane 1: Responsible for the establishment, maintenance, and optimization of terminal sessions, ensuring that terminal devices can access the network and provide corresponding quality of service. Control Plane 1 mainly includes the following functions: 1) Establishment and release of terminal sessions: Manage the initiation and termination of terminal sessions. 2) Mobility management: Manage the mobility of terminal devices, such as handovers, etc. 3) Resource allocation: Allocate frequency-domain resources and time-domain resources to meet the communication needs of terminals. 4) Quality of service control: Dynamically adjust resource allocation according to the network environment to ensure the specific quality of service of terminals. 5) Security management: Manage user security, including authentication, encryption, certification, etc.

[0243] In this wireless communication system, the control plane is also composed of the following logical functions: 1) Service Manager 11: Responsible for managing AI services and managing AI services through the unified scheduling of Resource Manager 12, Task Manager 13, Model Manager 14, and Access Manager 15. The AI service information in Service Manager 11 can come from Application Controller 4. 2) Resource Manager 12: Responsible for managing communication resources and computing resources, generating a scheduling policy based on the network environment, resource budget, and AI service requirements, and sending this scheduling policy to Task Manager 13. 3) Task Manager 13: Responsible for managing AI service-related tasks (abbreviation: AI tasks, that is, the above-mentioned model training tasks), and managing AI tasks through the unified scheduling of the routine manager and Model Manager 14. Further, Task Manager 13 can generate a corresponding scheduling policy for each AI task, or improve the existing scheduling policy for each AI task, and send this scheduling policy to the routine manager and Model Manager. 4) Model Manager 14: Responsible for managing AI models and realizing the management of AI models through the scheduling of the model base. According to the requirements of different AI tasks in the AI service, Model Manager 14 can be controlled by Service Manager 11 or Task Manager 13.

[0244] 5) Access Manager 15: Controlled by the Service Manager 11, and responsible for managing access devices, including terminal devices, distributed computing units, etc.

[0245] II. Computing Plane 2: Responsible for user data management, data collection, data storage, data collaboration, etc., so as to achieve efficient iteration of user data, reliable migration of network status, open data services, etc. The computing plane 2 mainly includes the following functions: 1) Automated integration of AI elements and their workflows: Manage AI element resources (such as computing power, algorithms, data, etc.) in the communication network, support the integration and interaction between AI elements, provide a local automated integration AI operating environment, achieve efficient operation of the AI workflow, and meet the real-time feedback of the operation results of AI services. 2) Efficient collaboration of intelligent services: Support the atomic abstraction, distributed deployment, collaboration, etc. of AI services.

[0246] In this wireless communication system, the computing plane 2 is also composed of the following logical functions: 1) Routine Manager 21: Responsible for managing the routines of AI tasks (abbreviated as AI routines, an AI task contains one or more AI routines), and scheduling the forwarding units 31 in the forwarding plane 3. In addition, the routine manager 21 can further optimize the scheduling policy received from the task manager 13. 2) Model Base 22: Responsible for the storage and scheduling of AI models, and sending the AI models to the terminal device 5 through the forwarding unit 31. 3) Distributed Computing Unit 23: Responsible for providing computing power for AI routines. The DCU can be deployed in the network as an independent logical function, or deployed in communication devices, such as deployed in a wireless access unit or a terminal device. Model training is performed by the DCU.

[0247] It should be noted that in the case where the DCU is not deployed in the terminal device, the local training sample data (i.e., the above-mentioned training data) of the terminal device can be uploaded to the base station (i.e., the above-mentioned first access network device), and the base station performs model training based on the local training sample data.

[0248] 3. Forwarding Plane 3: Also known as the user plane, responsible for the transmission of user data. The forwarding plane mainly includes the following functions: 1) Data transmission: Responsible for the transmission of user data between the terminal device and the network. 2) Packet forwarding: Responsible for forwarding data packets to ensure efficient routing of data packets between the terminal device and the destination. 3) Quality of Service processing: Manage quality of service parameters, including delay, throughput, reliability, etc. 4) Data optimization: Responsible for the optimization of data transmission, including data compression, data encryption, etc.

[0249] In this in-line communication system, the forwarding plane mainly consists of the following logical functions: 1) Forwarding unit 31: responsible for forwarding information or data between different logical functions. For example, forwarding the scheduling policy from the routine manager 21 to the radio access unit 32 so that the radio access unit 32 can transmit and control the data of the terminal device 5; forwarding the model data to each logical function for deploying and processing the AI service. Further, when the AI service is completed, the forwarding unit 31 forwards the information required by the application server 6 to the application server 6. 2) Radio access unit 32: responsible for the wireless communication between the terminal device 5 and the network.

[0250] Next, a wireless communication method adapted to the wireless communication system as Figure 5 shown is introduced. As Figure 6 shown, the method includes the following steps:

[0251] S1. At least one terminal device selected by the network to upload local training sample data (i.e., the above-mentioned training data) uploads the local training sample data in the at least one terminal device to the network.

[0252] S2. At least one terminal device selected by the network for local model update updates the local model based on the local training sample data.

[0253] S3. Send the updated model parameters (or the updated model) to the network.

[0254] S4. Based on the local training sample data received in S1, the DCU in the network performs local model update. Then, the updated local model and the model parameters (or the model itself) received in S2 are weighted-averaged and aggregated to obtain a new global model.

[0255] S5. If the updated global model can be used for model inference, the network sends the updated global model to the application server. Determine whether to perform the next round of iteration based on the number of iterations and / or whether the model converges. If the next round of iteration is required, the network sends the updated global model to at least one terminal device selected for local model update.

[0256] In the above steps, the UE sends the local training sample data to the DCU through S1, and the DCU performs local model update through S4. In addition, S2 and S4 are different in the time scale: on the one hand, the DCU and the base station can be deployed in the same place. Therefore, both S2 and S4 need to be subject to the same base station energy consumption limit; on the other hand, the DCU receives the local training sample data from the UE, while the base station receives the bit data stream from the UE, that is, the model or the model parameters.

[0257] If the DCU is deployed at the base station, the content carried in the bitstream data received by the base station is training sample data. Among them, the base station performs machine parsing on the received bitstream to obtain the training sample data, and then sends the training sample data to the DCU.

[0258] The wireless communication method provided by the embodiments of the present application can be implemented as including but not limited to the following Embodiment 1 to Embodiment 3.

[0259] Embodiment 1

[0260] Combined with the specific logical functions of the wireless communication system, the deployment of the federated learning training service includes the following processes, as Figure 7 shown:

[0261] S701. The service manager 11 receives the description information of the federated learning training service sent by the application controller 4, and receives the device access information and network environment information sent by the terminal device 5 through the access manager 15. Based on the received description information, device access information and network environment information, determine the deployment mode of the federated learning training service and the corresponding computing plane.

[0262] The application controller 4 sends the description information of the federated learning training service to the service manager 11. The description information includes the AI tasks, AI routines and business functions of the federated learning training service. Among them, the business function refers to the logical processing ability required to complete an AI routine.

[0263] The service manager 11 receives the device access information and network environment information through the access manager 15. Among them, the device access information includes the information of the terminal device and the DCU, such as the device ID, etc.; the network environment information includes the wireless channel environment between the terminal device and the DCU.

[0264] The service manager 11 determines the deployment mode of the federated learning training service and the corresponding computing plane 2 based on the received description information, device access information and network environment information. The deployment mode includes the following information:

[0265] Information 1: The location and required resources of the specific business function instance. The specific business function instance includes an AI routine server (i.e., the above-mentioned routine server) and an AI routine client. This information is optional information.

[0266] Information 2: The logical link between the AI routine server and the AI routine client, that is, the forwarding plane between the AI routine server and the AI routine client. Through this logical link, the AI routine server and the AI routine client know how to communicate through the forwarding plane.

[0267] Information 3: The performance metrics of the federated learning model training service, including model training latency, model accuracy, etc.

[0268] S702. The service manager 11 configures the computing plane 2 based on the deployment mode and the corresponding computing plane 2, and sends the configuration result of the computing plane 2 to the resource manager 12 to reserve communication resources and computing resources.

[0269] Based on the determined deployment mode and the corresponding computing plane, the service manager 11 configures the computing plane 2. For example, it configures the terminal devices and DCUs participating in the AI routine, the data forwarding order between the terminal devices and each DCU, and the DCU serving as the AI routine server. After successful configuration, the service manager 11 sends the configuration result to the resource manager 12 to reserve communication resources and computing resources.

[0270] S703. The service manager 11 sends the description information of the federated learning training service to the task manager 13 and the model manager 14.

[0271] S704. The resource manager 12 reserves the required communication resources and computing resources based on the configuration result of the computing plane, and sends the resource reservation result to the task manager 13.

[0272] S705. The task manager 13 determines the routine manager 21 and sends a request to the routine manager 21 to request the routine manager 21 to execute the corresponding AI routine.

[0273] The task manager 13 determines the routine manager 21 based on the AI routine in the received description information, and sends a request to the routine manager 21 to request the routine manager 21 to execute the corresponding AI routine. Additionally, when there are multiple AI routines in the description information, it determines the dependency relationship (i.e., the execution order of each AI routine) between the multiple AI routines and sends this dependency relationship to the routine manager 21.

[0274] S706. The task manager 13 determines the model requirements corresponding to the AI task according to the AI task in the description information, and sends the model requirements to the model base 22. The model requirements include the model ID, usage scenario, etc.

[0275] S707. The model manager 14 sends a model scheduling request to the model base 22 based on the received model requirements and the description information received from the service manager 11 in S703 to request the model base 22 to deploy the required model.

[0276] S708. The routine manager 21 sends the AI routine received in S705 (optionally, also including the dependency relationship between AI routines) to the AI routine server and the AI routine client through the forwarding unit 31.

[0277] An AI routine server can be associated with multiple AI routine clients, and the association relationship between the AI routine server and the AI routine clients can be dynamically updated by the task manager, that is, the task manager updates the association relationship between the AI routine server and the AI routine clients by updating the routine manager.

[0278] The logical link between the AI routine server and the AI routine clients is determined by the resource manager.

[0279] The AI routine server and the AI routine clients are determined by the service manager.

[0280] S709. The model base deploys the required models to the terminal device and the DCU through the forwarding unit.

[0281] Embodiment 2

[0282] Through Embodiment 1, the deployment process of the federated learning training service has been completed. The terminal device and the DCU can start the federated learning training. The iterative process of one federated learning training is as Figure 8 shown:

[0283] S801. The first type of terminal device selected by the resource manager 12 as an AI routine client sends local training sample data to the distributed computing unit 23 through the wireless access unit 32.

[0284] Other second type of terminal devices not selected by the resource manager 12 as AI routine clients perform local model updates based on the local training sample data. For example, they update the gradients of the model to obtain updated model parameters or an updated model.

[0285] Considering both the model training latency and the model accuracy, the process by which the resource manager 12 selects the first type of terminal device and the second type of terminal device is as follows:

[0286] During implementation, if the model error is approximately the mean square gradient norm of the loss function, then the initial model error is:

[0287]

[0288] In this way, the final model error after Γ rounds of iteration is:

[0289]

[0290] Based on the above formulas (1) and (2), the relationship between the final model error and the descent error α of any round of iteration τ can be obtained τ as:

[0291]

[0292] Based on formula (3), we can obtain:

[0293]

[0294] After Γ rounds of iteration, the total time delay is:

[0295]

[0296] where t τ is the time required for the τ -th round of iteration.

[0297] In order to reduce the total time delay, it is necessary to fix the time required for each iteration and reduce the total number of iterations. Since both the initial model error and the final model error are determined, therefore, is also fixed.

[0298] Therefore, in order to reduce the total number of iterations, it is necessary to increase α as much as possible τ , that is, to maximize the error reduction under the fixed time for each iteration, namely:

[0299]

[0300] In addition, since is the result of the (τ - 1)-th round of iteration and it is a fixed value during the τ -th round of iteration, therefore, in order to reduce the total number of iterations, it is necessary to minimize the model error of one round of iteration, namely:

[0301]

[0302] The lower bound of the global loss function of the model is F inf , then the upper bound of the model error of one round of iteration is:

[0303]

[0304] where M represents the number of distributed base stations, and B m is the size of the training sample data received by each DCU.

[0305] Furthermore, this upper bound can be transformed into a joint communication and computing resource allocation problem, namely:

[0306]

[0307] where T local represents the time required for the terminal device to upload local training sample data, T ul represents the time required for the terminal device to upload the local model or model parameters, B represents the strategy of each terminal device (i.e., uploading local training sample data, or uploading local model / model parameters), P represents the transmission power allocation of the terminal device, and f uDenote the computing frequency of the terminal device as C u Denote the CPU cycles for the terminal device to compute a sample of training data (i.e., the time to compute a sample of training data) as S, and S represents the size of the local sample training data Denote the transmission rate at which the terminal device in the m-th base station sends local sample training data

[0308] In the above manner, the performance optimization problem of federated learning model training is transformed into an optimization problem of network parameters, thereby determining the first type of terminal device and the second type of terminal device

[0309] S802. The second type of terminal device sends the updated model parameters (or the updated model) to the DCU through the radio access unit 32

[0310] S803. Based on the received local training sample data, the DCU performs local model update to obtain a new local model 1 (i.e., the above-mentioned local model). Then, the DCU performs weighted averaging and aggregation on the updated local model 1 and the received model parameters (or the model itself) to obtain a new local model 2 (i.e., the above-mentioned local model). Finally, the DCU sends the updated local model 2 to the AI routine server through the forwarding unit, where one of the multiple DCUs is selected by the resource manager as the AI routine server

[0311] S804. The AI routine server receives the local models 2 from each DCU and aggregates these local models 2 to obtain a new global model

[0312] S805. The AI routine server sends the updated global model to the first type of terminal device and the DCU through the forwarding unit for the next round of training iteration

[0313] S806. If the federated learning training task has ended, the AI routine server sends the updated global model to the model base through the forwarding unit so that the model base stores the updated global model

[0314] Embodiment III

[0315] The wireless communication system provided in this application includes a terminal device, a base station (i.e., the first type of access network device), and a core network. Then, the wireless communication method of this wireless communication system is as Figure 9 shown. Among them, the base station has functions such as a radio access unit, an access manager, a DCU, a forwarding unit, a model base, a model manager, a task manager, a routine manager, a resource manager, etc.; the core network has functions such as a service manager, an application controller, an application server, etc. As Figure 9 shown

[0316] S901. The service manager receives the description information of the federated learning model training service (i.e., the above-mentioned model training description information) from the application controller. The description information includes an AI task, an AI routine (i.e., the above-mentioned model training routine), and a business function.

[0317] Among them, the AI task refers to model iteration, that is, model update. One model iteration is called an AI task. The AI routine refers to the process required for model iteration. For example, local model training, model upload, global model update, etc. One AI task contains at least one AI routine. The business function refers to the logical processing capabilities required to complete an AI routine. For example, computing power, algorithms, data, etc.

[0318] S902. The service manager sends a model training requirement (i.e., the above-mentioned model training requirement information) to the base station. The model training requirement includes the model training delay (i.e., the above-mentioned model training delay information) and the model accuracy (i.e., the above-mentioned model training accuracy information).

[0319] Optionally, the model training requirement further includes a list of terminal devices (i.e., the above-mentioned device list), that is, at least one terminal device participating in the AI task.

[0320] S903. The base station sends Model 1 (i.e., the above-mentioned first model) to the terminal device, that is, the initial model for this AI task.

[0321] S904. The base station sends a model training capability request (i.e., the above-mentioned second message) to the terminal device to request the terminal device to report its model training capabilities.

[0322] Optionally, if the base station and the terminal device do not execute step 903, the base station also sends Model 1 to the terminal device.

[0323] In a possible implementation, the base station sends a terminal capability query message to the terminal device. This message is an RRC message. The message contains an indicator. If the value of the indicator is set to 1, it means that the base station requests the terminal device to report its model training capabilities. Further, the message also contains a model training type indicator, which is used to indicate that this model training is federated learning. At this time, this message is used to indicate that the base station requests the terminal device to report its model training capabilities in a federated learning environment.

[0324] In another possible implementation, the base station sends a terminal capability request message to the terminal device, and this message is an RRC message. There is an indicator in this message. If the value of this indicator is set to 1, it means that the base station requests the terminal device to report the model training capability. Further, there is a threshold value in this message, which is used to indicate that when the size of the local training sample data in the terminal device reaches or exceeds this threshold value, the terminal device needs to send the model training capability to the base station. For example, send the size of the local training sample data in the terminal device.

[0325] The above S903 and S904 are optional steps.

[0326] S905. The terminal device sends a model training capability report (i.e., the above model training capability information) to the base station. The model training capability includes the size of local training sample data, computing frequency, CPU cycles, etc.

[0327] The computing frequency refers to the clock frequency at which the CPU works, that is, the clock oscillation frequency of the CPU.

[0328] The CPU cycle refers to the time required to calculate a sample training data.

[0329] In a possible implementation, the terminal device sends a terminal capability information message to the base station, and this message is an RRC message. There is an indicator in this message, which is used to indicate that the terminal device reports the model training capability, and the following parameter information is included in this indicator: computing frequency, CPU cycles, size of local training sample data.

[0330] In another possible implementation, the terminal device sends a terminal capability response message to the base station, and this message is an RRC message. There is an indicator in this message, which is used to indicate that the terminal device reports the model training capability, and the following parameter information is included in this indicator: computing frequency, CPU cycles, size of local training sample data.

[0331] In another possible implementation, the terminal device sends the computing frequency and CPU cycles of the terminal device to the base station through the terminal capability information message. In addition, the terminal device sends the size of the local training sample data of the terminal device to the base station through the terminal capability response message.

[0332] In another possible implementation, when the terminal device sends the model training capability to the base station, it simultaneously sends the radio environment information of the terminal device to the base station. The radio environment information can be RSRP or RSRQ. Through this method, the model training capability of the terminal device is associated with the radio environment where the terminal device is located. For example, when the terminal device sends the model training capability to the base station multiple times, the corresponding radio environment may be different each time.

[0333] The base station selects the first type of terminal device to upload local training sample data, and / or selects the second type of terminal device to perform local model update, and sends a model training instruction (i.e., the first message above) to the terminal device.

[0334] The specific process of the base station selecting the first type of terminal device and the second type of terminal device is as in S801 in the second embodiment, which will not be elaborated here.

[0335] The model training instruction includes the model training strategy of the terminal device (i.e., the above model training strategy indication information), and the model training strategy is used to instruct the terminal device to upload local training sample data or perform local model update. For example, it instructs the first type of terminal device to upload local training sample data and instructs the second type of terminal device to perform local model update (i.e., upload the updated model or model parameters).

[0336] Optionally, when the terminal device needs to upload local training sample data, the model training instruction further includes the transmission power of the terminal device, that is, the transmit power required for the first type of terminal device to upload local training sample data to the base station.

[0337] Optionally, when the base station sends the model training instruction to the terminal device, it can also send the start time and end time corresponding to the model training strategy to the terminal device, that is, the model training instruction further includes the start time and end time. Among them, the start time and end time can be represented in an absolute timestamp or time offset manner.

[0338] Optionally, when the base station sends the model training instruction to the terminal device, it can also send the duration corresponding to the model training strategy to the terminal device, that is, the model training instruction further includes the duration. For example, the duration for the first type of terminal device to upload local training sample data is 2 seconds.

[0339] Optionally, the base station sends the above model training instruction to the terminal device through MAC CE. MAC CE includes the following parameters:

[0340] Model training strategy indication information. When this indication information is set to 1, the terminal device uploads local training sample data; when this indication information is set to 0, the terminal device performs local model update.

[0341] Transmission power index value, which is used to indicate that the terminal device uses the transmit power corresponding to this index value when uploading local training sample data.

[0342] Start time and end time. For example, MAC CE contains an indication information. When this indication information is set to 1, the start time and end time are absolute timestamps; when this indication information is set to 0, the start time and end time are time offsets.

[0343] Duration information.

[0344] In one example, the MAC CE format is as Figure 10 shown. Among them, the policy indication represents model training policy indication information, and the time indication represents the format of the start time and the end time.

[0345] Optionally, if the initial model is not sent to the terminal device in S903 and S904, then when the terminal device needs to perform local model update, the model training indication further includes Model 1, that is, the initial model of this AI task.

[0346] S907. The terminal device uploads local training sample data or the updated model to the base station.

[0347] Case 1. When the model training policy received by the terminal device is to send local training sample data, the terminal device uploads local training sample data to the base station.

[0348] Optionally, if the terminal device also receives a transmission power indication, the terminal device uploads local training sample data to the base station with this transmission power.

[0349] Optionally, if the terminal device also receives the start time and the end time, and the start time and the end time are absolute timestamps, the terminal device starts uploading local training sample data when the start time arrives, and uploads all local training sample data before the end time arrives. If there is still local training sample data not uploaded when the end time arrives, the terminal device stops uploading local training sample data. Among them, if the terminal device does not receive a transmission power indication, the terminal device can obtain the duration information for uploading local training sample data according to the start time and the end time, and determine the transmission power when uploading local training sample data to the base station according to this duration information. For example, the terminal device determines the transmission rate when uploading local training sample data to the base station according to this duration information, and then determines the corresponding transmission power according to this transmission rate.

[0350] Optionally, if the terminal device also receives a start time and an end time, and the start time and the end time are time offsets, the terminal device calculates the absolute timestamps of the start time and the end time when receiving the model training indication sent by the base station, starts uploading local training sample data when the start time arrives, and finishes uploading all local training sample data before the end time arrives. If there is still local training sample data not uploaded when the end time arrives, the terminal device stops uploading local training sample data. Among them, if the terminal device does not receive a transmission power indication, the terminal device can obtain the duration information for uploading local training sample data based on the start time and the end time, and determine the transmission power for uploading local training sample data to the base station according to the duration information. For example, the terminal device determines the transmission rate for uploading local training sample data to the base station according to the duration information, and then determines the corresponding transmission power according to the transmission rate.

[0351] Optionally, if the terminal device also receives duration information, the terminal device needs to upload local training sample data within the duration. If there is still local training sample data not uploaded within the duration, the terminal device stops uploading local training sample data. Among them, if the terminal device does not receive a transmission power indication, the terminal device can determine the transmission power for uploading local training sample data to the base station according to the duration information.

[0352] Optionally, if the terminal device cannot finish uploading local training sample data when the end time arrives or within the duration, the terminal device sends a failure message to the base station, and the failure message is used to indicate that there is still local training sample data not uploaded by the terminal device (or, the terminal device has not completely uploaded all local training sample data). Among them, optionally, the failure message contains an indication message, which is used to indicate that the reason for the failure of sending local training sample data is that the end time has arrived (or, it has not been possible to finish uploading local training sample data before the end time arrives). Optionally, the failure message contains information about the local training sample data not uploaded, for example, the size of the local training sample data not uploaded, or the proportion of the local training sample data not uploaded in all local training sample data.

[0353] Case 2: When the model training strategy received by the terminal device is to send the updated model, the terminal device updates the model based on the local training sample data to obtain Model 2 (i.e., the above-mentioned second model), and sends Model 2 or the parameters of Model 2 to the base station.

[0354] Optionally, if the terminal device also receives a start time and an end time, and the start time and the end time are absolute timestamps, the terminal device starts uploading Model 2 or the parameters of Model 2 when the start time arrives, and finishes uploading Model 2 or the parameters of Model 2 before the end time arrives. If the end time arrives but Model 2 or the parameters of Model 2 have not been fully uploaded, the terminal device sends a model upload failure message to the base station.

[0355] Optionally, if the terminal device also receives a start time and an end time, and the start time and the end time are time offsets, the terminal device calculates the absolute timestamps of the start time and the end time based on the model training indication sent by the base station, starts uploading Model 2 or the parameters of Model 2 when the start time arrives, and finishes uploading Model 2 or the parameters of Model 2 before the end time arrives. If the end time arrives but Model 2 or the parameters of Model 2 have not been fully uploaded, the user equipment sends a model upload failure message to the base station. Optionally, the failure message contains an indication message for indicating that the reason for the model transmission failure is that the end time has arrived (or, the model has not been fully uploaded before the end time arrives).

[0356] Optionally, if the terminal device also receives duration information, the terminal device needs to upload Model 2 or the parameters of Model 2 within the duration. If Model 2 or the parameters of Model 2 have not been fully uploaded within the duration, the terminal device sends a model upload failure message to the base station. Optionally, the failure message contains an indication message for indicating that the reason for the model transmission failure is that the model has not been fully uploaded within the duration).

[0357] S908. The base station updates Model 1 to obtain an updated local model.

[0358] If the base station only receives local training sample data sent by multiple terminal devices, the base station updates the model based on these local training sample data to obtain an updated local model.

[0359] If the base station only receives Model 2 or the parameters of Model 2 sent by multiple terminal devices, the base station performs weighted averaging and aggregation on the parameters of these Model 2s to obtain an updated local model.

[0360] If the base station receives local training sample data sent by at least one terminal device and Model 2 or the parameters of Model 2 sent by at least one other terminal device, the base station updates the model based on the local training sample data to obtain Model 3, and performs weighted averaging and aggregation on Model 3 and the parameters of Model 2 to obtain an updated local model.

[0361] S909. The base station sends the updated local model to the application server.

[0362] S909 is an optional step.

[0363] In the embodiments of the present application, a wireless communication system and a wireless communication method applied to the wireless communication system are designed to jointly design communication and computing, so that the network can provide corresponding computing, storage, communication and other resource guarantees for AI services.

[0364] The network architecture includes a control plane, a computing plane, and a forwarding plane. Among them, the control plane is mainly responsible for the establishment, maintenance, and optimization of user sessions, and is composed of a service manager, a resource manager, a task manager, a model manager, and an access manager; the computing plane is mainly responsible for data management, data acquisition, data storage, data collaboration, etc., and is composed of a routine manager, a model base, and a distributed computing unit; the forwarding plane is responsible for the transmission of user data and is composed of a forwarding unit and a wireless access unit.

[0365] The network selects a terminal device as a routine client, and the terminal device sends local training sample data to the DCU for local model update by the DCU. Other terminal devices that are not selected as routine clients perform local model update based on local training sample data.

[0366] Based on the description information of the federated learning training service, the network determines the deployment mode of the federated learning training service and the corresponding computing plane, and reserves corresponding communication resources and computing resources for the computing plane.

[0367] Based on the problem of minimizing the model error in one iteration of the federated learning model training, a network parameter optimization problem is constructed to accelerate the convergence speed of the federated learning model training.

[0368] Fourthly, to implement the above wireless communication method, a device 1100 (the first access network device or the terminal device) in the embodiments of the present application, as Figure 11 shown, may include at least one processor 1101 and at least one transceiver 1102 coupled to at least one processor 1101. The transceiver 1102 may include at least one separate receiving circuit system and transmitting circuit system, or at least one integrated receiving circuit system and transmitting circuit system. At least one processor 1101 may be a central processing unit (CPU, Central Processing Unit), a microprocessor (MPU, Micro Processor Unit), a digital signal processor (DSP, Digital Signal Processor), or a field programmable gate array (FPGA, Field-Programmable Gate Array), etc.

[0369] According to some embodiments of the present application, when the device 1100 is the first access network device, the first access network device includes a first transceiver; and

[0370] Receive a first message sent by a terminal device via the first transceiver, the first message includes model training capability information, and the first message includes at least one of the following: a terminal capability information message, a terminal capability response message;

[0371] Based on the first message, determine model training policy indication information, which is used to instruct the terminal device to upload training data or upload the locally updated model;

[0372] Send a MAC CE to the terminal device via the first transceiver, and the MAC CE includes model training policy indication information;

[0373] Receive model data sent by the terminal device via the first transceiver, and the model data is data determined based on the model training policy indication information and related to the update of the first model.

[0374] In some embodiments, the model training capability information includes at least one of the following: the size of the local training data of the terminal device; the computing frequency of the terminal device; the central processing unit (CPU) cycles of the terminal device.

[0375] In some embodiments, the first message further includes at least one of the following of the terminal device: reference signal received power (RSRP), reference signal received quality (RSRQ).

[0376] In some embodiments, the first processor is configured to:

[0377] Receive model training requirement information sent by a core network device via the first transceiver;

[0378] Based on the model training requirement information and the first message, determine the model training policy indication information.

[0379] In some embodiments, the model training requirement information includes at least one of the following: a device list, model training latency information of a model training task, model accuracy information of a model training task;

[0380] Wherein, the device list includes at least one terminal device participating in the model training task;

[0381] The model training task includes multiple iterative trainings, and the training of the first model is the Nth iterative training in the multiple iterative trainings.

[0382] In some embodiments, the MAC CE further includes at least one of the following:

[0383] The first power information, which is used to indicate the transmission power when the terminal device uploads training data;

[0384] The second power information, which is used to indicate the transmission power when the terminal device uploads the model;

[0385] The time information, which is used to indicate the application time of the model training strategy;

[0386] The time indication information, which is used to indicate the representation method adopted by the time information;

[0387] The first model.

[0388] In some embodiments, the time information includes at least one of the following: start time and end time; duration.

[0389] In some embodiments, the time indication information is used to indicate the representation method adopted by the start time and the end time, and the representation methods of the start time and the end time include one of the following: absolute timestamp; time offset relative to the first time, where the first time is the time when the terminal device receives the model training strategy indication information.

[0390] In some embodiments, the first processor is configured to: send a second message to the terminal device via the first transceiver, where the second message is used to trigger the sending of the first message, and the second message includes one of the following: terminal capability query message, terminal capability request message.

[0391] In some embodiments, the second message includes at least one of the following: model training capability request information; the first threshold, which is used to control the sending of the first message in combination with the size of the training data of the terminal device; the first model.

[0392] In some embodiments, the first processor is configured to: when the model training strategy instructs the terminal device to upload training data, receive the training data uploaded by the terminal device via the first transceiver; update the first model based on the training data to obtain a second model.

[0393] In some embodiments, the first processor is configured to: when the model training strategy instructs the terminal device to upload the locally updated model, receive the second model uploaded by the terminal device via the first transceiver; the second model is obtained by the terminal device updating the first model based on the training data.

[0394] In some embodiments, the first processor is configured to: receive local models sent by each of one or more second access network devices via a first transceiver; determine a global model based on the local models sent by each of the one or more second access network devices and the local model of the first access network device, where the local model of the first access network device is determined based on model data; and send the global model to a terminal device via the first transceiver.

[0395] In some embodiments, the terminal device is served jointly by a plurality of distributed access network devices, the plurality of distributed access network devices including a first access network device and a third access network device; the first processor is configured to: receive model training configuration information sent by a core network device via the first transceiver, the model training configuration information indicating that the first access network device participates in the Nth iteration training of a model training task; where the distributed access network device participating in the Mth iteration training of the model training task is the first access network device or the third access network device.

[0396] According to some embodiments of the present application, when the device 1100 is a terminal device, the terminal device includes a second transceiver; and

[0397] a second processor, coupled to the second transceiver; the second processor is configured to:

[0398] send a first message to the first access network device via the second transceiver, the first message including model training capability information, the first message including at least one of the following: a terminal capability information message, a terminal capability response message;

[0399] receive MAC CE sent by the first access network device via the second transceiver, the MAC CE including model training policy indication information, the model training indication information being determined based on the first message, the model training policy indication information being used to instruct the terminal device to upload training data or upload a locally updated model;

[0400] determine model data based on the model training policy indication information, the model data being related to the update of a first model;

[0401] send the model data to the first access network device via the second transceiver.

[0402] The description of the above device embodiments is similar to the description of the above method embodiments and has similar beneficial effects to the method embodiments. For technical details not disclosed in the device embodiments of the present application, please refer to the description of the method embodiments of the present application for understanding.

[0403] It should be noted that in the embodiments of the present application, if the above-mentioned wireless communication method is implemented in the form of software function modules and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiments of the present application, in essence or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods of the various embodiments of the present application. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), magnetic disks, or optical discs that can store program codes. In this way, the embodiments of the present application are not limited to any specific combination of hardware and software.

[0404] In a fifth aspect, to implement the above-mentioned wireless communication method, an embodiment of the present application provides an electronic device, including a memory and a processor. The memory stores a computer program that can run on the processor, and when the processor executes the program, it implements the steps in the above-mentioned wireless communication method provided in the embodiments.

[0405] In a sixth aspect, an embodiment of the present application provides a storage medium, that is, a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps in the above-mentioned wireless communication method provided in the embodiments.

[0406] It should be pointed out here that: the descriptions of the above storage medium and device embodiments are similar to the descriptions of the above method embodiments and have beneficial effects similar to those of the method embodiments. For the technical details not disclosed in the storage medium and device embodiments of the present application, please refer to the descriptions of the method embodiments of the present application for understanding.

[0407] It should be understood that the "one embodiment" or "an embodiment" mentioned throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the present application. Therefore, the appearances of "in one embodiment" or "in some embodiments" throughout the specification do not necessarily refer to the same embodiment. In addition, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of the present application, the magnitudes of the sequence numbers of the above processes do not mean the order of execution, and the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application. The sequence numbers of the embodiments of the present application above are only for description and do not represent the advantages and disadvantages of the embodiments.

[0408] It should be noted that in this text, the term "including", "comprising", or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or device that includes a series of elements not only includes those elements but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article, or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article, or device that includes such element.

[0409] In several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. Additionally, the couplings, direct couplings, or communication connections between the various components shown or discussed can be through some interfaces, and the indirect couplings or communication connections of devices or units can be electrical, mechanical, or other forms.

[0410] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units; they can be located in one place or distributed to multiple network units; some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0411] In addition, each functional unit in the embodiments of this application can be fully integrated into one processing unit, or each unit can be separately regarded as one unit, or two or more units can be integrated into one unit; the above-mentioned integrated units can be implemented in the form of hardware, or in the form of a combination of hardware and software functional units.

[0412] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps including the above method embodiments; and the aforementioned storage medium includes: various media such as removable storage devices, read-only memory (ROM), magnetic disks, or optical discs that can store program codes.

[0413] Alternatively, if the above integrated units of the present application are implemented in the form of software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiments of the present application essentially or the part that contributes to the related art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the methods of the various embodiments of the present application. The foregoing storage medium includes: various media that can store program codes, such as removable storage devices, ROMs, magnetic disks, or optical discs.

[0414] The above is only the implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered by the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A first access network device, the first access network device comprising a first transceiver; and a first processor coupled to the first transceiver; the first processor being configured to: Receiving, via the first transceiver, a first message sent by a terminal device, wherein the first message includes model training capability information, and the first message includes at least one of the following: a terminal capability information message and a terminal capability response message; Based on the first message, determine model training strategy indication information, where the model training strategy indication information is used to instruct the terminal device to upload training data or upload a locally updated model; Sending a media access control control element MAC CE to the terminal device via the first transceiver, where the MAC CE includes the model training strategy indication information; Model data sent by the terminal device is received via the first transceiver, where the model data is data determined based on the model training strategy indication information and is related to the update of the first model.

2. The first access network device according to claim 1, wherein the model training capability information comprises at least one of the following: The local training data size of the terminal device; The calculation frequency of the terminal device; The central processing unit CPU cycles of the terminal device.

3. According to the first access network device according to claim 1, the first message also includes at least one of the following of the terminal device: reference signal received power RSRP, reference signal received quality RSRQ.

4. The first access network device according to claim 1, wherein the MAC CE further comprises at least one of the following: first power information, where the first power information is used to indicate the transmission power of the terminal device when uploading training data; Second power information, where the second power information is used to indicate the transmission power of the terminal device when uploading the model; Time information, the time information is used to indicate the application time of the model training strategy; Time indication information, the time indication information is used to indicate the representation method adopted by the time information; The first model.

5. The first access network device according to claim 1, wherein the first processor is configured to: A second message is sent to the terminal device via the first transceiver, where the second message is used to trigger the sending of the first message, and the second message includes one of the following: a terminal capability query message and a terminal capability request message.

6. The first access network device according to claim 1, wherein the first processor is configured to: When the model training strategy instructs the terminal device to upload training data, receiving the training data uploaded by the terminal device via the first transceiver; The first model is updated based on the training data to obtain a second model.

7. The first access network device according to claim 1, wherein the first processor is configured to: When the model training strategy instructs the terminal device to upload a locally updated model, receiving, via the first transceiver, a second model uploaded by the terminal device; The second model is obtained after the terminal device updates the first model based on training data.

8. The first access network device according to claim 1, 6 or 7, wherein the first processor is configured to: Receiving, via the first transceiver, a local model sent by each second access network device in one or more second access network devices; Determine a global model based on a local model sent by each of the one or more second access network devices and a local model of the first access network device, wherein the local model of the first access network device is determined based on the model data; The global model is transmitted to the terminal device via the first transceiver.

9. The first access network device according to claim 1, wherein the terminal device is provided with services by a plurality of distributed access network devices, and the plurality of distributed access network devices include the first access network device and a third access network device; The first processor is configured to: Receiving, via the first transceiver, model training configuration information sent by the core network device, wherein the model training configuration information instructs the first access network device to participate in an Nth iteration training of a model training task; in, The distributed access network device participating in the Mth iterative training of the model training task is the first access network device or the third access network device.

10. A terminal device, comprising a second transceiver; and a second processor coupled to the second transceiver; the second processor being configured to: Sending a first message to a first access network device via the second transceiver, wherein the first message includes model training capability information, and the first message includes at least one of the following: a terminal capability information message and a terminal capability response message; Receiving, via the second transceiver, a MAC CE sent by the first access network device, where the MAC CE includes model training strategy indication information, where the model training strategy indication information is determined based on the first message, and where the model training strategy indication information is used to instruct the terminal device to upload training data or upload a locally updated model; Based on the model training strategy indication information, determine model data, where the model data is related to updating of the first model; The model data is sent to the first access network device via the second transceiver.

Images