First access point device, network device and communication method

Abstract

The invention discloses a first access point device, a network device and a communication method, and the method comprises the steps: receiving a plurality of pieces of first information sent by a plurality of pieces of first terminal equipment through the first access point device; performing information integration on the plurality of pieces of first information to obtain second information; the first access point device sends second information to the network device; the second information is used for the network equipment to determine a first model or determine a model segmentation strategy for the first model; the first access point device receives a first model sent by the network device, and segments the first model according to the model segmentation strategy to obtain a plurality of sub-models; the first access point device sends a plurality of sub-models to a plurality of second terminal devices; wherein the plurality of first terminal devices comprise a plurality of second terminal devices.

Description

Technical Field

[0001] This application relates to, but is not limited to, the field of communication technology, and in particular to a first access point device, a network device, and a communication method. Background Technology

[0002] Currently, deploying federated learning for model training in non-cellular Multiple-Input Multiple-Output (MIMO) networks faces numerous challenges. On the one hand, the varying computing capabilities of user devices make it difficult to uniformly deploy complete models. On the other hand, existing solutions typically rely on centralized servers for model aggregation and segmentation, which cannot adapt to the decentralized and multi-point access characteristics of non-cellular networks, resulting in low model training efficiency and high communication overhead. Summary of the Invention

[0003] This application provides at least one first access point device, a network device, and a communication method.

[0004] The technical solution of this application embodiment is implemented as follows: This application provides a first access point device for use in a non-cellular MIMO scenario. The first access point device includes: First transceiver; and A first processor, coupled to a first transceiver, wherein the first processor is configured to: The system receives multiple first messages sent by multiple first terminal devices via a first transceiver. The second piece of information is obtained by integrating multiple pieces of primary information. The second information is sent to the network device via the first transceiver; the second information is used by the network device to determine the first model, or to determine the model segmentation strategy for the first model. The first model is received from the network device via the first transceiver. Based on the model segmentation strategy, the first model is segmented to obtain multiple sub-models; Multiple sub-models are sent to multiple second terminal devices via the first transceiver; Among them, multiple first terminal devices include multiple second terminal devices.

[0005] In some embodiments, the multiple first pieces of information are multiple device information of multiple first terminal devices; the first processor is configured to: Information from multiple devices is integrated to obtain integrated information from multiple devices. The second piece of information includes: Information from multiple integrated devices; Device information for the first access point device.

[0006] In some embodiments, the device information includes one or more of the following: Equipment identification information; Channel environment information; Device communication capability information; Equipment computing capacity information.

[0007] In some embodiments, the first processor is configured to: The first transceiver receives computing plane configuration information sent by the network device based on the second information; the computing plane configuration information includes one or more of the following: Model segmentation strategy; Model aggregation order, which indicates the order of aggregation within an access point cluster and aggregation between access point clusters; First indication information, the first indication information is used to indicate multiple second terminal devices from multiple first terminal devices; The second instruction information is used to instruct the first access point device to perform model training using idle resources, which include one or more of communication idle resources, computing idle resources, and storage idle resources.

[0008] In some embodiments, the first processor is configured to: The first transceiver receives user plane configuration information sent by the second information transmission device of the network device; the user plane configuration information includes one or more of the following: Multiple synchronization parameters are associated with multiple second terminal devices, and these parameters are used to enable multiple second terminal devices to simultaneously perform training on the sub-model. Multiple third indication information, which are associated with multiple second terminal devices, are used to indicate the idle resources required by the second terminal device to perform model training. The idle resources include one or more of the following: communication idle resources, computing idle resources, and storage idle resources.

[0009] In some embodiments, the plurality of first terminal devices are a plurality of second terminal devices, and the plurality of first information are a plurality of activation data obtained by the plurality of second terminal devices performing forward propagation on the received sub-model using their own user data; the first processor is configured to: Multiple initial models are obtained by backpropagating multiple residual models using multiple activation data; the multiple residual models are the parts of the first model excluding multiple sub-models. Perform intra-cluster aggregation on multiple initial models to obtain the first global model; The second piece of information includes the model information of the first global model.

[0010] In some embodiments, the first processor is configured to: Given an initial model, perform one intra-cluster aggregation. After performing multiple intra-cluster aggregations, obtain the first global model. or, Given multiple initial models, perform one intra-cluster aggregation to obtain the first global model.

[0011] In some embodiments, the first processor is configured to: Under a first condition, multiple fourth indication messages are sent to multiple second terminal devices via a first transceiver; the multiple fourth indication messages are used to instruct the multiple second terminal devices to send multiple activation data. The first condition is that the first access point device receives multiple fifth indication messages sent by multiple second terminal devices to indicate that the model training is complete.

[0012] In some embodiments, the first processor is configured to: Multiple sixth instruction messages are sent to multiple second terminal devices via the first transceiver; each sixth instruction message includes a first time window, and each sixth instruction message is used to instruct the corresponding second terminal device to send the corresponding activation data after the model training is completed and the first time window is waited.

[0013] In some embodiments, the first processor is configured to: Multiple seventh indication messages are sent to multiple second terminal devices via the first transceiver; each seventh indication message includes a second time window, and each seventh indication message is used to instruct the corresponding second terminal device to send the corresponding activation data within the second time window after the model training is completed.

[0014] In some embodiments, the first processor is configured to: The first global model is sent to the network device via the first transceiver; Among them, the first global model and other global models sent by other access point devices associated with the network device are used by the network device to perform inter-cluster aggregation to obtain the second model, which is used to replace the first model to perform the next round of model training.

[0015] This application provides a network device for use in a non-cellular MIMO scenario. The network device includes: Second transceiver; and A second processor, coupled to a second transceiver, wherein the second processor is configured to: The second transceiver receives the second information sent by the first access point device. Perform inter-cluster aggregation based on the second information to obtain the first model, or determine the model segmentation strategy for the first model based on the second information; The first model is sent to the first access point device via the second transceiver, or the first model and model segmentation strategy are sent.

[0016] This application provides a communication method applied to a first access point device in a non-cellular MIMO scenario, the method comprising: Receive multiple first messages sent by multiple first terminal devices; integrate the multiple first messages to obtain second messages; Send second information to the network device; the second information is used by the network device to determine the first model, or to determine a model segmentation strategy for the first model; Receive the first model sent by the network device; divide the first model into multiple sub-models according to the model segmentation strategy; Multiple sub-models are sent to multiple second terminal devices; the multiple first terminal devices include multiple second terminal devices.

[0017] This application provides a communication method applied to a network device in a non-cellular MIMO scenario. The method includes: Receive the second information sent by the first access point device; Perform inter-cluster aggregation based on the second information to obtain the first model, or determine the model segmentation strategy for the first model based on the second information; Send the first model to the first access point device, or send the first model and model segmentation strategy.

[0018] In this embodiment, a first access point device receives multiple first pieces of information sent by multiple first terminal devices; integrates the multiple first pieces of information to obtain second information; the first access point device sends the second information to a network device; the second information is used by the network device to determine a first model, or to determine a model segmentation strategy for the first model; the first access point device receives the first model sent by the network device, and segments the first model according to the model segmentation strategy to obtain multiple sub-models; the first access point device sends the multiple sub-models to multiple second terminal devices; wherein, the multiple first terminal devices include multiple second terminal devices; in this application, the access point device receives information from multiple first terminal devices, integrates the information, and sends it to the network device, thereby enabling the network device to determine the model or model segmentation strategy based on this information. Then, the access point device segments the model into multiple sub-models according to the model segmentation strategy and distributes them to the second terminal devices. This approach, because the access point device undertakes part of the model forward propagation and all of the gradient backpropagation tasks, significantly improves the overall efficiency of model training compared to the traditional federated learning method where terminal devices directly upload complete model parameters.

[0019] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of this application. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the technical solutions of this application.

[0021] Figure 1 A flowchart illustrating a communication method provided in an embodiment of this application; Figure 2 This application provides an exemplary illustration of distributed cooperative transmission. Figure 1 ; Figure 3 This application provides an exemplary illustration of distributed cooperative transmission. Figure 2 ; Figure 4 This application provides an exemplary illustration of distributed cooperative transmission. Figure 3 ; Figure 5 A schematic diagram of an exemplary federated segmentation learning architecture provided for embodiments of this application; Figure 6 A flowchart illustrating an exemplary communication method provided in this application embodiment. Figure 1 ; Figure 7 A flowchart illustrating an exemplary communication method provided in this application embodiment. Figure 2 ; Figure 8 This is a schematic diagram of the composition of a device 800 provided in an embodiment of this application.

[0022] It should be noted that the terms "first" and "second" mentioned above are only used to distinguish between different options and do not represent the degree of superiority or inferiority of the options or their priority in the implementation process. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application are further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0025] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0026] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0027] This application provides a first access point device, a network device, and a communication method.

[0028] Figure 1 This application illustrates a communication method provided by an embodiment of the present application, applied to a first access point device. The method includes the following steps S100 to S1: S100: Receive multiple first messages sent by multiple first terminal devices; integrate the multiple first messages to obtain second messages.

[0029] In this embodiment of the application, the first access point device receives multiple first information messages sent by multiple first terminal devices, integrates the multiple first information messages, and obtains second information messages.

[0030] It should be noted that the first terminal device can be any user device with computing capabilities, such as a smartphone, tablet, smart wearable device, etc. These devices are deployed on the user plane and used to perform local model training; the first terminal device can be determined according to the actual situation, and this embodiment does not limit it.

[0031] In some embodiments, the first access point device may be connected to multiple first terminal devices, thereby enabling the first access point device to receive multiple first messages sent by the multiple first terminal devices with which it has a communication connection.

[0032] It should be noted that the first information sent by each first terminal device can be the device information of each first terminal device, or it can be the activation data obtained by each first terminal device after performing the previous round of model training. The first information sent by each first terminal device is different.

[0033] In some embodiments, after receiving multiple pieces of first information, the first access point device selects different information integration methods according to the different pieces of first information. If the first information is device information of a terminal device, the first access point device can summarize the multiple pieces of device information to obtain the summarized device information, i.e., the second information. Alternatively, if the first information is activation data of a terminal device, the first access point device can perform model training based on the activation data and then perform intra-cluster model aggregation to obtain a first global model, i.e., the second information. The first terminal device can be determined according to the actual situation, and this embodiment does not limit it.

[0034] S101. Send second information to the network device; the second information is used by the network device to determine the first model, or to determine the model segmentation strategy for the first model.

[0035] In this embodiment of the application, after receiving the second information, the first access point device sends the second information to the network device; the second information is used by the network device to determine the first model, or to determine a model segmentation strategy for the first model.

[0036] In some embodiments, if the second information is aggregated device information, then the second information is used by the network device to determine the model segmentation strategy for the first model, where the first model is the initial model, i.e., the model that has not been trained.

[0037] In other embodiments, if the second information is a first global model, then the second information is used by the network device to generate the first model, which can be a model obtained after training the initial model.

[0038] S102, Receive the first model sent by the network device; Divide the first model into multiple sub-models according to the model segmentation strategy.

[0039] In this embodiment of the application, the first access point device receives a first model sent by the network device; according to the model segmentation strategy, the first model is segmented to obtain multiple sub-models.

[0040] In some embodiments, when the second information is used to determine the model segmentation strategy, the network device also needs to send the model segmentation strategy to the first access point device.

[0041] It's important to note that the model segmentation strategy determines how the first model is split into multiple sub-models. For example, if the first model is a deep neural network, the model segmentation strategy might divide the first model into two parts at a specific layer (called the split point): one part is deployed on the first terminal device (user sub-model), and the other part is deployed on the first access point device (access point sub-model). The model segmentation strategy can be static or dynamic, and it can be adjusted in real time according to the capabilities of the first terminal device and the network environment.

[0042] It should also be noted that the model segmentation strategy can include sub-segmentation strategies for multiple second terminal devices. The sub-segmentation strategies for each second terminal device can be the same or different. For example, the sub-segmentation strategy for second terminal device A is to use the third layer of the first model as the segmentation point. In this case, the first three layers of the first model can be determined as the sub-model of second terminal device A. The sub-segmentation strategy for second terminal device B is to use the fourth layer of the first model as the segmentation point. In this case, the first four layers of the first model can be determined as the sub-model of second terminal device B. That is, when performing model segmentation for each second terminal device, it is based on a complete first model.

[0043] S103, Send multiple sub-models to multiple second terminal devices; the multiple first terminal devices include multiple second terminal devices.

[0044] In this embodiment of the application, the first access point device sends multiple sub-models to multiple second terminal devices; the multiple first terminal devices include multiple second terminal devices.

[0045] It should be noted that when the first access point device performs model segmentation for each second terminal device, it can obtain a sub-model and a remaining model corresponding to each second terminal device. Here, a sub-model and a remaining model constitute a complete first model. Therefore, after obtaining multiple sub-models corresponding to multiple second terminal devices, the first access point device can also obtain multiple remaining models corresponding to multiple second terminal devices. At this time, the first access point device can send multiple sub-models to multiple second terminal devices and then store multiple remaining models locally.

[0046] In some embodiments, the plurality of first terminal devices may be a plurality of second terminal devices, or the plurality of first terminal devices may include a plurality of second terminal devices.

[0047] The communication method proposed in this application will be described below through two implementation methods, namely, the first information being device information or activation data.

[0048] In one implementation, the first information is the device information of the first terminal device.

[0049] In some embodiments, the multiple first information are multiple device information of multiple first terminal devices. In this case, the above step S100 "integrating the multiple first information to obtain second information" includes: integrating the multiple device information to obtain integrated multiple device information.

[0050] It should be noted that the device information of the first terminal device may include at least one or more of the following: device identification information of the first terminal device, channel environment information of the first terminal device, device communication capability information of the first terminal device, and device computing capability information of the first terminal device.

[0051] Among them, the device identification information of the first terminal device is used to uniquely identify the identity of the first terminal device, such as: the Media Access Control (MAC) address, International Mobile Equipment Identity (IMEI), serial number, etc. of the first terminal device. The device identification information helps the first access point device or network device to correctly identify each terminal device participating in the training.

[0052] Channel environment information from the first terminal device represents the quality of the wireless link between the first terminal device and the first access point device. Examples include Received Signal Strength Indicator (RSSI), Signal-to-Noise Ratio (SNR), and Bit Error Rate (BER). Channel environment information is crucial for non-cellular MIMO networks because it directly impacts the reliability and efficiency of data transmission. By acquiring and analyzing channel environment information, network devices can dynamically adjust resource allocation strategies, such as selecting more suitable access point devices and optimizing data transmission rates, to improve overall communication performance and the convergence speed of federated learning.

[0053] The device communication capability information of the first terminal device represents its wireless communication capabilities, including the maximum supported data transmission rate, supported communication protocol versions, transmission power levels, and frequency band support. This information helps network devices assess the communication feasibility of each terminal device during the federated learning process, such as determining whether it has sufficient bandwidth to upload model parameters or whether it supports specific encoding methods to improve transmission efficiency. This information also allows for prioritizing terminal devices with stronger communication capabilities when deploying federated learning tasks, thereby improving the overall efficiency and stability of model training.

[0054] The device computing power information of the first terminal device indicates its computing resources when performing model training tasks locally, such as CPU frequency, memory size, GPU computing power, and storage space. This information helps network devices assess whether each terminal device can handle local model training tasks and whether load balancing or resource scheduling is necessary. For example, on terminal devices with lower computing power, the model structure can be simplified or the number of training epochs reduced to accommodate the device's computing capabilities while ensuring the overall progress and quality of federated learning.

[0055] It is understandable that the device information of the first terminal device can be used to better configure the model training task for network devices or access point devices. If the device information of a certain first terminal device indicates that the computing power of the first terminal device is weak, then it is advisable not to configure a model training task for the first terminal, or to configure a model training task with a lower computational load for it.

[0056] In some embodiments, the second information may include: integrated information on multiple devices and device information of the first access point device.

[0057] It should be noted that the device information of the first access point device may include at least one or more of the following: device identification information of the first access point device, channel environment information of the first access point device, device communication capability information of the first access point device, and device computing capability information of the first access point device.

[0058] Among them, the device identification information of the first access point device is used to uniquely identify the identity of the first access point device, such as the MAC address, IMEI, serial number, etc. of the first access point device. The device identification information helps the network device to correctly identify each access point device participating in the training.

[0059] Channel environment information from the first access point device (BNPB) indicates the quality of the wireless link between the BNPB and network devices, such as RSSI, SNR, and BER. This information is crucial for non-cellular MIMO networks because it directly impacts the reliability and efficiency of data transmission. By acquiring and analyzing this information, network devices can dynamically adjust resource allocation strategies, such as selecting more suitable BNPBs and optimizing data transmission rates, to improve overall communication performance and the convergence speed of federated learning.

[0060] The device communication capability information of the first access point device represents its wireless communication capabilities, including the maximum supported data transmission rate, supported communication protocol versions, transmission power levels, and frequency band support. This information helps network devices assess the communication feasibility of each access point device during the federated learning process, such as determining whether it has sufficient bandwidth to upload model parameters or whether it supports specific encoding methods to improve transmission efficiency. This information also allows for prioritizing access point devices with stronger communication capabilities when deploying federated learning tasks, thereby improving the overall efficiency and stability of model training.

[0061] The device computing power information of the first access point device indicates its computing resources when performing model training tasks locally, such as CPU frequency, memory size, GPU computing power, and storage space. This information helps network devices assess whether each access point device can handle local model training tasks and whether load balancing or resource scheduling is necessary. For example, on access point devices with lower computing power, the model structure can be simplified or the number of training epochs reduced to suit the device's computing capabilities, while ensuring the overall progress and quality of federated learning.

[0062] It is understandable that the device information of the first access point device can be used to better configure the model training task for the network device. If the device information of a certain first access point device indicates that the computing power of the first access point device is weak, then it is advisable not to configure a model training task for the first access point device, or to configure a model training task with lower computational load for it.

[0063] In some embodiments, after the first access point device sends the second information to the network device, the network device can determine the configuration information corresponding to each access point device based on the second information sent by the first access point device and the second information sent by other access point devices, and send the configuration information corresponding to each access point device to the corresponding access point device. At this time, the configuration information is used by each access point device to perform model training.

[0064] It should be noted that the configuration information may include configuration information for the computing plane and configuration information for the user plane.

[0065] In some embodiments, in this implementation, after step S101, the first access point device may further perform the following step: receiving the configuration information of the computing plane sent by the network device according to the second information.

[0066] In some embodiments, the configuration information of the computing plane includes one or more of the following: model segmentation strategy; model aggregation order, which indicates the order of aggregation within an access point cluster and aggregation between access point clusters; first indication information, which indicates a plurality of second terminal devices from a plurality of first terminal devices; and second indication information, which indicates the idle resources required by the first access point devices to perform model training, wherein the idle resources include one or more of communication idle resources, computing idle resources, and storage idle resources.

[0067] The model segmentation strategy refers to the method of splitting the first model into sub-models and the remaining models. Different terminal devices can be configured with the same model segmentation strategy or different model segmentation strategies, depending on the local computing power and network conditions of the terminal device. For example, for terminal devices with weak computing power, the network device can choose to keep more model layers on the access point device side, thereby reducing the computing burden on such terminal devices.

[0068] Model aggregation order refers to the order in which access point devices aggregate locally updated terminal device models during model training. Typically, intra-cluster aggregation occurs first, followed by inter-cluster aggregation. Intra-cluster aggregation refers to the aggregation operation between the initial models of all terminal devices served by the same access point device, while inter-cluster aggregation refers to model aggregation between different access point devices.

[0069] The first instruction information is used to select a number of second terminal devices from multiple first terminal devices to participate in the current round of model training. The first instruction information can be dynamically allocated based on factors such as the activity level, historical performance, and computing power of the terminal devices. For example, at a certain moment, the system may only allow some high-priority terminal devices to participate in model training to save communication and computing resources.

[0070] The second indication information is used to indicate the available idle resources of the first access point device when performing model training. Idle resources include idle communication resources (such as unused wireless spectrum), idle computing resources (such as idle CPU / GPU processing power), and idle storage resources (such as unused memory or hard disk space). The second indication information can help the access point device to allocate tasks reasonably under resource constraints and ensure the stability and efficiency of the model training process.

[0071] In some embodiments, after step S101, the first access point device may further perform the following steps: receiving user plane configuration information sent by the network device according to the second information.

[0072] In some embodiments, the configuration information of the user plane includes one or more of the following: multiple synchronization parameters associated with multiple second terminal devices, the multiple synchronization parameters being used for multiple second terminal devices to simultaneously perform training for a sub-model; multiple third indication information associated with multiple second terminal devices, the third indication information associated with each second terminal device being used to indicate the idle resources required by the second terminal device to perform model training, the idle resources including one or more of communication idle resources, computing idle resources, and storage idle resources.

[0073] The synchronization parameter is a timing or sequence parameter used to coordinate the initiation of model training by multiple second terminal devices in the same round. Multiple second terminal devices typically need to start local training at the same time point to ensure consistency in the aggregation process. In this application, due to the adoption of a federated segmented learning architecture, the second terminal devices only handle a portion of the forward propagation task of the sub-model. Therefore, it is necessary to accurately synchronize the training start times of multiple second terminal devices to ensure efficient uplink transmission of activation data and further processing at the access point device.

[0074] The third instruction information is used to inform the second terminal device of available idle resources during model training. This third instruction information is dynamically generated by the network device based on the second terminal device's historical usage, current load status, and anticipated future tasks, and is then sent to the second terminal device via user plane configuration information.

[0075] It should be noted that idle resources include three categories: communication idle resources, computing idle resources, and storage idle resources. Communication idle resources refer to the bandwidth and time window available for uplink or downlink data transmission on the wireless link for the second terminal device; computing idle resources refer to the CPU / GPU computing power available to the second terminal device, including currently unused processing units and computing power expected to be available in the future; storage idle resources refer to the memory or hard disk space that the second terminal device can use to cache activation data or other intermediate results.

[0076] In this implementation, the first access point device can first determine multiple second terminal devices from multiple first terminal devices based on the received configuration information, and then perform model segmentation once for each second terminal device according to the configuration information. After performing multiple model segmentation, multiple sub-models corresponding to multiple second terminal devices are obtained, and then the multiple sub-models are sent to the multiple second terminal devices accordingly.

[0077] In this implementation, the first model can be pre-sent by the network device, or the first model can be sent along with the configuration information.

[0078] In another implementation, the first information is the activation data of the first terminal device.

[0079] This implementation method can be used in combination with the previous implementation method, or it can be used alone. The following is an introduction to this implementation method from the perspective of combination.

[0080] Based on the former implementation, after receiving multiple sub-models, each second terminal device can use its own user data to perform forward propagation on the received sub-models to obtain activation data, that is, multiple second terminal devices obtain multiple activation data.

[0081] In some embodiments, when multiple second terminal devices send multiple activation data, the following sending methods may be used: The first transmission method: Under the first condition, the first access point device sends multiple fourth indication messages to multiple second terminal devices; the multiple fourth indication messages are used to instruct the multiple second terminal devices to send multiple activation data; wherein, the first condition is that the first access point device receives multiple fifth indication messages sent by the multiple second terminal devices to indicate that the model training is complete.

[0082] In this transmission method, after each second terminal device completes local model training, it sends a fifth indication message to the first access point device to notify the first access point device that the second terminal device has completed the model training task. If the first access point device receives multiple fifth indication messages, it can send a fourth indication message to each second terminal device to instruct the second terminal device to upload its own activation data; or, if the first access point device receives a fifth indication message, it can send a fourth indication message to the second terminal device that sent the fifth indication message to instruct the second terminal device to upload its own activation data.

[0083] The second sending method: The first access point device sends multiple sixth indication messages to multiple second terminal devices; each sixth indication message includes a first time window, and each sixth indication message is used to instruct the corresponding second terminal device to send the corresponding activation data after the model training is completed and the first time window is waited.

[0084] It should be noted that the sixth indication information can be carried in the user plane configuration information sent by the network device to the first access point device. Thus, the first access point device can forward multiple sixth indication information of multiple second terminal devices to multiple second terminal devices. For a certain second terminal device, after generating activation data, the second terminal device waits for a first time window according to the received sixth indication information, and then sends its own activation data to the first access network device.

[0085] The third sending method: The first access point device sends multiple seventh indication messages to multiple second terminal devices; each seventh indication message includes a second time window, and each seventh indication message is used to instruct the corresponding second terminal device to send the corresponding activation data within the second time window after the model training is completed.

[0086] It should be noted that the seventh indication information can be carried in the user plane configuration information sent by the network device to the first access point device. Thus, the first access point device can forward multiple seventh indication information of multiple second terminal devices to multiple second terminal devices. For a certain second terminal device, after generating activation data, the second terminal device sends its own activation data to the first access network device within the second time window according to the received seventh indication information.

[0087] The fourth method of transmission: After receiving its own activation data, each second terminal device immediately sends the activation data to the first access point device.

[0088] It is understandable that this application uses various methods to flexibly coordinate the timing of activation data reporting by different second terminal devices, thereby avoiding channel conflicts and resource competition, and thus improving the efficiency of model training.

[0089] In this implementation, the multiple first terminal devices are multiple second terminal devices. That is, the multiple first information of the multiple first terminal devices are multiple activation data sent by the multiple second terminal devices. At this time, the step S100 above, "integrating the multiple first information to obtain second information", includes: using the multiple activation data to backpropagate multiple remaining models to obtain multiple initial models; performing intra-cluster aggregation on the multiple initial models to obtain a first global model.

[0090] It should be noted that the second piece of information is the model information of the first global model.

[0091] It should also be noted that the multiple remaining models are the multiple remaining models obtained by the first access point device after performing model segmentation and obtaining multiple sub-models corresponding to multiple second terminal devices. The sub-models and remaining models corresponding to each second terminal device can form a complete first model.

[0092] In some embodiments, after receiving activation data sent by a second terminal device, the first access point device can first perform forward propagation based on the remaining model of the second terminal device stored in its own memory to generate a prediction result. Then, the first access point device calculates the loss function based on the error between the prediction result and the true value, and uses backpropagation to update the gradient of the remaining model stored in its own memory, thereby generating an initial model corresponding to the second terminal device. By performing the above steps on multiple second terminal devices, multiple initial models are obtained.

[0093] Based on the above embodiments, after obtaining multiple initial models, the first access point device can perform intra-cluster aggregation of the access point device (which may be indicated in the above model aggregation order) to obtain the first global model.

[0094] In some embodiments, when performing intra-cluster aggregation, the first access point device may perform intra-cluster aggregation once for each initial model obtained, and obtain a first global model by performing intra-cluster aggregation multiple times.

[0095] It should be noted that, when the first access point device generates the first two initial models, it performs an intra-cluster aggregation once to obtain a global model. Then, when the third initial model is generated, it performs an intra-cluster aggregation once on the third initial model and the obtained global model to obtain a second global model, and so on until the last initial model is obtained. Based on the last initial model, the last intra-cluster aggregation is performed, and the resulting global model is determined as the first global model.

[0096] In other embodiments, when performing intra-cluster aggregation, the first access point device may perform intra-cluster aggregation once to obtain a first global model, provided that multiple initial models have been obtained.

[0097] It should be noted that, when the first access point device generates multiple initial models corresponding to multiple second terminal devices, it performs an intra-cluster aggregation once to obtain the first global model.

[0098] In some embodiments, intra-cluster aggregation can be achieved by weighted averaging of the model parameters to be aggregated, or it can be achieved in other ways, which are not limited in this embodiment.

[0099] Based on the above embodiments, in this implementation, the step S101 of "sending second information to the network device" may include: sending a first global model to the network device, whereby the first global model is used to determine the first model.

[0100] Alternatively, this implementation is based on the steps executed after step S103 above. In this case, the first global model is used to determine the second model, which can be used to replace the first model to perform the next round of model training.

[0101] In this implementation, the first model / second model is obtained by the network device through inter-cluster aggregation based on the first global model sent by the first access point device and other global models sent by other access point devices.

[0102] It should be noted that when network devices perform inter-cluster aggregation of multiple global models, there are also two execution methods.

[0103] In some embodiments, when performing inter-cluster aggregation, the network device may perform inter-cluster aggregation once for each global model received, and obtain a first model / second model by performing multiple inter-cluster aggregations.

[0104] It should be noted that when the network device receives the first two global models, it performs an inter-cluster aggregation once to obtain an aggregated model. Then, when it receives the third global model, it performs an inter-cluster aggregation once on the third global model and the obtained aggregated model to obtain a second aggregated model, and so on until the last global model is obtained. When the last inter-cluster aggregation is performed based on the last global model, the obtained aggregated model is determined as the first model / second model.

[0105] In other embodiments, when performing inter-cluster aggregation, the network device may perform inter-cluster aggregation once to obtain a first model / second model if it receives multiple global models.

[0106] It should be noted that when a network device receives multiple global models corresponding to multiple second terminal devices, it performs an inter-cluster aggregation to obtain the first model / second model.

[0107] In some embodiments, inter-cluster aggregation can be achieved by weighting the model parameters to be aggregated, or it can be achieved in other ways, which are not limited in this embodiment.

[0108] In this application, the access point device receives information from multiple first terminal devices, integrates the information, and sends it to the network device. This enables the network device to determine the model or model segmentation strategy based on this information. Subsequently, the access point device segments the model into multiple sub-models according to the model segmentation strategy and distributes them to the second terminal devices. This approach significantly improves the overall efficiency of model training compared to traditional federated learning, where terminal devices directly upload complete model parameters, because the access point device undertakes part of the model forward propagation and all of the gradient backpropagation tasks.

[0109] The following describes the application of the embodiments of this application in a real-world scenario.

[0110] Currently, with the rapid development of intelligent user devices (also known as terminal devices) and the increasing prevalence of artificial intelligence (AI) applications, more and more terminal data is being sent to edge servers to participate in AI model training on the network side. However, because terminal data involves personal privacy or confidentiality agreements, uploading terminal data to servers can lead to the leakage of personal privacy. To protect personal privacy, Federated Learning (FL) is widely considered a potential technology.

[0111] Federated learning is a distributed model training method where user devices train the model locally and only need to send the updated model parameters to the server, without needing to send the local training data to the server. Federated learning includes the following main steps: Step 1: The server randomly selects multiple user devices and sends the initial model to the user devices.

[0112] Step 2: The user equipment trains the initial model based on local training data to obtain the updated model.

[0113] Step 3: The user device sends the updated model parameters (or the updated model itself) to the server.

[0114] Step 4: The server performs a weighted average and aggregation of 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.

[0115] In existing technologies, the communication process between user devices is synchronous, meaning that each user device needs to obtain the same initial model for each model training iteration. On the other hand, the computation process only occurs between the user devices and the centralized processing unit, meaning that only the user devices and the centralized processing unit in the network are aware of the model training, while other network devices are unaware of it.

[0116] Furthermore, when deploying federated learning in wireless networks, existing architectures struggle to support the optimization of model training.

[0117] In federated learning, model training parameters need to be transmitted via wireless links, so the quality of the wireless links affects the quality of model training. However, existing wireless resource allocation techniques typically only consider communication efficiency, without considering the efficiency of model training, such as model accuracy.

[0118] Due to limited wireless bandwidth, base stations typically select user equipment with better wireless link quality for local model training, rather than user equipment with limited wireless link quality, resulting in insufficient generalization performance of the model.

[0119] In federated learning, user devices train a model using their local data, update the model parameters (e.g., update the model's gradients or weights), and then send the updated model parameters to a server. The server aggregates the received model parameters to generate a global model, thus ensuring global consistency of the model. To ensure global model convergence, user devices participating in federated learning need to frequently upload and download updated model parameters via wireless communication links during the training process.

[0120] Cellular networks are considered a future network architecture. Unlike existing cell-centric wireless communication networks, cellular networks are a user-centric wireless network architecture.

[0121] Traditional wireless communication networks employ a cell-centric wireless network architecture. Cellular-free MIMO, on the other hand, is a user-centric wireless network architecture. By deploying a large number of distributed access points and introducing cooperation among these access points to effectively eliminate inter-cell interference, it breaks down the cell boundaries of traditional cellular networks, providing a completely new network coverage and user service model. Cellular-free MIMO redefines the cell concept in traditional cellular networks from the user's perspective, eliminating hard cell divisions and making the network a borderless cell. This allows users to seamlessly switch between different access points and achieves higher macro-diversity gain from distributed antennas. Furthermore, because access points are deployed close to the user, the communication distance between the user and the network is reduced, thereby improving spectral efficiency, reducing data transmission latency, and increasing energy efficiency.

[0122] The core idea of ​​cellular-free MIMO is to adopt a flexible cell concept at the user level. That is, the network dynamically selects access points in the vicinity of the user to serve the user based on the user's service needs, location, and mobility. From the user's perspective, as the terminal moves, its associated access points change dynamically; from the network's perspective, each access point is dynamically associated with multiple terminals.

[0123] When terminals reuse the same time-frequency resources, distributed cooperative transmission needs to flexibly support the scalability of both the user base and the number of cooperative access points. Distributed cooperative transmission has the following three methods: Method 1: Scalable Dynamic Cooperative Clustering (DCC).

[0124] refer to Figure 2The access point sends the received user signals to the Central Processing Unit (CPU), which then performs dynamic collaborative sending and receiving. The CPU is typically deployed in the cloud.

[0125] Method 2: Fully distributed collaboration.

[0126] refer to Figure 3 The access point locally performs coherent transmission and reception of multiple user signals, while the central processing unit performs downlink data distribution and uplink data merging.

[0127] Method 3: Dynamic collaboration based on fixed cluster deployment.

[0128] refer to Figure 4 The access point sends user signals to the Edge Distributed Unit (EDU), which handles the correlation of multiple user signals transmission and reception. User-centric distributed units (UCDUs) are deployed in the cloud, where user data processing (e.g., downlink data distribution and uplink data merging) is implemented.

[0129] However, since existing non-cellular networks were not designed with efficient AI capabilities and services in mind, directly deploying federated learning on existing non-cellular network architectures presents the following challenges: For federated learning, the frequent transmission of high-dimensional model parameters between a large number of user devices and servers leads to huge communication overhead, while the flexible multi-point access method of non-cellular networks will bring higher inter-user interference, thus exacerbating communication overhead. Therefore, federated learning needs to be more adapted to non-cellular network architectures to reduce communication overhead and improve service quality.

[0130] Due to the heterogeneous computing and storage capabilities of user devices, deploying a complete federated learning model on user devices may be challenging. Furthermore, the flexible user access methods of the non-cellular network architecture offer users greater freedom in how they participate in federated learning, and the computing power on the non-cellular network side also needs to be fully utilized to improve service quality. Therefore, the federated learning architecture needs to be adjusted to better adapt to the non-cellular network architecture.

[0131] The following introduces segmentation learning. Segmentation learning is a distributed model training method designed to provide a collaborative paradigm for deploying neural network models. Its core idea is to break down a complete neural network model into multiple modules (called sub-modules or sub-models), which can be distributed across different user devices and edge servers for processing. By exchanging their respective outputs, user devices and edge servers can collaboratively complete the neural network model training task. Segmentation learning includes the following main steps: Step 1: The server selects a single user device and sends the segmented sub-modules to the user device.

[0132] Step 2: The user equipment performs forward transmission of the submodule based on the local training data, that is: input the local training data, start the calculation from the input layer of the submodule and transmit it to the next layer, calculate layer by layer, and thus obtain the activation data.

[0133] Step 3: The user equipment sends the activation data to the server, and the server then forwards the remaining sub-modules (the modules in the global model excluding the user equipment sub-module).

[0134] Step 4: The server calculates the difference between the model's predicted value and the true value, i.e., calculates the model's loss function, and uses the chain rule to calculate the loss function in reverse from the output layer to the input layer, obtaining the gradient of each parameter (weight and bias) in the network, i.e., gradient backpropagation.

[0135] Step 5: The server sends the gradient to the user device to update the local sub-model.

[0136] It is worth noting that the server only stores and updates the sub-modules of the global model (i.e., the remaining sub-modules in step 3), while the sub-modules in the user devices are updated through the backpropagated gradients. If there are multiple user devices, the server will store different sub-modules and update them separately.

[0137] In existing technologies, to introduce segmentation learning techniques into federated learning training for multiple user devices, a common server is typically used to receive activation data from each user device, and all users employ the same segmentation method. Therefore, the sub-modules of each user device can be updated by sending average gradients, thus completing the multi-user federated learning training process. However, deploying segmentation-based federated learning in non-cellular networks is challenging due to its decentralized nature, making it difficult to support a model training process where multiple users use a unified model segmentation method.

[0138] Therefore, to address the aforementioned pain points, this application proposes a communication method that, while simultaneously considering federated learning and segmentation learning, jointly designs communication and computation, enabling the network to provide corresponding computation, storage, and communication resource guarantees for the service quality requirements of federated learning (specific requirements for latency, jitter, packet loss rate, reliability, etc.), thereby improving the deterministic service level agreement (SLA).

[0139] The basic concepts involved in the communication method proposed in this application are introduced below: 1. User cluster: The set of user devices (i.e., terminal devices mentioned in the above embodiments) associated with a single access point (i.e., the access network device mentioned in the above embodiments).

[0140] 2. Access point cluster: The set of access points associated with a single user equipment.

[0141] 3. Intra-cluster aggregation: Aggregate all models within a user / access point cluster.

[0142] 4. Inter-cluster aggregation: Aggregate models from different clusters.

[0143] 5. Global model: A complete neural network model (i.e., the first model or the second model mentioned in the above embodiments).

[0144] 6. Segmentation Point: The model segmentation strategy (i.e., the model segmentation strategy mentioned in the above embodiments). For example, a neural network with a total of 10 layers can be divided into sub-modules of the first 2 layers and sub-modules of the last 8 layers. The second layer is called the segmentation layer, i.e., the segmentation point. Therefore, the global model is divided into two different sub-modules through the segmentation point.

[0145] 7. User sub-model: A sub-module deployed on the user's device.

[0146] 8. Access Point Sub-model: The remaining sub-modules in the global model excluding the user sub-model.

[0147] The cellular-free MIMO-based federated segmentation learning architecture includes a central processing unit (CPU), an access point (AP), and user equipment (UE), and the following three planes provide functionality between the CPU, AP, and UE: 1. Control Plane: Responsible for establishing, maintaining, and optimizing federated segmentation learning, ensuring users can access the network and receiving the corresponding quality of service. The control plane mainly includes the following functions: (1) User access management: Associate access points and user equipment.

[0148] (2) Segmentation point management: Segmenting the model on the user device.

[0149] (3) Inter-cluster aggregation of access points: perform inter-cluster aggregation on the intra-cluster aggregation model of different access points.

[0150] 2. Computing Plane: Responsible for communication between access points and user devices in federated segmentation learning, updating the edge server computing model, and ensuring that the non-cellular network can support the federated learning and segmentation learning processes. The computing plane mainly includes the following functions: (1) Uplink and downlink communication: Receive the activation data transmitted by the user equipment uplink and transmit the user sub-model after inter-cluster aggregation downlink.

[0151] (2) Access point sub-model forward propagation: User activation data is propagated forward through the neural network at the corresponding segmentation point, and the corresponding loss function is calculated.

[0152] (3) Gradient backpropagation of access point model: Update the global models of different users deployed at the access point in reverse according to the loss function.

[0153] (4) Aggregation within the access point cluster: Aggregate different global models after the activation data is updated.

[0154] 3. User Plane: Responsible for the forward propagation of the user sub-model, generating and sending activation data to the associated access points. The user plane mainly includes the following functions: (1) User sub-model forward propagation: User data is forward propagated through the user sub-model to generate activation data.

[0155] (2) Uplink transmission of activation data: Uplink transmission of activation data to the associated access point.

[0156] (3) Downlink receiving updated user sub-model: Receive the user sub-model data after inter-cluster aggregation transmitted downlink from the associated access point.

[0157] In summary, the federated segmentation learning architecture based on cellular-free MIMO, such as Figure 5 As shown. In Figure 5 In this model, the user plane stores user datasets and sub-model data, while the computation plane and control plane store global model data associated with users. In the computation plane, forward propagation occurs on the access point sub-models, while gradient backpropagation occurs on the global model. Figure 5 In this architecture, a user equipment (UE) can be associated with multiple access points simultaneously, and these access points can be associated with different central processing units (CPUs). Furthermore, UEs are deployed on the user plane, access points on the compute plane, and CPUs on the control plane.

[0158] In one possible implementation, the access point is a Transmission Reception Point (TRP) or a server deployed on the TRP side, and the central processing unit is a Base Station (BS) or a server deployed on the Base Station side.

[0159] In another possible implementation, the access point is a base station or a server deployed on the base station side, and the central processing unit is an edge server or a network function device located in the core network.

[0160] Example 1: Combined with the specific training architecture of federated segmentation learning based on cellular-free MIMO, refer to Figure 6 The communication method proposed in this application includes the following steps: S600: The upper-layer server sends the model training description information, and the central processing unit receives the model training description information accordingly.

[0161] In some embodiments, the central processing unit receives description information of the federated segmentation learning training service from the upper-layer server, i.e., description information of model training.

[0162] The descriptive information for model training includes the AI ​​tasks, AI routines, and business functions of the training service.

[0163] Among them, AI task refers to the multi-round model update process of neural network model; AI routine refers to each step of executing AI task and its dependencies; business function refers to the logical processing capability required to complete an AI routine, that is, different participating devices and their processing flow and capabilities.

[0164] S601. The user equipment sends its access information, and the access point receives the user equipment's access information accordingly.

[0165] The user equipment access information includes active user equipment information, aggregated user equipment information, and access point information.

[0166] S602. The access point sends the user equipment's access information and network environment information, and the central processing unit receives the user equipment's access information and network environment information accordingly.

[0167] The network environment information includes the dynamic wireless channel environment between user equipment and access points, used to determine appropriate segmentation strategies and reasonable association relationships between user base stations. It also includes the idle communication and computing capabilities of user equipment and access points, such as idle transmission frequencies and power, as well as idle storage space and CPU processing frequency.

[0168] S603, the central processing unit determines the deployment mode of the corresponding federated segmentation learning and training service based on the service description information, user equipment access information and network environment information.

[0169] This deployment mode includes at least the following: 1. The access point IDs participating in the AI ​​routines, and the communication and computing capabilities required by each user device.

[0170] 2. Information on the communication and computing capabilities required for this AI routine.

[0171] 3. The user device IDs participating in the AI ​​routines, and the communication and computing capabilities required by each user device.

[0172] 4. Model segmentation strategy for user devices. Different user devices can be configured with the same model segmentation strategy or different model segmentation strategies.

[0173] S604. The central processing unit sends configuration information for the user plane and the computing plane, and the access point receives the configuration information for the user plane and the computing plane accordingly.

[0174] The configuration information for the computation plane includes: 1. The data aggregation order among the various access points participating in the AI ​​routine is as follows: first, aggregation within the access point cluster is performed, and then aggregation between access point clusters is performed according to the user equipment access logic.

[0175] 2. The association logic between access points and user equipment, i.e., which access points the user equipment connects to through the wireless air interface.

[0176] 3. Segmentation strategy between model sub-modules.

[0177] 4. Reserve corresponding communication, computing and storage resources.

[0178] The configuration information for the user plane includes: 1. Synchronization parameters between various user devices participating in the AI ​​routine, used to ensure that user devices start the AI ​​routine synchronously.

[0179] 2. Reserve corresponding communication, computing and storage resources.

[0180] S605. The access point sends user plane configuration information, and the user equipment receives the user plane configuration information accordingly.

[0181] S606: The central processing unit sends the global model, and the access point receives the global model accordingly.

[0182] The global model can be the initial model (i.e., the first model mentioned in the above embodiments), or the global model can be the model updated from the previous AI routine (i.e., the second model mentioned in the above embodiments).

[0183] It should be noted that, since each user device corresponds to a model, the global model of the compute plane storage is determined by the associated user device.

[0184] It should be noted that the central processing unit sends different user models to different access points according to the association logic (this model is the global user model after aggregation within and between access point clusters, which serves as the "initial" model for subsequent routines).

[0185] S607. The access point segments the global model according to the model segmentation strategy.

[0186] S608. The access point sends the user sub-model, and the user equipment receives the user sub-model accordingly.

[0187] It should be noted that the access point completes the segmentation of the user model sub-modules according to the model segmentation strategy and sends the user sub-models to the user equipment.

[0188] Example 2: After completing the above steps, the deployment process for the Federated Segmentation Learning and Training Service is finished. (Reference) Figure 7 The communication method proposed in this application may further include the following steps: S701. The user equipment performs forward propagation on the user sub-model to obtain activation data.

[0189] In some embodiments, each user device uses a local dataset to perform forward propagation on the user sub-model it receives, generating activation data (data calculated layer by layer from the input layer and output at the segmentation layer).

[0190] It should be noted that the user equipment uses the user model submodule deployed locally to complete the forward propagation of data; the submodule outputs intermediate data, namely activation data.

[0191] It should be noted that in the first round of training, user devices share some of the same model parameters. In subsequent training rounds, the model parameters are affected by the association logic in the previous round of training. For example, in the first round of training, different inter-cluster aggregation processes will be performed according to the association logic of each user device to generate different global user models, which will be used as the initial user models for the second round of training.

[0192] S702, the user equipment sends activation data, and the access point receives the activation data accordingly.

[0193] In some embodiments, the user equipment sends activation data to the access point, which then divides the access point sub-model based on the user segmentation point and the user global model updated in the last iteration, and performs forward propagation of the access point model.

[0194] It should be noted that after all user equipment completes local forward propagation, the access point controls the uplink transmission activation data to the corresponding associated access point. Several methods exist: 1. After completing local forward propagation, the user equipment sends an indication message to the access point to indicate that local forward propagation is complete. Once all user equipment participating in the AI ​​routine has completed local forward propagation, the access point instructs some or all user equipment to send activation data to the access point.

[0195] 2. In step 5 of Embodiment 1, the access point sends time-waiting information to the user equipment, instructing the user equipment to wait for the specified time after completing local forward propagation before sending activation data to the access point. That is, in step 5 of Embodiment 1, the synchronization parameters include time-waiting information.

[0196] 3. In step 5 of Embodiment 1, the access point sends time completion information to the user equipment, indicating that after completing local forward propagation, the user equipment should send activation data to the access point within the completion time. That is, in step 5 of Embodiment 1, the synchronization parameters include time completion information.

[0197] 4. After completing local forward propagation, the user equipment immediately transmits activation data uplink to the corresponding associated access point.

[0198] In some embodiments, the access point decodes the activation data uploaded by the user equipment according to the association logic, and completes further forward propagation of the user model according to the user model segmentation logic and the activation data, that is, forward propagation of the access point sub-model.

[0199] S703: The access point performs gradient backpropagation on the global model of each user and updates the global model of the associated users.

[0200] In some embodiments, each user model deployed on the access point undergoes gradient backpropagation; the access point stores the updated model data of the associated user.

[0201] S704. Aggregate all associated user models within the access point cluster.

[0202] In some embodiments, the access point averages the global model data of associated users stored within the access point.

[0203] In some embodiments, after the global models of all associated users have completed gradient backpropagation, the access point performs intra-cluster aggregation. In this approach, the access point only needs to perform intra-cluster aggregation once to obtain the access point cluster model (i.e., the first global model mentioned in the above embodiments).

[0204] Alternatively, once the global model of an associated user has completed gradient backpropagation, the access point immediately performs intra-cluster aggregation. In this approach, the access point needs to perform multiple intra-cluster aggregations to obtain the access point cluster model (i.e., the first global model mentioned in the above embodiments).

[0205] It should be noted that each access point is aggregated within the access point cluster to generate an access point cluster model.

[0206] S705, the access point sends the access point cluster model, and correspondingly, the central processing unit receives the access point cluster model.

[0207] In some embodiments, each access point sends an access point cluster model to the central processing unit, and the central processing unit performs inter-access point cluster aggregation based on the user's access point clusters to update the global model (i.e., the second model mentioned in the above embodiments).

[0208] In some embodiments, after the central processing unit has completed aggregation within all access point clusters, it begins aggregation between access point clusters to generate a global model.

[0209] Alternatively, the central processing unit can select the corresponding partial access point cluster model based on the association logic of each user device, perform inter-cluster aggregation, and generate a global model.

[0210] S706: The central processing unit sends the updated global model, and the access point receives the updated global model accordingly.

[0211] In some embodiments, the central processing unit sends the updated global model to the access point, which then performs the segmentation of the model sub-models and sends the user sub-models to the user equipment.

[0212] This application enables non-cellular networks to provide corresponding computing, storage, and communication resources for AI services through a joint design of communication and computing.

[0213] In the communication method of this application: 1. The control plane is responsible for user access management, segmentation point management, and aggregation between access point clusters; the computing plane is responsible for uplink and downlink communication, access point sub-model front-through, access point model gradient back-through, and aggregation within access point clusters; the user plane is responsible for user sub-model front-through, uplink sending of activation data, and downlink receiving of updated user sub-models.

[0214] 2. User equipment only receives and stores local user sub-models and uses local datasets to generate activation data; access points store global models of associated users, update the global models of corresponding users using the received user activation data, and then perform intra-cluster aggregation; the central processing unit stores global models of all users, performs inter-cluster aggregation based on the user's access point cluster information, and updates the global models of corresponding users.

[0215] 3. The sub-model training process includes the following steps: forward propagation of the model in the user equipment and access point; backward propagation of the model in the access point; aggregation of the model within the access point cluster and aggregation between access point clusters; updating the user global model and sub-model on the access point and user equipment.

[0216] 4. By transmitting activation data uplink from user equipment and user sub-model data downlink from access point, the communication overhead between uplink and downlink is reduced. By having the access point handle part of the model forward propagation and all of the gradient backpropagation, the computational load on the user side is reduced, and the utilization rate of computational resources on the access point side is improved.

[0217] According to some embodiments of this application, when device 800 is a first access network device, processor 801 in 800 may be referred to as a first processor, and transceiver 802 may be referred to as a first transceiver.

[0218] The first access point device includes: a first transceiver; and a first processor coupled to the first transceiver, wherein the first processor is configured to: The system receives multiple first messages sent by multiple first terminal devices via a first transceiver. The second piece of information is obtained by integrating multiple pieces of primary information. The second information is sent to the network device via the first transceiver; the second information is used by the network device to determine the first model, or to determine the model segmentation strategy for the first model. The first model is received from the network device via the first transceiver. Based on the model segmentation strategy, the first model is segmented to obtain multiple sub-models; Multiple sub-models are sent to multiple second terminal devices via the first transceiver; Among them, multiple first terminal devices include multiple second terminal devices.

[0219] In some embodiments, the multiple first pieces of information are multiple device information of multiple first terminal devices; the first processor is configured to: integrate the multiple device information to obtain integrated multiple device information; wherein, the second information includes: the integrated multiple device information; and the device information of the first access point device.

[0220] In some embodiments, the device information includes one or more of the following: device identification information; channel environment information; device communication capability information; and device computing capability information.

[0221] In some embodiments, the first processor is configured to: receive, via a first transceiver, configuration information of a computing plane sent by a network device according to second information; the configuration information of the computing plane includes one or more of the following: a model segmentation strategy; a model aggregation order, the model aggregation order being used to indicate the order of aggregation within an access point cluster and aggregation between access point clusters; first indication information, the first indication information being used to indicate a plurality of second terminal devices from a plurality of first terminal devices; and second indication information, the second indication information being used to indicate the idle resources required by the first access point devices to perform model training, the idle resources including one or more of communication idle resources, computing idle resources, and storage idle resources.

[0222] In some embodiments, the first processor is configured to: receive user plane configuration information transmitted by the network device via a first transceiver; the user plane configuration information includes one or more of the following: multiple synchronization parameters associated with multiple second terminal devices, the multiple synchronization parameters being used for the multiple second terminal devices to simultaneously perform training for a sub-model; multiple third indication information associated with multiple second terminal devices, the third indication information associated with each second terminal device being used to indicate the idle resources required for the second terminal device to perform model training, the idle resources including one or more of communication idle resources, computing idle resources, and storage idle resources.

[0223] In some embodiments, the plurality of first terminal devices are a plurality of second terminal devices, and the plurality of first information are a plurality of activation data obtained by the plurality of second terminal devices using their own user data to perform forward propagation on the received sub-models; the first processor is configured to: use the plurality of activation data to perform backpropagation on a plurality of remaining models to obtain a plurality of initial models; the plurality of remaining models are a subset of the first models other than the plurality of sub-models; perform intra-cluster aggregation on the plurality of initial models to obtain a first global model; wherein, the second information includes model information of the first global model.

[0224] In some embodiments, the first processor is configured to: perform intra-cluster aggregation once for each initial model obtained, and obtain a first global model by performing multiple intra-cluster aggregations; or, perform intra-cluster aggregation once for multiple initial models obtained, and obtain a first global model.

[0225] In some embodiments, the first processor is configured to: transmit a plurality of fourth indication messages to a plurality of second terminal devices via a first transceiver under a first condition; the plurality of fourth indication messages are used to instruct the plurality of second terminal devices to send a plurality of activation data; wherein the first condition is that the first access point device receives a plurality of fifth indication messages sent by the plurality of second terminal devices to indicate that model training is complete.

[0226] In some embodiments, the first processor is configured to: send a plurality of sixth indication messages to a plurality of second terminal devices via a first transceiver; each sixth indication message includes a first time window, and each sixth indication message is used to instruct the corresponding second terminal device to send corresponding activation data while waiting for the first time window after the model training is completed.

[0227] In some embodiments, the first processor is configured to: send a plurality of seventh indication messages to a plurality of second terminal devices via a first transceiver; each seventh indication message includes a second time window, and each seventh indication message is used to instruct the corresponding second terminal device to send corresponding activation data within a second time window after the model training is completed.

[0228] In some embodiments, the first processor is configured to: send a first global model to a network device via a first transceiver; wherein the first global model and other global models sent by other access point devices associated with the network device are used by the network device to perform inter-cluster aggregation to obtain a second model, and the second model is used to replace the first model to perform the next round of model training.

[0229] According to some embodiments of this application, when device 800 is a network device, processor 1401 in 800 may be referred to as a second processor, and transceiver 802 may be referred to as a second transceiver.

[0230] The network device includes a second transceiver; and a second processor coupled to the second transceiver; the second processor is configured to: The second transceiver receives the second information sent by the first access point device. Perform inter-cluster aggregation based on the second information to obtain the first model, or determine the model segmentation strategy for the first model based on the second information; The first model is sent to the first access point device via the second transceiver, or the first model and model segmentation strategy are sent.

[0231] The descriptions of the above device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0232] It should be noted that, in the embodiments of this application, if the above-mentioned wireless communication method is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, 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 to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0233] Fifthly, to implement the above-mentioned wireless communication method, embodiments of this application provide 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 wireless communication method provided in the above embodiments.

[0234] Sixthly, embodiments of this application provide a storage medium, namely a computer-readable storage medium, on which a computer program is stored, which, when executed by a processor, implements the steps in the wireless communication method provided in the above embodiments.

[0235] It should be noted that the descriptions of the storage medium and device embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0236] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in some embodiments" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, 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 this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0237] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0238] In the 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 in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0239] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0240] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0241] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented 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 of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0242] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0243] The above are merely embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A first access point device, applied in a non-cellular MIMO scenario, the first access point device comprising: First transceiver; and A first processor, coupled to the first transceiver, wherein the first processor is configured to: The system receives multiple first messages sent by multiple first terminal devices via the first transceiver. The multiple pieces of first information are integrated to obtain second information; The second information is sent to the network device via the first transceiver; the second information is used by the network device to determine the first model, or to determine a model segmentation strategy for the first model. The first model is received from the network device via the first transceiver; According to the model segmentation strategy, the first model is segmented to obtain multiple sub-models; The multiple sub-models are sent to multiple second terminal devices via the first transceiver; The plurality of first terminal devices include the plurality of second terminal devices.

2. The device according to claim 1, wherein, The plurality of first information refers to the plurality of device information of the plurality of first terminal devices; The first processor is configured to: The information from the multiple devices is integrated to obtain integrated information from the multiple devices. The second information includes: The integrated information of multiple devices; Device information of the first access point device.

3. The device according to claim 2, wherein, The device information includes one or more of the following: Equipment identification information; Channel environment information; Device communication capability information; Equipment computing capacity information.

4. The device according to claim 2 or 3, wherein, The first processor is configured to: The network device receives computing plane configuration information sent by the network device based on the second information via the first transceiver; the computing plane configuration information includes one or more of the following: The model segmentation strategy; Model aggregation order, which indicates the order of aggregation within an access point cluster and aggregation between access point clusters; First indication information, the first indication information being used to indicate the plurality of second terminal devices from the plurality of first terminal devices; The second indication information is used to instruct the first access point device to perform model training on the available resources, the available resources including one or more of communication available resources, computing available resources and storage available resources.

5. The device according to claim 1, wherein, The plurality of first terminal devices are the plurality of second terminal devices, and the plurality of first information are the plurality of activation data obtained by the plurality of second terminal devices using their own user data to perform forward propagation on the received sub-model; The first processor is configured to: The multiple activation data are used to backpropagate multiple remaining models to obtain multiple initial models; the multiple remaining models are the partial models other than the multiple sub-models in the first model. Perform intra-cluster aggregation on the multiple initial models to obtain the first global model; The second information includes the model information of the first global model.

6. The device according to claim 5, wherein, The first processor is configured to: Given an initial model, perform one intra-cluster aggregation. After performing multiple intra-cluster aggregations, the first global model is obtained. or, Given the multiple initial models, perform an intra-cluster aggregation to obtain the first global model.

7. The device according to claim 5, wherein, The first processor is configured to: Under a first condition, a plurality of fourth indication messages are sent to the plurality of second terminal devices via the first transceiver; the plurality of fourth indication messages are used to instruct the plurality of second terminal devices to send the plurality of activation data; The first condition is that the first access point device receives multiple fifth indication messages sent by the multiple second terminal devices to indicate that the model training is complete.

8. The device according to claim 5, wherein, The first processor is configured to: The first global model is sent to the network device via the first transceiver; Wherein, the first global model and other global models sent by other access point devices associated with the network device are used by the network device to perform inter-cluster aggregation to obtain a second model, and the second model is used to replace the first model to perform the next round of model training.

9. A network device for use in a non-cellular MIMO scenario, the network device comprising: Second transceiver; and A second processor, coupled to the second transceiver, wherein the second processor is configured to: The second transceiver receives the second information sent by the first access point device. Perform inter-cluster aggregation based on the second information to obtain the first model, or determine a model segmentation strategy for the first model based on the second information; The first model is sent to the first access point device via the second transceiver, or the first model and the model segmentation strategy are sent.

10. A communication method applied to a first access point device in a non-cellular MIMO scenario, the method comprising: Receive multiple first messages sent by multiple first terminal devices; The multiple pieces of first information are integrated to obtain second information; Send the second information to the network device; The second information is used by the network device to determine the first model, or to determine a model segmentation strategy for the first model; Receive a first model sent by the network device; segment the first model according to the model segmentation strategy to obtain multiple sub-models; Send the multiple sub-models to multiple second terminal devices; The plurality of first terminal devices includes the plurality of second terminal devices.

Images