WO2024027676A1 - 用于分层联邦学习网络中的切换的装置、方法和介质 - Google Patents

用于分层联邦学习网络中的切换的装置、方法和介质 Download PDF

Info

Publication number
WO2024027676A1
WO2024027676A1 PCT/CN2023/110473 CN2023110473W WO2024027676A1 WO 2024027676 A1 WO2024027676 A1 WO 2024027676A1 CN 2023110473 W CN2023110473 W CN 2023110473W WO 2024027676 A1 WO2024027676 A1 WO 2024027676A1
Authority
WO
WIPO (PCT)
Prior art keywords
user equipment
model
aggregation
intermediate node
global
Prior art date
Application number
PCT/CN2023/110473
Other languages
English (en)
French (fr)
Inventor
郑策
孙晨
Original Assignee
索尼集团公司
郑策
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 索尼集团公司, 郑策 filed Critical 索尼集团公司
Publication of WO2024027676A1 publication Critical patent/WO2024027676A1/zh

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W28/00—Network traffic management; Network resource management
    • H04W28/02—Traffic management, e.g. flow control or congestion control
    • H04W28/10—Flow control between communication endpoints
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W36/00—Hand-off or reselection arrangements
    • H04W36/24—Reselection being triggered by specific parameters
    • H04W36/32—Reselection being triggered by specific parameters by location or mobility data, e.g. speed data
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
    • H04W4/30—Services specially adapted for particular environments, situations or purposes
    • H04W4/40—Services specially adapted for particular environments, situations or purposes for vehicles, e.g. vehicle-to-pedestrians [V2P]
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W88/00—Devices specially adapted for wireless communication networks, e.g. terminals, base stations or access point devices
    • H04W88/02—Terminal devices
    • H04W88/04—Terminal devices adapted for relaying to or from another terminal or user

Definitions

  • the present disclosure relates to apparatus, methods and media for switching in hierarchical federated learning networks.
  • each user equipment accesses the base station through a wireless channel, uploads the locally learned model to the server through the base station, and the server aggregates the model and then distributes the aggregated model to each user equipment through the base station.
  • the user equipment may need to switch when performing a federated learning task.
  • the present disclosure provides apparatus, methods and media for switching in a hierarchical federated learning network.
  • a network-side electronic device for federated learning including a processing circuit configured to: determine a model aggregation time and a remaining service time for a user equipment, wherein the user A device is directly connected to the network or indirectly connected to the network via an intermediate node; when the model aggregation time and the remaining service time meet predetermined conditions, make a switching decision for the user equipment; and send The handover decision for the user equipment.
  • an intermediate node-side electronic device for federated learning including a processing circuit configured to: receive a handover decision for a user equipment from a network, wherein the handover decision It is made when the model aggregation time and the remaining service time for the user equipment meet predetermined conditions, and the user equipment is indirectly connected to the network via the intermediate node; the handover decision is sent to the User equipment.
  • a user equipment side electronic device for federated learning including a processing circuit configured to: receive a switching decision for the user equipment from the network, wherein the switching The decision is made when the model aggregation time and the remaining service time for the user equipment satisfy predetermined conditions. and the user equipment is directly connected to the network or indirectly connected to the network via an intermediate node; handover is performed based on the handover decision.
  • a network-side method for federated learning including: determining a model aggregation time and a remaining service time for a user equipment, wherein the user equipment is directly connected to the network or via an intermediate Nodes are indirectly connected to the network; when the model aggregation time and the remaining service time meet predetermined conditions, make a switching decision for the user equipment; and send the switching decision for the user equipment .
  • an intermediate node-side method for federated learning comprising: receiving a handover decision for a user equipment from a network, wherein the handover decision is made at model aggregation time and for the user equipment The switching decision is made when the remaining service time meets predetermined conditions, and the user equipment is indirectly connected to the network via the intermediate node; the switching decision is sent to the user equipment.
  • a user device side method for federated learning including: receiving a switching decision for a user device from a network, wherein the switching decision is made at model aggregation time and for the user It is made when the remaining service time of the device meets predetermined conditions, and the user equipment is directly connected to the network or indirectly connected to the network via an intermediate node; the handover decision is performed based on the handover decision.
  • a non-transitory computer-readable storage medium having stored thereon program instructions that, when executed by a processor, cause the processor to perform the method of the present disclosure.
  • a computer program product comprising program instructions that, when executed by a processor, cause the processor to perform a method of the present disclosure.
  • Figure 1 shows an exemplary structure of a traditional federated learning network.
  • Figure 2 illustrates an exemplary structure of a hierarchical federated learning network according to an embodiment of the present disclosure.
  • Figure 3 illustrates an exemplary federated learning process of a hierarchical federated learning network according to an embodiment of the present disclosure.
  • Figures 4A to 4C illustrate different handover scenarios for local UEs.
  • Figure 5 shows an exemplary handover process of a local UE according to an embodiment of the present disclosure.
  • Figures 6A to 6E show situations where the remaining service time of a local UE satisfies different conditions.
  • Figure 7 shows an exemplary handover process of a global UE according to an embodiment of the present disclosure.
  • Figures 8A to 8C show situations where the remaining service time of the global UE satisfies different conditions.
  • FIG 9 shows the 5G core network SBA architecture and some of its network functions (NF).
  • FIG. 10 is a block diagram illustrating an example of a schematic configuration of a computing device to which techniques of the present disclosure may be applied.
  • FIG. 11 is a block diagram illustrating a first example of a schematic configuration of a gNB to which the technology of the present disclosure can be applied.
  • FIG. 12 is a block diagram illustrating a second example of a schematic configuration of a gNB to which the technology of the present disclosure can be applied.
  • FIG. 13 is a block diagram showing an example of a schematic configuration of a smartphone to which the technology of the present disclosure can be applied.
  • FIG. 14 is a block diagram showing an example of a schematic configuration of a car navigation device to which the technology of the present disclosure can be applied.
  • Figure 1 shows an exemplary structure of a traditional federated learning network.
  • UEs 120 1 , 120 2 , 120 3 ...120 K-1 , 120 K are directly connected to the server 110 through a base station (not shown), and upload local models to the server 110 for global aggregation.
  • the process is as follows.
  • UE 120 accesses server 110 and obtains the initial global model w 0 through downlink transmission:
  • UE 120 uses the data stored locally to learn and completes the local iteration of the r+1 local model update:
  • the UE 120 transmits the learned local model through the uplink or gradient Upload to server 110.
  • the server 110 aggregates the local models collected from each UE to complete the update of the global model:
  • D i the data set of the i-th UE
  • D i the size of the data set
  • the server 110 distributes the updated global model w r+1 to the UE 120 again, and then repeats the above steps until the model converges.
  • the base station has limited coverage and cannot provide services to UEs outside the coverage area. Secondly, there are some areas with low communication rates within the coverage area of the base station. For UEs in these areas, the service quality cannot be guaranteed even if they are within the coverage area. In addition, the base station has limited uplink and downlink communication resources (such as frequency resources, number of carriers, etc.) and cannot support simultaneous access of too many UEs.
  • Figure 2 illustrates an exemplary structure of a hierarchical federated learning network according to an embodiment of the present disclosure.
  • the network structure consists of two layers: the first layer consists of the intermediate node 220 and the UE 230 it serves (the UE connected to the intermediate node is called a local UE in this article); the second layer consists of the global node 210 and the intermediate node 220 connected to it. It consists of UE 240 directly connected to the global node (this article refers to the UE directly connected to the global node as global UE).
  • the global node is implemented by a base station and/or a server connected to the base station.
  • the intermediate node can be a mobile vehicle mounted relay (Vehicle Mounted Relay, VMR), an unmanned aerial vehicle (Unmanned Aerial Vehicle, UAV), etc.
  • VMR Vehicle Mounted Relay
  • UAV Unmanned Aerial Vehicle
  • the intermediate nodes can be implemented as fixed-position roadside units (Road Side Unit, RSU), edge (Edge) nodes, etc.
  • FIG. 3 An exemplary federated learning process of a hierarchical federated learning network according to an embodiment of the present disclosure is described below with reference to FIG. 3 .
  • step S302 the local UE 230 (as a first-level node) performs, for example, k 1 local iterations, and then uploads the obtained local model to the intermediate node 220 (as a second-level node) for intermediate aggregation.
  • step S304 the intermediate node 220 performs an intermediate aggregation on the local models received from the local UE 230 to obtain an intermediate model.
  • the intermediate node 220 determines whether a total of, for example, k 2 intermediate aggregations has been completed. If the total number of intermediate aggregations is not k 2 times, the intermediate node 220 distributes the intermediate model to the local UE 230 . After receiving the intermediate model, the local UE 230 updates the local model to the intermediate model and performs a new round of local iteration. If the total number of intermediate aggregations has reached k 2 times, the intermediate node 220 uploads the intermediate model to the global node 210 for global aggregation.
  • step S308 if there is a global UE 240 directly served by the global node 210, the global UE 240 performs, for example, k 1 k 2 local iterations, and then uploads the obtained local model to the global node 210.
  • step S310 the global node 210 (or a network device such as a server connected to it) globally aggregates the intermediate model uploaded by the intermediate node 220 and the local model uploaded by the global UE 240 (if any), and distributes the obtained global model.
  • the intermediate node 220 updates the intermediate model to the global model, and distributes the global model to the local UEs 230 it serves.
  • the UE 230 and the global UE 240 update the local model to the global model.
  • local UEs are connected to intermediate nodes, which shortens the communication distance and ensures the quality of communication service.
  • multiple local UEs communicate directly with the intermediate node, and the intermediate node is connected to the global node.
  • the intermediate node serves multiple local UEs, each communication between it and the global node only uploads the aggregated intermediate model.
  • the amount of data is equivalent to the amount of data used by a global UE to communicate with a global node or a local UE to communicate with an intermediate node. This greatly reduces the load on global nodes and alleviates the problem of insufficient communication resources at global nodes.
  • this network structure makes full use of the functions of intermediate nodes, that is, the intermediate nodes not only serve as relays for transmission, but also participate in calculations as aggregators of the first-layer structure to complete the intermediate aggregation of the model.
  • the global node first initializes the global model w 0 and distributes the initialized global model to the global UE and intermediate nodes.
  • the intermediate node then distributes the initialized global model to local UEs.
  • the global node, global UE, intermediate node and local UE have the same federated learning model:
  • n i the number of local UEs served by intermediate node #i.
  • Each local UE performs local iteration based on locally stored data:
  • the local UE obtains local model parameters after k 1 local iterations and uploaded to the intermediate node. After receiving the local models uploaded by the local UEs of all its services, the intermediate node performs an intermediate aggregation:
  • r 1 k 1 r 2 , that is, every time an intermediate node completes an intermediate aggregation, the local UE it serves completes k 1 local model updates.
  • the intermediate node distributes the model obtained by intermediate aggregation (herein referred to as the intermediate model) to the local UEs it serves.
  • the intermediate node distributes the intermediate aggregated model to each local UE it serves.
  • the local UE updates the local model using the received intermediate model.
  • the intermediate node uploads the obtained intermediate model to the global node after completing the k 2 intermediate aggregation.
  • the global UE independently performs local iterations based on local data, which is similar to the process of equation (2):
  • the global UE obtains local model parameters after performing k 1 k 2 local iterations. and uploaded to the global node.
  • the global node After receiving the intermediate models uploaded by all intermediate nodes and the local models uploaded by all global UEs, the global node performs global aggregation:
  • r 2 k 2 r 3 , that is, every time a global node completes a global aggregation, its connected intermediate nodes complete k 2 intermediate aggregations.
  • the global node performs global model aggregation to obtain global model parameters. And distributed to each intermediate node and global UE.
  • the global node initializes the global model w 0 and distributes it to the global UE and intermediate nodes.
  • the intermediate node then distributes the initialized global model to local UEs.
  • the global node, global UE, intermediate node and global UE have the same learning model:
  • n i the number of local UEs served by intermediate node #i.
  • Each local UE performs local iteration based on locally stored data:
  • the local UE obtains local model parameters after k 1 local iterations And calculate the sum of the gradients of k 1 local iterations
  • the intermediate node After receiving the local gradients uploaded by the local UEs of all its services, the intermediate node performs an intermediate aggregation:
  • r 1 k 1 r 2 , that is, every time an intermediate node completes an intermediate aggregation, the local UE it serves completes k 1 local model updates.
  • the intermediate node distributes the intermediate aggregated model (herein referred to as the intermediate model) to each local UE it serves.
  • the local UE updates the local model using the received intermediate model.
  • the intermediate node uploads the obtained intermediate model to the global node after completing the k 2 intermediate aggregation.
  • the intermediate node obtains the intermediate model parameters after k 2 intermediate aggregations. And calculate the sum of the gradients of k 2 intermediate aggregations
  • the intermediate node will gradient Upload to the global node for global aggregation.
  • the global UE obtains local model parameters after performing k 1 k 2 local iterations. and upload the gradient To the global node:
  • the global node After receiving the intermediate models uploaded by all intermediate nodes and the local models uploaded by all global UEs, the global node performs global aggregation:
  • r 3 k 1 k 2 r 1 , that is, every time a global node completes global aggregation, its connected intermediate nodes complete k 2 intermediate aggregations.
  • the global node performs global model aggregation and updates, and distributes the aggregated model to each intermediate node and global UE.
  • Figure 4A shows that one or more local UEs are handed over from intermediate node #i to intermediate node #j, and before handover The scenario where intermediate node #j has local UE served. In this scenario, there is an intermediate model at the intermediate node #j, so there is no need to transmit the intermediate model from the intermediate node #i to the intermediate node #j.
  • UE#l directly performs handover when the handover conditions are met (for example, the received signal strength RSRP is less than a certain threshold). If the handover of UE#l occurs during the r 2 +1 intermediate aggregation of intermediate node #i (the r 2 intermediate aggregation has been completed, but the r 2 +1 intermediate aggregation has not been completed, that is, k 1 r 2 ⁇ r 1 ⁇ k 1 (r 2 +1)), it will lead to the following results.
  • intermediate node #i it has been disconnected from UE#l and cannot receive the local model of UE#l during the r 2 +1th intermediate aggregation.
  • intermediate node #j it can receive the local model uploaded by UE#l.
  • FIG. 4B shows a scenario in which all local UEs served by intermediate node #i are switched to intermediate node #j, and there is no local UE served by intermediate node #j before switching. This scenario may be caused by the movement of the intermediate node #i or the movement of the UE.
  • the UE switches from intermediate node #i to intermediate node #j.
  • the intermediate node #j performs the r 2 +1th intermediate aggregation, which requires the intermediate model after the r 2nd intermediate aggregation.
  • the gradient is uploaded, as shown in equation (10)
  • calculating the intermediate model after the r 2 +1th intermediate aggregation requires the r 2 intermediate model after the intermediate aggregation, and there is no r 2 at the intermediate node #j at this time.
  • the intermediate model after sub-intermediate aggregation. Even if you upload model parameters, you still need to use the last intermediate model parameters to determine the weights used in calculating the intermediate model.
  • intermediate node #i there are no intermediate models at intermediate node #j, only the global model initially received from the global node. Therefore, intermediate node #i needs to transmit its intermediate model to intermediate node #j.
  • FIG. 4C shows a scenario in which some local UEs served by intermediate node #i are switched to intermediate node #j, and there are no local UEs served by intermediate node #j before the handover.
  • both the intermediate node #i and the intermediate node #j lack a part of UEs to participate in the intermediate aggregation, which will lead to model divergence and reduced accuracy. That is, switching is performed during the intermediate aggregation process. Even if only some UEs are switched, it will cause the divergence of the intermediate model and reduce the accuracy of the global model.
  • some embodiments of the present disclosure enable the UE to perform handover after global aggregation and global model broadcast.
  • the models of each UE and the intermediate node are the same and are global models. There is no need for additional model transmission and the switching cost is minimal. If switching cannot be guaranteed after global aggregation and global model broadcast, some embodiments of the present disclosure enable switching after intermediate nodes complete intermediate aggregation. At this time, the intermediate node has the same model as all the local UEs it serves, and can ensure that the training service of the local UEs is not interrupted.
  • some embodiments of the present disclosure also extend the connection time of the UE and/or intermediate nodes to intermediate aggregation or global by increasing the transmission power, lowering the RSRP threshold, allocating more transmission resources (including time resources, frequency resources), etc. Aggregation is complete. This ensures service continuity and improves system performance.
  • T serve The estimated value of the remaining time for the global node or intermediate node to provide service to the UE;
  • T 1 The estimated value of the remaining time required for the global node to complete this round of global aggregation, that is, the time from the current moment to the global node completing this round of global aggregation;
  • T 2 The time required for the global node to complete the next round of global aggregation, that is, the time from the global node completing this round of global aggregation to the global node completing the next round of global aggregation;
  • T train T 1 + T 2 ;
  • FIG. 5 shows an exemplary handover process of a local UE according to an embodiment of the present disclosure.
  • UE#1 sends its own status information Info U to the intermediate node #i.
  • the UE's status information Info U may include channel status (such as RSRP), computing capabilities (such as CPU occupancy), local data information (such as the number of samples participating in model training, sample dimensions, etc.), power, location and mobility information (such as, speed, direction, time spent at a certain place, etc.) and one or more of other information.
  • channel status such as RSRP
  • computing capabilities such as CPU occupancy
  • local data information such as the number of samples participating in model training, sample dimensions, etc.
  • power location and mobility information (such as, speed, direction, time spent at a certain place, etc.) and one or more of other information.
  • the intermediate node #i transmits its own status information Info V and the status information of the local UEs it serves. Info U is sent to the global node.
  • the status information Info V of the intermediate node may include one of channel status (RSRP), computing power (such as CPU occupancy), location and movement information (such as speed, direction, residence time in a certain place, etc.) and other information or multiple items.
  • the channel status of the intermediate node may include the channel status between the intermediate node and the local UE and the channel status between the intermediate node and the global node.
  • step S503 the global node determines the remaining service time T serve for UE #1, where T serve is the remaining time for the intermediate node #i to provide services for the local UE.
  • step S504 if T serve satisfies the predetermined condition, the global node makes a switching decision.
  • the estimation of T serve can be determined by the global node based on the status information Info V of the intermediate node #i and the status information Info U of the UE #l.
  • the global node can estimate the link quality and connection time (such as the time when RSRP is greater than a certain threshold) between it and the intermediate node #i, and the link quality and connection time (such as the time when RSRP is greater than a certain threshold) between the intermediate node #i and UE#l time greater than a certain threshold), and then estimate the time T serve that the intermediate node #i can serve UE#l.
  • the global node may determine the time when the link between it and the intermediate node #i and the link between the intermediate node #i and UE#l simultaneously meet the corresponding requirements as the time when the intermediate node #i can serve the UE#l T serve .
  • the estimation of T 1 , T 2 , and T train can be determined by the global node based on the status information Info U of all UEs and the status information Info V of all intermediate nodes.
  • the estimation of t 1 , t 2 , and t train can be determined by the global node based on the status information Info V of the intermediate node #i and the status information Info U of all local UEs served by the intermediate node #i.
  • the estimation of t 1 , t 2 , t train can be determined by the intermediate node #i based on its own status information Info V and the status information Info U of all local UEs it serves, and sent to the global node in step S502 .
  • the global node and each intermediate node can be performed periodically, or it can also be triggered by certain triggering events, such as intermediate nodes or UE. Sudden movement etc.
  • step S506 the global node sends the handover decision to the intermediate node #i.
  • step S508 the intermediate node #i sends the received handover decision to the UE #1.
  • step S510 UE#1 performs handover based on the received handover decision.
  • the global node can send the switching decision immediately after making the switching decision, or it can send it when the global model is broadcast after the global aggregation is completed.
  • the handover decision can be sent in the traditional way, that is, sent to the intermediate node #i through the Uu link (Downlink), and then the intermediate node #i passes through PC5 (Sidelink) Sent to UE#l.
  • the global model is sent through broadcast and can be received by all intermediate nodes.
  • the handover decision of the global node is not in the form of broadcast, but is only sent to the UE#l that needs to perform handover and its connected intermediate node #i.
  • Figure 6A shows the situation where T serve > T train when UE#1 is a local UE.
  • the UE may participate in completing the global aggregation of this round and the next round. Therefore, switching will not be performed until the end of T train .
  • FIG. 6B shows the situation of T 1 ⁇ T serve ⁇ T train when UE#1 is a local UE.
  • the UE can participate in completing this round of global aggregation, but its service time cannot support the completion of the next round of global aggregation, so it will switch after this round of global aggregation.
  • the global node broadcasts and sends the global model after global aggregation in this round.
  • each UE and the intermediate node have the same model, and the handover only needs to consider the establishment and release of the communication link, without considering the transfer of the model, the divergence of the intermediate model, etc.
  • FIG. 6C shows the case where t train ⁇ T serve ⁇ T 1 when UE#1 is a local UE.
  • the service provided by the original intermediate node #i to the UE cannot complete the current round of global model aggregation, but can complete the intermediate aggregation of the current round and the next round of intermediate node #i. Therefore, switching will not be performed until the end of t train .
  • the global node may estimate the improved remaining service time T′ serve assuming that one or more of the following operations are performed: improve the UE Transmit power of one or more of #1, intermediate node #i, and global nodes; allocate more transmission resources to one or both of UE #1 and intermediate node #i; and reduce UE #1, intermediate nodes #i and the RSRP threshold of one or more of the global nodes. If T′ serve >T 1 , the global node performs the one or more operations and instructs UE#1 to switch after the end of this global aggregation.
  • Figure 6D shows the situation when t 1 ⁇ T serve ⁇ t train when UE#1 is a local UE.
  • the service provided by the original intermediate node #i to the UE can participate in completing the intermediate aggregation of the current round of intermediate node #i, but its service time cannot support the completion of the intermediate aggregation of the next round of intermediate node #i. Therefore, the switching is performed after the intermediate aggregation of the intermediate node #i in this round is completed.
  • the intermediate node #i sends the intermediate model after intermediate aggregation of this round to the UE it serves. At this time, the UE and the intermediate node #i have the same model.
  • the intermediate node #i needs to send the intermediate model to the intermediate node #j. If the UE switches to be served directly by the global node, the UE directly uploads the local model to the global node for global aggregation after completing k 1 k 2 local iterations.
  • FIG. 6E shows the case where T serve ⁇ t 1 when UE#1 is a local UE.
  • the service provided by the original intermediate node #i to UE #l cannot support the completion of intermediate aggregation by the intermediate node #i in this round. So just make the switch. If UE #l switches to intermediate node #j and intermediate node #j has no service user, intermediate node #i needs to send the intermediate model to intermediate node #j. If the UE switches to be served directly by the global node, the UE directly uploads the local model to the global node for global aggregation after completing k 1 k 2 local iterations.
  • the global node may estimate the improved remaining service time T′ serve assuming that one or more of the following operations are performed: improve the UE #l, middle section transmit power of one or more of point #i and global nodes; allocate more transmission resources to one or both of UE #1 and intermediate node #i; and reduce UE #1, intermediate node #i and global The RSRP threshold of one or more nodes. If T′ serve >t 1 , the global node performs the one or more operations and instructs UE#1 to switch after the end of this intermediate aggregation.
  • Figure 7 shows a global UE handover process according to an embodiment of the present disclosure.
  • step S791 UE#1 sends its own status information Info U to the global node.
  • step S792 the global node determines the remaining service time T serve for UE #1, where T serve is the remaining time for the global node to provide services for UE #1.
  • step S794 if T serve satisfies the predetermined condition, the global node makes a switching decision.
  • the estimate of T serve can be determined by the global node based on the status information Info U of UE#1.
  • the global node can estimate the link quality and connection time between it and UE#l (such as the time when RSRP is greater than a certain threshold), and then estimate the time T serve that the global node can serve UE#l. For example, the global node may determine the time when the link between it and UE#l meets the corresponding requirements as the time T serve when the global node can serve UE#l.
  • the estimation of T 1 , T 2 , and T train can be determined by the global node based on the status information Info U of all UEs and the status information Info V of all intermediate nodes.
  • the global node can be performed periodically, or it can also be triggered by certain triggering events, such as the sudden movement of the UE.
  • step S796 the global node sends the handover decision directly to UE#1.
  • step S798, UE#1 performs handover based on the received handover decision.
  • the global node can send the switching decision immediately after making the switching decision, or it can send it when the global model is broadcast after the global aggregation is completed.
  • the handover decision can be sent directly to UE#l through the Uu link (Downlink).
  • the global model is sent through broadcast, and can be received by each intermediate node and the global UE.
  • the handover decision of the global node is not in the form of broadcast, but is only sent to the UE#1 that needs to perform handover.
  • FIG. 8A shows the situation where T serve >T train when UE#1 is a global UE.
  • UE#1 can participate in completing the global aggregation of this round and the next round. Therefore, switching will not be performed until the end of T train .
  • FIG. 8B shows the situation of T 1 ⁇ T serve ⁇ T train when UE#1 is a global UE.
  • UE#l can participate in completing this round of global aggregation, but its service time cannot support the completion of the next round of global aggregation, so it will switch after this round of global aggregation.
  • the global node broadcasts and sends the global model after global aggregation in this round.
  • each UE#1 and the intermediate node have the same model, and the handover only needs to consider the establishment and release of the communication link, without considering the transfer of the model, the divergence of the intermediate model, etc.
  • FIG. 8C shows the case where T serve ⁇ T 1 when UE#1 is a global UE.
  • the service provided by the global node to UE#l cannot support the completion of this round of global aggregation. So just make the switch.
  • the global node may estimate the improved remaining service time T′ serve assuming that one or more of the following operations are performed: improve the UE transmit power of one or both of #1 and the global node; allocate more transmission resources to UE#1; and reduce the RSRP threshold of one or both of UE#1 and the global node. If T′ serve >T 1 , the global node performs the one or more operations and instructs UE#1 to switch after the end of this global aggregation.
  • Embodiments of the present disclosure can be applied to 5G core networks.
  • Figure 9 shows the 5G core network SBA (Service-based Architecture) architecture and some of its network functions (NF).
  • SBA Service-based Architecture
  • NF network functions
  • AF Application Function
  • application layer which can be internal applications of the operator or third-party AF (such as video servers and game servers).
  • NEF Network Exposure Function
  • NWDAF Network Data Analytics Function
  • NWDAF Network Data Analytics Function
  • AMF Access and Mobility Management Function
  • AMF Access and Mobility Management Function
  • PCF Policy Control function
  • NWDAF can analyze the movement of intermediate nodes and handover time to provide information to AF so that AF can calculate the optimal federated learning time information, and send the time information to AMF to affect the mobility management of UE to achieve efficient federated learning.
  • T serve , T 1 , T 2 , T train , t 1 , t 2 , t train can be estimated by NWDAF and the estimation results can be output to AF, or relevant information can be output to AF, and the AF can perform estimation.
  • the AF may also send the handover rules of the embodiment of the present disclosure to the PCF, and the PCF sends the mobility management policy to the AMF to control the handover of the UE.
  • the AF will send the rule that the UE should try to maintain the connection with the intermediate node before global aggregation or intermediate aggregation is completed to the PCF.
  • the PCF will further allocate more transmission resources, increase the transmission power, lower the RSRP threshold or reduce the transmission rate, etc.
  • the policy is sent to AMF.
  • AF can also obtain UE handover information through NEF to control gNB and federated learning applications in the cloud.
  • the base station can be implemented as any type of evolved Node B (eNB), gNB or TRP (Transmit Receive Point), such as macro eNB/gNB and small eNB/gNB.
  • eNB evolved Node B
  • gNB gNode B
  • TRP Transmit Receive Point
  • a small eNB/gNB may be an eNB/gNB that covers a smaller cell than a macro cell, such as a pico eNB/gNB, a micro eNB/gNB, and a home (femto) eNB/gNB.
  • the base station may be implemented as any other type of base station, such as NodeB and Base Transceiver Station (BTS).
  • BTS Base Transceiver Station
  • the base station may include: a main body (also referred to as a base station device) configured to control wireless communications; and one or more remote radio heads (RRH) disposed at a different place from the main body.
  • a main body also referred to as a base station device
  • RRH remote radio heads
  • various types of terminals to be described below may operate as base stations by temporarily or semi-persistently performing base station functions.
  • the user equipment may be implemented as a mobile terminal such as a smartphone, a tablet personal computer (PC), a notebook PC, a portable game terminal, a portable/dongle-type mobile router, and a digital camera, or a vehicle-mounted terminal such as a car navigation device.
  • the user equipment may also be implemented as a terminal performing machine-to-machine (M2M) communication (also known as a machine type communication (MTC) terminal).
  • M2M machine-to-machine
  • MTC machine type communication
  • base stations and user equipment may be implemented as various types of computing devices.
  • Computing device 700 includes processor 701, memory 702, storage 703, network interface 704, and bus 706.
  • the processor 701 may be, for example, a central processing unit (CPU) or a digital signal processor (DSP), and controls the functions of the server 700 .
  • the memory 702 includes random access memory (RAM) and read only memory (ROM), and stores data and programs executed by the processor 701 .
  • the storage device 703 may include storage media such as semiconductor memory and hard disk.
  • the network interface 704 is a wired communication interface used to connect the server 700 to the wired communication network 705 .
  • the wired communication network 705 may be a core network such as an Evolved Packet Core Network (EPC) or a Packet Data Network (PDN) such as the Internet.
  • EPC Evolved Packet Core Network
  • PDN Packet Data Network
  • Bus 706 connects processor 701, memory 702, storage device 703, and network interface 704 to each other.
  • Bus 706 may include two or more buses each having a different speed (such as a high speed bus and a low speed bus).
  • FIG. 11 is a block diagram illustrating a first example of a schematic configuration of a gNB to which the technology of the present disclosure can be applied.
  • gNB 800 includes one or more antennas 810 and base station equipment 820.
  • Base station equipment 820 and each antenna 810 may be connected via RF cables are connected to each other.
  • Antennas 810 each include a single or multiple antenna elements, such as multiple antenna elements included in a multiple-input multiple-output (MIMO) antenna, and are used by base station device 820 to transmit and receive wireless signals.
  • gNB 800 may include multiple antennas 810.
  • multiple antennas 810 may be compatible with multiple frequency bands used by gNB 800.
  • FIG. 11 shows an example in which gNB 800 includes multiple antennas 810, gNB 800 may also include a single antenna 810.
  • the base station device 820 includes a controller 821, a memory 822, a network interface 823, and a wireless communication interface 825.
  • the controller 821 may be, for example, a CPU or a DSP, and operates various functions of higher layers of the base station device 820 . For example, the controller 821 generates data packets based on the data in the signal processed by the wireless communication interface 825 and delivers the generated packets via the network interface 823 . The controller 821 may bundle data from multiple baseband processors to generate bundled packets, and deliver the generated bundled packets. The controller 821 may have logical functions to perform controls such as radio resource control, radio bearer control, mobility management, admission control, and scheduling. This control can be performed in conjunction with nearby gNB or core network nodes.
  • the memory 822 includes RAM and ROM, and stores programs executed by the controller 821 and various types of control data such as terminal lists, transmission power data, and scheduling data.
  • the network interface 823 is a communication interface used to connect the base station device 820 to the core network 824. Controller 821 may communicate with core network nodes or additional gNBs via network interface 823. In this case, the gNB 800 and the core network node or other gNBs may be connected to each other through logical interfaces such as the S1 interface and the X2 interface.
  • the network interface 823 may also be a wired communication interface or a wireless communication interface for a wireless backhaul line. If the network interface 823 is a wireless communication interface, the network interface 823 may use a higher frequency band for wireless communication than the frequency band used by the wireless communication interface 825 .
  • the wireless communication interface 825 supports any cellular communication scheme such as Long Term Evolution (LTE) and LTE-Advanced and provides wireless connectivity to terminals located in the cell of the gNB 800 via the antenna 810 .
  • Wireless communication interface 825 may generally include, for example, a baseband (BB) processor 826 and RF circuitry 827.
  • the BB processor 826 may perform, for example, encoding/decoding, modulation/demodulation, and multiplexing/demultiplexing, and perform layers such as L1, Medium Access Control (MAC), Radio Link Control (RLC), and Packet Data Convergence Protocol ( Various types of signal processing for PDCP)).
  • the BB processor 826 may have some or all of the above-mentioned logical functions.
  • the BB processor 826 may be a memory that stores a communication control program, or a module including a processor and related circuitry configured to execute the program.
  • the update program can cause the functionality of the BB processor 826 to change.
  • the module may be a card or blade that plugs into a slot of the base station device 820. Alternatively, the module may be a chip mounted on a card or blade.
  • the RF circuit 827 can To include, for example, mixers, filters, and amplifiers, and to transmit and receive wireless signals via antenna 810.
  • the wireless communication interface 825 may include multiple BB processors 826 .
  • multiple BB processors 826 may be compatible with multiple frequency bands used by gNB 800.
  • wireless communication interface 825 may include a plurality of RF circuits 827.
  • multiple RF circuits 827 may be compatible with multiple antenna elements.
  • FIG. 11 shows an example in which the wireless communication interface 825 includes multiple BB processors 826 and multiple RF circuits 827 , the wireless communication interface 825 may also include a single BB processor 826 or a single RF circuit 827 .
  • gNB 830 includes one or more antennas 840, base station equipment 850 and RRH 860.
  • the RRH 860 and each antenna 840 may be connected to each other via RF cables.
  • the base station equipment 850 and the RRH 860 may be connected to each other via high-speed lines such as fiber optic cables.
  • Antennas 840 each include a single or multiple antenna elements (such as multiple antenna elements included in a MIMO antenna) and are used by RRH 860 to transmit and receive wireless signals.
  • gNB 830 may include multiple antennas 840.
  • multiple antennas 840 may be compatible with multiple frequency bands used by gNB 830.
  • FIG. 12 shows an example in which gNB 830 includes multiple antennas 840, gNB 830 may also include a single antenna 840.
  • the base station device 850 includes a controller 851, a memory 852, a network interface 853, a wireless communication interface 855 and a connection interface 857.
  • the controller 851, the memory 852, and the network interface 853 are the same as the controller 821, the memory 822, and the network interface 823 described with reference to FIG. 11 .
  • the wireless communication interface 855 supports any cellular communication scheme (such as LTE and LTE-Advanced) and provides wireless communication to terminals located in the sector corresponding to the RRH 860 via the RRH 860 and the antenna 840.
  • the wireless communication interface 855 may generally include a BB processor 856, for example.
  • the BB processor 856 is the same as the BB processor 826 described with reference to FIG. 11 except that the BB processor 856 is connected to the RF circuit 864 of the RRH 860 via the connection interface 857.
  • the wireless communication interface 855 may include multiple BB processors 856.
  • multiple BB processors 856 may be compatible with multiple frequency bands used by gNB 830.
  • FIG. 12 shows an example in which the wireless communication interface 855 includes multiple BB processors 856, the wireless communication interface 855 may also include a single BB processor 856.
  • connection interface 857 is an interface for connecting the base station device 850 (wireless communication interface 855) to the RRH 860.
  • the connection interface 857 may also be a communication module for communication in the above-mentioned high-speed line that connects the base station device 850 (wireless communication interface 855) to the RRH 860.
  • RRH 860 includes a connection interface 861 and a wireless communication interface 863.
  • connection interface 861 is an interface for connecting the RRH 860 (wireless communication interface 863) to the base station device 850.
  • the connection interface 861 may also be a communication module used for communication in the above-mentioned high-speed line.
  • Wireless communication interface 863 transmits and receives wireless signals via antenna 840.
  • Wireless communication interface 863 may generally include RF circuitry 864, for example.
  • RF circuitry 864 may include, for example, mixers, filters, and amplifiers, and transmits and receives wireless signals via antenna 840 .
  • wireless communication interface 863 may include a plurality of RF circuits 864.
  • multiple RF circuits 864 may support multiple antenna elements.
  • FIG. 12 shows an example in which the wireless communication interface 863 includes a plurality of RF circuits 864, the wireless communication interface 863 may also include a single RF circuit 864.
  • the smart phone 900 includes a processor 901, a memory 902, a storage device 903, an external connection interface 904, a camera 906, a sensor 907, a microphone 908, an input device 909, a display device 910, a speaker 911, a wireless communication interface 912, one or more Antenna switch 915, one or more antennas 916, bus 917, battery 918, and auxiliary controller 919.
  • the processor 901 may be, for example, a CPU or a system on a chip (SoC), and controls functions of the application layer and other layers of the smartphone 900 .
  • the memory 902 includes RAM and ROM, and stores data and programs executed by the processor 901 .
  • the storage device 903 may include storage media such as semiconductor memory and hard disk.
  • the external connection interface 904 is an interface for connecting external devices, such as memory cards and Universal Serial Bus (USB) devices, to the smartphone 900 .
  • the camera 906 includes an image sensor such as a charge coupled device (CCD) and a complementary metal oxide semiconductor (CMOS) and generates a captured image.
  • Sensors 907 may include a group of sensors such as measurement sensors, gyroscope sensors, geomagnetic sensors, and acceleration sensors.
  • the microphone 908 converts the sound input to the smartphone 900 into an audio signal.
  • the input device 909 includes, for example, a touch sensor, a keypad, a keyboard, a button, or a switch configured to detect a touch on the screen of the display device 910, and receives an operation or information input from a user.
  • the display device 910 includes a screen such as a liquid crystal display (LCD) and an organic light emitting diode (OLED) display, and displays an output image of the smartphone 900 .
  • the speaker 911 converts the audio signal output from the smartphone 900 into sound.
  • the wireless communication interface 912 supports any cellular communication scheme such as LTE and LTE-Advanced, and performs wireless communication.
  • the wireless communication interface 912 may generally include a BB processor 913 and an RF circuit 914, for example.
  • the BB processor 913 can perform, for example, encoding/decoding, modulation/demodulation, and multiplexing/demultiplexing, and perform various types of signal processing for wireless communication.
  • RF circuitry 914 may include, for example, mixers, filters, and amplifiers, and transmit and receive wireless signals via antenna 916 .
  • the wireless communication interface 912 may have a BB processor 913 and A chip module of RF circuit 914. As shown in FIG.
  • the wireless communication interface 912 may include multiple BB processors 913 and multiple RF circuits 914 .
  • FIG. 13 shows an example in which the wireless communication interface 912 includes a plurality of BB processors 913 and a plurality of RF circuits 914, the wireless communication interface 912 may also include a single BB processor 913 or a single RF circuit 914.
  • the wireless communication interface 912 may support other types of wireless communication schemes, such as short-range wireless communication schemes, near field communication schemes, and wireless local area network (LAN) schemes.
  • the wireless communication interface 912 may include a BB processor 913 and an RF circuit 914 for each wireless communication scheme.
  • Each of the antenna switches 915 switches the connection destination of the antenna 916 between a plurality of circuits included in the wireless communication interface 912 (for example, circuits for different wireless communication schemes).
  • Antennas 916 each include a single or multiple antenna elements (such as multiple antenna elements included in a MIMO antenna) and are used by wireless communication interface 912 to transmit and receive wireless signals.
  • smartphone 900 may include multiple antennas 916.
  • FIG. 13 shows an example in which smartphone 900 includes multiple antennas 916 , smartphone 900 may also include a single antenna 916 .
  • smartphone 900 may include antennas 916 for each wireless communication scheme.
  • the antenna switch 915 may be omitted from the configuration of the smartphone 900 .
  • the bus 917 connects the processor 901, the memory 902, the storage device 903, the external connection interface 904, the camera 906, the sensor 907, the microphone 908, the input device 909, the display device 910, the speaker 911, the wireless communication interface 912 and the auxiliary controller 919 to each other. connect.
  • the battery 918 provides power to the various blocks of the smartphone 900 shown in Figure 13 via feeders, which are partially shown in the figure as dotted lines.
  • the auxiliary controller 919 operates the minimum necessary functions of the smartphone 900 in the sleep mode, for example.
  • the car navigation device 920 includes a processor 921, a memory 922, a global positioning system (GPS) module 924, a sensor 925, a data interface 926, a content player 927, a storage media interface 928, an input device 929, a display device 930, a speaker 931, a wireless Communication interface 933, one or more antenna switches 936, one or more antennas 937, and battery 938.
  • GPS global positioning system
  • the processor 921 may be, for example, a CPU or an SoC, and controls the navigation function and other functions of the car navigation device 920 .
  • the memory 922 includes RAM and ROM, and stores data and programs executed by the processor 921 .
  • the GPS module 924 measures the location (such as latitude, longitude, and altitude) of the car navigation device 920 using GPS signals received from GPS satellites.
  • Sensors 925 may include a set of sensors such as gyroscope sensors, geomagnetic sensors and air pressure sensor.
  • the data interface 926 is connected to, for example, the vehicle-mounted network 941 via a terminal not shown, and acquires data generated by the vehicle (such as vehicle speed data).
  • the content player 927 reproduces content stored in storage media, such as CDs and DVDs, which are inserted into the storage media interface 928 .
  • the input device 929 includes, for example, a touch sensor, a button, or a switch configured to detect a touch on the screen of the display device 930, and receives an operation or information input from a user.
  • the display device 930 includes a screen such as an LCD or an OLED display, and displays an image of a navigation function or reproduced content.
  • the speaker 931 outputs the sound of the navigation function or the reproduced content.
  • the wireless communication interface 933 supports any cellular communication scheme such as LTE and LTE-Advanced, and performs wireless communication.
  • Wireless communication interface 933 may generally include, for example, BB processor 934 and RF circuitry 935.
  • the BB processor 934 can perform, for example, encoding/decoding, modulation/demodulation, and multiplexing/demultiplexing, and perform various types of signal processing for wireless communications.
  • the RF circuit 935 may include, for example, a mixer, filter, and amplifier, and transmit and receive wireless signals via the antenna 937 .
  • the wireless communication interface 933 may also be a chip module on which the BB processor 934 and the RF circuit 935 are integrated. As shown in FIG.
  • the wireless communication interface 933 may include multiple BB processors 934 and multiple RF circuits 935 .
  • FIG. 14 shows an example in which the wireless communication interface 933 includes a plurality of BB processors 934 and a plurality of RF circuits 935, the wireless communication interface 933 may also include a single BB processor 934 or a single RF circuit 935.
  • the wireless communication interface 933 may support other types of wireless communication schemes, such as short-range wireless communication schemes, near field communication schemes, and wireless LAN schemes.
  • the wireless communication interface 933 may include a BB processor 934 and an RF circuit 935 for each wireless communication scheme.
  • Each of the antenna switches 936 switches the connection destination of the antenna 937 between a plurality of circuits included in the wireless communication interface 933, such as circuits for different wireless communication schemes.
  • the antennas 937 each include a single or multiple antenna elements (such as multiple antenna elements included in a MIMO antenna) and are used by the wireless communication interface 933 to transmit and receive wireless signals.
  • the car navigation device 920 may include a plurality of antennas 937 .
  • FIG. 14 shows an example in which the car navigation device 920 includes a plurality of antennas 937, the car navigation device 920 may also include a single antenna 937.
  • the car navigation device 920 may include an antenna 937 for each wireless communication scheme.
  • the antenna switch 936 may be omitted from the configuration of the car navigation device 920.
  • the battery 938 provides power to the various blocks of the car navigation device 920 shown in FIG. 14 via feeders, which are partially shown as dashed lines in the figure. Battery 938 accumulates power provided from the vehicle.
  • the technology of the present disclosure may also be implemented to include a car navigation device 920, a vehicle network 941, and a vehicle module In-vehicle system (or vehicle) 940 of one or more blocks in 942 .
  • the vehicle module 942 generates vehicle data such as vehicle speed, engine speed, and fault information, and outputs the generated data to the in-vehicle network 941 .
  • DSP digital signal processor
  • ASIC application specific integrated circuit
  • FPGA field-programmable gate array
  • a general purpose processor may be a microprocessor, but alternatively the processor may be any conventional processor, controller, microcontroller and/or state machine.
  • a processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors combined with a DSP core, and/or any other such configuration.
  • the functionality described herein may be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored on or transmitted as one or more instructions or code on a non-transitory computer-readable medium. Other examples and implementations are within the scope and spirit of the disclosure and appended claims. For example, given the nature of software, the functions described above may be performed using software executed by a processor, hardware, firmware, hardwiring, or any combination of these. Features that implement a function may also be physically located at various locations, including being distributed such that portions of the function are implemented at different physical locations.
  • Non-transitory computer-readable media can be any available non-transitory media that can be accessed by a general purpose computer or special purpose computer.
  • non-transitory computer-readable media may include RAM, ROM, EEPROM, flash memory, CD-ROM, DVD or other optical disk storage, magnetic disk storage or other magnetic storage devices, or can be used Any other medium that carries or stores the desired program code components in the form of instructions or data structures and can be accessed by a general or special purpose computer or a general or special purpose processor.
  • the present disclosure also includes the following embodiments.
  • a network-side electronic device for federated learning including a processing circuit, the processing circuit being configured as:
  • determining a model aggregation time and a remaining service time T serve for a user device connected directly to the network or indirectly via an intermediate node;
  • the status information of the user equipment includes one or more of the following: channel status, computing power, local data information, power, location and mobility information.
  • processing circuit is further configured to receive status information of the intermediate node, and the model aggregation time and the remaining service time T serve are further based on the The status information of the intermediate nodes is determined.
  • status information of the intermediate node includes one or more of the following: channel status, computing capability, location, and mobility information.
  • the model aggregation time includes the remaining time T 1 required for this global aggregation and the time T 2 required for the next global aggregation, and
  • T 1 ⁇ T serve ⁇ T 1 +T 2 a handover decision is made for the user equipment, and the handover decision instructs the user equipment to perform handover after the end of this global aggregation.
  • T′ serve >T 1 the one or more operations are performed, and a switching decision is made for the user equipment.
  • the switching decision instructs the user equipment to switch after the end of this global aggregation.
  • model aggregation time includes the remaining time t 1 required for this intermediate aggregation at the intermediate node and the time t 2 required for the next intermediate aggregation.
  • T′ serve >t 1 the one or more operations are performed, and a switching decision is made for the user equipment, where the switching decision instructs the user equipment to switch after the end of this intermediate aggregation.
  • An intermediate node-side electronic device for federated learning comprising a processing circuit configured to:
  • a handover decision for the user equipment is received from the network, wherein the handover decision is made when the model aggregation time and the remaining service time T serve for the user equipment satisfy a predetermined condition, and the user equipment via the Intermediate nodes are indirectly connected to said network;
  • processing circuit is further configured to:
  • processing circuit is further configured to:
  • the model aggregation time and the remaining service time T serve are determined based on at least the status information of the user equipment and the status information of the intermediate node.
  • processing circuit is further configured to:
  • the other parts of the model aggregation time and the remaining service time T serve are determined based on at least the status information of the user equipment and the status information of the intermediate node.
  • a user equipment side electronic device for federated learning comprising a processing circuit, the processing circuit being configured to:
  • the handover decision is made when the model aggregation time and the remaining service time T serve for the user equipment meet predetermined conditions, and the user equipment is directly connected to the network or indirectly connected to the network via an intermediate node;
  • Handover is performed based on the handover decision.
  • processing circuit is further configured to:
  • processing circuit is further configured to:
  • the model aggregation time and the remaining service time T serve are determined based on at least the status information of the user equipment.
  • a network-side method for federated learning including:
  • determining a model aggregation time and a remaining service time T serve for a user device connected directly to the network or indirectly via an intermediate node;
  • An intermediate node-side method for federated learning including:
  • a handover decision for the user equipment is received from the network, wherein the handover decision is made when the model aggregation time and the remaining service time T serve for the user equipment satisfy a predetermined condition, and the user equipment via the Intermediate nodes are indirectly connected to said network;
  • a user-device-side method for federated learning including:
  • the handover decision is made when the model aggregation time and the remaining service time T serve for the user equipment meet predetermined conditions, and the user equipment is directly connected to the network or indirectly connected to the network via an intermediate node;
  • Handover is performed based on the handover decision.
  • a non-transitory computer-readable storage medium having stored thereon program instructions that, when executed by a processor, cause the processor to perform the method of any one of embodiments 20 to 22.
  • a computer program product comprising program instructions which, when executed by a processor, cause the processor to perform a method according to any one of embodiments 20 to 22.

Abstract

本公开涉及用于分层联邦学习网络中的切换的装置、方法和介质。提供了一种用于联邦学习的网络侧电子装置，包括处理电路，所述处理电路被配置为：确定模型聚合时间和针对用户设备的剩余服务时间，其中所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；在所述模型聚合时间和所述剩余服务时间满足预定条件的情况下，做出针对所述用户设备的切换决定；以及发送针对所述用户设备的所述切换决定。

Description

用于分层联邦学习网络中的切换的装置、方法和介质
相关申请的交叉引用
本申请要求于2022年8月5日递交、申请号为202210936728.X、名称为“分层联邦学习网络中的切换”的中国专利申请的优先权，其全部内容通过引用并入本文。
技术领域
本公开涉及用于分层联邦学习网络中的切换的装置、方法和介质。
背景技术
在联邦学习网络中，各用户设备(UE)通过无线信道接入基站，将本地所学习到的模型通过基站上传至服务器，服务器进行聚合后再通过基站将聚合的模型分发至各用户设备。然而，由于用户设备的移动性、用户设备与基站之间的无线信道的变化等，用户设备可能需要在执行联邦学习任务时进行切换。
发明内容
本公开提供了用于分层联邦学习网络中的切换的装置、方法和介质。
根据本公开的一个方面，提供了一种用于联邦学习的网络侧电子装置，包括处理电路，所述处理电路被配置为：确定模型聚合时间和针对用户设备的剩余服务时间，其中所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；在所述模型聚合时间和所述剩余服务时间满足预定条件的情况下，做出针对所述用户设备的切换决定；以及发送针对所述用户设备的所述切换决定。
根据本公开的又一个方面，提供了一种用于联邦学习的中间节点侧电子装置，包括处理电路，所述处理电路被配置为：从网络接收针对用户设备的切换决定，其中所述切换决定是在模型聚合时间和针对所述用户设备的剩余服务时间满足预定条件的情况下做出的，并且所述用户设备经由所述中间节点间接连接到所述网络；将所述切换决定发送至所述用户设备。
根据本公开的又一个方面，提供了一种用于联邦学习的用户设备侧电子装置，包括处理电路，所述处理电路被配置为：接收来自网络的针对用户设备的切换决定，其中所述切换决定是在模型聚合时间和针对所述用户设备的剩余服务时间满足预定条 件的情况下做出的，并且所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；基于所述切换决定进行切换。
根据本公开的又一个方面，提供了一种用于联邦学习的网络侧方法，包括：确定模型聚合时间和针对用户设备的剩余服务时间，其中所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；在所述模型聚合时间和所述剩余服务时间满足预定条件的情况下，做出针对所述用户设备的切换决定；以及发送针对所述用户设备的所述切换决定。
根据本公开的又一个方面，提供了一种用于联邦学习的中间节点侧方法，包括：从网络接收针对用户设备的切换决定，其中所述切换决定是在模型聚合时间和针对所述用户设备的剩余服务时间满足预定条件的情况下做出的，并且所述用户设备经由所述中间节点间接连接到所述网络；将所述切换决定发送至所述用户设备。
根据本公开的又一个方面，提供了一种用于联邦学习的用户设备侧方法，包括：接收来自网络的针对用户设备的切换决定，其中所述切换决定是在模型聚合时间和针对所述用户设备的剩余服务时间满足预定条件的情况下做出的，并且所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；基于所述切换决定进行切换。
根据本公开的又一个方面，提供了一种非暂态计算机可读存储介质，其上存储了程序指令，所述程序指令在由处理器执行时使处理器执行本公开的方法。
根据本公开的又一个方面，提供了一种计算机程序产品，包括程序指令，所述程序指令在由处理器执行时使处理器执行本公开的方法。
附图说明
当结合附图考虑实施例的以下具体描述时，可以获得对本公开更好的理解。在各附图中使用了相同或相似的附图标记来表示相同或者相似的部件。各附图连同下面的具体描述一起包含在本说明书中并形成说明书的一部分，用来例示说明本公开的实施例和解释本公开的原理和优点。
图1示出了传统联邦学习网络的示例性结构。
图2示出了根据本公开的实施例的分层联邦学习网络的示例性结构。
图3示出了根据本公开的实施例的分层联邦学习网络的示例性联邦学习过程。
图4A至4C示出了局部UE的不同切换场景。
图5示出了根据本公开的实施例的局部UE的示例性切换流程。
图6A至6E示出了局部UE的剩余服务时间满足不同条件的情况。
图7示出了根据本公开的实施例的全局UE的示例性切换流程。
图8A至8C示出了全局UE的剩余服务时间满足不同条件的情况。
图9示出了5G核心网SBA架构及其部分网络功能(NF)。
图10是示出可以应用本公开的技术的计算设备的示意性配置的示例的框图。
图11是示出可以应用本公开的技术的gNB的示意性配置的第一示例的框图。
图12是示出可以应用本公开的技术的gNB的示意性配置的第二示例的框图。
图13是示出可以应用本公开的技术的智能电话的示意性配置的示例的框图。
图14是示出可以应用本公开的技术的汽车导航设备的示意性配置的示例的框图。
具体实施方式
在下文中，将参照附图详细地描述本公开的优选实施例。注意，在本说明书和附图中，用相同的附图标记来表示具有基本上相同的功能和结构的结构元件，并且省略对这些结构元件的重复说明。
图1示出了传统联邦学习网络的示例性结构。在传统联邦学习网络中，UE 1201、1202、1203…120K-1、120K通过基站(未示出)直接连接到服务器110，并且将本地模型上传至服务器110进行全局聚合，具体流程如下。
首先，UE 120接入服务器110，通过下行链路传输获取初始全局模型w0：
其中，表示第k个UE的初始本地模型。
然后，UE 120使用存储在本地的数据进行学习，完成第r+1次本地模型更新的本地迭代:

其中，表示第k个UE在第r次迭代后的本地模型，表示第k个UE的梯度，η表示学习率(Learning Rate)，表示第k个UE的损失函数。
然后，UE 120通过上行链路，将学习到的本地模型或梯度上传至服务器110。服务器110将收集到的来自各UE的本地模型进行聚合，完成全局模型的更新：
其中pk为来自于各UE的本地模型的权值，通常设置为Di表示第i个UE的数据集，|Di|表示数据集的大小。
最后，服务器110将更新后的全局模型wr+1再次分发到UE 120，然后重复上述步骤直至模型收敛。
然而，该网络结构有其相应的局限性。首先，基站的覆盖范围有限，对于覆盖范围外的UE无法提供服务。其次，基站覆盖范围内存在一些通信速率较低的地区，对于该地区的UE，即使在覆盖范围内，服务质量无法保证。此外，基站的上下行链路通信资源(如频率资源、载波个数等)有限，无法支持过多UE的同时接入。
图2示出了根据本公开的实施例的分层联邦学习网络的示例性结构。该网络结构由两层组成：第一层由中间节点220与其服务的UE 230(本文将连接到中间节点的UE称为局部UE)组成；第二层由全局节点210与其连接的中间节点220以及与全局节点直接连接的UE 240(本文将直接连接到全局节点的UE称为全局UE)组成。
全局节点由基站和/或与基站连接的服务器实现。中间节点可以由具有移动性的车载中继(Vehicle Mounted Relay，VMR)，无人机(Unmanned Aerial Vehicle，UAV)等。替代地，中间节点可以被实现为位置固定的路边单元(Road Side Unit，RSU)，边缘(Edge)节点等实现。
下面结合图3说明根据本公开的实施例的分层联邦学习网络的示例性联邦学习过程。
在步骤S302，局部UE 230(作为一级节点)进行例如k1次本地迭代，再将得到的本地模型上传至中间节点220(作为二级节点)进行中间聚合。在步骤S304，中间节点220对从局部UE 230接收到的本地模型进行一次中间聚合，获得中间模型。
在步骤S306，中间节点220判断是否完成了总共例如k2次中间聚合。如果中间聚合的总次数不够k2次，则中间节点220将中间模型分发给局部UE 230。局部UE 230在接收到中间模型之后将本地模型更新为中间模型并进行新一轮的本地迭代。如果中间聚合的总次数已达k2次，则中间节点220将中间模型上传至全局节点210进行全局聚合。
在步骤S308，如果存在全局节点210直接服务的全局UE 240，则全局UE 240进行例如k1k2次本地迭代，再将得到的本地模型上传至全局节点210。
在步骤S310，全局节点210(或其连接的服务器等网络设备)对中间节点220上传的中间模型以及全局UE 240上传的本地模型(如果有的话)进行全局聚合，并将得到的全局模型分发给中间节点220和全局UE 240(如果有的话)。中间节点220在接收到全局模型之后，将中间模型更新为全局模型，并将全局模型分发给其服务的局部UE 230。局部 UE 230和全局UE 240在接收到全局模型之后，将本地模型更新为全局模型。
在本公开的实施例的分层联邦学习网络结构中，局部UE与中间节点连接，缩短了通信距离，保障了通信服务质量。此外，多个局部UE与中间节点直接通信，中间节点再与全局节点连接。尽管中间节点服务于多个局部UE，但其与全局节点之间的每次通信只上传聚合得到的中间模型。其数据量相当于一个全局UE与全局节点通信或者一个局部UE与中间节点通信的数据量。大大降低了全局节点的负载，缓解了全局节点处通信资源不足的问题。此外，该网络结构充分利用了中间节点的功能，即，中间节点不只是单纯作为中继进行传输，同时也作为第一层结构的聚合者参与计算，完成模型的中间聚合。
下面将详细介绍根据本公开的实施例的局部UE、中间节点、全局UE和全局节点中的具体联邦学习处理。假设有C个全局UE，M个中间节点，第i个中间节点服务的局部UE的个数为ni。在联邦学习过程中，UE、中间节点可以通过上传模型参数或者梯度来上传联邦学习模型。因此，下面将分别说明上传模型参数和梯度的联邦学习过程。
首先介绍上传模型参数的联邦学习过程。全局节点首先初始化全局模型w0，并将初始化的全局模型分发至全局UE和中间节点。中间节点再将初始化的全局模型分发至局部UE。此时，全局节点、全局UE、中间节点和局部UE具有相同的联邦学习模型：
其中，
----中间节点#i的模型参数，
----中间节点#i服务的第l个局部UE的模型参数，
ni----中间节点#i服务的局部UE的个数。
各局部UE基于本地存储的数据进行本地迭代：
其中，
----中间节点#i服务的第l个局部UE在第r1+1次本地迭代后的本地模型参数，
----中间节点#i服务的第l个局部UE在第r1次本地迭代后的本地模型参数，
----中间节点#i服务的第l个局部UE在第r1+1次本地迭代时的本地梯度，
----中间节点#i服务的第l个局部UE在第r1+1次本地迭代时的本地损失 函数。
局部UE经过k1次的本地迭代后得到本地模型参数并上传至中间节点。中间节点在接收到其所有服务的局部UE上传的本地模型后，进行一次中间聚合：
其中，
----中间节点#i第r2+1次中间聚合后的中间模型参数，
----中间节点#i第r2次中间聚合后的中间模型参数，
pi,l----中间节点#i服务的第l个局部UE上传的本地模型对应的权值，其通常需要利用上一次的中间模型作为参考来确定，
r1＝k1r2,即中间节点每完成一次中间聚合，其服务的局部UE完成k1次本地模型更新。
然后，中间节点将进行中间聚合得到的模型(此处称为中间模型)分发至其服务的局部UE。
重复式(2)和(3)的过程。当r2不是k2的整数倍时，中间节点将中间聚合后的模型分发至其服务的各个局部UE。局部UE使用接收到的中间模型更新本地模型。当r2是k2的整数倍时，中间节点在完成第k2次中间聚合后，将得到的中间模型上传至全局节点。
此外，平行于式(2)和(3)的过程，全局UE基于本地数据单独进行本地迭代，其类似于式(2)的过程：
其中，
----第j个全局UE在第r1+1次本地迭代后的本地模型参数，
----第j个全局UE在第r1次本地迭代后的本地模型参数，
----第j个全局UE在第r1+1次本地迭代时的本地梯度，
----第j个全局UE在第r1+1次本地迭代时的本地损失函数。
全局UE进行k1k2次本地迭代后得到本地模型参数并上传至全局节点。
全局节点在接收到所有中间节点上传的中间模型和所有全局UE上传的本地模型后，进行全局聚合：
其中，
----全局节点在进行第r3+1次全局聚合后的全局模型参数，
----第i个中间节点上传的中间模型参数，
----第j个全局UE上传的本地模型参数，
pi----第i个中间节点上传的中间模型对应的权值，其通常需要利用上一次的全局模型作为参考来确定，
pc,j----第j个全局UE上传的本地模型对应的权值，其通常需要利用上一次的全局模型作为参考来确定，
r2＝k2r3，即全局节点每完成一次全局聚合，其连接的中间节点完成k2次中间聚合。
全局节点进行全局模型聚合得到全局模型参数并分发至各中间节点和全局UE。
重复式(2)至(5)的过程直至全局模型收敛。
接下来介绍UE、中间节点上传梯度的联邦学习过程。首先，全局节点初始化全局模型w0，并分发至全局UE和中间节点。中间节点再将初始化的全局模型分发至局部UE。此时全局节点、全局UE、中间节点以及全局UE具有相同的学习模型：
其中，
----中间节点#i的本地模型参数，
----中间节点#i服务的第l个UE的本地模型参数，
ni----中间节点#i服务的局部UE的个数。
各局部UE基于本地存储的数据，进行本地迭代：
其中，
----中间节点#i服务的第l个局部UE在第r1+1次本地迭代后的本地模型参数，
----中间节点#i服务的第l个局部UE在第r1次本地迭代后的本地模型参数，
----中间节点#i服务的第l个局部UE在第r1+1次本地迭代时的本地梯度，
----中间节点#i服务的第l个局部UE在第r1+1次本地迭代时的本地损失函数。
局部UE经过k1次本地迭代后得到本地模型参数并计算k1次本地迭代的梯度的和
局部UE上传梯度至中间节点。中间节点在接收到其所有服务的局部UE上传的本地梯度后，进行一次中间聚合：
其中，
----中间节点#i第r2+1次中间聚合后的中间模型参数，
----中间节点#i第r2次中间聚合后的中间模型参数，
----中间节点#i第r2+1次中间聚合时的梯度，
pi,l----中间节点#i服务的第l个局部UE上传的本地模型(梯度)对应的权值，
r1＝k1r2，即中间节点每完成一次中间聚合，其服务的局部UE完成k1次本地模型更新。
重复式(7)至(9)的过程。当r2不是k2的整数倍时，中间节点将中间聚合后的模型(此处称之为中间模型)分发至其服务的各个局部UE。局部UE使用接收到的中间模型更新本地模型。当r2是k2的整数倍时，中间节点在完成第k2次中间聚合后，将得到的中间模型上传至全局节点。
中间节点经过k2次的中间聚合后得到中间模型参数并计算k2次中间聚合的梯度的和
中间节点将梯度上传至全局节点进行全局聚合。
此外，平行于式(7)至(10)的过程，全局UE基于本地数据单独进行本地迭代， 其类似于式(7)的过程：
其中，
----第j个全局UE在第r1+1次模型更新后的本地模型参数，
----第j个全局UE在第r1次模型更新前的本地模型参数，
----第j个全局UE在第r1次模型更新时的本地梯度，
----第j个全局UE在第r1次模型更新时的本地损失函数，
全局UE进行k1k2次本地迭代后得到本地模型参数并上传梯度至全局节点：
全局节点在接收到所有中间节点上传的中间模型和所有全局UE上传的本地模型后，进行全局聚合：
其中，
----全局节点在进行第r3+1次全局聚合后的全局模型参数，
----第i个中间节点上传的中间模型梯度，
----第j个全局UE上传的本地模型梯度，
pi----第i个中间节点上传的中间模型对应的权值，
pc,j----第j个全局UE上传的本地模型对应的权值，
r3＝k1k2r1，即全局节点每完成一次全局聚合，其连接的中间节点完成k2次中间聚合。
全局节点进行全局模型聚合更新，并将聚合后的模型分发至各中间节点和全局UE。
重复式(7)至(13)的过程直至全局模型收敛。
由于UE和中间节点的移动性、无线信道的变化等，UE可能需要在执行联邦学习任务时进行切换。图4A示出了一个或多个局部UE从中间节点#i切换到中间节点#j，且切换前 中间节点#j有服务的局部UE的场景。在该场景中，中间节点#j处有中间模型，所以不需要从中间节点#i向中间节点#j传输中间模型。
假设不考虑联邦学习，UE#l在满足切换条件时直接进行切换(如接收信号强度RSRP小于一定门限)。如果UE#l的切换发生在中间节点#i的第r2+1次中间聚合的过程中(已完成第r2次中间聚合，但未完成第r2+1次中间聚合，即k1r2<r1<k1(r2+1))，则会导致如下结果。
首先，对于中间节点#i而言，其已与UE#l断开连接，在进行第r2+1次中间聚合时无法接收到UE#l的本地模型。其次，对于中间节点#j而言，其可以接收UE#l上传的本地模型。但是UE#l在r1＝k1(r2+1)次本地迭代后上传的本地模型是基于在r1＝k1r2时从中间节点#i(而不是中间节点#j)接收到的中间模型训练得到的。
一种现有方案是中间节点#j丢弃并不使用UE#l在r1＝k1(r2+1)次本地迭代后上传的本地模型。因此，对于UE#l，只有其在中间节点#i(或中间节点#j)的第r2+1次中间聚合过程中(从第r2次中间聚合的模型发送至UE#l，到UE#l完成k1(r2+1)次本地迭代并上传)完全在相应中间节点的覆盖范围内，才可参与该中间节点的第r2+1次中间聚合。
另一种现有方案是中间节点#j将UE#l在r1＝k1(r2+1)次本地迭代后上传的本地模型用于其第r2+1次中间聚合。但这会导致中间节点#j处中间模型的发散，从而导致系统性能(如全局模型收敛速度)或者全局模型准确度的下降。
图4B示出了中间节点#i服务的所有局部UE切换到中间节点#j，且切换前中间节点#j没有服务的局部UE的场景。该场景可以由中间节点#i的移动导致，也可由UE的移动导致。
假设在完成r2次的中间聚合后，UE从中间节点#i切换到中间节点#j。接下来中间节点#j进行第r2+1次中间聚合，其需要第r2次中间聚合后的中间模型。如果是上传梯度，如式(10)所示，计算第r2+1次中间聚合后的中间模型需要第r2次中间聚合后的中间模型，而此时中间节点#j处没有第r2次中间聚合后的中间模型。即使是上传模型参数，也需要利用上一次的中间模型参数确定计算中间模型时用到的权值。
在该场景中，中间节点#j处没有中间模型，只有最初从全局节点接收到的全局模型。因此，中间节点#i需要将其中间模型传输给中间节点#j。
图4C示出了中间节点#i服务的部分局部UE切换到中间节点#j，且切换前中间节点#j没有服务的局部UE的场景。与切换前相比，切换后，中间节点#i和中间节点#j都缺少一部分UE参与中间聚合，因此会导致模型发散，准确度降低。即在中间聚合过程中进行切换，即使只有部分UE进行切换也会导致中间模型的发散，降低全局模型的准确性。
针对上述问题，本公开的一些实施例使得UE在全局聚合且全局模型广播后进行切换。此时各UE和中间节点的模型相同，均为全局模型，无需进行额外的模型传输，切换代价最小。如不能保证在全局聚合且全局模型广播后进行切换，本公开的一些实施例使得在中间节点完成中间聚合后进行切换。此时中间节点与其服务的所有局部UE的模型相同，且能保证局部UE的训练服务不中断。此外，本公开的一些实施例还通过提高发送功率、降低RSRP门限、分配更多的传输资源(包括时间资源、频率资源)等方法，延长UE和/或中间节点的连接时间至中间聚合或全局聚合完成。从而保障服务的连续性，提高系统性能。
为了说明的简便，首先定义以下变量：
Tserve---全局节点或中间节点对UE提供服务的剩余时间的估计值；
T1---全局节点完成本轮全局聚合所需的剩余时间的估计值，即从当前时刻到全局节点完成本轮全局聚合的时间；
T2---全局节点完成下轮全局聚合所需的时间，即从全局节点完成本轮全局聚合起到全局节点完成下一轮全局聚合的时间；
Ttrain---全局节点完成本轮及下轮全局聚合所需的剩余时间，即从当前时刻到全局节点完成下一轮全局聚合的时间，即Ttrain＝T1+T2；
t1---中间节点完成本轮中间聚合所需的剩余时间，即从当前时刻到中间节点完成本轮中间节点聚合的时间；
t2---中间节点完成下轮中间聚合所需的剩余时间,即从中间节点完成本轮中间节点聚合起到中间节点完成下一轮中间节点聚合的时间；
ttrain---中间节点完成本轮及下一轮全局聚合所需的剩余时间，即从当前时刻到中间节点完成下一轮中间节点聚合的时间,即ttrain＝t1+t2。
首先讨论UE#l是由中间节点#i服务的局部UE的情况，UE#l可以切换为由中间节点#j服务或者直接由全局节点服务。图5示出了根据本公开的实施例的局部UE的示例性切换流程。
在步骤S501，UE#l将自身的状态信息InfoU发送至中间节点#i。UE的状态信息InfoU可以包括信道状态(例如RSRP)、计算能力(例如CPU占用率)、本地数据信息(例如参与模型训练的样本数量、样本维度等)、电量、位置和移动信息(例如，速度、方向、在某地的停留时间等)以及其他信息中的一项或多项。
在步骤S502，中间节点#i将自身的状态信息InfoV以及其服务的局部UE的状态信息 InfoU发送至全局节点。中间节点的状态信息InfoV可以包括信道状态(RSRP)、计算能力(例如CPU占用率)、位置和移动信息(例如，速度、方向、在某地的停留时间等)以及其他信息中的一项或多项。中间节点的信道状态可以包括中间节点与局部UE之间的信道状态以及中间节点与全局节点之间的信道状态。
在步骤S503，全局节点确定针对UE#l的剩余服务时间Tserve，Tserve为中间节点#i为该局部UE提供服务的剩余时间。在步骤S504，在Tserve满足预定条件的情况下，全局节点做出切换决定。
对于Tserve的估计，可由全局节点根据中间节点#i的状态信息InfoV以及UE#l的状态信息InfoU确定。全局节点可以估计其与中间节点#i之间的链路质量和连接时间(如RSRP大于一定阈值的时间)，以及中间节点#i和UE#l之间的链路质量和连接时间(如RSRP大于一定阈值的时间)，进而估计出中间节点#i可服务UE#l的时间Tserve。例如，全局节点可以将其与中间节点#i之间的链路以及中间节点#i和UE#l之间的链路同时满足相应要求的时间确定为中间节点#i可服务UE#l的时间Tserve。
对于T1，T2，Ttrain的估计，可以由全局节点根据所有UE的状态信息InfoU和所有中间节点的状态信息InfoV确定。对于t1，t2，ttrain的估计，可由全局节点根据中间节点#i的状态信息InfoV以及中间节点#i服务的所有局部UE的状态信息InfoU确定。替代地，对于t1，t2，ttrain的估计，可以由中间节点#i根据自身的状态信息InfoV以及其服务的所有局部UE的状态信息InfoU确定，并在步骤S502发送至全局节点。
对于Tserve，T1，T2，Ttrain，t1，t2，ttrain的估计，全局节点和各中间节点可以周期性的进行，也可由某些触发事件触发，如中间节点或者UE的突然移动等。
在步骤S506，全局节点将切换决定发送至中间节点#i。在步骤S508，中间节点#i将接收到的切换决定发送至UE#l。在步骤S510，UE#l基于接收到的切换决定进行切换。
全局节点在做出切换决定后，可以立即发送切换决定，或者可以在全局聚合结束后，广播全局模型时发送。在基站-VMR-UE的分层联邦学习结构中，切换决定的发送可通过传统的方式，即通过Uu链路(Downlink)发送至中间节点#i，再由中间节点#i经PC5(Sidelink)发送至UE#l。全局模型是通过广播发送的，各中间节点均可接收。但是全局节点的切换决定并不是广播的形式，而是仅发送至需要执行切换的UE#l及其连接的中间节点#i。
接下来将参照图6A-6E分别讨论UE#l是局部UE时Tserve所满足的条件，以及在相应条件下是否做出切换决定。
图6A示出了UE#l是局部UE时Tserve>Ttrain的情况。在这种情况下，UE可参与完成本轮和下轮的全局聚合。因此，在Ttrain结束前暂不进行切换。
图6B示出了UE#l是局部UE时T1<Tserve<Ttrain的情况。在这种情况下，UE可参与完成本轮全局聚合，但其服务时间无法支持下一轮全局聚合的完成，则在本轮全局聚合后切换。全局节点将本轮全局聚合后的全局模型广播发送。此时，各UE和中间节点具有相同的模型，切换只需考虑通信链路的建立与释放，无需考虑模型的转移，中间模型的发散等。
图6C示出了UE#l是局部UE时ttrain<Tserve<T1的情况。在这种情况下，原中间节点#i对UE提供的服务不能完成本轮全局模型聚合，但可以完成本轮和下轮的中间节点#i的中间聚合。因此，ttrain结束前暂不进行切换。
在本公开的一些实施例中，在Tserve<T1的情况下，全局节点可以在假设执行以下各个操作中的一个或多个操作的情况下估计提高的剩余服务时间T′serve：提高UE#l、中间节点#i和全局节点中的一个或多个的发送功率；为UE#l和中间节点#i中的一个或二者分配更多的传输资源；以及降低UE#l、中间节点#i和全局节点中的一个或多个的RSRP门限。如果T′serve>T1，则全局节点执行所述一个或多个操作，并且指示UE#l在本次全局聚合结束后进行切换。
图6D示出了UE#l是局部UE时t1<Tserve<ttrain的情况。在这种情况下原中间节点#i对UE提供的服务可参与完成本轮的中间节点#i的中间聚合，但其服务时间无法支持下一轮中间节点#i的中间聚合完成。因此，在本轮中间节点#i的中间聚合结束后进行切换。中间节点#i将本轮中间聚合后的中间模型发送至其服务的UE。此时，UE与中间节点#i具有相同的模型。若UE切换到中间节点#j并且中间节点#j无服务用户，则中间节点#i需要将中间模型发送至中间节点#j。若UE切换到由全局节点直接服务，则UE在完成k1k2次本地迭代后直接将本地模型上传至全局节点进行全局聚合。
图6E示出了UE#l是局部UE时Tserve<t1的情况。在这种情况下，原中间节点#i对UE#l提供的服务无法支持本轮中间节点#i中间聚合的完成。因此，直接进行切换。若UE#l切换到中间节点#j并且中间节点#j无服务用户，则中间节点#i需要将中间模型发送至中间节点#j。若UE切换到由全局节点直接服务，则UE在完成k1k2次本地迭代后直接将本地模型上传至全局节点进行全局聚合。
在本公开的一些实施例中，在Tserve<t1的情况下，全局节点可以在假设执行以下各个操作中的一个或多个操作的情况下估计提高的剩余服务时间T′serve：提高UE#l、中间节 点#i和全局节点中的一个或多个的发送功率；为UE#l和中间节点#i中的一个或二者分配更多的传输资源；以及降低UE#l、中间节点#i和全局节点中的一个或多个的RSRP门限。如果T′serve>t1，则全局节点执行所述一个或多个操作，并且指示UE#l在本次中间聚合结束后进行切换。接下来讨论UE#l是由全局节点直接服务的全局UE的情况，UE#l可以切换为由中间节点#j服务。图7示出了根据本公开的实施例的全局UE的切换流程。
在步骤S791，UE#l将自身的状态信息InfoU发送至全局节点。在步骤S792，全局节点确定针对UE#l的剩余服务时间Tserve，Tserve为全局节点为UE#l提供服务的剩余时间。在步骤S794，在Tserve满足预定条件的情况下，全局节点做出切换决定。
对于Tserve的估计，可由全局节点根据UE#l的状态信息InfoU确定。全局节点可以估计其与UE#l之间的链路质量和连接时间(如RSRP大于一定阈值的时间)，进而估计出全局节点可服务UE#l的时间Tserve。例如，全局节点可以将其与UE#l之间的链路满足相应要求的时间确定为全局节点可服务UE#l的时间Tserve。
对于T1，T2，Ttrain的估计，可以由全局节点根据所有UE的状态信息InfoU和所有中间节点的状态信息InfoV确定。对于Tserve，T1，T2，Ttrain的估计，全局节点可以周期性的进行，也可由某些触发事件触发，如UE的突然移动等。
在步骤S796，全局节点将切换决定直接发送至UE#l。在步骤S798，UE#l基于接收到的切换决定进行切换。
全局节点在做出切换决定后，可以立即发送切换决定，或者可以在全局聚合结束后，广播全局模型时发送。切换决定的发送可通过Uu链路(Downlink)直接发送至UE#l。全局模型是通过广播发送的，各中间节点和全局UE均可接收。但是全局节点的切换决定并不是广播的形式，而是仅发送至需要执行切换的UE#l。
接下来将参照图8A-8C分别讨论UE#l是全局UE时Tserve所满足的条件，以及在相应条件下是否做出切换决定。
图8A示出了UE#l是全局UE时Tserve>Ttrain的情况。在这种情况下，UE#l可参与完成本轮和下轮的全局聚合。因此，在Ttrain结束前暂不进行切换。
图8B示出了UE#l是全局UE时T1<Tserve<Ttrain的情况。在这种情况下，UE#l可参与完成本轮全局聚合，但其服务时间无法支持下一轮全局聚合的完成，则在本轮全局聚合后切换。全局节点将本轮全局聚合后的全局模型广播发送。此时，各UE#l和中间节点具有相同的模型，切换只需考虑通信链路的建立与释放，无需考虑模型的转移，中间模型的发散等。
图8C示出了UE#l是全局UE时Tserve<T1的情况。在这种情况下，全局节点对UE#l提供的服务无法支持本轮全局聚合的完成。因此，直接进行切换。
在本公开的一些实施例中，在Tserve<T1的情况下，全局节点可以在假设执行以下各个操作中的一个或多个操作的情况下估计提高的剩余服务时间T′serve：提高UE#l和全局节点中的一个或二者的发送功率；为UE#l分配更多的传输资源；以及降低UE#l和全局节点中的一个或二者的RSRP门限。如果T′serve>T1，则全局节点执行所述一个或多个操作，并且指示UE#l在本次全局聚合结束后进行切换。
本公开的实施例可以应用于5G核心网。图9示出了5G核心网SBA(Service-based Architecture)架构及其部分网络功能(NF)。
AF(Application Function，应用功能)指应用层的各种服务，可以是运营商内部的应用、也可以是第三方的AF(如视频服务器、游戏服务器)。
NEF(Network Exposure Function，网络开放功能)位于5G核心网和外部第三方应用功能体之间，负责管理对外开放网络数据。所有的外部应用想要访问5G核心网内部数据都必须要通过NEF。
NWDAF(Network Data Analytics Function，网络数据分析功能)可以收集数据，执行分析并将分析结果提供给其他网络功能，如NEF。
AMF(Access and Mobility Management Function，接入和移动性管理功能)负责注册、连接、可达性、移动性及与安全和访问管理和业务授权。
PCF(Policy Control function，策略控制功能)提供其负责的所有移动性，UE访问选择和PDF会话相关的策略。
NWDAF可以分析中间节点的移动以及切换的时间从而为AF提供信息以便AF计算最佳的联邦学习时间信息，并将时间信息发送到AMF从而影响UE的移动管理以便实现高效的联邦学习。例如，可由NWDAF估计Tserve，T1，T2，Ttrain，t1，t2，ttrain并将估计结果输出至AF，或者输出相关的信息至AF，由AF进行估计。
另外，AF也可以将本公开的实施例的切换规则发送至PCF并且由PCF将移动管理策略发送至AMF从而控制UE的切换。例如，AF将在全局聚合或中间聚合未完成前应尽量保持UE与中间节点的连接的规则发送到PCF，PCF进一步将分配更多的传输资源、提高发送功率、降低RSRP门限或者降低传输速率等策略发送至AMF。
此外，AF还可以通过NEF获取UE的切换信息从而控制gNB以及云端的联邦学习应用。
<应用示例>
本公开的技术能够应用于各种产品。基站可以被实现为任何类型的演进型节点B(eNB)、gNB或TRP(Transmit Receive Point)，诸如宏eNB/gNB和小eNB/gNB。小eNB/gNB可以为覆盖比宏小区小的小区的eNB/gNB，诸如微微eNB/gNB、微eNB/gNB和家庭(毫微微)eNB/gNB。代替地，基站可以被实现为任何其它类型的基站，诸如NodeB和基站收发台(BTS)。基站可以包括：被配置为控制无线通信的主体(也称为基站设备)；以及设置在与主体不同的地方的一个或多个远程无线头端(RRH)。另外，下面将描述的各种类型的终端均可以通过暂时地或半持久性地执行基站功能而作为基站工作。
用户设备可以被实现为移动终端(诸如智能电话、平板个人计算机(PC)、笔记本式PC、便携式游戏终端、便携式/加密狗型移动路由器和数字摄像装置)或者车载终端(诸如汽车导航设备)。用户设备还可以被实现为执行机器对机器(M2M)通信的终端(也称为机器类型通信(MTC)终端)。
此外，基站和用户设备均可以被实现为各种类型的计算设备。
[关于计算设备的应用示例]
图10是示出可以应用本公开的技术的计算设备700的示意性配置的示例的框图。计算设备700包括处理器701、存储器702、存储装置703、网络接口704以及总线706。
处理器701可以为例如中央处理单元(CPU)或数字信号处理器(DSP)，并且控制服务器700的功能。存储器702包括随机存取存储器(RAM)和只读存储器(ROM)，并且存储数据和由处理器701执行的程序。存储装置703可以包括存储介质，诸如半导体存储器和硬盘。
网络接口704为用于将服务器700连接到有线通信网络705的有线通信接口。有线通信网络705可以为诸如演进分组核心网(EPC)的核心网或者诸如因特网的分组数据网络(PDN)。
总线706将处理器701、存储器702、存储装置703和网络接口704彼此连接。总线706可以包括各自具有不同速度的两个或更多个总线(诸如高速总线和低速总线)。
[关于基站的应用示例]
(第一应用示例)
图11是示出可以应用本公开的技术的gNB的示意性配置的第一示例的框图。gNB800包括一个或多个天线810以及基站设备820。基站设备820和每个天线810可以经由 RF线缆彼此连接。
天线810中的每一个均包括单个或多个天线元件(诸如包括在多输入多输出(MIMO)天线中的多个天线元件)，并且用于基站设备820发送和接收无线信号。如图11所示，gNB 800可以包括多个天线810。例如，多个天线810可以与gNB 800使用的多个频带兼容。虽然图11示出其中gNB 800包括多个天线810的示例，但是gNB 800也可以包括单个天线810。
基站设备820包括控制器821、存储器822、网络接口823以及无线通信接口825。
控制器821可以为例如CPU或DSP，并且操作基站设备820的较高层的各种功能。例如，控制器821根据由无线通信接口825处理的信号中的数据来生成数据分组，并经由网络接口823来传递所生成的分组。控制器821可以对来自多个基带处理器的数据进行捆绑以生成捆绑分组，并传递所生成的捆绑分组。控制器821可以具有执行如下控制的逻辑功能：该控制诸如为无线资源控制、无线承载控制、移动性管理、接纳控制和调度。该控制可以结合附近的gNB或核心网节点来执行。存储器822包括RAM和ROM，并且存储由控制器821执行的程序和各种类型的控制数据(诸如终端列表、传输功率数据以及调度数据)。
网络接口823为用于将基站设备820连接至核心网824的通信接口。控制器821可以经由网络接口823而与核心网节点或另外的gNB进行通信。在此情况下，gNB 800与核心网节点或其它gNB可以通过逻辑接口(诸如S1接口和X2接口)而彼此连接。网络接口823还可以为有线通信接口或用于无线回程线路的无线通信接口。如果网络接口823为无线通信接口，则与由无线通信接口825使用的频带相比，网络接口823可以使用较高频带用于无线通信。
无线通信接口825支持任何蜂窝通信方案(诸如长期演进(LTE)和LTE-先进)，并且经由天线810来提供到位于gNB 800的小区中的终端的无线连接。无线通信接口825通常可以包括例如基带(BB)处理器826和RF电路827。BB处理器826可以执行例如编码/解码、调制/解调以及复用/解复用，并且执行层(例如L1、介质访问控制(MAC)、无线链路控制(RLC)和分组数据汇聚协议(PDCP))的各种类型的信号处理。代替控制器821，BB处理器826可以具有上述逻辑功能的一部分或全部。BB处理器826可以为存储通信控制程序的存储器，或者为包括被配置为执行程序的处理器和相关电路的模块。更新程序可以使BB处理器826的功能改变。该模块可以为插入到基站设备820的槽中的卡或刀片。可替代地，该模块也可以为安装在卡或刀片上的芯片。同时，RF电路827可 以包括例如混频器、滤波器和放大器，并且经由天线810来传送和接收无线信号。
如图11所示，无线通信接口825可以包括多个BB处理器826。例如，多个BB处理器826可以与gNB 800使用的多个频带兼容。如图11所示，无线通信接口825可以包括多个RF电路827。例如，多个RF电路827可以与多个天线元件兼容。虽然图11示出其中无线通信接口825包括多个BB处理器826和多个RF电路827的示例，但是无线通信接口825也可以包括单个BB处理器826或单个RF电路827。
(第二应用示例)
图12是示出可以应用本公开的技术的gNB的示意性配置的第二示例的框图。gNB 830包括一个或多个天线840、基站设备850和RRH 860。RRH 860和每个天线840可以经由RF线缆而彼此连接。基站设备850和RRH 860可以经由诸如光纤线缆的高速线路而彼此连接。
天线840中的每一个均包括单个或多个天线元件(诸如包括在MIMO天线中的多个天线元件)并且用于RRH 860发送和接收无线信号。如图12所示，gNB 830可以包括多个天线840。例如，多个天线840可以与gNB 830使用的多个频带兼容。虽然图12示出其中gNB 830包括多个天线840的示例，但是gNB 830也可以包括单个天线840。
基站设备850包括控制器851、存储器852、网络接口853、无线通信接口855以及连接接口857。控制器851、存储器852和网络接口853与参照图11描述的控制器821、存储器822和网络接口823相同。
无线通信接口855支持任何蜂窝通信方案(诸如LTE和LTE-先进)，并且经由RRH 860和天线840来提供到位于与RRH 860对应的扇区中的终端的无线通信。无线通信接口855通常可以包括例如BB处理器856。除了BB处理器856经由连接接口857连接到RRH 860的RF电路864之外，BB处理器856与参照图11描述的BB处理器826相同。如图12所示，无线通信接口855可以包括多个BB处理器856。例如，多个BB处理器856可以与gNB 830使用的多个频带兼容。虽然图12示出其中无线通信接口855包括多个BB处理器856的示例，但是无线通信接口855也可以包括单个BB处理器856。
连接接口857为用于将基站设备850(无线通信接口855)连接至RRH 860的接口。连接接口857还可以为用于将基站设备850(无线通信接口855)连接至RRH 860的上述高速线路中的通信的通信模块。
RRH 860包括连接接口861和无线通信接口863。
连接接口861为用于将RRH 860(无线通信接口863)连接至基站设备850的接口。 连接接口861还可以为用于上述高速线路中的通信的通信模块。
无线通信接口863经由天线840来传送和接收无线信号。无线通信接口863通常可以包括例如RF电路864。RF电路864可以包括例如混频器、滤波器和放大器，并且经由天线840来传送和接收无线信号。如图12所示，无线通信接口863可以包括多个RF电路864。例如，多个RF电路864可以支持多个天线元件。虽然图12示出其中无线通信接口863包括多个RF电路864的示例，但是无线通信接口863也可以包括单个RF电路864。
[关于终端的应用示例]
(第一应用示例)
图13是示出可以应用本公开的技术的智能电话900的示意性配置的示例的框图。智能电话900包括处理器901、存储器902、存储装置903、外部连接接口904、摄像装置906、传感器907、麦克风908、输入装置909、显示装置910、扬声器911、无线通信接口912、一个或多个天线开关915、一个或多个天线916、总线917、电池918以及辅助控制器919。
处理器901可以为例如CPU或片上系统(SoC)，并且控制智能电话900的应用层和另外层的功能。存储器902包括RAM和ROM，并且存储数据和由处理器901执行的程序。存储装置903可以包括存储介质，诸如半导体存储器和硬盘。外部连接接口904为用于将外部装置(诸如存储卡和通用串行总线(USB)装置)连接至智能电话900的接口。
摄像装置906包括图像传感器(诸如电荷耦合器件(CCD)和互补金属氧化物半导体(CMOS))，并且生成捕获图像。传感器907可以包括一组传感器，诸如测量传感器、陀螺仪传感器、地磁传感器和加速度传感器。麦克风908将输入到智能电话900的声音转换为音频信号。输入装置909包括例如被配置为检测显示装置910的屏幕上的触摸的触摸传感器、小键盘、键盘、按钮或开关，并且接收从用户输入的操作或信息。显示装置910包括屏幕(诸如液晶显示器(LCD)和有机发光二极管(OLED)显示器)，并且显示智能电话900的输出图像。扬声器911将从智能电话900输出的音频信号转换为声音。
无线通信接口912支持任何蜂窝通信方案(诸如LTE和LTE-先进)，并且执行无线通信。无线通信接口912通常可以包括例如BB处理器913和RF电路914。BB处理器913可以执行例如编码/解码、调制/解调以及复用/解复用，并且执行用于无线通信的各种类型的信号处理。同时，RF电路914可以包括例如混频器、滤波器和放大器，并且经由天线916来传送和接收无线信号。无线通信接口912可以为其上集成有BB处理器913和 RF电路914的一个芯片模块。如图13所示，无线通信接口912可以包括多个BB处理器913和多个RF电路914。虽然图13示出其中无线通信接口912包括多个BB处理器913和多个RF电路914的示例，但是无线通信接口912也可以包括单个BB处理器913或单个RF电路914。
此外，除了蜂窝通信方案之外，无线通信接口912可以支持另外类型的无线通信方案，诸如短距离无线通信方案、近场通信方案和无线局域网(LAN)方案。在此情况下，无线通信接口912可以包括针对每种无线通信方案的BB处理器913和RF电路914。
天线开关915中的每一个在包括在无线通信接口912中的多个电路(例如用于不同的无线通信方案的电路)之间切换天线916的连接目的地。
天线916中的每一个均包括单个或多个天线元件(诸如包括在MIMO天线中的多个天线元件)，并且用于无线通信接口912传送和接收无线信号。如图13所示，智能电话900可以包括多个天线916。虽然图13示出其中智能电话900包括多个天线916的示例，但是智能电话900也可以包括单个天线916。
此外，智能电话900可以包括针对每种无线通信方案的天线916。在此情况下，天线开关915可以从智能电话900的配置中省略。
总线917将处理器901、存储器902、存储装置903、外部连接接口904、摄像装置906、传感器907、麦克风908、输入装置909、显示装置910、扬声器911、无线通信接口912以及辅助控制器919彼此连接。电池918经由馈线向图13所示的智能电话900的各个块提供电力，馈线在图中被部分地示为虚线。辅助控制器919例如在睡眠模式下操作智能电话900的最小必需功能。
(第二应用示例)
图14是示出可以应用本公开的技术的汽车导航设备920的示意性配置的示例的框图。汽车导航设备920包括处理器921、存储器922、全球定位系统(GPS)模块924、传感器925、数据接口926、内容播放器927、存储介质接口928、输入装置929、显示装置930、扬声器931、无线通信接口933、一个或多个天线开关936、一个或多个天线937以及电池938。
处理器921可以为例如CPU或SoC，并且控制汽车导航设备920的导航功能和另外的功能。存储器922包括RAM和ROM，并且存储数据和由处理器921执行的程序。
GPS模块924使用从GPS卫星接收的GPS信号来测量汽车导航设备920的位置(诸如纬度、经度和高度)。传感器925可以包括一组传感器，诸如陀螺仪传感器、地磁传感 器和空气压力传感器。数据接口926经由未示出的终端而连接到例如车载网络941，并且获取由车辆生成的数据(诸如车速数据)。
内容播放器927再现存储在存储介质(诸如CD和DVD)中的内容，该存储介质被插入到存储介质接口928中。输入装置929包括例如被配置为检测显示装置930的屏幕上的触摸的触摸传感器、按钮或开关，并且接收从用户输入的操作或信息。显示装置930包括诸如LCD或OLED显示器的屏幕，并且显示导航功能的图像或再现的内容。扬声器931输出导航功能的声音或再现的内容。
无线通信接口933支持任何蜂窝通信方案(诸如LTE和LTE-先进)，并且执行无线通信。无线通信接口933通常可以包括例如BB处理器934和RF电路935。BB处理器934可以执行例如编码/解码、调制/解调以及复用/解复用，并且执行用于无线通信的各种类型的信号处理。同时，RF电路935可以包括例如混频器、滤波器和放大器，并且经由天线937来传送和接收无线信号。无线通信接口933还可以为其上集成有BB处理器934和RF电路935的一个芯片模块。如图14所示，无线通信接口933可以包括多个BB处理器934和多个RF电路935。虽然图14示出其中无线通信接口933包括多个BB处理器934和多个RF电路935的示例，但是无线通信接口933也可以包括单个BB处理器934或单个RF电路935。
此外，除了蜂窝通信方案之外，无线通信接口933可以支持另外类型的无线通信方案，诸如短距离无线通信方案、近场通信方案和无线LAN方案。在此情况下，针对每种无线通信方案，无线通信接口933可以包括BB处理器934和RF电路935。
天线开关936中的每一个在包括在无线通信接口933中的多个电路(诸如用于不同的无线通信方案的电路)之间切换天线937的连接目的地。
天线937中的每一个均包括单个或多个天线元件(诸如包括在MIMO天线中的多个天线元件)，并且用于无线通信接口933传送和接收无线信号。如图14所示，汽车导航设备920可以包括多个天线937。虽然图14示出其中汽车导航设备920包括多个天线937的示例，但是汽车导航设备920也可以包括单个天线937。
此外，汽车导航设备920可以包括针对每种无线通信方案的天线937。在此情况下，天线开关936可以从汽车导航设备920的配置中省略。
电池938经由馈线向图14所示的汽车导航设备920的各个块提供电力，馈线在图中被部分地示为虚线。电池938累积从车辆提供的电力。
本公开的技术也可以被实现为包括汽车导航设备920、车载网络941以及车辆模块 942中的一个或多个块的车载系统(或车辆)940。车辆模块942生成车辆数据(诸如车速、发动机速度和故障信息)，并且将所生成的数据输出至车载网络941。
结合本公开所述的各种示意性的块和部件可以用被设计来执行本文所述的功能的通用处理器、数字信号处理器(DSP)、ASIC、FPGA或其它可编程逻辑设备、离散门或晶体管逻辑、离散硬件部件或它们的任意组合来实现或执行。通用处理器可以是微处理器，但是可替代地，处理器可以是任何传统的处理器、控制器、微控制器和/或状态机。处理器也可以被实现为计算设备的组合，例如DSP与微处理器、多个微处理器、结合DSP核的一个或多个微处理器和/或任何其它这样的配置的组合。
本文所述的功能可以在硬件、由处理器执行的软件、固件或它们的任意组合中实现。如果在由处理器执行的软件中实现，则功能可以被存储在非暂态计算机可读介质上或者被传输作为非暂态计算机可读介质上的一个或多个指令或代码。其它示例和实现在本公开和所附权利要求的范围和精神内。例如，鉴于软件的本质，以上所述的功能可以使用由处理器执行的软件、硬件、固件、硬连线或这些中的任意的组合来执行。实现功能的特征也可以被物理地置于各种位置处，包括被分布使得功能的部分在不同物理位置处实现。
此外，包含于其它部件内的或者与其它部件分离的部件的公开应当被认为是示例性的，因为潜在地可以实现多种其它架构以达成同样的功能，包括并入全部的、大部分的、和/或一些的元件作为一个或多个单一结构或分离结构的一部分。
非暂态计算机可读介质可以是能够被通用计算机或专用计算机存取的任何可用的非暂态介质。举例而言而非限制地，非暂态计算机可读介质可以包括RAM、ROM、EEPROM、闪速存储器、CD-ROM、DVD或其它光盘存储、磁盘存储或其它磁存储设备、或能够被用来承载或存储指令或数据结构形式的期望的程序代码部件和能够被通用或专用计算机或者通用或专用处理器存取的任何其它介质。
本公开的先前描述被提供来使本领域技术人员能够制作或使用本公开。对本公开的各种修改对本领域技术人员而言是明显的，本文定义的通用原理可以在不脱离本公开的范围的情况下应用到其它变形。因此，本公开并不限于本文所述的示例和设计，而是对应于与所公开的原理和新特征一致的最宽范围。
本公开还包括如下实施方式。
1.一种用于联邦学习的网络侧电子装置，包括处理电路，所述处理电路被配置为：
确定模型聚合时间和针对用户设备的剩余服务时间Tserve，其中所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；
在所述模型聚合时间和所述剩余服务时间Tserve满足预定条件的情况下，做出针对所述用户设备的切换决定；以及
发送针对所述用户设备的所述切换决定。
2.如实施方式1所述的电子装置，其中，所述处理电路还被配置为：
接收来自所述中间节点的中间聚合模型；
至少基于所述中间聚合模型，生成全局聚合模型；以及
广播所述全局聚合模型。
3.如实施方式1所述的电子装置，其中，所述处理电路还被配置为接收所述用户设备的状态信息，并且所述模型聚合时间和所述剩余服务时间Tserve是至少基于所述用户设备的状态信息确定的。
4.如实施方式3所述的电子装置，其中，所述用户设备的状态信息包括以下各项中的一项或多项：信道状态、计算能力、本地数据信息、电量、位置和移动信息。
5.如实施方式3所述的电子装置，其中，所述处理电路还被配置为接收所述中间节点的状态信息，并且所述模型聚合时间和所述剩余服务时间Tserve是还基于所述中间节点的状态信息确定的。
6.如实施方式5所述的电子装置，其中，所述中间节点的状态信息包括以下各项中的一项或多项：信道状态、计算能力、位置和移动信息。
7.如实施方式1所述的电子装置，其中，
所述模型聚合时间包括本次全局聚合所需的剩余时间T1和下次全局聚合所需的时间T2，并且
如果T1<Tserve<T1+T2，则做出针对所述用户设备的切换决定，并且所述切换决定指示所述用户设备在本次全局聚合结束后进行切换。
8.如实施方式7所述的电子装置，其中，如果Tserve<T1，则所述处理电路还被配置为在假设执行以下各个操作中的一个或多个操作的情况下估计提高的剩余服务时间T′serve：
提高所述用户设备、所述中间节点和全局节点中的一个或多个的发送功率；
为所述用户设备和所述中间节点中的一个或二者分配更多的传输资源；以及
降低所述用户设备、所述中间节点和全局节点中的一个或多个的RSRP门限，
其中如果T′serve>T1，则执行所述一个或多个操作，并且做出针对所述用户设备的切换决定，所述切换决定指示所述用户设备在本次全局聚合结束后进行切换。
9.如实施方式1或7所述的电子装置，其中，所述模型聚合时间包括所述中间节点处的本次中间聚合所需的剩余时间t1和下次中间聚合所需的时间t2，并且
如果t1<Tserve<t1+t2，则做出针对所述用户设备的切换决定，并且所述切换决定指示所述用户设备在本次中间聚合结束后进行切换。
10.如实施方式9所述的电子装置，其中，如果Tserve<t1，则做出针对所述用户设备的切换决定，并且所述切换决定指示所述用户设备立即进行切换。
11.如实施方式9所述的电子装置，其中，如果Tserve<t1，则所述处理电路还被配置为在假设执行以下各个操作中的一个或多个操作的情况下估计提高的剩余服务时间T′serve：
提高所述用户设备、所述中间节点和全局节点中的一个或多个的发送功率；
为所述用户设备和所述中间节点中的一个或二者分配更多的传输资源；以及
降低所述用户设备、所述中间节点和全局节点中的一个或多个的RSRP门限，
其中如果T′serve>t1，则执行所述一个或多个操作，并且做出针对所述用户设备的切换决定，所述切换决定指示所述用户设备在本次中间聚合结束后进行切换。
12.如实施方式9所述的电子装置，其中，如果所述用户设备切换到另一中间节点，并且所述另一中间节点处没有中间聚合模型，则将所述中间节点处的中间聚合模型发送至所述另一中间节点。
13.一种用于联邦学习的中间节点侧电子装置，包括处理电路，所述处理电路被配置为：
从网络接收针对用户设备的切换决定，其中所述切换决定是在模型聚合时间和针对所述用户设备的剩余服务时间Tserve满足预定条件的情况下做出的，并且所述用户设备经由所述中间节点间接连接到所述网络；
将所述切换决定发送至所述用户设备。
14.如实施方式13所述的电子装置，其中，所述处理电路还被配置为：
接收来自所述用户设备的本地模型，至少基于所述本地模型生成中间聚合模型，将所述中间聚合模型发送至所述网络；以及
接收来自所述网络的全局聚合模型，将所述全局聚合模型发送至所述用户设备。
15.如实施方式13所述的电子装置，其中，所述处理电路还被配置为：
接收所述用户设备的状态信息；
将所述用户设备的状态信息和所述中间节点的状态信息发送至所述网络，
其中，所述模型聚合时间和所述剩余服务时间Tserve是至少基于所述用户设备的状态信息和所述中间节点的状态信息确定的。
16.如实施方式13所述的电子装置，其中，所述处理电路还被配置为：
接收所述用户设备的状态信息；
至少基于所述用户设备的状态信息和所述中间节点的状态信息确定所述模型聚合时间的至少一部分；
将所述用户设备的状态信息、所述中间节点的状态信息以及所述模型聚合时间的所述至少一部分发送至所述网络,
其中，所述模型聚合时间的其它部分和所述剩余服务时间Tserve是至少基于所述用户设备的状态信息和所述中间节点的状态信息确定的。
17.一种用于联邦学习的用户设备侧电子装置，包括处理电路，所述处理电路被配置为：
接收来自网络的针对用户设备的切换决定，其中所述切换决定是在模型聚合时间和针对所述用户设备的剩余服务时间Tserve满足预定条件的情况下做出的，并且所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；
基于所述切换决定进行切换。
18.如实施方式17所述的电子装置，其中，所述处理电路还被配置为：
利用本地数据训练本地模型；
将所述本地模型发送至所述网络或者所述中间节点；以及
接收来自所述网络的全局聚合模型；
将所述本地模型更新为所述全局聚合模型。
19.如实施方式17所述的电子装置，其中，所述处理电路还被配置为：
将所述用户设备的状态信息发送至所述网络或者所述中间节点，
其中，所述模型聚合时间和所述剩余服务时间Tserve是至少基于所述用户设备的状态信息确定的。
20.一种用于联邦学习的网络侧方法，包括：
确定模型聚合时间和针对用户设备的剩余服务时间Tserve，其中所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；
在所述模型聚合时间和所述剩余服务时间Tserve满足预定条件的情况下，做出针对所述用户设备的切换决定；以及
发送针对所述用户设备的所述切换决定。
21.一种用于联邦学习的中间节点侧方法，包括：
从网络接收针对用户设备的切换决定，其中所述切换决定是在模型聚合时间和针对所述用户设备的剩余服务时间Tserve满足预定条件的情况下做出的，并且所述用户设备经由所述中间节点间接连接到所述网络；
将所述切换决定发送至所述用户设备。
22.一种用于联邦学习的用户设备侧方法，包括：
接收来自网络的针对用户设备的切换决定，其中所述切换决定是在模型聚合时间和针对所述用户设备的剩余服务时间Tserve满足预定条件的情况下做出的，并且所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；
基于所述切换决定进行切换。
23.一种非暂态计算机可读存储介质，其上存储了程序指令，所述程序指令在由处理器执行时使处理器执行根据实施方式20至22中任一项所述的方法。
24.一种计算机程序产品，包括程序指令，所述程序指令在由处理器执行时使处理器执行根据实施方式20至22中任一项所述的方法。

Claims (24)

  1. 一种用于联邦学习的网络侧电子装置，包括处理电路，所述处理电路被配置为：
    确定模型聚合时间和针对用户设备的剩余服务时间Tserve，其中所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；
    在所述模型聚合时间和所述剩余服务时间Tserve满足预定条件的情况下，做出针对所述用户设备的切换决定；以及
    发送针对所述用户设备的所述切换决定。
  2. 如权利要求1所述的电子装置，其中，所述处理电路还被配置为：
    接收来自所述中间节点的中间聚合模型；
    至少基于所述中间聚合模型，生成全局聚合模型；以及
    广播所述全局聚合模型。
  3. 如权利要求1所述的电子装置，其中，所述处理电路还被配置为接收所述用户设备的状态信息，并且所述模型聚合时间和所述剩余服务时间Tserve是至少基于所述用户设备的状态信息确定的。
  4. 如权利要求3所述的电子装置，其中，所述用户设备的状态信息包括以下各项中的一项或多项：信道状态、计算能力、本地数据信息、电量、位置和移动信息。
  5. 如权利要求3所述的电子装置，其中，所述处理电路还被配置为接收所述中间节点的状态信息，并且所述模型聚合时间和所述剩余服务时间Tserve是还基于所述中间节点的状态信息确定的。
  6. 如权利要求5所述的电子装置，其中，所述中间节点的状态信息包括以下各项中的一项或多项：信道状态、计算能力、位置和移动信息。
  7. 如权利要求1所述的电子装置，其中，
    所述模型聚合时间包括本次全局聚合所需的剩余时间T1和下次全局聚合所需的时间T2，并且
    如果T1<Tserve<T1+T2，则做出针对所述用户设备的切换决定，并且所述切换决定指示所述用户设备在本次全局聚合结束后进行切换。
  8. 如权利要求7所述的电子装置，其中，如果Tserve<T1，则所述处理电路还被配置为在假设执行以下各个操作中的一个或多个操作的情况下估计提高的剩余服务 时间T′serve：
    提高所述用户设备、所述中间节点和全局节点中的一个或多个的发送功率；
    为所述用户设备和所述中间节点中的一个或二者分配更多的传输资源；以及
    降低所述用户设备、所述中间节点和全局节点中的一个或多个的RSRP门限，
    其中如果T′serve>T1，则执行所述一个或多个操作，并且做出针对所述用户设备的切换决定，所述切换决定指示所述用户设备在本次全局聚合结束后进行切换。
  9. 如权利要求1或7所述的电子装置，其中，所述模型聚合时间包括所述中间节点处的本次中间聚合所需的剩余时间t1和下次中间聚合所需的时间t2，并且
    如果t1<Tserve<t1+t2，则做出针对所述用户设备的切换决定，并且所述切换决定指示所述用户设备在本次中间聚合结束后进行切换。
  10. 如权利要求9所述的电子装置，其中，如果Tserve<t1，则做出针对所述用户设备的切换决定，并且所述切换决定指示所述用户设备立即进行切换。
  11. 如权利要求9所述的电子装置，其中，如果Tserve<t1，则所述处理电路还被配置为在假设执行以下各个操作中的一个或多个操作的情况下估计提高的剩余服务时间T′serve：
    提高所述用户设备、所述中间节点和全局节点中的一个或多个的发送功率；
    为所述用户设备和所述中间节点中的一个或二者分配更多的传输资源；以及
    降低所述用户设备、所述中间节点和全局节点中的一个或多个的RSRP门限，
    其中如果T′serve>t1，则执行所述一个或多个操作，并且做出针对所述用户设备的切换决定，所述切换决定指示所述用户设备在本次中间聚合结束后进行切换。
  12. 如权利要求9所述的电子装置，其中，如果所述用户设备切换到另一中间节点，并且所述另一中间节点处没有中间聚合模型，则将所述中间节点处的中间聚合模型发送至所述另一中间节点。
  13. 一种用于联邦学习的中间节点侧电子装置，包括处理电路，所述处理电路被配置为：
    从网络接收针对用户设备的切换决定，其中所述切换决定是在模型聚合时间和针对所述用户设备的剩余服务时间Tserve满足预定条件的情况下做出的，并且所述用户设备经由所述中间节点间接连接到所述网络；
    将所述切换决定发送至所述用户设备。
  14. 如权利要求13所述的电子装置，其中，所述处理电路还被配置为：
    接收来自所述用户设备的本地模型，至少基于所述本地模型生成中间聚合模型，将所述中间聚合模型发送至所述网络；以及
    接收来自所述网络的全局聚合模型，将所述全局聚合模型发送至所述用户设备。
  15. 如权利要求13所述的电子装置，其中，所述处理电路还被配置为：
    接收所述用户设备的状态信息；
    将所述用户设备的状态信息和所述中间节点的状态信息发送至所述网络，
    其中，所述模型聚合时间和所述剩余服务时间Tserve是至少基于所述用户设备的状态信息和所述中间节点的状态信息确定的。
  16. 如权利要求13所述的电子装置，其中，所述处理电路还被配置为：
    接收所述用户设备的状态信息；
    至少基于所述用户设备的状态信息和所述中间节点的状态信息确定所述模型聚合时间的至少一部分；
    将所述用户设备的状态信息、所述中间节点的状态信息以及所述模型聚合时间的所述至少一部分发送至所述网络,
    其中，所述模型聚合时间的其它部分和所述剩余服务时间Tserve是至少基于所述用户设备的状态信息和所述中间节点的状态信息确定的。
  17. 一种用于联邦学习的用户设备侧电子装置，包括处理电路，所述处理电路被配置为：
    接收来自网络的针对用户设备的切换决定，其中所述切换决定是在模型聚合时间和针对所述用户设备的剩余服务时间Tserve满足预定条件的情况下做出的，并且所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；
    基于所述切换决定进行切换。
  18. 如权利要求17所述的电子装置，其中，所述处理电路还被配置为：
    利用本地数据训练本地模型；
    将所述本地模型发送至所述网络或者所述中间节点；以及
    接收来自所述网络的全局聚合模型；
    将所述本地模型更新为所述全局聚合模型。
  19. 如权利要求17所述的电子装置，其中，所述处理电路还被配置为：
    将所述用户设备的状态信息发送至所述网络或者所述中间节点，
    其中，所述模型聚合时间和所述剩余服务时间Tserve是至少基于所述用户设备的 状态信息确定的。
  20. 一种用于联邦学习的网络侧方法，包括：
    确定模型聚合时间和针对用户设备的剩余服务时间Tserve，其中所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；
    在所述模型聚合时间和所述剩余服务时间Tserve满足预定条件的情况下，做出针对所述用户设备的切换决定；以及
    发送针对所述用户设备的所述切换决定。
  21. 一种用于联邦学习的中间节点侧方法，包括：
    从网络接收针对用户设备的切换决定，其中所述切换决定是在模型聚合时间和针对所述用户设备的剩余服务时间Tserve满足预定条件的情况下做出的，并且所述用户设备经由所述中间节点间接连接到所述网络；
    将所述切换决定发送至所述用户设备。
  22. 一种用于联邦学习的用户设备侧方法，包括：
    接收来自网络的针对用户设备的切换决定，其中所述切换决定是在模型聚合时间和针对所述用户设备的剩余服务时间Tserve满足预定条件的情况下做出的，并且所述用户设备直接连接到所述网络或者经由中间节点间接连接到所述网络；
    基于所述切换决定进行切换。
  23. 一种非暂态计算机可读存储介质，其上存储了程序指令，所述程序指令在由处理器执行时使处理器执行根据权利要求20至22中任一项所述的方法。
  24. 一种计算机程序产品，包括程序指令，所述程序指令在由处理器执行时使处理器执行根据权利要求20至22中任一项所述的方法。
PCT/CN2023/110473 2022-08-05 2023-08-01 用于分层联邦学习网络中的切换的装置、方法和介质 WO2024027676A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210936728.XA CN117560722A (zh) 2022-08-05 2022-08-05 分层联邦学习网络中的切换
CN202210936728.X 2022-08-05

Publications (1)

Publication Number Publication Date
WO2024027676A1 true WO2024027676A1 (zh) 2024-02-08

Family

ID=89817247

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/110473 WO2024027676A1 (zh) 2022-08-05 2023-08-01 用于分层联邦学习网络中的切换的装置、方法和介质

Country Status (2)

Country Link
CN (1) CN117560722A (zh)
WO (1) WO2024027676A1 (zh)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111836321A (zh) * 2020-07-27 2020-10-27 北京邮电大学 一种基于联邦学习和边缘计算的小区切换方法
WO2021107831A1 (en) * 2019-11-28 2021-06-03 Telefonaktiebolaget Lm Ericsson (Publ) Performing a handover procedure
WO2021238274A1 (zh) * 2020-05-28 2021-12-02 浪潮电子信息产业股份有限公司 一种分布式深度学习的梯度信息更新方法及相关装置
CN114363911A (zh) * 2021-12-31 2022-04-15 哈尔滨工业大学（深圳） 一种部署分层联邦学习的无线通信系统及资源优化方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021107831A1 (en) * 2019-11-28 2021-06-03 Telefonaktiebolaget Lm Ericsson (Publ) Performing a handover procedure
WO2021238274A1 (zh) * 2020-05-28 2021-12-02 浪潮电子信息产业股份有限公司 一种分布式深度学习的梯度信息更新方法及相关装置
CN111836321A (zh) * 2020-07-27 2020-10-27 北京邮电大学 一种基于联邦学习和边缘计算的小区切换方法
CN114363911A (zh) * 2021-12-31 2022-04-15 哈尔滨工业大学（深圳） 一种部署分层联邦学习的无线通信系统及资源优化方法

Also Published As

Publication number Publication date
CN117560722A (zh) 2024-02-13

Similar Documents

Publication Publication Date Title
US11284290B2 (en) Terminal device, communication control device, base station, gateway device, control device, method, and recording medium
US11757510B2 (en) Electronic devices and communication methods
WO2019128795A1 (zh) 无线通信系统中的装置和方法、计算机可读存储介质
WO2018171580A1 (zh) 用于无线通信的电子装置以及无线通信方法
US9867140B2 (en) Random access techniques for fixed devices in mobile broadband networks
CN108605198B (zh) 电子装置、信息处理设备和信息处理方法
KR102480438B1 (ko) 전자 디바이스 및 라디오 통신 방법
US20180124857A1 (en) Radio access network device, data processing method, and ip packet processing method
JP6855838B2 (ja) 端末装置、方法及びプログラム
WO2018223983A1 (zh) 频谱管理装置和方法、频谱协调装置和方法以及电子设备
US20230385688A1 (en) Electronic device and method for federated learning
WO2021197269A1 (zh) 一种临近服务的数据传输方法、设备及系统
JP6848440B2 (ja) 端末装置、基地局装置、方法及び記録媒体
WO2017104281A1 (ja) 装置、方法及びプログラム
US20240073848A1 (en) Network Slice in a Wireless Network
WO2017167082A1 (zh) 电子设备和用于电子设备的方法、信息处理设备
WO2024027676A1 (zh) 用于分层联邦学习网络中的切换的装置、方法和介质
WO2016002332A1 (ja) 装置、方法及びプログラム
WO2021164675A1 (zh) 电子设备、无线通信方法和计算机可读存储介质
CN113938985A (zh) 一种通信方法及装置
WO2024083034A1 (zh) 用于无线通信的电子设备和方法、计算机可读存储介质
WO2022083505A1 (zh) 电子设备和通信方法
WO2022205630A1 (zh) 电子设备、通信方法以及计算机可读存储介质
WO2023208043A1 (zh) 用于无线通信系统的电子设备、方法和存储介质
WO2021043150A1 (zh) 用于无线通信的电子设备和方法、计算机可读存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23849383

Country of ref document: EP

Kind code of ref document: A1