Unmanned aerial vehicle aided training method, device, equipment, medium and program product

Abstract

The invention relates to the technical field of Internet of Things, and provides an unmanned aerial vehicle aided training method, device and equipment, a medium and a program product. The method comprises the following steps: receiving global model parameters sent by a federated learning server, and sending the global model parameters to a base station cell; receiving a local model parameter sent by the base station cell, and sending the local model parameter to a federated learning server, so that the federated learning server updates a global model based on the local model parameter; wherein the local model parameters are obtained by performing local training based on the global model parameters and local data collected by participants of the base station cell. In this way, the unmanned aerial vehicle group can effectively participate in federated learning, and the problem that the model training effect is poor due to insufficient unmanned aerial vehicle computing resources is avoided.

Description

Technical Field

[0001] The present application relates to the field of Internet of Things technology, and in particular to a method, device, equipment, medium and program product for drone-assisted training. Background Art

[0002] As an advanced open integrated network that integrates communication, computing, intelligence, storage and learning functions, the Internet of Vehicles can build a seamless information exchange platform between users, vehicles and roadside units (RSU). This interconnected ecosystem not only catalyzes the birth of a series of transportation applications, but also significantly improves the overall efficiency of the transportation system.

[0003] Faced with the massive amount of dynamic data emerging in the Internet of Vehicles, artificial intelligence technology has been applied to the Internet of Vehicles scenario to build data-driven models, analyze and process data, and then realize intelligent traffic prediction and generate route optimization suggestions. However, traditional model training and learning methods often rely on centrally transmitting data to a central server for model training. This approach has gradually revealed its limitations against the backdrop of a surge in IoT devices and explosive growth in data volume: the server's storage and computing resources are difficult to effectively cope with massive data, thus restricting the efficiency of model training.

[0004] The federated learning algorithm utilizes the computing and storage capabilities of IoT devices and can directly conduct model training on the device side. The Internet of Vehicles service provider (i.e., the owner of the model) provides a federated learning server and assembles a team of IoT devices such as parked vehicles and RSUs as participants in federated learning (FL). Participants can collect local data from the surrounding environment and become key nodes for data collection and processing. At the beginning of each training cycle of the model, participants can receive the global model parameters of the federated learning server, and then start local model training on their own unique local data sets. After completing local training, participants will feed back the updated local model parameters to the federated learning server, and the federated learning server will perform model aggregation based on the local model parameters of each participant, thereby updating the global model. The federated learning server can then send the global model parameters of the updated global model to each participant again, starting a new round of local model training iterations. This process is repeated until the global model of the federated learning server reaches the preset accuracy standard.

[0005] In the prior art, when drones participate in federated learning, they usually act as participating nodes or federated learning servers. If drones, parked vehicles, RSUs, etc. all act as participating nodes in federated learning and directly participate in the training of the local model, this requires the drones to have sufficient computing resources to execute the model training task, which easily increases the design complexity and production cost of the drones. In addition, since the drones need to directly participate in the local model training for a long time, this will reduce the flexibility of their execution of other tasks. If the drones are used as federated learning servers, a large amount of data needs to be aggregated on the drones, which will cause insufficient computing resources of the drones. At the same time, the drones are easily affected by factors such as weather, battery power, and signal interference, which easily leads to the interruption of time-consuming training tasks, and then leads to poor training effects of the global model.

[0006] Therefore, in the prior art, it is difficult for drones to effectively participate in federated learning, which easily leads to poor training effects of the model. Summary of the Invention

[0007] The embodiments of the present application provide a method, device, equipment, medium and program product for drone-assisted training to solve the technical problem that it is difficult for drones to effectively participate in federated learning and easily leads to poor training effects of the model.

[0008] In a first aspect, the embodiments of the present application provide a method for drone-assisted training, which is applied to a group of drones. The method includes: receiving global model parameters sent by a federated learning server and sending the global model parameters to a base station cell; receiving local model parameters sent by the base station cell and sending the local model parameters to the federated learning server, so that the federated learning server updates the global model based on the local model parameters; wherein, the local model parameters are obtained by local training based on the global model parameters and local data collected by participants in the base station cell.

[0009] In one embodiment, the group of drones includes multiple drones, and the number of base station cells is at least one; sending the global model parameters to the base station cell includes: constructing an initial set of drone coalition partitions; the initial set of drone coalition partitions includes multiple drone coalition groups, and each drone coalition group includes at least one drone; determining the preference degree of each base station cell for each drone coalition group; a preference degree is determined according to the importance and training duration of the corresponding base station cell, and the importance is determined according to the data sampling quality of the participants in the corresponding base station cell; based on the preference degree of each base station cell for each drone coalition group, according to the second-price auction principle, determining the target base station cell of each drone coalition group; based on each drone coalition group, sending the global model parameters to the target base station cell corresponding to each drone coalition group.

[0010] In one embodiment, based on the preference degree of each base station cell for each UAV coalition group, according to the second-price auction principle, determining the target base station cell for each UAV coalition group includes: based on the preference degree of each base station cell for each UAV coalition group, according to the second-price auction principle, determining the revenue of each UAV coalition group; determining the service cost of each UAV coalition group; based on the revenue of each UAV coalition group and the service cost of each UAV coalition group, determining the maximum profit of each UAV coalition group and the target base station cell of each UAV coalition group; the target base station cell is the base station cell that pays the revenue to the corresponding UAV coalition group so that the corresponding UAV coalition group obtains the maximum profit.

[0011] In one embodiment, after determining the maximum profit of each UAV coalition group and the target base station cell of each UAV coalition group, it further includes: based on the maximum profit of each UAV coalition group, determining the total system profit; the total system profit is the sum of the maximum profits of all UAV coalition groups, and each maximum profit is determined based on the energy loss of the corresponding UAV coalition group and the valuation of the corresponding target base station cell, and the valuation is determined based on the preference degree of the corresponding target base station cell for the corresponding UAV coalition group; based on each UAV coalition group, performing coalition combination to obtain an updated first UAV coalition partition set; the first UAV coalition partition set includes multiple updated first UAV coalition groups, and each first UAV coalition group includes at least one UAV; determining the first total system profit of the first UAV coalition partition set; judging whether the first total system profit is greater than the total system profit; if the first total system profit is greater than the total system profit, then based on each first UAV coalition group, respectively sending the global model parameters to the target base station cells corresponding to each first UAV coalition group; if the first total system profit is less than or equal to the total system profit, then based on each UAV coalition group, respectively sending the global model parameters to the target base station cells corresponding to each UAV coalition group.

[0012] In one embodiment, after determining the maximum profit of each UAV coalition group and the target base station cell of each UAV coalition group, it further includes: determining the total system profit based on the maximum profit of each UAV coalition group; the total system profit is the sum of the maximum profits of all UAV coalition groups, and each maximum profit is determined based on the energy consumption of the corresponding UAV coalition group and the valuation of the corresponding target base station cell, and the valuation is determined based on the preference degree of the corresponding target base station cell for the corresponding UAV coalition group; performing coalition group splitting based on each UAV coalition group to obtain an updated second UAV coalition division set; the second UAV coalition division set includes multiple updated second UAV coalition groups, and each second UAV coalition group includes at least one UAV; determining the second total system profit of the second UAV coalition division set; determining whether the second total system profit is greater than the total system profit; if the second total system profit is greater than the total system profit, then based on each second UAV coalition group, sending the global model parameters to the target base station cell corresponding to each second UAV coalition group respectively; if the second total system profit is less than or equal to the total system profit, then based on each UAV coalition group, sending the global model parameters to the target base station cell corresponding to each UAV coalition group respectively.

[0013] In one embodiment, receiving the local model parameters sent by the base station cell and sending the local model parameters to the federated learning server includes: based on each UAV coalition group, respectively receiving the local model parameters sent by each target base station cell and sending the local model parameters sent by each target base station cell to the federated learning server.

[0014] In a second aspect, an apparatus for UAV-assisted training provided by an embodiment of the present application includes: a first transmission module, configured to receive the global model parameters sent by the federated learning server and send the global model parameters to the base station cell; a second transmission module, configured to receive the local model parameters sent by the base station cell and send the local model parameters to the federated learning server, so that the federated learning server updates the global model based on the local model parameters; wherein, the local model parameters are obtained through local training based on the global model parameters and the local data collected by the participants of the base station cell.

[0015] In a third aspect, an electronic device provided by an embodiment of the present application includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the method for UAV-assisted training as described in any of the above.

[0016] In a fourth aspect, a non-transitory computer-readable storage medium provided by an embodiment of the present application stores a computer program, and when the computer program is executed by a processor, it implements the method for UAV-assisted training as described in any of the above.

[0017] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program which, when executed by a processor, implements the method for drone-assisted training as described in any of the above.

[0018] In the method, apparatus, device, medium and program product for drone-assisted training provided by the embodiments of the present application, in the vehicle-to-everything (V2X) federated learning scenario, a drone group is used as a relay between a federated learning server and participants in a base station cell, responsible for distributing the global model parameters of the federated learning server to the base station cell. Thus, the participants in the base station cell can perform local model training based on the global model parameters and the locally collected data, update to obtain local model parameters, and send the received local model parameters to the federated learning server, so that the federated learning server can update the global model based on the local model parameters. In this way, the drone group can effectively participate in federated learning without acting as either a federated learning participant or a federated learning server, which can not only avoid an increase in the design complexity and production cost of drones, but also improve the flexibility of drones to perform other tasks. At the same time, since the drone group only needs to transmit a small amount of model parameters between the federated learning server and the participants in the base station cell and does not need to directly perform local model or global model training, the problem of poor model training effect caused by insufficient computing resources of drones can be avoided. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0020] Figure 1 It is a schematic flowchart of the method for drone-assisted training provided by an embodiment of the present application.

[0021] Figure 2 It is a schematic structural diagram of the apparatus for drone-assisted training provided by an embodiment of the present application.

[0022] Figure 3 It is a schematic structural diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following will clearly and completely describe the technical solutions in this application in conjunction with the accompanying drawings in the embodiments of this application. Apparently, the described embodiments are some, but not all, of the embodiments of this application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts fall within the scope of protection of this application.

[0024] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of the method for drone-assisted training provided by the embodiments of this application. As Figure 1 shown, in the embodiments of this application, the method for drone-assisted training is applied to a drone group. The method for drone-assisted training includes steps S110 to S120, and the specific steps are as follows: S110: Receive the global model parameters sent by the federated learning server and send the global model parameters to the base station cell.

[0025] S120: Receive the local model parameters sent by the base station cell and send the local model parameters to the federated learning server so that the federated learning server updates the global model based on the local model parameters.

[0026] Among them, the local model parameters are obtained through local training based on the global model parameters and the local data collected by the participants in the base station cell.

[0027] Optionally, the number of base station cells is at least one.

[0028] Considering factors such as link failure and node resource unavailability, this embodiment introduces a drone group as a relay between the federated learning server and the participants in the base station cell to improve the communication efficiency between the server and each participant in the federated learning scenario.

[0029] Specifically, in the vehicle-to-everything (V2X) federated learning scenario, a federated learning server for federated learning can be set at the network edge. The federated learning server can publish crowdsourcing tasks to the participants in each base station cell, generate global model parameters, and send them to the drone group.

[0030] After receiving the global model parameters sent by the federated learning server, the drone group sends the global model parameters to each base station cell respectively; the participants in each base station cell are responsible for collecting and labeling local data, and performing local training based on the global model parameters and the collected local data to obtain local model parameters; after the local training is completed, each base station cell can upload the local model parameters of each participant to the drone group, and the drone group can send the aggregated local model parameters of the participants in each base station cell to the federated learning server.

[0031] Further, the federated learning server can aggregate the local model parameters of each participant to update the global model.

[0032] In the method for drone-assisted training provided by the embodiments of this application, in the vehicle-to-everything (V2X) federated learning scenario, a drone group is used as a relay between the federated learning server and the participants in the base station cell, responsible for sending the global model parameters of the federated learning server to the base station cell. As a result, the participants in the base station cell can perform local model training based on the global model parameters and the collected local data, update to obtain the local model parameters, and send the received local model parameters to the federated learning server, so that the federated learning server can update the global model based on the local model parameters. In this way, the drone group can effectively participate in federated learning without being used as a federated learning participant or a federated learning server, which can not only avoid the increase in the design complexity and production cost of the drone, but also improve the flexibility of the drone to perform other tasks. At the same time, since the drone group only needs to transmit a small amount of model parameters between the federated learning server and the participants in the base station cell and does not need to directly perform local model or global model training, the problem of poor model training effect caused by insufficient computing resources of the drone can be avoided.

[0033] In some embodiments, the drone group includes multiple drones, and the number of base station cells is at least one. Sending the global model parameters to the base station cell includes: constructing an initial drone coalition partition set; the initial drone coalition partition set includes multiple drone coalition groups, and each drone coalition group includes at least one drone; determining the preference degree of each base station cell for each drone coalition group; a preference degree is determined according to the importance and training duration of the corresponding base station cell, and the importance is determined according to the data sampling quality of the participants in the corresponding base station cell; based on the preference degree of each base station cell for each drone coalition group, according to the second-price auction principle, determining the target base station cell of each drone coalition group; based on each drone coalition group, sending the global model parameters to the target base station cell corresponding to each drone coalition group respectively.

[0034] In some embodiments, based on the preference degree of each base station cell for each drone coalition group, according to the second-price auction principle, determining the target base station cell of each drone coalition group includes: based on the preference degree of each base station cell for each drone coalition group, according to the second-price auction principle, determining the income of each drone coalition group; determining the service cost of each drone coalition group; based on the income of each drone coalition group and the service cost of each drone coalition group, determining the maximum profit of each drone coalition group and the target base station cell of each drone coalition group; the target base station cell is the base station cell that pays the income to the corresponding drone coalition group so that the corresponding drone coalition group obtains the maximum profit.

[0035] Specifically, assume there are I heterogeneous wireless base station cells, and each base station cell has participants involved in the local model training process of federated learning. The participants can be RSUs, moving vehicles, Internet of Things devices, etc.

[0036] For each base station cell, since the data sampling rates and data sampling qualities of each participant in this base station cell are different, the individual importance of each participant in this base station cell will be different due to the differences in data sampling rates and data sampling qualities. There may be a problem of insufficient data during the training process of federated learning. Participants with low data sampling rates and low data sampling qualities will also have lower individual importance in this base station cell; each participant can upload its data sampling rate to the federated learning server through the drone group, and the federated learning server can calculate the individual importance of each participant based on this.

[0037] Optionally, the individual importance of a participant in the base station cell is defined as a concave function of the data sampling rate. Then the individual importance of the participant in the base station cell is calculated as follows: ; where represents the data sampling rate of the participant in the base station cell .

[0038] The importance of a base station cell is determined according to the data sampling qualities of its participants. Therefore, based on the above formula, the importance of the base station cell is defined as the sum of the individual importances of each participant in this base station cell. Then the importance of the base station cell is calculated as follows: ; where represents the total number of participants in the base station cell .

[0039] Assume that M drones are deployed between the I base station cells and the federated learning server (i.e., the drone group includes M drones) to facilitate the training process of federated learning. After each participant in each base station cell completes local training, it sends its local model parameters to the drone group through the corresponding base station cell, and the drones in the drone group can aggregate the local model parameters and forward them to the federated learning server.

[0040] For each drone in the drone group, the energy loss generated when the drone performs a federated learning task includes flight energy loss, computing energy loss, communication energy loss, hovering energy loss, and circuit energy loss. The calculation methods for various energy losses are as follows: (1) Flight energy loss: Assume that drone m flies from the warehouse to the base station cell and assists in reporting the local model parameters. Then, the flight energy loss of drone m depends on the location of the base station cell and the drone warehouse. Therefore, the flight energy loss of drone m is calculated as follows: ; Among them, represents the flight speed of drone m; represents the energy required for drone m to fly once per unit time; represents the flight energy weight of the base station cell ; represents the flight energy weight of the storage warehouse where drone m is located; represents the distance between the storage warehouse where drone m is located and the base station cell .

[0041] Among them, the flight energy weight depends on the terrain of the base station cell and the drone warehouse.

[0042] (2) Computing energy loss: The computing energy loss is the loss generated when drone m completes a single iteration task (i.e., aggregates the local model parameters of the participants in the base station cell it serves). Therefore, the computing energy loss of drone m for completing a single iteration task is calculated as follows: ; Among them, is the number of CPU cycles required for drone m to complete the aggregation of local model parameters; is the computing capacity of drone m; is a constant coefficient.

[0043] (3) Communication energy loss: The communication energy loss is formed due to the parameter transmission between different entities in the system model (i.e., the participants in the base station cell in federated learning, drones, and the federated learning server).

[0044] Assume that the communication mode between entities is OFDMA (Orthogonal Frequency Division Multiple Access). The energy required for the UAV m to transmit the local model parameters in any base station cell to the federated learning server in each iteration via the wireless channel is: ; where represents the total data volume of the aggregated local model parameters; is the transmission rate when the UAV m sends the aggregated local model parameters to the federated learning server; represents the transmission power of the UAV m.

[0045] Meanwhile, the UAV needs energy to receive the global model parameters of the federated learning server. The energy consumption generated when the UAV m receives the global model parameters of the federated learning server once is: ; where is the data volume of the global model parameters; is the downlink transmission rate from the federated learning server to the UAV m; is the receiving power of the UAV m.

[0046] In addition, the UAV also needs energy to receive the local model parameters from the federated learning participants in the base station cell. The energy consumption generated when the UAV m receives the local model parameters of the federated learning participants in the base station cell once is: ; where is the receiving power of the UAV m; represents the total number of participants in the base station cell; represents the data volume of the local model parameters of the accepted participants; represents the transmission rate for the participants.

[0047] Similarly, the UAV also needs energy to send the global model parameters of the federated learning server to the federated learning participants in the base station cell. The energy consumption generated when the UAV m sends the global model parameters to the federated learning participants in the base station cell once is: ; where represents the transmission power of the UAV m; represents the base station cell The total number of participants; The amount of data for the global model parameters; Indicates the transmission rate when the drone m sends the global model parameters to the participants.

[0048] The above four types of energy losses are all communication energy losses.

[0049] (4) In order to keep the drone m hovering above the base station cell to assist in the federated learning training, the drone m will also generate hovering energy and circuit energy losses .

[0050] After the federated learning training in the base station cell completes the iterative task, the federated learning server will pay corresponding rewards. The amount of the reward is related to the importance of the base station cell and the task completion time. Therefore, in order for the base station cell to obtain a higher reward, it is necessary for the drone to improve the communication efficiency, thereby promoting its assistance in completing the training. Considering that some drones may not have enough energy resources to support a base station cell to complete the entire federated learning process, therefore, this embodiment considers completing it by establishing a drone coalition.

[0051] Specifically, this embodiment will design an auction algorithm according to the determined drone coalition division set to solve the allocation problem between each drone coalition group and each base station cell. It is set that each drone coalition group is an auctioneer that conducts an auction and provides communication resources to promote the federated learning training process, and each base station cell is defined as a buyer that submits a bid and pays a reward to the drone coalition group.

[0052] Each base station cell sets a preference degree for the drone coalition group that can provide services to itself. The preference degree is related to the importance of the base station cell and the completion time of the federated learning iterative task. Therefore, the preference degree can be determined according to the importance and training duration of the corresponding base station cell. The setting of the preference relationship is to quantify the preferences of different base station cells for different drone coalition groups in different scenarios. For example, for road safety applications, the IoT terminal is more inclined to select a drone coalition group that is closer, but for entertainment services, the IoT terminal is more inclined to select a drone coalition group that can provide higher bandwidth. Therefore, the parameter weights can be reasonably adjusted for different scenarios to adjust the preference degree.

[0053] Each base station cell will only bid for the drone coalition group that can support it to complete the federated learning iterative task (that is, satisfy , is the maximum number of iterations that the drone m in the drone coalition group can complete, is the total number of iterations specified by the federated learning server). In this embodiment, the base station cell for a certain drone coalition group The valuation of for this UAV coalition group preference degree , and the calculation formula of the preference degree is as follows: ; Among them, , are weight factors; represents the unit price paid by the federated learning server to the base station cell ; is the flight time for UAV m to reach the base station cell ; is the importance of the base station cell .

[0054] In order to encourage the participants of the base station cell to submit real valuations, this embodiment adopts the second-price auction mechanism to determine the actual price paid by the winning bidder (i.e., the winning base station cell). Suppose represents the price paid by the base station cell to the UAV coalition group after winning the bid (i.e., the income of the UAV coalition group ), so the utility of the base station cell after winning the bid is defined as . Only when the UAV coalition group is allocated to the base station cell (i.e., the target base station cell of the UAV coalition group is the base station cell ) does it have a real utility, otherwise the utility is .

[0055] As the auctioneer, the goal of the UAV coalition group is to maximize individual profit. For the UAV coalition group , the income it obtains is the actual price paid by the base station cell , and its service cost is the energy loss generated when the base station cell provides the federated learning task. The calculation formula of is as follows: Among them, is the number of iterations supported by UAV m in the base station cell ; is the cost coefficient of each unit of energy consumed by UAV m; is for multiple UAVs to form a UAV coalition group The collaborative cost generated when. It should be noted that not every drone in the drone coalition group necessarily supports the maximum number of iterations. If the remaining number of iterations is less than the maximum number of iterations that the drone can support, the drone only needs to support the remaining number of iterations.

[0056] It should be noted that when multiple drones form a drone coalition group When Each drone within the drone coalition group needs to cooperate with each other, which will generate a collaborative cost : .

[0057] The collaborative cost is related to the size of the drone coalition group and is generated by the communication between the drones within the drone coalition group. When two or more drones cooperate to form a drone coalition group to jointly support the federated learning training in the base station cell As long as the nearest drone reaches the base station cell , the federated learning process will start. The drone closest to the target base station cell in the drone coalition group aggregates the local model parameters uploaded by the participants in the target base station cell and transmits the aggregated local model parameters to the federated learning server, and completes the maximum number of iterations under the given energy capacity of the drone; then, the second closest drone in the same drone coalition group takes over, continues to aggregate the local model parameters, and sends the aggregated local model parameters to the federated learning server; similarly, after the second closest drone completes the task, the third closest drone takes over and continues this process. If the maximum number of iterations that the drone can support is greater than the remaining number of iterations in the federated learning training process, the drone only needs to support the remaining number of iterations.

[0058] For fairness, since each drone in the drone coalition group completes a different number of iterations, the profit obtained by each drone in the drone coalition group depends on the number of iterations it completes. Each drone m in the drone coalition group The maximum number of iterations that can be supported to facilitate the federated learning training process in the base station cell is: ; where is the total energy possessed by the drone m in the drone coalition group .

[0059] Therefore, the profit obtained by the drone coalition group assisting the base station cell to complete the federated learning training is: .

[0060] The total system profit obtained by all drone coalition groups within the system is: ; wherein, represents the allocation vector of the drone coalition group and the base station cell . If the value of is 1, it means that the drone coalition group is allocated to the base station cell . As can be seen from the above formula, the total system profit is the sum of the maximum profits of all drone coalition groups. Each maximum profit is determined based on the energy consumption of the corresponding drone coalition group and the valuation of the corresponding target base station cell, and the valuation is determined based on the preference degree of the corresponding target base station cell for the corresponding drone coalition group.

[0061] Based on the above definitions, the allocation algorithm process between the drone coalition group and the base station cell is as follows: Step 1: Construct an initial set of drone coalition partitions , and the initial set of drone coalition partitions includes multiple drone coalition groups (i.e., to ), each drone coalition group includes at least one drone, and let the total system profit , , be the total number of base station cells.

[0062] Step 2: Each base station cell that needs to participate in the federated learning training process calculates its preference degree for each drone coalition group according to the preference degree calculation formula, and submits the corresponding preference degree as a bid for the drone coalition group that can support it to complete the iterative training.

[0063] Step 3: Each drone coalition group that has received a bid determines its obtained income and the base station cells that provide bids according to the second-price auction principle.

[0064] It can be understood that the income of the drone coalition group can be determined according to the preference degree of the base station cell for this drone coalition group : Since the preference degree is the bid submitted by the base station cell, if the preference degrees of multiple base station cells for this drone coalition group are different, in order to ensure the maximization of the individual profit of the drone, the highest preference degree can be used as the income of the drone coalition group ; If some base station cells have the same preference for this UAV coalition group , each base station cell can continue to bid on the basis of the highest preference, and take the final maximum preference as the income of the UAV coalition group . At this time, , will be greater than the initial bids submitted by each base station cell.

[0065] Step 4: Use the calculation formula to calculate the service cost for the UAV coalition group to provide services to the corresponding target base station cell.

[0066] Step 5: Use the calculation formula to calculate the profit that the UAV coalition group can obtain on the base station cells that offer bids, and further calculate the maximum profit of the UAV coalition group and the corresponding target base station cell of the UAV coalition group .

[0067] Step 6: Determine whether there is a situation where multiple UAV coalition groups compete for the same base station cell; if so, assign the UAV coalition group with a higher valuation to the base station cell; if not, obtain the allocation relationship between the UAV coalition group and the corresponding target base station cell according to Step 5.

[0068] Step 7: Delete the allocated target base station cell and the UAV coalition group from the initial UAV coalition division set, let (the equal sign here represents assignment), and at the same time add the maximum profit obtained by the UAV coalition group to assist the corresponding target base station cell to complete the federated learning training to the current total system profit to obtain the updated total system profit.

[0069] Step 8: Determine whether is equal to 0; if it is 0, output the sum of the maximum profits of all UAV coalition groups in the system (i.e., the total system profit) and the allocation scheme of each UAV coalition group and the corresponding target base station cell; if it is not 0, return to Step 2 until all base station cells and all UAV coalition groups are allocated.

[0070] Furthermore, for each UAV coalition group after the allocation is completed, the UAV coalition group can send the global model parameters to its corresponding target base station cell.

[0071] An embodiment of the present application proposes a method for assisted training based on drone collaboration in a vehicle-to-everything (V2X) federated learning scenario. The drone group acts as a relay between the federated learning server and the model training participants in the base station cell, responsible for sending the global model parameters updated by the edge server to each base station cell for local model updates in each base station cell. At the same time, it aggregates the local model parameters after each base station cell completes iteration and forwards them to the edge server for global model parameter training. Since the drones only act as communication relays, it reduces the exposure of data during transmission, which helps improve data security and user privacy protection. At the same time, it allows for flexible adjustment and optimization of algorithms on the edge server without modifying the software of the drones or other terminal devices, making system upgrade and maintenance more convenient. In addition, by quantifying the data sampling quality and training duration of the base station cell to characterize the preference degree of the base station cell for the drone coalition group, the allocation problem of the drone coalition group is characterized as an auction problem between the drone coalition group and the base station cell. At the same time, in order to motivate the drone coalition group to complete the federated training task, the embodiment of the present application adopts the second-price auction mechanism to achieve efficient allocation between the drone coalition group and the base station cell.

[0072] In some embodiments, after determining the maximum profit of each drone coalition group and the target base station cell of each drone coalition group, it further includes: based on the maximum profit of each drone coalition group, determining the total system profit; the total system profit is the sum of the maximum profits of all drone coalition groups, and each maximum profit is determined based on the energy consumption of the corresponding drone coalition group and the valuation of the corresponding target base station cell, and the valuation is determined based on the preference degree of the corresponding target base station cell for the corresponding drone coalition group; based on each drone coalition group, performing coalition combination to obtain an updated first drone coalition partition set; the first drone coalition partition set includes multiple updated first drone coalition groups, and each first drone coalition group includes at least one drone; determining the first total system profit of the first drone coalition partition set; judging whether the first total system profit is greater than the total system profit; if the first total system profit is greater than the total system profit, then based on each first drone coalition group, sending the global model parameters to the target base station cell corresponding to each first drone coalition group respectively; if the first total system profit is less than or equal to the total system profit, then based on each drone coalition group, sending the global model parameters to the target base station cell corresponding to each drone coalition group respectively.

[0073] Understandably, if two independent drone coalition groups cooperate to form a larger drone coalition group, and it can be ensured that the profit obtained by the combined larger drone coalition group is greater than the profits when the two drone coalition groups operate independently, then such a cooperative effect can be considered additive. Similarly, two drone coalition groups can be combined to jointly serve a base station cell. When the price paid by the base station cell remains unchanged, if the service cost of the combined drone coalition group (such as energy loss and cooperation cost) increases, then the effect of this combination is negative. Therefore, to maximize the overall interests of the drone coalition group, this embodiment proposes a mechanism for merging and splitting drone coalition groups.

[0074] The specific process of the drone coalition group merging mechanism is as follows: Step 9: Initialize the coalition partition set , where is the maximum number of possible combinations of drone coalitions.

[0075] Step 10: Calculate the total system profit that can be obtained under the current combination partition and the allocation scheme between each target base station cell and the drone coalition group.

[0076] Among them, the current combination partition can be the allocation scheme of all current target base station cells and all drone coalition groups obtained in Step 8.

[0077] Step 11: Design the coalition merging mechanism. Under the current combination partition, for each drone coalition group , consider merging with other drone coalition groups to generate an updated first drone coalition partition set .

[0078] Step 12: Calculate the first total system profit of the new first drone coalition partition set , and determine the allocation scheme between each base station cell and the drone coalition in the first drone coalition partition set .

[0079] Step 13: Judge the profit change of the merged first drone coalition partition set; if , then officially merge the drone coalition groups participating in the merger to form a new drone coalition group, and update the current total system profit (i.e., ) and the new drone coalition partition set (i.e., ).

[0080] Step 14: Judge the profit change of the merged first drone coalition partition set; if , the current UAV coalition division set is maintained.

[0081] Step 15: Repeat Steps 10 to 14 to continue the next iteration. In the given current UAV coalition division set any UAV coalition group will consider forming a new UAV coalition group jointly with another UAV coalition group . After considering all possible groupings, the merging mechanism terminates. At the end of the merging mechanism, the algorithm returns the final grouping result .

[0082] In some embodiments, after determining the maximum profit of each UAV coalition group and the target base station cell of each UAV coalition group, it further includes: determining the total system profit based on the maximum profit of each UAV coalition group; the total system profit is the sum of the maximum profits of all UAV coalition groups, and each maximum profit is determined based on the energy loss of the corresponding UAV coalition group and the valuation of the corresponding target base station cell, and the valuation is determined based on the preference degree of the corresponding target base station cell for the corresponding UAV coalition group; performing coalition group splitting based on each UAV coalition group to obtain an updated second UAV coalition division set; the second UAV coalition division set includes multiple updated second UAV coalition groups, and each second UAV coalition group includes at least one UAV; determining the second total system profit of the second UAV coalition division set; judging whether the second total system profit is greater than the total system profit; if the second total system profit is greater than the total system profit, then based on each second UAV coalition group, sending the global model parameters to the target base station cells corresponding to each second UAV coalition group respectively; if the second total system profit is less than or equal to the total system profit, then based on each UAV coalition group, sending the global model parameters to the target base station cells corresponding to each UAV coalition group respectively.

[0083] Similarly, the specific process of the UAV coalition group splitting mechanism is as follows: Step 16: Design of the coalition splitting mechanism. For ease of understanding, based on Step 15 here, from the perspective of splitting the UAV coalition groups , the initialized UAV coalition group is the set of UAV coalition groups that may split .

[0084] Step 17: For all possible splitting cases of the UAV coalition group , similar to the merging mechanism, in order to evaluate the total system profit, assume that after the coalition group splitting, all UAVs are finally divided into the second UAV coalition division set , and calculate the second UAV coalition partition set of the second total system profit .

[0085] Step 18: If the original total system profit can be improved by more coalition splits, i.e., satisfying , then the original UAV coalition group will be officially split into , forming a new UAV coalition group (i.e., ), update the current total system profit (i.e., ) and the new UAV coalition partition set (i.e., ).

[0086] Step 19: If the original total system profit cannot be improved by more coalition splits, i.e., satisfying , then keep the current UAV coalition partition set.

[0087] Step 20: Given the new partition , repeat Steps 16 to 19. Any UAV coalition group will consider splitting and re-evaluating the total system profit. If the total system profit of the current grouping is greater than that after splitting, then the current UAV coalition partition set will be maintained without further splitting. When all possible coalition splits have been considered, the splitting mechanism terminates.

[0088] Step 21: At the end of the merge and split algorithms, return the final UAV coalition partition set .

[0089] Optionally, after completing the allocation of the initial UAV coalition partition set, the merge and split algorithms may not be executed, or only the merge or split algorithm may be executed, or the merge and split algorithms may be executed alternately. If the merge or split algorithm needs to be executed, then according to the final UAV coalition partition set obtained after executing the merge or split algorithm, services can be provided for each base station cell.

[0090] The method for UAV-assisted training provided in the embodiments of the present application designs a scheme for establishing coalitions based on multiple UAVs to jointly cooperate to assist base station cells in completing the update of model parameters; at the same time, in order to adapt to the dynamically changing training requirements of base station cells, the embodiments of the present application design a coalition merge and split mechanism, and decide whether to perform inter-coalition mergers or split the coalition into finer-grained coalition groups according to the energy consumption of the UAV coalition group and the valuation of the base station cell. The embodiments of the present application consider the situation where a single UAV may lack the ability to stay stably in the air throughout the federated learning training process, and use the mutual cooperation between UAVs to jointly complete the federated learning training, which can maximize the utilization of UAV resources.

[0091] In some embodiments, receiving the local model parameters sent by the receiving base station cell and sending the local model parameters to the federated learning server includes: based on each drone coalition group, respectively receiving the local model parameters sent by each target base station cell and sending the local model parameters sent by each target base station cell to the federated learning server.

[0092] An embodiment of the present application also provides a device for drone-assisted training. Please refer to Figure 2 , Figure 2 FIG. is a schematic structural diagram of the device for drone-assisted training provided by the embodiment of the present application. In the embodiment of the present application, the device for drone-assisted training includes a first transmission module 210 and a second transmission module 220.

[0093] The first transmission module 210 is configured to receive the global model parameters sent by the federated learning server and send the global model parameters to the base station cell.

[0094] The second transmission module 220 is configured to receive the local model parameters sent by the base station cell and send the local model parameters to the federated learning server, so that the federated learning server updates the global model based on the local model parameters.

[0095] Among them, the local model parameters are obtained by local training based on the global model parameters and the local data collected by the participants of the base station cell.

[0096] In some embodiments, the drone group includes multiple drones, and the number of base station cells is at least one.

[0097] The first transmission module 210 is configured to construct an initial drone coalition division set; the initial drone coalition division set includes multiple drone coalition groups, and each drone coalition group includes at least one drone; determine the preference degree of each base station cell for each drone coalition group; a preference degree is determined according to the importance and training duration of the corresponding base station cell, and the importance is determined according to the data sampling quality of the participants of the corresponding base station cell; based on the preference degree of each base station cell for each drone coalition group, according to the second-price auction principle, determine the target base station cell of each drone coalition group; based on each drone coalition group, send the global model parameters to the target base station cell corresponding to each drone coalition group respectively.

[0098] In some embodiments, the first transmission module 210 is configured to determine the revenue of each drone coalition group based on the preference degree of each base station cell for each drone coalition group according to the second-price auction principle; determine the service cost of each drone coalition group; based on the revenue of each drone coalition group and the service cost of each drone coalition group, determine the maximum profit of each drone coalition group and the target base station cell of each drone coalition group; the target base station cell is the base station cell that pays revenue to the corresponding drone coalition group so that the corresponding drone coalition group obtains the maximum profit.

[0099] In some embodiments, the first transmission module 210 is configured to determine the total system profit based on the maximum profit of each drone coalition group; the total system profit is the sum of the maximum profits of all drone coalition groups, and each maximum profit is determined based on the energy consumption of the corresponding drone coalition group and the valuation of the corresponding target base station cell, and the valuation is determined based on the preference degree of the corresponding target base station cell for the corresponding drone coalition group; based on each drone coalition group, perform coalition combination to obtain an updated first drone coalition partition set; the first drone coalition partition set includes a plurality of updated first drone coalition groups, and each first drone coalition group includes at least one drone; determine the first total system profit of the first drone coalition partition set; determine whether the first total system profit is greater than the total system profit; if the first total system profit is greater than the total system profit, then based on each first drone coalition group, send the global model parameters to the target base station cells corresponding to each first drone coalition group respectively; if the first total system profit is less than or equal to the total system profit, then based on each drone coalition group, send the global model parameters to the target base station cells corresponding to each drone coalition group respectively.

[0100] In some embodiments, the first transmission module 210 is configured to determine the total system profit based on the maximum profit of each UAV coalition group; the total system profit is the sum of the maximum profits of all UAV coalition groups, and each maximum profit is determined based on the energy consumption of the corresponding UAV coalition group and the valuation of the corresponding target base station cell, and the valuation is determined based on the preference degree of the corresponding target base station cell for the corresponding UAV coalition group; perform coalition group splitting based on each UAV coalition group to obtain an updated second UAV coalition division set; the second UAV coalition division set includes multiple updated second UAV coalition groups, and each second UAV coalition group includes at least one UAV; determine the second total system profit of the second UAV coalition division set; determine whether the second total system profit is greater than the total system profit; if the second total system profit is greater than the total system profit, then based on each second UAV coalition group, send the global model parameters to the target base station cell corresponding to each second UAV coalition group respectively; if the second total system profit is less than or equal to the total system profit, then based on each UAV coalition group, send the global model parameters to the target base station cell corresponding to each UAV coalition group respectively.

[0101] In some embodiments, the second transmission module 220 is configured to receive the local model parameters sent by each target base station cell based on each UAV coalition group, and send the local model parameters sent by each target base station cell to the federated learning server.

[0102] Embodiments of the present application further provide an electronic device. Figure 3 It is a schematic structural diagram of the electronic device provided by the embodiments of the present application. As Figure 3 shown, the electronic device may include: a processor 310, a communication interface 320, a memory 330, and a communication bus 340. Among them, the processor 310, the communication interface 320, and the memory 330 complete mutual communication through the communication bus 340. The processor 310 can call the logical instructions in the memory 330 to execute the method for UAV-assisted training.

[0103] In addition, when the logical instructions in the above-mentioned memory 330 can be implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0104] The embodiments of the present application also provide a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the method for drone-assisted training provided by the above-mentioned various methods.

[0105] The embodiments of the present application also provide a computer program product. The computer program product includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the method for drone-assisted training provided by the above-mentioned various methods.

[0106] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative labor.

[0107] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the technical solution, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disks, optical discs, etc., and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for drone-assisted training, characterized in that: Applied to a group of drones, the method comprises: Receiving global model parameters sent by the federated learning server, and sending the global model parameters to the base station cell; Receiving local model parameters sent by the base station cell, and sending the local model parameters to the federated learning server, so that the federated learning server updates the global model based on the local model parameters; The local model parameters are obtained by local training based on the global model parameters and local data collected by participants in the base station cell.

2. The method for drone-assisted training according to claim 1, characterized in that: The drone group includes a plurality of drones, and the number of the base station cells is at least one; The sending the global model parameters to a base station cell includes: Constructing an initial drone alliance partition set; the initial drone alliance partition set includes a plurality of drone alliance groups, each of which includes at least one drone; Determine the preference of each base station cell to each drone alliance group; the preference is determined according to the importance and training duration of the corresponding base station cell, and the importance is determined according to the data sampling quality of the participants of the corresponding base station cell; Based on the preference of each base station cell for each drone alliance group, determine the target base station cell of each drone alliance group according to the second price auction principle; Based on each of the drone alliance groups, the global model parameters are respectively sent to the target base station cells corresponding to each of the drone alliance groups.

3. The method for drone-assisted training according to claim 2, characterized in that: Based on the preference of each base station cell for each drone alliance group, according to the second price auction principle, determining the target base station cell of each drone alliance group includes: Based on the preference of each base station cell for each drone alliance group, the income of each drone alliance group is determined according to the second price auction principle; determining a service cost for each of said drone alliance groups; Based on the income of each drone alliance group and the service cost of each drone alliance group, the maximum profit of each drone alliance group and the target base station cell of each drone alliance group are determined; the target base station cell is the base station cell that pays income to the corresponding drone alliance group so that the corresponding drone alliance group obtains the maximum profit.

4. The method for drone-assisted training according to claim 3, characterized in that: After determining the maximum profit of each drone alliance group and the target base station cell of each drone alliance group, the method further includes: Based on the maximum profit of each drone alliance group, determine the total profit of the system; the total profit of the system is the sum of the maximum profits of all the drone alliance groups, each of the maximum profits is determined based on the energy loss of the corresponding drone alliance group and the valuation of the corresponding target base station cell, and the valuation is determined based on the preference of the corresponding target base station cell for the corresponding drone alliance group; Based on each of the drone alliance groups, alliance groups are merged to obtain an updated first drone alliance partition set; the first drone alliance partition set includes multiple updated first drone alliance groups, and each of the first drone alliance groups includes at least one drone; Determine a first system total profit of the first drone alliance partition set; Determining whether the total profit of the first system is greater than the total profit of the system; If the total profit of the first system is greater than the total profit of the system, based on each of the first drone alliance groups, the global model parameters are sent to the target base station cells corresponding to each of the first drone alliance groups respectively; If the total profit of the first system is less than or equal to the total profit of the system, then based on each of the drone alliance groups, the global model parameters are sent to the target base station cells corresponding to each of the drone alliance groups respectively.

5. The method for drone-assisted training according to claim 3, characterized in that: After determining the maximum profit of each drone alliance group and the target base station cell of each drone alliance group, the method further includes: Based on the maximum profit of each drone alliance group, determine the total profit of the system; the total profit of the system is the sum of the maximum profits of all the drone alliance groups, each of the maximum profits is determined based on the energy loss of the corresponding drone alliance group and the valuation of the corresponding target base station cell, and the valuation is determined based on the preference of the corresponding target base station cell for the corresponding drone alliance group; Based on each of the drone alliance groups, the alliance group is split to obtain an updated second drone alliance partition set; the second drone alliance partition set includes a plurality of updated second drone alliance groups, and each of the second drone alliance groups includes at least one drone; Determine a second system total profit of the second drone alliance partition set; Determining whether the second system total profit is greater than the system total profit; If the second system total profit is greater than the system total profit, based on each of the second drone alliance groups, the global model parameters are sent to the target base station cells corresponding to each of the second drone alliance groups respectively; If the second system total profit is less than or equal to the system total profit, then based on each of the drone alliance groups, the global model parameters are sent to the target base station cells corresponding to each of the drone alliance groups respectively.

6. The method for drone-assisted training according to claim 2, characterized in that: The receiving the local model parameters sent by the base station cell, and sending the local model parameters to the federated learning server, includes: Based on each of the drone alliance groups, local model parameters sent by each of the target base station cells are received respectively, and the local model parameters sent by each of the target base station cells are sent to the federated learning server.

7. A drone-assisted training device, characterized in that: include: A first transmission module, configured to receive global model parameters sent by a federated learning server, and send the global model parameters to a base station cell; A second transmission module is used to receive the local model parameters sent by the base station cell, and send the local model parameters to the federated learning server, so that the federated learning server updates the global model based on the local model parameters; The local model parameters are obtained by local training based on the global model parameters and local data collected by participants in the base station cell.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the method for drone-assisted training according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for drone-assisted training as claimed in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for drone-assisted training as claimed in any one of claims 1 to 6 is implemented.

Images