CN117289725B

CN117289725B - Unmanned plane distributed general calculation integrated resource scheduling method and device

Info

Publication number: CN117289725B
Application number: CN202311590741.5A
Authority: CN
Inventors: 杜军; 江炳青; 侯向往; 张华蕾; 王劲涛
Original assignee: Tsinghua University
Current assignee: Tsinghua University
Priority date: 2023-11-27
Filing date: 2023-11-27
Publication date: 2024-02-27
Anticipated expiration: 2043-11-27
Also published as: CN117289725A

Abstract

The application relates to an unmanned aerial vehicle distributed general calculation integrated resource scheduling method, an unmanned aerial vehicle distributed general calculation integrated resource scheduling device, computer equipment and a storage medium. The method comprises the following steps: transmitting global model parameters to each follower unmanned aerial vehicle, and receiving Euclidean norms, total energy consumption and energy virtual queues of the local model parameters fed back by each follower unmanned aerial vehicle; aiming at each follower unmanned aerial vehicle, determining the scheduling probability of the follower unmanned aerial vehicle based on Euclidean norms, total energy consumption and energy virtual queues of local model parameters corresponding to the follower unmanned aerial vehicle; determining target follower unmanned aerial vehicles in the follower unmanned aerial vehicles according to the scheduling probability of the follower unmanned aerial vehicles, and training a preset global model based on the target follower unmanned aerial vehicles to obtain a trained global model; and under the condition that the trained global model meets the preset training stop condition, determining the trained global model as an unmanned aerial vehicle scheduling model. By adopting the method, the unmanned aerial vehicle dispatching efficiency can be improved.

Description

Unmanned plane distributed general calculation integrated resource scheduling method and device

Technical Field

The application relates to the technical field of unmanned aerial vehicles, in particular to an unmanned aerial vehicle distributed general calculation integrated resource scheduling method, an unmanned aerial vehicle distributed general calculation integrated resource scheduling device, computer equipment and a storage medium.

Background

Unmanned aerial vehicles benefit from characteristics such as formation nature, dynamic nature, mobility, have advantages such as deployment is quick, the action is controllable, the network is nimble, have important effect in all fields. In particular, the unmanned aerial vehicle or the unmanned aerial vehicle cluster executes the federal learning task based on the unmanned aerial vehicle distributed general calculation integrated resource scheduling method. Wherein, unmanned aerial vehicle cluster contains a pilot unmanned aerial vehicle and a plurality of follower unmanned aerial vehicles.

In the existing unmanned aerial vehicle distributed general calculation integrated resource scheduling method, a scheduled follower unmanned aerial vehicle transmits local update parameters to a pilot unmanned aerial vehicle. The pilot unmanned aerial vehicle trains the global model according to the local updating parameters of the scheduled follower unmanned aerial vehicle. And under the condition that the global model meets the preset training stop condition, the pilot unmanned aerial vehicle determines the global model as an unmanned aerial vehicle scheduling model. The pilot unmanned aerial vehicle schedules each follower unmanned aerial vehicle to execute the federal learning task based on the unmanned aerial vehicle scheduling model.

However, in the existing unmanned aerial vehicle distributed general calculation integrated resource scheduling method, as the attribute of each follower unmanned aerial vehicle is different, the follower unmanned aerial vehicle is scheduled according to the same probability, so that the aggregated aggregate model parameters deviate, and the unmanned aerial vehicle scheduling efficiency is poor.

Disclosure of Invention

In view of the foregoing, it is desirable to provide a method, an apparatus, a computer device, and a computer-readable storage medium for scheduling resources for distributed computing integration of unmanned aerial vehicles.

In a first aspect, the present application provides a method for scheduling distributed general calculation integrated resources of an unmanned aerial vehicle, where the method is applied to a pilot unmanned aerial vehicle, and the method includes:

transmitting global model parameters to each follower unmanned aerial vehicle, and receiving Euclidean norms, total energy consumption and energy virtual queues of the local model parameters fed back by each follower unmanned aerial vehicle;

determining, for each follower unmanned aerial vehicle, a scheduling probability of the follower unmanned aerial vehicle based on the euclidean norm of the local model parameter, the total energy consumption and the energy virtual queue corresponding to the follower unmanned aerial vehicle;

determining target follower unmanned aerial vehicles in the follower unmanned aerial vehicles according to the scheduling probability of the follower unmanned aerial vehicles, and training a preset global model based on the target follower unmanned aerial vehicles to obtain a trained global model;

under the condition that the trained global model meets a preset training stop condition, determining the trained global model as an unmanned plane scheduling model; the unmanned aerial vehicle scheduling model is used for scheduling the follower unmanned aerial vehicle to execute tasks.

In one embodiment, the determining, for each of the follower unmanned aerial vehicles, the scheduling probability of the follower unmanned aerial vehicle based on the euclidean norm of the local model parameter, the total energy consumption, and the energy virtual queue corresponding to the follower unmanned aerial vehicle includes:

determining a data set duty cycle of the follower unmanned aerial vehicle for each follower unmanned aerial vehicle; the data set duty ratio is the proportion of the local data set of the follower unmanned aerial vehicle and all data sets of all the follower unmanned aerial vehicles;

and carrying out data processing on the Euclidean norm of the local model parameter, the total energy consumption, the energy virtual queue and the data set duty ratio according to a preset probability algorithm to obtain the scheduling probability of the follower unmanned aerial vehicle.

In one embodiment, the determining, in each follower unmanned aerial vehicle, the target follower unmanned aerial vehicle according to the scheduling probability of each follower unmanned aerial vehicle includes:

determining a target follower unmanned aerial vehicle in each follower unmanned aerial vehicle according to a preset non-replacement sampling algorithm and scheduling probability of each follower unmanned aerial vehicle;

and updating the scheduling probability of each target follower unmanned aerial vehicle to be zero.

In one embodiment, the training a preset global model based on each target follower unmanned aerial vehicle to obtain a trained global model includes:

receiving local model parameters of each target follower unmanned aerial vehicle through a sight channel; the local model parameters of each target follower unmanned aerial vehicle are aggregated in the sight channel to obtain aggregated model parameters;

and updating parameters of a preset global model based on the aggregate model parameters, the transmission noise and the power scaling factor to obtain a trained global model.

In one embodiment, after the training a preset global model based on each target follower unmanned aerial vehicle to obtain a trained global model, the method further includes:

judging whether the trained global model meets a preset training stopping condition or not;

and under the condition that the trained global model does not meet the training stop condition, executing the step of sending global model parameters to each follower unmanned aerial vehicle until the trained global model meets the training stop condition.

In a second aspect, the present application provides a method for scheduling distributed general calculation integrated resources of an unmanned aerial vehicle, where the method is applied to a follower unmanned aerial vehicle, and the method includes:

Receiving global model parameters sent by a pilot unmanned aerial vehicle, and executing a preset random gradient descent algorithm according to the global model parameters to obtain Euclidean norms of local model parameters;

determining the total energy consumption of the follower unmanned aerial vehicle according to the transmission attribute of the follower unmanned aerial vehicle and the property of the central processing unit;

determining an energy virtual queue based on the total energy constraint condition of the follower unmanned aerial vehicle;

transmitting the euclidean norms of the local model parameters, the total energy consumption, and the virtual queue of energy to the pilot drone.

In one embodiment, the determining the total energy consumption of the follower unmanned aerial vehicle according to the transmission attribute of the follower unmanned aerial vehicle and the property of the central processor includes:

determining the corresponding calculated energy consumption of the follower unmanned aerial vehicle according to the property of the central processor of the follower unmanned aerial vehicle;

determining the communication energy consumption corresponding to the follower unmanned aerial vehicle according to the transmission power and the power scaling factor of the follower unmanned aerial vehicle;

and determining the total energy consumption corresponding to the follower unmanned aerial vehicle according to the calculated energy consumption and the communication energy consumption.

In a third aspect, the present application further provides an unmanned aerial vehicle distributed general calculation integrated resource scheduling device, the device is applied to a pilot unmanned aerial vehicle, the device includes:

The receiving module is used for sending global model parameters to each follower unmanned aerial vehicle and receiving Euclidean norms, total energy consumption and energy virtual queues of the local model parameters fed back by each follower unmanned aerial vehicle;

the first determining module is used for determining scheduling probability of each follower unmanned aerial vehicle based on Euclidean norms, total energy consumption and the energy virtual queues of the local model parameters corresponding to the follower unmanned aerial vehicle;

the training module is used for determining target follower unmanned aerial vehicles in the follower unmanned aerial vehicles according to the scheduling probability of the follower unmanned aerial vehicles, and training a preset global model based on the target follower unmanned aerial vehicles to obtain a trained global model;

the second determining module is used for determining the trained global model as an unmanned plane scheduling model under the condition that the trained global model meets a preset training stopping condition; the unmanned aerial vehicle scheduling model is used for scheduling the follower unmanned aerial vehicle to execute tasks.

In a fourth aspect, the present application further provides an unmanned aerial vehicle distributed general calculation integrated resource scheduling device, the device is applied to a follower unmanned aerial vehicle, and the device includes:

The execution module is used for receiving global model parameters sent by the pilot unmanned aerial vehicle, and executing a preset random gradient descent algorithm according to the global model parameters to obtain Euclidean norms of the local model parameters;

a third determining module, configured to determine total energy consumption of the follower unmanned aerial vehicle according to a transmission attribute of the follower unmanned aerial vehicle and a property of a central processor;

a fourth determining module, configured to determine an energy virtual queue based on a total energy constraint condition of the follower unmanned aerial vehicle;

and the transmission module is used for transmitting the local model parameters, the total energy consumption and the energy virtual queue to the pilot unmanned aerial vehicle.

In a fifth aspect, the present application also provides a computer device comprising a memory and a processor, the memory storing a computer program, the processor implementing the following steps when executing the computer program:

In a sixth aspect, the present application further provides a computer device comprising a memory and a processor, the memory storing a computer program, the processor implementing the following steps when executing the computer program:

In a seventh aspect, the present application also provides a computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of:

In an eighth aspect, the present application also provides a computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of:

The unmanned aerial vehicle distributed general calculation integrated resource scheduling method, the unmanned aerial vehicle distributed general calculation integrated resource scheduling device, the computer equipment and the storage medium send global model parameters to all the unmanned aerial vehicles of the followers, and receive Euclidean norms, total energy consumption and energy virtual queues of the local model parameters fed back by the unmanned aerial vehicles of the followers; determining, for each follower unmanned aerial vehicle, a scheduling probability of the follower unmanned aerial vehicle based on the euclidean norm of the local model parameter, the total energy consumption and the energy virtual queue corresponding to the follower unmanned aerial vehicle; determining target follower unmanned aerial vehicles in the follower unmanned aerial vehicles according to the scheduling probability of the follower unmanned aerial vehicles, and training a preset global model based on the target follower unmanned aerial vehicles to obtain a trained global model; under the condition that the trained global model meets a preset training stop condition, determining the trained global model as an unmanned plane scheduling model; the unmanned aerial vehicle scheduling model is used for scheduling the follower unmanned aerial vehicle to execute tasks. By adopting the method, the scheduling probability of the follower unmanned aerial vehicle is determined through the Euclidean norm, the total energy consumption and the energy virtual queue of the local model parameters corresponding to the follower unmanned aerial vehicle, and then the target follower unmanned aerial vehicle with better performance is obtained by screening the follower unmanned aerial vehicles based on the scheduling probability. According to the target follower man-machine training global model with better performance, the unmanned aerial vehicle scheduling model is obtained, and unmanned aerial vehicle scheduling efficiency is improved.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the related art, the drawings that are required to be used in the embodiments or the related technical descriptions will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and other drawings may be obtained according to the drawings without inventive effort for a person having ordinary skill in the art.

Fig. 1 is an application environment diagram of a distributed general calculation integrated resource scheduling method of an unmanned aerial vehicle in one embodiment;

fig. 2 is a flow chart of a method for a pilot unmanned aerial vehicle to perform unmanned aerial vehicle distributed general calculation integrated resource scheduling in an embodiment;

FIG. 3 is a flow chart illustrating the steps of determining scheduling probabilities in one embodiment;

FIG. 4 is a flow chart illustrating steps for determining a target follower drone in one embodiment;

FIG. 5 is a flow diagram of the steps for training a global model in one embodiment;

FIG. 6 is a flowchart illustrating steps for determining whether the global model satisfies a training stop condition in one embodiment;

fig. 7 is a flow chart of a method for executing a distributed general calculation integrated resource scheduling method of a unmanned aerial vehicle by a follower unmanned aerial vehicle in an embodiment;

FIG. 8 is a flow chart illustrating the steps of determining total energy consumption in one embodiment;

fig. 9 is a flow chart of a method for scheduling distributed general calculation integrated resources of a drone in another embodiment;

fig. 10 is a block diagram of a distributed general calculation integrated resource scheduling device of a drone in one embodiment;

fig. 11 is a block diagram of a distributed general calculation integrated resource scheduling device of a drone in another embodiment;

fig. 12 is an internal structural diagram of a computer device in one embodiment.

Detailed Description

In order to make the objects, technical solutions and advantages of the present application more apparent, the present application will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the present application.

The unmanned aerial vehicle distributed general calculation integrated resource scheduling method provided by the embodiment of the application can be applied to an unmanned aerial vehicle scheduling cluster 100 shown in fig. 1. Wherein the drone dispatch cluster 100 includes one pilot drone 110 and a plurality of follower drones 120. The pilot drone 110 communicates with each follower drone 120, transmitting data.

In an exemplary embodiment, as shown in fig. 2, a method for scheduling distributed computing integrated resources of an unmanned aerial vehicle is provided, and the method is applied to the pilot unmanned aerial vehicle 110 (the pilot unmanned aerial vehicle is omitted from the description below) in fig. 1, and includes the following steps 202 to 208. Wherein:

Step 202, sending global model parameters to each follower unmanned aerial vehicle, and receiving Euclidean norms, total energy consumption and energy virtual queues of the local model parameters fed back by each follower unmanned aerial vehicle.

The total energy consumption comprises calculation energy consumption and communication energy consumption, and represents the energy consumption of the follower unmanned aerial vehicle in the process of executing the federal learning task. The virtual energy queue characterizes the gap between the accumulated total energy consumption and the energy constraint to the current training round.

In implementation, the pilot unmanned aerial vehicle broadcasts global model parameters to all the follower unmanned aerial vehicles, so that each follower unmanned aerial vehicle performs local model training based on the global model parameters to obtain Euclidean norms of the local model parameters. Each follower unmanned aerial vehicle determines the total energy consumption and the energy virtual queue corresponding to the follower unmanned aerial vehicle based on the attribute information of the follower unmanned aerial vehicle. Each follower drone transmits the euclidean norm, the total energy consumption, and the virtual queue of energy of the local model parameters of that follower drone to the pilot drone through a control channel. The pilot unmanned aerial vehicle receives Euclidean norms, total energy consumption and energy virtual queues of local model parameters fed back by each follower unmanned aerial vehicle through a control channel.

Optionally, the global model parameters are determined according to the requirements of the task, and the embodiment of the application does not limit the global model parameters.

Step 204, for each follower unmanned aerial vehicle, determining a scheduling probability of the follower unmanned aerial vehicle based on the euclidean norm, the total energy consumption and the energy virtual queue of the local model parameters corresponding to the follower unmanned aerial vehicle.

In implementation, a probability algorithm is preset in the pilot unmanned aerial vehicle. Aiming at each follower unmanned aerial vehicle, the pilot unmanned aerial vehicle carries out data processing on Euclidean norms, total energy consumption and energy virtual queues of local model parameters corresponding to the follower unmanned aerial vehicle according to a preset probability algorithm, and the scheduling probability of the follower unmanned aerial vehicle is obtained.

Step 206, determining target follower unmanned aerial vehicles in the follower unmanned aerial vehicles according to the scheduling probability of the follower unmanned aerial vehicles, and training a preset global model based on the target follower unmanned aerial vehicles to obtain a trained global model.

In implementation, the pilot drone is preset with a scheduling number. And determining target follower unmanned aerial vehicles conforming to the dispatching quantity in each follower unmanned aerial vehicle according to the dispatching probability of each follower unmanned aerial vehicle by the pilot unmanned aerial vehicle. And then, the pilot unmanned aerial vehicle trains a preset global model based on local model parameters of each target follower unmanned aerial vehicle, and a trained global model is obtained. And then, the pilot unmanned aerial vehicle judges whether the trained global model meets the preset training stop condition.

Optionally, the scheduling number is determined according to the training requirement of the global model and the number of the follower unmanned aerial vehicles, and the embodiment of the application does not limit the scheduling number.

Alternatively, the training stop condition may be that the training round reaches a preset training round threshold, or the global model converges.

And step 208, determining the trained global model as the unmanned plane scheduling model under the condition that the trained global model meets the preset training stop condition.

The unmanned aerial vehicle scheduling model is used for scheduling the follower unmanned aerial vehicle to execute tasks.

In implementation, the pilot unmanned aerial vehicle determines the trained global model as the unmanned aerial vehicle scheduling model under the condition that the trained global model meets a preset training stop condition. And the pilot unmanned aerial vehicle schedules the follower unmanned aerial vehicle to execute tasks according to the unmanned aerial vehicle scheduling model. The follower unmanned aerial vehicle can be a target follower unmanned aerial vehicle in the training process, and can also be other follower unmanned aerial vehicles. The number of scheduled follower drones is not limited.

According to the unmanned aerial vehicle distributed general calculation integrated resource scheduling method, the scheduling probability of the unmanned aerial vehicle of the follower is determined through the Euclidean norm, the total energy consumption and the energy virtual queue of the local model parameters corresponding to the unmanned aerial vehicle of the follower, and then the unmanned aerial vehicle of each follower is screened based on the scheduling probability, so that the unmanned aerial vehicle of the target follower with better performance is obtained. According to the target follower man-machine training global model with better performance, the unmanned aerial vehicle scheduling model is obtained, and unmanned aerial vehicle scheduling efficiency is improved.

In an exemplary embodiment, as shown in FIG. 3, the specific process of step 204 includes steps 302 through 304. Wherein:

step 302, for each follower drone, determining a data set duty cycle of the follower drone.

Wherein the data set duty cycle is the ratio of the local data set of the follower unmanned aerial vehicle to the total data set of all the follower unmanned aerial vehicles. The data set duty ratio characterizes unbalance of data distribution among the follower unmanned aerial vehicles, and the data set duty ratio characterizes that the local model parameters obtained by the training of the follower unmanned aerial vehicles with larger data scale proportion have larger contribution to the global model training.

In the implementation, the pilot unmanned aerial vehicle performs addition processing on the data volume of the local data set of each follower unmanned aerial vehicle to obtain the data volume of all the data sets of all the follower unmanned aerial vehicles. Then, the pilot unmanned aerial vehicle carries out proportional operation on the data quantity of the local data set and the data quantity of all the data sets of each follower unmanned aerial vehicle to obtain the data set duty ratio of the follower unmanned aerial vehicle.

In an exemplary embodiment, use ofIndicate->Personal follower unmanned plane->Indicate->A local data set of the individual follower drones. And the pilot unmanned aerial vehicle performs addition operation on the data volume of the local data sets of all the follower unmanned aerial vehicles according to a preset data set aggregation algorithm to obtain the data volume of all the data sets of all the follower unmanned aerial vehicles. The data set aggregation algorithm is as follows equation (1):

（1）

Wherein, in the above formula (1),is a natural number greater than 0, +.>Represents the nth follower drone, < ->Representing the local data set of the nth follower drone, D representing the data volume of the complete data set,/->Indicate->Person to->The modulo of the data volume of the local data set of the individual follower drones is added.

And step 304, carrying out data processing on Euclidean norms, total energy consumption, energy virtual queues and data set duty ratios of the local model parameters according to a preset probability algorithm to obtain the scheduling probability of the follower unmanned aerial vehicle.

In implementation, a probability algorithm is preset in the pilot unmanned aerial vehicle. And the pilot unmanned aerial vehicle calculates Euclidean norms, total energy consumption, energy virtual queues and data set duty ratios of the local model parameters of the follower unmanned aerial vehicle according to a probability algorithm, and the scheduling probability of the follower unmanned aerial vehicle is obtained.

Wherein, the probability algorithm is shown in the following formula (2):

（2）

in the above formula (2), n represents the nth follower unmanned aerial vehicle, the training round is indexed using T e {1,2, …, T } which is the communication round, and T is also the training round.Euclidean norms of local model parameters for follower drone- >For the data set duty cycle +.>For energy virtual queues, +.>Representation assurance->The optimal value of the lagrangian multiplier of (c) can be obtained by a one-dimensional search algorithm. />Representing total energy consumption->Is a weight coefficient. />The probability of scheduling the follower unmanned plane at the t-th round is provided.

In an exemplary embodiment, the present application explores a completely new metric to evaluate the contribution of local updates to global model training. For ease of further analysis, assume a loss functionFor L-smoothness, i.eAnd->The following formula (3) is satisfied:

（3）

wherein, in the above formula (3),and->Respectively is a function->At->Point and->The value of the point is set to be,is a function->Transpose of gradient, ++>Is->Point and->Distance between points, ++>Is a function->At the point->About vector->The first two terms of the second order taylor expansion of (c).

Based on the formula (3), the convergence performance of the federal learning system based on air calculation under a probability scheduling mechanism is determined. Specifically: the device schedule set is a set of target follower drones. Given an optimal global learning modelAnd device schedule set->The convergence speed of the t-th round is shown in the following formula (4):

（4）

wherein,，/>learning rate for global model in training round t, < > >Is the global optimum, < >>Global penalty value for training round t+1,>global penalty value for training round t, < >>The square of the euclidean norm of the aggregate model parameter is trained for the t-th round. />Indicating the L-slip constant. />For the scheduling probability of the nth follower unmanned aerial vehicle trained on the t th round, N is the total number of the follower unmanned aerial vehicles, and +.>Data set duty cycle for nth follower drone, +.>The square of euclidean norms of the local model parameters of the nth follower drone in the t-th round of training. />Square of default power scaling factor, +.>Is the square of the power scaling factor in the t-th training. />The expected gap between the global loss value and the optimal value is t+1, and the optimal value is bounded by four factors. First item->Representing the expected difference of the global loss value from the optimal value for the t-round, while the second term +.>And third item->The fourth term ++relating to the second norm of the aggregate model parameter (true global gradient)>Variance from aggregate model parameters->Related to the following. Further observing the formula (4), the first three terms are irrelevant to the scheduling strategy and can be used as constants; while the variance of the global aggregation gradient depends on the scheduling design and needs to be optimized. Notably, the true global gradient indicates the direction in which the global loss function drops most rapidly. Thus, the smaller the variance of the global aggregation gradient, the faster the global loss function decreases. In summary, equation (4) establishes a link between the gradient variance and the model convergence speed, thus introducing the concept of gradient variance, i.e. >The method measures the difference between the local model parameters (local gradient) and the aggregate model parameters (real global gradient), provides an effective measurement method for the importance of local update, and the smaller gradient difference has larger contribution to the convergence of the global model.

In this application, the primary goal of federal learning is to achieve the desired training performance while minimizing energy. Based on the convergence analysis result given by the formula (3) and consideration of minimized energy consumption, the method evaluates the importance of local model parameter (local gradient) update through analysis of gradient gap, and simultaneously evaluates the energy consumption by considering channel conditions and unmanned aerial vehicle computing power. In order to improve training efficiency and minimize energy consumption, a follower unmanned aerial vehicle with excellent channel conditions, strong computing power and small gradient difference needs to be selected to participate in training. Thus, by introducing weight coefficientsTo balance the two factors of gradient gap and energy consumption. Specifically, & gt>Is determined by considering the relative magnitude relationship of the two components so that they reach a similar scale. Furthermore, since the real global gradient is inaccessible during training and is independent of the device scheduling policy, the +_ in the gradient gap is ignored >An item. The device scheduling problem may be expressed as formula set (5).

（5）

Wherein, in the formula group (5),and (3) representing the scheduling probability from the 1 st follower unmanned aerial vehicle to the N th follower unmanned aerial vehicle, wherein N is the number of the follower unmanned aerial vehicles. />Representing the minimum. />Virtual energy queues for the nth follower drone in the t-th training. />Euclidean norms for the local model parameters of the nth follower drone in the t-th round of training, +.>The data set duty cycle for the nth follower drone. />Is a weight coefficient. />For the scheduling probability of the nth follower drone in the t-th training round, +.>Representing the total energy consumption of the nth follower drone in the t-th round of training. T is the total training round. />Indicating whether the follower drone N (N e N) is involved in the gradient upload of the present round, ++>And the energy consumption of the nth unmanned aerial vehicle is represented. Since the formula set (5) is a convex function, the algorithm of the scheduling probability of the follower unmanned aerial vehicle is obtained through KKT (Kuhn-Tucker conditions, coulomb) condition solution, namely the formula (2).

The above equation (2) shows that the optimal scheduling probability mainly depends on three factors, namely the local data duty ratioLocal gradient (local model parameters) >Euclidean norm and total energy consumption +.>. Specifically, the data set duty ratio (local data duty ratio) reflects the imbalance of data distribution among the follower unmanned aerial vehicles, and the local model parameters (local gradients) obtained by the training of the follower unmanned aerial vehicles with larger data scale proportion have larger contribution to the global model training. In order to accelerate model convergence, it is desirable to schedule follower drones with larger gradient norms as much as possible. Because of the variability in channel conditions and computing power between follower drones, it is reasonable to assign a lower probability to a drone with greater energy consumption. Therefore, the pilot unmanned aerial vehicle can obtain reasonable scheduling probability distribution by comprehensively considering the update importance, the channel condition and the computing power.

In the embodiment, the scheduling probability of the follower unmanned aerial vehicle is determined through the Euclidean norm of the local model parameters, the total energy consumption, the energy virtual queue and the data set duty ratio, so that the relation between the heterogeneity and the energy consumption of the follower unmanned aerial vehicle and the scheduling probability is clarified, the efficiency of training the global model is improved, and the energy efficiency of unmanned aerial vehicle scheduling is improved.

In an exemplary embodiment, as shown in fig. 4, the specific process of determining the target follower drone in each follower drone according to the scheduling probability of each follower drone in step 206 includes steps 402 to 404. Wherein:

Step 402, determining a target follower unmanned aerial vehicle in each follower unmanned aerial vehicle according to a preset non-return sampling algorithm and scheduling probability of each follower unmanned aerial vehicle.

In implementation, a pilot unmanned aerial vehicle is preset with a non-return sampling algorithm. The pilot unmanned aerial vehicle determines a target follower unmanned aerial vehicle in each follower unmanned aerial vehicle based on a preset non-return sampling algorithm and scheduling probability of each follower unmanned aerial vehicle.

And step 404, updating the scheduling probability of each target follower unmanned aerial vehicle to zero.

In an implementation, in order to enrich the data set of the global model training, after determining that the target follower unmanned aerial vehicle is completed, the scheduling probability of the target follower unmanned aerial vehicle is updated to zero.

In an exemplary embodiment, due to channel distortion problems introduced by over-the-air computing, multiple follower drones need to be scheduled to mitigate the corresponding effects. In particular, by scheduling multiple follower drones, the variance of the channel noise will be reduced toThereby reducing the effect of channel distortion to some extent. Wherein (1)>Square of default power scaling factor, +.>Is the square of the power scaling factor in the t-th training. / >And (5) representing a target follower unmanned aerial vehicle set in the t-th round of training. In order to realize a low-complexity scheduling algorithm, the pilot unmanned aerial vehicle sequentially selects a plurality of follower unmanned aerial vehicles based on a preset non-return sampling strategy according to the optimal scheduling probability distribution represented by the formula (2). Specifically, when s-1 follower drones are selected,the scheduling probability of the selected follower drone (which is the target follower drone) will be set to 0. Therefore, the conditional dispatch profile for selecting the next follower drone is shown in the following formula (6):

（6）

wherein, in the above formula (6),and characterizing the scheduling probability of the next follower unmanned aerial vehicle. />For the target follower drone that has been selected. t is the current training wheel number. />The scheduling probability of the follower man-machine in the formula (2). />Is a natural number from 1 to s-1. />For a follower drone that is not selected. />The scheduling probability for the j-th non-selected follower drone. Thus, the scaled local gradient of the selected follower drone (target follower drone) may be expressed as equation (7), equation (7) as follows:

（7）

wherein, in the above formula (7), Local model parameters for the nth target follower unmanned aerial vehicle->Data quantity representing the complete data set, +.>Representing the scheduling probability of the next follower unmanned aerial vehicle, < ->Scaling the local gradient for the nth target follower drone. />Data volume of the local data set of the nth target follower drone. It is noted that the mean value of the local model parameters (local gradients) selected at this time still satisfies an unbiased estimate of the true global gradient during the aggregation.

In the embodiment, the target follower unmanned aerial vehicle is determined through the non-return sampling method and the scheduling probability of each follower unmanned aerial vehicle, the target follower unmanned aerial vehicle with better performance is determined, and the scheduling efficiency of the unmanned aerial vehicle is improved.

In an exemplary embodiment, as shown in fig. 5, the specific process of training the preset global model based on each target follower unmanned aerial vehicle in step 206 to obtain the trained global model includes steps 502 to 504. Wherein:

step 502, receiving local model parameters of each target follower unmanned aerial vehicle through a sight line channel.

The local model parameters of each target follower unmanned aerial vehicle are aggregated in the sight line channel, and the aggregated model parameters are obtained.

In implementation, each target follower drone transmits local model parameters of the target follower drone to the pilot drone through a Line-of-Sight (LoS) channel. And aggregating the local model parameters of each follower unmanned aerial vehicle in the sight line channel to obtain aggregated model parameters. The pilot drone then receives the aggregate model parameters over the line-of-sight channel.

And step 504, updating parameters of a preset global model based on the aggregate model parameters, the transmission noise and the power scaling factor to obtain a trained global model.

In an implementation, a model update algorithm is preset in the pilot unmanned aerial vehicle. And the pilot unmanned aerial vehicle model updating algorithm performs data processing on the aggregate model parameters, the transmission noise and the power scaling factors, updates the parameters of a preset global model and obtains the trained global model.

The model updating algorithm is shown in the following formula (8):

（8）

wherein, in the above formula (8),global model parameters for the current training round (t-th round,)>Is the global model parameter of the previous training round (t-1). />For the learning rate of the global model in the current training round,/for the global model in the current training round>Is the power scaling factor in the current training round. / >To represent the set of target follower drones in the t-th round of training. />Is an aggregate model parameter. />Is transmission noise, i.e., additive gaussian noise with zero mean and variance. />Unmanned pilotThe machine will receive a weighted sum of the transmitted signals doped with channel fading and noise.

In an exemplary embodiment, the unmanned aerial vehicle scheduling algorithm of the present application is applied in the unmanned aerial vehicle scheduling cluster 100. The drone dispatch cluster 100 includes 1 pilot drone andindividual follower unmanned aerial vehicle UAV (Unmanned Aerial Vehicle )/(unmanned aerial vehicle)>. In the collaborative training process of each round, the selected follower unmanned aerial vehicle (target follower unmanned aerial vehicle) trains and shares a global model with the pilot unmanned aerial vehicle, and training rounds are indexed by using T epsilon {1,2, …, T }, wherein T is a communication total round and is also a training total round. Because all unmanned aerial vehicles are used as a cluster in the task execution process, modeling is performed based on a three-dimensional Cartesian coordinate system. In particular, the three-dimensional coordinates of the pilot drone and the follower drone in the t-th round of training may be expressed as +.>And. However, due to the influence of the dynamic environment, the unmanned aerial vehicle group lacks the ability to maintain a fixed topology, resulting in a constant change in the relative position between the follower unmanned aerial vehicle and the pilot unmanned aerial vehicle, i.e. the dynamic characteristics of the environment result in a constant change in the spatial structure of the fabric within the cluster. Since unmanned aerial vehicles typically fly at high altitudes, the unmanned aerial vehicle clusters encounter few obstacles. Thus, it is assumed that the communication link between each follower drone and the pilot drone is predominantly dominated by the line-of-sight channel to achieve reliable communication. According to the free space path loss model, in the t-th round of updating, the channel gain between the follower unmanned aerial vehicle n and the pilot unmanned aerial vehicle is as follows:

（9）

Wherein, in the above formula (9).For reference distance->Channel gain at>Representing the distance between the pilot unmanned aerial vehicle and the follower unmanned aerial vehicle, i.e.。/>Representing the channel gain between follower drone n and pilot drone, +.>Is a distance decay index. In the course of setting up the federal learning system, a local data set based on the follower unmanned plane n +.>The local loss function of (2) is shown in the following formula (10):

（10）

wherein, in the above formula (10),for data sample->For quantifying the training samples +.>Model parameters>Real tag->Prediction error between the two. Data sample->Is->Data sample,/->Is the local data set of the follower unmanned aerial vehicle n. The task of the unmanned aerial vehicle cluster is to obtain an optimal global model parameter +_by minimizing the global loss function of federal learning>The following equation (11) represents a process of optimizing the global model parameters: />

（11）

Wherein, in the above formula (11),。/>is the optimal global model parameter.For the data volume of the complete data set, +.>Is the local data set of the nth follower drone. />Is a global model parameter.

In a conventional federal averaging strategy, the probability of each node participating in training is Equal. However, in heterogeneous environments, this strategy typically leads to global gradient bias. To overcome this problem, the present application employs a probabilistic scheduling framework for scheduling design, wherein each follower drone is assigned a specific probability, denoted as，/>The probability of participating in the round of training is represented. Thus, the global model parameter update of the pilot drone may be performed by equation (12):

（12）

wherein,for learning rate in training round t, coefficient +.>The method is used for guaranteeing unbiased gradient aggregation of the pilot unmanned aerial vehicle in the t-th round of training. />For the data volume of the complete data set, +.>Is the local data set of the nth follower drone.The scheduling probability of the nth follower unmanned aerial vehicle in the training process of the nth round is given. />The model parameters are local model parameters of the nth follower unmanned aerial vehicle in the t-th training process. />To represent the target in the t-th trainingA set of follower drones. />For global model parameters in training round t, < ->Is a global model parameter in the t-1 th round of training.

In order to further improve communication efficiency, the target follower unmanned aerial vehicle adopts an air computing transmission mode in the federal learning system to access an uplink channel. It is assumed that the transmissions of all target follower drones are strictly synchronized. In order to mitigate the effects of channel fading and noise introduced during waveform superposition in the air computation, each target follower drone n transmits power Its local gradient information (local model parameters) is transmitted. Therefore, the pilot drone weights the transmission signal received with the channel fading and noise and is shown in the following equation (13):

（13）

wherein, in the above formula (13),，/>and represents additive gaussian noise with zero mean and variance. />Is the channel gain between follower drone n and pilot drone. />The scheduling probability of the nth follower unmanned aerial vehicle in the training process of the nth round is given. />Local model parameters of the nth follower unmanned aerial vehicle in the t-th training process. />To represent the set of target follower drones in the t-th round of training. In order to combat the effects of channel fading and noise, a channel inversion strategy is used for gradient aggregation. Specifically, the transmission power of the follower drone n in the t-round +.>The following formula (14) shows:

（14）

wherein, in the formula (14),is the corresponding power scaling factor. />Is the channel gain between follower drone n and pilot drone. Thus, the model update algorithm for the global model is shown in equation (8) above.

In the embodiment, the aggregation model parameters are received based on the sight line channel, the low-delay parameter uploading is realized by utilizing the superposition characteristic of the sight line channel, and the parameters of the global model are updated based on the aggregation model parameters, so that the efficiency of training the global model is improved.

In one exemplary embodiment, in the event that the trained global model does not meet the training stop condition, the training of the global model is continued until the trained global model meets the training stop condition. As shown in fig. 6, after the step 206 is performed, the specific processing procedure of the method for scheduling distributed computing integrated resources of the unmanned aerial vehicle further includes steps 602 to 604. Wherein:

step 602, determining whether the trained global model meets a preset training stop condition.

The training stopping condition is that the training round reaches a preset training round threshold or the global model converges.

In implementation, a training round threshold is preset in the pilot unmanned aerial vehicle. The pilot unmanned aerial vehicle judges whether the training round reaches a preset training round threshold value and judges whether the trained global model converges. Under the condition that the trained global model converges or the training round reaches a preset training round threshold, the pilot unmanned aerial vehicle determines that the trained global model meets a preset training stop condition. Under the conditions that the trained global model does not converge and the training round does not reach the preset training round threshold, the pilot unmanned aerial vehicle determines that the trained global model does not meet the preset training stop condition.

Optionally, the training round threshold is determined according to a training requirement of the global model, and the training round threshold is not limited in the embodiment of the present application.

Step 604, executing the step of sending the global model parameters to each follower unmanned aerial vehicle until the trained global model meets the training stop condition under the condition that the trained global model does not meet the training stop condition.

In implementation, in the case where the trained global model does not meet the training stop condition, the pilot drone performs step 202 above until the trained global model meets the training stop condition. The specific implementation process of step 202 is described in detail in the above embodiments, and the embodiments of the present application are not described herein again.

In this embodiment, under the condition that the trained global model does not meet the training stop condition, the global model is trained again, and the accuracy of the global model is improved through multiple times of training of the global model, so that the scheduling efficiency of the unmanned aerial vehicle is improved.

In an exemplary embodiment, as shown in fig. 7, a method for scheduling resources of a distributed computing system of a drone is provided, and the method is applied to the follower drone 120 (the follower drone is omitted from the following description) in fig. 1, and the method includes the following steps 702 to 708. Wherein:

Step 702, receiving a global model parameter sent by a pilot unmanned aerial vehicle, and executing a preset random gradient descent algorithm according to the global model parameter to obtain a euclidean norm of a local model parameter.

In an implementation, a follower drone receives global model parameters sent by the drone. And then, the follower unmanned aerial vehicle executes a preset random gradient descent algorithm according to the global model parameters, and calculates local model parameters corresponding to the follower unmanned aerial vehicle. And the follower unmanned aerial vehicle carries out vector length calculation on the local model parameters to obtain Euclidean norms of the local model parameters.

In an exemplary embodiment, the follower drone receives the global model parametersThen, local small lot local data in the local data set based on the follower drone, i.e +.>Executing random gradient descent algorithm, calculating local model gradient (local model parameter)>. Wherein (1)>Is local data in small local batches. The local model gradient calculation is shown in the following formula (15):

（15）/>

wherein, in the above formula (15),is the t-th round of slave dataset +.>A subset of randomly selected tagged data samples. />Is->Data samples. / >For data sample->Gradient of the loss function at t-1 round. Limited by communication resources and storage capacity, each round can only select a part of the capable follower unmanned aerial vehicles for gradient uploading, wherein +.>Representing the set of follower drones selected by the t-th round.

Step 704, determining the total energy consumption of the follower unmanned aerial vehicle according to the transmission attribute of the follower unmanned aerial vehicle and the property of the central processing unit.

In an implementation, the follower drone includes a CPU (Central Processing Unit ). And determining the calculated energy consumption of the follower unmanned aerial vehicle according to the CPU frequency and the CPU fan revolution required by the data sample training. The follower drone then determines communication energy consumption based on the transmission attributes. And the follower unmanned aerial vehicle performs addition operation on the calculated energy consumption and the communication energy consumption to obtain the total energy consumption of the follower unmanned aerial vehicle.

Step 706, determining an energy virtual queue based on total energy constraints of the follower drone.

In an implementation, the follower drone determines an energy virtual queue from the gap between the accumulated energy consumption and the total energy constraint.

In an exemplary embodiment, to satisfy the total energy constraint of the device, the present application constructs a virtual queue for a follower drone n∈n To represent the difference between the accumulated energy consumption and the energy constraint in the t-round training and the update mode of the virtual queueThe following formula (16) shows:

（16）

wherein, in the above formula (16), the initial value，/>Indicating whether UAV N (n.epsilon.N) is involved in the gradient upload of the present round, wherein ∈N>And (5) indicating to participate in the gradient uploading of the round, otherwise, not participating in the gradient uploading of the round. />Representing the total energy constraint of follower drone n, +.>Representing the average energy constraint for each round.

Step 708, transmitting the Euclidean norms of the local model parameters, the total energy consumption and the energy virtual queues to the pilot drone.

In an implementation, the follower drone transmits the euclidean norms of the local model parameters, the total energy consumption, and the virtual queue of energy to the pilot drone over a control channel.

In this embodiment, the euclidean norms, the total energy consumption and the energy virtual queues of the local model parameters corresponding to the follower unmanned aerial vehicle are transmitted to the pilot unmanned aerial vehicle, so that the follower unmanned aerial vehicle can determine the scheduling probability of each follower unmanned aerial vehicle, and the aggregation model parameters become accurate.

In an exemplary embodiment, as shown in FIG. 8, the specific process of step 704 includes steps 802 through 806. Wherein:

Step 802, determining the corresponding calculation energy consumption of the follower unmanned aerial vehicle according to the property of the central processing unit of the follower unmanned aerial vehicle.

The properties of the CPU include CPU frequency and CPU fan speed.

In implementation, the follower unmanned aerial vehicle performs data processing on the CPU frequency and the fan rotation speed of the CPU of the follower unmanned aerial vehicle according to a preset calculation energy consumption algorithm to obtain the corresponding calculation energy consumption of the follower unmanned aerial vehicle.

In an exemplary embodiment, in view of the limited energy supply of the drone, how to minimize the energy consumption of the overall training is critical, including communication energy consumption and computing energy consumption. The present application describes computing power using CPU frequency. Suppose the CPU frequency of the follower unmanned plane n in the t-round isThe number of CP revolutions required for training a data sample is +.>. Therefore, in the t-th training, the calculation energy consumption algorithm of the follower unmanned aerial vehicle n is shown in the following formula (17):

（17）

wherein, in the above formula (17),is an effective switched capacitance value that depends on the chip architecture. Specifically, & gt>Is time-varying during the training process, as the computational load of different drones is constantly changing. />The CPU frequency of the follower unmanned plane n. / >The number of CP revolutions required for training one data sample. />And calculating energy consumption for the follower unmanned aerial vehicle n.

Step 804, determining the communication energy consumption corresponding to the follower unmanned aerial vehicle according to the transmission power and the power scaling factor of the follower unmanned aerial vehicle.

In implementation, a communication energy consumption algorithm preset by the follower unmanned aerial vehicle carries out data calculation on the transmission power and the power scaling factor of the follower unmanned aerial vehicle to obtain the communication energy consumption corresponding to the follower unmanned aerial vehicle. The communication energy consumption algorithm is shown in the following formula (18):

（18）

wherein, in the above formula (18),communication energy consumption generated by the transmission gradient of the follower unmanned plane n in the t-th round. />Representing the transmission power of the follower drone n in the t-th round. />Is the corresponding power scaling factor. />Is the channel gain between follower drone n and pilot drone.

And step 806, determining the total energy consumption corresponding to the follower unmanned aerial vehicle according to the calculated energy consumption and the communication energy consumption.

In implementation, the follower unmanned aerial vehicle performs addition operation on the calculated energy consumption and the communication energy consumption, and the total energy consumption of the follower unmanned aerial vehicle is obtained. The calculation process of the total energy consumption is shown in the following formula (19):

（19）

wherein, in the above formula (19), Is the total energy consumption of the follower unmanned plane. />Communication energy consumption generated by the transmission gradient of the follower unmanned plane n in the t-th round. />And calculating energy consumption for the follower unmanned aerial vehicle n.

In the embodiment, the total energy consumption of the follower unmanned aerial vehicle is determined through the CPU property and the transmission property of the follower unmanned aerial vehicle, so that the energy consumption of the follower unmanned aerial vehicle in the training process is clear, and the follow-up determination of the scheduling probability of the follower unmanned aerial vehicle is facilitated.

In an exemplary embodiment, fig. 9 is a schematic flow chart of a method for scheduling distributed computing integrated resources of a drone in another embodiment, which includes:

step 901, a pilot unmanned aerial vehicle broadcasts global model parameters.

In an implementation, a pilot drone broadcasts global model parameters to each follower drone.

In step 902, all the follower unmanned aerial vehicles perform local model training, and the total energy consumption and the energy virtual queue of each follower unmanned aerial vehicle are determined.

In practice, each follower man-machine receives a global model parameter and performs a random gradient descent algorithm based on the global model parameter to obtain a local model gradient (local model parameter). Then, each follower unmanned plane carries out vector length calculation on the local model gradient to obtain the Euclidean norm of the local model gradient. Each follower unmanned aerial vehicle determines the total energy consumption of the follower unmanned aerial vehicle according to the transmission attribute of the follower unmanned aerial vehicle and the property of the central processing unit. Then, each follower unmanned aerial vehicle determines an energy virtual queue based on the total energy constraint condition of the follower unmanned aerial vehicle.

In step 903, all follower drones upload the euclidean norms, total energy consumption, and energy virtual queues of the local model gradients.

In an implementation, each follower drone uploads the euclidean norm of the local model gradient, the total energy consumption, and the energy virtual queues to the pilot drone over a control channel.

In step 904, the pilot unmanned aerial vehicle calculates scheduling probability according to a formula.

In an implementation, a pilot drone receives euclidean norms, total energy consumption, and energy virtual queues for local model gradients for each follower drone. And then, aiming at each follower unmanned aerial vehicle, the pilot unmanned aerial vehicle processes Euclidean norms, total energy consumption and energy virtual queues of the local model gradient of the follower unmanned aerial vehicle according to a preset probability algorithm (formula), and the scheduling probability of the follower unmanned aerial vehicle is obtained.

In step 905, the pilot drone selects a specified number of follower drones according to the scheduling probability samples.

In an implementation, a pilot unmanned aerial vehicle determines a target follower unmanned aerial vehicle among the follower unmanned aerial vehicles according to scheduling probabilities of the follower unmanned aerial vehicles.

Step 906, aggregating gradient information of the specified number of follower unmanned aerial vehicles, and updating global model parameters based on the gradient information.

In practice, each target follower drone transmits a local model gradient to the pilot drone over a line-of-sight channel. The local model gradients of the target follower unmanned aerial vehicle are aggregated in the sight line channel to obtain aggregation model parameters. And updating parameters of a preset global model by the pilot unmanned aerial vehicle based on the aggregated model parameters to obtain a trained global model.

Step 907, it is determined whether the maximum update round number is converged or reached.

In implementation, the pilot unmanned aerial vehicle determines whether the trained global model converges or whether the training round number reaches a maximum update round number. If the trained global model converges or the training round number reaches the maximum updating round number, the pilot unmanned aerial vehicle determines that the trained global model is an unmanned aerial vehicle model, and the training is finished. If the trained global model does not converge and the training round number does not reach the maximum update round number, the pilot unmanned aerial vehicle executes step 901 until the trained global model converges or the training round number reaches the maximum update round number.

It should be understood that, although the steps in the flowcharts related to the embodiments described above are sequentially shown as indicated by arrows, these steps are not necessarily sequentially performed in the order indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least some of the steps in the flowcharts described in the above embodiments may include a plurality of steps or a plurality of stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of the steps or stages is not necessarily performed sequentially, but may be performed alternately or alternately with at least some of the other steps or stages.

Based on the same inventive concept, the embodiment of the application also provides an unmanned aerial vehicle distributed general calculation integrated resource scheduling device for realizing the unmanned aerial vehicle distributed general calculation integrated resource scheduling method. The implementation scheme of the solution to the problem provided by the device is similar to the implementation scheme recorded in the method, so the specific limitation in the embodiments of the one or more unmanned aerial vehicle distributed general calculation integrated resource scheduling devices provided below can be referred to the limitation of the unmanned aerial vehicle distributed general calculation integrated resource scheduling method hereinabove, and the description is omitted here.

In an exemplary embodiment, as shown in fig. 10, there is provided a distributed general computing integrated resource scheduling apparatus 1000 for a drone, including: a receiving module 1001, a first determining module 1002, a training module 1003, and a second determining module 1004, wherein:

and the receiving module 1001 is configured to send global model parameters to each follower unmanned aerial vehicle, and receive euclidean norms, total energy consumption and energy virtual queues of the local model parameters fed back by each follower unmanned aerial vehicle.

A first determining module 1002, configured to determine, for each follower unmanned aerial vehicle, a scheduling probability of the follower unmanned aerial vehicle based on the euclidean norm of the local model parameter, the total energy consumption, and the energy virtual queue corresponding to the follower unmanned aerial vehicle.

The training module 1003 is configured to determine a target follower unmanned aerial vehicle in each follower unmanned aerial vehicle according to the scheduling probability of each follower unmanned aerial vehicle, and train a preset global model based on each target follower unmanned aerial vehicle, so as to obtain a trained global model.

A second determining module 1004, configured to determine the trained global model as an unmanned plane scheduling model when the trained global model meets a preset training stop condition; the unmanned aerial vehicle scheduling model is used for scheduling the follower unmanned aerial vehicle to execute tasks.

In an exemplary embodiment, the first determining module 1002 includes:

a first determining sub-module for determining, for each follower unmanned aerial vehicle, a data set duty cycle of the follower unmanned aerial vehicle; the data set duty cycle is the ratio of the local data set of the follower drone to the total data set of the total follower drone.

And the first processing sub-module is used for carrying out data processing on the Euclidean norm of the local model parameter, the total energy consumption, the energy virtual queue and the data set duty ratio according to a preset probability algorithm to obtain the scheduling probability of the follower unmanned aerial vehicle.

In an exemplary embodiment, training module 1003 includes a second determination sub-module and a first training sub-module. Wherein the second determination submodule includes:

And the third determining sub-module is used for determining the target follower unmanned aerial vehicle in each follower unmanned aerial vehicle according to a preset non-return sampling algorithm and the scheduling probability of each follower unmanned aerial vehicle.

And the first updating sub-module is used for updating the scheduling probability of each target follower unmanned aerial vehicle to zero.

In an exemplary embodiment, training module 1003 includes a second determination sub-module and a first training sub-module. Wherein the first training submodule includes:

the first receiving sub-module is used for receiving local model parameters of each target follower unmanned aerial vehicle through a sight channel; and aggregating the local model parameters of each target follower unmanned aerial vehicle in the sight line channel to obtain aggregated model parameters.

And the second updating sub-module is used for updating the parameters of the preset global model based on the aggregate model parameters, the transmission noise and the power scaling factor to obtain the trained global model.

In an exemplary embodiment, the unmanned plane distributed general computing integrated resource scheduling apparatus 1000 further includes:

the judging module is used for judging whether the trained global model meets the preset training stopping condition.

And the second execution module is used for executing the step of sending the global model parameters to each follower unmanned aerial vehicle under the condition that the trained global model does not meet the training stop condition until the trained global model meets the training stop condition.

In an exemplary embodiment, as shown in fig. 11, there is provided a distributed computing integrated resource scheduling apparatus 1100 for a drone, including: an execution module 1101, a third determination module 1102, a fourth determination module 1103, and a transmission module 1104, wherein:

the execution module 1101 is configured to receive a global model parameter sent by the pilot unmanned aerial vehicle, and execute a preset random gradient descent algorithm according to the global model parameter, so as to obtain a euclidean norm of the local model parameter.

A third determining module 1102 is configured to determine total energy consumption of the follower unmanned aerial vehicle according to the transmission attribute of the follower unmanned aerial vehicle and the property of the central processor.

A fourth determining module 1103 is configured to determine an energy virtual queue based on the total energy constraint condition of the follower drone.

A transmission module 1104 is configured to transmit the euclidean norms of the local model parameters, the total energy consumption, and the energy virtual queues to the pilot drone.

In an exemplary embodiment, the third determining module 1102 is configured to determine, according to a property of a central processor of the follower unmanned aerial vehicle, a corresponding calculated energy consumption of the follower unmanned aerial vehicle; determining the communication energy consumption corresponding to the follower unmanned aerial vehicle according to the transmission power and the power scaling factor of the follower unmanned aerial vehicle; and determining the total energy consumption corresponding to the follower unmanned aerial vehicle according to the calculated energy consumption and the communication energy consumption.

All or part of each module in the unmanned aerial vehicle distributed general calculation integrated resource scheduling device can be realized by software, hardware and a combination thereof. The above modules may be embedded in hardware or may be independent of a processor in the computer device, or may be stored in software in a memory in the computer device, so that the processor may call and execute operations corresponding to the above modules.

In an exemplary embodiment, a computer device, which may be a terminal, is provided, and an internal structure thereof may be as shown in fig. 12. The computer device includes a processor, a memory, an input/output interface, a communication interface, a display unit, and an input means. The processor, the memory and the input/output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input/output interface. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input/output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used for carrying out wired or wireless communication with an external terminal, and the wireless mode can be realized through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. The computer program, when executed by the processor, implements a distributed general calculation integrated resource scheduling method for unmanned aerial vehicles. The display unit of the computer device is used for forming a visual picture, and can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, can also be a key, a track ball or a touch pad arranged on the shell of the computer equipment, and can also be an external keyboard, a touch pad or a mouse and the like.

It will be appreciated by those skilled in the art that the structure shown in fig. 12 is merely a block diagram of some of the structures associated with the present application and is not limiting of the computer device to which the present application may be applied, and that a particular computer device may include more or fewer components than shown, or may combine certain components, or have a different arrangement of components.

In an embodiment, there is also provided a computer device comprising a memory and a processor, the memory having stored therein a computer program, the processor implementing the steps of the method embodiments described above when the computer program is executed.

In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored which, when executed by a processor, carries out the steps of the method embodiments described above.

Those skilled in the art will appreciate that implementing all or part of the above described methods may be accomplished by way of a computer program stored on a non-transitory computer readable storage medium, which when executed, may comprise the steps of the embodiments of the methods described above. Any reference to memory, database, or other medium used in the various embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, high density embedded nonvolatile Memory, resistive random access Memory (ReRAM), magnetic random access Memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric Memory (Ferroelectric Random Access Memory, FRAM), phase change Memory (Phase Change Memory, PCM), graphene Memory, and the like. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, and the like. By way of illustration, and not limitation, RAM can be in the form of a variety of forms, such as Static Random access memory (Static Random access memory AccessMemory, SRAM) or dynamic Random access memory (Dynamic Random Access Memory, DRAM), and the like. The databases referred to in the various embodiments provided herein may include at least one of relational databases and non-relational databases. The non-relational database may include, but is not limited to, a blockchain-based distributed database, and the like. The processors referred to in the embodiments provided herein may be general purpose processors, central processing units, graphics processors, digital signal processors, programmable logic units, quantum computing-based data processing logic units, etc., without being limited thereto.

The technical features of the above embodiments may be arbitrarily combined, and all possible combinations of the technical features in the above embodiments are not described for brevity of description, however, as long as there is no contradiction between the combinations of the technical features, they should be considered as the scope of the description.

The above examples only represent a few embodiments of the present application, which are described in more detail and are not to be construed as limiting the scope of the present application. It should be noted that it would be apparent to those skilled in the art that various modifications and improvements could be made without departing from the spirit of the present application, which would be within the scope of the present application. Accordingly, the scope of protection of the present application shall be subject to the appended claims.

Claims

1. A method for scheduling distributed general calculation integrated resources of an unmanned aerial vehicle, wherein the method is applied to a pilot unmanned aerial vehicle, and the method comprises:

transmitting global model parameters to each follower unmanned aerial vehicle, and receiving Euclidean norms, total energy consumption and energy virtual queues of the local model parameters fed back by each follower unmanned aerial vehicle; the energy virtual queue is the difference between the accumulated energy consumption and the total energy constraint of the follower unmanned aerial vehicle;

under the condition that the trained global model meets a preset training stop condition, determining the trained global model as an unmanned plane scheduling model; the unmanned aerial vehicle scheduling model is used for scheduling the follower unmanned aerial vehicle to execute tasks;

the determining, for each follower unmanned aerial vehicle, a scheduling probability of the follower unmanned aerial vehicle based on the euclidean norm of the local model parameter, the total energy consumption, and the energy virtual queue corresponding to the follower unmanned aerial vehicle includes:

Performing data processing on the Euclidean norm of the local model parameter, the total energy consumption, the energy virtual queue and the data set duty ratio according to a preset probability algorithm to obtain the scheduling probability of the follower unmanned aerial vehicle; the probability algorithm isThe method comprises the steps of carrying out a first treatment on the surface of the Wherein n represents the nth follower unmanned aerial vehicle, the training round is indexed by using T epsilon {1,2, …, T }, and T is the training total round; />Euclidean norms, for the local model parameters of the follower unmanned aerial vehicle->For the data set duty cycle, +.>For the energy virtual queue, +.>Representation assurance->The optimal value of the Lagrangian multiplier is obtained through a one-dimensional search algorithm; />Representing the total energy consumption->Is a weight coefficient; />And the scheduling probability of the follower unmanned aerial vehicle at the t-th round is given.

2. The method of claim 1, wherein the determining a target follower drone among the follower drones according to the scheduling probabilities of the follower drones, comprises:

3. The method of claim 1, wherein training a preset global model based on each of the target follower unmanned aerial vehicles to obtain a trained global model comprises:

4. The method of claim 1, wherein after training a preset global model based on each of the target follower unmanned aerial vehicles to obtain a trained global model, the method further comprises:

5. A distributed general calculation integrated resource scheduling method for unmanned aerial vehicles, wherein the method is applied to a follower unmanned aerial vehicle, the method comprising:

determining an energy virtual queue based on the total energy constraint condition of the follower unmanned aerial vehicle; the energy virtual queue is the difference between the accumulated energy consumption and the total energy constraint of the follower unmanned aerial vehicle;

6. The method of claim 5, wherein determining the total energy consumption of the follower drone based on the transmission attributes of the follower drone and the nature of the central processor, comprises:

7. An unmanned aerial vehicle distributed general calculation integrated resource scheduling device, wherein the device is applied to a pilot unmanned aerial vehicle, the device comprises:

the receiving module is used for sending global model parameters to each follower unmanned aerial vehicle and receiving Euclidean norms, total energy consumption and energy virtual queues of the local model parameters fed back by each follower unmanned aerial vehicle; the energy virtual queue is the difference between the accumulated energy consumption and the total energy constraint of the follower unmanned aerial vehicle;

the second determining module is used for determining the trained global model as an unmanned plane scheduling model under the condition that the trained global model meets a preset training stopping condition; the unmanned aerial vehicle scheduling model is used for scheduling the follower unmanned aerial vehicle to execute tasks;

performing data processing on the Euclidean norm of the local model parameter, the total energy consumption, the energy virtual queue and the data set duty ratio according to a preset probability algorithm to obtain the scheduling probability of the follower unmanned aerial vehicle; the probability algorithm isThe method comprises the steps of carrying out a first treatment on the surface of the Wherein n represents the nth follower unmanned aerial vehicle, the training round is indexed by using T epsilon {1,2, …, T }, and T is the training total round; />Euclidean norms, for the local model parameters of the follower unmanned aerial vehicle->For the data set duty cycle, +.>For the energy virtual queue, +.>Representation assurance->The optimal value of the Lagrangian multiplier is obtained through a one-dimensional search algorithm; / >Representing the total energy consumption->For the rightA weight coefficient; />And the scheduling probability of the follower unmanned aerial vehicle at the t-th round is given.

8. The apparatus of claim 7, wherein the apparatus further comprises:

the judging module is used for judging whether the trained global model meets preset training stopping conditions or not;

9. An unmanned aerial vehicle distributed general calculation integrated resource scheduling device, wherein the device is applied to a follower unmanned aerial vehicle, the device comprises:

a fourth determining module, configured to determine an energy virtual queue based on a total energy constraint condition of the follower unmanned aerial vehicle; the energy virtual queue is the difference between the accumulated energy consumption and the total energy constraint of the follower unmanned aerial vehicle;

10. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any one of claims 1 to 4 or 5 to 6 when the computer program is executed.