Method for task cooperation of unmanned aerial vehicle cluster

Abstract

The application provides a UAV cluster task cooperation method, comprising the following steps: model splitting, splitting the model into two parts, delivering to different groups of UAVs for training, and adding corresponding auxiliary functions to the split model; model training, in a specific global round, each UAV receives the model and auxiliary function aggregated in the last round; group leader selection, selecting a high-reputation node as a large group leader node, which is responsible for aggregating and distributing parameters in federated learning; auxiliary function, extracting the potential features of input information through a self-encoder, and using the potential features to reconstruct an output result with the same size as the input data and with representative features. By introducing auxiliary functions into the split model, the application realizes parallel training of the split model, improves the accuracy of cluster task cooperation and the invulnerability of the cluster, and reduces the communication burden between nodes.

Description

TECHNICAL FIELD

[0001] The present application relates to a UAV cluster task cooperation method, belonging to the technical field of machine learning. BACKGROUND

[0002] In a UAV cluster, in order to efficiently process the acquired data, the cluster will adopt a machine learning algorithm to extract effective feature information from a large data set.

[0003] The traditional idea is to centralize learning by collecting data collected by each node in the cluster to the base station. However, limited bandwidth and dynamic wireless channels can greatly affect the transmission efficiency of data, and the load and bandwidth required during transmission can also affect the effect of model training.

[0004] Therefore, it is necessary to provide a UAV cluster task cooperation method to solve the above problems. SUMMARY

[0005] The purpose of the present application is to provide a UAV cluster task cooperation method, which can solve the problems of large data transmission load, weak cluster invulnerability and privacy leakage in cluster task cooperation.

[0006] To achieve the above purpose, the present application provides a UAV cluster task cooperation method, comprising the following steps:

[0007] Step 1, model splitting, splitting the model into two parts and delivering it to different groups of UAVs for training, while adding corresponding auxiliary functions to the split model;

[0008] Step 2, model training, in a specific global round, each UAV receives the model and auxiliary function aggregated in the last round;

[0009] Step 3, group leader selection, the selection of large group leader nodes is to select high reputation nodes, and the high reputation nodes are responsible for aggregation and distribution of parameters in federated learning;

[0010] Step 4, auxiliary function, the latent features of input data are extracted through an autoencoder, which are used to reconstruct output results with the same size as input data and with characteristic features.

[0011] As a further improvement of the present application, in step 1, each UAV node in the cluster has pre-stored a complete initial model when it starts, and the two parts of the network model after splitting are M S and M C , and the corresponding auxiliary functions are AF S and AF C , after grouping the UAV cluster according to the grouping algorithm, the group is equally divided into server group GS and client group GC, and the two groups are trained respectively.S and M C .

[0012] As a further improvement of the present invention, in step 1, after a certain number of iterations, the training models are exchanged between groups to maximize the use of the local datasets of each node.

[0013] As a further improvement of the present invention, in step 2, in a specific global round t, each drone s∈GS in the server group GS t Received the model after the previous round of global aggregation and auxiliary functions Each drone c∈GC in the client group GC t Received the model after the previous round of global aggregation and auxiliary functions And let

[0014]

[0015] As a further improvement of the present invention, step 2 specifically includes the following steps:

[0016] Step 21: Each drone node c in the client group GC is in the previous round of client model. The forward propagation is performed in parallel, passing the intermediate data obtained from its training to the auxiliary function. Depend on Perform a simple backpropagation;

[0017] Step 22: Each UAV node s in the server group GS executes the auxiliary function. Perform a simple forward propagation to pass the output training data to the server-side model. Perform backpropagation;

[0018] Step 23: The server group GS and the client group GC process forward propagation and backward propagation in parallel, and at the end of each iteration, they exchange their parameters to optimize the auxiliary function and model, and then distribute the optimized model and auxiliary function parameters.

[0019] Step 24: Repeat steps 21-23.

[0020] As a further improvement of the present invention, in step 21, the parameters are updated based on the local loss function to obtain... and

[0021]

[0022] Where, η t Let be the learning rate in round t.

[0023] As a further improvement of the present application, in step 22, the updated parameters are obtained and

[0024]

[0025] wherein, η t is the learning rate of the t-th round.

[0026] As a further improvement of the present application, in step 23, the optimized model and the auxiliary function of the client group GC are updated as

[0027]

[0028] As a further improvement of the present application, in step 3, the calculation formula of the reputation T is as follows:

[0029] T = α1C1 + α2C2 + α3C3,

[0030] wherein, the meanings of the parameters are as follows:

[0031] C1: the approximation degree of the parameters uploaded by the node and the updated parameters after aggregation;

[0032] C2: the difference factor of the node dataset;

[0033] C3: the state of the node itself;

[0034] α1, α2, α3: weighted parameters, satisfying α1 + α2 + α3 = 1, and the values are set according to the influence degree of the parameters on the reputation.

[0035] As a further improvement of the present application, in step 4, the auxiliary function adopts a self-encoder, and the self-encoder comprises an encoder, a latent feature representation module and a decoder.

[0036] The present application has the beneficial effects that: the present application realizes parallel training of the split model by introducing an auxiliary function into the split model. While improving the task cooperation accuracy of the cluster and the anti-destroying ability of the cluster, the communication burden between nodes is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The algorithm architecture diagram of the unmanned aerial vehicle cluster task cooperation method.

[0038] Figure 2 The split federated learning parallel training flowchart of the unmanned aerial vehicle cluster task cooperation method. DETAILED DESCRIPTION

[0039] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be described in detail below with reference to the drawings and specific embodiments.

[0040] Here, it should be noted that, in order to avoid obscuring the present application due to unnecessary details, only structures and / or processing steps closely related to the solutions of the present application are shown in the drawings, and other details not closely related to the present application are omitted.

[0041] In addition, it should also be noted that the term "comprise", "include" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device.

[0042] As shown in Figure 1 and Figure 2 , the present application discloses a UAV cluster task cooperation method, which can realize P2P-based node communication in a UAV cluster, without the participation of a base station or a server, thereby reducing the communication burden between nodes. At the same time, the algorithm adopts a combination of federated learning and split learning, with the help of auxiliary functions, to realize parallel training of task cooperation models by each node, thereby improving the accuracy of cluster task cooperation and the invulnerability of the cluster.

[0043] The UAV cluster task cooperation method comprises the following steps:

[0044] Step 1, model splitting, splitting the model into two parts and delivering them to different groups of UAVs for training, while adding corresponding auxiliary functions to the split models to realize parallel training of the split models.

[0045] Suppose that each UAV node in the cluster has pre-stored a complete initial model (i.e. the two parts M S ,M C , corresponding auxiliary functions AF S , AF C ) when it departs, and after the UAV cluster is grouped according to the grouping algorithm, the groups are equally divided into a server group GS and a client group GC according to the node distribution, and the two groups train the models M S ,M C , respectively.

[0046] In particular, after a certain number of iteration rounds, the training models are exchanged between groups to maximize the use of local data sets of each node.

[0047] Step 2, model training, in a certain global round t, each UAV s∈GS treceive the model after the global aggregation of the previous round and the auxiliary function each drone c in the client group GC t receive the model after the global aggregation of the previous round and the auxiliary function and let

[0048]

[0049] The model training steps are as follows:

[0050] Step 21, each drone node c in the client group GC performs forward propagation on the model after the client model of the previous round and passes the intermediate data obtained by training to the auxiliary function by performing simple back propagation. Based on the local loss function, the parameters are updated to obtain and

[0051]

[0052] where η t is the learning rate of the t-th round.

[0053] Step 22, each drone node s in the server group GS performs the auxiliary function performing simple forward propagation, passing the output training data to the server model performing back propagation and updating the parameters to obtain and

[0054]

[0055] where η t is the learning rate of the t-th round.

[0056] Step 23, the GC and GS groups process forward propagation and back propagation in parallel, and at the end of each iteration, exchange their respective parameters to optimize the auxiliary function and the model, and distribute the optimized model and auxiliary function parameters.

[0057] Taking the GC group as an example, the optimized model and auxiliary function are updated to

[0058]

[0059] Step 24, repeat the process of step 21-step 23, GC and GS two groups, each containing a plurality of small groups. After each node trains a given model locally, it will send the trained model parameters to the temporary group leader of the group to which it belongs, and the group leader is responsible for the next step of parameter transmission. Each temporary group leader will select a grandmaster according to certain group leader selection criteria, responsible for aggregating the model parameters of each group node, and updating the new global model parameters to update the group leader, and the group leader will issue each node. By adopting such a hierarchical federated learning structure, the communication burden of the group leader node as the parameter aggregation and forwarding can be reduced, the node endurance time can be prolonged, and the influence of the node on the cluster task cooperation can be avoided.

[0060] Step 3, group leader selection, the selection of large group leader node is to select a high reputation node, which is responsible for aggregating and distributing parameters in federated learning. The calculation formula of reputation T is as follows:

[0061] T = a1C1 + a2C2 + a3C3,

[0062] Wherein, the meanings of each parameter are as follows:

[0063] C1: the degree of approximation of the parameters uploaded by the node to the updated parameters after aggregation;

[0064] C2: the difference factor of the node dataset;

[0065] C3: the state of the node itself (including power, coordinates, etc.);

[0066] a1, a2, a3: weighted parameters, satisfying a1+a2+a3=1, the value is set according to the influence degree of the parameter on the reputation.

[0067] When initially determining the reputation value, the data set index C2 collected by the unmanned aerial vehicle node and the weight C3 of the state are mainly considered. At the same time, the reputation is updated in the iteration process, the selection of the group leader is optimized according to the change of the node reputation T, and the change and maintenance of the group leader are determined according to the needs.

[0068] Step 4, auxiliary function, the auxiliary function uses an autoencoder. The autoencoder is a kind of neural network used in semi-supervised learning and unsupervised learning, which can extract the potential features (encoding) of the input data, and use it to reconstruct the output result (decoding) with the same size as the input data and with the characteristic features. Therefore, the autoencoder can usually be used as a feature extraction module in the neural network to improve the information extraction ability of the network to the input picture.

[0069] The main components of the autoencoder are three: the encoder, the latent feature representation module and the decoder. The input image is converted into an intermediate tensor vector representing the latent features by the encoder. The feature information embodied in the vector is extracted by the network of the encoder and can be learned and used by the decoder. The intermediate vector is input into the decoder module to obtain an output image of the same dimension and size as the input image. This process is similar to image reconstruction or feature repair, and can also be regarded as image compression and decompression. In addition, the feature dimension of the output image can be modified to become a reconstructed image that is more conducive to neural network learning features.

[0070] In the present application, more is used to realize the same dimension and size of input and output to supplement the missing of the two-part model after the selected model of federated learning is split.

[0071] Figure 1 The algorithm architecture in the present application, each unmanned aerial vehicle carries a split model and an auxiliary function. The complete model is split into two parts, each corresponding to an auxiliary function. After splitting the model, the corresponding auxiliary function is used to realize the parallel training of the two groups of models during training. Each unmanned aerial vehicle can realize the splicing of the split model and the auxiliary function, and train the model using the local data set, while maintaining the global parameter aggregation mechanism of federated learning in the global iteration process to speed up the convergence depth of the model.

[0072] Figure 2 The split federated learning parallel training flowchart is shown in the figure. For forward propagation, the crushed data is transmitted to realize the connection of the model and the auxiliary function. For backward propagation, the gradient of the crushed data is transmitted to realize the update of the model and the auxiliary function. During model training, the GC group transmits the model M C The output crushed data is transmitted to the auxiliary function AF C to obtain highly characteristic data, which is directly subjected to image classification or target detection; the data set collected by the GS group is first transmitted to the auxiliary function AF S to obtain intermediate features, which are input into the M S model to obtain the model result. Thus, both end models can perform complete forward propagation and backward propagation, and the computational load is balanced. More importantly, true parallel training of the two groups can be realized.

[0073] In summary, according to the algorithm of the present application, by applying the federated learning of P2P to the unmanned aerial vehicle cluster, using the combination of federated learning and split learning, and introducing auxiliary functions in the split model, the accuracy and invulnerability of cluster task cooperation can be improved. In practical applications, the algorithm of the present application has high accuracy, low device requirements and small training cost.

[0074] The above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalent replaced without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for collaborative tasks in a drone swarm, characterized in that, Includes the following steps: Step 1: Model Splitting. The model is split into two parts and distributed to different groups of drones for training. Corresponding auxiliary functions are added to the split models. Each drone node in the cluster has pre-stored a complete initial model upon departure. The two parts of the split network model are M... S and M C The corresponding auxiliary function is AF. S and AF C After the drone swarm is grouped according to the grouping algorithm, the groups are divided into a server group (GS) and a client group (GC) based on the node distribution. The two groups respectively enable and train model M. S and M C ; Step 2, Model Training: In a specific global round, each UAV receives the model and auxiliary functions from the previous round of global aggregation; specifically, in a specific global round t, each UAV s∈GS in the server group GS... t Received the model after the previous round of global aggregation and auxiliary functions Each drone c∈GC in the client group GC t Received the model after the previous round of global aggregation and auxiliary functions And let Step 3: Group leader selection. The selection of group leader nodes involves selecting high-reputation nodes, which are responsible for aggregating and distributing parameters in federated learning. Step 4: Auxiliary function. The autoencoder extracts the latent features of the input data and uses them to reconstruct the output results with the same size as the input data and with representative features.

2. The UAV swarm mission collaboration method according to claim 1, characterized in that: In step 1, after a certain number of iterations, the training models are exchanged between groups to maximize the use of the local datasets of each node.

3. The UAV swarm mission collaboration method according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 21: Each drone node c in the client group GC is in the previous round of client model. The forward propagation is performed in parallel, passing the intermediate data obtained from its training to the auxiliary function. Depend on Perform a simple backpropagation; Step 22: Each UAV node s in the server group GS executes the auxiliary function. Perform a simple forward propagation to pass the output training data to the server-side model. Perform backpropagation; Step 23: The client group GC and the server group GS process forward propagation and backward propagation in parallel, and at the end of each iteration, they exchange their parameters to optimize the auxiliary function and model, and then distribute the optimized model and auxiliary function parameters. Step 24: Repeat steps 21-23.

4. The UAV swarm mission collaboration method according to claim 3, characterized in that: In step 21, the parameters are updated based on the local loss function to obtain... and Where, η t Let be the learning rate in round t.

5. The UAV swarm mission collaboration method according to claim 3, characterized in that: In step 22, the updated parameters are obtained. and Where, η t Let be the learning rate in round t.

6. The UAV swarm mission collaboration method according to claim 3, characterized in that: In step 23, the client group's GC-optimized model and helper functions will be updated to...

7. The UAV swarm mission collaboration method according to claim 1, characterized in that: In step 3, the formula for calculating reputation T is as follows: T = α1C1 + α2C2 + α3C3, The meanings of each parameter are as follows: C1: The degree of similarity between the parameters uploaded by the node and the parameters updated after aggregation; C2: Factors contributing to the differences in node datasets; C3: The node's own state; α1, α2, α3: Weighting parameters that satisfy α1 + α2 + α3 = 1. The values ​​are set according to the degree of influence of the parameters on reputation.

8. The UAV swarm mission collaboration method according to claim 1, characterized in that: In step 4, the auxiliary function employs an autoencoder, which includes an encoder, a latent feature representation module, and a decoder.

Images