Federal learning method, device and equipment for multi-base-station unmanned vehicle

Abstract

The invention discloses a federated learning method, device and equipment for a multi-base-station unmanned vehicle. The method comprises the following steps: 1, constructing a base station set and an unmanned vehicle set; 2, constructing a first edge model and an initial global network model; 3, broadcasting the first edge model to each unmanned vehicle; 4, training each local model of each unmanned vehicle to obtain updating parameters of each local model; 5, compressing the local model update parameters, and uploading the compressed local model update parameters to the corresponding base stations for aggregation to obtain a second edge model; and 6, replacing the first edge model with the second edge model, returning to the third step, repeatedly executing the third step to the fifth step for the first time, and the like, during repeated execution for the bth time, if b is equal to 2, 3,..., E-1, correspondingly replacing the bth edge model with the (b + 1) th edge model, returning to the third step until the target edge models are obtained, and aggregating the target edge models to obtain the target edge models. Therefore, the target global network model with high performance is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of mobile communications, and in particular, to a federated learning method, apparatus, and device for multi-base station unmanned vehicles. Background Art

[0002] As a new machine learning paradigm, federated learning can support multiple distributed data holders to collaboratively train a specific neural network model. Compared with the traditional centralized model training method, in federated learning, the local original data does not need to leave the local storage of each distributed party. Therefore, the risk of data leakage can be greatly reduced, and the rights and interests of each party can be better protected. In addition, in the traditional centralized model training method, uploading the corresponding original data of each distributed party to the central server will consume a large amount of wireless spectrum resources and cause serious network delays, and this problem does not exist in federated learning. Moreover, in federated learning, the training of the network model is carried out in parallel at each party, and no longer overly relies on the computing and storage capabilities of the central server. Therefore, the computing efficiency is also greatly improved. It can be seen that such a scenario where multiple distributed participants work together efficiently is very suitable for the collaborative and efficient operation of multiple distributed unmanned vehicles guaranteed by cellular base stations in an intelligent transportation system. Based on this, in recent years, research on federated learning related to connected unmanned vehicles has emerged continuously. However, to effectively ensure that connected unmanned vehicles can bring the expected benefits, there are still many inadaptations in the existing work.

[0003] Specifically, considering that the model parameters need to be frequently exchanged between the unmanned vehicle and the base station during the federated learning process, in order to reduce the restriction of the scarce wireless spectrum resources in reality, various model compression and sparsification methods have been widely studied. However, the existing related research work does not fully consider the inherent statistical characteristics of the federated learning process. Therefore, the model compression effect still needs to be improved. In addition, in order to introduce more diverse sample data into the federated learning process and thus obtain a high-performance federated learning model, related research has proposed a federated learning framework across multiple base stations, but it is the remote cloud that acts as the parameter server to aggregate the edge models formed by multiple base stations. As we know, different from manned vehicles, unmanned vehicles have extremely high requirements for real-time guidance and perception. Therefore, the latency generated by the model parameters traveling to and from the remote cloud will be intolerable for the operation of connected unmanned vehicles in practice.

[0004] Furthermore, the existing federated learning mechanisms are mainly divided into synchronous federated learning and asynchronous federated learning. During the implementation of synchronous federated learning, it is necessary to wait for all the autonomous vehicles to complete the training of the local model and upload the corresponding local model update parameters before performing global model aggregation. However, due to the differences in computing and storage capabilities and the number of samples in the local dataset among autonomous vehicles, the time required to complete one local model training varies. If synchronous federated learning is applied, it is necessary to wait for the autonomous vehicle that takes the longest time to complete the local model training and upload the corresponding local model update parameters. In this way, the real-time update ability of the trained model will be seriously affected. Furthermore, considering the complex driving environment and wireless transmission characteristics, it may be impossible to collect the local model update parameters from all autonomous vehicles for a long time in reality. Different from synchronous federated learning, in fully asynchronous federated learning, as long as any autonomous vehicle completes the training of the local model and uploads the corresponding local model update parameters, the parameter server will perform a global model aggregation. Obviously, the already scarce wireless spectrum resources are difficult to effectively support such frequent model parameter exchanges in fully asynchronous federated learning. Summary of the Invention

[0005] The object of the present invention is to provide a federated learning method, device and equipment for multi-base-station autonomous vehicles, to solve the straggler problem that may be caused by vehicle failures, traffic accidents, extreme weather, etc. It mainly includes how to integrate the advantages of synchronous federated learning and fully asynchronous federated learning, construct a multi-base-station autonomous vehicle federated learning framework, how to avoid the negative impact of out-of-date local model updates and autonomous vehicles with too few local dataset samples on the performance of the trained model, and how to fully exploit the inherent statistical characteristics of federated learning to reconcile the contradiction between the frequent model parameter exchanges and the scarce wireless spectrum resources during the federated learning process.

[0006] To solve the above technical problems, a technical solution adopted by the present invention is to provide a federated learning method for multi-base-station unmanned vehicles. The method includes the following steps: First, construct a base station set and an unmanned vehicle set. The base station set includes M base stations, and the unmanned vehicle set includes N unmanned vehicles. Second, based on a hybrid aggregation strategy, construct M first edge models and an initial global network model. The M first edge models correspond to the M base stations one by one, and the initial global network model is formed by aggregating the M first edge models. Third, broadcast the M first edge models to each unmanned vehicle in the unmanned vehicle set. Fourth, use the M first edge models to train the respective local models of each unmanned vehicle to obtain the updated parameters of each local model. Fifth, compress the updated parameters of each local model and upload them to the corresponding base station for edge model aggregation to obtain M second edge models. Sixth, use the M second edge models to replace the M first edge models, return to the third step, and repeat the execution of the third to fifth steps for the first time. By analogy, when repeating the execution for the bth time, b = 2, 3,..., E - 1, correspondingly use M (b + 1)th edge models to replace the M bth edge models and return to the third step for broadcasting to each unmanned vehicle in the unmanned vehicle set until M target edge models are obtained. Then, perform global model aggregation on the M target edge models to obtain a target global network model.

[0007] The present invention also provides a federated learning device for a multi-base station unmanned vehicle. The device includes: a set construction unit for constructing a base station set and an unmanned vehicle set, where the base station set includes M base stations and the unmanned vehicle set includes N unmanned vehicles; a model construction unit for constructing M first edge models and an initial global network model based on a hybrid aggregation strategy; the M first edge models respectively correspond to the M base stations one by one, and the initial global network model is formed by aggregating the M first edge models; a distribution unit for broadcasting the M first edge models to each unmanned vehicle in the unmanned vehicle set; a model training unit for using the M first edge models to respectively train each local model of each unmanned vehicle to obtain update parameters of each local model; a model update unit for compressing the update parameters of each local model and uploading them to the corresponding base station for edge model aggregation to obtain M second edge models; the model update unit is further configured to replace the M first edge models with the M second edge models, so that the distribution unit, the model training unit, and the model update unit are further configured to repeat the corresponding operations for the first time according to the M second edge models, and so on. When the distribution unit, the model training unit, and the model update unit repeat the corresponding operations for the b-th time, b = 2, 3,..., E - 1, correspondingly, the model update unit is further configured to replace the M b-th edge models in the distribution unit with M (b + 1)-th edge models, and the distribution unit broadcasts the M (b + 1)-th edge models to each unmanned vehicle in the unmanned vehicle set until M target edge models are obtained, and then aggregates the M target edge models to obtain a target global network model.

[0008] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above method are implemented.

[0009] The beneficial effects of the present invention are as follows: The present invention discloses a federated learning method, device, and equipment for multi-base station unmanned vehicles. The method includes the following steps: First, construct a base station set and an unmanned vehicle set. The base station set includes M base stations, and the unmanned vehicle set includes N unmanned vehicles. Second, based on a hybrid aggregation strategy, construct M first edge models and an initial global network model. The M first edge models correspond to the M base stations one by one, and the initial global network model is formed by aggregating the M first edge models. Third, broadcast the M first edge models to each unmanned vehicle in the unmanned vehicle set. Fourth, use the M first edge models to train each local model of each unmanned vehicle respectively to obtain update parameters for each local model. Fifth, after compressing the update parameters of each local model respectively, upload them to the corresponding base station for edge model aggregation to obtain M second edge models. Sixth, use the M second edge models to replace the M first edge models, and return to the third step. When repeating the execution of the third to fifth steps for the first time, and so on, when repeating the execution for the b-th time, b = 2, 3,..., E - 1, correspondingly use the M (b + 1)-th edge models to replace the M b-th edge models and return to the third step for broadcasting to each unmanned vehicle in the unmanned vehicle set. Until M target edge models are obtained, perform global model aggregation on the M target edge models to obtain a target global network model. This method can effectively address the differences in computing and storage capabilities and local dataset sample data among multiple unmanned vehicles within a single base station, and can also solve the problem of stragglers that may be caused by vehicle failures, traffic accidents, extreme weather, etc. in reality, as well as the overly frequent model parameter exchanges in fully asynchronous federated learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 is the overall flowchart of a federated learning method for multi-base station unmanned vehicles according to the present invention;

[0011] Figure 2 is the data flow diagram of a federated learning method for multi-base station unmanned vehicles according to the present invention;

[0012] Figure 3 is the schematic diagram of the principle of a federated learning method for multi-base station unmanned vehicles according to the present invention;

[0013] Figure 4 is the specific flowchart of step S3 in a federated learning method for multi-base station unmanned vehicles according to the present invention;

[0014] Figure 5 is the specific flowchart of step S5 in a federated learning method for multi-base station unmanned vehicles according to the present invention;

[0015] Figure 6It is a comparison result graph of the training error of a federated learning method for multi-base-station unmanned vehicles in the present invention compared with synchronous federated learning and fully asynchronous federated learning;

[0016] Figure 7 It is a comparison result graph of the training accuracy of a federated learning method for multi-base-station unmanned vehicles in the present invention compared with synchronous federated learning and fully asynchronous federated learning;

[0017] Figure 8 It is a weight design strategy in a federated learning method for multi-base-station unmanned vehicles in the present invention, and it is a comparison result graph of the training error compared with the weight design strategy based only on the number of samples and only on the parameter age;

[0018] Figure 9 It is a comparison result graph of the training error when different values are taken for the time period in a federated learning method for multi-base-station unmanned vehicles in the present invention;

[0019] Figure 10 It is a comparison result graph of the training accuracy when different values are taken for the time period in a federated learning method for multi-base-station unmanned vehicles in the present invention;

[0020] Figure 11 It is a comparison result graph of the training error between the random distribution and the uniform distribution of unmanned vehicles in a federated learning method for multi-base-station unmanned vehicles in the present invention;

[0021] Figure 12 It is a comparison result graph of the training accuracy between the random distribution and the uniform distribution of unmanned vehicles in a federated learning method for multi-base-station unmanned vehicles in the present invention;

[0022] Figure 13 It is a block diagram of the composition of a federated learning device for multi-base-station unmanned vehicles in the present invention. Detailed implementation manners

[0023] To facilitate the understanding of the present invention, the present invention will be described in more detail below in conjunction with the accompanying drawings and specific embodiments. The preferred embodiments of the present invention are shown in the accompanying drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure content of the present invention more thorough and comprehensive.

[0024] It should be noted that unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not used to limit the present invention. The term "and / or" used in this specification includes any and all combinations of one or more of the related listed items.

[0025] As shown Figure 1 in the figure, a federated learning method for a multi-base station unmanned vehicle according to the present invention is shown, including the following steps:

[0026] Step S1: Construct a base station set and an unmanned vehicle set. The base station set includes M base stations, and the unmanned vehicle set includes N unmanned vehicles.

[0027] Step S2: Based on a hybrid aggregation strategy, construct M first-edge models and an initial global network model; the M first-edge models correspond to the M base stations one by one, and the initial global network model is formed by aggregating the M first-edge models.

[0028] Step S3: Broadcast the M first-edge models to each unmanned vehicle in the unmanned vehicle set.

[0029] Step S4: Use the M first-edge models to train each local model of each unmanned vehicle respectively to obtain update parameters of each local model.

[0030] Step S5: After compressing the update parameters of each local model respectively, upload them to the corresponding base station for edge model aggregation to obtain M second-edge models.

[0031] Step S6: Use the M second-edge models to replace the M first-edge models, and return to Step S3. For the first time, repeat Steps S3 to S5. By analogy, when repeating for the bth time, b = 2, 3,..., E - 1, correspondingly use the M (b + 1)-th edge models to replace the M b-th edge models and return to Step S3 for broadcasting to each unmanned vehicle in the unmanned vehicle set until M target edge models are obtained, and then perform global model aggregation on the M target edge models to obtain a target global network model.

[0032] For the convenience of description, hereinafter, the first-edge model, the b-th edge model, and the (b + 1)-th edge model are uniformly referred to as the z-th edge model. Wherein, z = 1, 2, 3,..., E.

[0033] In this application, by constructing a base station set and an unmanned vehicle set, it is convenient to associate each unmanned vehicle with the corresponding base station; and by constructing M first-edge models and an initial global network model, each first-edge model can be associated with the corresponding base station. Further, after each unmanned vehicle in the unmanned vehicle set obtains the relevant parameters of the corresponding first-edge model, each unmanned vehicle asynchronously trains its corresponding local model to obtain the updated parameters of the corresponding local model, and completes the aggregation of the edge models within a single base station at the network edge to obtain the second-edge model corresponding to each base station; then, according to step S6, edge iteration is performed. After obtaining the target edge models corresponding to each base station, global network model aggregation among multiple base stations is completed at the network edge, thereby constructing a target global network model applicable to multi-unmanned vehicle collaboration. The present invention focuses on the hot issue of multi-unmanned vehicle collaboration, proposes a highly practical and effective solution, and has good application prospects. This federated learning method can reconcile the realistic contradiction between the frequent model parameter exchanges in the federated learning process and the scarce radio spectrum resources, and can ensure the good performance of the obtained target global network model while solving the communication bottleneck problem.

[0034] It should be noted that each time the operations of step S3 to step S5 are executed, that is, an edge iteration is performed on the z-th edge model to obtain the (z + 1)-th edge model. Until the last edge iteration, the target edge models corresponding to each base station can be obtained.

[0035] The following combines Figures 1 to 8 , and further details the present application with specific embodiments.

[0036] Figure 1 FIG. shows the schematic flowchart of the federated learning method for multi-base station unmanned vehicles provided by the embodiment of the present application, which is described in detail as follows:

[0037] Step S1: Construct a base station set and an unmanned vehicle set. The base station set includes M base stations, and the unmanned vehicle set includes N unmanned vehicles.

[0038] It should be noted that each base station can correspond to multiple unmanned vehicles, while each unmanned vehicle can only correspond to one base station, and each unmanned vehicle is also correspondingly provided with a local model. It is worth noting that in reality, within the coverage area of a base station, there may be multiple unmanned vehicles, and an unmanned vehicle may also be within the coverage areas of multiple base stations. However, in this application, each unmanned vehicle has only one associated base station. That is to say, each unmanned vehicle can only upload the updated parameters of the corresponding local model to the associated base station.

[0039] In this application, it is defined that and respectively represent the base station set composed of M base stations and the unmanned vehicle set composed of N unmanned vehicles.

[0040] Furthermore, the local dataset corresponding to the local model of the driverless vehicle is where |D n | represents the number of samples in the local dataset D n , X i represents the i-th data sample, and Y i is the label output corresponding to the i-th data sample X i .

[0041] Secondly, it is set that the driverless vehicle has one and only one associated base station and the multiple driverless vehicles associated with the base station (including the driverless vehicle ) can be defined as the m-th subset of driverless vehicles, and the m-th subset of driverless vehicles can be expressed as

[0042] It should be noted that, to serve emerging in-vehicle applications, N driverless vehicles spanning M base stations need to jointly train a neural network model determined by a d-dimensional parameter vector .

[0043] Step S2: Based on the hybrid aggregation strategy, construct M first-edge models and an initial global network model; the M first-edge models correspond to the M base stations one by one, and the initial global network model is formed by aggregating the M first-edge models.

[0044] It should be noted that the hybrid aggregation strategy includes the federated learning strategy based on periodic aggregation and the synchronous federated learning strategy. Since each base station is associated with multiple driverless vehicles, and each driverless vehicle is only associated with one base station. Therefore, multiple driverless vehicles associated within a single base station, based on the federated learning strategy of periodic aggregation, aggregate the local models corresponding to the multiple driverless vehicles to form the edge model corresponding to this base station; while between multiple base stations, based on the synchronous federated learning strategy, the edge models corresponding to each of them are aggregated to form the global network model.

[0045] Specifically, as Figure 4 shown, step S2 includes the following sub-steps

[0046] Step S21: Perform edge aggregation on the local models of multiple driverless vehicles associated with each of the M base stations to obtain M first-edge models.

[0047] In this application, for each base station in the base station set, each base station has multiple associated driverless vehicles.

[0048] Among them, based on the federated learning strategy of periodic aggregation, the driverless vehicles associated with the base station Aggregate the local models corresponding to multiple associated driverless vehicles to form a base station A corresponding first edge model.

[0049] Step S22: Perform global model aggregation on the M first edge models to obtain an initial global network model.

[0050] In this application, based on the synchronous federated learning strategy, globally aggregate the first edge models corresponding to each of the M base stations to form an initial global network model.

[0051] It should be noted that the initial global network model is a neural network model determined by a d-dimensional parameter vector Subsequent steps are mainly for training this model to obtain the target global network model actually needed.

[0052] In practical applications, to obtain the target global network model, during the process of training the initial global network model, the corresponding loss function can be expressed as:

[0053]

[0054] Among them, w represents the global network model used in one global iteration; l(w, X i ) represents the loss function value calculated based on the i-th data sample X i ; D = ∪ n∈N D n represents the total amount of sample data from N driverless vehicles, corresponding to a size of |D|.

[0055] In the existing federated learning paradigm, the local data sets of each driverless vehicle are private, and the network model training process is carried out locally based on its own computing and storage capabilities, and then the updated parameters of the trained local model are uploaded to the parameter server for aggregation. Based on this, for the driverless vehicle , the local target loss function it uses can be expressed as:

[0056]

[0057] Among them, w represents the local model corresponding to the driverless vehicle in one edge iteration.

[0058] In subsequent edge iterations, after each driverless vehicle completes local model training using the corresponding local target loss function, it uploads the updated parameters of the corresponding local model to the associated base station via a wireless link for edge model aggregation, thereby obtaining a new edge model corresponding to each of the M base stations.

[0059] Step S3: Broadcast the M first edge models to each driverless vehicle in the set of driverless vehicles.

[0060] Combine Figure 2 and Figure 3 Before the first edge iteration starts, the parameter server 1 broadcasts the initial global network model w 0 down to M base stations 2 in the base station set via the downlink, where the parameter server 1 is served by one of the M base stations 2. Then, each of the M base stations 2 broadcasts its corresponding first edge model down to the corresponding subset of unmanned vehicles, that is, broadcasts the M first edge models to each unmanned vehicle in the unmanned vehicle set .

[0061] It should be noted that since the downlink has sufficient channel bandwidth, the transmission rate is usually high. Therefore, the time delay of the parameter server 1 broadcasting the initial global network model w 0 can be ignored.

[0062] Step S4: Train each local model of each unmanned vehicle to obtain the updated parameters of each local model.

[0063] As Figure 2 shown, for each unmanned vehicle, if it completes the local model training within a fixed time period, it uploads the corresponding local model updated parameters to the associated base station 2 via the wireless link, otherwise it continues the local model training without being disturbed. In this way, it can effectively cope with the differences in the computing and storage capabilities and local dataset sample data of multiple unmanned vehicles in a single base station 2, and can also solve the problem of stragglers that may be caused by vehicle failures, traffic accidents, extreme weather, etc. in reality, as well as the overly frequent model parameter exchange in fully asynchronous federated learning.

[0064] It should be noted that due to the differences in computing processing capabilities and the number of local dataset samples, etc., the time required for each unmanned vehicle to complete one local model training is not the same.

[0065] In this embodiment, it is set that the time required for the unmanned vehicle to complete one local model training is t n ∈[t min , t max , where t min and t max are respectively the shortest and longest times required for all unmanned vehicles to complete one local model training.

[0066] Combine Figure 2 and Figure 3, within each time period T, it is set that the parameter server 1 performs edge model aggregation once (that is, performs one edge iteration, which is also to perform the specific operations from step S4 to step S5 to obtain a new edge model), and all unmanned vehicles need to perform L local iterations to complete one local model training.

[0067] In this embodiment, when each unmanned vehicle in the set receives the first edge model , it starts the first edge iteration.

[0068] Among them, the local models corresponding to each unmanned vehicle respectively start local model training to obtain the local model update parameters correspondingly. Then, after compressing the local model update parameters, they are uploaded to their respective corresponding base stations. Each base station performs edge model aggregation based on the multiple local model update parameters received to obtain the corresponding second edge model.

[0069] It should be noted that for the m-th subset of unmanned vehicles associated with the base station if the unmanned vehicles in the m-th subset of unmanned vehicles are trained based on synchronous federated learning, it is necessary to wait for all unmanned vehicles to complete local model training and upload the corresponding local model updates before edge model aggregation can be performed. However, since the communication between the unmanned vehicle and the base station is through a wireless link, there will be many uncertain and unpredictable factors. In reality, vehicle failures, traffic accidents, and weather and other reasons are very likely to cause some vehicles to be unable to complete local model training and upload the corresponding local model updates to the associated base station in a short time. In addition, the too frequent model parameter exchange in fully asynchronous federated learning is also not applicable to the multiple unmanned vehicles in the m-th subset of unmanned vehicles

[0070] to cooperate in training.

[0071] Based on this, aiming at the characteristics of multiple unmanned vehicles in a single base station, the present invention proposes a federated learning method based on periodic aggregation, which allows the multiple unmanned vehicles associated with a single base station to perform local model training asynchronously.

[0072] Specifically, it is set that the base station performs edge model aggregation once every fixed time period T. If the unmanned vehicles in the m-th subset of unmanned vehicles complete local model training within the time period T, they upload the corresponding local model update parameters to the base station ; otherwise, they continue local model training without being disturbed.

[0073] ​In this embodiment, it is set that M base stations all need to complete E edge iterations, and edge model aggregation is performed once every time period T, and finally M target edge models are obtained; after obtaining the M target edge models, the M base stations upload them to the parameter server 1 for subsequent global model aggregation.

[0074] As described above, to obtain the target edge model, all base stations need to complete E edge iterations. In this way, after T×E time, all base stations will obtain the target edge model simultaneously, that is, there is no possible straggler problem. In addition, the communication between base stations is through stable and reliable wired transmission. Therefore, multiple base stations can be trained based on synchronous federated learning. In summary, in the federated learning method considered in the present invention, multiple unmanned vehicles associated with a single base station are trained collaboratively based on periodic aggregation federated learning to obtain their respective corresponding target edge models, and multiple base stations are trained based on synchronous federated learning to obtain the target global network model.

[0075] In this embodiment, in the z-th edge iteration, the unmanned vehicle first performs local model training based on its own local dataset D n If the training is completed within the time period T, then the corresponding local model update is calculated Otherwise, continue local model training without being disturbed. Among them, represents the local model update parameter corresponding to the unmanned vehicle in the z-th edge iteration, and the calculation method is:

[0076]

[0077] Among them, is the updated local model obtained by the unmanned vehicle in the z-th edge iteration based on the local dataset D n and its own computing and storage resources, and E is the total number of edge iterations.

[0078] Step S5: Compress each local model update parameter respectively and upload it to the corresponding base station for edge model aggregation to obtain M second edge models.

[0079] It should be noted that the communication between the driverless vehicle and the base station is carried out through wireless transmission. Considering the scarce wireless spectrum resources in reality and the frequent model parameter exchange between the driverless vehicle and the base station during the federated learning process, before each driverless vehicle uploads the d-dimensional local model update parameters to the associated base station, the present invention method compresses them through a corresponding compression function to obtain each compressed parameter. Then, each compressed parameter is uploaded to the corresponding base station, and a decompression function is used to decompress each compression function to obtain each parameter approximation. Finally, M base stations respectively perform weighted averaging on the multiple parameter approximations corresponding to them, so as to obtain M second edge models.

[0080] Essentially, the compression processing method proposed by the present invention is based on the inherent statistical characteristics of the federated learning process, which is mainly reflected in two aspects. On the one hand, the change of model parameters during the local model training process is slow and continuous rather than abrupt. Therefore, the local model update parameters, as the difference between the edge model parameters corresponding to two adjacent edge iterations, can be compressed, or can be compressed after sparse transformation. On the other hand, the local model update parameters of two adjacent times are similar, that is to say, the current local model update parameters can be mapped to the subspace where the previous local model update parameters are located for compression.

[0081] Specifically, before the local model training corresponding to each driverless vehicle in the m-th subset of driverless vehicles starts, the base station will use a same shared matrix with all the driverless vehicles in the m-th subset of driverless vehicles Each element in the shared matrix Q independently takes values from the same Gaussian distribution and each row is normalized; where represents a Gaussian distribution with a mean of 0 and a variance of 1 / d. In this way, each row in the shared matrix Q is randomly distributed on the surface of the d-dimensional unit sphere, and the mean of all rows is 0. In addition, any two rows of the shared matrix Q are uncorrelated. That is to say, any proportion of row vectors extracted from the shared matrix Q will form an orthogonal basis.

[0082] As described above, in the z-th edge iteration, before the driverless vehicle n uploads the d-dimensional local model update parameters , it is necessary to compress them into -dimensional vectors through a compression function , that is, to obtain the compressed parameter

[0083] In practice, is determined by the uplink communication resources available to the driverless vehicle n and the corresponding wireless transmission channel state. That is to say, the original local model update parameters There are d parameters in it. However, in reality, due to the limitations of the uplink communication resources available to the driverless vehicle n and the corresponding constraints of the wireless transmission channel state, only the local model update parameters can be compressed into compressed parameters with only parameters for uploading.

[0084] Among them, the compression function is designed based on the corresponding compression matrix , and the compression matrix consists of rows selected from the shared matrix Q.

[0085] Specifically, it is assumed that the available wireless channel transmission bandwidth of the base station is B m , and it is assumed that the uplink transmission adopts orthogonal frequency division multiple access technology. Each driverless vehicle in equally enjoys the available wireless spectrum resources of the base station . That is to say, the process of multiple driverless vehicles in uploading the network model parameters to the base station m through the wireless link is non-interfering, and the available wireless transmission channel bandwidth of each driverless vehicle is where

[0086] is the number of driverless vehicles in . n , the additive noise in the actual transmission process satisfies the distribution CN(0,σ 2 ), and the corresponding uplink channel transmission gain . In this way, in the z-th edge iteration, for the driverless vehicle , the corresponding wireless transmission channel rate is:

[0087]

[0088] Furthermore, it is assumed that in the z-th edge iteration, the delay constraint for the driverless vehicle to upload the network model parameters to the base station m is , then the driverless vehicle can upload at most bits to the base station m. Among them, is calculated as follows:

[0089]

[0090] Referring to the existing federated learning methods, it is assumed that in the federated learning process, it takes 32 bits to upload a network model parameter via a wireless link. Based on this, the number of parameters that the autonomous vehicle n can upload in the z-th edge iteration can be obtained as follows:

[0091]

[0092] Furthermore, define as the set of indices of the selected rows. Based on the determined index set the compression matrix can be obtained as:

[0093]

[0094] Next, the compression function can be designed as:

[0095]

[0096] As described above, before the autonomous vehicle n uploads the local model update parameter it will be compressed by the compression function In this way, based on the compression function determined in the above formula, the compressed version of the original local model update parameter i.e., the compressed parameter actually transmitted wirelessly can be obtained as:

[0097]

[0098] Furthermore, corresponding to the compression function the decompression function can be designed as:

[0099]

[0100] As described above, after the associated base station receives the compressed parameter uploaded by the autonomous vehicle n, it will be decompressed by a decompression function and the corresponding parameter approximation can be obtained as:

[0101]

[0102] where is the parameter approximation after the local model update parameter has gone through the compression and decompression processes.

[0103] It can be found that as long as the compression matrix and the compression function Decompression and compression function The solution will come naturally. Therefore, the difficulty in constructing the proposed model compression processing method based on the intrinsic statistical characteristics of federated learning lies in how to select rows from the shared matrix Q to form the compression matrix

[0104] In the actual federated learning process, after the z-th edge iteration is completed, the base station m will broadcast the latest aggregated edge model and the calculated index set to each unmanned vehicle in. In this way, when each unmanned vehicle in participates in the next edge iteration, it can determine the corresponding compression matrix based on the index set

[0105] As for how to determine the index set In the present invention, it is mainly based on the similarity between the continuously updated local model parameters. That is to say, the current updated local model parameters can be mapped to the subspace where the previous updated local model parameters are located for model compression.

[0106] Specifically, in the proposed model compression method based on the intrinsic statistical characteristics of federated learning, the index set can be obtained by solving problem P 1 Problem P 1 can be expressed as:

[0107]

[0108] s.t.P = Q T (I, :)(Q(I, :)Q T (I, :)) -1 Q(I, :);

[0109] Among them, problem P 1 can be understood as how to select the index set I to maximize the norm on the dimensional subspace formed by the rows corresponding to the index set I in the shared matrix Q. Obviously, the solution of problem P 1 is NP-hard (non-deterministic polynomial). In practice, a sub-optimal solution of problem P 1 can also be obtained based on classical methods such as matching pursuit in the field of compressive sensing.

[0110] In the proposed model compression method based on the intrinsic statistical characteristics of federated learning, the base station with strong computing and storage capabilities is responsible for solving problem P 1 ​, the obtained index set I is also broadcast by the base station to each driverless vehicle. Therefore, for driverless vehicles with limited computing, storage, and communication resources, such model compression design will not increase their burden.

[0111] In this embodiment, before the driverless vehicle n uploads the d-dimensional local model update parameters to the associated base station, it will be compressed through the corresponding compression function and expressed as:

[0112]

[0113] where is the local model update parameter and the compressed parameter obtained after compression, and there is

[0114] Correspondingly, after the associated base station receives the uploaded by the driverless vehicle n, it will be decompressed through a decompression function to obtain the corresponding parameter approximation:

[0115]

[0116] where is the local model update parameter and the parameter approximation after compression and decompression.

[0117] As described above, represents the m-th subset of driverless vehicles associated with the base station . Since the local model training of each driverless vehicle is asynchronous, further, it is set that represents the subset of driverless vehicles that have completed local model training in the m-th subset of driverless vehicles at the z-th edge iteration. In this way, the weighted average of multiple parameter approximations associated with the base station can be obtained as:

[0118]

[0119] where is the weight corresponding to the parameter approximation associated with the base station .

[0120] Furthermore, it is set that represents the (z + 1)-th edge model of the base station m obtained in the z-th edge iteration, then there is:

[0121]

[0122] It should be noted that in the existing related work on federated learning, when the base station m performs edge model aggregation, the weights are often designed based on the size of the local dataset of the unmanned vehicle n, that is, the weights are designed as:

[0123]

[0124] where, is the number of samples in.

[0125] For the base station m, such a weight design during edge model aggregation can indeed well balance the number of samples of each unmanned vehicle in. However, as mentioned above, in the present invention, multiple unmanned vehicles associated with the same base station are co-trained based on the federated learning strategy of periodic aggregation. Therefore, although they are all associated with the same base station, the training of the local models may be based on different versions of the edge models. In this case, to avoid the updated parameters of the outdated local models from affecting the currently aggregated edge model, when designing the weights it is not only necessary to consider the number of samples of each unmanned vehicle in, but also the freshness of the updated parameters of the local models.

[0126] Based on this, to numerically represent the freshness of the updated parameters of the local models, its parameter age is further defined as follows:

[0127]

[0128] where, represents the last edge model aggregation participated by the unmanned vehicle n. In the present invention, the specific calculation method of is:

[0129]

[0130] Obviously, in the z-th edge iteration, the parameter age corresponding to the parameter approximation of the unmanned vehicle n can be understood as the time since its last participation in the edge model aggregation. The smaller the parameter age, the newer the parameter approximation , and correspondingly, the larger the parameter age means that the unmanned vehicle n has not participated in the edge model aggregation for a long time.

[0131] To avoid the situation where the outdated local model update parameters and the self-driving vehicles with a small number of local dataset samples overly affect the aggregation of the edge model, considering the number of local dataset samples of each self-driving vehicle and the parameter age of the parameter approximation values of the local model update parameters comprehensively, the weight design strategy is as follows:

[0132]

[0133] Among them, \(e\) is the natural logarithm. Obviously, in such a weight design strategy, the more local dataset samples a self-driving vehicle has, the smaller the corresponding parameter age, and the greater the weight when participating in the edge model aggregation. Thus far, the weight strategy corresponding to the single-base-station multi-self-driving vehicle federated learning based on periodic aggregation has been designed.

[0134] Specifically, combining Figure 3 and Figure 5 , step S5 includes the following sub-steps:

[0135] Step S51: Use each compression function to perform compression processing on each local model update parameter respectively to obtain each compressed parameter.

[0136] Step S52: Upload each compressed parameter to the corresponding base station.

[0137] Step S53: Use each decompression function to decompress each compressed parameter to obtain each parameter approximation value.

[0138] Step S54: Use the multiple parameter approximation values corresponding to each of the \(M\) base stations to perform edge model aggregation to obtain \(M\) second edge models.

[0139] In this embodiment, in the first edge iteration, through the above specific operations, the second edge models corresponding to each of the \(M\) base stations can be obtained. After that, in step S6, the \(M\) base stations can broadcast the second edge models corresponding to them to each self-driving vehicle to start the next round of edge iteration.

[0140] Step S6: Use the \(M\) second edge models to replace the \(M\) first edge models, return to step S3, and repeat step S3 to step S5 for the first time. And so on, when repeating for the \(b\)th time, \(b = 2, 3,\cdots, E - 1\), correspondingly use the \(M\) \((b + 1)\)th edge models to replace the \(M\) \(b\)th edge models and return to step S3 for broadcasting to each self-driving vehicle in the set of self-driving vehicles until \(M\) target edge models are obtained, and then perform global model aggregation on the \(M\) target edge models to obtain the target global network model.

[0141] In this embodiment, after obtaining the second edge model, the second edge iteration is performed, that is, the second edge model is used to replace the first edge model, and the specific operations of steps S3 to S5 are performed to obtain the third edge model; and so on, until after the E-th edge iteration, the (E + 1)-th edge model can be obtained. That is, the target edge model

[0142] Combined with Figure 2 , the following will focus on how to design the corresponding weight strategy when the parameter server 1 aggregates the edge models from multiple base stations 2.

[0143] Specifically, the base station After experiencing E edge iterations, the target edge model will be obtained And then upload it to the parameter server 1. After the parameter server 1 receives the target edge models uploaded by each base station 2 , it aggregates to form the target global network model as follows:

[0144]

[0145] As described above, in the formation process of each edge model, the local dataset sample quantity and parameter age corresponding to each unmanned vehicle have been fully considered. Therefore, in the present invention, each base station 2 has the same weight value when the global model is aggregated.

[0146] In some embodiments, after obtaining the corresponding M edge models for each edge iteration, a global iteration is performed to obtain a corresponding global network model. For example, after performing the z-th edge iteration to obtain the corresponding M (z + 1)-th edge models, the z-th global iteration is performed. The M (z + 1)-th edge models obtained are aggregated for the global model to obtain a (z + 1)-th global network model. Then, enter the (z + 1)-th edge iteration. The parameter server broadcasts the (z + 1)-th global network model to each base station immediately, and then uses the M (z + 1)-th edge models to replace the M z-th edge models and return to step S3, and repeat steps S3 to S5. Then, perform the (z + 1)-th global iteration, and so on, until the target global network model that meets the actual needs is obtained.

[0147] In this application, the acquisition of the target global network model comprehensively considers the unmanned vehicles from multiple base stations, and the aggregation of the edge model and the global network model is completed at the network edge. Therefore, a high-performance network model can be provided in real time online to serve emerging in-vehicle applications such as online perception of the driving environment, real-time prediction of traffic flow, and early warning of vehicle accidents.

[0148] Furthermore, the present invention compares the training effects of different federated learning mechanisms based on the publicly available MNIST dataset. During the simulation process, the MNIST dataset is divided into a training dataset and a test dataset. Among them, the training dataset contains 60,000 samples, and the test dataset contains 10,000 samples.

[0149] In addition, in order to fairly and objectively compare the performance of different federated learning mechanisms, it is set that the available wireless communication resources within a fixed time in all federated learning mechanisms are the same. More specifically, as described above, in the present invention, multiple unmanned vehicles within a single base station are trained based on federated learning with periodic aggregation. It is assumed that the period is set to T = t max / 16, where t max is the longest time required to complete one local model training among all unmanned vehicles.

[0150] In this way, for the federated learning method proposed in this application, if the number of available communication resources for each edge model aggregation is U, then the number of available communication resources for synchronous federated learning is 16U. It is set that the fully asynchronous federated learning aggregates V max times within the time t AFL Then the number of available communication resources for each time is 16UT 1 / t max V AFL .

[0151] As Figure 6 and Figure 7 shown, whether judged by error or accuracy, compared with synchronous federated learning and fully asynchronous federated learning, the proposed federated learning mechanism based on the hybrid aggregation strategy has more superior performance. This is because the proposed method integrates the advantages of synchronous federated learning and fully asynchronous federated learning, which can not only effectively address the straggler problem in synchronous federated learning but also overcome the overly frequent model parameter exchanges in fully asynchronous federated learning. In addition, it can be found that as the total training time increases, the performance of the models corresponding to each federated learning mechanism gradually converges. However, relatively speaking, the federated learning method proposed in this application always demonstrates better performance.

[0152] As Figure 8As shown, compared with only considering the parameter age of the parameter approximation of the local model update parameters and only considering the number of samples in the local dataset when aggregating edge models, the proposed weight design strategy demonstrates superior performance whether the comparison criterion is error or accuracy. This is because, although all are associated with the same base station, due to the different times required to complete a local model training, the versions of the edge models on which each unmanned vehicle's local model training is based are also different in the federated learning framework based on periodic aggregation. The proposed weight design strategy comprehensively considers the parameter age of the parameter approximation of the local model update parameters and the number of samples in the local dataset of each unmanned vehicle, which can effectively avoid the negative impact of outdated local model updates and unmanned vehicles with too few samples on the performance of the trained model.

[0153] Furthermore, as described above, in the present invention, multiple unmanned vehicles associated with the same base station are co-trained based on federated learning with periodic aggregation. Based on this, this application also compares the impact of different values of the time period T on the performance of the proposed method.

[0154] As Figure 9 and Figure 10 shown, regardless of the value of the time period T, as the model training time increases, both the training error and the test error gradually decrease, that is, both the training accuracy and the test accuracy gradually increase. Further careful observation reveals that when T = t max / 4, the convergence speed of the proposed algorithm is the fastest. However, in terms of the performance to which the final model converges, when T = t max / 16, the performance of the proposed method is the best. This is because if the value of the time period T is too large, then the number of edge iterations within a fixed time is very limited, which naturally affects the performance of the obtained model.

[0155] Furthermore, combining Figure 11 and Figure 12 , the following analyzes the impact of the uniform distribution and random distribution of N unmanned vehicles on the performance of the federated learning method proposed in this application.

[0156] Specifically, the uniform distribution of N unmanned vehicles means that each base station is associated with the same number of N / M unmanned vehicles. Correspondingly, the random distribution means that N unmanned vehicles are randomly distributed within the coverage of M base stations. It can be seen from the figure that whether it is a uniform distribution or a random distribution, as the number of available communication resources increases, both the training error and the test error gradually decrease. This is because, as the number of available communication resources increases, the communication resources available to each unmanned vehicle will increase accordingly. Naturally, the compression rate of the local model update can be reduced and more parameters can be uploaded, thereby improving the performance of the trained model.

[0157] Furthermore, upon closer inspection, it can be found that when the number of communication resources is fixed, compared with random distribution, the federated learning method proposed in this application has better performance when N unmanned vehicles are evenly distributed.

[0158] Based on the same inventive concept, as Figure 13 shown, the present invention also provides a federated learning device for multi-base-station unmanned vehicles, and the device includes:

[0159] A set construction unit 101, configured to construct a base station set and an unmanned vehicle set, where the base station set includes M base stations, and the unmanned vehicle set includes N unmanned vehicles.

[0160] A model construction unit 102, configured to construct M first edge models and an initial global network model based on a hybrid aggregation strategy; the M first edge models respectively correspond to the M base stations one by one, and the initial global network model is formed by aggregating the M first edge models.

[0161] A distribution unit 103, configured to broadcast the M first edge models to each unmanned vehicle in the unmanned vehicle set.

[0162] A model training unit 104, configured to use the M first edge models to respectively train each local model of each unmanned vehicle to obtain update parameters of each local model.

[0163] A model update unit 105, configured to compress the update parameters of each local model and then upload them to the corresponding base stations for edge model aggregation respectively to obtain M second edge models.

[0164] Furthermore, the model update unit 105 is further configured to use the M second edge models to replace the M first edge models, so that the distribution unit 103, the model training unit 104, and the model update unit 105 are further configured to repetitively execute corresponding operations for the first time according to the M second edge models, and so on. When the distribution unit 103, the model training unit 104, and the model update unit 105 repetitively execute corresponding operations for the bth time, b = 2, 3,..., E - 1, correspondingly, the model update unit 105 is further configured to use the M (b + 1) edge models to replace the M b edge models in the distribution unit 103, and the distribution unit 103 broadcasts the M (b + 1) edge models to each unmanned vehicle in the unmanned vehicle set until after obtaining M target edge models, the M target edge models are subjected to global model aggregation to obtain a target global network model.

[0165] In this application, other technical features in the above-mentioned federated learning device for multi-base-station unmanned vehicles are the same as the features disclosed in the above method embodiment, and will not be elaborated herein.

[0166] Based on the same inventive concept, the present invention further provides an electronic device, which may specifically be a control device or control system inside a smart device, or an external device communicating with the smart device. It may be a desktop computer, a portable computer, a smart phone, a tablet computer, a personal digital assistant (PDA), a server, etc.

[0167] Specifically, the device may include a processor and a memory. Further, the processor may be a general-purpose processor, such as a central processing unit (CPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, which can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present application may be directly embodied as being executed by a hardware processor, or executed by a combination of hardware and software modules in the processor.

[0168] Further, the memory, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. The memory may include at least one type of storage medium, for example, it may include flash memory, a hard disk, a multimedia card, a card-type memory, a random access memory (RAM), a static random access memory (SRAM), a programmable read-only memory (PROM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic memory, a magnetic disk, an optical disk, etc. The memory is any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory in the embodiments of the present application may also be a circuit or any other device capable of implementing a storage function, for storing program instructions and / or data.

[0169] It can be seen that the present invention discloses a federated learning method, device and equipment for multi-base-station unmanned vehicles. The method includes the following steps: First step, construct a base station set and an unmanned vehicle set. The base station set includes M base stations, and the unmanned vehicle set includes N unmanned vehicles; Second step, based on a hybrid aggregation strategy, construct M first edge models and an initial global network model; The M first edge models correspond to the M base stations one by one, and the initial global network model is formed by aggregating the M first edge models; Third step, broadcast the M first edge models to each unmanned vehicle in the unmanned vehicle set; Fourth step, use the M first edge models to train each local model of each unmanned vehicle respectively to obtain the updated parameters of each local model; Fifth step, after compressing the updated parameters of each local model respectively, upload them to the corresponding base station for edge model aggregation to obtain M second edge models; Sixth step, use the M second edge models to replace the M first edge models, and return to the third step. The third to fifth steps are repeatedly executed for the first time, and so on. When the third to fifth steps are repeatedly executed for the bth time, b = 2, 3,..., E - 1, correspondingly use the M (b + 1)th edge models to replace the M bth edge models and return to the third step for broadcasting to each unmanned vehicle in the unmanned vehicle set. Until M target edge models are obtained, aggregate the M target edge models to obtain a target global network model. This method can not only effectively cope with the differences in the computing and storage capabilities of multiple unmanned vehicles within a single base station and the sample data of local data sets, but also solve the problem of stragglers that may be caused by vehicle failures, traffic accidents, extreme weather, etc. in reality, as well as the overly frequent model parameter exchange in fully asynchronous federated learning.

[0170] The above are only the embodiments of the present invention, and do not limit the protection scope of the present invention accordingly. Any equivalent structural transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, is included in the protection scope of the present invention.

Claims

1. A federated learning method for multi-base station unmanned vehicles, characterized in that: The method comprises the following steps: The first step is to construct a base station set and an unmanned vehicle set, wherein the base station set includes M base stations and the unmanned vehicle set includes N unmanned vehicles; The second step is to construct M first edge models and an initial global network model based on a hybrid aggregation strategy; the M first edge models correspond to the M base stations one by one, and the initial global network model is formed by aggregating the M first edge models; Step 3: broadcast the M first edge models to each unmanned vehicle in the unmanned vehicle set; Step 4: using the M first edge models, respectively train the local models of the unmanned vehicles to obtain update parameters of the local models; Step 5: compress the update parameters of each local model and upload them to the corresponding base station for edge model aggregation to obtain M second edge models; In the sixth step, the M second edge models are used to replace the M first edge models, and the process returns to the third step. The third to fifth steps are repeated for the first time, and so on. When the execution is repeated for the bth time, b=2, 3, ..., E-1, and the M (b+1)th edge models are used to replace the M bth edge models. The process returns to the third step for broadcasting to each unmanned vehicle in the unmanned vehicle set until M target edge models are obtained. The M target edge models are then aggregated into a global model to obtain a target global network model.

2. The federated learning method according to claim 1, characterized in that: The second step includes: Performing edge aggregation on local models of multiple unmanned vehicles associated with each of the M base stations to obtain the M first edge models; Perform global model aggregation on the M first edge models to obtain the initial global network model.

3. The federated learning method according to claim 1, characterized in that: Before the third step, it also includes: The initial global network model is broadcasted to each base station in the base station set.

4. The federated learning method according to claim 1, characterized in that: The fifth step comprises: Using compression functions to compress the local model update parameters respectively to obtain compression parameters; Uploading the compression parameters to their corresponding base stations; Decompressing the compressed parameters using decompression functions to obtain approximate values ​​of the parameters; Edge model aggregation is performed using a plurality of parameter approximations corresponding to each of the M base stations to obtain the M second edge models.

5. The federated learning method according to claim 4, characterized in that: The method for obtaining the compression functions is: Calculate the number of parameters that can be uploaded by each unmanned vehicle to obtain each index set; Obtaining compression matrices corresponding to the unmanned vehicles according to the shared matrices used by the unmanned vehicles and the index sets; Based on the compression matrices, the compression functions corresponding to the unmanned vehicles are designed.

6. The federated learning method according to claim 5, characterized in that: The designing the compression functions corresponding to the unmanned vehicles based on the compression matrices includes: Design the compression function corresponding to the unmanned vehicle n; in the zth edge iteration, we have: Where z = 1, 2, 3, ..., E; represents the compression function corresponding to the unmanned vehicle n in the zth edge iteration; Represents the compression matrix corresponding to the unmanned vehicle n in the zth edge iteration.

7. The federated learning method according to claim 5, characterized in that: The performing edge model aggregation using the multiple parameter approximations corresponding to each of the M base stations includes: The edge model is aggregated using multiple parameter approximations corresponding to base station m; in the zth edge iteration, there are: Where z = 1, 2, 3, ..., E; represents the (z+1)th edge model of the base station m obtained in the zth edge iteration; represents a zth edge model of the base station m used in the zth edge iteration; represents a weighted average of a plurality of parameter approximations associated with the base station m in the zth edge iteration, and the weighted average It can be expressed as: in, represents the approximate value of the parameter corresponding to the unmanned vehicle n; is the approximate value of the parameter corresponding to the unmanned vehicle n The weight of Represents a subset of unmanned vehicles consisting of multiple unmanned vehicles associated with the base station m in the zth edge iteration.

8. The federated learning method according to claim 7, characterized in that: The method further comprises: Designing the weights of the parameter approximations corresponding to the respective unmanned vehicles according to the number of samples in the local data set of the respective unmanned vehicles and the parameter age of the local model update parameters; Among them, the approximate parameter value corresponding to the unmanned vehicle n is The weight is: Where, e is the natural logarithm; |D n | represents the number of samples in the local data set of the unmanned vehicle n; represents the parameter age, and the parameter age It can be expressed as: in, represents the edge model aggregation that the unmanned vehicle n participated in last time, and represents a subset of unmanned vehicles consisting of multiple unmanned vehicles associated with the base station m in the z'th edge iteration.

9. A federated learning device for a multi-base station unmanned vehicle, characterized in that: The device comprises: A set construction unit, used to construct a base station set and an unmanned vehicle set, wherein the base station set includes M base stations, and the unmanned vehicle set includes N unmanned vehicles; A model building unit, configured to build M first edge models and an initial global network model based on a hybrid aggregation strategy; the M first edge models correspond to the M base stations one by one, respectively, and the initial global network model is formed by aggregating the M first edge models; A sending unit, configured to broadcast the M first edge models to each unmanned vehicle in the unmanned vehicle set; A model training unit, used to train each local model of each unmanned vehicle using the M first edge models to obtain update parameters of each local model; A model updating unit is used to compress the local model update parameters and upload them to the corresponding base station for edge model aggregation to obtain M second edge models; the model updating unit is also used to replace the M first edge models with the M second edge models, so that the sending unit, the model training unit and the model updating unit are also used to repeat the corresponding operations for the first time according to the M second edge models, and so on, when the sending unit, the model training unit and the model updating unit repeat the corresponding operations for the bth time, b=2,3,...,E-1, the corresponding model updating unit is also used to replace the M bth edge models in the sending unit with M (b+1)th edge models, and the sending unit broadcasts the M (b+1)th edge models to each unmanned vehicle in the unmanned vehicle set until M target edge models are obtained, and then the M target edge models are aggregated into global models to obtain a target global network model.

10. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method according to any one of claims 1 to 8 when executing the computer program.

Images