A federated learning method, apparatus, and device for multi-base station unmanned vehicles

Abstract

This invention discloses a federated learning method, apparatus, and device for multi-base station unmanned vehicles. The method includes: first, constructing a base station set and an unmanned vehicle set; second, constructing a first edge model and an initial global network model; third, broadcasting the first edge model to each unmanned vehicle; fourth, training each local model of each unmanned vehicle to obtain update parameters for each local model; fifth, compressing the update parameters of each local model and uploading them to the corresponding base station for aggregation to obtain a second edge model; sixth, replacing the first edge model with the second edge model and returning to the third step. The third to fifth steps are repeated for the first time, and so on. When the b-th iteration is repeated, b = 2, 3, ..., E-1, the (b+1)-th edge model is replaced with the b-th edge model and the process returns to the third step until the target edge model is obtained. The target edge models are then aggregated to obtain a high-performance target global network model.

Description

Technical Field

[0001] This invention relates to the field of mobile communication technology, and in particular to a federated learning method, apparatus, and device for multi-base station unmanned vehicles. Background Technology

[0002] As an emerging machine learning paradigm, federated learning enables multiple distributed data holders to collaboratively train specific neural network models. Compared to traditional centralized model training methods, in federated learning, the original data does not need to leave the local storage of each distributed participant, thus significantly reducing the risk of data leakage and better protecting the rights and interests of each participant. Furthermore, in traditional centralized model training methods, uploading the corresponding original data from each distributed participant to the central server consumes a large amount of wireless spectrum resources and causes severe network latency, a problem that does not exist in federated learning. Moreover, in federated learning, network model training is performed in parallel at each participant, no longer overly relying on the computing and storage capabilities of the central server, thus significantly improving computational efficiency. It is evident that such a scenario of multiple distributed participants working collaboratively and efficiently is highly compatible with the collaborative and efficient operation of multiple distributed autonomous vehicles ensured by cellular base stations in intelligent transportation systems. Based on this, research related to federated learning for connected autonomous vehicles has emerged continuously in recent years. However, to truly ensure that connected autonomous vehicles achieve their expected benefits, existing work still has many shortcomings.

[0003] Specifically, considering the frequent exchange of model parameters between autonomous vehicles and base stations during federated learning, various model compression and sparsification methods have been extensively studied to reduce the constraints of scarce wireless spectrum resources. However, existing research has not fully considered the inherent statistical characteristics of federated learning, thus the model compression effect needs further improvement. In addition, to introduce more diverse sample data into the federated learning process and obtain high-performance federated learning models, related research has proposed a multi-base station federated learning framework. However, this relies on a remote cloud server acting as a parameter server to aggregate edge models formed by multiple base stations. We know that, unlike manned vehicles, autonomous vehicles have extremely high requirements for real-time guidance and perception. Therefore, the latency caused by the exchange of model parameters with the remote cloud would be unacceptable for the operation of interconnected autonomous vehicles in practice.

[0004] Furthermore, existing federated learning mechanisms are mainly divided into synchronous federated learning and asynchronous federated learning. Synchronous federated learning requires waiting for all autonomous vehicles to complete their local model training and upload their corresponding local model update parameters before global model aggregation. However, due to differences in computing and storage capabilities and the number of samples in their local datasets, the time required for each autonomous vehicle to complete local model training varies. Applying synchronous federated learning would require waiting for the vehicle with the longest training time to complete its local model training and upload its corresponding local model update parameters, severely impacting the real-time update capability of the trained model. Moreover, considering the complex driving environment and the characteristics of wireless transmission, in reality, it may take a long time to collect local model update parameters from all autonomous vehicles. Unlike synchronous federated learning, in fully asynchronous federated learning, as soon as any autonomous vehicle completes its local model training and uploads its corresponding local model update parameters, the parameter server performs a global model aggregation. Clearly, the already scarce wireless spectrum resources can hardly effectively support such frequent model parameter exchanges in fully asynchronous federated learning. Summary of the Invention

[0005] The purpose of this invention is to provide a federated learning method, apparatus, and device for multi-base station unmanned vehicles, addressing the problem of lagging vehicles that may be caused by vehicle malfunctions, traffic accidents, and extreme weather. This mainly includes how to integrate the advantages of synchronous federated learning and fully asynchronous federated learning to construct a federated learning framework for multi-base station unmanned vehicles; how to avoid outdated local model updates and the negative impact of unmanned vehicles with insufficient 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 frequent model parameter exchanges and scarce wireless spectrum resources during the federated learning process.

[0006] To address the aforementioned technical problems, this invention provides a federated learning method for multi-base station unmanned vehicles (UAVs). The method includes the following steps: First, constructing a base station set and an UAV set, wherein the base station set includes M base stations and the UAV set includes N UAVs; Second, based on a hybrid aggregation strategy, constructing M first-order edge models and an initial global network model; the M first-order edge models correspond one-to-one with the M base stations, and the initial global network model is formed by aggregating the M first-order edge models; Third, broadcasting the M first-order edge models to each UAV in the UAV set; Fourth, using the M first-order edge models to train each local model of each UAV. The process involves several steps: 1) obtaining update parameters for each local model; 2) compressing these local model update parameters and uploading them to the corresponding base station for edge model aggregation to obtain M second-order edge models; 3) replacing the M first-order edge models with the M second-order edge models and returning to step 3. This process is repeated for the first time, from step 3 to step 5, and so on. For the b-th repetition (b = 2, 3, ..., E-1), the M (b+1)-th edge models are used to replace the M b-th edge models, and the process returns to step 3. This is then broadcast to each unmanned vehicle in the unmanned vehicle set until M target edge models are obtained. Finally, these M target edge models are aggregated globally to obtain the target global network model.

[0007] This invention also provides a federated learning device for multi-base station unmanned vehicles, the device comprising: a set construction unit for constructing a base station set and an unmanned vehicle set, the base station set including M base stations and the unmanned vehicle set including 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 each corresponding one-to-one with the M base stations, and the initial global network model being 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 training each local model of each unmanned vehicle using the M first-edge models to obtain update parameters for each local model; and a model update unit for compressing the update parameters of each local model and uploading them to the corresponding [unmanned vehicle set]. The base station performs edge model aggregation to obtain M second-order edge models. The model update unit is also used to replace the M first-order edge models with the M second-order edge models. The distribution unit, the model training unit, and the model update unit are also used to repeatedly perform the corresponding operation for the first time based on the M second-order edge models. Similarly, when the distribution unit, the model training unit, and the model update unit repeatedly perform the corresponding operation for the bth time, b = 2, 3, ..., E-1, the model update unit is also used to replace the M b-th edge models in the distribution unit with the M (b+1)-th edge models. 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. Then, the M target edge models are globally aggregated to obtain the 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, wherein the processor executes the computer program to perform the steps of the method described above.

[0009] The beneficial effects of this invention are as follows: This invention discloses a federated learning method, apparatus, and device for multi-base station unmanned vehicles. The method includes the following steps: First, constructing 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; Second, based on a hybrid aggregation strategy, constructing M first-order edge models and an initial global network model; the M first-order edge models correspond one-to-one with the M base stations, and the initial global network model is formed by aggregating the M first-order edge models; Third, broadcasting the M first-order edge models to each unmanned vehicle in the unmanned vehicle set; Fourth, using the M first-order edge models, training each local model of each unmanned vehicle to obtain... The process involves several steps: First, obtaining the updated parameters of each local model. Second, compressing these parameters and uploading them to the corresponding base station for edge model aggregation to obtain M second-order edge models. Third, replacing the M first-order edge models with the M second-order edge models and returning to step three. This process is repeated for the first time, from step three to step five, and so on. For the b-th repetition (b = 2, 3, ..., E-1), the M (b+1)-th edge models are used to replace the M b-th edge models, and the process returns to step three. This is then broadcast to all autonomous vehicles in the vehicle set until M target edge models are obtained. Finally, these M target edge models are aggregated globally to obtain the target global network model. This method effectively addresses the differences in computing and storage capabilities among multiple autonomous vehicles within a single base station and the variations in local dataset sample data. It also solves the problem of vehicles falling behind due to vehicle malfunctions, traffic accidents, and extreme weather, as well as the excessively frequent model parameter exchanges in fully asynchronous federated learning. Attached Figure Description

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

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

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

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

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

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

[0016] Figure 7 This is a comparison of the training accuracy of a federated learning method for multi-base station unmanned vehicles according to the present invention compared with synchronous federated learning and fully asynchronous federated learning.

[0017] Figure 8 This is a weight design strategy in a federated learning method for multi-base station unmanned vehicles according to the present invention, compared with the training error comparison results of weight design strategies based solely on sample size and based solely on parameter age;

[0018] Figure 9 This is a comparison of training errors when the time period takes different values ​​in the federated learning method for multi-base station unmanned vehicles of the present invention.

[0019] Figure 10 This is a comparison of training accuracy when the time period takes different values ​​in the federated learning method for multi-base station unmanned vehicles of the present invention.

[0020] Figure 11 This is a comparison of training errors for randomly distributed and uniformly distributed unmanned vehicles in a federated learning method for multi-base station unmanned vehicles according to the present invention.

[0021] Figure 12 This is a comparison of the training accuracy of randomly distributed and uniformly distributed unmanned vehicles in a federated learning method for multi-base station unmanned vehicles according to the present invention.

[0022] Figure 13 This is a block diagram of a federated learning device for multi-base station unmanned vehicles according to the present invention. Detailed Implementation

[0023] To facilitate understanding of the present invention, a more detailed description is provided below with reference to the accompanying drawings and specific embodiments. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[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 one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0025] like Figure 1 The diagram illustrates a federated learning method for multi-base station unmanned vehicles according to the present invention, comprising 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 the hybrid aggregation strategy, construct M first-edge models and an initial global network model; the M first-edge models correspond one-to-one with the M base stations, and the initial global network model is formed by aggregating the M first-edge models.

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

[0029] Step S4: Using M first-order edge models, train each local model of each autonomous vehicle to obtain the update parameters of each local model.

[0030] Step S5: 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.

[0031] Step S6: Replace the M first edge models with M second edge models, return to step S3, and repeat steps S3 to S5 for the first time. In the b-th repetition, b = 2, 3, ..., E-1, replace the M b-th edge models with M (b+1)-th edge models and return to step S3. This is used to broadcast to each autonomous vehicle in the autonomous vehicle set until M target edge models are obtained. Then, perform global model aggregation on the M target edge models to obtain the target global network model.

[0032] For ease of description, the 1st edge model, the bth edge model, and the (b+1)th edge model will be uniformly represented as the zth edge model. Where z = 1, 2, 3, ..., E.

[0033] This application constructs a set of base stations and a set of unmanned vehicles (UAVs) to facilitate the association between each UAV and its corresponding base station; and constructs M first-level edge models and an initial global network model to associate each first-level edge model with its corresponding base station. Further, after each UAV in the set obtains the relevant parameters of its corresponding first-level edge model, each UAV asynchronously trains its corresponding local model to obtain corresponding local model update parameters. Edge model aggregation within a single base station is then performed at the network edge to obtain the second-level edge model corresponding to each base station. Following step S6, edge iteration is performed. After obtaining the target edge model 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 suitable for multi-UAV collaboration. This invention focuses on the hot topic of multi-UAV collaboration, proposing a practical and effective solution with excellent application prospects. This federated learning method can reconcile the practical contradiction between frequent model parameter exchanges during federated learning and scarce wireless spectrum resources, ensuring good performance of the obtained target global network model while solving communication bottlenecks.

[0034] It should be noted that each time steps S3 to 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 model corresponding to each base station can be obtained.

[0035] The following is combined with Figures 1 to 8 The present application will be further described in detail with reference to specific embodiments.

[0036] Figure 1 A flowchart illustrating the federated learning method for multi-base station unmanned vehicles provided in this application is shown below, in detail:

[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 autonomous vehicles, while each autonomous vehicle can only correspond to one base station, and each autonomous vehicle also has a corresponding local model. It is worth noting that in reality, a base station's coverage area may include multiple autonomous vehicles, and a single autonomous vehicle may be within the coverage area of ​​multiple base stations. However, in this application, each autonomous vehicle has only one associated base station. That is, each autonomous vehicle can only upload its corresponding local model update parameters to its associated base station.

[0039] In this application, the following are defined and Let M represent the set of base stations consisting of M base stations and the set of unmanned vehicles consisting of N unmanned vehicles, respectively.

[0040] Furthermore, driverless cars The local dataset corresponding to the local model is Among them, |D n | represents the local dataset D n The number of samples in X i Let Y represent the i-th data sample. i For the i-th data sample X i The corresponding label is output.

[0041] Secondly, setting up driverless cars There is one and only one associated base station. And with base station Multiple related driverless vehicles (including driverless cars) The set of autonomous vehicles can be defined as the m-th subset, and the m-th subset can be represented as...

[0042] It should be noted that, in order to serve emerging vehicle applications, N autonomous vehicles spanning M base stations need to collaboratively train a parameter vector of d dimensions. The neural network model that determines this.

[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 one-to-one with the M base stations, 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 a periodic aggregation-based federated learning strategy and a synchronous federated learning strategy. Since each base station is associated with multiple autonomous vehicles, and each autonomous vehicle is associated with only one base station, the periodic aggregation-based federated learning strategy aggregates the local models of multiple autonomous vehicles associated with a single base station to form the edge model corresponding to that base station. Meanwhile, the synchronous federated learning strategy aggregates the corresponding edge models of multiple base stations to form a global network model.

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

[0046] Step S21: Perform edge aggregation on the local models of multiple unmanned 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 unmanned vehicles.

[0048] Among them, the federated learning strategy based on periodic aggregation will be combined with the base station The local models of multiple related autonomous vehicles are aggregated to form a base station. The corresponding first edge model.

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

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

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

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

[0053]

[0054] Where w represents the global network model used in a single global iteration; l(w,X) i ) represents the data based on the i-th data sample X i The calculated loss function value; D = ∪ n∈N D n Let |D| represent the total amount of sample data from N autonomous vehicles.

[0055] In existing federated learning paradigms, each autonomous vehicle's local dataset is private. The network model training process is performed locally based on its own computing and storage capabilities, and then the trained local model updates its parameters and is uploaded to a parameter server for aggregation. Based on this, for autonomous vehicles... In this regard, the local objective loss function used can be expressed as:

[0056]

[0057] Where w represents the autonomous vehicle in one edge iteration The corresponding local model.

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

[0059] Step S3: Broadcast the M first-order edge models to each autonomous vehicle in the autonomous vehicle set.

[0060] Combination Figure 2 and Figure 3 Before the first edge iteration begins, parameter server 1 broadcasts the initial global network model w0 to the M base stations 2 in the base station set via the downlink, where parameter server 1 is one of the M base stations 2. Then, the M base stations 2 each transmit their corresponding first edge model. The broadcast is sent to the corresponding subset of autonomous vehicles, that is, the M first-order edge models are broadcast to the set of autonomous vehicles. The various driverless vehicles in the system.

[0061] It should be noted that, due to the sufficient channel bandwidth of the downlink, the transmission rate is usually high. Therefore, the time delay of parameter server 1 broadcasting the initial global network model w0 can be ignored.

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

[0063] like Figure 2 As shown, for each autonomous vehicle, if it completes local model training within a fixed time period, it uploads the corresponding local model update parameters to its associated base station 2 via a wireless link; otherwise, it continues local model training without interruption. This effectively addresses the differences in computing and storage capabilities and local dataset sample data among multiple autonomous vehicles within a single base station 2. It also solves the problem of vehicles falling behind due to vehicle malfunctions, traffic accidents, and extreme weather, as well as the excessively frequent model parameter exchanges inherent in fully asynchronous federated learning.

[0064] It should be noted that due to differences in computing power and the number of local dataset samples, the time required for each autonomous vehicle to complete a local model training varies.

[0065] In this embodiment, an unmanned vehicle is set up. The time required to complete one local model training is t. n ∈[t min ,t max ], where t min and t max These represent the shortest and longest time required to complete a local model training for all autonomous vehicles, respectively.

[0066] Combination Figure 2 and Figure 3Within each time period T, the parameter server 1 performs an edge model aggregation once (i.e., performs an edge iteration, which is the specific operation of steps S4 to S5 to obtain a new edge model), and all autonomous vehicles need to perform L local iterations to complete a local model training.

[0067] In this embodiment, when the unmanned vehicles gather Each drone in the process receives the first edge model Then, the first edge iteration begins.

[0068] In this process, each autonomous vehicle's local model begins training to obtain corresponding local model update parameters. These parameters are then compressed and uploaded to their respective base stations. Each base station then aggregates the received local model update parameters to obtain the corresponding second edge model.

[0069] It should be noted that, for base stations The associated m-th unmanned vehicle subset If the m-th unmanned vehicle set In If an autonomous vehicle is trained using synchronous federated learning, it needs to wait for all... Edge model aggregation can only be performed after all autonomous vehicles have completed local model training and uploaded the corresponding local model updates.

[0070] However, since autonomous vehicles communicate with base stations via wireless links, many uncertainties and unpredictable factors exist. In reality, vehicle malfunctions, traffic accidents, weather, and other reasons may prevent some vehicles from completing local model training and uploading the corresponding local model updates to the associated base station within a short period of time. Furthermore, the overly frequent model parameter exchanges inherent in fully asynchronous federated learning are also unsuitable for the m-th subset of autonomous vehicles. Multiple unmanned vehicles are conducting collaborative training.

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

[0072] Specifically, setting up base stations Edge model aggregation is performed once every fixed time period T, if the m-th unmanned vehicle subset If each autonomous vehicle completes local model training within time period T, it will then upload the corresponding local model update parameters to the base station. Otherwise, local model training will continue without interruption.

[0073] In this embodiment, each of the M base stations is required to complete E edge iterations, and edge model aggregation is performed once every time period T to finally obtain M target edge models. After obtaining the M target edge models, the M base stations upload them to parameter server 1 for subsequent global model aggregation.

[0074] As mentioned 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 simultaneously obtain the target edge model, meaning there is no potential "falling behind" problem. Furthermore, communication between base stations is achieved through stable and reliable wired transmission. Therefore, training between multiple base stations can be based on synchronous federated learning. In summary, in the federated learning method considered in this invention, multiple autonomous vehicles associated with a single base station collaboratively train to obtain their respective target edge models based on periodic aggregation federated learning, while multiple base stations train together based on synchronous federated learning to obtain the target global network model.

[0075] In this embodiment, in the z-th edge iteration, the autonomous vehicle Firstly, based on its own local dataset D n If local model training is completed within time period T, the corresponding local model update is calculated and obtained. Otherwise, continue local model training without interruption. Indicates driverless car The local model update parameters corresponding to the z-th edge iteration are calculated as follows:

[0076]

[0077] in, For driverless cars Based on the local dataset D in the z-th edge iteration n The updated local model trained using its own computational and storage resources, where E is the total number of edge iterations.

[0078] Step S5: 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.

[0079] It should be noted that communication between the autonomous vehicles and base stations is via wireless transmission. Considering the scarcity of wireless spectrum resources and the frequent exchange of model parameters between the autonomous vehicles and base stations during federated learning, this invention compresses the local model update parameters (d-dimensional) of each autonomous vehicle before uploading them to the associated base station using a corresponding compression function to obtain compressed parameters. Then, these compressed parameters are uploaded to the corresponding base station, and decompression functions are used to decompress the compression functions, obtaining approximate values ​​for each parameter. Finally, the M base stations perform a weighted average of their respective approximate parameter values ​​to obtain M second-order edge models.

[0080] Essentially, the compression method proposed in this invention is based on the inherent statistical characteristics of the federated learning process, which are mainly reflected in two aspects. Firstly, the changes in model parameters during local model training are slow and continuous rather than abrupt. Therefore, the difference between the edge model parameters corresponding to two adjacent edge iterations of the local model update parameters can be compressed, or in other words, it can be compressed after sparse transformation. Secondly, the local model update parameters of two adjacent iterations are similar; that is, the current local model update parameters can be mapped to the subspace containing the previous local model update parameters for compression.

[0081] Specifically, in the m-th subset of autonomous vehicles Before the local model training for each autonomous vehicle in the system begins, the base station Will be with the m-th unmanned vehicle subset All driverless cars in the system use the same shared matrix. Each element in the shared matrix Q independently takes values ​​from the same Gaussian distribution. Furthermore, each row has been normalized; among them, Let Q represent a Gaussian distribution with a mean of 0 and a variance of 1 / d. In this way, each row of the shared matrix Q is randomly distributed on the surface of a d-dimensional unit sphere, and the mean of all rows is 0. Furthermore, any two rows of the shared matrix Q are uncorrelated. That is, drawing row vectors of any proportion from the shared matrix Q will form an orthogonal basis.

[0082] As mentioned above, in the z-th edge iteration, the autonomous vehicle uploads the d-dimensional local model update parameters. Previously, it was necessary to use a compression function. Compress it into A 3D vector, which yields the compression parameters.

[0083] In practice, The determination of depends on the uplink communication resources available to the autonomous vehicle n and the corresponding wireless transmission channel status. In other words, the original local model update parameters... There are d parameters. However, in reality, due to the limitations of uplink communication resources available to the autonomous vehicle n and the constraints of the corresponding wireless transmission channel status, the parameters can only be updated using the local model. Compressed to only Compression parameters of each parameter Upload now.

[0084] Among them, compression function Based on the corresponding compression matrix The designed, and compression matrix Then selected from the shared matrix Q Line composition.

[0085] Specifically, setting up base stations The available wireless channel transmission bandwidth is B m The uplink transmission is configured to use orthogonal frequency division multiple access (OFDMA) technology. Each driverless vehicle in the system enjoys equal access to base stations. Available wireless spectrum resources. In other words... The process of multiple autonomous vehicles uploading network model parameters to base station m via wireless links is independent of each other, and the available wireless transmission channel bandwidth for each autonomous vehicle is [missing information]. in, for The number of driverless cars in the country.

[0086] Furthermore, setting up driverless cars The transmission power is p n The additive noise in the actual transmission process satisfies the distribution CN(0,σ). 2 The corresponding uplink channel transmission gain In this way, in the z-th edge iteration, for the autonomous vehicle In this regard, the corresponding wireless transmission channel rate is:

[0087]

[0088] Furthermore, let's define the autonomous vehicle in the z-th edge iteration. The time delay constraint for uploading network model parameters to base station m is: Then driverless car Maximum upload Bits to base station m. Wherein, The specific calculation method is as follows:

[0089]

[0090] Referring to existing federated learning methods, it is assumed that 32 bits are required to upload a network model parameter via a wireless link during the federated learning process. 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 For the selected A set of row indices. Based on a defined set of indices. You can then obtain the compression matrix. for:

[0093]

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

[0095]

[0096] As mentioned above, the autonomous vehicle n uploads local model update parameters. Previously, it would be done through a compression function. Compression is then performed. In this way, based on the compression function determined in the above equation... This allows you to obtain the original local model update parameters. The compressed version, i.e., the compression parameters of the actual wireless transmission. for:

[0097]

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

[0099]

[0100] As mentioned above, the associated base station receives the compression parameters uploaded by the unmanned vehicle n. Then, it will be processed by a decompression function. Decompress it to obtain the corresponding approximate parameter values:

[0101]

[0102] in, This refers to updating parameters for the local model. Approximate parameter values ​​after compression and decompression processes.

[0103] It can be observed that as long as the compression matrix is ​​determined... Compression function and decompression functions Solving this problem will then become straightforward. Therefore, the challenge in constructing the proposed model compression method based on the inherent statistical properties of federated learning lies in how to select from the shared matrix Q... Rows form a compressed matrix

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

[0105] As for how to determine the index set This invention primarily relies on the similarity between consecutive local model update parameters. In other words, the current local model update parameters can be mapped to the subspace containing the previous local model update parameters for model compression.

[0106] Specifically, in the proposed model compression method based on the inherent statistical properties of federated learning, the index set This can be obtained by solving problem P1, which can be expressed as:

[0107]

[0108] stP=Q T (I,:)(Q(I,:)Q T (I,:)) -1 Q(I,:);

[0109] Problem P1 can be understood as how to select the index set I to maximize The rows formed by the corresponding index set I in the shared matrix Q The norm on the dimensional subspace. Obviously, the solution to problem P1 is NP-hard (non-deterministic polynomial). In practice, suboptimal solutions to problem P1 can also be obtained based on classic matching pursuit methods in the field of compressed sensing.

[0110] In the proposed model compression method based on the inherent statistical properties of federated learning, the base station with strong computing and storage capabilities is responsible for solving problem P1, and the obtained index set I is also broadcast by the base station to each autonomous vehicle. Therefore, for autonomous vehicles with limited computing, storage and communication resources, such model compression design will not increase their burden.

[0111] In this embodiment, the local model update parameters of dimension d are uploaded to the n-th dimension of the autonomous vehicle. Before reaching the associated base station, it will be processed through the corresponding compression function. Compress it, and it will be represented as:

[0112]

[0113] in, This refers to updating parameters for the local model. The compression parameters obtained after compression, and have

[0114] Correspondingly, the associated base station receives the data uploaded by the unmanned vehicle n. Then, it will be processed by a decompression function. Decompress it to obtain the corresponding approximate parameter values:

[0115]

[0116] in, This refers to updating parameters for the local model. Approximate parameter values ​​after compression and decompression.

[0117] As mentioned above, Indicates connection with base station The associated m-th subset of autonomous vehicles. Since the local model training for each autonomous vehicle is asynchronous, further, we set... This represents the subset of m-th unmanned vehicles at the z-th edge iteration. A subset of autonomous vehicles that have completed local model training. In this way, base station data can be obtained. The weighted average of the approximate values ​​of the associated parameters is:

[0118]

[0119] in, For base stations Related parameter approximations The corresponding weights.

[0120] Furthermore, set Let represent the (z+1)th edge model of base station m obtained in the z-th edge iteration, then we have:

[0121]

[0122] It is worth noting that in existing federated learning work, when base station m performs edge model aggregation, the weights The design is often based on the size of the local dataset of the autonomous vehicle n, that is, the weights Designed as follows:

[0123]

[0124] in, for The number of samples in the sample.

[0125] For base station m, this weight design during edge model aggregation does indeed strike a good balance. The number of samples for each autonomous vehicle. However, as mentioned above, in this invention, multiple autonomous vehicles associated with the same base station are collaboratively trained based on a periodic aggregation federated learning strategy. Therefore, although they are all associated with the same base station, the training of the local model may be based on different versions of the edge model. In this case, to avoid outdated local model update parameters affecting the currently aggregated edge model, weights are designed... At that time, it is not only necessary to consider In addition to considering the sample size of each autonomous vehicle, attention should also be paid to the local model update parameters. The freshness.

[0126] Based on this, parameters are updated for the numerical local model. The degree of freshness is further defined by the parameter age as follows:

[0127]

[0128] in, This represents the edge model aggregation that the autonomous vehicle n participated in the last time. In this invention, The specific calculation method is as follows:

[0129]

[0130] Clearly, in the z-th edge iteration, the approximate parameter values ​​of the autonomous vehicle n are... The corresponding parameter age This can be understood as the time since it last participated in edge model aggregation; the smaller the parameter age, the more approximate the parameter value. The newer the parameter, the older it is, which means that the autonomous vehicle n has not participated in edge model aggregation for a long time.

[0131] To avoid outdated local model update parameters and autonomous vehicles with small local datasets excessively influencing the aggregation of edge models, the following weight design strategy is adopted, taking into account both the number of local dataset samples for each autonomous vehicle and the parameter age of the approximate values ​​of the local model update parameters:

[0132]

[0133] Where e is the natural logarithm. Clearly, in this weighting strategy, the more local dataset samples the autonomous vehicle has, the younger the corresponding parameter age, and the greater the weight it receives when participating in edge model aggregation. Thus, the weighting strategy for single-base station multi-autonomous vehicle federated learning based on periodic aggregation has been designed.

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

[0135] Step S51: Compress the update parameters of each local model using each compression function to obtain the compressed parameters.

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

[0137] Step S53: Use each decompression function to decompress each compression parameter and obtain approximate values ​​for each parameter.

[0138] Step S54: Use the approximate values ​​of multiple parameters corresponding to each of the M base stations to aggregate the edge models and obtain M second edge models.

[0139] In this embodiment, in the first edge iteration, through the above specific operations, the second edge model corresponding to each of the M base stations can be obtained. Then, in step S6, the M base stations can broadcast their respective second edge models to each unmanned vehicle to start the next round of edge iteration.

[0140] Step S6: Replace the M first edge models with M second edge models, return to step S3, and repeat steps S3 to S5 for the first time. In the b-th repetition, b = 2, 3, ..., E-1, replace the M b-th edge models with M (b+1)-th edge models and return to step S3. This is used to broadcast to each autonomous vehicle in the autonomous vehicle set until M target edge models are obtained. 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, a second edge iteration is performed, that is, the first edge model is replaced with the second edge model, and the specific operations of steps S3-S5 are performed to obtain the third edge model; and so on, until the Eth edge iteration is performed, at which point the (E+1)th edge model is obtained. That is, target edge model

[0142] Combination Figure 2 The following section will focus on how to design the weighting strategy when parameter server 1 aggregates edge models from multiple base stations 2.

[0143] Specifically, base stations The target edge model will be obtained after E edge iterations. Then it is uploaded to parameter server 1. Parameter server 1 receives the target edge models uploaded by each base station 2. Then, the networks are aggregated to form the target global network model, as follows:

[0144]

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

[0146] In some embodiments, after obtaining M corresponding edge models in each edge iteration, a global iteration is performed to obtain a corresponding global network model. For example, after obtaining M (z+1)th edge models in the z-th edge iteration, the z-th global iteration is performed, and the obtained M (z+1)th edge models are aggregated into a global model to obtain a (z+1)th global network model. Then, in the (z+1)th edge iteration, the parameter server immediately broadcasts the (z+1)th global network model to each base station, and then replaces the M z-th edge models with the M (z+1)th edge models, returning to step S3, and repeating steps S3 to S5. Then, the (z+1)th global iteration is performed, and so on, until a target global network model that meets the actual needs is obtained.

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

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

[0149] In addition, to ensure a fair and objective comparison of the performance of different federated learning mechanisms, it is assumed that the available wireless communication resources within a fixed time period are the same for all federated learning mechanisms. More specifically, as described above, in this invention, multiple unmanned vehicles within a single base station are trained using periodic aggregation-based federated learning, assuming the period is set to T = t. max / 16, where tmax The longest time required to complete a local model training session for all autonomous vehicles.

[0150] Therefore, for the federated learning method proposed in this application, if the number of communication resources available for each edge model aggregation is U, then the number of communication resources available for synchronous federated learning is 16U. Assuming fully asynchronous federated learning occurs at time t... max Intrapolymer V AFL Each time, the amount of communication resources available is 16UT1 / t. max V AFL .

[0151] like Figure 6 and Figure 7 As shown, whether using error or accuracy as the evaluation criterion, the proposed federated learning mechanism based on a hybrid aggregation strategy exhibits superior performance compared to synchronous federated learning and fully asynchronous federated learning. This is because the proposed method integrates the advantages of both synchronous and fully asynchronous federated learning, effectively addressing the laggard problem in synchronous federated learning while overcoming the excessively frequent model parameter exchanges in fully asynchronous federated learning. Furthermore, it can be observed that as the total training time increases, the performance of the models corresponding to each federated learning mechanism gradually converges, but relatively speaking, the federated learning method proposed in this application consistently demonstrates better performance.

[0152] like Figure 8 As shown, compared to edge model aggregation which only considers the parameter age of the approximate parameters of the local model update and the sample size of the local dataset, the proposed weight design strategy demonstrates superior performance, whether measured by error or accuracy. This is because, although all vehicles are associated with the same base station, the time required to complete a local model training varies, and the edge model versions used for training the local models of each autonomous vehicle in the periodic aggregation-based federated learning framework are also different. The proposed weight design strategy comprehensively considers the parameter age of the approximate parameters of the local model update and the sample size of the local dataset for each autonomous vehicle, effectively avoiding the negative impact of outdated local model updates and vehicles with insufficient sample sizes on the performance of the trained model.

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

[0154] like Figure 9 and Figure 10As shown, regardless of the value of the time period T, as the model training time increases, both the training error and the testing error gradually decrease, meaning that both the training accuracy and the testing accuracy gradually increase. Closer observation reveals that when T = t... max When T = t, the proposed algorithm converges fastest. However, in terms of the final model's convergence performance, when T = t... max The proposed method performs best when the time period T is 16. This is because if the time period T is too large, the number of edge iterations within a fixed time is very limited, which naturally affects the performance of the obtained model.

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

[0156] Specifically, a uniform distribution of N autonomous vehicles means that each base station is associated with the same number of N / M autonomous vehicles. Conversely, a random distribution means that N autonomous vehicles are randomly distributed within the coverage area of ​​M base stations. As shown in the figure, regardless of whether the distribution is uniform or random, the training and testing errors gradually decrease as the amount of available communication resources increases. This is because, with an increase in available communication resources, each autonomous vehicle has more communication resources at its disposal. Naturally, this allows for a reduction in the compression rate of local model updates, enabling the uploading of more parameters and thus improving the performance of the trained model.

[0157] Furthermore, closer observation reveals that when the number of communication resources is fixed, the federated learning method proposed in this application performs better than random distribution when N unmanned vehicles are evenly distributed.

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

[0159] The set construction unit 101 is used to 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.

[0160] The model building unit 102 is used to build M first-edge models and an initial global network model based on a hybrid aggregation strategy; the M first-edge models correspond one-to-one with the M base stations, and the initial global network model is formed by aggregating the M first-edge models.

[0161] The distribution unit 103 is used to broadcast M first edge models to each autonomous vehicle in the autonomous vehicle set.

[0162] The model training unit 104 is used to train each local model of each autonomous vehicle using M first-order edge models to obtain the update parameters of each local model.

[0163] The model update unit 105 is used to compress the update parameters of each local model and then upload them to the corresponding base station for edge model aggregation to obtain M second edge models.

[0164] Furthermore, the model update unit 105 is also used to replace the M first edge models with M second edge models, so that the sending unit 103, the model training unit 104, and the model update unit 105 are also used to repeatedly perform the corresponding operation for the first time based on the M second edge models, and so on. When the sending unit 103, the model training unit 104, and the model update unit 105 repeatedly perform the corresponding operation for the bth time, b = 2, 3, ..., E-1, the corresponding model update unit 105 is also used to replace the M b-th edge models in the sending unit 103 with M (b+1)-th edge models. The sending unit 103 broadcasts the M (b+1)-th edge models to each autonomous vehicle in the autonomous vehicle set until the M target edge models are obtained. Then, the M target edge models are global model aggregated to obtain the target global network model.

[0165] In this application, the other technical features of the federated learning device for multi-base station unmanned vehicles are the same as those disclosed in the above method embodiments, and will not be repeated here.

[0166] Based on the same inventive concept, the present invention also provides an electronic device, which can be a control device or control system inside a smart device, or an external device that communicates with a smart device. The device can be a desktop computer, a portable computer, a smartphone, 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), digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the methods disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules in the processor.

[0168] Furthermore, 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. Memory can include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic memory, magnetic disk, optical disk, etc. Memory is any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited to this. The memory in the embodiments of this application can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0169] Therefore, this invention discloses a federated learning method, apparatus, and device for multi-base station unmanned vehicles. The method includes the following steps: First, constructing 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; Second, based on a hybrid aggregation strategy, constructing M first-order edge models and an initial global network model; the M first-order edge models correspond one-to-one with the M base stations, and the initial global network model is formed by aggregating the M first-order edge models; Third, broadcasting the M first-order edge models to each unmanned vehicle in the unmanned vehicle set; Fourth, using the M first-order edge models, training each local model of each unmanned vehicle to obtain each... The steps are as follows: First, update the parameters of the local model. Second, compress the updated parameters of each local model and upload them to the corresponding base station for edge model aggregation to obtain M second-order edge models. Third, replace the M first-order edge models with the M second-order edge models and return to step three. Repeat steps three through five for the first time, and so on. For the b-th repetition, b = 2, 3, ..., E-1, replace the M b-th edge models with M (b+1)-th edge models and return to step three. This is then broadcast to all autonomous vehicles in the autonomous vehicle set until M target edge models are obtained. Finally, aggregate the M target edge models globally to obtain the target global network model. This method effectively addresses the differences in computing and storage capabilities of multiple autonomous vehicles within a single base station and the differences in local dataset sample data. It also solves the problem of lagging behind due to vehicle malfunctions, traffic accidents, and extreme weather, as well as the excessively frequent model parameter exchanges in fully asynchronous federated learning.

[0170] The above are merely embodiments of the present invention and do not limit the scope of protection of the present invention. Any equivalent structural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are included within the scope of protection of the present invention.

Claims

1. A federated learning method for multi-base station unmanned vehicles, characterized in that, The method includes the following steps: The first step is to construct a base station set and an unmanned vehicle set, wherein the base station set includes... The unmanned vehicle cluster includes one base station. A driverless car; The second step is to build based on a hybrid aggregation strategy. A first-order edge model and an initial global network model; the The first edge model is respectively related to the... Each base station corresponds one-to-one with the others; the initial global network model is derived from the [previous model]. The first edge model is aggregated to form a; each base station can correspond to multiple unmanned vehicles, while each unmanned vehicle can only correspond to one base station, and each unmanned vehicle also corresponds to a local model; The third step is to... The first edge model is broadcast to each autonomous vehicle in the set of autonomous vehicles; Fourth step, using the aforementioned The first edge model is used to train each local model of each unmanned vehicle to obtain the update parameters of each local model. The fifth step involves compressing the update parameters of each local model and uploading them to the corresponding base station for edge model aggregation to obtain... The second edge model; Step 6, using the aforementioned The second edge model replaces the The first edge model is returned to the third step, and the third to fifth steps are repeated for the first time, and so on, until the first edge model is reached. During repeated execution E-1 represents the maximum number of times, corresponding to the use of The first Edge model replacement The first The edge model returns to the third step, which is used to broadcast to each autonomous vehicle in the set of autonomous vehicles until it receives... After the target edge model, the The target edge models are aggregated globally to obtain the target global network model.

2. The federated learning method according to claim 1, characterized in that, The second step includes: Regarding the Edge aggregation is performed on the local models of multiple unmanned vehicles associated with each base station to obtain the... The first edge model; Regarding the The first edge model is used to perform global model aggregation to obtain the initial global network model.

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

4. The federated learning method according to claim 1, characterized in that, The fifth step includes: Each local model update parameter is compressed using a compression function to obtain the compressed parameters. Each compression parameter is uploaded to its corresponding base station; The compression parameters are decompressed using the decompression functions to obtain approximate values ​​for each parameter; Using the above The edge model is aggregated by combining multiple approximate parameter values ​​corresponding to each base station to obtain the... The second edge model.

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

6. The federated learning method according to claim 5, characterized in that, The step of designing the compression functions corresponding to each unmanned vehicle based on the compression matrices includes: Design driverless cars The corresponding compression function; where, in the first... In the next edge iteration, we have: ； in, E is the maximum number of times; Indicated in the first The unmanned vehicle mentioned in the next edge iteration The corresponding compression function; Indicated in the first The unmanned vehicle mentioned in the next edge iteration The corresponding compression matrix.

7. The federated learning method according to claim 5, characterized in that, The use of the The edge model is aggregated from multiple approximate parameter values ​​corresponding to each base station, including: Using base stations Multiple corresponding parameter approximations are aggregated into an edge model; where, in the first... In the next edge iteration, we have: ； in, E is the maximum number of times; Indicated in the first The base station obtained in the next edge iteration The Edge model; Indicated in the first The base station used in the next edge iteration The Edge model; Indicated in the first In the next edge iteration, with the base station The weighted average of multiple related parameter approximations, and the weighted average It can be represented as: ； in, The unmanned vehicle The corresponding approximate parameter values; For the unmanned vehicle Approximate values ​​of the corresponding parameters The weights; Indicated in the first In the next edge iteration, with the base station A subset of autonomous vehicles consisting of multiple related autonomous vehicles.

8. The federated learning method according to claim 7, characterized in that, The method further includes: Based on the number of local dataset samples and the parameter age of the local model update parameters for each unmanned vehicle, the weights of the corresponding parameter approximations for each unmanned vehicle are designed. Among them, the unmanned vehicle Approximate values ​​of the corresponding parameters The weights are: ； in, It is the natural logarithm; The unmanned vehicle The number of samples in the local dataset; The parameter age represents the age of the parameter. It can be represented as: ； in, The unmanned vehicle The order of the last participation in edge model aggregation, and ; Indicates the first In the next edge iteration, with the base station A subset of autonomous vehicles consisting of multiple related autonomous vehicles.

9. A federated learning device for multi-base station unmanned vehicles, characterized in that, The device includes: An assembly construction unit is used to construct a base station set and an unmanned vehicle set, wherein the base station set includes... The unmanned vehicle cluster includes one base station. A driverless car; Model building units are used to build models based on hybrid aggregation strategies. A first-order edge model and an initial global network model; the The first edge model is respectively related to the... Each base station corresponds one-to-one with the others; the initial global network model is derived from the [previous model]. The first edge model is aggregated to form a; each base station can correspond to multiple unmanned vehicles, while each unmanned vehicle can only correspond to one base station, and each unmanned vehicle also corresponds to a local model; The sending unit is used to send the... The first edge model is broadcast to each autonomous vehicle in the set of autonomous vehicles; Model training unit, used to utilize the The first edge model is used to train each local model of each unmanned vehicle to obtain the update parameters of each local model. The model update unit is used to compress the update parameters of each local model and then upload them to the corresponding base station for edge model aggregation to obtain... The second edge model; the model update unit is also used to use the... The second edge model replaces the The first edge model, such that the distribution unit, the model training unit, and the model update unit are further configured to, according to the first edge model, enable the distribution unit, the model training unit, and the model update unit to further configure the distribution unit, the model training unit, and the model update unit ... the model training unit, and the model update unit to, The second edge model repeats the corresponding operation for the first time, and so on, with the distribution unit, the model training unit, and the model update unit performing the operation repeatedly. When the corresponding operation is repeated several times. The model update unit is also used to use The first The edge model replaces the sending unit in The first Edge model, which is sent by the issuing unit The first The edge model is broadcast to each autonomous vehicle in the set of autonomous vehicles until it is obtained. After the target edge model, the The target edge models are aggregated globally to obtain the target global network model.

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

Images