Layered aggregation federated learning method for unreliable vehicle user scheduling

Abstract

The invention provides a hierarchical aggregation federated learning method and system for unreliable vehicle user scheduling. The method comprises the following steps: in a vehicle layer, establishing a mixed integer nonlinear programming model through a semi-asynchronous aggregation mechanism, jointly optimizing an aggregation time point, a scheduling strategy and communication computing resources, ensuring that a local model is effectively updated, and reducing energy consumption and time overhead. In a base station layer, a screening mechanism based on auction is introduced, each base station competes with a resource state according to model quality, and an optimal base station subset is selected to execute global synchronous aggregation. According to the method, through semi-asynchronous aggregation, an auction mechanism and layered optimization, overall optimization of federal learning communication, calculation and model performance in a vehicle moving environment is realized.

Description

Technical Field

[0001] This invention relates to the field of federated learning technology, and in particular to a hierarchical aggregated federated learning method for unreliable vehicle user scheduling. Background Technology

[0002] With the rapid development of artificial intelligence technology, Federated Learning (FL), as a distributed learning paradigm that can achieve multi-party collaborative training while protecting data privacy, has shown great application potential in fields such as the Internet of Vehicles (IoV). In a typical vehicle federated learning scenario, intelligent vehicles, as user devices, use locally collected traffic data to train models and, through interaction with edge base stations or roadside units (RSUs) and cloud servers, jointly construct a global prediction model.

[0003] However, existing federated learning mechanisms face the following significant challenges in practical vehicle network deployments:

[0004] First, the unreliability and mobility of vehicle users. During operation, the communication status of vehicle equipment is severely limited by speed, road conditions, and base station coverage. This high mobility makes it highly susceptible to connection interruptions or disconnections during federated learning (i.e., unreliable users), significantly impacting training continuity. Second, the inefficiency of synchronous aggregation mechanisms. Traditional synchronous federated learning mechanisms require waiting for all selected users to upload their models before aggregation can begin. In vehicle scenarios, due to differences in computing power, remaining energy consumption, and channel bandwidth among different vehicles, the system is often limited by training the slowest "stragglers," leading to a significant reduction in overall training efficiency. Although a hierarchical aggregation architecture (cloud-edge-vehicle) has been proposed to alleviate long-distance communication pressure, an effective mechanism for dynamically managing resources and balancing model performance and energy efficiency at the edge layer remains lacking. Furthermore, in resource-constrained environments, accurately selecting high-quality vehicles for scheduling and rationally allocating power and bandwidth resources are crucial for ensuring federated learning convergence. At the same time, since the parameter server (base station / RSU) itself also has operating costs, without a scientific incentive and pricing mechanism, it is difficult to ensure that each edge node and vehicle equipment participates in the federated training mission honestly and efficiently.

[0005] To address the problems existing in the current federated learning training mechanism, there is an urgent need to design a hierarchical federated learning scheduling mechanism that can dynamically adjust aggregation time for unreliable vehicle users and achieve optimized allocation of resources and rewards. Summary of the Invention

[0006] This invention provides a hierarchical aggregation federated learning method and system for scheduling unreliable vehicle users, which solves the defects of unreliable users, frequent disconnections and low synchronization aggregation efficiency caused by vehicle mobility in the Internet of Vehicles environment, and realizes the performance improvement of federated learning in complex traffic scenarios.

[0007] In a first aspect, the present invention provides a hierarchical aggregation federated learning method for unreliable vehicle user scheduling, the method comprising:

[0008] Regional network segmentation: The vehicle network is divided into multiple sets; wherein each set includes a parameter server (such as a base station or roadside unit) that communicates with a cloud server and at least one vehicle user that communicates with the key parameter server;

[0009] Parameter server determination: Before each round of global iteration, the bids of each parameter server are determined through an auction mechanism, and multiple parameter servers participating in the scheduling in this round are determined based on the bids;

[0010] Local model training: Using the local dataset corresponding to each vehicle in the set of parameter servers participating in the scheduling, and based on the joint optimization constraints of dynamic aggregation time, transmission power and bandwidth, the local model corresponding to each vehicle is trained respectively.

[0011] Regional model aggregation: Perform edge model aggregation on the local models in the set to which the parameter servers participating in the scheduling belong, respectively, to obtain intermediate model parameters;

[0012] Global model aggregation: The intermediate model parameters are sent to the cloud server, where the regional model parameters are aggregated globally to obtain global model parameters.

[0013] Looping iteration: The global model parameters are broadcast to multiple parameter servers in the vehicle network that will participate in the next round of scheduling, and the next round of global iteration begins until the preset iteration end condition is met.

[0014] According to the method provided by the present invention, the step of determining the winning parameter server through an auction mechanism includes:

[0015] To determine the bidding utility, based on the previous round of training results, determine the joint optimization objective function of performance, time, and energy consumption for the neighbor set to which each of the key parameters belongs. Value, as a reference for the effectiveness of bidding;

[0016] To determine the winning parameter server, cloud servers tend to be selected. A higher-valued parameter server, combined with model accuracy constraints and communication latency limitations, ensures that at least one server is selected in each round. Given a number of servers, the scheduling priorities are sorted in descending order and the winning set is determined.

[0017] To determine the payment of cloud server revenue to parameter server, a pricing function based on critical value rules is adopted. The payment revenue for each winning base station is calculated and set according to the utility increment of the winning base station and the critical bid with the smallest ratio in the unselected set.

[0018] According to the method provided by the present invention, the resource scheduling and allocation during the local model training process includes:

[0019] The optimization objective is constructed and modeled as a joint optimization function encompassing model performance, aggregation time, and communication energy consumption;

[0020] Solving the above optimization problem requires consideration of physical constraints, including that vehicles must complete training uploads before leaving the base station's communication coverage area, and that the sum of the bandwidth allocations for all selected vehicles does not exceed the base station's total bandwidth.

[0021] The Alternating Multiplier Method (ADMM) is used to decompose the mixed-integer nonlinear programming problem into three subproblems: vehicle scheduling, resource allocation (power and bandwidth), and dynamic aggregation time, which are optimized separately to address the unreliability of vehicle users.

[0022] According to the method provided by the present invention, the global model aggregation step includes:

[0023] The received intermediate model parameters are temporarily stored in the buffer of the cloud server;

[0024] When the number of models collected in the buffer reaches a preset threshold, or the training time of the current model exceeds a preset time window threshold, the models in the buffer are globally aggregated to obtain the global model parameters for the next round.

[0025] This invention provides a hierarchical aggregation federated learning system for unreliable vehicle user scheduling, the system comprising:

[0026] The set partitioning module is used to divide the vehicle network into multiple sets containing parameter servers and vehicle users;

[0027] Auction scheduling module: Used to determine the set of base stations participating in the global aggregation in this round through auction mechanism and scheduling priority;

[0028] Joint optimization module: used to perform joint optimization solutions for vehicle scheduling, resource allocation, and dynamic aggregation time at the edge layer;

[0029] Layered aggregation module: performs edge model aggregation based on parameter servers and global model aggregation based on cloud servers respectively;

[0030] The global aggregation module is used to send intermediate model parameters to the cloud server. The cloud server performs global model aggregation on the intermediate model parameters to obtain the global model parameters.

[0031] Loop control module: Used for parameter distribution and controlling global iteration until the convergence condition is met.

[0032] Furthermore, 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 program to implement the hierarchical aggregate federated learning method for unreliable vehicle user scheduling as described above.

[0033] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the hierarchical aggregate federated learning method for unreliable vehicle user scheduling as described above.

[0034] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the hierarchical aggregated federated learning method for unreliable vehicle user scheduling as described above.

[0035] The beneficial effects of the technical solutions provided by some embodiments of the present invention include at least the following:

[0036] 1) The present invention provides a hierarchical aggregation federated learning method for scheduling unreliable vehicle users. It designs a dynamic scheduling mechanism for vehicle users and effectively solves the "fallback" problem caused by high-speed movement and unstable connection of vehicle users by adopting a semi-asynchronous aggregation mechanism and dynamically setting the waiting aggregation time. It avoids the defect that the system efficiency is limited by the slowest node in the traditional synchronous aggregation mechanism and improves the communication efficiency and convergence speed of model training.

[0037] 2) This invention achieves its goal by constructing a joint optimization objective function that encompasses performance, time, and energy consumption. Furthermore, the Alternating Multiplier Method (ADMM) was used to solve the Mixed Integer Nonlinear Programming (MINLP) problem, enabling refined management of transmit power, bandwidth, and computing resources. This reduced the overall energy efficiency of vehicle equipment while ensuring model performance, optimizing resource allocation and energy consumption balance. Simultaneously, by introducing vehicle dwell time constraints, the system ensured that scheduled vehicles could complete their tasks before leaving the base station's coverage area, significantly reducing the model training failure rate caused by vehicle disconnections.

[0038] 3) This invention improves the performance of the upper-level auction mechanism design. The value-guided cloud server prioritizes training base stations with higher performance and lower communication overhead, achieving optimal resource allocation at the regional network scale and reducing the social cost of the federated learning system. The auction rules introduce a pricing function based on a "critical value" rule, ensuring that the payment received by the winning base station matches its incremental utility and the level of market competition. This not only incentivizes honest participation from edge nodes but also guarantees the economic feasibility of the incentive mechanism.

[0039] 4) This invention adopts the idea of ​​regional set partitioning, which transforms the complex whole network scheduling into local optimization of multiple parameter server sets. Combining the two-layer architecture of edge aggregation and global aggregation, it significantly reduces the communication load and computing pressure of cloud servers. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0041] Figure 1 A schematic diagram illustrating the principle of the federated learning framework provided for embodiments of the present invention;

[0042] Figure 2 A flowchart illustrating a hierarchical aggregation federated learning method for unreliable vehicle user scheduling, provided as an embodiment of the present invention;

[0043] Figure 3 A schematic diagram of the structure of a hierarchical aggregation federated learning system for unreliable vehicle user scheduling provided as an embodiment of the present invention;

[0044] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0046] Currently, intelligent vehicles are mostly deployed in clusters on urban roads or highways, supporting autonomous driving and intelligent transportation tasks such as real-time traffic prediction, cooperative obstacle avoidance, and vehicle trajectory planning through machine learning technology. Unlike fixed-location computing nodes, the communication window between vehicle users and roadside units (RSUs) or base stations is limited during movement, and the bandwidth of wireless communication links is strictly constrained by environmental interference and multipath effects. Especially in federated learning scenarios that require frequent exchange of large-scale model parameters, directly migrating the traditional federated learning framework to the highly dynamic vehicle network scenario is not the optimal choice.

[0047] Existing research based on federated learning frameworks generally focuses on the direct interaction between vehicles and the cloud, which has high communication overhead. However, in the actual process of advancement, two major challenges still need to be addressed: (1) Heterogeneity of computing and communication. Differences in computing power and communication module performance among different brands or models of intelligent vehicles will directly lead to significant time differences in local model update speed and model transmission efficiency, causing interference to the overall training process and fluctuations in model accuracy. (2) Unreliability of vehicle users. Due to the characteristics of high-speed movement, the service time of a single vehicle in a specific roadside unit coverage area is naturally limited, which can easily lead to potential interruptions or disconnections in the model training process, adversely affecting the continuity and stability of training.

[0048] This invention proposes a novel hierarchical aggregation federated learning framework for scheduling unreliable vehicle users. This framework, aided by edge computing nodes (such as RSUs), determines scheduling priorities based on each vehicle user's training capabilities, communication capabilities, and dwell time, thereby selecting "higher-quality" participants for global training. Under this mechanism, each participating vehicle must complete data collection and local model training within its limited communication coverage window. Roadside units, as key nodes in the neighbor set, are responsible for collecting models from vehicles within the set and performing initial aggregation (edge ​​aggregation). Furthermore, addressing the differences in vehicle connection density and mobility across different road segments, this invention designs a server-selected aggregation mode based on an auction mechanism. Within the limited vehicle-to-ground / vehicle-to-cloud communication window, this allows more high-quality model updates to participate in global aggregation, significantly increasing aggregation frequency and ensuring training continuity.

[0049] Please see Figure 1 , Figure 1 A schematic diagram of the principle of the federated learning framework provided for embodiments of the present invention, as shown below. Figure 1 As shown, the system consists of a cloud server, multiple base stations, and vehicle users. It consists of multiple regions, each containing one and only one base station. Specifically, the cloud server ① initializes and distributes tasks: at the beginning of each iteration, the cloud server is responsible for initializing or updating global model parameters and distributing computation tasks; ② local training: vehicle users use their local datasets to train local models based on the received model parameters; ③ edge aggregation: after completing training, the selected vehicles send their local model parameters to their respective base stations via a link. The base stations perform weighted aggregation of the collected local models to obtain intermediate model parameters; ④ global aggregation: the cloud server collects the model parameters of the winning base station. When the number of models collected in the buffer reaches a preset threshold, or the current iteration duration exceeds a preset time window, the cloud server performs global model aggregation and updates the global model parameters.

[0050] The technical solution of the present invention will be described below with reference to specific embodiments.

[0051] Example 1

[0052] Please see Figure 2 , Figure 2 One of the flowcharts for a hierarchical aggregation federated learning method for unreliable vehicle user scheduling provided as an embodiment of the present invention includes:

[0053] S101. Divide the vehicle network into multiple sets according to the coverage of the base stations;

[0054] S102. Determine the vehicle users selected by each base station and calculate the objective function to determine the vehicle users participating in the scheduling in this round;

[0055] S103. Train the corresponding local models using the local datasets of the vehicle users in the set participating in the scheduling.

[0056] S104. Perform edge model aggregation on the local model parameters uploaded by vehicle users in the neighbor set to which the parameter server belongs, and obtain intermediate model parameters.

[0057] S105. Send the intermediate model parameters to the cloud server. The cloud server determines the winning base station through auction. The intermediate model parameters of the winning base station are then aggregated globally to obtain the global model parameters.

[0058] S106. Broadcast the global model parameters to the base station and enter the next round of global iteration until the preset iteration end condition is met.

[0059] This invention designs a dynamic scheduling mechanism. In each round of global training, the vehicle users participating in scheduling are determined based on the performance and energy consumption of each base station in the set. This prioritizes scheduling vehicle users who contribute more significantly to model performance and communication efficiency to participate in intermediate model aggregation, effectively balancing model accuracy and communication efficiency. Furthermore, this invention designs an auction mechanism for base stations. By calculating payment rewards through a pricing function based on a critical value, it ensures that the compensation received by the winning base station matches the incremental utility it contributes. Then, the winning base station uploads parameters to the cloud server, and finally, global model aggregation is achieved on the cloud server. This incentivizes edge nodes to participate honestly and improves the scalability of the system in large-scale vehicle networks.

[0060] In S101 of this embodiment, the vehicle network is divided into multiple sets, each set including a base station, with which vehicle users can exchange model parameters.

[0061] Specifically, because the signal coverage of a base station is limited, the number of vehicle users that can directly communicate with it is also limited. Therefore, the number of vehicle users that can communicate can be determined based on the coverage of the base station. Furthermore, the base stations in any two sets do not overlap; that is, for any two base stations in the vehicle network… .

[0062] In S102 of this embodiment, the base station determines the selected vehicle users and decides which vehicle users' local models participate in model aggregation.

[0063] First, identify the possible factors influencing base station selection. These include the total amount of local data from vehicle users; users with larger data volumes may contribute more to the model. Channel conditions also play a role; vehicles with good channel quality experience higher transmission efficiency. Computational capabilities are another factor; stronger capabilities allow for faster local training. Due to the mobility of vehicles, local training and model uploading are typically completed only once within the base station's coverage area. Therefore, each training round requires rescheduling participating vehicles and allocating communication resources to them. Simultaneously, the parameter server needs to update the model based on the dynamic aggregation time; only vehicles that complete training and upload within the aggregation time can participate in model aggregation. To ensure the effectiveness of the global model, as many vehicles as possible should be included in model aggregation. An increased number of participating vehicles often means a larger overall data scale, which is beneficial for improving model performance. However, due to the heterogeneity among vehicles, their communication and computational conditions differ significantly. Too many participating vehicles will increase communication and computational energy consumption and prolong overall time consumption. In addition, the setting of the dynamic aggregation time is also crucial: if the aggregation time is set too short, there may not be enough vehicles participating in the aggregation, thereby weakening the model performance; if the aggregation time is too long, too many vehicles may be introduced, resulting in a waste of communication and computing resources.

[0064] For example, for each base station, we statistically analyze its vehicle user data sample size, allocable bandwidth resources, dynamic aggregation time, current channel quality, and other indicators. We normalize these indicators and calculate the balance between model performance and cost through weighted summation. The mathematical expression of this balance is as follows:

[0065]

[0066] in, It is a binary variable representing whether the vehicle user has selected an edge server. Represents transmission power, Represents bandwidth, For semi-asynchronous aggregation time, An accuracy model representing a base station. Representative cost model.

[0067] Accuracy Model The expression is as follows:

[0068]

[0069] On behalf of vehicle users Local data volume and These are the fitting parameters.

[0070] The cost model includes time cost and energy cost, expressed as follows:

[0071]

[0072] This represents the dynamic waiting time for aggregation. and For balance coefficient, This refers to the communication energy consumption of the parameter server.

[0073] In the provided framework, vehicle users need to upload and download model parameters through communication with a parameter server. Orthogonal Frequency Division Multiplexing (OFDM) technology is used for communication between vehicle users, therefore interference is not considered in the communication model. Each vehicle maintains a constant wireless channel gain during one training round, and the channel gain changes in different rounds. Considering the vehicle-to-everything (V2X) scenario, the channel gain between the vehicle user and the parameter server in one round is modeled as follows:

[0074]

[0075] In the formula, For vehicle users Real-time distance to the parameter server; large-scale fading effect caused by shadowing. With path loss Common description, among which It follows a log-normal distribution, while the path loss model is... Small-scale fading effect It then follows a Rayleigh distribution.

[0076] As the training rounds progress, the vehicle's position changes, and the link between it and the parameter server will experience varying degrees of channel attenuation. Therefore, the data transmission rate will differ between rounds. Based on Shannon's theorem, the vehicle... The transmission rate to the parameter server can be calculated as follows:

[0077]

[0078] in, Indicates vehicle user The allocated communication bandwidth This represents the power used during data transmission. This is defined as channel noise power. The total bandwidth that the parameter server can provide is denoted as... For any vehicle user The energy consumption generated by communicating with the parameter server when completing a single round of training can be expressed as:

[0079]

[0080] In the formula, Vehicle users Unique transmission energy efficiency coefficient, This indicates the number of model parameters to be transmitted. Additionally, vehicle users... The expression for calculating the communication time during a training round is as follows:

[0081]

[0082] When scheduling vehicle model training, the movement characteristics of the vehicles must be fully considered. First, we assume that all vehicles travel in the same direction and that the speed of each vehicle remains constant within a single training round. At the start of each training round, the system updates the set of participating vehicles and their corresponding speed information.

[0083] When the vehicle density on the road is low, the distribution of vehicles within the communication coverage area of ​​the parameter server can be approximated by a Poisson distribution; while when the vehicle density is high, their location distribution is closer to a uniform distribution.

[0084] To more realistically reflect the dynamic vehicle environment, the speed of each vehicle is set to follow a truncated Gaussian distribution. Once a vehicle enters the communication range of the parameter server, it reports its position and speed information to the server. Based on this real-time information, the parameter server calculates the duration the vehicle can remain within the communication coverage area, using the following formula:

[0085]

[0086] in, This indicates the communication radius of the parameter server. Representative vehicle The average driving speed, and the spatial coordinates of the vehicle are set as follows. The location of the base station is defined as .

[0087] In S103 of this embodiment, the base station sends the initial model parameters for this round to vehicle users within its coverage area. In its local dataset Local training is performed on the target platform, with the goal of minimizing its local loss function in each local iteration. The local model is obtained, and the expression for the loss function is as follows:

[0088]

[0089] in, The model parameters represent the parameters of the l-th local iteration. For the sample The corresponding loss value. To optimize this loss function, the gradient needs to be calculated:

[0090]

[0091] in, On behalf of vehicle users The gradient in the l-th local iteration, Representative sample The gradient of the corresponding loss function.

[0092] With learning rate Update the local model parameters using the following formula:

[0093]

[0094] in, This represents the model parameters for the (l+1)th local iteration.

[0095] In S104 of this embodiment, for each vehicle user participating in scheduling determined in S102, local model parameters are uploaded to the base station for intermediate model aggregation.

[0096] Assume that the K base stations participating in the scheduling in this round of global iteration have been determined. For each base station For each k=1,2,...,K, perform the following local training and intermediate model aggregation steps simultaneously:

[0097] a. Base station Receive the initial model parameters sent by the cloud server and distribute the initial model parameters to each vehicle user in its set;

[0098] b. The vehicle users in this set perform [actions] on their respective local datasets. In the local training round, after local training is completed, the selected vehicle user transmits the trained local model to the base station. ;

[0099] c. When the base station After receiving all the trained local models, Star Perform intermediate model aggregation, and then redistribute the aggregated intermediate models to each vehicle user. Repeat steps b and c until the process is complete. Edge aggregation yields the final intermediate model parameters;

[0100] d. via base station Will be completed The intermediate model parameters obtained from edge aggregation are transmitted to the cloud server.

[0101] In step S105 of this embodiment, the cloud server performs global model aggregation using a weighted average method based on the intermediate model parameters sent by the winning base station to obtain new global model parameters. Specifically, the base station publishes a bidding price according to the optimization objective function Q, and the cloud server formulates auction rules to select the set of winning base stations. For each winning base station in the set The algorithm will search for a critical node among the base stations that have not won. This node represents the optimal alternative the system can choose if the current winner exits the competition, and its unit utility cost (cost). With utility The ratio of ( ) is considered as the benchmark threshold for pricing.

[0102] The pricing function used in this invention is not a simple fixed-price payment, but rather achieves payment through utility equivalence compensation. The setting and the utility increment provided by the winning base station This means that the better the training performance of vehicles covered by the base station (e.g., large data volume, high accuracy, low energy consumption), the higher its baseline payment level. Furthermore, the payment function introduces a critical ratio as a regulating factor. When external competition is intense (the critical ratio is small), it indicates the existence of efficient substitutes in the market, and the payment function will lower the premium to save social costs; conversely, if the winning base station has a significant advantage or resource scarcity, the payment function will provide higher rewards to incentivize its continued participation. The expression for the payment function is as follows:

[0103]

[0104] In S106 of this embodiment, the cloud server receives... After receiving the intermediate model from the winning base station, the cloud server will perform global model aggregation on the received intermediate model parameters, and use the aggregated global model parameters as the initial model parameters for the next round of global training, entering the next round of global iteration, until the preset iteration termination condition is reached.

[0105] This invention determines dynamic aggregation time and user scheduling and resource allocation strategies by maximizing the formulation of an optimization problem, and constructs the following optimization objective:

[0106]

[0107]

[0108]

[0109]

[0110]

[0111]

[0112] in , , B represents the total bandwidth that the base station can be allocated. The constraint is a binary number. When, it means no user is selected. The model, When, it represents the selection of users. The model is designed to limit the bandwidth allocated by the base station to all vehicles to no more than the total bandwidth B. Furthermore, selected users must complete a training round and upload model parameters before leaving the base station's communication coverage area. The determination of the dynamic aggregation time must ensure that at least one user completes a training round and uploads the model before global model aggregation.

[0113] The aforementioned optimization problem is a mixed-integer nonlinear programming (MINLP) problem. MINLP problems are powerful in handling optimization problems involving integer variables and nonlinear constraints. The alternating multiplier method is employed for solving this problem. This method introduces multipliers to integrate the constraints of the original problem into the objective function, iteratively approximating the optimal solution. This effectively addresses complex issues in federated learning, such as user scheduling, resource allocation, and aggregation time determination.

[0114] Example 2

[0115] Please see Figure 3 , Figure 3 A schematic diagram of a hierarchical aggregation federated learning system for unreliable vehicle user scheduling, provided as an embodiment of the present invention, is shown. The system includes:

[0116] The set partitioning module 201 is used to divide the vehicle network into multiple sets; each set includes a base station that communicates with the cloud server and a group of vehicle users participating in training.

[0117] Auction scheduling module 202 is used to determine the winning base station participating in global model aggregation, and determines the participants in this round of aggregation based on the critical value auction rule. One base station;

[0118] The local training module 203 is used to train the corresponding local model for each vehicle user in the set to which the base station belongs.

[0119] The joint optimization module 204 is used to perform edge aggregation on the local model of the selected vehicle user in each base station set to obtain intermediate model parameters;

[0120] The global aggregation module 205 is used to send intermediate model parameters to the cloud server. The cloud server performs global model aggregation on the intermediate model parameters to obtain global model parameters.

[0121] The iterative loop module 206 is used to broadcast global model parameters to the vehicle network and enter the next round of global iteration until the preset iteration end condition is reached.

[0122] The hierarchical aggregation federated learning system for unreliable vehicle user scheduling described above can be referred to in correspondence with the hierarchical aggregation federated learning method for unreliable vehicle user scheduling described in the various method embodiments above.

[0123] Figure 4The schematic diagram of an electronic device provided in an embodiment of the present invention includes: a processor 310, a communication interface 320, a memory 330, and a communication bus 340. The processor 310, communication interface 320, and memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute a hierarchical aggregation federated learning method for unreliable vehicle user scheduling provided by the methods described above.

[0124] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute a hierarchical aggregated federated learning method for unreliable vehicle user scheduling provided by the above methods.

[0126] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform a hierarchical aggregated federated learning method for unreliable vehicle user scheduling provided by the methods described above.

[0127] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0128] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hierarchical aggregated federated learning method for unreliable vehicle user scheduling, characterized in that, The method includes: Vehicle-level scheduling steps: The cloud server sends computing tasks and an initial global model to the parameter server, which then distributes them to vehicle user devices within the coverage area. The vehicle users participating in the scheduling use their corresponding local datasets to train their respective local models. Base station layer aggregation steps: The parameter server sets a dynamic waiting time T through a resource management algorithm, and selects a group of vehicle users who can complete training and model upload within this time. The local model aggregation of the selected vehicle users is completed at the parameter server to obtain intermediate model parameters. Cloud auction steps: Before each round of global iteration, determine a set of parameter servers selected by the cloud server, and determine multiple parameter servers participating in the scheduling in this round based on the auction results; Global aggregation step: The intermediate model parameters are sent to the cloud server, and the intermediate model parameters are aggregated globally on the cloud server to obtain the global model parameters; The global model parameters are broadcast to all parameter servers in the vehicle network to enter the next round of global iteration until the preset iteration end condition is met.

2. The hierarchical aggregation federated learning method for unreliable vehicle user scheduling according to claim 1, characterized in that, The parameter server schedules and allocates resources to vehicle users by maximizing an objective function Q, which is modeled as an accuracy model. -Cost Model : It is a binary variable representing whether the vehicle user has selected an edge server. Represents transmission power, Represents bandwidth, This is a semi-asynchronous aggregation time function used to quantize complex multi-objective optimization problems in federated learning.

3. The hierarchical aggregation federated learning method for unreliable vehicle user scheduling according to claim 2, characterized in that, The accuracy model for determining the parameter server based on the amount of local data of the vehicle users participating in the scheduling includes: Determine the amount of data in the local dataset for each vehicle user. ; Determine the accuracy model of the parameter server : By the amount of data in the local dataset and Fitting the relationship and determining the parameters and .

4. The hierarchical aggregation federated learning method for unreliable vehicle user scheduling according to claim 2, characterized in that, The cost model of the parameter server consists of time cost and energy cost, including: This represents the dynamic waiting time for aggregation. and For balance coefficient, This refers to the communication energy consumption of the parameter server.

5. The hierarchical aggregation federated learning method for unreliable vehicle user scheduling according to claim 2, characterized in that, The optimization problem has the following constraints, including: Vehicle selection variables ,when When =1, the vehicle user is selected; The sum of the bandwidth allocated to all vehicles by the base station cannot exceed the total bandwidth. ; The selected vehicles must complete the training and upload before leaving the base station's communication coverage area (i.e., during the stay time); The dynamic aggregation time must ensure that at least one vehicle user completes one round of training and uploading.

6. The hierarchical aggregation federated learning method for unreliable vehicle user scheduling according to claim 1, characterized in that, The specific winner selection logic of the auction mechanism includes: The cloud server initializes an empty set of selected base stations. ; Under the conditions that the relationship between the number of iterations and accuracy is satisfied, and the communication time and channel capacity meet the maximum allowable time limit, the bid with the lowest cost is cyclically searched from the qualified bid set to be added to the set. Until the set size reaches the preset minimum number of base stations. ; Set the decision variable corresponding to the selected bid to 1, and remove the selected bid from the candidate set.

7. The hierarchical aggregation federated learning method for unreliable vehicle user scheduling according to claim 6, characterized in that, The auction mechanism also includes a reward payment calculation step based on a threshold rule, including: Determine the critical bid: For each selected base station, find the critical bid that minimizes its unit utility ratio in the set of unselected bids; The cloud server calculates its payment based on the utility increment of the winning parameter server and the cost-utility ratio of the critical competitive bid through a pricing function. The pricing function aims to ensure that the payment received by the winning parameter server is positively correlated with its performance contribution to the system and to provide dynamic compensation based on the degree of market competition.

8. A hierarchical aggregation federated learning system for unreliable vehicle user scheduling, characterized in that, The system includes: The set partitioning module is used to divide the vehicle network into multiple sets; each set includes a base station that communicates with the cloud server and a group of vehicle users participating in training. The auction scheduling module is used to determine the winning base station participating in global model aggregation, and determines the participants in this round of aggregation based on the critical value auction rule. One base station; The local training module is used to train the corresponding local model for each vehicle user in the set to which the base station belongs. The joint optimization module is used to perform edge aggregation on the local models of selected vehicle users in each base station set to obtain intermediate model parameters; The global aggregation module is used to send intermediate model parameters to the cloud server. The cloud server performs global model aggregation on the intermediate model parameters to obtain the global model parameters. The loop iteration module is used to broadcast global model parameters to the vehicle network, enter the next round of global iteration, and control the global iteration until convergence.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the hierarchical aggregated federated learning method for unreliable vehicle user scheduling as described in any one of claims 1 to 7.

Images