Vehicle networking federated learning system and method for constructing joint optimization model

Abstract

The application discloses a federated learning system for Internet of Vehicles and a method for constructing a joint optimization model, and the federated learning system for Internet of Vehicles comprises a vehicle terminal, a roadside unit and a V2X wireless communication link; the V2X wireless communication link is disconnected in the vehicle, the V2X wireless communication link is connected to the roadside unit, and the roadside unit is connected to the vehicle terminal; the method for constructing a joint optimization model comprises the following steps: step S1, constructing an optimization objective function; step S2, establishing a constraint system of the optimization objective function; step S3, constructing a coupling relationship between federated learning and communication according to the objective function and the constraint system; step S4, constructing an optimization problem according to step S3; and step S5, solving the optimization problem by using an adaptive alternating optimization algorithm; and the application solves the problems that the prior art cannot simultaneously consider communication reliability and resource utilization efficiency and lacks a federated learning optimization mechanism capable of performing real-time power regulation and bandwidth allocation according to vehicle dynamic characteristics.

Description

Technical Field

[0001] This invention relates to the field of vehicle networking and intelligent transportation systems, and particularly to a method for constructing a federated learning system for vehicle networking and a joint optimization model. Background Technology

[0002] With the rapid development of vehicle-to-everything (V2X) and intelligent transportation systems, data interaction between vehicles and roadside infrastructure is becoming increasingly frequent. Massive amounts of multi-source, heterogeneous data provide a rich information foundation for applications such as autonomous driving, vehicle-road cooperative perception, traffic prediction, and intelligent scheduling. However, these applications heavily rely on machine learning (ML) models for rapid processing and real-time decision-making regarding vehicle perception data. Traditional centralized machine learning methods require uploading data to the cloud for unified training, which not only incurs extremely high communication loads but also poses serious privacy and security risks.

[0003] To address the aforementioned issues, Federated Learning (FL) has been introduced into the connected vehicle environment as a distributed intelligent training mechanism. Its core idea is to achieve distributed model collaborative optimization by training models locally in each vehicle and uploading only parameter updates without uploading the original data, thereby effectively reducing communication overhead while ensuring data privacy. However, applying Federated Learning to the connected vehicle environment still faces many challenges, primarily including the unreliability of wireless communication and the inefficiency of communication resource allocation.

[0004] First, the high-speed mobility and complex channel environment in vehicular networks make wireless communication links extremely unstable. During vehicle operation, signal fading is frequent, interference fluctuations are severe, and connection interruptions are highly probable, leading to model upload failures or excessive latency. Traditional federated learning algorithms typically assume stable and available communication links. When some vehicles are unable to upload model updates due to channel fading or bandwidth limitations, the server's global aggregation process will experience data loss, resulting in slower model convergence and decreased accuracy. Furthermore, the significant differences in signal-to-noise ratio (SNR) among different vehicles make it difficult for fixed power allocation strategies to guarantee communication reliability, especially in low SNR environments where model update loss is particularly prominent.

[0005] Secondly, communication resources (such as bandwidth, power, and resource blocks) in the vehicle-to-everything (V2X) environment are highly limited, while the number of vehicles is large and their service loads vary significantly. Existing federated learning methods often employ uniform or random allocation strategies, failing to dynamically adjust resource allocation based on the vehicle's channel state and model importance. This approach not only wastes resources but also causes low-channel-quality nodes to occupy excessive bandwidth, thereby exacerbating communication latency and energy consumption, significantly impacting the convergence speed of the global model and the overall system efficiency.

[0006] Therefore, achieving reliable transmission and adaptive resource allocation in dynamic vehicular networks has become a key issue in improving the performance of federated learning. To address this problem, academia has proposed a series of improvement schemes, such as vehicle selection mechanisms, power control optimization, and joint scheduling methods. For example, J. Zhao et al. proposed a dynamic vehicle selection algorithm based on a feedback mechanism to alleviate model bias caused by data heterogeneity; H. Xiao et al. studied a joint vehicle selection and resource optimization method, improving system efficiency by considering vehicle position and speed characteristics. However, these schemes still suffer from insufficient scalability and real-time performance in large-scale, rapidly changing vehicular network environments.

[0007] In summary, existing technologies cannot simultaneously ensure both communication reliability and resource utilization efficiency, lacking a federated learning optimization mechanism capable of real-time power adjustment and bandwidth allocation based on vehicle dynamic characteristics. This problem is particularly prominent in 5G-V2X (Vehicle-to-Everything) network environments. To address this, this invention proposes a federated learning method for vehicle-to-everything networks based on Adaptive Resource Alternation Optimization (ARAO). By jointly optimizing power allocation and resource block allocation, it achieves simultaneous improvement in communication reliability and learning performance, providing a new technical approach for robust training of intelligent learning systems in vehicle-to-everything networks.

[0008] Existing technology 1: Current connected vehicle federated learning systems typically employ the traditional federated averaging (FedAvg) algorithm, implementing distributed training and aggregation of models in mobile edge computing (MEC) or vehicle-to-infrastructure (V2I) environments. The core idea of ​​this approach is that each vehicle trains its model locally using its own collected dataset. After a certain number of iterations, the model parameters (or gradients) are uploaded to a roadside unit (RSU) or a central server. The server then performs a weighted average of the model parameters uploaded by each vehicle to obtain new global model parameters.

[0009] The updated global model is then distributed to the vehicle side for the next round of training.

[0010] At the communication level, this scheme assumes that vehicles maintain a stable connection with the RSU via 5G or LTE-V2X wireless channels. Each vehicle shares limited communication resources, such as bandwidth, power, and resource blocks (RBs), to complete model uploading and global aggregation. Within this framework, some studies (such as J. Zhang et al. and H. Xiao et al.) have proposed methods for jointly optimizing vehicle selection and resource allocation, by selecting vehicles with better channel conditions in each training round or using heuristic allocation strategies to improve system efficiency. Furthermore, to address model uploading failures caused by channel fluctuations, some improved algorithms introduce fixed power control or channel threshold detection mechanisms during the model uploading phase to improve the uploading success rate and the stability of model aggregation.

[0011] In summary, the typical technical solution of the existing technology is mainly based on the FedAvg framework, which realizes federated model training through fixed or simple power control and bandwidth allocation, and has the advantages of simple implementation and low communication overhead.

[0012] The drawback of existing technology 1: Existing federated learning methods for vehicular networks based on the FedAvg framework typically rely on fixed communication and computing resource allocation mechanisms to achieve local model training and roadside unit aggregation on the vehicle side. However, in the highly dynamic 5G-V2X vehicular network environment, these methods have significant shortcomings in communication reliability, resource allocation efficiency, and model convergence performance, making it difficult to meet the comprehensive requirements of vehicular networks for real-time performance, stability, and high-precision learning.

[0013] (1) Insufficient communication reliability leads to a decline in model aggregation performance. The FedAvg algorithm is designed based on the assumption that the communication link between the vehicle and the RSU is stable, with sufficient bandwidth and controllable latency. However, in the actual vehicle network environment, vehicles move at high speeds, signal blockage and multipath interference occur frequently, causing instantaneous changes in channel state and drastic fluctuations in signal-to-noise ratio. As a result, vehicles are prone to packet loss, delays or interruptions when uploading local model parameters, causing some nodes' model updates to be unable to participate in global aggregation. When the aggregated data is incomplete, the global model is prone to bias towards vehicles with high channel quality, resulting in unbalanced model training, slower convergence speed and decreased final accuracy. Especially in low signal coverage areas or dense traffic scenarios, the packet loss rate increases significantly, and model fluctuations are more obvious. The research results show that when the channel reliability drops below 0.8, the global model accuracy decreases by about 10% to 15% compared to a stable environment, which fully demonstrates that communication fluctuations have a significant impact on federated learning performance.

[0014] (2) Fixed resource allocation leads to low system utilization and energy efficiency. Traditional FedAvg algorithms generally adopt a static average bandwidth or fixed power allocation strategy during the communication phase, meaning all vehicles occupy the same communication resources in the same training round. However, there are significant differences in channel quality, location information, and load status among vehicles. This one-size-fits-all allocation method fails to dynamically adjust according to the real-time channel status or task importance of the nodes. As a result, vehicles with poor channel conditions frequently fail to upload due to insufficient bandwidth or power mismatch, while vehicles with good channel conditions cannot fully utilize idle resources, resulting in a decrease in overall bandwidth utilization. At the same time, the fixed transmit power strategy is difficult to guarantee transmission success rate in weak channel environments, while excessively high power in strong channel environments causes energy waste and channel interference, reducing the overall energy efficiency of the system. In addition, due to unreasonable resource allocation, the upload delay of each node varies greatly, further causing asynchronous global aggregation time, thus affecting the convergence stability and real-time performance of the model.

[0015] In summary, existing FedAvg-based federated learning methods for vehicular networks suffer from two major problems in the dynamic communication links and complex channel conditions of 5G-V2X environments: poor communication reliability and inefficient resource allocation. These shortcomings make it difficult to balance transmission stability and model convergence performance in practical deployments, severely limiting the effectiveness of federated learning in intelligent computing scenarios for vehicular networks.

[0016] Existing technology two: Existing technology two mainly addresses the problems of low communication efficiency and slow model convergence in the traditional FedAvg algorithm by proposing a federated learning scheme that jointly optimizes vehicle selection and resource allocation. This type of scheme dynamically selects participating vehicles by simultaneously considering factors such as vehicle channel state, location, speed, and data quality during model training, and optimizes the allocation of communication resources accordingly, aiming to improve model convergence performance and communication efficiency under limited bandwidth conditions.

[0017] The typical technical solution can be summarized as follows: In each round of federated learning training, the central server (or roadside unit) prioritizes vehicles with stable channel conditions, large data volume, or high update contribution to participate in training based on the channel information, signal-to-noise ratio, and local data scale uploaded by the vehicles. Subsequently, the server uses optimization algorithms (such as linear programming, convex optimization, or heuristic algorithms) to allocate radio resource blocks (RBs) to minimize the total system latency or maximize the update rate of the global model. For example, M. Yang et al. proposed two algorithms to quantify the noise ratio and strategically select vehicles, thereby minimizing the adverse effects of noise labels while minimizing additional communication overhead; T. Yang et al. proposed a dynamic connection priority strategy, which prioritizes allocating more resources to vehicles with higher communication link quality to improve the overall communication success rate and learning accuracy.

[0018] In summary, the second existing technology adds a vehicle selection and resource joint optimization mechanism to the traditional federated learning architecture, which can improve the utilization rate of communication resources and learning efficiency to a certain extent, and has high theoretical research value and engineering feasibility.

[0019] The shortcomings of existing technology 2 are as follows: 1. Joint vehicle selection and resource allocation methods typically rely on periodic state acquisition and centralized optimization to generate the participating vehicles and resource schemes for the current round. However, the vehicle network channel fluctuates drastically in milliseconds due to vehicle movement and shadows / interference. The time overhead of optimization calculation and distribution means that the scheme is no longer compatible with the actual channel state when it reaches the execution plane, resulting in "outdated decisions". This lag directly causes two types of mismatch: one is resource mismatch, for example, a high SNR vehicle may enter an obstructed area at the execution time and lose bandwidth / power guarantee, or a low SNR vehicle may be assigned to an RB that is difficult to support uploading, resulting in uploading failure, latency accumulation and increased energy consumption.

[0020] 2. Learning mismatch: Upload gaps and delays result in incomplete global aggregation samples and shifted contribution weights, exacerbating convergence oscillations and accuracy decline under non-independent and identically distributed (Non-IID) conditions. Training iterations are forced to slow down to wait for "lagging" nodes, further amplifying latency and instability. Since existing solutions are mostly driven by communication layer objectives (instantaneous rate / latency) and lack joint modeling for the federated learning convergence process, decision lag not only impairs link reliability but also accumulates as bias and divergence risks on the aggregation side. The root cause of these problems lies in the fact that the non-convex coupling between fast, multi-dimensional state inputs (location, velocity, SNR, interference) and high-dimensional discrete / continuous decision variables (power, RB, participation set) requires long computation and convergence times, making it difficult to meet the millisecond-level closed-loop requirements of 5G-V2X scenarios. Therefore, it is difficult to achieve a stable, low-latency, and scalable federated training process in real-world deployments. Summary of the Invention

[0021] This invention addresses the problems in existing technologies that cannot simultaneously ensure communication reliability and resource utilization efficiency, and lack federated learning optimization mechanisms capable of real-time power adjustment and bandwidth allocation based on vehicle dynamic characteristics. It proposes a method for building federated learning systems and joint optimization models for vehicle networks.

[0022] The technical solution of the present invention is as follows:

[0023] The vehicle-to-everything (V2X) federated learning system includes: vehicle terminal, roadside unit, and V2X wireless communication link; the vehicle disconnects from the V2X wireless communication link, the V2X wireless communication link connects to the roadside unit, and the roadside unit connects to the vehicle terminal.

[0024] Preferably, the vehicle terminal reports channel status information to the roadside unit (RSU) via the V2X wireless communication uplink. After receiving the reported channel status, the RSU performs an ARAO alternating optimization mechanism based on link quality and resource constraints to obtain the transmission power and resource block allocation scheme for the current wheel of the vehicle. The vehicle completes model training locally according to the allocated uplink resources and uploads the model update via the V2X wireless communication link. The RSU determines the success of the uploaded model update and performs global model aggregation on the set of vehicles that have successfully received the update. After the model aggregation is completed, the global model is broadcast to all vehicles by the RSU via the downlink.

[0025] On the other hand, a joint optimization model method includes the following steps:

[0026] Step S1: Construct the optimization objective function;

[0027] Step S2: Establish the constraint system for optimizing the objective function;

[0028] Step S3: Based on the objective function and constraint system, construct the coupling relationship between federated learning and communication;

[0029] Step S4: Based on step S3, construct the optimization problem;

[0030] Step S5: Solve the optimization problem using the adaptive alternating optimization algorithm.

[0031] Preferably, step S1 includes the following steps:

[0032] Step S11: Assume there are N vehicles participating in the training in the system, and each vehicle i is in the local dataset D. i Independently optimize its local loss function:

[0033]

[0034] in, This represents the loss term for sample j;

[0035] Step S12: At communication round t, the vehicle's local model is updated as follows:

[0036]

[0037] in, Indicates vehicle In communication rounds Local model parameters, For vehicles Update model parameters after local training For learning rate, For vehicles Based on its empirical loss function on its local dataset This represents the gradient of the loss function with respect to the model parameters.

[0038] Step S13: After receiving the parameters uploaded by the vehicles during the aggregation phase, the Roadside Unit (RSU) forms a global model, the expression of which is:

[0039]

[0040] in, These are the global model parameters obtained by aggregating the roadside unit (RSU) at communication round t. Let be the set of vehicles that successfully uploaded and updated in communication round t. Let i be the amount of local data for vehicle i. This represents the total data for all vehicles. This indicates whether the model update for vehicle i was successfully received; 1 indicates success and 0 indicates failure. Update the local model for vehicle i;

[0041] Step S14: The global objective function is:

[0042]

[0043] Where F(ω) represents the global loss target of the system, For vehicles Local loss function, decision variables Indicates vehicle Uplink transmit power, decision variables Indicates vehicle Use resource blocks The binary variable, 0 indicates unallocated, and 1 indicates allocated.

[0044] Preferably, the constraint system in step S2 is as follows:

[0045]

[0046] Where K is the total number of available resource blocks in the system. , The maximum permissible transmission power for the vehicle. Let SNR be the instantaneous SNR of vehicle i at time t. The minimum SNR threshold required for successful upload, η, is the minimum reliable upload probability required by the system, which is given by the channel model:

[0047] in, The distance between the vehicle and the RSU, α is the path loss index, β is the noise power, and β is a constant related to the channel characteristics.

[0048] Preferably, the upper bound of the expected loss of the system after T rounds of communication in step S3 is:

[0049]

[0050] Where T represents the communication round. These are the global model parameters after the Tth round of communication. For optimal model parameters, For the global loss function in gradient at, Representing the mathematical expectation, the constant C is related to the learning rate and data heterogeneity;

[0051] The expectation of this gradient norm is influenced by the probability of a successful upload, and its expression is:

[0052]

[0053] in, This formula represents the expected probability that vehicle i will successfully upload data, and it indicates the distance... Vehicles that are farther away or have lower power are more likely to fail to upload data. Therefore, power allocation and RB allocation directly determine the model's convergence efficiency.

[0054] Preferably, the joint optimization problem of power control and RB allocation in step S4 is as follows:

[0055]

[0056] in, Represents the global loss function. For vehicles The transmission power, For vehicles Does it occupy a resource block? binary allocation variables, Represents resource block Assigned to vehicles , Represents resource block Unassigned vehicles , This indicates the total number of vehicles uploaded during the federated learning process. This represents the total number of available resource blocks in the system. Indicates the first The set of vehicles participating in uploading during wheel communication. For vehicles At any moment Instantaneous SNR, The minimum SNR threshold required for successful upload, This represents the minimum reliable upload probability required by the system. This represents the vehicle's maximum permissible transmission power.

[0057] Preferably, step S5 includes the following steps:

[0058] Step S51: Employ variable relaxation and block optimization strategies: Use binary variables... Relaxed to continuous variables binary variables Let k represent the proportion of resource block k used by vehicle i. The coupling constraints are decoupled using the Lagrange relaxation method, and the problem is transformed into two independent convex subproblems.

[0059] Step S52: The two independent convex subproblems are: the power optimization subproblem and the resource allocation subproblem;

[0060] Step S53: In order to balance the optimization of power and bandwidth, the present invention adopts an adaptive alternating iterative mechanism to solve two independent convex subproblems.

[0061] Compared with existing technologies, the beneficial effects of this invention on the federated learning system for vehicle networks and the method for constructing a joint optimization model are as follows:

[0062] 1. This invention addresses the problem of insufficient communication reliability. To address the issue that the FedAvg algorithm, which assumes a stable channel, cannot handle upload failures caused by high-speed vehicle movement, this invention designs a channel-aware adaptive power control mechanism. This mechanism dynamically adjusts the transmit power based on the real-time signal-to-noise ratio and interference level of the vehicle, thereby effectively improving the upload success rate of nodes with low channel quality, reducing model update loss, and ensuring the continuity and robustness of global aggregation.

[0063] 2. This invention addresses the problems of inefficient resource allocation and decision-making lag. To resolve the decision-making delays inherent in traditional static allocation strategies and existing joint optimization algorithms, this invention constructs an alternating iterative framework for joint power and bandwidth optimization. This framework takes the real-time channel state of the vehicle and its upload requirements as input, and achieves rapid dynamic adjustment of power allocation and resource block allocation through a low-complexity alternating optimization algorithm. Unlike centralized global optimization, this invention's alternating update mechanism can complete resource reconfiguration in near-millisecond time during each training round, thereby significantly improving the system's real-time adaptability and resource utilization.

[0064] 3. This invention improves the stability and accuracy of model convergence. Addressing the imbalance in aggregation caused by communication latency and data heterogeneity, this invention introduces a robust aggregation weight mechanism during the global update phase. This mechanism performs weighted corrections based on the credibility of the model uploaded by each vehicle and the link status, effectively mitigating the impact of abnormal nodes or packet loss nodes. This allows the global model to maintain stable convergence even under dynamic channel conditions, thus improving overall training accuracy.

[0065] 4. This invention can simultaneously enhance communication reliability, improve resource utilization, and optimize model convergence performance in dynamic vehicle networking environments. It overcomes the technical limitations of the existing FedAvg algorithm's fragile communication and the poor real-time performance of joint optimization algorithms, providing a deployable, scalable, robust, and efficient solution for distributed intelligent learning under 5G-V2X networks.

[0066] 5. This invention has high real-time performance, and the alternating optimization subproblems are all convex problems, which can be solved in milliseconds.

[0067] 6. This invention has high reliability, and adaptive power adjustment significantly improves the upload success rate of low-channel-quality nodes.

[0068] 7. This invention features high resource utilization, with RB allocation and power decision-making coordinated to maximize system capacity.

[0069] 8. This invention has high convergence, and the performance optimization of the communication layer directly promotes the rapid and stable convergence of the model layer.

[0070] 9. After joint optimization of power and resources, the probability of successful vehicle upload is increased as follows: Therefore, the expected convergence speed of the global model is accelerated, and its theoretical bound can be expressed as: ,in, The formula, representing the average upload success rate, indicates that higher communication reliability leads to faster model convergence, demonstrating that this invention achieves effective collaboration between the physical and algorithmic layers. Furthermore, theoretical analysis and simulation verification show that this alternating optimization algorithm can quickly converge to a stable solution within a finite number of steps, with a complexity of approximately [missing information - likely a number]. Compared to traditional global search algorithms Significantly reduced, meeting the needs of millisecond-level real-time decision-making. Attached Figure Description

[0071] Figure 1 This is a system structure model diagram of the present invention.

[0072] Figure 2 This is a flowchart of the algorithm of the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0074] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0075] The vehicle-to-everything (V2X) federated learning system includes: vehicle terminal, roadside unit, and V2X wireless communication link; the vehicle disconnects from the V2X wireless communication link, the V2X wireless communication link connects to the roadside unit, and the roadside unit connects to the vehicle terminal.

[0076] In this implementation scheme, the vehicle terminal reports channel status information to the roadside unit (RSU) via the V2X wireless communication uplink. After receiving the reported channel status, the RSU performs an ARAO alternating optimization mechanism based on link quality and resource constraints to obtain the transmit power and resource block allocation scheme for the current wheel of the vehicle. The vehicle completes model training locally according to the allocated uplink resources and uploads the model update via the V2X wireless communication link. The RSU determines the success of the uploaded model update and performs global model aggregation on the set of vehicles that have successfully received the update. After the model aggregation is completed, the global model is broadcast to all vehicles by the RSU via the downlink.

[0077] A method for a joint optimization model includes the following steps:

[0078] Step S1: Construct the optimization objective function;

[0079] Step S2: Establish the constraint system for optimizing the objective function;

[0080] Step S3: Based on the objective function and constraint system, construct the coupling relationship between federated learning and communication;

[0081] Step S4: Based on step S3, construct the optimization problem;

[0082] Step S5: Solve the optimization problem using the adaptive alternating optimization algorithm.

[0083] Step S1 of this implementation plan includes the following steps:

[0084] Step S11: Assume there are N vehicles participating in the training in the system, and each vehicle i is in the local dataset D. i Independently optimize its local loss function:

[0085]

[0086] in, This represents the loss term for sample j;

[0087] Step S12: At communication round t, the vehicle's local model is updated as follows:

[0088]

[0089] in, Indicates vehicle In communication rounds Local model parameters, For vehicles Update model parameters after local training For learning rate, For vehicles Based on its empirical loss function on its local dataset This represents the gradient of the loss function with respect to the model parameters.

[0090] Step S13: After receiving the parameters uploaded by the vehicles during the aggregation phase, the Roadside Unit (RSU) forms a global model, the expression of which is:

[0091]

[0092] in, These are the global model parameters obtained by aggregating the roadside unit (RSU) at communication round t. Let be the set of vehicles that successfully uploaded and updated in communication round t. Let i be the amount of local data for vehicle i. This represents the total data for all vehicles. This indicates whether the model update for vehicle i was successfully received; 1 indicates success and 0 indicates failure. Update the local model for vehicle i;

[0093] Step S14: The global objective function is:

[0094]

[0095] Where F(ω) represents the global loss target of the system, For vehicles Local loss function, decision variables Indicates vehicle Uplink transmit power, decision variables Indicates vehicle Use resource blocks The binary variable, 0 indicates unallocated, and 1 indicates allocated.

[0096] The constraint system for step S2 of this implementation plan is as follows:

[0097]

[0098] Where K is the total number of available resource blocks in the system. , The maximum permissible transmission power for the vehicle. Let SNR be the instantaneous SNR of vehicle i at time t. The minimum SNR threshold required for successful upload, η, is the minimum reliable upload probability required by the system, which is given by the channel model:

[0099] in, The distance between the vehicle and the RSU, α is the path loss index, β is the noise power, and β is a constant related to the channel characteristics.

[0100] In step S3 of this implementation scheme, the upper bound of the expected loss of the system after T rounds of communication is:

[0101]

[0102] Where T represents the communication round. These are the global model parameters after the Tth round of communication. For optimal model parameters, For the global loss function in gradient at, Representing the mathematical expectation, the constant C is related to the learning rate and data heterogeneity;

[0103] The expectation of this gradient norm is influenced by the probability of a successful upload, and its expression is:

[0104]

[0105] in, This formula represents the expected probability that vehicle i will successfully upload data, and it indicates the distance... Vehicles that are farther away or have lower power are more likely to fail to upload data. Therefore, power allocation and RB allocation directly determine the model's convergence efficiency.

[0106] In step S4 of this implementation scheme, the joint optimization problem of power control and RB allocation is:

[0107]

[0108] in, Represents the global loss function. For vehicles The transmission power, For vehicles Does it occupy a resource block? binary allocation variables, Represents resource block Assigned to vehicles , Represents resource block Unassigned vehicles , This indicates the total number of vehicles uploaded during the federated learning process. This represents the total number of available resource blocks in the system. Indicates the first The set of vehicles participating in uploading during wheel communication. For vehicles At any moment Instantaneous SNR, The minimum SNR threshold required for successful upload, This represents the minimum reliable upload probability required by the system. This represents the vehicle's maximum permissible transmission power.

[0109] Step S5 of this implementation plan includes the following steps:

[0110] Step S51: Employ variable relaxation and block optimization strategies: Use binary variables... Relaxed to continuous variables binary variables Let k represent the proportion of resource block k used by vehicle i. The coupling constraints are decoupled using the Lagrange relaxation method, and the problem is transformed into two independent convex subproblems.

[0111] Step S52: The two independent convex subproblems are: the power optimization subproblem and the resource allocation subproblem;

[0112] Step S53: To balance the optimization of power and bandwidth, an adaptive alternating iterative mechanism is used to solve the two independent convex subproblems.

[0113] When this implementation plan is implemented,

[0114] Existing federated learning systems for vehicle-to-everything (V2X) networks still suffer from significant problems in the highly dynamic communication environment of 5G-V2X, including unstable communication, inefficient resource utilization, and slow learning convergence. These problems are limited by factors such as large communication link fluctuations, frequent channel interference, and significant differences in node distribution. Traditional FedAvg algorithms, employing fixed bandwidth and power allocation strategies, struggle to cope with dynamic channel changes, leading to model upload failures and incomplete aggregated data, thus affecting model accuracy and training stability. While joint vehicle selection and resource allocation optimization algorithms theoretically improve communication efficiency, their centralized optimization structure is computationally complex and decision-making is lagging, making it difficult to meet millisecond-level communication feedback requirements. This results in a mismatch between resource scheduling and the actual channel state, further degrading model convergence performance.

[0115] Based on the above shortcomings, the purpose of this invention is to propose a method for constructing a federated learning system and a joint optimization model for vehicle-to-everything (V2X) networks, in order to solve the key problems existing in the current technology in dynamic V2X environments. The main technical problems to be solved by this invention include:

[0116] (1) Addressing the problem of insufficient communication reliability. To address the issue that the FedAvg algorithm assumes a stable channel and cannot handle upload failures caused by high-speed vehicle movement, this invention designs a channel-aware adaptive power control mechanism. This mechanism can dynamically adjust the transmission power based on the real-time signal-to-noise ratio and interference level of the vehicle, thereby effectively improving the upload success rate of nodes with low channel quality, reducing model update loss, and ensuring the continuity and robustness of global aggregation.

[0117] (2) Solving the problems of inefficient resource allocation and decision lag. To address the decision delay issues of traditional static allocation strategies and existing joint optimization algorithms, this invention constructs an alternating iterative framework for joint optimization of power and bandwidth. This framework takes the real-time channel state of the vehicle and its upload requirements as input, and achieves rapid dynamic adjustment of power allocation and resource block allocation through a low-complexity alternating optimization algorithm. Unlike centralized global optimization, the alternating update mechanism of this invention can complete resource reconfiguration in nearly milliseconds during each training round, thereby significantly improving the system's real-time adaptability and resource utilization.

[0118] (3) Improve the stability and accuracy of model convergence. To address the imbalance problem caused by communication delay and data heterogeneity, this invention introduces a robust aggregation weight mechanism in the global update stage. The weighted correction is performed based on the credibility of the model uploaded by each vehicle and the link status, which effectively reduces the impact of abnormal nodes or packet loss nodes, so that the global model can still maintain stable convergence under dynamic channel conditions and improve the overall training accuracy.

[0119] Through the above technical solutions, the present invention can simultaneously enhance communication reliability, improve resource utilization, and optimize model convergence performance in dynamic vehicle networking environments. It overcomes the technical limitations of the existing FedAvg algorithm's fragile communication and the poor real-time performance of the joint optimization algorithm, and provides a deployable, scalable, robust, and efficient solution for distributed intelligent learning under 5G-V2X networks.

[0120] (I) Overall System Structure

[0121] This invention addresses the federated learning scenario in vehicle-to-everything (V2X) communication. It comprises three parts: a vehicle terminal, a roadside unit, and a V2X wireless communication link. A federated learning and adaptive resource optimization mechanism is run on this link, forming a closed loop of "local training—uplink upload—global aggregation—downlink distribution." The overall system structure is as follows:

[0122] 1. Vehicle terminals (Clients): Each with its own private dataset Train a local model on the local model to obtain a local model or an incremental model. Vehicles periodically report channel status (such as SNR, instantaneous channel gain) to the RSU. The system uses side information such as location / velocity for resource optimization and completes uplink transmission for model updates according to the configuration after the resource and power strategies are determined.

[0123] 2. Roadside Unit (RSU) (Parameter Server): Responsible for: 1. Collecting vehicle status and performing power-resource block alternation optimization, issuing transmit power and bandwidth / RB allocation for the current wheel; 2. Updating the local set upon successful reception. Perform global model aggregation and generate a new round of global models. Global model deployment is performed via the downlink. Under the baseline FedAvg mechanism, the aggregation of RSUs can be represented as:

[0124]

[0125] in, Let be the set of vehicles that successfully uploaded and updated in communication round t. This represents the amount of local data (number of samples) for vehicle i. This represents the total data for all vehicles. This indicates whether the model update from vehicle i was successfully received (1 for success, 0 for failure). RSU simultaneously maintains a joint objective and constraints for learning convergence and communication reliability, which drive the alternating optimization process.

[0126] 3. V2X Wireless Communication Link: The vehicle and RSU exchange local and global models via a V2X wireless link. The uplink is used for uploading the local model, following a block fading model and a path loss model, with an instantaneous SNR of [missing value]. ,in, For transmission power, For channel gain, This represents noise power. The probability of successful transmission of the link is determined by a threshold. Determined in conjunction with channel statistics, it can be expressed as:

[0127]

[0128] Where di(t) is the distance between the vehicle and the RSU, α is the path loss exponent, and σ is the distance between the vehicle and the RSU. 2 Let denoted as noise variance, and β as a constant related to channel characteristics. Under constraints such as total power, number of RBs, and reliability threshold, the RSU allocates power and RBs to each vehicle to improve the probability of successful upload and suppress aggregation bias caused by uplink latency and packet loss.

[0129] The system's runtime sequence is based on the collaborative operation of three parts: vehicle terminals, roadside units (RSUs), and V2X wireless communication links. Vehicle terminals report channel state information, including instantaneous SNR, channel gain, and dynamic parameters such as position and speed, to the RSUs via the V2X uplink. Upon receiving this information, the RSUs execute the ARAO alternating optimization mechanism proposed in this invention, based on link quality and resource constraints, to obtain the transmit power and resource block allocation scheme for each vehicle's current wheel. Subsequently, each vehicle completes model training locally according to the allocated uplink resources and uploads the updated model via the V2X wireless link. The RSUs determine the success of the upload and perform global model aggregation on the set of successfully received vehicles. The aggregated global model is then broadcast by the RSUs to all vehicles via the downlink as the initial model for the next round of training. The closed-loop process described above, which involves "status reporting → ARAO optimization → local training and uploading → success determination and aggregation → global model distribution", runs continuously and iteratively under the conditions of dynamic changes in the vehicle set and rapid fluctuations in the channel. This enables a robust and efficient federated learning convergence process within the constraints of reliable transmission probability and resource budget.

[0130] (ii) Constructing the optimization problem

[0131] To achieve efficient training and robust aggregation of FL (Flexible Interconnect) in vehicular networks under dynamic wireless channels, this invention establishes a joint optimization model to simultaneously minimize the global learning loss function while ensuring communication reliability and resource utilization efficiency. This model fully characterizes the training process at the vehicle end, channel state changes, and resource constraints.

[0132] 1. Optimize the construction of the objective function

[0133] Assume there are N vehicles participating in the training in the system, and each vehicle i is in the local dataset D. iIndependently optimize its local loss function:

[0134]

[0135] in, This represents the loss term for sample j;

[0136] At communication round t, the vehicle's local model is updated as follows:

[0137]

[0138] in, Indicates vehicle In communication rounds Local model parameters, For vehicles Update model parameters after local training For learning rate, For vehicles Based on its empirical loss function on its local dataset This represents the gradient of the loss function with respect to the model parameters.

[0139] After receiving parameters uploaded by vehicles during the aggregation phase, the Roadside Unit (RSU) forms a global model, the expression of which is:

[0140]

[0141] in, These are the global model parameters obtained by aggregating the roadside unit (RSU) at communication round t. Let be the set of vehicles that successfully uploaded and updated in communication round t. Let i be the amount of local data for vehicle i. This represents the total data for all vehicles. This indicates whether the model update for vehicle i was successfully received; 1 indicates success and 0 indicates failure. Update the local model for vehicle i;

[0142] The global objective function is:

[0143]

[0144] Where F(ω) represents the global loss target of the system, For vehicles Local loss function, decision variables Indicates vehicle Uplink transmit power, decision variables Indicates vehicle Use resource blocks The binary variable, 0 indicates unallocated, and 1 indicates allocated.

[0145] 2. Constraints and Variable Definitions To simultaneously consider power, bandwidth resources, and communication reliability, this invention establishes the following constraint system:

[0146]

[0147] Where K is the total number of available resource blocks in the system, and P max The maximum permissible transmission power for the vehicle. Let SNR be the instantaneous SNR of vehicle i at time t. The minimum SNR threshold required for successful upload, The minimum reliable upload probability required by the system is given by the channel model:

[0148] in, For the distance between the vehicle and the RSU, For path loss index, For noise power, This is a constant related to channel characteristics.

[0149] 3. The Coupling Relationship Between Federated Learning and Communication

[0150] The convergence speed of the FL model depends on the number of vehicles that successfully upload and the channel conditions. Based on the derivation in the paper, the upper bound of the expected loss after T rounds of communication is:

[0151]

[0152] Where T represents the communication round. These are the global model parameters after the Tth round of communication. For optimal model parameters, For the global loss function in gradient at, Representing the mathematical expectation, the constant C is related to the learning rate and data heterogeneity;

[0153] The expectation of this gradient norm is influenced by the probability of a successful upload, and its expression is:

[0154]

[0155] in, This formula represents the expected probability that vehicle i will successfully upload data, and it indicates the distance... Vehicles that are farther away or have lower power are more likely to fail to upload data. Therefore, power allocation and RB allocation directly determine the model's convergence efficiency.

[0156] 4. Construct the optimization problem

[0157] Considering the above factors, this invention defines the joint optimization problem of power control and RB allocation as follows:

[0158]

[0159] This problem is a mixed-integer nonconvex optimization problem (MINLP), which has high complexity when solved directly and is difficult to execute in real time in highly dynamic vehicle-to-everything (V2X) environments. Therefore, this invention employs a convex relaxation and alternating optimization strategy for solving the problem, and its mathematical principles and implementation process will be explained in detail in the following sections.

[0160] (III) Adaptive Alternating Optimization Algorithm

[0161] because Given discrete variables, this problem belongs to mixed-integer nonlinear programming, and the direct solution complexity is O(n log n). This method is not suitable for online solutions. To reduce the complexity of the solution, this invention employs a variable relaxation and block optimization strategy: binary variables... Relaxed to continuous variables , representing the proportion of resource block k used by vehicle i; the coupling constraints are decoupled using the Lagrangian relaxation method, transforming them into two independent convex subproblems.

[0162] 1. Power Optimization Subproblem

[0163] In a fixed resource allocation scheme, Under the condition of optimal transmission power set It can be modeled as:

[0164]

[0165] in, This represents the bandwidth of resource block k. Channel gain when vehicle i uses resource block k. This is the noise power spectral density. This is about... The standard convex optimization problem can be solved using the KKT (Karush-Kuhn-Tucker) conditions. Construct the Lagrangian function:

[0166]

[0167] in, This represents a Lagrange multiplier used to handle total power constraints. variable The Lagrangian function is used for solving convex optimization problems. Setting the partial derivatives to zero, we obtain the closed-form of the optimal power solution:

[0168]

[0169] in This indicates a nonnegative projection operation. Let i be the equivalent channel gain of vehicle i on its occupied resource block. Let λ be the bandwidth of the resource block obtained by vehicle i, and ln2 be the logarithmic base-changing factor, which transforms the natural logarithm into a logarithm with base 2. λ can be obtained from the total power constraint. The result is obtained through a bisection iterative method. This closed expression indicates that the power optimization satisfies a "water-filling" structure: vehicles with higher channel gain and lower noise are allocated higher power.

[0170] 2. Resource Allocation Subproblem

[0171] The transmit power of all vehicles is fixed at the optimal transmit power set. In this case, the RB allocator problem can be expressed as:

[0172]

[0173] in The optimal transmit power of vehicle i, obtained from the power optimization sub-problem in the previous stage, is constrained. This means that each RB can be allocated to at most one vehicle within an uplink time slot, constraining... This means that each vehicle can obtain at most one RB within a scheduling cycle, which is a constraint. This indicates that constraints will be continuously relaxed for convex optimization solutions. This problem is a linearly constrained convex optimization problem, and the optimal solution can be obtained using the Lagrange multiplier method or a greedy allocation algorithm. When the channel bandwidth... When the values ​​are identical and fixed, the closed-form optimal allocation solution is:

[0174]

[0175] in, This indicates that the vehicle index with the highest speed is selected. This means that resource block k is allocated to the vehicle i with the highest speed. This means that no other vehicles will be allocated the resource block. This closed-form solution shows that each resource block is always allocated to the vehicle node with the highest current channel gain (or highest effective rate) to maximize instantaneous system capacity.

[0176] 3. Adaptive Alternating Iterative Mechanism

[0177] To achieve coordinated optimization of power and bandwidth, this invention employs an alternating update strategy. The iterative process of the algorithm is as follows: Step 1: Initialization: Set the initial power... With bandwidth allocation Set an upper limit for the number of iterations. Let the convergence limit of the algorithm be set. Step 2: Power Optimization Phase: Under fixed conditions Solve below Step 3: Resource Allocation Phase: In fixed... Solve below Step 4: Convergence check: If

[0178]

[0179] Then stop the iteration; otherwise, return to step 2.

[0180] Step 5: Output the optimal solution: Obtain the joint optimal solution .

[0181] The key technical point of this invention lies in proposing a federated learning method for vehicular networks based on Adaptive Resource Alternation Optimization (ARAO). By dynamically adjusting the allocation of power and bandwidth resources, it solves the problems of communication reliability and model convergence in highly dynamic vehicular network environments. The core innovations to be protected include:

[0182] 1. A joint adaptive optimization mechanism for power and resources achieves a dual improvement in communication and learning performance;

[0183] 2. A robust federated aggregation mechanism that corrects the impact of incomplete uploaded data on the global model through weighted adjustments.

Claims

1. A federated learning system for vehicle networking, characterized in that: include: Vehicle terminal, roadside unit, and V2X wireless communication link; The vehicle disconnects from the V2X wireless communication link, which is connected to the roadside unit, and the roadside unit is connected to the vehicle terminal.

2. The federated learning system for vehicle networking according to claim 1, characterized in that, The vehicle terminal reports channel status information to the roadside unit (RSU) via the V2X wireless communication uplink. After receiving the reported channel status, the RSU performs an ARAO alternating optimization mechanism based on link quality and resource constraints to obtain the vehicle's transmit power and resource block allocation scheme for the current wheel. The vehicle completes model training locally according to the allocated uplink resources and uploads the model update via the V2X wireless communication link. The RSU determines the success of the uploaded model update and performs global model aggregation on the set of vehicles that successfully receive the update. The global model after model aggregation is completed is broadcast to all vehicles by the RSU via the downlink.

3. A method for a joint optimization model, characterized in that, Includes the following steps: Step S1: Construct the optimization objective function; Step S2: Establish the constraint system for optimizing the objective function; Step S3: Based on the objective function and constraint system, construct the coupling relationship between federated learning and communication; Step S4: Based on step S3, construct the optimization problem; Step S5: Solve the optimization problem using the adaptive alternating optimization algorithm.

4. The method for constructing a joint optimization model according to claim 3, characterized in that, Step S1 includes the following steps: Step S11: Assume there are N vehicles participating in the training in the system, and each vehicle i is in the local dataset D. i Independently optimize its local loss function: ； in, This represents the loss term for sample j; Step S12: At communication round t, the vehicle's local model is updated as follows: ； Where, ω i (t) Indicates vehicle In communication rounds The local model parameters, ω i (t+1) For vehicles Update model parameters after local training For learning rate, For vehicles Based on its empirical loss function on its local dataset This represents the gradient of the loss function with respect to the model parameters. Step S13: After receiving the parameters uploaded by the vehicles during the aggregation phase, the Roadside Unit (RSU) forms a global model, the expression of which is: ； in, These are the global model parameters obtained by aggregating the roadside unit (RSU) at communication round t. Let be the set of vehicles that successfully uploaded and updated in communication round t. Let i be the amount of local data for vehicle i. This represents the total data for all vehicles. This indicates whether the model update for vehicle i was successfully received; 1 indicates success and 0 indicates failure. Update the local model for vehicle i; Step S14: The global objective function is: ； Where F(ω) represents the global loss target of the system, For vehicles Local loss function, decision variables Indicates vehicle Uplink transmit power, decision variables Indicates vehicle Use resource blocks The binary variable, 0 indicates unallocated, and 1 indicates allocated.

5. The method for constructing a joint optimization model according to claim 3, characterized in that, The constraint system for step S2 is as follows: ； Where K is the total number of available resource blocks in the system. , The maximum permissible transmission power for the vehicle. Let SNR be the instantaneous SNR of vehicle i at time t. The minimum SNR threshold required for successful upload, η, is the minimum reliable upload probability required by the system. The success probability is given by the channel model, and its expression is: ; in, The distance between the vehicle and the RSU, α is the path loss index, β is the noise power, and β is a constant related to the channel characteristics.

6. The method for constructing a joint optimization model according to claim 3, characterized in that, The upper bound of the expected loss of the system after T rounds of communication in step S3 is: ； Where T represents the communication round. These are the global model parameters after the Tth round of communication. For optimal model parameters, For the global loss function in gradient at, Representing the mathematical expectation, the constant C is related to the learning rate and data heterogeneity; The expectation of this gradient norm is influenced by the probability of a successful upload, and its expression is: ； in, This formula represents the expected probability that vehicle i will successfully upload data, and it indicates the distance... Vehicles that are farther away or have lower power are more likely to fail to upload data. Therefore, power allocation and RB allocation directly determine the model's convergence efficiency.

7. The method for constructing a joint optimization model according to claim 3, characterized in that, The joint optimization problem of power control and RB allocation in step S4 is as follows: ； in, Represents the global loss function. For vehicles The transmission power, For vehicles Does it occupy a resource block? binary allocation variables, Represents resource block Assigned to vehicles , Represents resource block Unassigned vehicles , This indicates the total number of vehicles uploaded during the federated learning process. This represents the total number of available resource blocks in the system. Indicates the first The set of vehicles participating in uploading during wheel communication. For vehicles At any moment Instantaneous SNR, The minimum SNR threshold required for successful upload, This represents the minimum reliable upload probability required by the system. This represents the vehicle's maximum permissible transmission power.

8. The method for constructing a joint optimization model according to claim 3, characterized in that, Step S5 includes the following steps: Step S51: Employ variable relaxation and block optimization strategies: Use binary variables... Relaxed to continuous variables binary variables Let k represent the proportion of resource block k used by vehicle i. The coupling constraints are decoupled using the Lagrange relaxation method, and the problem is transformed into two independent convex subproblems. Step S52: The two independent convex subproblems are: the power optimization subproblem and the resource allocation subproblem; Step S53: To balance the optimization of power and bandwidth, an adaptive alternating iterative mechanism is used to solve the two independent convex subproblems.

Images