Federal learning training method based on vehicle state perception suitable for dynamic vehicle networking

Abstract

The present application belongs to the technical field of Internet of Vehicles and communication security, and discloses a federated learning training method based on vehicle state perception and applicable to dynamic Internet of Vehicles. The method collects vehicle running state information of vehicle clients, preliminarily screens candidate vehicle clients, calculates comprehensive scores of the screened vehicle clients, adopts a client selection strategy combining selection of the top-ranked clients with random supplementation to determine a set of vehicle clients participating in the current round of federated learning, configures local training parameters for the selected vehicle clients according to vehicle speed and computing capacity, performs local training on local data sets to obtain model updates, performs weighted aggregation on the model updates of the effective participating vehicle clients at the server end, introduces a momentum mechanism to smooth the model update process, and obtains new global model parameters. This method effectively improves the stability, robustness and training efficiency of federated learning in a high-dynamic Internet of Vehicles environment.

Description

Technical Field

[0001] This invention belongs to the field of vehicle networking and communication security technology, specifically, it relates to a federated learning training method based on vehicle state perception suitable for dynamic vehicle networking. Background Technology

[0002] With the rapid development of vehicle-to-everything (V2X) technology and artificial intelligence, federated learning, due to its advantage of enabling multi-node collaborative training without exposing raw data, has been widely applied in the field of V2X model training and optimization. Several solutions have been proposed to address issues such as data heterogeneity, privacy protection, and model training efficiency in the V2X environment.

[0003] Chinese patent CN120416804A discloses a "Method and System for Dynamic Protection of Vehicle-to-Everything (V2X) Data in Federated Learning Process." This method constructs a V2X model on a server and distributes it to multiple clients. It determines the clients participating in training based on evaluation scores and a dynamic client selection algorithm, and adjusts the learning rate for different modalities of data, thereby achieving balanced training of multimodal data. This approach improves the training effect and adaptability of multimodal V2X models to some extent. However, this method primarily focuses on the balance of different modalities of data during the learning process. Its dynamic client selection process emphasizes model evaluation results and does not fully consider the differences in vehicle states during real-world operation, such as vehicle speed, network connection stability, and changes in the operating environment. Furthermore, this method assumes relatively stable communication conditions for clients, making it difficult to adapt to the frequent topology changes caused by high-speed vehicle movement in V2X scenarios. In dynamic scenarios, there is still a risk of training interruption and unstable updates.

[0004] Chinese patent CN 119670840A discloses a "Federated Semi-Supervised Learning Method and System for Vehicle Networks," which improves model training efficiency and system real-time performance by generating pseudo-label samples on the vehicle client and combining them with a supervised model on the server for multi-round aggregate training. This method reduces reliance on large amounts of manually labeled data by introducing a semi-supervised learning mechanism, enhancing the system's adaptability in complex traffic environments. However, this approach primarily improves model training efficiency from the perspective of learning paradigm; the client's participation method and local training process still use a uniform or fixed configuration, failing to differentiate scheduling based on the operating states of different vehicles. Furthermore, this method does not model vehicle mobility and communication instability, making it difficult to guarantee the continuity and stability of the federated learning training process in the highly dynamic operating environment of vehicle networks, potentially affecting the overall convergence performance of the model.

[0005] Chinese patent CN 118821910A discloses a "Method, System, and Device for Balancing Training of Multimodal Vehicle Networking Models for Federated Learning," which improves the security and privacy protection of vehicle network data during federated learning by introducing differential privacy protection and a multi-level aggregation mechanism. This solution primarily addresses the privacy leakage risks during model parameter uploading and aggregation, enhancing the security and reliability of the vehicle network federated learning system. However, the method's technical focus is on data privacy protection and secure transmission, without addressing the optimization of the federated learning client selection strategy. Furthermore, this solution assumes a relatively stable node model update process and does not consider the impact of high-speed vehicle movement and frequent disconnections on training scheduling and model aggregation quality, making it difficult to achieve efficient and stable model training in dynamic vehicle network environments.

[0006] In summary, existing federated learning methods in dynamic vehicular network (V2N) scenarios typically select clients for training randomly or statically, failing to consider vehicle speed, connection stability, battery level, and other operational states. This can easily lead to communication interruptions or training failures, and ignores the differences in vehicle operating states. Vehicles move at high speeds in road environments, and network connections change frequently. Traditional federated learning assumes clients are always online, making it difficult to guarantee the continuity and stability of the training process and adapt to the highly dynamic topology of V2N networks. Furthermore, different vehicles have significantly different computing capabilities and operating states; uniform local training parameter configurations can easily lead to resource waste or degraded model update quality, resulting in unreasonable allocation of training resources. In highly dynamic V2N environments, frequent disconnections and low-quality updates affect the convergence speed and final performance of the global model, leading to insufficient model convergence efficiency and stability. Therefore, a federated learning training method that can perceive vehicle operating states and adapt to the dynamic environment of V2N networks is needed to further improve the practicality and engineering application value of federated learning in V2N networks. Summary of the Invention

[0007] In view of the above-mentioned technical problems and defects, the purpose of this invention is to provide a federated learning training method based on vehicle state perception suitable for dynamic vehicle networks. By introducing vehicle operating state information, the method enables adaptive selection of the federated learning client and dynamic adjustment of local training parameters, thereby improving the stability, robustness and training efficiency of federated learning in highly dynamic vehicle network environments.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A federated learning training method based on vehicle state perception suitable for dynamic vehicle-to-everything (V2X) networks, comprising the following steps:

[0010] Step 1. Before the federated learning training begins, the server initializes the global model and distributes it to each vehicle client.

[0011] Step 2. In each round of federated learning, the server first collects the vehicle operating status information of the vehicle client and obtains the network connection stability index based on the vehicle speed mapping.

[0012] Step 3. The server performs an initial screening of candidate vehicle clients based on vehicle remaining battery power and network connection stability indicators, eliminating vehicle clients that do not meet the operating conditions; then, it calculates a comprehensive score for the screened vehicle clients, and adopts a client selection strategy that combines selecting the top-ranked clients with random supplementation to determine the set of vehicle clients participating in this round of federated learning.

[0013] Step 4. For the selected vehicle client, the server adaptively configures local training parameters based on the vehicle speed and computing power, including the number of local training rounds, batch size, maximum number of training batches, and training interruption probability.

[0014] Step 5. The vehicle client receives the global model parameters sent by the server, performs local training on the local dataset according to the adaptively configured local training parameters, and uploads the updated model to the server.

[0015] Step 6. After receiving the model update uploaded by the vehicle client, the server performs weighted aggregation on the model updates that are validly involved in the vehicle client, and introduces a momentum mechanism to smooth the model update process, thereby obtaining new global model parameters.

[0016] Step 7. Repeat the above process until the preset training termination condition is met.

[0017] As a preferred embodiment of the present invention, the vehicle operating status information in step 2 includes at least: vehicle speed, vehicle computing power, vehicle remaining battery power, number of times the vehicle has participated in federated learning training in the past, and network connection stability index. The network connection stability is obtained by mapping the vehicle speed, and its calculation method adopts a piecewise function.

[0018] As a preferred embodiment of the present invention, the vehicle client comprehensive scoring in step 3 Among them, speed score Fairness score , Indicates vehicle speed, Indicators of network connection stability Indicates the vehicle's computing power, Indicates the vehicle's remaining battery power. Indicates the number of times a vehicle has participated in federal learning and training in its history. Indicates the vehicle client number. This indicates the federal learning round.

[0019] As a preferred embodiment of the present invention, the local training rounds of the vehicle client in step 4 for:

[0020] ;

[0021] Batch size of vehicle client for:

[0022] ;

[0023] Maximum number of training batches for vehicle clients in a single round of federated learning for:

[0024] ;

[0025] Training interruption probability for:

[0026] ;

[0027] When random variable And satisfy If the vehicle client is found to have experienced a training or communication interruption in the current round, its model update will not participate in the aggregation of this round.

[0028] As a preferred embodiment of the present invention, step 6 is in the... In the round of federated learning, the server first determines the aggregation weight of each vehicle client based on the amount of local data on each vehicle client: ;in: This indicates the number of local data samples that vehicle client i participated in training for in the current round. This represents the sum of local data on the client side for all participating vehicles in the current round. Indicates vehicle client Weighting coefficients in this round of aggregation Indicates the first The set of vehicle clients participating in training during the round of federated learning;

[0029] The server performs a weighted summation of the model updates uploaded by each vehicle client based on the aggregation weight, to obtain the aggregated update amount for this round: ;in: Indicates the first In federated learning, the global update direction is formed by the updates of all valid vehicle client models. Indicates vehicle client The amount of parameter updates relative to the global model.

[0030] As a preferred embodiment of the present invention, step 6 introduces a momentum mechanism during the aggregation process. The server updates the global model parameters based on the momentum variable to obtain the global model parameters for the next round of federated learning. The update method is as follows:

[0031] ;

[0032] ;

[0033] in: Indicates the momentum variable in the previous round of federated learning. Let be the momentum coefficient, satisfying ; The first The global model parameters at the start of each round of federated learning.

[0034] Advantages and beneficial effects of the present invention:

[0035] (1) Improve the stability of federated learning training in a dynamic environment of vehicle networking. This invention introduces a sense of vehicle operating state.

[0036] The proposed mechanism comprehensively considers vehicle speed, battery level, and network connection stability during the client selection phase, effectively avoiding communication interruptions and model update failures caused by frequent participation in training by highly mobile or unstable vehicles. Experiments have demonstrated that the method of this invention achieves a significantly higher average number of effectively participating clients during training compared to traditional federated learning methods, thus significantly improving the training stability of federated learning in the dynamic environment of vehicle networks.

[0037] (2) Adaptive scheduling of the local training process to improve resource utilization efficiency. Addressing the issue of significant differences in vehicle computing power, battery levels, and operating states in the vehicle-to-everything (V2X) environment, this invention adaptively configures local training parameters based on vehicle operating states, enabling different vehicle clients to participate in training within their own resource constraints. Experimental results show that the total number of local training steps in different rounds dynamically changes with the client's state, indicating that this invention can effectively reduce invalid computation and redundant communication, thereby improving the overall resource utilization efficiency of the system.

[0038] (3) Enhance the fairness and robustness of the federated learning training process. This invention constructs a comprehensive scoring model that integrates vehicle mobility characteristics, resource status and historical participation, and adopts a client selection strategy that combines optimal selection with random supplementation to avoid the problem of selecting the same batch of vehicle clients for training in a long period of time, thereby improving the fairness of different vehicle clients participating in federated learning, and enhancing the robustness of the system under dynamic vehicle conditions.

[0039] (4) Improve the smoothness and convergence reliability of the model aggregation process. This invention introduces a weighting mechanism based on local data volume during the server-side model aggregation process, combined with a momentum smoothing update method. This effectively alleviates the oscillation problem of model parameter updates when vehicle clients participate in dynamic changes in the set and training interruptions occur. Experiments have confirmed that the model update norm and momentum norm changes smoothly using the method of this invention, enabling the global model to converge more stably towards the optimal direction, thus improving the training efficiency and convergence reliability of federated learning in the vehicle network environment.

[0040] (5) Reduce system resource waste caused by invalid training rounds in dynamic vehicle networking environment. This invention effectively reduces invalid model updates and repeated training caused by frequent client disconnections or training interruptions by eliminating vehicle clients with unstable operation status or limited resources during the client selection stage and adaptively controlling the training intensity during the local training stage, thereby reducing the overall system resource waste and improving the effectiveness of the federated learning training process.

[0041] (6) Enhance the scalability of the federated learning system under varying vehicle scale. This invention enables the federated learning training process to adapt to the changing number of vehicles over time through a client selection and dynamic scheduling mechanism based on vehicle operating status. Even when the number of vehicle clients increases or decreases, it can still maintain a stable training process and model update quality, thereby improving the system's scalability in large-scale vehicle network environments. Attached Figure Description

[0042] Other objects and results of the invention will become more apparent and readily understood with reference to the following description taken in conjunction with the accompanying drawings. In the drawings:

[0043] Figure 1 This is a schematic diagram of the overall architecture provided by the present invention;

[0044] Figure 2 This is a flowchart of the filtering and selection process for federated learning clients based on vehicle operating status, as described in this invention.

[0045] Figure 3 This is a flowchart of the adaptive configuration process for local training parameters on the vehicle client of the present invention. Detailed Implementation

[0046] To enable those skilled in the art to better understand the technical solutions and advantages of the present invention, the present application will be described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of protection of the present invention.

[0047] like Figures 1 to 3 As shown, this embodiment provides a federated learning training method based on vehicle state perception suitable for dynamic vehicle networks, including the following steps:

[0048] Step 1. Initialize the global model on the server side and distribute it to multiple vehicle clients:

[0049] In this embodiment, the local training data for the vehicle client uses a public dataset (the publicly available handwritten digit image dataset MNIST), and is partitioned according to the requirements of the connected vehicle federated learning experiment. To simulate the uneven distribution of data in a connected vehicle environment, the public dataset is partitioned into non-independent identically distributed datasets based on the vehicle clients, resulting in differences in the amount and distribution of data held by different vehicle clients.

[0050] It should be noted that the specific type of the dataset in this embodiment does not constitute a limitation on the present invention. Those skilled in the art can select other public or private datasets suitable for federated learning according to the actual application scenario.

[0051] Step 2. The server collects vehicle operating status information in each round of federated learning:

[0052] In this embodiment, during each round of federated learning, the server collects vehicle operation status information from currently active vehicle clients. This vehicle operation status information includes at least: vehicle speed, vehicle computing power, vehicle remaining battery power, the number of times the vehicle has participated in federated learning training in the past, and network connection stability indicators mapped from vehicle speed. The vehicle speed is dynamically generated by the traffic simulation platform. In the simulation environment, the vehicle computing power and vehicle remaining battery power are initialized within a preset range (this embodiment uses random initialization because the vehicle's remaining battery power cannot be obtained in the simulation environment) to simulate the heterogeneity of vehicle hardware resources. The collected vehicle location information and lane information are retained as expandable status information for subsequent functional expansion but do not participate in the client selection and scheduling calculations of the current round.

[0053] Specifically, in this embodiment, the vehicle client... In the At the start of federated learning, the vehicle's operating state vector is constructed:

[0054]

[0055] in: Indicates vehicle speed, Indicators of network connection stability Indicates the vehicle's computing power, Indicates the vehicle's remaining battery power. Indicates the number of times a vehicle has participated in federal learning and training in its history. Indicates the vehicle client number. This indicates the federal learning round.

[0056] In this embodiment, the network connection stability of the vehicle is obtained by mapping the vehicle speed, and its calculation method uses a piecewise function defined as:

[0057]

[0058] This mapping relationship is used to reflect the operational characteristics of decreased communication stability and increased probability of training interruption when the vehicle is moving at high speed.

[0059] Step 3. The server filters and selects candidate vehicle clients based on the vehicle's operating status:

[0060] In this embodiment, the server performs preliminary screening of vehicle clients based on the vehicle's remaining battery power and network connection stability. The corresponding vehicle client is removed from the candidate set when any of the following conditions are met:

[0061] or

[0062] Then, for the vehicles that passed the initial screening... Construct a comprehensive scoring function. First, define the speed score. With fairness score :

[0063]

[0064]

[0065] Based on this, calculate the vehicle client's comprehensive score. :

[0066]

[0067] in: Indicates vehicle speed, Indicators of network connection stability Indicates the vehicle's computing power, Indicates the vehicle's remaining battery power. Indicates the number of times a vehicle has participated in federal learning and training in its history. Indicates the vehicle client number. This indicates the federal learning round.

[0068] Finally, the server ranks the vehicle clients based on the comprehensive score and adopts a hybrid selection strategy that combines top-ranked vehicles with random supplementation: it prioritizes selecting the top-ranked vehicle clients as core participating clients, and then randomly selects some vehicle clients from the remaining candidate vehicle clients. After merging, a final set of participating clients for this round of federated learning is formed, so as to balance the quality of model updates and the fairness of participation.

[0069] Step 4. For the selected vehicle client, adaptively configure local training parameters and perform local model training:

[0070] In this embodiment, the selected vehicle client... According to the vehicle in the The local training parameters are adaptively configured according to the running state of each round of federated learning; wherein, the local training parameters include at least: the number of local training rounds. Batch size Maximum number of training batches and probability of training interruption , .

[0071] The above training parameters are derived from vehicle speed. Vehicle computing power The overall mapping relationship, jointly determined, is expressed as follows:

[0072]

[0073] Among them, the function It is defined using a segmented mapping method.

[0074] The local training rounds for the vehicle client are defined as follows:

[0075]

[0076] This design ensures that the intensity of local training is reduced when the vehicle is moving at high speed or the connection is unstable, while the number of training rounds is increased when the vehicle speed is low and the computing power is strong.

[0077] The batch size for the vehicle client is defined as:

[0078]

[0079] By adjusting the batch size, while ensuring model training efficiency, excessive invalid computations caused by unstable communication in highly mobile vehicles can be avoided.

[0080] The maximum number of training batches for a vehicle client in a single round of federated learning is defined as:

[0081]

[0082] This parameter is used to limit the length of local training time to prevent frequent interruptions to the vehicle before training is complete.

[0083] The probability of training interruption is defined as:

[0084]

[0085] The training parameters are determined in a segmented configuration based on the vehicle's operating status. The higher the vehicle speed or the lower the computing power, the lower the local training intensity and the higher the probability of training interruption.

[0086] In this embodiment, when the random variable And satisfy If the vehicle client is found to have experienced a training or communication interruption in the current round, its model update will not participate in the aggregation of this round.

[0087] Specifically, in this embodiment, the local training parameters are adaptively configured as shown in Table 1.

[0088] Table 1: Local Training Parameter Adaptive Configuration Table

[0089] State interval Local training rounds Batch size Maximum number of training batches Training interruption probability High speed 1 64 8 0.6 medium speed 1 64 15 0.35 Low speed, high computing power 3 96 40 0.05 Low speed ordinary 3 80 40 0

[0090] Step 5. The vehicle client uploads local model updates to the server:

[0091] In this embodiment, the vehicle client is in the first... The round receives global model parameters sent by the server. and in its local dataset Perform local training to obtain model updates:

[0092]

[0093] in: Indicates the server side at the 1st The global model parameters issued at the start of the round, Indicates vehicle client The model parameters obtained after training on its local dataset Indicates vehicle client The amount of parameter updates relative to the global model.

[0094] Step 6. After receiving the model update uploaded by the vehicle client, the server performs weighted aggregation on the model updates that effectively participated in the vehicle client's update, and introduces a momentum mechanism to smooth the model update process, resulting in new global model parameters:

[0095] In this embodiment, the server side is in the... In the first round of federated learning, the local model updates uploaded by participating clients are aggregated to obtain new global model parameters. The model aggregation process includes a weighted aggregation step and a momentum smoothing update step, specifically: in the first round... In the round of federated learning, the set of participating vehicle clients is denoted as For any participating vehicle client To reflect the differences in the contribution of different vehicle clients to the global model training, the server determines the aggregation weight of each vehicle client based on the amount of local data they possess. For participating vehicle clients... Its aggregation weight is defined as:

[0096]

[0097] in: This indicates the number of local data samples that vehicle client i participated in training for in the current round. This represents the sum of local data on the client side for all participating vehicles in the current round. Indicates vehicle client The weighting coefficients in this round of aggregation Generally take , Indicates the first The set of vehicle clients participating in training during the round of federal learning.

[0098] Based on this, a momentum mechanism is introduced to smooth the model update:

[0099] Based on the aforementioned weights, the server performs a weighted summation of the model updates uploaded by each vehicle client to obtain the aggregated update volume for this round:

[0100]

[0101] in: Indicates the first In a round of federated learning, the global update direction is formed by the updates of all valid vehicle client models.

[0102] To mitigate the impact of dynamic changes in the client and training interruptions on model convergence in the connected vehicle environment, a momentum mechanism is introduced during model aggregation. The server maintains momentum variables. Its update method is defined as:

[0103]

[0104] in: Represents the momentum variable in the previous round of federated learning. Let be the momentum coefficient, satisfying Used to control the weight ratio between historical and current updated information. This indicates the proportion of the current update quantity in the momentum update.

[0105] The server updates the global model parameters based on the momentum variable to obtain the global model parameters for the next round of federated learning.

[0106]

[0107] in: Indicates the first By introducing momentum variables to smooth model updates at the start of federated learning, the global model parameters can be optimized, effectively mitigating model update oscillations caused by fluctuations in vehicle client states and changes in the participating set, thus improving the convergence stability of the model in the dynamic environment of vehicle networking.

[0108] Step 7. Repeat the above process until the preset training termination condition is met.

[0109] Furthermore, in this embodiment, a federated learning experimental system for vehicle networking is constructed, and the overall system structure is as follows: Figure 1 As shown, the system includes a federated learning server and multiple vehicle clients. The federated learning server includes a vehicle operation status acquisition module, a vehicle client filtering and selection module, a model aggregation and update module, and a global model management module. It is mainly used for global model initialization, vehicle operation status acquisition, vehicle client filtering and selection, and model aggregation and update. The vehicle clients are used for local model training and model update uploading.

[0110] In this embodiment, the vehicle operating environment is simulated through a traffic simulation platform. The simulation platform can generate real-time speed, location information and lane information of the vehicle on the road, which is used to simulate the operating characteristics of high-speed vehicle movement and frequent changes in network topology in the Internet of Vehicles environment.

[0111] Furthermore, in this embodiment, the experimental environment and parameter configuration are shown in Table 2.

[0112] Table 2: Experimental Environment and Parameter Configuration Table

[0113] project Parameter settings Dataset MNIST handwritten digit dataset Number of vehicle clients 20 Global training rounds 40 Data distribution method Non-independent and identically distributed (Non-IID) Initial model structure Convolutional Neural Network (CNN) model Vehicle client selection method FedAvg / Method of this invention Traffic simulation platform SUMO Vehicle status parameters Speed, power consumption, computing power, and historical participation count Aggregation method Weighted aggregation + momentum mechanism

[0114] Furthermore, in this embodiment, the vehicle client selection, local training, model aggregation process, and performance changes during the multi-round federated learning process are shown in Tables 3 and 4.

[0115] Table 3: Key Indicator Recording Table of Federated Training Process in this Invention

[0116] Rounds Selected number of clients Top Client Random Client Total training steps Number of disconnections Number of valid participants 1 4 v_4,v_0,v_2 v_3 480 0 4 2 7 gen_49237_5,v_2,v_3 v_2 840 0 7 3 10 v_3,gen_43638_2,gen_49237_5 gen_72022_8 870 1 9 4 12 gen_72022_8,gen_69748_7,gen_49237_5 gen_92487_9 1200 2 10 5 15 gen_44575_3,v_2,gen_72022_8 gen_25817_14 1560 2 13

[0117] Table 4: Performance Changes

[0118] Rounds Update norm Momentum Norm accuracy loss Time taken per round (s) Communication volume (MB) 1 4.4575 0.4458 0.207 2.200 9.88 3.43 2 2.9909 0.6607 0.446 2.021 15.63 6.01 3 2.2734 0.8002 0.337 1.802 17.43 7.73 4 2.0675 0.8940 0.351 1.745 22.72 8.58 5 1.8330 0.9389 0.386 1.896 29.06 11.16

[0119] As shown in Tables 3 and 4, during the multi-round federated learning process, the model update norm of the proposed method generally shows a decreasing trend, while the momentum norm gradually increases and tends to stabilize. This indicates that after introducing the vehicle state-aware scheduling and momentum smoothing aggregation mechanism, the model parameter update process is more stable and controllable (with good training stability). Even with dynamic changes in vehicle client participation states and training interruptions, the model accuracy, although fluctuating to some extent, maintains an overall upward trend, and the overall model loss decreases, demonstrating the good training robustness and optimization effectiveness of the proposed method. Furthermore, even with vehicle disconnections, the proposed method can still maintain a high number of effective participating clients. As the number of effective participating clients increases, the training time and communication volume per round increase accordingly, indicating that the proposed method can fully utilize available vehicle resources to complete effective model updates, verifying its engineering feasibility and training stability in dynamic vehicle-to-everything (V2X) environments.

[0120] In this embodiment, to verify the technical advantages of the method of the present invention in a dynamic vehicle-to-everything (V2X) environment, comparative experiments were conducted using the traditional federated learning method FedAvg and the federated learning training method based on vehicle operation status perception proposed in this invention, under the same experimental environment, dataset partitioning method, and model structure. The two methods maintained consistency in initial model parameters, training rounds, and dataset partitioning methods, differing only in the client selection strategy, local training parameter configuration method, and model aggregation mechanism. Table 5 shows that, in terms of final model performance, the method of the present invention outperforms the traditional FedAvg method in both test accuracy and test loss metrics. Specifically, the method of the present invention maintains a high level of test accuracy while further reducing test loss, indicating that by introducing vehicle operation status perception scheduling and adaptive local training mechanisms, the model parameter update process is more reasonable, contributing to improved model optimization and generalization performance. Further, Table 6 shows that, in a dynamic V2X environment, the method of the present invention exhibits significant advantages in several key training metrics. Specifically, the method of this invention can significantly reduce the average number of disconnected clients per round and increase the average number of effectively participating clients, making the federated learning training process more continuous and stable. Simultaneously, by introducing a momentum smoothing aggregation mechanism, the model update norm remains within a stable range, effectively suppressing parameter update oscillations caused by dynamic changes in client participation status. Although the training time and communication volume per round increase accordingly, these overheads mainly come from the increase in the number of effectively participating clients and the reasonable configuration of local training intensity. This indicates that the method of this invention can fully utilize available vehicle resources to complete high-quality model updates in dynamic vehicular network environments, thus making it more suitable for application scenarios involving high-speed vehicle movement and unstable communication conditions.

[0121] Table 5: Comparison of Model Performance of Different Federated Learning Methods

[0122] method Final test accuracy Final test loss FedAvg 0.97 0.12 The method of this invention (VARO-FL) 0.98 0.08

[0123] Table 6: Performance Comparison Results of the Invention Method and Existing Federated Learning Methods

[0124] index FedAvg This invention (VARO-FL) Minimum test loss 0.0669 0.1037 Achieving 95% accuracy in rounds 16 25 Average number of clients disconnecting per round 2.82 1.23 Average number of active clients 6.31 12.75 Average training time per round (s) 10.69 25.99 Total communication volume (MB) 211.18 437.80 Average model update norm 1.262 1.233 Momentum Norm 0 0.639 (stable)

[0125] In this embodiment, due to the susceptibility of vehicle clients to disconnection in dynamic vehicular network environments, the vehicle clients participating in training are different in each round, resulting in randomness. Furthermore, the local training intensity needs to be dynamically adjusted. Therefore, the model performance does not exhibit a monotonic trend. Federated learning itself allows for "performance rebound," and this is frequently observed in the classic FedAvg method. For example, the test loss reaches its lowest value in the 15th round of training, while the loss value slightly increases in the 40th round. This demonstrates that the federated learning training process commonly exhibits a phenomenon where "the loss bottoms out in a single round, and then fluctuates within a stable convergence range." In this embodiment, for the FedAvg method, its test loss once dropped to a minimum of 0.0669 in a certain round, but eventually converged to 0.12, indicating that in dynamic vehicular network environments, the algorithm's performance significantly declines in later stages due to the dynamic environment. The VARO-FL method provided by this invention has a historical minimum test loss of 0.1037 (slightly higher than the minimum single-round loss of FedAvg), but eventually converges stably to 0.08. This result shows that this method is more stable in later training stages and can significantly improve the reliability of the final model.

[0126] It should be noted that the VARO-FL method provided in this invention achieves a higher minimum loss per round than FedAvg. The core reason lies in the momentum aggregation mechanism introduced in this invention. This mechanism essentially suppresses parameter oscillations during model training based on historical update information. Its design goal is not to pursue "optimal performance per round," but rather to ensure overall convergence stability in dynamic environments. Therefore, it does not guarantee that each round will be "better." In the dynamic vehicle-to-everything (V2X) scenario, the model's stable performance at the end of training is more valuable for practical applications and engineering significance than the instantaneous minimum loss of a particular round. This invention, by introducing a vehicle operation status-aware scheduling strategy and a momentum smoothing aggregation mechanism, effectively avoids the problem of significant performance degradation in the later stages of the model, keeping the test loss at a low and stable level, thus adapting to the actual deployment needs of dynamic V2X environments.

[0127] In particular, the method of the present invention significantly reduces invalid model updates and improves the effectiveness of model aggregation by introducing a client selection mechanism for vehicle operation status perception, even in situations where vehicles are moving at high speeds and communication is unstable. Furthermore, by introducing a momentum smoothing update mechanism, the model parameter update process is made more stable, reducing the risk of fluctuations during training.

[0128] In summary, this embodiment verifies that the method of the present invention can effectively improve the stability, robustness, and engineering applicability of the federated learning training process in the dynamic operating environment of the Internet of Vehicles.

[0129] The above-described specific examples are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A federated learning training method based on vehicle state perception suitable for dynamic vehicle networks, characterized in that, The method includes the following steps: Step 1. Before the federated learning training begins, the server initializes the global model and distributes it to each vehicle client. Step 2. In each round of federated learning, the server first collects the vehicle operating status information of the vehicle client and obtains the network connection stability index based on the vehicle speed mapping. Step 3. The server performs an initial screening of candidate vehicle clients based on vehicle remaining battery power and network connection stability indicators, eliminating vehicle clients that do not meet the operating conditions; then, it calculates a comprehensive score for the screened vehicle clients, and adopts a client selection strategy that combines selecting the top-ranked clients with random supplementation to determine the set of vehicle clients participating in this round of federated learning. Step 4. For the selected vehicle client, the server adaptively configures local training parameters based on the vehicle speed and computing power, including the number of local training rounds, batch size, maximum number of training batches, and training interruption probability. Step 5. The vehicle client receives the global model parameters sent by the server, performs local training on the local dataset according to the adaptively configured local training parameters, and uploads the updated model to the server. Step 6. After receiving the model update uploaded by the vehicle client, the server performs weighted aggregation on the model updates that are validly involved in the vehicle client, and introduces a momentum mechanism to smooth the model update process, thereby obtaining new global model parameters. Step 7. Repeat the above process until the preset training termination condition is met.

2. The federated learning training method based on vehicle state perception suitable for dynamic vehicle networks according to claim 1, characterized in that, The vehicle operating status information in step 2 includes at least: vehicle speed, vehicle computing power, vehicle remaining battery power, number of times the vehicle has participated in federated learning training in history, and network connection stability index. The network connection stability is obtained by mapping the vehicle speed, and its calculation method adopts a piecewise function.

3. The federated learning training method based on vehicle state perception suitable for dynamic vehicle networks according to claim 2, characterized in that, Step 3, vehicle client comprehensive scoring Among them, speed score Fairness score , Indicates vehicle speed, Indicators of network connection stability Indicates the vehicle's computing power, Indicates the vehicle's remaining battery power. Indicates the number of times a vehicle has participated in federal learning and training in its history. Indicates the vehicle client number. This indicates the federal learning round.

4. The federated learning training method based on vehicle state perception suitable for dynamic vehicle networks according to claim 3, characterized in that, The local training rounds of the vehicle client in step 4 for: ； Batch size of vehicle client for: ； Maximum number of training batches for vehicle clients in a single round of federated learning for: ； Training interruption probability for: ； When random variable And satisfy If the vehicle client is found to have experienced a training or communication interruption in the current round, its model update will not participate in the aggregation of this round.

5. A federated learning training method based on vehicle state perception suitable for dynamic vehicle networks according to claim 4, characterized in that, Step 6 is in In the round of federated learning, the server first determines the aggregation weight of each vehicle client based on the amount of local data on each vehicle client: ;in: This indicates the number of local data samples that vehicle client i participated in training for in the current round. This represents the sum of local data on the client side for all participating vehicles in the current round. Indicates vehicle client Weighting coefficients in this round of aggregation Indicates the first The set of vehicle clients participating in training during the round of federated learning; The server performs a weighted summation of the model updates uploaded by each vehicle client based on the aggregation weight, to obtain the aggregated update amount for this round: ;in: Indicates the first In federated learning, the global update direction is formed by the updates of all valid vehicle client models. Indicates vehicle client The amount of parameter updates relative to the global model.

6. The federated learning training method based on vehicle state perception suitable for dynamic vehicle networks according to claim 5, characterized in that, Step 6 introduces a momentum mechanism during the aggregation process. The server updates the global model parameters based on the momentum variable to obtain the global model parameters for the next round of federated learning. The update method is as follows: ； ； in: Represents the momentum variable in the previous round of federated learning. Let be the momentum coefficient, satisfying ; The first The global model parameters at the start of each round of federated learning.

Images