CN-121745225-B - Federal learning training method suitable for dynamic vehicle networking based on vehicle state sensing

CN121745225BCN 121745225 BCN121745225 BCN 121745225BCN-121745225-B

Abstract

The invention belongs to the technical field of vehicle networking and communication safety, and discloses a federal learning training method suitable for dynamic vehicle networking based on vehicle state sensing, which is characterized in that vehicle running state information of vehicle clients is collected, preliminary screening is carried out on candidate vehicle clients, comprehensive scores are calculated on the screened vehicle clients, a client selection strategy combining the selection of the front score ranking with random replenishment is adopted, and a vehicle client set of the present wheel participating in federal learning is determined; and the server side carries out weighted aggregation on the model update effectively participated in the vehicle client side, and introduces a momentum mechanism to carry out smoothing treatment on the model update process so as to obtain new global model parameters. The method effectively improves the stability, robustness and training efficiency of federal learning in a high-dynamic car networking environment.

Inventors

WANG SHAOQIANG
Min Haojia
LI SHUTONG
Qian Jiapeng
ZHENG JINTAO

Assignees

长春大学

Dates

Publication Date: 20260508
Application Date: 20260225

Claims (6)

1. The federal learning training method suitable for dynamic vehicle networking based on vehicle state sensing is characterized by comprising the following steps of: Step 1, before federal learning training begins, a server side initializes a global model and transmits the global model to each vehicle client side; Step 2, in each round of federal learning, the server side firstly collects vehicle running state information of the vehicle client side and obtains a network connection stability index based on vehicle speed mapping; Step 3, the server terminal performs preliminary screening on candidate vehicle clients based on the vehicle residual capacity and the network connection stability index, and eliminates the vehicle clients which do not meet the running conditions; then, calculating comprehensive scores of the screened vehicle clients, and determining a vehicle client set participating in federal learning in the round by adopting a client selection strategy combining the selection of the front score rank and random replenishment; Step 4, for the selected vehicle client, the server adaptively configures local training parameters according to the vehicle speed and the vehicle computing capacity, wherein the local training parameters comprise local training rounds, batch processing sizes, maximum training batch times and training interruption probability; Step 5, the vehicle client receives global model parameters issued by the server, performs local training on a local data set according to the self-adaptively configured local training parameters, and uploads the updated model parameters to the server; step 6, after receiving the model update uploaded by the vehicle client, the server side carries out weighted aggregation on the model update effectively participated in the vehicle client, and introduces a momentum mechanism to carry out smoothing treatment on the model update process so as to obtain new global model parameters; and 7, repeating the process until reaching the preset training termination condition.
2. The federal learning training method based on vehicle state sensing and applicable to dynamic Internet of vehicles according to claim 1, wherein the vehicle running state information in the step 2 at least comprises vehicle speed, vehicle computing power, vehicle residual quantity, the number of times of participation of vehicle history in federal learning training and network connection stability indexes, wherein the network connection stability is obtained by vehicle speed mapping, and the computing mode adopts a piecewise function.
3. The federal learning training method for dynamic internet of vehicles based on vehicle state awareness according to claim 2, wherein the vehicle client comprehensive score in step3 Wherein the velocity score Fairness score , Indicating the speed of the vehicle, An index of network connection stability, Representing the computing power of the vehicle, Indicating the residual electric quantity of the vehicle, Representing the number of times the history of the vehicle participates in the federal learning training, The number of the vehicle client is indicated, Indicating federal learning rounds.
4. A federal learning training method for dynamic internet of vehicles based on vehicle state sensing according to claim 3, wherein the local training round of the vehicle client in step 4 The method comprises the following steps: ; batch size of vehicle clients The method comprises the following steps: ; maximum training batch number of vehicle client in single-round federal learning The method comprises the following steps: ; Training outage probabilities The method comprises the following steps: ; When the random variable And satisfy the following And when the training or communication interruption of the vehicle client side in the current round is judged, the model update of the vehicle client side does not participate in the aggregation of the present round.
5. The federal learning training method for dynamic internet of vehicles based on vehicle state sensing according to claim 4, wherein the step 6 is at the following In the round federation learning, a server side firstly determines the aggregation weight of each vehicle client according to the local data quantity of the vehicle client: And wherein: a local data sample number representing the training of the vehicle client i in the current round, Representing the sum of the local data quantity of all the participating vehicle clients in the current round, Representing vehicle clients Weight coefficient in the present round of aggregation, Represent the first A vehicle client set participating in training in the round federal learning; the server terminal performs weighted summation on model updates uploaded by each vehicle client terminal based on the aggregation weight to obtain the aggregation update quantity of the round: And wherein: Is shown in the first In the round federal learning, a global update direction is formed by all valid vehicle client model updates together, Representing vehicle clients The amount of parameter updates relative to the global model.
6. The federal learning training method based on vehicle state sensing and suitable for dynamic Internet of vehicles according to claim 5, wherein in step 6, a momentum mechanism is introduced in the aggregation process, and the server terminal updates the global model parameters based on momentum variables to obtain global model parameters of next federal learning, wherein the momentum variables are The updating mode of (a) is as follows: ; ; Wherein: representing momentum variable in the previous federal learning, Is momentum coefficient, satisfy ; Respectively the first Global model parameters at the beginning of the round federal learning.

Description

Federal learning training method suitable for dynamic vehicle networking based on vehicle state sensing Technical Field The invention belongs to the technical field of Internet of vehicles and communication safety, and particularly relates to a federal learning training method applicable to dynamic Internet of vehicles based on vehicle state sensing. Background With the rapid development of the Internet of vehicles technology and artificial intelligence, federal learning is widely applied to the field of Internet of vehicles model training and optimization due to the advantage of realizing multi-node collaborative training under the condition that original data is not exposed. Aiming at the problems of data isomerism, privacy protection, model training efficiency and the like in the environment of the Internet of vehicles, corresponding solutions have been proposed by a plurality of people. Chinese patent CN120416804a discloses a "dynamic protection method and system for internet of vehicles data in federal learning process", in which, in the method, an internet of vehicles model is built at a server end and distributed to a plurality of clients, the clients participating in training are determined according to an evaluation score and a dynamic client selection algorithm, and learning rates are adjusted for different mode data, so as to realize balance training of multi-mode data. According to the scheme, training effect and adaptability of the multi-mode car networking model are improved to a certain extent. However, the method mainly focuses on the balance of different modal data in the learning process, the dynamic client selection process focuses on the model evaluation result, and the state difference of the vehicle in the real running process, such as the running speed of the vehicle, the stability of network connection, the running environment change and other factors, are not fully considered. Meanwhile, the method assumes that the client has relatively stable communication conditions, is difficult to adapt to the problem of frequent topology change caused by high-speed movement of vehicles in the Internet of vehicles, and still has the risks of training interruption and unstable updating under a dynamic scene. Chinese patent CN 119670840a discloses a "federal semi-supervised learning method and system for car networking", which generates a pseudo tag sample at a vehicle client and performs multi-round aggregation training by combining with a supervision model at a server, so as to improve model training efficiency and system instantaneity. By introducing a semi-supervised learning mechanism, the method reduces the dependence on a large amount of manual annotation data and enhances the adaptability of the system in a complex traffic environment. Meanwhile, the method does not model the mobility and communication instability of the vehicle, so that the continuity and stability of the federal learning training process are difficult to ensure under the high-dynamic running environment of the Internet of vehicles, and the overall convergence effect of the model is easy to be influenced. Chinese patent CN 118821910a discloses a "federally learning-oriented multi-mode internet of vehicles model balanced training method, system and device", which improves the security and privacy protection level of internet of vehicles data in the federally learning process by introducing differential privacy protection and multi-stage aggregation mechanism. The scheme mainly solves the privacy leakage risk in the uploading and aggregation process of the model parameters, and enhances the safety and reliability of the federal learning system of the Internet of vehicles. Meanwhile, the scheme assumes that the node model updating process is relatively stable, the influence of running states such as high-speed movement and frequent disconnection of vehicles on training scheduling and model aggregation quality is not considered, and efficient and stable model training is still difficult to realize under a dynamic Internet of vehicles environment. In summary, it can be seen that the existing federal learning method generally randomly or statically selects the client side participating in training in the dynamic scene of the internet of vehicles, and does not consider the running states such as vehicle speed, connection stability, electric quantity level, etc., so that communication interruption or training failure is easily caused, and the difference of the running states of the vehicles is ignored. Vehicles move at high speed in a road environment, network connection changes frequently, a client is assumed to be stable and online by traditional federal learning, continuity and stability of a training process are difficult to ensure, and the method is difficult to adapt to the high dynamic topological characteristic of the Internet of vehicles. In addition, the calculation capability and the running state of different vehicles