An internet of vehicles federated learning method, a software defined network controller and a system

Abstract

The application provides a vehicle networking federated learning method, a software defined network controller and a system. The method comprises the following steps: receiving current communication parameters returned by at least one RSU; constructing a training energy consumption model according to the current communication parameters of at least one target RSU; determining target communication parameters of the at least one target RSU based on a preset constraint condition and the training energy consumption model; configuring training parameters and the target communication parameters for the at least one target RSU, and sending an intelligent driving model to the target RSU for training. Thus, since the software defined network controller can determine the target communication parameters corresponding to the minimum resource energy consumption for the at least one target RSU, and configure appropriate training parameters, the RSU can greatly reduce the consumption during model training, thereby improving the model training efficiency. Moreover, since the model training process is implemented in the RSU, compared with direct training at the vehicle end, the stability of vehicle communication can also be ensured.

Description

Technical Field

[0001] This invention relates to the field of vehicle networking and intelligent driving technology, and in particular to a vehicle networking federated learning method, a software-defined network controller, and a system. Background Technology

[0002] Currently, intelligent connected vehicles are developing towards Level 2 / Level 3 autonomous driving. These vehicles are equipped with numerous sensors, such as millimeter-wave radar, lidar, and cameras. The data collected by these sensors can help train deep learning models related to autonomous driving.

[0003] With increasing user awareness of privacy, more and more users are unwilling to upload their data to the cloud. This phenomenon is particularly evident in the field of connected vehicles. To address these issues, federated learning has been proposed as a candidate solution.

[0004] However, existing federated learning systems consider mobile user scenarios, which are very different from vehicle-to-everything (V2X) scenarios. For example, the intelligent driving models trained in V2X are very large, so training models on the vehicle side is very energy-intensive, which in turn affects the training efficiency of intelligent driving models. Summary of the Invention

[0005] To address the problems existing in the prior art, embodiments of the present invention provide a federated learning method, a software-defined network controller, and a system for vehicle networking, in order to solve or partially solve the technical problem that the high energy consumption of model training affects the training efficiency of intelligent driving models when using federated learning to train deep learning models for autonomous driving.

[0006] A first aspect of the present invention provides a federated learning method for vehicle-to-everything (V2X) networks, applied in a software-defined network controller, the method comprising:

[0007] Receive current communication parameters returned by at least one target roadside unit (RSU);

[0008] Construct a training energy consumption model based on the current communication parameters of the at least one target RSU;

[0009] The target communication parameters of the at least one target RSU are determined based on preset constraints and the training energy consumption model; the target communication parameters are used to minimize the resource energy consumption required for the target RSU to train the intelligent driving model.

[0010] Configure training parameters and target communication parameters for the at least one target RSU, and send the intelligent driving model to the target RSU for training.

[0011] In the above scheme, the current communication parameters include: the available CPU frequency, CPU capacitance coefficient, transmit power of the target RSU, channel gain, communication bandwidth, and available time slots between the target RSU and the software-defined network controller; the training energy consumption model includes: a communication energy consumption model and a computational energy consumption model; the step of constructing a training energy consumption model based on the current communication parameters of at least one target RSU includes:

[0012] The uplink communication rate of the target RSU is determined based on the communication bandwidth, the transmit power, the channel gain between the target RSU and the software-defined network controller, and the noise power.

[0013] The uplink communication delay of the target RSU is determined based on the memory occupied by the intelligent driving model and the uplink communication rate of the target RSU.

[0014] The communication energy consumption model of the target RSU is determined based on the uplink communication delay of the target RSU and the transmit power.

[0015] The computational energy consumption model is determined based on the training parameters, the CPU capacitance coefficient, and the maximum transmit power.

[0016] In the above scheme, determining the communication energy consumption model of the target RSU based on the uplink communication delay and the transmit power includes:

[0017] According to the formula Determine the communication energy consumption model for the nth target RSU Where, p n The transmit power of the nth target RSU, The uplink communication delay is the delay for the nth target RSU.

[0018] In the above scheme, determining the computational energy consumption model based on the training parameters, the CPU capacitor coefficient, and the available CPU frequency includes:

[0019] Obtain the number of training samples and the number of local iterations from the training parameters;

[0020] According to the formula Determine the computational energy consumption model for the nth target RSU Among them, a n f is the CPU capacitance coefficient of the nth target RSU. n D is the CPU available frequency of the nth target RSU. n L represents the number of training samples for the nth target RSU, and L represents the number of local iterations.

[0021] In the above scheme, the target communication parameters of at least one target RSU are determined based on a preset constraint function and the training energy consumption model, including:

[0022] Construct a training energy consumption objective function based on the training energy consumption model;

[0023] Under the constraints, the global optimal solution algorithm is used to find the global optimal solution for the training energy consumption objective function; the global optimal solution is the target communication parameters of at least one target RSU.

[0024] In the above scheme, constructing the training energy consumption objective function based on the training energy consumption model includes:

[0025] According to the formula Construct the training energy consumption objective function; wherein,

[0026] The energy consumption model for calculating the nth target RSU; Let be the communication energy consumption model for the nth target RSU; p is the transmit power vector, containing the transmit power of at least one target RSU; τ is the target time slot vector, containing the target time slot of at least one target RSU; T c.max This is the first duration threshold.

[0027] In the above scheme, the constraints specifically include:

[0028]

[0029] in, T is the local training duration of the nth target RSU. c,max τ is the preset first duration threshold. n For the time slot allocated to the nth target RSU, T max The preset second duration threshold; p is the uplink communication duration of the nth target RSU. n Let n be the transmit power of the nth target RSU. The maximum transmit power of the nth target RSU.

[0030] In the above scheme, after sending the intelligent driving model to the target RSU for training, the method further includes:

[0031] Receive the trained intelligent driving model sent by the at least one target RSU;

[0032] Each of the trained intelligent driving models is aggregated to obtain a global model;

[0033] If it is determined that the global model meets the convergence condition, a termination training instruction is sent to the at least one target RSU.

[0034] A second aspect of the present invention provides a software-defined network controller, the software-defined network controller comprising:

[0035] A receiving unit is used to receive current communication parameters returned by at least one target RSU;

[0036] A construction unit is used to construct a training energy consumption model based on the current communication parameters of the at least one target RSU;

[0037] A determining unit is configured to determine the target communication parameters of the at least one target RSU based on preset constraints and the training energy consumption model; the target communication parameters are used to minimize the resource energy consumption required for the target RSU to train the intelligent driving model;

[0038] The configuration unit is used to configure training parameters and target communication parameters for the at least one target RSU, and send the intelligent driving model to the target RSU for training.

[0039] A third aspect of the present invention provides a connected vehicle federated learning system, the system comprising: a software-defined network controller and multiple RSUs;

[0040] The software-defined network controller is used to identify at least one target RSU among the multiple RSUs and send a parameter query message to the at least one target RSU;

[0041] The target RSU is used to query the message and return the current communication parameters based on the parameters;

[0042] The software-defined network controller is further configured to construct a training energy consumption model based on the current communication parameters of at least one target RSU; determine target communication parameters of at least one target RSU based on preset constraints and the training energy consumption model; the target communication parameters are used to minimize the resource energy consumption required by the target RSU to train the intelligent driving model; configure training parameters and target communication parameters for at least one target RSU; and send the intelligent driving model to the target RSU for training.

[0043] This invention provides a federated learning method, a software-defined network controller, and a system for vehicle-to-everything (V2X) communication. The method includes: receiving current communication parameters returned by at least one target Roadside Unit (RSU); constructing a training energy consumption model based on the current communication parameters of the at least one target RSU; determining target communication parameters for the at least one target RSU based on preset constraints and the training energy consumption model; the target communication parameters are used to minimize the resource energy consumption required by the target RSU to train an intelligent driving model; configuring training parameters and target communication parameters for the at least one target RSU; and sending the intelligent driving model to the target RSU for training. Thus, since the software-defined network controller can determine the target communication parameters corresponding to the minimum resource energy consumption for at least one target RSU, by configuring suitable training parameters, the resources consumed by the RSU during model training can be significantly reduced, thereby improving model training efficiency. Furthermore, since the model training process is implemented in the RSU, the training speed is faster than directly training the intelligent driving model on the vehicle side, and it does not affect the stability of vehicle communication. Attached Figure Description

[0044] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0045] Figure 1 A schematic diagram of the architecture of a vehicle-to-everything (V2X) federated learning system according to an embodiment of the present invention is shown.

[0046] Figure 2 A schematic diagram of a vehicle-to-everything (V2X) federated learning method according to an embodiment of the present invention is shown.

[0047] Figure 3 A schematic diagram of the overall structure of a software network definition controller according to an embodiment of the present invention is shown. Detailed Implementation

[0048] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0049] Currently, an increasing number of users are unwilling to upload data from their vehicle's sensors to the cloud because this data contains a great deal of user privacy. For example, cockpit camera data can capture images inside the cabin; if uploaded to the cloud, the user's every move inside the car could be leaked. Similarly, external camera data, if it captures pedestrians while driving, could expose their trajectory information if uploaded to the cloud. Although vehicle sensors can collect a large amount of data, and this data can be used to train driving assistance or autonomous driving models, it cannot be uploaded to the cloud to protect the privacy of users and pedestrians, thus preventing the training of vehicle driving assistance or intelligent driving models.

[0050] Based on this, the present invention provides a vehicle-to-everything (V2X) federated learning system. In this system, vehicles do not need to upload their own data; they only need to train an intelligent driving model locally using their own data and then upload it to the cloud. This method allows for the training of intelligent driving models while protecting user privacy.

[0051] However, in existing connected vehicle federated learning systems, the training of intelligent driving models is generally carried out on the vehicle side, which has the following drawbacks: First, vehicles move at high speeds, so the frequency of network handover during model training is higher (for example, the need to continuously collect sample data), which will have a certain impact on the stability of communication; Second, the amount of data in intelligent driving models is huge, while the computing resources on the vehicle side are limited, so the training efficiency of the model will also be affected.

[0052] To address the aforementioned problems, the present invention provides a vehicle-to-everything (V2X) federated learning system, such as... Figure 1 As shown, the system includes: a Software Defined Network (SDN) controller and multiple Road Side Units (RSUs). Among them,

[0053] Software-defined network controller (SDN) is used to identify at least one target RSU among multiple RSUs and send parameter query messages to at least one target RSU.

[0054] The target RSU is used to query the message based on parameters to return the current communication parameters;

[0055] The software-defined network controller is also used to construct a training energy consumption model based on the current communication parameters of at least one target RSU; determine the target communication parameters of at least one target RSU based on preset constraints and the training energy consumption model; the target communication parameters are used to minimize the resource energy consumption required for the target RSU to train the intelligent driving model; configure training parameters and target communication parameters for at least one target RSU, and send the intelligent driving model to the target RSU for training.

[0056] In this system, the SDN manages all RSUs, and data is transmitted wirelessly between the SDN and all RSUs, as well as between the RSUs and the vehicles. Furthermore, to avoid communication conflicts between the RSUs and the SDN, the communication between the SDN and each RSU in this invention employs time-division multiplexing.

[0057] To further improve the efficiency of sending SDN-related instructions (such as parameter acquisition instructions and training instructions), SDN and each RSU communicate using the OpenFlow communication protocol. SDN sends relevant instructions to the RSU via OpenFlow messages.

[0058] During training, the model in the SDN is called the global model. The global model is sent to each RSU for local training; this process is called local training or local iteration. The model trained by each RSU is called a local model. Then, the RSU uploads the trained local models to the SDN, and the SDN aggregates all the local models to obtain the global model.

[0059] Understandably, given the large amount of data in intelligent driving models, SDN and RSU can use the FTP protocol to send intelligent driving models in order to improve transmission efficiency.

[0060] Specifically, the execution logic of the vehicle-to-everything (V2X) federated learning system in this invention is as follows:

[0061] S1 and SDN send query commands to all RSUs via OpenFlow messages to check whether there is data related to the model training task in the RSUs.

[0062] S2 and RSU return query results to SDN based on the query request. The query results include two items: whether training data is included and the amount of training data. For example, the query result can be: including training data and the corresponding amount of training data; the query result can also be: not including training data, in which case the corresponding amount of training data is 0.

[0063] S3 and SDN identify the RSU containing the training data as the target RSU and continue to send parameter query messages to at least one target RSU via OpenFlow messages.

[0064] S4. At least one target RSU is used to return the current communication parameters based on the parameter query message; the current communication parameters include: the CPU available frequency of the target RSU, the CPU capacitance coefficient, the maximum transmit power, the channel gain between the target RSU and the software-defined network controller, the communication bandwidth, and the available time slots.

[0065] S5. SDN constructs a training energy consumption model based on the current communication parameters of at least one target RSU; determines the target communication parameters of at least one target RSU based on preset constraints and the training energy consumption model; the target communication parameters are used to minimize the resource energy consumption required for the target RSU to train the intelligent driving model; configures training parameters and target communication parameters for at least one target RSU, and sends the intelligent driving model to the target RSU for training.

[0066] The specific implementation method of SDN constructing a training energy consumption model based on the current communication parameters and determining the target communication parameters of at least one target RSU based on preset constraints and the training energy consumption model will be described in detail in subsequent SDN-side embodiments, so it will not be repeated here.

[0067] The vehicular network federated learning system provided by this invention reduces training energy consumption by adjusting the RSU transmit power, time slot allocation, and setting the maximum number of training samples. The distributed RSU network of this invention can avoid the impact of communication terminals on model training due to the high-speed movement of vehicles; furthermore, since the model training process takes place on the RSU rather than on the vehicle, the reliability of the communication link is improved when the model is uploaded to the SDN.

[0068] Based on the same inventive concept as the foregoing embodiments, this invention also provides a federated learning method for vehicle networks, applied in the aforementioned Software-Defined Networking (SDN) control, such as... Figure 2 As shown, the method mainly includes the following steps:

[0069] S210, receive current communication parameters returned by at least one target roadside unit (RSU).

[0070] Since there are multiple Roadside Units (RSUs), but not all of them are currently undergoing training tasks, the method further includes the following steps before receiving the current communication parameters returned by at least one target Roadside Unit (RSU):

[0071] SDN sends query commands to all RSUs via OpenFlow messages to check if any data related to the model training task exists within the RSUs. The RSUs then return the query results to SDN, which include two items: whether training data is present and the amount of training data. For example, the query result could be: "Training data present" and the corresponding amount of training data; or it could be: "Training data not present," in which case the amount of training data would be 0. The query results can be in JSON format.

[0072] The RSU containing the amount of training data is identified as the target RSU, and parameter query messages are continued to be sent to at least one target RSU via OpenFlow messages;

[0073] Receive current communication parameters returned by at least one target RSU based on a parameter query message; the current communication parameters include: the CPU available frequency of the target RSU, the CPU capacitance coefficient, the maximum transmit power, the channel gain between the target RSU and the software-defined network controller, the communication bandwidth, and the available time slots.

[0074] S211, Construct a training energy consumption model based on the current communication parameters of the at least one target RSU.

[0075] After receiving the current communication parameters sent by at least one target RSU, the SDN can construct a training energy consumption model based on the current communication parameters of at least one target RSU. The training energy consumption model includes: a communication energy consumption model and a computational energy consumption model; constructing the training energy consumption model based on the current communication parameters of at least one target RSU includes:

[0076] The uplink communication rate of the target RSU is determined based on the communication bandwidth, maximum transmit power, channel gain between the target RSU and the software-defined network controller, and noise power.

[0077] The uplink communication latency of the target RSU is determined based on the memory occupied by the intelligent driving model and the uplink communication rate of the target RSU.

[0078] The communication energy consumption model of the target RSU is determined based on the uplink communication delay and maximum transmit power of the target RSU.

[0079] The energy consumption calculation model is determined based on the training parameters, CPU capacitance coefficient, and maximum transmit power.

[0080] Specifically, let N = {1, 2, ..., n, ..., N} be the set of all target RSUs. For the nth target RSU, assume that the available CPU frequency of the nth target RSU is f. n The transmission power is p n The number of training samples is D n The channel gain between the SDN and the SDN is h. n The available real-time slot is τ n The communication bandwidth is B, and the CPU capacitance coefficient is a. n The noise power is N0, the number of CPU clock cycles required to process one sample is c, the number of local iterations is L, and the memory occupied by the intelligent driving model is m.

[0081] Therefore, the uplink communication rate of the target RSU is determined based on the communication bandwidth, maximum transmit power, channel gain between the target RSU and the software-defined network controller, and noise power, including:

[0082] According to the formula Determine the uplink communication rate r of the target RSUn .

[0083] The uplink communication latency of the target RSU is determined based on the memory occupied by the intelligent driving model and the uplink communication rate of the target RSU, including:

[0084] According to the formula Determine the uplink communication delay of the target RSU

[0085] In one implementation, the communication energy consumption model of the target RSU is determined based on the uplink communication delay and maximum transmit power of the target RSU, including:

[0086] According to the formula Determine the communication energy consumption model for the nth target RSU in,

[0087] p n Let n be the transmit power of the nth target RSU (which can be understood as the current transmit power). The uplink communication delay is for the nth target RSU.

[0088] In one implementation, the energy consumption calculation model is determined based on training parameters, CPU capacitance coefficient, and transmit power, including:

[0089] Obtain the number of training samples and the number of local iterations from the training parameters;

[0090] According to the formula Determine the computational energy consumption model for the nth target RSU in,

[0091] a n f is the CPU capacitance coefficient of the nth target RSU. n D is the CPU available frequency for the nth target RSU. n L represents the number of training samples for the nth target RSU, and L represents the number of local iterations.

[0092] This determined the training energy consumption model.

[0093] S212, based on preset constraints and the training energy consumption model, determine the target communication parameters of the at least one target RSU; the target communication parameters are used to minimize the resource energy consumption required for the target RSU to train the intelligent driving model.

[0094] Once the communication energy consumption model and computational energy consumption model are determined, the training energy consumption model is also determined. In one implementation, the target communication parameters for at least one target RSU are determined based on a preset constraint function and the training energy consumption model, including:

[0095] Construct a training energy consumption objective function based on the training energy consumption model;

[0096] Under constraints, the global optimal solution algorithm is used to find the global optimal solution for the training energy consumption objective function; the global optimal solution is the target communication parameters of at least one target RSU.

[0097] In one implementation, constructing a training energy consumption objective function based on the training energy consumption model includes:

[0098] According to the formula Construct a training energy consumption objective function; where,

[0099] The energy consumption model for calculating the nth target RSU. Let be the communication energy consumption model for the nth target RSU; p is the transmit power vector, containing the transmit power of at least one target RSU; τ is the target time slot vector, containing the target time slot of at least one target RSU; T c.max This is the first duration threshold.

[0100] The constraints are as follows:

[0101]

[0102] in, T is the local training duration of the nth target RSU. c,max τ is the preset first duration threshold. n For the time slot allocated to the nth target RSU, T max The preset second duration threshold; p is the uplink communication duration of the nth target RSU. n Let n be the transmit power of the nth target RSU. The maximum transmit power of the nth target RSU.

[0103] Specifically, the objective function represents the total training energy consumption of all target RSUs, and the first constraint represents the local training time of at least one target RSU. It should be less than the first duration threshold T c.max The second constraint includes the total training time T for all target RSUs. c,max +∑ n∈N τ n It should be less than the second duration threshold T max The third constraint states that the uplink communication delay of the target RSU should be less than the allocated time slot τ. n The fourth constraint states that the target RSU's transmit power should meet the constraint of maximum transmit power (it should be less than the maximum transmit power).

[0104] Once the objective function and constraints are determined, a global optimal solution algorithm can be used to solve the objective function, thereby obtaining the global optimal solution. The global optimal solution contains multiple sets of solutions, each containing the target communication parameters of at least one objective RSU. The global optimal solution algorithm can be any global operations research algorithm such as CPLEX, and is not limited here.

[0105] The target communication parameters determined in this way are equivalent to the optimal communication parameters of the target RSU. Therefore, when the target RSU trains the model based on the target communication parameters, it has the minimum communication energy consumption and computational energy consumption. Thus, the target communication parameters can minimize the resource energy consumption required for the target RSU to train the intelligent driving model.

[0106] S213, configure training parameters and target communication parameters for the at least one target RSU, and send the intelligent driving model to the target RSU for training.

[0107] After determining the target communication parameters, SDN will configure training parameters and target communication parameters for at least one target RSU and send the intelligent driving model to the target RSU for training.

[0108] In one implementation, after sending the intelligent driving model to the target RSU for training, the method further includes:

[0109] Receive a trained intelligent driving model sent by at least one target RSU;

[0110] Each trained intelligent driving model is aggregated to obtain a global model;

[0111] If the global model is determined to meet the convergence condition, a termination training instruction is sent to at least one target RSU.

[0112] This involves aggregating each trained intelligent driving model to obtain a global model, which specifically includes:

[0113] Obtain the learning parameters from each trained intelligent driving model;

[0114] The average value of each learning parameter is determined, and the average value of each learning parameter is used as the optimal learning parameter for the global model, thus obtaining the global model.

[0115] The learning parameters may include weights, learning rate, regularization parameters, etc.

[0116] Specifically, SDN can send training code to the RSU via OpenFlow messages, while configuring parameters such as CPU frequency, transmit power, maximum number of training samples, time slots, number of local iterations, sample loss threshold, and sample allocation ratio for the RSU to complete training. The autonomous driving model file to be trained is then transferred to the target RSU using the FTP protocol.

[0117] At least one target RSU uses the collected vehicle sensor data to train the intelligent driving model, and then returns the trained intelligent driving model to the SDN via the FTP protocol.

[0118] In order to improve the freshness of the training set, when the amount of data collected exceeds the maximum data capacity, the target RSU will reconstruct the training sample set according to the sample loss threshold and quantity allocation ratio configured by SDN and the loss value of all samples.

[0119] For example, suppose the sample loss threshold is set to 0.1, and the number allocation ratios are 80% and 20% respectively. The maximum number of samples that the target RSU can use is 180. If a car passes by and the target RSU collects 20 more data samples, then the total number of samples is 200. Next, 180 samples need to be selected from these 200 samples to reconstruct the dataset.

[0120] First, the target RSU process inputs these 200 samples into the global model used in the previous iteration, thereby obtaining the loss values ​​for each of the 200 samples. Then, the samples are sorted in ascending order of loss value. Since the sample loss threshold is 0.1, these 200 samples can be divided into two sets based on this threshold: a set of samples with loss values ​​less than 0.1 (referred to as the first set) and a set of samples with loss values ​​greater than 0.1 (referred to as the second set).

[0121] Assume the first set has 160 samples and the second set has 40 samples. Since the data distribution ratio is 80% and 20%, and the maximum sample size is 180, the sample sizes for the two sets are 144 and 36 respectively. Finally, 144 samples are randomly sampled from the first set and 36 samples are randomly sampled from the second set.

[0122] If the total number of samples in a set is less than the number of samples to be collected, then the entire set will be sampled, and the missing samples will be supplemented by another set. This sampling method can improve the freshness of the data and ensure the convergence speed of training.

[0123] In this way, SDN and the target RSU continuously perform global iterations until the global model meets the convergence requirements. When the global model meets the prediction accuracy of the test set, SDN sends a stop training instruction to the target RSU via OpenFlow messages. At this point, the entire training process ends.

[0124] Since the software-defined network controller can determine the target communication parameters corresponding to the minimum resource energy consumption for at least one target RSU, by configuring appropriate training parameters, the RSU can significantly reduce the resources consumed by model training, thus improving model training efficiency; and since the model training process is implemented in the RSU, it will not affect the stability of vehicle communication compared to training autonomous driving directly on the vehicle.

[0125] The vehicular network federated learning system provided by this invention reduces training energy consumption by adjusting the RSU transmit power, time slot allocation, and setting a maximum number of training samples. The distributed RSU network avoids the impact of high-speed vehicle movement on model training caused by communication terminals; furthermore, since the model training process takes place on the RSU rather than on the vehicle, the reliability of the communication link is improved when the model is uploaded to the SDN. Although setting a maximum number of training samples may have some impact on training performance (because the number of samples used for training is limited), this invention, by setting a sample update strategy, allows the RSU to continuously update samples, ensuring the freshness of the training set and mitigating the impact on training results and training speed.

[0126] Based on the same inventive concept as in the foregoing embodiments, this embodiment also provides a software-defined network controller, such as... Figure 3 As shown, the software-defined network controller includes:

[0127] The receiving unit 31 is used to receive the current communication parameters returned by at least one target RSU;

[0128] Construction unit 32 is used to construct a training energy consumption model based on the current communication parameters of the at least one target RSU;

[0129] The determining unit 33 is used to determine the target communication parameters of the at least one target RSU based on preset constraints and the training energy consumption model; the target communication parameters are used to minimize the resource energy consumption required for the target RSU to train the intelligent driving model;

[0130] Configuration unit 34 is used to configure training parameters and target communication parameters for the at least one target RSU, and send the intelligent driving model to the target RSU for training.

[0131] Since the device described in the embodiments of this invention is a controller used to implement the vehicle-to-everything (V2X) federated learning method of this invention, those skilled in the art can understand the specific structure and variations of the device based on the method described in the embodiments of this invention, and therefore will not be described in detail here. All devices used in the methods of the embodiments of this invention fall within the scope of protection of this invention.

[0132] Through one or more embodiments of the present invention, the present invention has the following beneficial effects or advantages:

[0133] This invention provides a federated learning method, a software-defined network controller, and a system for vehicle-to-everything (V2X) communication. The method includes: receiving current communication parameters returned by at least one target Roadside Unit (RSU); constructing a training energy consumption model based on the current communication parameters of the at least one target RSU; determining target communication parameters for the at least one target RSU based on preset constraints and the training energy consumption model; the target communication parameters are used to minimize the resource energy consumption required by the target RSU to train an intelligent driving model; configuring training parameters and target communication parameters for the at least one target RSU; and sending the intelligent driving model to the target RSU for training. Thus, since the software-defined network controller can determine the target communication parameters corresponding to the minimum resource energy consumption for at least one target RSU, by configuring suitable training parameters, the resources consumed by the RSU during model training can be significantly reduced, thereby improving model training efficiency. Furthermore, since the model training process is implemented in the RSU, it does not affect the stability of vehicle communication compared to directly training intelligent driving on the vehicle side.

[0134] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0135] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A federated learning method for vehicle-to-everything (V2X) networks, characterized in that, When applied in a software-defined network controller, the method includes: Receive current communication parameters returned by at least one target roadside unit (RSU); Construct a training energy consumption model based on the current communication parameters of the at least one target RSU; The target communication parameters of the at least one target RSU are determined based on preset constraints and the training energy consumption model; the target communication parameters are used to minimize the resource energy consumption required for the target RSU to train the intelligent driving model. Configure training parameters and target communication parameters for the at least one target RSU, and send the intelligent driving model to the target RSU for training; wherein... Based on a preset constraint function and the training energy consumption model, at least one target communication parameter for a target RSU is determined, including: Construct a training energy consumption objective function based on the training energy consumption model; Under the constraints, the global optimal solution algorithm is used to find the global optimal solution for the training energy consumption objective function; the global optimal solution is the target communication parameters of at least one target RSU.

2. The method as described in claim 1, characterized in that, The current communication parameters include: the available CPU frequency, CPU capacitance coefficient, transmit power of the target RSU, channel gain, communication bandwidth, and available time slots between the target RSU and the software-defined network controller; the training energy consumption model includes: a communication energy consumption model and a computational energy consumption model; constructing the training energy consumption model based on the current communication parameters of at least one target RSU includes: The uplink communication rate of the target RSU is determined based on the communication bandwidth, the transmit power, the channel gain between the target RSU and the software-defined network controller, and the noise power. The uplink communication delay of the target RSU is determined based on the memory occupied by the intelligent driving model and the uplink communication rate of the target RSU. The communication energy consumption model of the target RSU is determined based on the uplink communication delay of the target RSU and the transmit power. The computational energy consumption model is determined based on the training parameters, the CPU capacitance coefficient, and the maximum transmit power.

3. The method as described in claim 2, characterized in that, The step of determining the communication energy consumption model of the target RSU based on the uplink communication delay and the transmit power includes: According to the formula Determine the first n Communication energy consumption model for a target RSU ;in, For the first n The transmit power of each target RSU For the first n Uplink communication latency of each target RSU.

4. The method as described in claim 2, characterized in that, Determining the computational energy consumption model based on the training parameters, the CPU capacitance coefficient, and the available CPU frequency includes: Obtain the number of training samples and the number of local iterations from the training parameters; According to the formula Determine the first n Computational energy consumption model for each target RSU ;in, For the first n CPU capacitor coefficient of each target RSU For the first n CPU available frequency for each target RSU For the first n Number of training samples for each target RSU L Let be the number of local iterations.

5. The method as described in claim 1, characterized in that, The step of constructing the training energy consumption objective function based on the training energy consumption model includes: According to the formula Construct the training energy consumption objective function; wherein, For the first n Computational energy consumption model for one target RSU; For the first n Communication energy consumption model for one target RSU; p is the transmit power vector, which contains the transmit power of at least one target RSU; The target time slot vector contains the target time slot of at least one target RSU; This is the first duration threshold.

6. The method as described in claim 1, characterized in that, The constraints specifically include: ； in, For the first n Local training time for each target RSU, The preset first duration threshold, For the first n The time slots allocated to each target RSU The preset second duration threshold; For the first n Uplink communication duration for each target RSU For the first n The transmit power of each target RSU For the first n The maximum transmit power of each target RSU.

7. The method as described in claim 1, characterized in that, After sending the intelligent driving model to the target RSU for training, the method further includes: Receive the trained intelligent driving model sent by the at least one target RSU; Each of the trained intelligent driving models is aggregated to obtain a global model; If it is determined that the global model meets the convergence condition, a termination training instruction is sent to the at least one target RSU.

8. A software-defined network controller, characterized in that, The software-defined network controller includes: A receiving unit is used to receive current communication parameters returned by at least one target RSU; A construction unit is used to construct a training energy consumption model based on the current communication parameters of the at least one target RSU; A determining unit is configured to determine the target communication parameters of the at least one target RSU based on preset constraints and the training energy consumption model; the target communication parameters are used to minimize the resource energy consumption required for the target RSU to train the intelligent driving model; A configuration unit is used to configure training parameters and target communication parameters for the at least one target RSU, and to send the intelligent driving model to the target RSU for training. Based on a preset constraint function and the training energy consumption model, at least one target communication parameter for a target RSU is determined, including: Construct a training energy consumption objective function based on the training energy consumption model; Under the constraints, the global optimal solution algorithm is used to find the global optimal solution for the training energy consumption objective function; the global optimal solution is the target communication parameters of at least one target RSU.

9. A vehicle-to-everything (V2X) federated learning system, characterized in that, The system includes: a software-defined network controller and multiple RSUs; The software-defined network controller is used to identify at least one target RSU among the multiple RSUs and send a parameter query message to the at least one target RSU; The target RSU is used to query the message and return the current communication parameters based on the parameters; The software-defined network controller is further configured to: construct a training energy consumption model based on the current communication parameters of at least one target RSU; determine target communication parameters for at least one target RSU based on preset constraints and the training energy consumption model; the target communication parameters are used to minimize the resource energy consumption required by the target RSU to train the intelligent driving model; configure training parameters and target communication parameters for at least one target RSU; and send the intelligent driving model to the target RSU for training. Based on a preset constraint function and the training energy consumption model, at least one target communication parameter for a target RSU is determined, including: Construct a training energy consumption objective function based on the training energy consumption model; Under the constraints, the global optimal solution algorithm is used to find the global optimal solution for the training energy consumption objective function; the global optimal solution is the target communication parameters of at least one target RSU.

Images