A Method and System for Training Vehicle-Road Cooperative Distributed Models Based on Federated Learning

By employing federated learning and multi-dimensional weighted aggregation mechanisms in the vehicle-road cooperative system, the problems of data privacy leakage and communication bandwidth consumption in the training of the vehicle-road cooperative model are solved, ensuring the stability and accuracy of the model in complex environments and complying with legal and regulatory requirements.

CN122674809BActive Publication Date: 2026-11-17ZHEJIANG ELECTROMECHANICAL VOCATIONAL & TECH COLLEGE
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202611169868.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-04
Publication Date
2026-11-17
Estimated Expiration
2046-08-04

AI Technical Summary

Technical Problem

Existing vehicle-road cooperative model training suffers from high risks of data privacy leakage, high communication bandwidth consumption, and decreased model accuracy due to heterogeneous nodes.

Method used

A distributed model training method for vehicle-road cooperative systems based on federated learning is adopted. Local traffic perception data is collected by roadside units and vehicle-mounted units for local training. Multi-dimensional adaptive weighted aggregation is performed using node stability, data quality and communication quality. The weighted aggregation operation is performed by the edge computing server, and differential privacy encryption and homomorphic encryption are used to protect data privacy.

Benefits of technology

This achieves data transmission without leaving the local area, reduces communication pressure, protects data privacy, improves the stability and accuracy of the model in complex environments, and complies with legal and regulatory requirements.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122674809B_ABST
    Figure CN122674809B_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for training a distributed vehicle-road cooperative model based on federated learning, relating to the fields of intelligent transportation and artificial intelligence. This invention aims to address the high risk of data privacy leakage and high communication bandwidth consumption inherent in existing centralized training models. This invention uses roadside units and vehicle-mounted units to locally collect traffic perception data and complete model training. It calculates aggregation weight factors based on multiple dimensions, including node stability, data quality, and communication quality, and then uses an edge computing server to perform weighted aggregation and update the global model. This invention effectively suppresses the contamination of the global model by low-quality nodes while protecting data privacy and reducing communication overhead, thereby improving model accuracy and system robustness.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and artificial intelligence technology, specifically to a training method and system for a vehicle-road cooperative distributed model based on federated learning, electronic equipment, and computer-readable storage medium. Background Technology

[0002] With the rapid development of vehicle-road cooperative technology, autonomous driving and intelligent traffic management have placed extremely high demands on the accuracy and real-time performance of algorithm models. Current model training typically adopts a "centralized" model, which requires all vehicles and roadside equipment to upload the raw video and radar point cloud data they collect to a cloud server for centralized training.

[0003] However, existing technologies have the following significant drawbacks: the raw data contains sensitive information such as vehicle trajectories, faces, and license plates, and directly uploading it to the cloud poses a risk of violating laws and regulations such as the Data Security Law; the massive amount of video and radar data transmission consumes huge amounts of network bandwidth, leading to increased communication latency and affecting the normal operation of real-time services; existing general federated learning algorithms usually assume that the nodes participating in training are homogeneous, that is, the data quality and computing power of each node are similar. If the model parameters of all nodes are simply averaged and aggregated, the noisy gradients uploaded by low-quality nodes (such as vehicles with occluded sensors) will pollute the global model, resulting in a decrease in model accuracy. Summary of the Invention

[0004] The purpose of this invention is to solve the technical problems of high data privacy leakage risk, large communication bandwidth consumption, and model accuracy reduction caused by heterogeneous nodes in the training of existing vehicle-road cooperative models.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for training a distributed vehicle-road cooperative model based on federated learning, applied to a vehicle-road cooperative system comprising roadside units, on-board units, and edge computing servers, the method comprising:

[0007] The roadside unit and the vehicle-mounted unit respectively collect local traffic perception data and use the local traffic perception data to train a preset local model to obtain local model parameters.

[0008] Determine the aggregation weight factor of the roadside unit and the vehicle-mounted unit in this training;

[0009] The aggregation weight factor is calculated based on the node stability dimension, data quality dimension, and communication quality dimension. The node stability dimension is determined based on whether the node is a fixed deployment. The data quality dimension includes at least the image clarity evaluation index of the collected data.

[0010] The roadside unit and the vehicle-mounted unit send the local model parameters and the corresponding aggregate weight factors to the edge computing server;

[0011] The edge computing server performs a weighted aggregation operation on each of the received local model parameters according to each of the aggregation weight factors, and updates the global model parameters.

[0012] Preferably, determining the aggregation weight factor for the roadside unit and the vehicle-mounted unit in this training includes:

[0013] For the node stability dimension, a first preset stability weight is assigned to the roadside unit and a second preset stability weight is assigned to the vehicle unit, wherein the first preset stability weight is higher than the second preset stability weight;

[0014] Alternatively, the second preset stability weight can be dynamically adjusted based on the moving speed of the vehicle-mounted unit; the higher the moving speed, the lower the second preset stability weight.

[0015] Preferably, the data quality dimension also includes data distribution entropy;

[0016] The calculation method for the image sharpness evaluation index includes:

[0017] Detect image contrast or edge gradient information in the local traffic perception data;

[0018] When the image contrast or edge gradient information is lower than a preset sharpness threshold, it is determined that there is visual interference in the current environment, and the value of the data quality dimension in the aggregation weight factor is reduced.

[0019] Preferably, the communication quality dimension is determined based on the vehicle wireless communication channel status between the roadside unit or the vehicle-mounted unit and the edge computing server, and the vehicle wireless communication channel status includes reference signal received power or signal-to-interference-plus-noise ratio;

[0020] When the state of the vehicle wireless communication channel is below the communication threshold, the aggregation weight factor is reduced to prevent communication delay from affecting the convergence speed of the global model.

[0021] Preferably, the local model includes at least one of the following neural network structures:

[0022] The convolutional neural network-long short-term memory network used for traffic flow prediction takes traffic video frames or radar point cloud sequences at consecutive time steps as input and outputs traffic flow predictions for future time steps.

[0023] A graph convolutional neural network for abnormal behavior detection takes a vehicle trajectory topology graph as input and outputs an abnormal probability value for vehicles driving in the wrong direction or pedestrians entering the area.

[0024] Preferably, the roadside unit and the vehicle-mounted unit send the local model parameters and the corresponding aggregated weight factors to the edge computing server, including:

[0025] Differential privacy encryption or homomorphic encryption is applied to the gradient information in the local model parameters to generate encrypted gradients;

[0026] The encryption gradient and the aggregate weight factor in plaintext form are sent to the edge computing server.

[0027] Preferably, the method further includes:

[0028] The edge computing server distributes the updated global model parameters to the roadside unit and the vehicle unit for use in the next round of local training.

[0029] The edge computing server periodically synchronizes the global model parameters to the cloud control platform, which then performs cross-regional model scheduling or version management.

[0030] A vehicle-road cooperative system based on federated learning includes:

[0031] The local training module is used to collect local traffic perception data and use the local traffic perception data to train a preset local model to obtain local model parameters.

[0032] The weight calculation module is used to determine the aggregate weight factor in this training; wherein, the aggregate weight factor is calculated based on the node stability dimension, data quality dimension and communication quality dimension, the node stability dimension is determined based on whether the node is a fixed deployment, and the data quality dimension includes at least the image sharpness evaluation index of the collected data.

[0033] The data transmission module is used to send the local model parameters and the corresponding aggregation weight factors to the edge computing server so that the edge computing server can perform a weighted aggregation operation.

[0034] An electronic device, comprising:

[0035] processor;

[0036] Memory, used to store computer programs;

[0037] Communication interface, used to enable vehicle wireless communication;

[0038] The processor implements the steps of the method when executing the computer program.

[0039] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method as described.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] This invention achieves data transmission without leaving the local machine through a federated learning architecture, transmitting only model parameters, thus avoiding the leakage of original sensitive data and complying with the requirements of laws and regulations such as the Data Security Law.

[0042] Compared to transmitting raw video and radar data, this invention reduces the amount of data transmitted by transmitting only model parameters, effectively alleviating the communication pressure in vehicle-road cooperative scenarios.

[0043] This invention introduces a multi-dimensional adaptive weighted aggregation mechanism based on node stability, data quality, and communication quality, which effectively suppresses the pollution of the global model by low-quality nodes and solves the model convergence problem in heterogeneous node scenarios.

[0044] This invention, by dynamically adjusting aggregation weights, can adapt to complex and ever-changing traffic environments and communication conditions, ensuring stable operation of the model in adverse weather conditions such as fog and heavy rain, as well as in high-speed moving scenarios. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0046] Figure 2 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] Example 1

[0049] Reference Figure 1 This embodiment provides a method for training a distributed vehicle-road cooperative model based on federated learning, which is applied to a vehicle-road cooperative system that includes roadside units (RSUs), on-board units (OBUs), edge computing servers, and a cloud control platform.

[0050] Roadside units are deployed at traffic intersections or on gantries, equipped with high-precision cameras, millimeter-wave radar, and computing units. They are characterized by fixed locations, ample power supply, and wide viewing angles. The computing units of roadside units typically utilize edge computing boxes, configured with an 8-core ARM processor, 16GB of RAM, and a 512GB solid-state drive, capable of supporting real-time video stream processing and deep learning model inference. Vehicle-mounted units are installed on moving vehicles, equipped with onboard sensors and vehicle-to-everything (V2X) communication modules. They are characterized by high-speed mobility, limited computing resources, and limited viewing angles. Vehicle-mounted units typically have a 4-core ARM processor, 8GB of RAM, and a 256GB solid-state drive, primarily responsible for local data acquisition and lightweight model training. Edge computing servers are deployed on the roadside or at base stations, responsible for aggregating model parameters within the area. They are typically configured with a 32-core x86 processor, 128GB of RAM, and a 2TB solid-state drive, capable of simultaneously processing model parameter aggregation requests from hundreds of nodes. The cloud control platform is deployed in the cloud and is responsible for cross-regional model scheduling and version management. It adopts a distributed cloud computing architecture and can support the operation of vehicle-road cooperative systems nationwide.

[0051] The core of this embodiment lies in proposing a hierarchical federated learning architecture of "end-edge-cloud" and introducing an adaptive weighted aggregation mechanism. This architecture distributes model training tasks across three layers: end-side devices (roadside units and vehicle-mounted units), edge computing servers, and cloud control platforms. This fully leverages the computing resource advantages of each layer while simultaneously addressing the issues of data privacy protection and limited communication bandwidth. The specific method includes the following steps:

[0052] Local model initialization and data acquisition

[0053] An initial global model is pre-distributed to each roadside unit and vehicle-mounted unit via a cloud server or edge computing server. This initial global model can be a base model pre-trained on a public traffic dataset, such as the KITTI or BDD100K datasets. The pre-training process uses large-scale labeled data for supervised learning, enabling the model to possess basic traffic object detection and behavior recognition capabilities. The parameter size of the initial global model is adjusted according to the application scenario; for traffic flow prediction tasks, the model parameter size is approximately 5 million to 10 million; for abnormal behavior detection tasks, the model parameter size is approximately 10 million to 20 million.

[0054] The roadside unit and the vehicle-mounted unit collect local traffic perception data respectively.

[0055] Specifically, the roadside unit collects fixed-point traffic video streams and radar trajectory data at intersections. The video stream resolution is 1920×1080, the frame rate is 30fps, and the millimeter-wave radar has a detection range of 200 meters and an angular resolution of 1°. The vehicle-mounted unit collects forward-looking video streams and surrounding vehicle trajectory data during vehicle movement. The video stream resolution is 1280×720, the frame rate is 25fps, and the vehicle-mounted radar has a detection range of 150 meters.

[0056] To meet the model training requirements, local traffic perception data is preprocessed before being used as training samples. Preprocessing operations include normalization, denoising, and frame alignment. Normalization maps image pixel values ​​from the 0-255 range to the 0-1 range; the specific formula is as follows:

[0057]

[0058] in, These are the original image pixel values. These are the normalized image pixel values.

[0059] The denoising operation employs a Gaussian filtering algorithm to remove noise points from the image, with a filter kernel size of 3×3 and a standard deviation of 1.0. Frame alignment synchronizes video frames with radar point cloud data in time, ensuring that the image and radar data correspond at the same moment. Synchronization accuracy is required to be within 10 milliseconds to guarantee the spatiotemporal consistency of the data.

[0060] Local training

[0061] Roadside units and vehicle-mounted units train pre-set local models using local traffic perception data to obtain local model parameters. During this process, each node does not upload raw data; it only updates the model weights locally using the stochastic gradient descent (SGD) algorithm. This "data not leaving the local machine" training mode fundamentally solves the problem of data privacy leakage and complies with the requirements of laws and regulations such as the Data Security Law and the Personal Information Protection Law.

[0062] The hyperparameter settings for local training are as follows: BatchSize is set to 32 to 64. For roadside units with sufficient computing resources, BatchSize can be set to 64; for vehicle-mounted units with limited computing resources, BatchSize can be set to 32. The learning rate is set to 0.001 to 0.01, with an initial learning rate of 0.01. After every 10 local iterations, the learning rate decays to 0.9 times its original value. The number of local iterations is 5 to 10. The number of local iterations for roadside units can be set to 10, and the number of local iterations for vehicle-mounted units can be set to 5.

[0063] The loss function for local training is selected based on different application scenarios. For traffic flow prediction tasks, the mean squared error (MSE) loss function is used, with the following formula:

[0064]

[0065] in, For the sample size, This represents the actual traffic flow value. This represents the traffic flow value predicted by the model.

[0066] For the abnormal behavior detection task, the binary cross-entropy (BCE) loss function is used, and the specific formula is as follows:

[0067] in, The sample labels are 1 for abnormal and 0 for normal. This represents the anomaly probability value predicted by the model.

[0068] During local training, each node independently calculates gradients and updates model parameters without needing to communicate with other nodes, greatly reducing communication overhead.

[0069] Calculation of aggregate weighting factor

[0070] In traditional federated learning, a weighted average is typically used based on the number of samples, meaning that nodes with more samples have greater weights in their model parameters. However, in vehicle-to-everything (V2X) scenarios, the heterogeneity of nodes is significant, with roadside units and vehicle-mounted units exhibiting substantial differences in data quality, computing power, and communication conditions. Simply using a weighted average based on sample size can lead to noisy gradients from low-quality nodes severely contaminating the global model and causing a decrease in model accuracy.

[0071] To address the heterogeneity issue in V2X scenarios, this embodiment designs a multi-dimensional weighting mechanism. The aggregated weighting factors for roadside units and vehicle-mounted units in this training process are determined. These aggregated weighting factors are calculated based on node stability, data quality, and communication quality dimensions.

[0072] The specific calculation formula can be expressed as follows:

[0073]

[0074] in, Let i be the aggregation weight factor of the i-th node. , , These are normalized scores for node stability, data quality, and communication quality, respectively. , , These are the preset coefficients for each dimension. The values ​​of each dimension's coefficients are adjusted according to the actual application scenario. For example, in an urban road scenario, data quality has a significant impact on model accuracy, so the values ​​can be adjusted accordingly. Set it to 0.4. Set it to 0.4. Set to 0.2; In highway scenarios, communication quality has a significant impact on system real-time performance, so it can be... Set it to 0.3. Set it to 0.3. Set it to 0.4.

[0075] All normalized scores range from 0 to 1, ensuring that the aggregate weight factor also ranges from 0 to 1. After calculating the aggregate weight factor for all nodes, the weights need to be normalized so that the sum of the weights for all nodes is 1. The specific formula is as follows:

[0076] in, This represents the total number of nodes participating in this training. This is the normalized aggregate weight factor.

[0077] Node stability dimension

[0078] The node stability dimension is determined based on whether the node is a fixed deployment. Since roadside units are located in fixed positions, the data they collect has spatiotemporal continuity and can provide stable traffic state information. Therefore, a first preset stability weight is assigned to roadside units, for example... .

[0079] Because the vehicle-mounted unit is in motion, its data acquisition environment changes rapidly, its viewing angle is limited, and it may suffer from signal obstruction. Therefore, a second preset stability weight is assigned to the vehicle-mounted unit. Preferably, the second preset stability weight is dynamically adjusted according to the vehicle-mounted unit's moving speed; the higher the moving speed, the lower the second preset stability weight. This is because data collected from high-speed moving vehicles has stronger instantaneousness, poorer spatiotemporal continuity, and a relatively smaller contribution to the global model.

[0080] The dynamic adjustment formula is:

[0081]

[0082] in, This represents the vehicle's speed, measured in meters per second (m / s). For example, when the vehicle is stationary (v=0), When the vehicle is traveling at a speed of 10 m / s (approximately 36 km / h), When the vehicle is traveling at a speed of 30 m / s (approximately 108 km / h), .

[0083] Data quality dimensions

[0084] Data quality dimensions include at least image sharpness evaluation metrics and data distribution entropy. Data quality directly affects the effectiveness of model training; low-quality data can cause the model to learn incorrect features, thereby reducing the model's prediction accuracy.

[0085] Image sharpness evaluation metrics: Combining the hardware characteristics of the camera, the system detects image contrast or edge gradient information in local traffic perception data in real time.

[0086] This embodiment uses the Laplacian operator to calculate image variance to evaluate image sharpness. The specific steps are as follows:

[0087] Convert a color image to a grayscale image;

[0088] The grayscale image is convolved using a 3×3 Laplacian kernel.

[0089] Calculate the variance of the convolution result; the larger the variance, the clearer the image.

[0090] Data distribution entropy: Calculates the category distribution entropy of local data. A higher entropy value indicates richer data samples, better diversity, and higher weights. The formula for calculating data distribution entropy is:

[0091] in, The total number of data categories, denoted as the proportion of the k-th data type in the local dataset.

[0092] Normalize the data distribution entropy to between 0 and 1 to obtain the data distribution entropy score:

[0093]

[0094] Final data quality score It is obtained by weighting the image sharpness score and the data distribution entropy score:

[0095]

[0096] in, This is the weighting coefficient for image sharpness, which can be set to 0.7. Rate the image sharpness.

[0097] Communication quality dimensions

[0098] The communication quality dimension is determined based on the V2X channel status between the roadside unit or vehicle-mounted unit and the edge computing server. V2X communication adopts C-V2X (Cellular Vehicle-to-Everything) technology, operates in the 5.9GHz frequency band, and supports both direct communication and cellular communication modes.

[0099] V2X channel status includes Reference Signal Received Power (RSRP) or Signal-to-Interference-plus-Noise Ratio (SINR). RSRP typically ranges from -140 dBm to -44 dBm, and SINR typically ranges from -10 dB to 30 dB. When the V2X channel status is below the communication threshold, the aggregation weighting factor is reduced.

[0100] This mechanism can prevent nodes with poor communication quality from experiencing packet loss or excessive latency when transmitting parameters, thereby avoiding slowing down the convergence speed of the entire federated learning system.

[0101] Model Structure and Application Scenarios

[0102] The local model includes at least one of the following neural network architectures: a convolutional neural network-long short-term memory network (CNN-LSTM) for traffic flow prediction, and a graph convolutional neural network (GCN) for abnormal behavior detection.

[0103] Traffic flow prediction scenarios

[0104] The local model employs a convolutional neural network-long short-term memory network (CNN-LSTM) for traffic flow prediction. This model combines the spatial feature extraction capabilities of CNNs with the time series modeling capabilities of LSTMs, enabling it to simultaneously capture the spatial distribution characteristics and temporal evolution patterns of traffic flow.

[0105] Input: Traffic video frames (for spatial feature extraction) or radar point cloud sequences at consecutive time steps. The input sequence is 12 time steps long, with each time step spaced 5 minutes apart, representing traffic data from the past hour.

[0106] Processing: The CNN layer extracts spatial features from each frame, while the LSTM layer processes temporal features to capture the evolution of traffic flow. The CNN part consists of 3 convolutional layers and 2 pooling layers. The convolutional kernels are 3×3 with a stride of 1, and the pooling kernels are 2×2 with a stride of 2. The LSTM part consists of 2 LSTM layers, each containing 128 hidden units. Finally, the prediction result is output through a fully connected layer.

[0107] Output: Traffic flow forecasts for future time steps (e.g., congestion index, vehicle volume). The output sequence is 6 time steps long, predicting traffic conditions for the next 30 minutes.

[0108] Intrinsic Relationships: Traffic flow exhibits significant spatiotemporal dependence. The traffic state at a given moment and location is related not only to the past traffic state of that location but also to the traffic state of the surrounding area. The CNN-LSTM structure can simultaneously extract spatial distribution features and temporal evolution features, thereby accurately predicting future traffic conditions.

[0109] Abnormal behavior detection scenarios

[0110] The local model employs a graph neural network (GCN) for anomaly detection. Graph neural networks are capable of effectively processing graph-structured data, capturing the topological relationships and interaction patterns between nodes.

[0111] Input: A vehicle trajectory topology graph, where nodes represent vehicles and edges represent the relative distances or interactions between vehicles. Each node's features include the vehicle's position, speed, acceleration, and heading angle.

[0112] Processing: The GCN layer aggregates features from neighboring nodes to learn interaction patterns between vehicles. The GCN part consists of two graph convolutional layers, each containing 64 hidden units. Finally, an anomaly probability value is output through a fully connected layer and a sigmoid activation function.

[0113] Output: The probability value of an abnormal event such as a vehicle driving in the wrong direction, a pedestrian running into the vehicle, or abnormal parking. When the probability value is greater than 0.5, it is considered an abnormal behavior.

[0114] Intrinsic Relationships: Abnormal behavior is usually manifested as an abnormal topological relationship with surrounding vehicles (such as the opposite relative speed vector between a vehicle traveling in the wrong direction and the normal traffic flow). GCN can effectively capture such changes in topology, thereby accurately detecting abnormal behavior.

[0115] Edge aggregation and global update

[0116] The roadside units and vehicle-mounted units send their local model parameters and corresponding aggregated weight factors to the edge computing server. To protect privacy, before sending, each node performs differential privacy encryption (adding Laplacian noise) or homomorphic encryption on the gradient information in its local model parameters to generate encrypted gradients.

[0117] Differential privacy encryption protects data privacy while ensuring model usability by adding random noise to the gradient. The scale of the Laplace noise is determined according to the privacy budget. Adjustments will be made, and the specific formula is as follows:

[0118]

[0119] in, For the original gradient, Sensitivity to gradients, For privacy budget. Privacy budget The smaller the value, the greater the added noise and the higher the level of privacy protection, but the model accuracy will decrease. In practical applications, it can be... Set it to between 1.0 and 10.0.

[0120] Homomorphic encryption allows computations to be performed on encrypted data, and the decrypted result is identical to the result performed on the plaintext data. This embodiment employs the Paillier homomorphic encryption algorithm, which supports additive homomorphic operations and can meet the requirements of weighted aggregation in federated learning.

[0121] The encrypted gradient and the aggregated weight factors in plaintext are sent to the edge computing server. The aggregated weight factors are transmitted in plaintext because they do not contain any sensitive information and the data volume is very small.

[0122] The edge computing server performs a weighted aggregation operation on the received local model parameters based on various aggregation weight factors, and updates the global model parameters. The aggregation formula is:

[0123]

[0124] in, For the updated global model parameters, These are the local model parameters for the i-th node. This is the normalized aggregate weight factor.

[0125] Subsequently, the edge computing server distributes the updated global model parameters to roadside units and vehicle-mounted units for the next round of local training. In addition, the edge server periodically synchronizes the model to the cloud control platform for cross-regional scheduling. The synchronization cycle can be set to once every 24 hours. The cloud control platform merges the global models from different regions to generate a more generalized nationwide model, which is then distributed to the various edge computing servers.

[0126] Training termination conditions include: the global model's loss function value falling below a preset threshold, or reaching a preset maximum number of training epochs. The preset threshold can be set to 0.01, and the maximum number of training epochs can be set to 100 epochs. When the training termination conditions are met, the training process ends, and the final global model can be deployed to roadside units and vehicle-mounted units for real-time traffic prediction and abnormal behavior detection.

[0127] Example 2

[0128] Reference Figure 2 This embodiment also provides a vehicle-road cooperative system based on federated learning, including:

[0129] Local training module: This module is used to collect local traffic perception data and train a pre-built local model using this data to obtain local model parameters. Deployed on roadside units and vehicle-mounted units, this module integrates data acquisition, preprocessing, and model training functions.

[0130] Weight Calculation Module: This module determines the aggregate weight factors in this training process. It integrates the calculation logic for stability, data quality (including image sharpness detection), and communication quality. This module is also deployed on the roadside unit and the vehicle-mounted unit, performing weight calculation locally without uploading any raw data.

[0131] Data transmission module: This module sends local model parameters and corresponding aggregate weight factors to the edge computing server. It employs V2X communication technology and supports encrypted transmission to ensure the security and reliability of data transmission.

[0132] Example 3

[0133] Furthermore, the present invention also provides an electronic device, including a processor, a memory, and a communication interface. The processor is connected to the memory and the communication interface via a bus. The memory stores a computer program, and when the processor executes the computer program, it implements the methods described in any of the above embodiments. The communication interface is specifically a V2X communication module (such as a C-V2X module or a DSRC module), supporting both direct communication and cellular communication modes.

[0134] Through the above-mentioned solution, this invention leverages the hardware advantages of roadside units and vehicle-mounted units, and through a multi-dimensional weighted aggregation mechanism, effectively solves the problems of data heterogeneity and communication bandwidth limitations in V2X scenarios while protecting data privacy, significantly improving the prediction accuracy of the model under complex road conditions.

[0135] The embodiments described above are merely examples of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application.

Claims

1. A method for training a distributed vehicle-road cooperative model based on federated learning, applied to a vehicle-road cooperative system comprising roadside units, on-board units, and edge computing servers, characterized in that, The method includes: The roadside unit and the vehicle-mounted unit respectively collect local traffic perception data and use the local traffic perception data to train a preset local model to obtain local model parameters. Determine the aggregate weighting factor for the roadside unit and the vehicle-mounted unit in this training; The aggregation weight factor is calculated based on the node stability dimension, data quality dimension, and communication quality dimension. The node stability dimension is determined based on whether the node is a fixed deployment. The data quality dimension includes at least the image clarity evaluation index of the collected data. The roadside unit and the vehicle-mounted unit send the local model parameters and the corresponding aggregate weight factors to the edge computing server; The edge computing server performs a weighted aggregation operation on each of the received local model parameters according to each of the aggregation weight factors, and updates the global model parameters; The determination of the aggregate weight factor for the roadside unit and the vehicle-mounted unit in this training includes: For the node stability dimension, a first preset stability weight is assigned to the roadside unit and a second preset stability weight is assigned to the vehicle unit, wherein the first preset stability weight is higher than the second preset stability weight; Alternatively, the second preset stability weight can be dynamically adjusted based on the moving speed of the vehicle-mounted unit; the higher the moving speed, the lower the second preset stability weight.

2. The method according to claim 1, characterized in that, The data quality dimension also includes data distribution entropy; The calculation method for the image sharpness evaluation index includes: Detect image contrast or edge gradient information in the local traffic perception data; When the image contrast or edge gradient information is lower than a preset sharpness threshold, it is determined that there is visual interference in the current environment, and the value of the data quality dimension in the aggregation weight factor is reduced.

3. The method according to claim 1, characterized in that, The communication quality dimension is determined based on the vehicle wireless communication channel status between the roadside unit or the vehicle-mounted unit and the edge computing server. The vehicle wireless communication channel status includes the reference signal received power or the signal-to-interference-plus-noise ratio. When the state of the vehicle wireless communication channel is below the communication threshold, the aggregation weight factor is reduced to prevent communication delay from affecting the convergence speed of the global model.

4. The method according to claim 1, characterized in that, The local model includes at least one of the following neural network structures: The convolutional neural network-long short-term memory network used for traffic flow prediction takes traffic video frames or radar point cloud sequences at consecutive time steps as input and outputs traffic flow predictions for future time steps. A graph convolutional neural network for abnormal behavior detection takes a vehicle trajectory topology graph as input and outputs an abnormal probability value for vehicles driving in the wrong direction or pedestrians entering the area.

5. The method according to claim 1, characterized in that, The roadside unit and the vehicle-mounted unit send the local model parameters and the corresponding aggregate weight factors to the edge computing server, including: Differential privacy encryption or homomorphic encryption is applied to the gradient information in the local model parameters to generate encrypted gradients; The encryption gradient and the aggregate weight factor in plaintext form are sent to the edge computing server.

6. The method according to claim 1, characterized in that, The method further includes: The edge computing server distributes the updated global model parameters to the roadside unit and the vehicle unit for use in the next round of local training. The edge computing server periodically synchronizes the global model parameters to the cloud control platform, which then performs cross-regional model scheduling or version management.

7. A vehicle-road cooperative system based on federated learning, using the vehicle-road cooperative distributed model training method as described in claim 1, characterized in that, include: The local training module is used to collect local traffic perception data and use the local traffic perception data to train a preset local model to obtain local model parameters. The weight calculation module is used to determine the aggregate weight factor in this training; wherein, the aggregate weight factor is calculated based on the node stability dimension, data quality dimension and communication quality dimension, the node stability dimension is determined based on whether the node is a fixed deployment, and the data quality dimension includes at least the image sharpness evaluation index of the collected data. The data transmission module is used to send the local model parameters and the corresponding aggregation weight factors to the edge computing server so that the edge computing server can perform a weighted aggregation operation.

8. An electronic device, characterized in that, include: processor; Memory, used to store computer programs; Communication interface, used to enable vehicle wireless communication; When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • Unmanned vehicle blind area risk early warning system based on vehicle infrastructure cooperation

    CN121281266A

  • Internet of vehicles data sharing method based on homomorphic encryption enhanced federated learning

    CN122247588A

  • ETC vehicle-road cooperation device and system

    CN122347835A