METHOD AND SYSTEM FOR DISTRIBUTED MODEL TRAINING

Abstract

The present application relates to a distributed model training method performed on the vehicle side, comprising: Step A: Acquiring data related to a defined target; Step B: Training a first model on the vehicle side using the data to determine initial update parameters of the first model, wherein initialization parameters of the first model are received from a central server or cloud platform; Step C: Sending the first update parameters of the first model to the central server or cloud platform, so that the central server or cloud platform can aggregate the update parameters received from multiple participating vehicles to update the global model and determine second update parameters for the first model;and step D: Receiving the second update parameters from the central server or cloud platform to update the first model on the vehicle side. The present application further relates to a computer program product, a domain controller, and a distributed model training system.

Description

AREA OF INVENTION

[0001] The present application relates to the training of vehicle models, more precisely a distributed model training procedure that is executed on the vehicle side, a computer program product, a domain controller and a distributed model training system. STATE OF THE ART

[0002] The long-tail problem (LTP) typically refers to a situation where, in certain domains or systems, a few frequently occurring events or situations consume the majority of attention and resources, while a large number of rare or unusual events are neglected; this collection of rare events is known as the "long tail." The long-tail problem manifests itself in different forms and has varying impacts across different domains.

[0003] In the field of autonomous driving, for example, the long-tail problem refers to scenarios or events that occur rarely but can have a significant impact on the safety and reliability of the system. Such problems can include extreme weather conditions, unusual traffic incidents, or specific road conditions. Because these long-tail scenarios occur only infrequently, the autonomous driving system may not have enough data to learn and adapt to these situations, resulting in poor system performance in such scenarios. The presence of the long-tail problem limits the widespread adoption of autonomous driving technologies, as it can lead to system failure or faulty decisions at critical moments.

[0004] To solve the “long-tail” problem, extensive data collection for vehicles is required, however, such comprehensive data collection is difficult to implement in practice, for example due to data protection restrictions, communication costs and compliance problems related to the collection of cartographic data. REVELATION OF THE INVENTION

[0005] The inventors of the present application have recognized that by means of distributed model training between several vehicles with local data, a global model can be built on the basis of virtually merged data without the need to exchange local individual or sample data, but only by exchanging model parameters (or intermediate results), thereby achieving a balance between data protection and joint data processing.

[0006] According to one aspect of the present application, a distributed model training method is provided, which is executed on the vehicle side, comprising: (Step A) acquiring data related to a defined target; (Step B) training a first model on the vehicle side using the data to determine first update parameters of the first model, wherein initialization parameters of the first model are received from a central server or cloud platform; (Step C) sending the first update parameters of the first model to the central server or cloud platform, so that the central server or cloud platform can aggregate the update parameters received from multiple participating vehicles to update a global model and determine second update parameters for the first model;and (step D) receiving the second update parameters from the central server or cloud platform to update the first model on the vehicle side.

[0007] As a supplement to or alternative to the solution mentioned above, the procedure mentioned above may also include: repeatedly performing steps B to D until the first model achieves an expected performance or a termination condition is met.

[0008] As a supplement to or alternative to the solution mentioned above, the defined objective of the method mentioned above is to improve an autonomous driving system, and step A comprises the following: acquiring initial data using vehicle sensors, wherein the initial data is related to the improvement of the autonomous driving system and the initial data includes dynamic characteristics of the vehicle as well as driving environment-related information relating to the vehicle, wherein the dynamic characteristics include the position, speed and acceleration of the vehicle, and wherein the driving environment-related information relating to the vehicle includes the driving trajectories, information about surrounding vehicles (position, size, relative speed, etc.).), lane information (including curvature, width, shape of lane markings) as well as driving-related information (traffic lights, speed limit signs, zebra crossings, etc.).

[0009] As a supplement to or alternative to the solution mentioned above, the defined objective of the procedure mentioned above is to improve a navigation system, and step A comprises the following: acquiring second data related to the improvement of the navigation system, wherein the second data includes GPS position, speed, direction of travel and traffic conditions.

[0010] As a supplement to or alternative to the solution mentioned above, the defined objective of the procedure mentioned above is to optimize energy efficiency, and step A includes the following: collecting third-party data related to energy efficiency, wherein the third-party data includes battery state, energy consumption data, driving behavior, speed, acceleration, and driving mode.

[0011] As a supplement or alternative to the solution mentioned above, the first update parameters in the aforementioned procedure are the weight of the first model.

[0012] As a supplement to or alternative to the solution mentioned above, the procedure mentioned above also includes the following: registering the vehicle in a federal learning network and ensuring that the vehicle can communicate securely with the central server or cloud platform.

[0013] As a supplement to or alternative to the solution mentioned above, step C of the procedure mentioned above includes the following: encrypted uploading of the first update parameters of the first model to the central server or cloud platform, so that the central server or cloud platform can collect model updates from multiple vehicles registered in the federal learning network and aggregate them using a federal learning algorithm to improve the global model.

[0014] As a supplement to or alternative to the solution mentioned above, the aforementioned procedure includes achieving the expected performance or fulfilling the termination condition of the first model under one or more of the following conditions: the first model achieves a predetermined accuracy or other performance characteristic on a validation or test dataset; the first model is stopped after a predetermined number of iterations or training rounds; the performance improvement of the first model falls below a certain threshold over several successive iterations; the training process is automatically terminated after reaching a predetermined time; and the training must be stopped under certain conditions due to limited computing resources or memory.

[0015] According to another aspect of the present application, a computer program product is provided which comprises a computer program, wherein the computer program product, when executed by a processor, implements the aforementioned method.

[0016] According to another aspect of the present application, a domain controller is provided, comprising a memory, a processor and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the aforementioned method.

[0017] According to another aspect of the present application, a distributed model training system is provided, wherein the system comprises: a central server or cloud platform; and multiple vehicles registered with the central server or cloud platform to participate in federated learning, each vehicle comprising a domain controller as described above.

[0018] The distributed model training solution for vehicles described in the embodiments of the present application uses data related to a defined target to train a first model (e.g., a machine learning model) on the vehicle side and sends initial update parameters of the first model (instead of the original data used for training) to a central server or cloud platform. This means the original data never leaves the local device (i.e., the vehicle), thus contributing to data privacy and security. Furthermore, only the model update is transmitted between the vehicle and the central server or cloud platform, reducing the data transmission volume compared to directly transmitting the training data.Finally, the first model is trained on distributed devices, thereby realizing edge computing, which can reduce the need for centralized data servers. DESCRIPTION OF THE FIGURES

[0019] The following detailed description, with reference to the figures, further clarifies the aforementioned and other purposes and advantages of the present application, whereby identical or similar essential elements are marked with identical reference symbols. Fig. Figure 1 shows a flowchart of a distributed model training procedure performed on the vehicle side according to an embodiment of the present application; Fig. Figure 2 shows a schematic representation of a distributed model training system according to an embodiment of the present application; and Fig. Figure 3 shows a schematic representation of the structure of a domain controller according to an embodiment of the present application. DETAILED DESCRIPTION OF THE EXECUTION FORMS

[0020] The distributed model training solution for vehicles described in the various embodiments of the present application is explained in detail below with reference to the corresponding figures.

[0021] Fig. Figure 1 shows a flowchart of a distributed model training procedure 1000 performed on the vehicle side according to an embodiment of the present application. As in Fig. As shown in Figure 1, the procedure 1000 comprises the following steps: In step S110, data relating to a defined goal is collected with regard to the vehicle; In step S120, a first model is trained on the vehicle side using the data to determine the first update parameters of the first model, whereby an initialization parameter of the first model is received from a central server or a cloud platform; In step S130, the first update parameters of the first model are sent to the central server or cloud platform, so that the central server or cloud platform can aggregate the update parameters received from multiple participating vehicles to update a global model and determine second update parameters for the first model; and In step S140, the second update parameters are received from the central server or cloud platform to update the first model on the vehicle side.

[0022] In one embodiment, although not in Fig. As shown in Figure 1, the aforementioned procedure 1000 can further include, for example, the following: Before executing step S110, the use case and the objective of using distributed learning (federal learning) in the vehicle environment must be clearly defined. In one embodiment, the defined objective includes tasks such as improving an autonomous driving system, extending a navigation system, or optimizing energy efficiency.

[0023] In step S110, data related to a defined goal is acquired from the vehicle. In an embodiment where the defined goal is the improvement of an autonomous driving system, step S110 may include the following: acquiring initial data using the vehicle's sensors, wherein the initial data is related to the improvement of the autonomous driving system and includes the vehicle's dynamic characteristics as well as driving environment information, wherein the dynamic characteristics include the vehicle's position, speed, and acceleration, and wherein the driving environment information includes the vehicle's trajectories and information about the surrounding driving environment (position, size, relative speed, etc.).), lane information (including curvature, width, shape of lane markings) as well as driving-related information (traffic lights, speed limit signs, crosswalks, etc.). In an embodiment where the defined goal is to improve a navigation system, step S110 may include: acquiring second data related to improving the navigation system, wherein the second data includes GPS position, speed, direction of travel, and traffic conditions, etc. In an embodiment where the defined goal is to optimize energy efficiency, step S110 may include: acquiring third data related to energy efficiency, wherein the third data includes battery state, energy consumption data, driving behavior, speed, acceleration, and driving mode, etc.

[0024] After the data related to the defined goal has been acquired, in step S120 a first model is trained on the vehicle side using the data to determine the first update parameters of the first model, whereby the initialization parameters of the first model are received from a central server or a cloud platform.

[0025] In the context of the present application, the first model should be suitable for a specific use case, such as a machine learning model (including a deep learning model). Furthermore, the first model must, in one or more embodiments, take into account factors such as the computational and memory limitations of the onboard system. For example, certain models (such as deep learning models) place higher demands on the hardware, so lightweight models are preferentially chosen when computing resources are limited. The data volume also influences the model selection: some models, such as decision trees, are suitable for small datasets, while deep learning models may require large datasets.

[0026] In one embodiment, the first update parameters of the first model are the weights of the first model. In other words, initialization weights W(t) are received from the central server or cloud platform to initialize the first model. Subsequently, based on the training data, a training update of the first model takes place on the local device (i.e., on the vehicle side), obtaining updated parameters, namely the updated weights, which are then sent from the vehicle to the central server or cloud platform.

[0027] In one embodiment, the aforementioned method 1000 may further include: registering the vehicle in a federated learning network and ensuring that the vehicle can communicate securely with the central server or cloud platform. In other words, the vehicle must undergo an authentication process to ensure that only authorized vehicles can join the federated learning network. In one embodiment, to ensure communication security, a secure key exchange protocol (such as Diffie-Hellman key exchange) may be used to generate a shared key for subsequent communication encryption. In one embodiment, the vehicle must obtain a digital certificate from a trusted Certificate Authority (CA) to prove its identity.

[0028] In step S130, the first update parameters of the first model are sent to the central server or cloud platform. This allows the central server or cloud platform to aggregate the update parameters received from multiple participating vehicles in order to update the global model and determine second update parameters for the first model (such as an update value for the weights). In one embodiment, the central server or cloud platform uses a federated learning algorithm to aggregate the update parameters uploaded by the participating vehicles and thereby improve the global model.

[0029] In the context of the present application, federated learning, as a novel paradigm of machine learning, represents a method that enables a procedure for training models under the condition of protecting data privacy.It is particularly suitable for the field of autonomous driving, for example: (1) Since autonomous driving systems need to process a large amount of sensitive personal and environmental data, federated learning enables the training of more accurate models by sharing model updates without having to share the original data; (2) Since autonomous driving systems operate in different regions and under different conditions, federated learning can improve the generalizability of the models by integrating data from different sources; (3) Federalized learning enables computations to be performed directly on devices on the vehicle side, thereby reducing the need for data transmissions and thus decreasing both latency and bandwidth usage.

[0030] The federated learning algorithm is a key technology for implementing distributed machine learning, enabling multiple participants to collaboratively train a model while maintaining the data privacy of each party. In one or more implementations, the federated learning algorithm can include: federated mean algorithm (FedAvg), horizontal federated learning, vertical federated learning, federated transfer learning, etc.

[0031] After the global model has been updated by the central server or cloud platform, the updated global model is sent back to all participating vehicles. Therefore, in step S140, the second update parameters are received from the central server or cloud platform to update the first model on the vehicle side. This updated model now integrates the knowledge learned by all participating vehicles.

[0032] In one or more embodiments, the aforementioned method 1000 may further comprise: repeatedly performing steps S120 to S140 until the first model achieves an expected performance or a termination condition is met.The following are some common termination conditions and performance evaluation methods: (1) Performance metrics: The model achieves a predefined accuracy or other performance characteristic on a validation or test dataset; (2) Number of training rounds: The model is stopped after a predefined number of iterations or training rounds; (3) Performance improvement threshold: If the model's performance improvement falls below a certain threshold over several consecutive iterations, the model is considered to have converged; (4) Time limit: The training process is automatically terminated after a predefined time has elapsed; (5) Resource limit: Due to limited computing resources or memory, training must be stopped under certain conditions.

[0033] In one embodiment, the evaluation of model performance can be performed both on the vehicle side and on a cloud server. For example, when evaluating the first model on the vehicle side, the autonomous driving system simultaneously executes two versions of the control algorithms: a main system that actually controls the vehicle, and a shadow system used exclusively for learning and evaluation purposes. The shadow system uses the same sensor inputs and control commands as the main system but does not influence the actual control of the vehicle. The first model is integrated into the shadow system, and its performance can be evaluated by comparing the predictions or decisions of the shadow system with those of the main system. Of course, the evaluation of the global model's performance can also be performed on the cloud server, as in...utilizing the advantages of cloud-based data and computational examples.

[0034] Furthermore, it is readily apparent to those skilled in the art that the aforementioned method 1000 for distributed model training on the vehicle side, according to one or more embodiments of the present application, can be implemented by a computer program. For example, the computer program is contained in a computer program product, wherein, when executed by a processor, the computer program performs the method 1000 for distributed model training on the vehicle side according to one or more embodiments of the present application. As another example, if a computer-readable storage medium (such as a USB drive) on which the computer program is stored is connected to a computer, the method 1000 for distributed model training on the vehicle side, according to one or more embodiments of the present application, can be performed by executing the computer program.

[0035] It was on Fig. Reference is made to Figure 2, which shows a schematic representation of a distributed model training system according to an embodiment of the present application. As in Fig. As shown in Figure 2, the system comprises the following: a central server or cloud platform; and multiple vehicles registered with the central server or cloud platform to participate in federal learning.

[0036] In one embodiment, a vehicle's driver assistance system (such as automatic emergency braking, lane keeping assist, or adaptive cruise control) can be improved using federated learning. First, the specific functions and performance indicators to be improved in the driver assistance system must be determined; then, data from vehicle sensors (such as cameras, radar, LiDAR, etc.) are used.) data is collected during the journey, including but not limited to the vehicle's position, speed, acceleration, and surroundings; the vehicle is then registered in a federated learning network, and its ability to communicate securely with the central server or cloud platform is ensured; the driver assistance model is then trained in the vehicle's local system using the collected data, which may require model simplification to accommodate the vehicle's computing capabilities; subsequently, the model updates trained in the vehicle (such as weights, etc.) are applied.) encrypted and uploaded to the central server or cloud platform; the central server or cloud platform then collects the model updates from the vehicles and performs an aggregation using a federated learning algorithm to improve the global model; subsequently, the central server or cloud platform distributes the aggregated and updated global model to all vehicles participating in federated learning to update their local models; and finally, the model is optimized through continuous iteration to adapt to new driving scenarios and conditions.

[0037] In the aforementioned embodiment, appropriate privacy measures must be implemented throughout the entire process, such as differential privacy and secure data transmission protocols. Furthermore, user feedback on the driver assistance system can be collected to further adapt and optimize the model.

[0038] Given the large number of vehicles on the end-user side, available data transmission bandwidth can be limited. Therefore, in some implementations, content is cached on servers located closer to the users (vehicles) via a Content Delivery Network (CDN). This optimizes communication between the vehicles and the central server or cloud platform, reduces transmission distance and latency, and improves bandwidth utilization. In one or more implementations, the communication protocol can also be optimized, for example, by adapting the communication protocols to specific application requirements and eliminating unnecessary overhead and features. This can further reduce bandwidth usage and increase the efficiency of federated learning.

[0039] See Fig. 3, in which a schematic representation of the structure of a domain controller 3000 according to an embodiment of the present application is shown. As in Fig. As shown in Figure 3, the domain controller 3000 comprises a memory 310 and a processor 320, with a computer program stored in the memory 310. In one embodiment, the processor 320 executes the computer program to implement the distributed model training method 1000 on the vehicle side. Current in-vehicle domain controllers typically have powerful CPU and GPU computing capacity, so that by having the processor 320 execute the computer program to implement the distributed model training method 1000 on the vehicle side, the available computing resources can be used effectively.

[0040] In summary, the distributed model training solution for vehicles described in the embodiments of the present application uses data related to a defined target to train a first model on the vehicle side and sends initial update parameters of this first model (instead of the original data used for training) to a central server or cloud platform. Thus, the original data does not leave the local device (i.e., the vehicle), and this distributed model training solution contributes to data privacy and data security. Furthermore, only the model update is transmitted between the vehicle and the central server or cloud platform, reducing the data transmission volume compared to directly transmitting the training data.Finally, the first model is trained on distributed vehicle devices, thereby realizing edge computing, which can reduce the need for centralized data servers (i.e., decentralized training takes place).

[0041] The examples mentioned above mainly illustrate the distributed model training solution of the embodiments of the present application. Although only some embodiments of the present application have been described, those skilled in the art will understand that the present application can be implemented in many other forms without departing from its spirit and scope. Therefore, the examples and embodiments shown are considered illustrative and not limiting, and the present application can cover various modifications and substitutions without departing from the spirit and scope of the application as defined by the claims.

Claims

[1] Method for distributed model training that is performed on the vehicle side, characterized by , that the procedure includes the following: Step A: Collecting data related to a defined goal; Step B: Training an initial model on the vehicle side using the data to determine initial update parameters of the first model, with initialization parameters of the first model being received from a central server or cloud platform; Step C: Sending the first update parameters of the first model to the central server or cloud platform, so that the central server or cloud platform can aggregate the update parameters received from multiple participating vehicles to update a global model and determine second update parameters for the first model; and Step D: Receiving the second update parameters from the central server or cloud platform to update the first model on the vehicle side. [2] Method according to claim 1, further comprising: Repeated execution of step B through step D until the first model achieves an expected performance or a termination condition is met. [3] Method according to claim 1, wherein the defined objective is the improvement of an autonomous driving system and step A comprises: acquiring first data using sensors of the vehicle, wherein the first data are related to the improvement of the autonomous driving system and include the dynamic characteristics of the vehicle as well as driving environment-related information relating to the vehicle, wherein the dynamic characteristics include the position, speed and acceleration of the vehicle. [4] Method according to claim 1, wherein the defined objective is the improvement of a navigation system and step A comprises: acquiring second data related to the improvement of the navigation system, wherein the second data comprises GPS position, speed, direction of travel and traffic conditions. [5] Method according to claim 1, wherein the defined objective is the optimization of energy efficiency and step A comprises: acquiring third-party data related to energy efficiency, wherein the third-party data includes battery state, energy consumption data, driving behavior, speed, acceleration and driving mode. [6] Method according to claim 1, wherein the first update parameters are the weights of the first model. [7] Method according to claim 1, further comprising: Registering the vehicle in a federal learning network and ensuring, that the vehicle can communicate securely with the central server or cloud platform. [8] The method of claim 7, wherein step C comprises: encrypted uploading of the first update parameters of the first model to the central server or cloud platform, so that the central server or The cloud platform can collect model updates from all vehicles registered in the federal learning network and aggregate them using a federal learning algorithm to improve the global model. [9] The method of claim 2, wherein achieving the expected performance or fulfilling the termination condition of the first model comprises one or more of the following conditions: The first model achieves a predefined accuracy or other performance characteristic on a validation or test data set; The first model is generated after a predetermined number of iterations or Training laps stopped; The performance improvement of the first model falls below a certain threshold over several successive iterations; The training process ends automatically after a predetermined time has elapsed; and Training may need to be stopped under certain conditions due to limited computing resources or storage space. [10] Computer program product, comprising a computer program, characterized by that the computer program, when executed by a processor, implements the method according to one of claims 1 to 9. [11] Domain controller comprising memory, a processor and a computer program stored in memory, characterized by , that the processor executes the computer program to implement the method according to any one of claims 1 to 9. [12] Distributed model training system, characterized by , that the system includes the following: a central server or cloud platform; and multiple vehicles registered with the central server or cloud platform to participate in federated learning, each vehicle comprising a domain controller according to claim 11.

Images