Driving model training method and device, storage medium and equipment

Abstract

The invention provides a driving model training method and device, a storage medium and equipment, and the method comprises the steps: training a first driving model issued by an edge server by each vehicle end according to the respective driving data and a preset number of training times, and obtaining a trained first model parameter difference value; each edge server trains a first driving model according to first model parameter difference values uploaded by a plurality of corresponding vehicle ends, obtains a third driving model and edge aggregation times, and re-issues the third driving model to each vehicle end; updating the first driving model of each vehicle end and obtaining a new first model parameter difference value again until the edge aggregation times are equal to the edge aggregation times threshold value, and obtaining a second model parameter difference value; and the cloud server trains the first driving model according to the second model parameter difference values uploaded by the plurality of edge servers to obtain a target vehicle driving model. The training burden of the driving model can be reduced, and the training efficiency can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of driving model training, and in particular to a driving model training method, apparatus, storage medium and device. Background Art

[0002] With the rapid development of autonomous driving technology, autonomous vehicles are becoming more and more popular. When an autonomous vehicle is on the road, data from the driving process can be collected through information collection devices such as cameras and sensors. However, since autonomous vehicles have a lot of data, they also collect a lot of data. Therefore, using such a large amount of data for driving model training will place a heavy burden on the equipment running the training, resulting in low training efficiency. Therefore, the existing technology has the technical defects of heavy burden on driving model training and low training efficiency. Summary of the invention

[0003] The purpose of this application is to overcome the shortcomings and deficiencies in the prior art and to provide a driving model training method, apparatus, storage medium and device that can share the burden of training the driving model through the vehicle side, edge server and cloud server, thereby improving the training efficiency of the driving model.

[0004] A first aspect of an embodiment of the present application provides a driving model training method, which is applied to a driving model training system, including multiple vehicle terminals, several edge servers and a cloud server, and the method includes:

[0005] Each of the edge servers sends the first driving model to a corresponding number of vehicle terminals;

[0006] Each of the vehicle terminals trains the first driving model according to its own driving data and a preset number of training times to obtain a trained second driving model, and obtains a first model parameter difference according to the second driving model and the first driving model; each of the vehicle terminals uploads the first model parameter difference to a corresponding edge server;

[0007] Each of the edge servers trains the first driving model according to the first model parameter differences uploaded by the corresponding plurality of vehicle terminals to obtain a third driving model and an edge aggregation number; if the edge aggregation number is less than a preset edge aggregation number threshold, the third driving model is re-sent to each of the vehicle terminals to update the first driving model of each vehicle terminal and re-acquire a new first model parameter difference until the edge aggregation number is equal to the edge aggregation number threshold, thereby obtaining a fourth driving model, and obtaining a second model parameter difference according to the fourth driving model and the first driving model; each of the edge servers uploads the second model parameter difference to the cloud server;

[0008] The cloud server trains the first driving model according to the second model parameter differences uploaded by the plurality of edge servers to obtain a target vehicle driving model.

[0009] The second aspect of the embodiment of the present application provides a driving model training method, wherein the cloud server sends the obtained target vehicle driving model to multiple vehicle terminals through a plurality of edge servers to improve the autonomous driving performance of each of the vehicle terminals.

[0010] A third aspect of an embodiment of the present application provides a driving model training device, which is applied to a driving model training system, including multiple vehicle terminals, several edge servers and a cloud server, and the device includes:

[0011] A driving model sending module, used to drive each of the edge servers to send the first driving model to a corresponding number of vehicle terminals;

[0012] A vehicle-side training module, used to drive each of the vehicle-sides to train the first driving model according to its own driving data and a preset number of training times, to obtain a trained second driving model, and to obtain a first model parameter difference according to the second driving model and the first driving model; each of the vehicle-sides uploads the first model parameter difference to a corresponding edge server;

[0013] The edge server training module is used to drive each of the edge servers to train the first driving model according to the first model parameter difference values ​​uploaded by the corresponding plurality of vehicle terminals, and obtain a third driving model and an edge aggregation number; if the edge aggregation number is less than a preset edge aggregation number threshold, the third driving model is re-sent to each of the vehicle terminals to update the first driving model of each vehicle terminal and re-acquire a new first model parameter difference value, until the edge aggregation number is equal to the edge aggregation number threshold, a fourth driving model is obtained, and a second model parameter difference value is obtained according to the fourth driving model and the first driving model; each of the edge servers uploads the second model parameter difference value to the cloud server;

[0014] The cloud server training module is used to drive the cloud server to train the first driving model according to the second model parameter differences uploaded by several edge servers to obtain the target vehicle driving model.

[0015] A fourth aspect of an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the driving model training method as described above are implemented.

[0016] The fifth aspect of an embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the steps of the driving model training method as described above when executing the computer program.

[0017] Compared with the related art, each vehicle end of the present application performs model training according to its own driving data, and then the edge server performs aggregate training according to the first model parameter difference obtained by the training of each vehicle end, until the edge aggregation times are equal to the edge aggregation times threshold, and obtains the second model parameter difference that integrates the training results of multiple vehicle ends, and the cloud server then performs aggregate training on the first driving model according to the second model parameter difference uploaded by each edge server to obtain the target vehicle driving model. By having the vehicle end, edge server and cloud server be responsible for part of the model training respectively, the vehicle end, edge server and cloud server share the burden of training the driving model, and the data for training each vehicle end is only its own driving data, and the data for training each edge server is only the first model parameter difference uploaded by the corresponding number of vehicle ends, and the data that the cloud server needs to train is the second model parameter difference uploaded by each edge server, so the training burden of the vehicle end, edge server and cloud server is very small, so that the vehicle end, edge server and cloud server can quickly train the driving model under the state of small burden, which can improve the training efficiency of the driving model.

[0018] In order to provide a clearer understanding of the present application, the specific implementation of the present application will be described below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic diagram of a training framework for a driving model training method according to an embodiment of the present application.

[0020] Figure 2 This is a flowchart of a driving model training method according to an embodiment of the present application.

[0021] Figure 3 This is a schematic diagram of module connections of a driving model training device according to an embodiment of the present application.

[0022] 100. Driving model training device; 101. Driving model sending module; 102. Vehicle-side training module; 103. Edge server training module; 104. Cloud server training module. DETAILED DESCRIPTION

[0023] In order to make the objectives, technical solutions and advantages of the present application more clear, the embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0024] It should be clear that the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present application.

[0025] When the following description relates to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances. The singular forms of "a", "said" and "the" used in the present application and the appended claims are also intended to include the majority form, unless the context clearly indicates other meanings. The words "if" / "if" used herein can be interpreted as "at the time of" or "when" or "in response to determination".

[0026] In addition, in the description of this application, unless otherwise specified, "plurality" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects are in an "or" relationship.

[0027] See also Figure 1 , which is a training framework diagram of a driving model training method of an embodiment of the present application. The training framework of the present application can be divided into a vehicle user end, an edge layer and a cloud layer, wherein each vehicle end of the vehicle user end performs model training according to its own driving data, and then the edge server of the edge layer performs aggregation training according to the first model parameter difference obtained by the training of each vehicle end, until the edge aggregation times are equal to the edge aggregation times threshold, and the second model parameter difference that integrates the training results of multiple vehicle ends is obtained, and the cloud server of the cloud layer then performs aggregation training on the first driving model according to the second model parameter difference uploaded by each edge server. Among them, Figure 1 Vehicle 1 and vehicle M / N are both vehicle terminals, and the roadside unit is a medium device for the edge server to communicate with the network, that is, the edge server can exchange data with the vehicle terminal and the cloud server respectively through the roadside unit. The roadside unit is a device with network communication function set near the road, such as a base station set near the road.

[0028] See also Figure 2 , Figure 2This is a flowchart of a driving model training method according to an embodiment of the present application, which is applied to a driving model training system, including multiple vehicle terminals, several edge servers and a cloud server. The method includes:

[0029] S1: Each of the edge servers sends a first driving model to a corresponding number of vehicle terminals.

[0030] Among them, the edge server refers to a server device that has the ability to train models and exchanges data with the vehicle end through the roadside unit. The first driving model refers to the driving model stored in the edge server. Before training begins, the first driving model is the initial driving model, and the first driving model is stored in the edge server and the cloud server respectively. The several vehicle ends corresponding to the edge server refer to the vehicle ends that can exchange data with the edge server. Since the vehicle end has high flexibility, and the first driving models stored in the edge servers of different road locations and terrains may not be exactly the same, therefore, the first driving model is stored in the edge server, which can reduce the storage load caused by storing multiple first driving models on the vehicle end. When the vehicle end enters the data interaction range of the edge server, the edge server can send the first driving model to the vehicle end to drive the vehicle end to train the first driving model according to the driving data.

[0031] S2: Each of the vehicle terminals trains the first driving model according to its own driving data and a preset number of training times to obtain a trained second driving model, and obtains a first model parameter difference according to the second driving model and the first driving model; each of the vehicle terminals uploads the first model parameter difference to the corresponding edge server.

[0032] The preset number of training times can be stored in the vehicle end or in the edge server. Since different road locations and terrains have different requirements for the number of training times, and a unified number of training times is conducive to improving the balance of training the first driving model at each vehicle end, preferably, the number of training times is preset in the edge server, and when the edge server sends the first driving model to the vehicle end, the edge server sends the number of training times to the vehicle end.

[0033] The second driving model refers to a driving model obtained by training the first driving model according to the number of training times.

[0034] The first model parameter difference is used to represent the parameter difference between the second driving model and the first driving model, that is, it represents the model parameter difference between the first driving model in the vehicle end before and after training. Wherein, obtaining the first model parameter difference according to the second driving model and the first driving model includes: the vehicle end determines the model parameter difference between the second driving model and the first driving model as the first model parameter difference.

[0035] For example, the first model parameter difference can be obtained by the following formula:

[0036] p m =w 0 -w

[0037] Among them, p m is the first model parameter difference, w 0 is the first driving model of the vehicle, and w is the second driving model of the vehicle.

[0038] S3: Each of the edge servers trains the first driving model according to the first model parameter differences uploaded by the corresponding number of vehicle terminals to obtain a third driving model and an edge aggregation number; if the edge aggregation number is less than a preset edge aggregation number threshold, the third driving model is re-sent to each of the vehicle terminals to update the first driving model of each vehicle terminal and re-acquire a new first model parameter difference until the edge aggregation number is equal to the edge aggregation number threshold, thereby obtaining a fourth driving model, and obtaining a second model parameter difference according to the fourth driving model and the first driving model; each of the edge servers uploads the second model parameter difference to the cloud server.

[0039] The edge servers train the first driving model according to the first model parameter differences uploaded by the corresponding plurality of vehicle terminals to obtain a third driving model and the number of edge aggregations, including:

[0040] S31: Each of the edge servers obtains an optimal weight of each of the first model parameter differences according to a preset learning algorithm and a plurality of the first model parameter differences.

[0041] For example, the optimal weights of the first model parameter differences are obtained by the following formula:

[0042]

[0043] in, is the optimal weight of the mth first model parameter difference, ∈ is the strategy selection, ∈∈[0,1], μ0 is the prior fixed weight, μ0=1 / m, μ m is the initial weight of the mth first model parameter difference, p m is the mth first model parameter difference.

[0044] S32: Each of the edge servers obtains a first gradient descent coefficient according to each of the first model parameter differences and their optimal weights.

[0045] For example, the first gradient descent coefficient can be obtained by the following formula:

[0046]

[0047] Among them, d r is the first gradient descent coefficient, which is used to indicate the gradient descent direction of the aggregate training.

[0048] S33: Each edge server trains the first driving model according to the first gradient descent coefficient and a preset first training step size to obtain the third driving model, and determines the number of times the third driving model is obtained as the edge aggregation number.

[0049] For example, the third driving model can be obtained by the following formula:

[0050] W n (r+1)←W n (r)-η r d r

[0051] Among them, W n (r+1) is the third driving model of the nth edge server, W n (r) is the first driving model of the nth edge server, η r is the first training step.

[0052] The fourth driving model is the third driving model obtained by the last aggregation training when the number of edge aggregations is equal to the edge aggregation number threshold.

[0053] The obtaining of the second model parameter difference according to the fourth driving model and the first driving model includes: the edge server determining the difference between the model parameters of the fourth driving model and the model parameters of the first driving model as the second model parameter difference.

[0054] S4: The cloud server trains the first driving model according to the second model parameter differences uploaded by the edge servers to obtain a target vehicle driving model.

[0055] Wherein, step S4 may include:

[0056] The cloud server trains the first driving model according to the second model parameter difference uploaded by several edge servers to obtain a fifth driving model and a cloud aggregation number; if the cloud aggregation number is less than a preset cloud aggregation number threshold, the fifth driving model is re-sent to several edge servers to update the first driving model of each edge server and re-acquire a new second model parameter difference until the cloud aggregation number is equal to the cloud aggregation number threshold, thereby obtaining the target vehicle driving model.

[0057] The number of cloud aggregation times refers to the number of times the cloud server obtains the fifth driving model. Specifically, the cloud server trains the first driving model according to the difference of the second model parameters uploaded by the edge servers to obtain the fifth driving model and the number of cloud aggregation times, including:

[0058] S41: The cloud server obtains the optimal weight of each of the second model parameter differences according to a preset learning algorithm and a plurality of the second model parameter differences.

[0059] S42: The cloud server obtains a second gradient descent coefficient according to each of the second model parameter differences and their optimal weights.

[0060] S43: The cloud server trains the first driving model according to the second gradient descent coefficient and a preset second training step size to obtain the fifth driving model, and determines the number of times the fifth driving model is obtained as the cloud aggregation number.

[0061] Therefore, each edge server performs parameter aggregation training based on the first model parameter difference of each vehicle end that reflects the difference between the second driving model and the first driving model to obtain a fourth driving model, and then each edge server uploads the second model parameter difference that reflects the difference between the fourth driving model and the first driving model to the cloud server, so that the cloud server performs parameter aggregation training to obtain the target vehicle driving model.

[0062] Specifically, when the cloud aggregation times are less than the cloud aggregation times threshold, the cloud server will re-send the fifth driving model to each edge server as the new first driving model of each edge server. At this time, the edge aggregation times of the edge server are reset to zero, and then the edge server will also send the new first driving model to each vehicle end to obtain the new first model parameter difference obtained by each vehicle end local training to receive the new first driving model. The edge server will also train the new first driving model stored in the new first model parameter difference to obtain a new third driving model and edge aggregation times, and when the edge aggregation times are less than the edge aggregation times threshold, the new third driving model will be re-sent to each of the vehicle ends for training, and the execution will be repeated until the edge aggregation times are equal to the edge aggregation times threshold, and the third driving model obtained by the last aggregation training is determined as the new fourth driving model, and then the second model parameter difference is obtained according to the new fourth driving model and the new first driving model, and then uploaded to the cloud server for the cloud server to perform the next aggregation training, and obtain a new fifth driving model and the cumulative cloud aggregation times. When the cloud aggregation times are equal to the cloud aggregation times threshold, the fifth driving model obtained by the last aggregation training of the cloud server is the target vehicle driving model.

[0063] Compared with the related art, each vehicle end of the present application performs model training according to its own driving data, and then the edge server performs aggregate training according to the first model parameter difference obtained by the training of each vehicle end, until the edge aggregation times are equal to the edge aggregation times threshold, and obtains the second model parameter difference that integrates the training results of multiple vehicle ends, and the cloud server then performs aggregate training on the first driving model according to the second model parameter difference uploaded by each edge server to obtain the target vehicle driving model. By having the vehicle end, edge server and cloud server be responsible for part of the model training respectively, the vehicle end, edge server and cloud server share the burden of training the driving model, and the data for training each vehicle end is only its own driving data, and the data for training each edge server is only the first model parameter difference uploaded by the corresponding number of vehicle ends, and the data that the cloud server needs to train is the second model parameter difference uploaded by each edge server, so the training burden of the vehicle end, edge server and cloud server is very small, so that the vehicle end, edge server and cloud server can quickly train the driving model under the state of small burden, which can improve the training efficiency of the driving model.

[0064] The second embodiment of the present application also provides a driving model training method, including: the cloud server sends the obtained target vehicle driving model to multiple vehicle terminals through several edge servers to improve the autonomous driving performance of each of the vehicle terminals.

[0065] See also Figure 3 The third embodiment of the present application provides a driving model training device 100, which is applied to a driving model training system, including multiple vehicle terminals, several edge servers and a cloud server, and the device includes:

[0066] A driving model sending module 101 is used to drive each of the edge servers to send the first driving model to a corresponding number of vehicle terminals;

[0067] The vehicle-side training module 102 is used to drive each of the vehicle-sides to train the first driving model according to the respective driving data and a preset number of training times, to obtain a trained second driving model, and to obtain a first model parameter difference according to the second driving model and the first driving model; each of the vehicle-sides uploads the first model parameter difference to a corresponding edge server;

[0068] The edge server training module 103 is used to drive each of the edge servers to train the first driving model according to the first model parameter difference values ​​uploaded by the corresponding plurality of vehicle terminals, and obtain a third driving model and an edge aggregation number; if the edge aggregation number is less than a preset edge aggregation number threshold, the third driving model is re-sent to each of the vehicle terminals to update the first driving model of each vehicle terminal and re-acquire a new first model parameter difference value, until the edge aggregation number is equal to the edge aggregation number threshold value, a fourth driving model is obtained, and a second model parameter difference value is obtained according to the fourth driving model and the first driving model; each of the edge servers uploads the second model parameter difference value to the cloud server;

[0069] The cloud server training module 104 is used to drive the cloud server to train the first driving model according to the second model parameter differences uploaded by the edge servers to obtain the target vehicle driving model.

[0070] It should be noted that the driving model training device 100 provided in the third embodiment of the present application only uses the division of the above-mentioned functional modules as an example when executing the driving model training method. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the driving model training device 100 provided in the third embodiment of the present application and the driving model training method of the first embodiment of the present application belong to the same concept. The implementation process thereof is detailed in the method embodiment and will not be repeated here.

[0071] The fourth embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the driving model training method as described above are implemented.

[0072] The fifth embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, and the processor implements the steps of the driving model training method as described above when executing the computer program.

[0073] The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present application. Ordinary technicians in this field can understand and implement it without creative work.

[0074] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0075] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the function selected in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 function selected in a box or multiple boxes.

[0076] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 steps for the function selected in a box or multiple boxes.

[0077] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0078] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0079] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0080] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0081] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A driving model training method, characterized in that: Applied to a driving model training system, including multiple vehicle terminals, several edge servers and a cloud server, the method includes: Each of the edge servers sends the first driving model to a corresponding number of vehicle terminals; Each of the vehicle terminals trains the first driving model according to its own driving data and a preset number of training times to obtain a trained second driving model, and obtains a first model parameter difference according to the second driving model and the first driving model; each of the vehicle terminals uploads the first model parameter difference to a corresponding edge server; Each of the edge servers trains the first driving model according to the first model parameter differences uploaded by the corresponding plurality of vehicle terminals to obtain a third driving model and an edge aggregation number; if the edge aggregation number is less than a preset edge aggregation number threshold, the third driving model is re-sent to each of the vehicle terminals to update the first driving model of each vehicle terminal and re-acquire a new first model parameter difference until the edge aggregation number is equal to the edge aggregation number threshold, thereby obtaining a fourth driving model, and obtaining a second model parameter difference according to the fourth driving model and the first driving model; each of the edge servers uploads the second model parameter difference to the cloud server; The cloud server trains the first driving model according to the second model parameter differences uploaded by the plurality of edge servers to obtain a target vehicle driving model.

2. The driving model training method according to claim 1, characterized in that: Each of the edge servers trains the first driving model according to the first model parameter differences uploaded by the corresponding plurality of vehicle terminals to obtain a third driving model and the number of edge aggregations, including: Each of the edge servers obtains an optimal weight of each of the first model parameter differences according to a preset learning algorithm and a plurality of the first model parameter differences; Each of the edge servers obtains a first gradient descent coefficient according to each of the first model parameter differences and their optimal weights; Each of the edge servers trains the first driving model according to the first gradient descent coefficient and a preset first training step size to obtain the third driving model, and determines the number of times the third driving model is obtained as the edge aggregation number.

3. The driving model training method according to claim 1, characterized in that: The obtaining a second model parameter difference according to the fourth driving model and the first driving model includes: The edge server determines a difference between a model parameter of the fourth driving model and a model parameter of the first driving model as the second model parameter difference.

4. The driving model training method according to claim 1, characterized in that: The obtaining a first model parameter difference according to the second driving model and the first driving model includes: The vehicle end determines a difference between a model parameter of the second driving model and a model parameter of the first driving model as the first model parameter difference.

5. The driving model training method according to claim 1, characterized in that: The cloud server trains the first driving model according to the second model parameter differences uploaded by the plurality of edge servers to obtain a target vehicle driving model, including: The cloud server trains the first driving model according to the second model parameter difference uploaded by several edge servers to obtain a fifth driving model and a cloud aggregation number; if the cloud aggregation number is less than a preset cloud aggregation number threshold, the fifth driving model is re-sent to several edge servers to update the first driving model of each edge server and re-acquire a new second model parameter difference until the cloud aggregation number is equal to the cloud aggregation number threshold, thereby obtaining the target vehicle driving model.

6. The driving model training method according to claim 5, characterized in that: The cloud server trains the first driving model according to the second model parameter differences uploaded by the plurality of edge servers to obtain a fifth driving model and a cloud aggregation number, including: The cloud server obtains the optimal weight of each of the second model parameter differences according to a preset learning algorithm and a plurality of the second model parameter differences; The cloud server obtains a second gradient descent coefficient according to each of the second model parameter differences and their optimal weights; The cloud server trains the first driving model according to the second gradient descent coefficient and a preset second training step size to obtain the fifth driving model, and determines the number of times the fifth driving model is obtained as the cloud aggregation number.

7. A driving model training method, characterized in that: include: The cloud server sends the target vehicle driving model obtained according to the driving model training method according to any one of claims 1 to 6 to multiple vehicle terminals through several edge servers to improve the autonomous driving performance of each of the vehicle terminals.

8. A driving model training device, characterized in that: Applied to the driving model training system, including multiple vehicle terminals, several edge servers and cloud servers, the device includes: A driving model sending module, used to drive each of the edge servers to send the first driving model to a corresponding number of vehicle terminals; A vehicle-side training module, used to drive each of the vehicle-sides to train the first driving model according to its own driving data and a preset number of training times, to obtain a trained second driving model, and to obtain a first model parameter difference according to the second driving model and the first driving model; each of the vehicle-sides uploads the first model parameter difference to a corresponding edge server; The edge server training module is used to drive each of the edge servers to train the first driving model according to the first model parameter difference values ​​uploaded by the corresponding plurality of vehicle terminals, and obtain a third driving model and an edge aggregation number; if the edge aggregation number is less than a preset edge aggregation number threshold, the third driving model is re-sent to each of the vehicle terminals to update the first driving model of each vehicle terminal and re-acquire a new first model parameter difference value, until the edge aggregation number is equal to the edge aggregation number threshold, a fourth driving model is obtained, and a second model parameter difference value is obtained according to the fourth driving model and the first driving model; each of the edge servers uploads the second model parameter difference value to the cloud server; The cloud server training module is used to drive the cloud server to train the first driving model according to the second model parameter differences uploaded by several edge servers to obtain the target vehicle driving model.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the driving model training method according to any one of claims 1 to 7 are implemented.

10. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable by the processor, wherein the processor implements the steps of the driving model training method as claimed in any one of claims 1 to 7 when executing the computer program.

Images