Air-ground integrated multi-domain fused automobile high-level automatic driving system

Abstract

The invention discloses a heaven and earth integrated multi-domain fused automobile high-level automatic driving system, which realizes more comprehensive and non-blind area sensing and positioning of an automatic driving automobile through multi-domain information fusion of a satellite platform, a cloud platform, an automobile end platform and the like, and promotes comprehensive upgrading of comprehensive efficiency of an intelligent traffic system. Based on cloud error broadcasting and a high and low orbit satellite fusion enhancement technology, the problem of automobile communication loss caused by the fact that single ground communication is limited by base station coverage is solved; based on a cloud reinforcement learning architecture, the communication network stability of the space-ground integrated system is improved through adaptive switching of a communication network; a road semantic map and a decision control experience pool are established based on vehicle high-precision trajectory data collected by a platform, and the limitation of the perception ability of a bicycle is made up through cloud federal learning scene experience sharing; the vehicle end dynamically updates a cloud road semantic map through local map matching; and information support is provided for realizing a chassis multi-domain fusion technology based on high-precision positioning.

Description

Technical Field

[0001] The present invention belongs to the field of vehicle engineering and transportation engineering, and relates to a high-level automatic driving system for automobiles that integrates earth, sky and multiple domains. Background Art

[0002] With the rapid development of vehicle-to-everything (V2X) communication technology, information interconnection has been achieved between vehicles, roads, clouds, and satellites. How to effectively utilize the integrated communication network of vehicles, roads, satellites, and clouds to build a seamless information interaction system, improve the safety and comfort of high-level autonomous driving, and promote the comprehensive upgrade of the comprehensive performance of intelligent transportation systems has become a key issue.

[0003] Traditional single ground communication cannot meet the high real-time and high safety requirements of autonomous driving due to reliability problems such as weak communication signals and limited network coverage, which limits its widespread application in autonomous driving. For example, the autonomous driving system based on cellular network vehicle networking technology (CV2X) has data transmission delays or interruptions due to insufficient cellular network coverage in remote areas, urban edges, and building obstructions. It cannot cope with complex dynamic driving scenarios, which reduces the system's scenario adaptability and system reliability. At the same time, the autonomous driving system based on cellular network vehicle networking technology (CV2X) lacks multi-domain information integration capabilities, making it difficult to fully realize multi-domain information coordination from autonomous driving perception, decision-making to control processes. There is an information island problem caused by insufficient information sharing, and it is impossible to respond quickly and comprehensively to changes in driving scenarios. In summary, the existing system has limitations in scenario adaptability and multi-domain information integration, which restricts its further development.

[0004] The space-ground integrated autonomous driving system combines the ground communication network with the satellite communication network to achieve high-reliability seamless coverage of the communication network. Through satellite signals, it provides accurate time synchronization and reliable communication transmission between multiple domains, and provides high-precision positioning as the basic data support for the integration of multiple domains such as intelligent driving, chassis, cockpit, and power. It provides information support such as traffic, weather, and road conditions, and provides new solutions for autonomous driving in complex traffic scenarios. Summary of the invention

[0005] The present invention provides a high-level automatic driving system that integrates multiple domains of the sky and the earth. By integrating multi-domain information such as satellite platforms, cloud platforms, and vehicle-side platforms, and integrating different data sources, it realizes more comprehensive perception and positioning of automatic driving vehicles without blind spots, improves the safety and reliability of planning and control, and promotes the comprehensive upgrade of the comprehensive efficiency of intelligent transportation systems. Based on cloud-based error broadcasting and high- and low-orbit satellite fusion enhancement technology, the problem of vehicle communication loss caused by single ground communication limited by base station coverage is solved, and the high-precision positioning and high-security communication requirements required for autonomous driving are guaranteed; based on the cloud-based reinforcement learning architecture, the stability of the communication network of the space-ground integrated system is improved through adaptive switching of the communication network; based on the high-precision vehicle trajectory data collected by the platform, a road semantic map and a decision-making control experience pool are established, and through cloud-based federated learning scenario experience sharing, the limitations of single-vehicle perception capabilities are compensated, and redundant safety and long-tail problems that cannot be solved by single-vehicle intelligence are solved; the vehicle side realizes dynamic updating of the cloud-based road semantic map by matching the semantic map provided by the cloud with the local driving map constructed by the vehicle itself, thereby improving the cross-scenario integration capability of the entire autonomous driving system; based on high-precision positioning, the vehicle is provided with traffic, weather, road conditions and other information in the driving scene, supporting the realization of chassis multi-domain fusion technology.

[0006] The technical solution of the high-level automatic driving system with integrated space-ground multi-domain integration of the present invention includes three main contents and eight fields, among which the three contents are vehicle-side, cloud-side, and satellite-side, and the eight fields are intelligent cockpit field, intelligent driving field, powertrain field, intelligent chassis field, knowledge management field, cloud-side control field, satellite service field, and information communication field. Among them, the intelligent cockpit field, intelligent driving field, powertrain field, and intelligent chassis field are deployed on the vehicle-side, the cloud-side control field and knowledge management field are deployed on the cloud, the satellite service field is deployed on the high-orbit and low-orbit satellite-side, and the information communication field serves between the vehicle-side, cloud-side, and satellite-side.

[0007] For the satellite side, it is composed of a satellite service domain and related support platforms. Through satellite signals, accurate time synchronization and reliable communication transmission between multiple domains are provided, high-precision positioning is provided as the basic data support for multi-domain integration, and additional data support such as traffic information, weather changes, and road condition updates are provided. The satellite service domain provides high-precision positioning information through satellite communication. The related support platforms include but are not limited to a traffic management platform that monitors traffic information in real time and provides a complete traffic network, a logistics platform that provides real-time logistics network information, a map platform that provides navigation services and regularly updates offline map information, a positioning platform that provides high-precision positioning services and location information sharing functions, and a meteorological platform that provides real-time meteorological monitoring and analysis.

[0008] The cloud is mainly used for information transmission and further processing, and realizes centralized information management and knowledge extraction through different cloud control applications. The cloud consists of a cloud control domain and a knowledge management domain, and realizes efficient information exchange between the vehicle, the cloud, and the satellite through Vehicle-to-Network (V2N) communication combined with satellite communication.

[0009] The cloud control domain, based on the three cloud control levels of edge cloud, regional cloud, and central cloud, realizes cloud control applications with different functions through three cloud control empowerments: connected vehicle empowerment, traffic management and control, and traffic data empowerment. Cloud control applications include communication network monitoring applications, communication fusion enhancement applications, network switching strategies, high-precision bird's-eye views without blind spots, semantic bird's-eye view matrices, and driving safety field construction.

[0010] The communication network monitoring application is mainly used to continuously monitor the status of the communication network, including signal strength, delay, bandwidth, and packet loss, etc. When a network failure or performance degradation occurs, the communication network monitoring application can detect and issue an alarm in time to solve the problem through subsequent network switching applications.

[0011] The communication fusion enhancement application refers to sending a phase deviation representing the satellite clock error, satellite orbit error, and phase delay error to a high-orbit satellite through a satellite ground reference station Satellite Ground Reference Station (SGRS). The high-orbit satellite further calculates the satellite ephemeris error through the phase deviation data and sends it to the low-orbit satellite set corresponding to the high-orbit satellite. Finally, the low-orbit satellite returns the navigation enhancement information to the satellite ground base station Satellite Ground Station (SGS), and uses the navigation enhancement information to offset the orbit error from the satellite during the satellite positioning process.

[0012] The network switching strategy trains the switching strategy of the satellite network and the cellular network through deep reinforcement learning, and the deep reinforcement learning is described by a tuple (S, P, A, R), where:

[0013] S represents the input state of deep reinforcement learning:

[0014] S=[s cell ,s sat ,d cell ,d sat ,b cell ,b sat ,p lat ,p long ,c,h]

[0015] Among them, s cell Indicates the current cellular network signal strength, s satIndicates the current satellite network signal strength, d cell Indicates the current cellular network delay, d sat Indicates the current satellite network delay, b cell Indicates the current cellular network bandwidth, b sat Indicates the current satellite network bandwidth, p lat Indicates the current vehicle's latitude, p long represents the current longitude of the vehicle, c represents the current network connection status, c=0 represents the current connection to the cellular network, c=1 represents the current connection to the satellite network, and h represents the communication quality history in the recent period of time, which is represented by the communication network delay.

[0016] P represents the deep reinforcement learning state transfer equation, and the present invention uses a neural network for fitting.

[0017] A represents the output of the deep reinforcement learning communication network switching model, A=0 means maintaining the current communication network connection, A=1 means switching to the cellular network, and A=2 means switching to the satellite network.

[0018] R represents the reward function (environmental feedback) after executing the communication network switching strategy output:

[0019] R t =w1R stab +w2R cost +w3R quality

[0020] Among them, R t represents the environmental feedback at time t, w1, w2, and w3 represent the weighted weights of each sub-reward function, R stab represents the connection stability related reward function, R cost represents the switching cost-related reward function, R quality represents the reward function related to the communication service quality. Table 1 describes the values ​​of each sub-reward function in detail, where △d represents the delay difference after executing the switching strategy c, θ d represents the delay threshold, △b represents the bandwidth difference after executing switching strategy c, θ b Indicates the bandwidth threshold.

[0021] Table 1

[0022]

[0023]

[0024] The high-precision bird's-eye view without blind spots receives high-precision positioning provided by the satellite service domain deployed on the satellite through the Satellite Ground Station (SGS) responsible for receiving and sending satellite signals, and establishes a road semantic map based on the vehicle trajectory data collected by the platform under high-precision positioning. The high-precision bird's-eye view without blind spots contains 4 channels of information, namely the drivable area of ​​the road, lane boundaries, dynamic obstacles (vehicles, pedestrians), and traffic infrastructure (signals, traffic signs).

[0025] The semantic bird's-eye view matrix divides the high-precision bird's-eye view without blind spots into matrices with different semantics according to different channels. The size of the semantic bird's-eye view matrix without blind spots is [0,1] W×H×C , where W represents the width of the blind-spot-free semantic bird's-eye view matrix, H represents the height of the blind-spot-free semantic bird's-eye view matrix, and C represents the number of channels of the blind-spot-free semantic bird's-eye view matrix.

[0026] The driving safety field, based on a high-precision bird's-eye view without blind spots, constructs a dynamic driving safety field to guide vehicle driving behavior. The driving safety field is modeled using the Gaussian equation, and the driving risk of the intelligent connected vehicle in the environment is evaluated by describing the dynamic transfer process of the risk center of the driving safety field during the movement of the intelligent connected vehicle.

[0027] The knowledge management domain obtains the neural network parameters of the intelligent connected vehicles in the vehicle-side group through Vehicle-to-Network (V2N) communication, and screens the neural network parameters through the intelligent chassis feedback of the vehicle-side to perform local neural network Actor-Critic parameter aggregation. The knowledge management domain further trains the scene causal reasoning model in the cross-scene training sequence, refines the transferable scene factors in each scene, and establishes a scene experience library. When a new vehicle is connected to the system, the knowledge management domain pushes the beyond-visual-range road semantic map of the driving environment of the intelligent connected vehicle, and predicts the scene factors of the vehicle-side environment based on the scene causal reasoning model, and pushes parameter experience to the vehicle-side based on the neural network parameters corresponding to the scene factors in the scene experience library, so as to achieve rapid access to the vehicle-side.

[0028] For the vehicle side, it includes multiple vehicle-side groups, each of which includes N smart connected vehicles. Each smart connected vehicle includes a smart cockpit domain, a smart driving domain, a powertrain domain, and a smart chassis domain.

[0029] The smart cockpit domain is mainly used to improve driving safety, enhance user experience and provide personalized services, and includes information systems, interactive systems, and networking systems. The information system displays vehicle information and entertainment content in real time through a high-resolution display screen, allowing the driver to fully understand the vehicle status. The interactive system realizes comprehensive information interaction between the vehicle and the driver through an intuitive user interface and voice recognition. The networking system realizes seamless connection between the vehicle and the external network through the Internet of Vehicles technology.

[0030] The intelligent driving domain includes a perception processing module, a trajectory prediction module, and a reinforcement learning module. The perception processing module obtains the blind-spot-free semantic bird's-eye view provided by the cloud control domain through Vehicle-to-Network (V2N) communication, and cuts it according to the high-precision positioning provided by the satellite service domain. The size of the cropped blind-spot-free semantic bird's-eye view matrix is ​​[0,1] w×h×c , where w represents the width of the cropped blind-free semantic bird's-eye view matrix, h represents the height of the cropped blind-free semantic bird's-eye view matrix, and c represents the number of channels of the cropped blind-free semantic bird's-eye view matrix. The cropped blind-free semantic bird's-eye view is stacked with the sensor information as part of the input of the reinforcement learning module. At the same time, the perception processing module uses the vehicle-side sensor to build a real-time dynamic map and upload it, and realizes real-time local update of the offline map through dynamic matching based on high-precision positioning. The trajectory prediction module stacks the blind-free semantic bird's-eye view provided by the cloud control domain in multiple time steps, and combines the scene factors provided by the knowledge management domain to realize the multi-time-step trajectory prediction of dynamic obstacles around the intelligent networked vehicle as another part of the input of the reinforcement learning module. The reinforcement learning module stacks the input multi-time steps provided by the perception processing module, and integrates it with the multi-time-step trajectory prediction results of the surrounding dynamic obstacles provided by the trajectory prediction module as a complete reinforcement learning input. Further, the reinforcement learning module constructs a fusion reward function based on the driving safety field provided by the cloud control domain to guide the training process of the vehicle-side reinforcement learning neural network. Finally, the reinforcement learning module outputs control instructions to the powertrain domain.

[0031] The powertrain domain mainly outputs power control according to the control instructions of the intelligent driving domain. Through the synergy of the battery unit, motor unit and electronic control unit, the overall performance and efficiency of the powertrain domain are improved to ensure the power performance, energy management and safety of electric vehicles during driving. The battery unit provides a continuous and strong power supply through a high-energy-density battery pack to achieve long-range and high-efficiency of the vehicle. The motor unit provides fast-response acceleration performance and a sensitive driving experience through efficient motors and advanced rotor design. The electronic control unit achieves intelligent coordination between the motor and the battery through advanced control algorithms to optimize power output and energy management.

[0032] The intelligent chassis domain, mainly through the intelligent chassis subsystem, optimizes the driving comfort of the vehicle, and provides the traffic, weather, road conditions and other information of the driving scene for the car based on high-precision positioning, supporting the realization of chassis multi-domain fusion technology. The intelligent chassis subsystem includes but is not limited to the steering subsystem, (directyaw control) DYC subsystem, suspension subsystem and drive subsystem with state coupling and cost coupling. The steering subsystem and the DYC subsystem calculate the lateral control amount of the intelligent networked vehicle, the suspension subsystem calculates the vertical control amount of the intelligent networked vehicle, and the drive subsystem calculates the longitudinal control amount of the intelligent networked vehicle. Through information transmission, the intelligent chassis domain is connected to the three major fields of the intelligent cockpit domain, the intelligent driving domain, and the powertrain domain. First, through the interaction in the intelligent cockpit domain, the intelligent chassis domain is dynamically adjusted to realize the interaction between the intelligent cockpit domain and the intelligent chassis domain. Secondly, the present invention uses chassis feedback as part of the state quantity input of the intelligent driving domain to realize the interaction between the intelligent driving domain and the intelligent chassis domain. Finally, the steering subsystem and DYC subsystem of the intelligent chassis domain calculate the yaw torque of the intelligent connected vehicle based on the output of the powertrain domain and the preview error between the actual path and the expected path. The drive subsystem obtains the expected target speed and calculates the four-wheel drive or braking torque of the intelligent connected vehicle based on the output of the powertrain domain and the current speed of the intelligent connected vehicle, thereby realizing the interaction between the powertrain domain and the intelligent chassis domain.

[0033] The information communication domain serves between the vehicle side, the cloud side, and the satellite side. The information communication domain implements intersatellite communication, information monitoring, high-precision positioning, and time synchronization services on the satellite platform. The information communication domain implements information storage, data analysis, experience summary, and communication enhancement services on the cloud platform. The information communication domain implements scene recognition, information sharing, experience reception, and network switching services on the vehicle side platform.

[0034] The present invention provides a high-level autonomous driving system that integrates the earth and the sky and integrates multiple domains, and the implementation process includes the following steps:

[0035] Step 1: Build a satellite platform to provide accurate time synchronization and reliable communication transmission between multiple domains through satellite signals, provide high-precision positioning as the basic data support for multi-domain integration, and provide additional data support such as traffic information, weather changes, and road condition updates. Deploy the satellite service domain on the satellite platform, and the satellite service domain provides high-precision positioning information through satellite communication.

[0036] Step 2: Build a cloud platform and deploy the cloud control domain and knowledge management domain on the cloud platform. Centralized information management and knowledge extraction are achieved through different cloud control applications. The cloud consists of the cloud control domain and the knowledge management domain. Through Vehicle-to-Network (V2N) communication combined with satellite communication, efficient information exchange between the vehicle, the cloud, and the satellite is achieved.

[0037] Step 3: Build the vehicle-side platform and deploy the intelligent cockpit domain, intelligent driving domain, powertrain domain, and intelligent chassis domain on the vehicle-side platform. The vehicle-side platform includes multiple vehicle-side groups, each of which includes N intelligent connected vehicles, and each intelligent connected vehicle includes the intelligent cockpit domain, intelligent driving domain, powertrain domain, and intelligent chassis domain.

[0038] Step 4: Build the information and communication domain to serve the vehicle side, cloud side, and satellite side. The information and communication domain implements inter-satellite communication, information monitoring, high-precision positioning, and time synchronization services on the satellite platform. The information and communication domain implements information storage, data analysis, experience summary, and communication enhancement services on the cloud platform. The information and communication domain implements scene recognition, information sharing, experience reception, and network switching services on the vehicle side platform.

[0039] Step 5: The group of intelligent connected vehicles on the vehicle-side platform integrates the complete reinforcement learning state input based on the perception processing module and trajectory prediction module in the intelligent driving domain.

[0040] Step 6: The intelligent connected vehicle group on the vehicle-side platform further constructs a fusion reward function based on the driving safety field provided by the cloud control domain to guide the training process of the intelligent connected vehicle reinforcement learning model.

[0041] Step 7: The intelligent connected car on the vehicle platform uploads its own neural network Actor-Critic parameters to the knowledge management domain of the cloud platform during the reinforcement learning training process. In different scenarios where the intelligent connected car is located, the knowledge management domain selects the neural network Actor-Critic parameters involved in parameter aggregation according to the intelligent chassis domain, and aggregates the local neural network Actor-Critic parameters based on federated learning. In addition, the cross-scenario training sequence of the neural network of the intelligent connected car is saved in the knowledge management domain, and the scenario causal reasoning model is trained based on the cross-scenario sequence, the scenario factors corresponding to different scenarios are refined, and a scenario experience library is established.

[0042] Step 8: The knowledge management domain of the cloud platform sends the aggregated local neural network Actor-Critic parameters to the corresponding vehicle-side platform intelligent connected vehicles, and the cycle continues until the network converges. When a new intelligent connected vehicle on the vehicle-side platform is connected to the high-level autonomous driving system that integrates the earth and the sky and integrates multiple domains, the knowledge management domain of the cloud platform predicts the scene factors of the scene based on the observation sequence of the intelligent connected vehicle, and retrieves the corresponding scene experience from the scene experience library to push the experience based on federated learning, so as to realize the rapid access of intelligent connected vehicles in the system.

[0043] Preferably, in step 1, the relevant support platform includes but is not limited to a traffic management platform that monitors traffic information in real time and provides a complete traffic network, a logistics platform that provides real-time logistics network information, a map platform that provides navigation services and regularly updates offline map information, a positioning platform that provides high-precision positioning services and location information sharing functions, and a meteorological platform that provides real-time meteorological monitoring and analysis.

[0044] Preferably, in step 2, the cloud control domain, based on the three cloud control levels of edge cloud, regional cloud, and central cloud, realizes cloud control applications with different functions through three cloud control empowerments, namely, connected vehicle empowerment, traffic management and control, and traffic data empowerment. Cloud control applications include communication network monitoring applications, communication fusion enhancement applications, network switching strategies, high-precision bird's-eye views without blind spots, semantic bird's-eye view matrices, and driving safety field construction.

[0045] Preferably, the communication network monitoring application is mainly used to continuously monitor the status of the communication network, including signal strength, delay, bandwidth, and packet loss, etc. When a network failure or performance degradation occurs, the communication network monitoring application can promptly detect and issue an alarm to solve the problem through subsequent network switching applications.

[0046] Preferably, the communication fusion enhancement application refers to sending a phase deviation representing the satellite clock error, satellite orbit error, and phase delay error to a high-orbit satellite through a satellite ground reference station Satellite Ground Reference Station (SGRS). The high-orbit satellite further calculates the satellite ephemeris error through the phase deviation data and sends it to the low-orbit satellite set corresponding to the high-orbit satellite. Finally, the low-orbit satellite returns the navigation enhancement information to the satellite ground base station Satellite Ground Station (SGS), and uses the navigation enhancement information to offset the orbit error from the satellite during the satellite positioning process.

[0047] Preferably, the network switching strategy trains the switching strategy of the satellite network and the cellular network through deep reinforcement learning, and the deep reinforcement learning is described by a tuple (S, P, A, R), where:

[0048] S represents the input state of deep reinforcement learning:

[0049] S=[s cell ,s sat ,d cell ,d sat ,b cell ,b sat ,p lat ,p long ,c,h]

[0050] Among them, s cell Indicates the current cellular network signal strength, s sat Indicates the current satellite network signal strength, d cell Indicates the current cellular network delay, d sat Indicates the current satellite network delay, b cell Indicates the current cellular network bandwidth, b sat Indicates the current satellite network bandwidth, p lat Indicates the current vehicle's latitude, p long represents the current longitude of the vehicle, c represents the current network connection status, c=0 represents the current connection to the cellular network, c=1 represents the current connection to the satellite network, and h represents the communication quality history in the recent period of time, which is represented by the communication network delay.

[0051] P represents the deep reinforcement learning state transfer equation, and the present invention uses a neural network for fitting.

[0052] A represents the output of the deep reinforcement learning communication network switching model, A=0 means maintaining the current communication network connection, A=1 means switching to the cellular network, and A=2 means switching to the satellite network.

[0053] R represents the reward function (environmental feedback) after executing the communication network switching strategy output:

[0054] R t =w1R stab +w2R cost +w3R quality

[0055] Among them, R t represents the environmental feedback at time t, w1, w2, and w3 represent the weighted weights of each sub-reward function, R stab represents the connection stability related reward function, R cost represents the switching cost-related reward function, R quality represents the reward function related to the communication service quality. Table 1 describes the values ​​of each sub-reward function in detail, where △d represents the delay difference after executing the switching strategy c, θ d represents the delay threshold, △b represents the bandwidth difference after executing switching strategy c, θb Indicates the bandwidth threshold.

[0056] Table 1

[0057]

[0058] Preferably, the high-precision bird's-eye view without blind spots receives high-precision positioning provided by the satellite service domain deployed on the satellite through the Satellite Ground Station (SGS) responsible for receiving and sending satellite signals, and establishes a road semantic map based on the vehicle trajectory data under high-precision positioning collected by the platform. The high-precision bird's-eye view without blind spots contains 4 channels of information, namely, the drivable area of ​​the road, lane boundaries, dynamic obstacles (vehicles, pedestrians), and traffic infrastructure (signals, traffic signs).

[0059] Preferably, the semantic bird's-eye view matrix divides the high-precision bird's-eye view without blind spots into matrices with different semantics according to different channels. The size of the semantic bird's-eye view matrix without blind spots is [0,1] W×H×C , where W represents the width of the blind-spot-free semantic bird's-eye view matrix, H represents the height of the blind-spot-free semantic bird's-eye view matrix, and C represents the number of channels of the blind-spot-free semantic bird's-eye view matrix.

[0060] Preferably, the driving safety field is based on a high-precision bird's-eye view without blind spots to construct a dynamic driving safety field to guide vehicle driving behavior. The driving safety field is modeled using the Gaussian equation, and the driving risk of the intelligent connected vehicle in the environment is evaluated by describing the dynamic transfer process of the driving safety field risk center during the movement of the intelligent connected vehicle.

[0061] Preferably, the knowledge management domain obtains the neural network parameters of the intelligent connected vehicles in the vehicle-side group through Vehicle-to-Network (V2N) vehicle-cloud communication, and screens the neural network parameters through the intelligent chassis feedback of the vehicle-side to perform local neural network Actor-Critic parameter aggregation. The knowledge management domain further trains the scene causal reasoning model in the cross-scene training sequence, refines the transferable scene factors under each scene, and establishes a scene experience library. When a new vehicle is connected to the system, the knowledge management domain pushes the beyond-visual-range road semantic map of the driving environment of the intelligent connected vehicle, and predicts the scene factors of the vehicle-side environment based on the scene causal reasoning model, and pushes parameter experience to the vehicle-side based on the neural network parameters corresponding to the scene factors in the scene experience library, so as to achieve rapid access to the vehicle-side.

[0062] Preferably, in step 3, the smart cockpit domain is mainly used to improve driving safety, enhance user experience and provide personalized services, and includes an information system, an interactive system, and a network system. The information system displays vehicle information and entertainment content in real time through a high-resolution display screen, so that the driver can fully understand the vehicle status. The interactive system realizes comprehensive information interaction between the vehicle and the driver through an intuitive user interface and voice recognition. The network system realizes seamless connection between the vehicle and the external network through the Internet of Vehicles technology.

[0063] Preferably, in step 3, the intelligent driving domain includes a perception processing module, a trajectory prediction module, and a reinforcement learning module. The perception processing module obtains the blind-spot-free semantic bird's-eye view provided by the cloud control domain through Vehicle-to-Network (V2N) communication, and cuts it according to the high-precision positioning provided by the satellite service domain. The trajectory prediction module stacks the blind-spot-free semantic bird's-eye view provided by the cloud control domain in multiple time steps, and combines the scene factors provided by the knowledge management domain to realize the multi-time-step trajectory prediction of the dynamic obstacles around the intelligent networked vehicle as another part of the input of the reinforcement learning module. The reinforcement learning module stacks the input provided by the perception processing module in multiple time steps, and integrates it with the multi-time-step trajectory prediction results of the surrounding dynamic obstacles provided by the trajectory prediction module into a complete reinforcement learning input. Furthermore, the reinforcement learning module constructs a fusion reward function based on the driving safety field provided by the cloud control domain to guide the training process of the vehicle-side reinforcement learning neural network. Finally, the reinforcement learning module outputs control instructions to the powertrain domain.

[0064] Preferably, in step 3, the powertrain domain mainly outputs power control according to the control instructions of the intelligent driving domain, and improves the overall performance and efficiency of the powertrain domain through the synergy of the battery unit, motor unit and electronic control unit, ensuring the power performance, energy management and safety of the electric vehicle during driving. The battery unit provides a continuous and strong power supply through a high-energy-density battery pack to achieve long-range and high-efficiency of the vehicle. The motor unit provides fast-response acceleration performance and a sensitive driving experience through efficient motors and advanced rotor design. The electronic control unit achieves intelligent coordination between the motor and the battery through advanced control algorithms to optimize power output and energy management.

[0065] Preferably, in step 3, the intelligent chassis domain, mainly through the intelligent chassis subsystem, optimizes the driving comfort of the vehicle, and provides the vehicle with traffic, weather, road conditions and other information of the driving scene based on high-precision positioning, supporting the realization of chassis multi-domain fusion technology. The intelligent chassis subsystem includes but is not limited to the steering subsystem, (direct yaw control) DYC subsystem, suspension subsystem and drive subsystem with state coupling and cost coupling. The steering subsystem and the DYC subsystem calculate the lateral control amount of the intelligent networked vehicle, the suspension subsystem calculates the vertical control amount of the intelligent networked vehicle, and the drive subsystem calculates the longitudinal control amount of the intelligent networked vehicle. Through information transmission, the intelligent chassis domain is connected to the three major fields of the intelligent cockpit domain, the intelligent driving domain, and the powertrain domain. First, through the interaction in the intelligent cockpit domain, the intelligent chassis domain is dynamically adjusted to realize the interaction between the intelligent cockpit domain and the intelligent chassis domain. Secondly, the present invention uses chassis feedback as part of the state quantity input of the intelligent driving domain to realize the interaction between the intelligent driving domain and the intelligent chassis domain. Finally, the steering subsystem and DYC subsystem of the intelligent chassis domain calculate the yaw torque of the intelligent connected vehicle based on the output of the powertrain domain and the preview error between the actual path and the expected path. The drive subsystem obtains the expected target speed and calculates the four-wheel drive\braking torque of the intelligent connected vehicle based on the output of the powertrain domain and the current speed of the intelligent connected vehicle, thereby realizing the interaction between the powertrain domain and the intelligent chassis domain.

[0066] Preferably, in step 5, the integrated complete reinforcement learning state input includes two parts. In the first part, the perception processing module obtains the blind-spot-free semantic bird's-eye view provided by the cloud control domain through Vehicle-to-Network (V2N) communication, and crops it according to the high-precision positioning provided by the satellite service domain. The cropped blind-spot-free semantic bird's-eye view matrix size is [0,1] w×h×c , where w represents the width of the cropped blind-spot-free semantic bird's-eye view matrix, h represents the height of the cropped blind-spot-free semantic bird's-eye view matrix, and c represents the number of channels of the cropped blind-spot-free semantic bird's-eye view matrix. The cropped blind-spot-free semantic bird's-eye view is stacked with the sensor information as part of the input of the reinforcement learning module. In the second part, the trajectory prediction module stacks the blind-spot-free semantic bird's-eye view provided by the cloud control domain in multiple time steps, and combines the scenario factors provided by the knowledge management domain to realize the multi-time-step trajectory prediction of the dynamic obstacles around the intelligent connected vehicle as another part of the input of the reinforcement learning module. The reinforcement learning module stacks the input provided by the perception processing module in multiple time steps, and integrates it with the multi-time-step trajectory prediction results of the surrounding dynamic obstacles provided by the trajectory prediction module into a complete reinforcement learning input.

[0067] Preferably, in step 6, the fusion reward function is considered from two perspectives: driving safety and driving expectation, wherein the driving safety-related reward function is constructed based on the driving risk of the intelligent connected vehicle predicted by the driving safety location provided by the cloud control domain. The driving expectation reward function is constructed based on the lateral error and heading angle deviation between the actual path of the intelligent connected vehicle and the expected path.

[0068] Beneficial effects of the present invention:

[0069] (1) The present invention provides a high-level autonomous driving system that integrates multiple domains and the sky. By integrating information from multiple domains such as satellite platforms, cloud platforms, and vehicle-side platforms, and integrating different data sources, the system achieves more comprehensive perception and positioning of autonomous vehicles without blind spots, improves the safety and reliability of planning and control, and promotes the comprehensive upgrade of the comprehensive efficiency of intelligent transportation systems.

[0070] (2) Based on cloud error broadcasting and high-orbit and low-orbit satellite fusion enhancement technology, the problem of vehicle communication loss caused by single ground communication being limited by base station coverage is solved, ensuring the high-precision positioning and high-security communication requirements required for autonomous driving; based on the cloud reinforcement learning architecture, the communication network stability of the space-ground integrated system is improved through adaptive switching of the communication network;

[0071] (3) Based on the high-precision vehicle trajectory data collected by the platform, a road semantic map and a decision-making control experience pool are established. Through cloud-based federated learning scenario experience sharing, the limitations of single-vehicle perception capabilities are compensated, and redundant safety and long-tail problems that single-vehicle intelligence cannot solve are solved.

[0072] (4) The vehicle side dynamically updates the semantic map of the road in the cloud by matching the semantic map provided by the cloud with the local driving map constructed by the vehicle itself. High-precision positioning can provide the car with information such as traffic, weather, and road conditions in the driving scene, improve the cross-scenario integration capability of the entire autonomous driving system, and support the realization of chassis multi-domain fusion technology, which is of great value to the implementation of high-level autonomous driving technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 Schematic diagram of the high-level autonomous driving system with integrated space-ground multi-domain integration proposed by the present invention;

[0074] Figure 2 A communication diagram of a high-level autonomous driving system with integrated space-ground multi-domain integration proposed by the present invention;

[0075] Figure 3 Schematic diagram of the neural network structure of the network switching strategy proposed by the present invention;

[0076] Figure 4 Schematic diagram of a semantic bird's-eye view without blind spots used in the present invention. DETAILED DESCRIPTION

[0077] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings, but the content of the present invention is not limited thereto.

[0078] The present invention provides a high-level autonomous driving system that integrates the earth and the sky and integrates multiple domains. Figure 1 As shown, high-level autonomous driving based on the integration of space and earth and multi-domains can be realized in complex driving scenarios. The communication diagram of the high-level autonomous driving system based on the integration of space and earth and multi-domains is shown in Figure 2 As shown, the specific steps include:

[0079] (1) Build a satellite platform to provide accurate time synchronization and reliable communication transmission between multiple domains through satellite signals, provide high-precision positioning as the basic data support for multi-domain integration, and provide additional data support such as traffic information, weather changes, and road condition updates. Deploy the satellite service domain on the satellite platform, and the satellite service domain provides high-precision positioning information through satellite communication. The relevant support platforms include but are not limited to traffic management platforms that monitor traffic information in real time and provide a complete traffic network, logistics platforms that provide real-time logistics network information, map platforms that provide navigation services and regularly update offline map information, positioning platforms that provide high-precision positioning services and location information sharing functions, and meteorological platforms that provide real-time meteorological monitoring and analysis.

[0080] (2) Build a cloud platform and deploy the cloud control domain and knowledge management domain on the cloud platform. Centralized information management and knowledge extraction are achieved through different cloud control applications. The cloud consists of a cloud control domain and a knowledge management domain. Vehicle-to-Network (V2N) communication combined with satellite communication enables efficient information exchange between the vehicle, the cloud, and the satellite. The cloud control domain, based on the three cloud control levels of edge cloud, regional cloud, and central cloud, implements cloud control applications with different functions through three cloud control empowerments: connected vehicle empowerment, traffic management and control, and traffic data empowerment. Cloud control applications include communication network monitoring applications, communication fusion enhancement applications, network switching strategies, high-precision bird's-eye views without blind spots, semantic bird's-eye view matrices, and driving safety field construction.

[0081] Communication network monitoring applications are mainly used to continuously monitor the status of communication networks, including signal strength, latency, bandwidth, and packet loss. When a network fails or performance degrades, the communication network monitoring application can detect and issue an alarm in time to solve the problem through subsequent network switching applications.

[0082] Communication fusion enhancement application refers to sending phase deviations representing satellite clock errors, satellite orbit errors, and phase delay errors to high-orbit satellites through the Satellite Ground Reference Station (SGRS). The high-orbit satellite further calculates the satellite ephemeris error through the phase deviation data and sends it to the low-orbit satellite set corresponding to the high-orbit satellite. Finally, the low-orbit satellite returns the navigation enhancement information to the Satellite Ground Station (SGS), and uses the navigation enhancement information to offset the orbit error from the satellite during the satellite positioning process.

[0083] Network switching strategy, the network switching strategy neural network structure used is as follows Figure 3 As shown in Figure 1, the switching strategy between satellite network and cellular network is trained by deep reinforcement learning. Deep reinforcement learning is described by a tuple (S, P, A, R), where:

[0084] S represents the input state of deep reinforcement learning:

[0085] S=[s cell ,s sat ,d cell ,d sat ,b cell ,b sat ,p lat ,p long ,c,h]

[0086] Among them, s cell Indicates the current cellular network signal strength, s sat Indicates the current satellite network signal strength, d cell Indicates the current cellular network delay, d sat Indicates the current satellite network delay, b cell Indicates the current cellular network bandwidth, b sat Indicates the current satellite network bandwidth, p lat Indicates the current vehicle's latitude, p long represents the current longitude of the vehicle, c represents the current network connection status, c=0 represents the current connection to the cellular network, c=1 represents the current connection to the satellite network, and h represents the communication quality history in the recent period of time, which is represented by the communication network delay.

[0087] P represents the deep reinforcement learning state transfer equation, and the present invention uses a neural network for fitting.

[0088] A represents the output of the deep reinforcement learning communication network switching model, A=0 means maintaining the current communication network connection, A=1 means switching to the cellular network, and A=2 means switching to the satellite network.

[0089] R represents the reward function (environmental feedback) after executing the communication network switching strategy output:

[0090] R t =w1R stab +w2R cost +w3R quality

[0091] Among them, R t represents the environmental feedback at time t, w1, w2, and w3 represent the weighted weights of each sub-reward function, R stab represents the connection stability related reward function, R cost represents the switching cost-related reward function, R quality represents the reward function related to the communication service quality. Table 1 describes the values ​​of each sub-reward function in detail, where △d represents the delay difference after executing the switching strategy c, θ d represents the delay threshold, △b represents the bandwidth difference after executing switching strategy c, θ b Indicates the bandwidth threshold.

[0092] Table 1

[0093]

[0094] High-precision bird's-eye view without blind spots, such as Figure 4 As shown, the high-precision positioning provided by the satellite service domain deployed on the satellite is received through the Satellite Ground Station (SGS) responsible for receiving and sending satellite signals, and the road semantic map is established based on the vehicle trajectory data collected by the platform under high-precision positioning. The high-precision bird's-eye view map without blind spots contains 4 channels of information, namely the drivable area of ​​the road, lane boundaries, dynamic obstacles (vehicles, pedestrians), and traffic infrastructure (signals, traffic signs).

[0095] Semantic bird's-eye view matrix, divides the high-precision bird's-eye view without blind spots into matrices with different semantics according to different channels. The size of the semantic bird's-eye view matrix without blind spots is [0,1] W×H×C , where W represents the width of the blind-spot-free semantic bird's-eye view matrix, H represents the height of the blind-spot-free semantic bird's-eye view matrix, and C represents the number of channels of the blind-spot-free semantic bird's-eye view matrix.

[0096] Driving safety field, based on high-precision bird's-eye view without blind spots, constructs a dynamic driving safety field to guide vehicle driving behavior. The driving safety field is modeled using Gaussian equations, and by describing the dynamic transfer process of the driving safety field risk center in the movement of intelligent connected vehicles, the driving risk of intelligent connected vehicles in the environment is evaluated.

[0097] The knowledge management domain obtains the neural network parameters of the intelligent connected vehicles in the vehicle-side group through Vehicle-to-Network (V2N) vehicle-cloud communication, and screens the neural network parameters through the intelligent chassis feedback of the vehicle-side to aggregate the local neural network Actor-Critic parameters. The knowledge management domain further trains the scene causal reasoning model in the cross-scene training sequence, refines the transferable scene factors in each scene, and establishes a scene experience library. When a new vehicle is connected to the system, the knowledge management domain pushes the beyond-visual-range road semantic map of the driving environment of the intelligent connected vehicle, and predicts the scene factors of the vehicle-side environment based on the scene causal reasoning model, and pushes parameter experience to the vehicle-side based on the neural network parameters corresponding to the scene factors in the scene experience library, so as to achieve rapid access to the vehicle-side.

[0098] (3) Build a vehicle-side platform and deploy the intelligent cockpit domain, intelligent driving domain, powertrain domain, and intelligent chassis domain on the vehicle-side platform. The vehicle-side platform includes multiple vehicle-side groups, each of which includes N intelligent connected vehicles, and each intelligent connected vehicle includes an intelligent cockpit domain, an intelligent driving domain, a powertrain domain, and an intelligent chassis domain.

[0099] The smart cockpit domain is mainly used to improve driving safety, enhance user experience and provide personalized services, including information systems, interactive systems, and network systems. The information system displays vehicle information and entertainment content in real time through a high-resolution display screen, allowing the driver to fully understand the vehicle status. The interactive system realizes comprehensive information interaction between the vehicle and the driver through an intuitive user interface and voice recognition. The network system realizes seamless connection between the vehicle and the external network through the Internet of Vehicles technology.

[0100] The intelligent driving domain includes a perception processing module, a trajectory prediction module, and a reinforcement learning module. The perception processing module obtains the blind-spot-free semantic bird's-eye view provided by the cloud control domain through Vehicle-to-Network (V2N) communication, and cuts it according to the high-precision positioning provided by the satellite service domain. At the same time, the perception processing module uses the vehicle-side sensor to build a real-time dynamic map and upload it, and realizes real-time local update of the offline map based on dynamic matching that relies on high-precision positioning. The trajectory prediction module stacks the blind-spot-free semantic bird's-eye view provided by the cloud control domain in multiple time steps, and combines the scene factors provided by the knowledge management domain to realize the multi-time-step trajectory prediction of dynamic obstacles around the intelligent networked vehicle as another part of the input of the reinforcement learning module. The reinforcement learning module stacks the input provided by the perception processing module in multiple time steps, and integrates it with the multi-time-step trajectory prediction results of the surrounding dynamic obstacles provided by the trajectory prediction module as a complete reinforcement learning input. Furthermore, the reinforcement learning module constructs a fusion reward function based on the driving safety field provided by the cloud control domain to guide the training process of the vehicle-side reinforcement learning neural network. Finally, the reinforcement learning module outputs control instructions to the powertrain domain.

[0101] The powertrain domain mainly outputs power control according to the control instructions of the intelligent driving domain. Through the synergy of the battery unit, motor unit and electronic control unit, the overall performance and efficiency of the powertrain domain are improved to ensure the power performance, energy management and safety of electric vehicles during driving. The battery unit provides a continuous and strong power supply through a high-energy-density battery pack to achieve long-range and high-efficiency of the vehicle. The motor unit provides fast-response acceleration performance and a sensitive driving experience through efficient motors and advanced rotor design. The electronic control unit uses advanced control algorithms to achieve intelligent coordination between the motor and the battery to optimize power output and energy management.

[0102] The intelligent chassis domain mainly optimizes the driving comfort of the vehicle through the intelligent chassis subsystem, and provides the car with traffic, weather, road conditions and other information of the driving scene based on high-precision positioning, supporting the realization of chassis multi-domain fusion technology. The intelligent chassis subsystem includes but is not limited to the steering subsystem, (directyaw control) DYC subsystem, suspension subsystem and drive subsystem with state coupling and cost coupling. The steering subsystem and the DYC subsystem calculate the lateral control amount of the intelligent networked vehicle, the suspension subsystem calculates the vertical control amount of the intelligent networked vehicle, and the drive subsystem calculates the longitudinal control amount of the intelligent networked vehicle. Through information transmission, the intelligent chassis domain is connected to the three major fields of the intelligent cockpit domain, the intelligent driving domain, and the powertrain domain. First, through the interaction in the intelligent cockpit domain, the intelligent chassis domain is dynamically adjusted to realize the interaction between the intelligent cockpit domain and the intelligent chassis domain. Secondly, the present invention uses chassis feedback as part of the state quantity input of the intelligent driving domain to realize the interaction between the intelligent driving domain and the intelligent chassis domain. Finally, the steering subsystem and DYC subsystem of the intelligent chassis domain calculate the yaw torque of the intelligent connected vehicle based on the output of the powertrain domain and the preview error between the actual path and the expected path. The drive subsystem obtains the expected target speed and calculates the four-wheel drive\braking torque of the intelligent connected vehicle based on the output of the powertrain domain and the current speed of the intelligent connected vehicle, thereby realizing the interaction between the powertrain domain and the intelligent chassis domain.

[0103] (4) The group of intelligent connected vehicles on the vehicle-side platform integrates the complete reinforcement learning state input based on the perception processing module and trajectory prediction module in the intelligent driving domain.

[0104] (5) The intelligent connected vehicle group on the vehicle-side platform further constructs a fusion reward function based on the driving safety scene provided by the cloud control domain to guide the training process of the intelligent connected vehicle reinforcement learning model. The fusion reward function considers driving safety and driving expectations. The driving safety-related reward function is constructed based on the driving risk of the intelligent connected vehicle predicted by the driving safety scene provided by the cloud control domain. The driving expectation reward function is constructed based on the lateral error and heading angle deviation between the actual path of the intelligent connected vehicle and the expected path.

[0105] (6) During the reinforcement learning training process, the intelligent connected car on the vehicle platform uploads the parameters of its own neural network Actor-Critic to the knowledge management domain of the cloud platform. In different scenarios where the intelligent connected car is located, the knowledge management domain selects the neural network Actor-Critic parameters involved in parameter aggregation according to the intelligent chassis domain, and aggregates the local neural network Actor-Critic parameters based on federated learning. The cross-scenario training sequence of the neural network of the intelligent connected car is saved in the knowledge management domain, and the scenario causal reasoning model is trained based on the cross-scenario sequence to extract the scenario factors corresponding to different scenarios and establish a scenario experience library.

[0106] (7) The knowledge management domain of the cloud platform sends the aggregated local neural network Actor-Critic parameters to the corresponding vehicle-side platform intelligent connected vehicles, and the cycle continues until the network converges. When a new intelligent connected vehicle on the vehicle-side platform is connected to the high-level autonomous driving system that integrates the earth and the sky and integrates multiple domains, the knowledge management domain of the cloud platform predicts the scene factors of the scene based on the observation sequence of the intelligent connected vehicle, and retrieves the corresponding scene experience from the scene experience library to push the experience based on federated learning, so as to realize the rapid access of the intelligent connected vehicle to the system.

[0107] In summary, the present invention proposes a high-level automatic driving system for vehicles with integrated earth-ground multi-domains. Through the fusion of multi-domain information such as satellite platforms, cloud platforms, and vehicle-side platforms, different data sources are integrated to achieve more comprehensive and blind-spot-free perception and positioning of automatic driving vehicles, improve the safety and reliability of planning and control, and promote the comprehensive upgrade of the comprehensive efficiency of intelligent transportation systems. Based on cloud error broadcasting and high-orbit and low-orbit satellite fusion enhancement technology, the problem of automobile communication loss caused by single ground communication limited by base station coverage is solved, and the high-precision positioning and high-security communication requirements required for automatic driving are guaranteed; based on the cloud reinforcement learning architecture, the communication network stability of the earth-ground integrated system is improved through adaptive switching of the communication network; based on the high-precision vehicle trajectory data collected by the platform, a road semantic map and a decision-making control experience pool are established, and through the cloud federated learning scene experience sharing, the limitations of the perception ability of a single vehicle are compensated, and the redundant safety and long-tail problems that cannot be solved by single-vehicle intelligence are solved; the vehicle side realizes the dynamic update of the cloud road semantic map by matching the semantic map provided by the cloud with the local driving map constructed by the vehicle itself. High-precision positioning can provide cars with information such as traffic, weather, and road conditions in driving scenarios, improve the cross-scenario integration capabilities of the entire autonomous driving system, and support the implementation of chassis multi-domain fusion technology, which is of great value to the implementation of high-level autonomous driving technology.

[0108] The series of detailed descriptions listed above are only specific descriptions of feasible implementation methods of the present invention. They are not intended to limit the scope of protection of the present invention. All equivalent methods or changes that do not deviate from the technical creation of the present invention should be included in the scope of protection of the present invention.

Claims

1. A high-level autonomous driving system for cars that integrates the earth and the sky and integrates multiple domains, characterized by: It includes three platforms and eight domains; the three platforms are the vehicle-side platform, the cloud platform and the satellite platform; the eight domains are the smart cockpit domain, the smart driving domain, the powertrain domain, the smart chassis domain, the knowledge management domain, the cloud control domain, the satellite service domain and the information communication domain; The vehicle-side platform includes multiple vehicle-side groups, each of which includes N intelligent networked vehicles, and each intelligent networked vehicle deploys an intelligent cockpit domain, an intelligent driving domain, a powertrain domain, and an intelligent chassis domain; the cloud platform deploys a cloud control domain and a knowledge management domain; The satellite platform deploys a satellite service domain; the information communication domain realizes information interconnection between the vehicle-side platform, the cloud platform, and the satellite platform; The satellite platform: based on the deployed satellite service domain, provides accurate time synchronization and reliable communication transmission between multiple domains through satellite signals, provides high-precision positioning as the basic data support for multi-domain fusion, and provides traffic information, weather changes, and road condition update data support; The cloud platform: based on the deployed cloud control domain and knowledge management domain, realizes information transmission and processing, and realizes centralized information management and knowledge extraction through different cloud control applications; The vehicle-side platform: based on the deployed smart cockpit domain, smart driving domain, powertrain domain, and smart chassis domain, can realize dynamic update of the cloud road semantic map, realize status interaction with the driver, realize trajectory prediction, realize motor control and energy management, and realize vehicle driving comfort.

2. The high-level automatic driving system for vehicles integrating earth, sky and multiple domains according to claim 1 is characterized in that: The satellite platform includes a satellite service domain and related supporting platforms; the satellite service domain provides high-precision positioning information through satellite communication; the related supporting platforms include a traffic management platform that monitors traffic information in real time and provides a complete traffic network, a logistics platform that provides real-time logistics network information, a map platform that provides navigation services and regularly updates offline map information, a positioning platform that provides high-precision positioning services and location information sharing functions, and a meteorological platform that provides real-time meteorological monitoring and analysis.

3. The high-level automatic driving system for vehicles integrating earth, sky and multiple domains according to claim 1 is characterized in that: The cloud platform includes a cloud control domain and a knowledge management domain, and realizes efficient information exchange between the vehicle side, the cloud side, and the satellite side through Vehicle-to-Network (V2N) communication combined with satellite communication; The cloud control domain, based on three cloud control levels of edge cloud, regional cloud, and central cloud, realizes cloud control applications with different functions through three cloud control empowerments: connected vehicle empowerment, traffic management and control, and traffic data empowerment; The knowledge management domain obtains the neural network parameters of the intelligent connected vehicles in the vehicle-side group through Vehicle-to-Network (V2N) communication, and screens the neural network parameters through the intelligent chassis feedback of the vehicle-side to perform local neural network Actor-Critic parameter aggregation; the knowledge management domain trains the scene causal reasoning model in the cross-scene training sequence, refines the transferable scene factors under each scene, and establishes a scene experience library. When a new vehicle is connected to the system, the knowledge management domain pushes the beyond-visual-range road semantic map of the driving environment of the intelligent connected vehicle, and predicts the scene factors of the vehicle-side environment based on the scene causal reasoning model, and pushes parameter experience to the vehicle-side based on the neural network parameters corresponding to the scene factors in the scene experience library, so as to realize rapid access to the vehicle-side.

4. The high-level automatic driving system for vehicles with integrated space-ground multi-domain integration according to claim 3 is characterized in that: The cloud control application includes communication network monitoring application, communication fusion enhancement application, providing network switching strategy, providing high-precision bird's-eye view without blind spots, providing semantic bird's-eye view matrix, and building a driving safety field; The communication network monitoring application is used to continuously monitor the status of the communication network, including signal strength, latency, bandwidth and packet loss. When a network failure or performance degradation occurs, the communication network monitoring application can promptly detect and issue an alarm so as to solve the problem through a subsequent network switching application; The communication fusion enhancement application sends a phase deviation representing a satellite clock error, a satellite orbit error, and a phase delay error to a high-orbit satellite through a satellite ground reference station. The high-orbit satellite calculates a satellite ephemeris error through the phase deviation data and sends it to a low-orbit satellite set corresponding to the high-orbit satellite. Finally, the low-orbit satellite returns the navigation enhancement information to the satellite ground base station, and the orbit error from the satellite during the satellite positioning process is offset by the navigation enhancement information. The network switching strategy trains the switching strategy of the satellite network and the cellular network through deep reinforcement learning; The high-precision bird's-eye view without blind spots receives high-precision positioning provided by the satellite service domain deployed on the satellite platform through a satellite ground base station responsible for receiving and sending satellite signals, and establishes a road semantic map based on the vehicle trajectory data under high-precision positioning collected by the platform; the high-precision bird's-eye view without blind spots contains 4 channels of information, namely, the drivable area of ​​the road, lane boundaries, dynamic obstacles, and traffic infrastructure; The semantic bird's-eye view matrix divides the high-precision bird's-eye view without blind spots into matrices with different semantics according to different channels. The size of the semantic bird's-eye view matrix without blind spots is [0,1] W×H×C , where W represents the width of the blind-spot semantic bird's-eye view matrix, H represents the height of the blind-spot semantic bird's-eye view matrix, and C represents the number of channels of the blind-spot semantic bird's-eye view matrix; The driving safety field, based on a high-precision bird's-eye view without blind spots, constructs a dynamic driving safety field to guide vehicle driving behavior. The driving safety field is modeled using the Gaussian equation. By describing the dynamic transfer process of the risk center of the driving safety field during the movement of the intelligent connected vehicle, the driving risk of the intelligent connected vehicle in the environment in which it is located is evaluated.

5. The high-level automatic driving system for vehicles integrating earth, sky and multiple domains according to claim 4 is characterized in that: The deep reinforcement learning in the network switching strategy is described by a tuple (S, P, A, R), where: S represents the input state of deep reinforcement learning: S=[s cell ,s sat ,d cell ,d sat ,b cell ,b sat ,p lat ,p long ,c,h] Among them, s cell Indicates the current cellular network signal strength, s sat Indicates the current satellite network signal strength, d cell Indicates the current cellular network delay, d sat Indicates the current satellite network delay, b cell Indicates the current cellular network bandwidth, b sat Indicates the current satellite network bandwidth, p lat Indicates the current vehicle's latitude, p long represents the current longitude of the vehicle, c represents the current network connection status, c=0 represents the current connection to the cellular network, c=1 represents the current connection to the satellite network, and h represents the communication quality history in the recent period, which is represented by the communication network delay; P represents the deep reinforcement learning state transfer equation, which is fitted using a neural network; A represents the output of the deep reinforcement learning communication network switching model, A=0 means maintaining the current communication network connection, A=1 means switching to the cellular network, and A=2 means switching to the satellite network; R represents the reward function after executing the communication network switching strategy output: R t =w1R stab +w2R cost +w3R quality Among them, R t represents the environmental feedback at time t, w1, w2 and w3 represent the weighted weights of each sub-reward function, R stab represents the connection stability related reward function, R cost represents the switching cost-related reward function, R quality Represents the reward function related to communication service quality.

6. The high-level automatic driving system for vehicles integrating earth, sky and multiple domains according to claim 1 is characterized in that: The smart cockpit domain in the vehicle-side platform is used to improve driving safety, enhance user experience and provide personalized services, including information systems, interactive systems and network systems; The information system displays vehicle information and entertainment content in real time through a high-resolution display screen, allowing the driver to have a comprehensive understanding of the vehicle status; the interactive system realizes comprehensive information interaction between the vehicle and the driver through an intuitive user interface and voice recognition; the network system realizes seamless connection between the vehicle and the external network through vehicle networking technology.

7. The high-level automatic driving system for vehicles integrating earth, sky and multiple domains according to claim 1 is characterized in that: The intelligent driving domain in the vehicle-side platform includes a perception processing module, a trajectory prediction module, and a reinforcement learning module; The perception processing module obtains the blind-spot-free semantic bird's-eye view provided by the cloud control domain through Vehicle-to-Network (V2N) communication, and crops it according to the high-precision positioning provided by the satellite service domain. The cropped blind-spot-free semantic bird's-eye view matrix size is [0,1] w×h×c , where w represents the width of the cropped blind-spot-free semantic bird's-eye view matrix, h represents the height of the cropped blind-spot-free semantic bird's-eye view matrix, and c represents the number of channels of the cropped blind-spot-free semantic bird's-eye view matrix. The cropped blind-spot-free semantic bird's-eye view is stacked with the sensor information as part of the input of the reinforcement learning module. At the same time, the perception processing module uses the vehicle-side sensor to build a real-time dynamic map and upload it. Through dynamic matching based on high-precision positioning, real-time local update of the offline map is realized; The trajectory prediction module stacks the blind-zone-free semantic bird's-eye view provided by the cloud control domain in multiple time steps, and combines the scene factors provided by the knowledge management domain to achieve multi-time-step trajectory prediction of dynamic obstacles around the intelligent connected vehicle as another part of the input of the reinforcement learning module; The reinforcement learning module stacks the input provided by the perception processing module in multiple time steps, and integrates it with the multi-time-step trajectory prediction results of the surrounding dynamic obstacles provided by the trajectory prediction module into a complete reinforcement learning input. The reinforcement learning module builds a fusion reward function based on the driving safety field provided by the cloud control domain to guide the training process of the vehicle-side reinforcement learning neural network. Finally, the reinforcement learning module outputs control instructions to the powertrain domain.

8. The high-level automatic driving system for vehicles integrating earth, sky and multiple domains according to claim 1 is characterized in that: The powertrain domain in the vehicle-side platform outputs power control according to the control instructions of the intelligent driving domain, and through the synergy of the battery unit, motor unit and electronic control unit, improves the overall performance and efficiency of the powertrain domain, ensuring the power performance, energy management and safety of electric vehicles during driving; The battery unit provides a continuous and powerful power supply through a high energy density battery pack, achieving long driving range and high performance of the vehicle; The motor unit provides fast-response acceleration performance and a responsive driving experience through a highly efficient electric motor and advanced rotor design; The electronic control unit achieves intelligent coordination between the electric motor and the battery through advanced control algorithms, optimizing power output and energy management.

9. The high-level automatic driving system for vehicles with integrated space-ground multi-domain integration according to claim 1 is characterized in that: The smart chassis domain in the vehicle-side platform is connected to the smart cockpit domain, the smart driving domain, and the powertrain domain. First, through the interaction in the smart cockpit domain, the smart chassis domain is dynamically adjusted to realize the interaction between the smart cockpit domain and the smart chassis domain. Secondly, chassis feedback is used as part of the state quantity input of the intelligent driving domain to realize the interaction between the intelligent driving domain and the intelligent chassis domain; finally, the steering subsystem and DYC subsystem of the intelligent chassis domain calculate the yaw torque of the intelligent connected vehicle based on the output of the powertrain domain and the preview error between the actual path and the expected path; the drive subsystem obtains the expected target speed and calculates the four-wheel drive or braking torque of the intelligent connected vehicle based on the output of the powertrain domain and the current speed of the intelligent connected vehicle, thereby realizing the interaction between the powertrain domain and the intelligent chassis domain.

10. The high-level automatic driving system for vehicles integrating earth, sky and multiple domains according to claim 1, characterized in that: The information communication domain realizes inter-satellite communication, information monitoring, high-precision positioning and time synchronization services on the satellite platform; the information communication domain realizes information storage, data analysis, experience summary, and communication enhancement services on the cloud platform; the information communication domain realizes scene recognition, information sharing, experience reception, and network switching services on the vehicle-side platform.

Images