Intelligent driving function algorithm safety self-evolution method based on data closed loop

Abstract

The present application belongs to the technical field of intelligent driving function, and particularly relates to a method for safe self-evolution of intelligent driving function algorithm based on data closed loop, comprising: obtaining the current vehicle state based on the real-time data obtained by the vehicle sensor by using the current grey box model; generating the preliminary control instruction based on the current vehicle state through MPC optimization; performing safety check and optimization on the preliminary control instruction by using the control barrier function and quadratic programming, so as to ensure that the control instruction U safe Under the premise of meeting the safety constraint, the initial control instruction U initial is as close as possible safe ; the control instruction U safe is converted into the actual execution action of the vehicle; the grey box model is the sum of the vehicle state physical model not containing system noise and the data-driven model for describing the nonlinear characteristics and time-varying characteristics of the vehicle; the data-driven model in the grey box model is updated and optimized after a period of time, and can realize accurate prediction of the vehicle state in a complex and changeable environment.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field related to intelligent driving functions, and more specifically, relates to a data closed-loop based intelligent driving function algorithm safety self-evolution method. BACKGROUND

[0002] The initial stage of intelligent driving function development often faces the problem of data insufficiency and scene limitation. In order to quickly release usable function software methods and enter testing and verification as soon as possible, the system usually starts with a minimum viable product (MVP) as a starting point, aiming to realize basic functions such as adaptive cruise control, lane keeping, and automatic parking under ideal conditions through simple and stable algorithms. This method allows developers to quickly verify core functions and collect data feedback to provide the basis for subsequent iteration and improvement.

[0003] As vehicle systems are gradually applied to more complex and diverse real driving environments, the initial model relying only on fixed rules and static parameters has been difficult to meet the changing needs. In this case, data closed-loop based self-evolution becomes particularly important. Data closed-loop involves collecting actual driving data, conducting model updates and parameter optimization based on data, and updating the system to improve performance and adaptability. At the same time, when the system performs any algorithm optimization or parameter adjustment, it is also necessary to ensure that the safety of the system is not reduced due to the optimization process. Through the data closed-loop based safety self-evolution method, it helps the vehicle system to maintain driving safety and robustness when facing complex environments and uncertain factors.

[0004] Under this background, some new technical solutions have emerged. Patent "CN116300850A" proposes an automatic driving human-like safety self-evolution method and system based on data mechanism fusion, which extracts real human driving data features and iteratively extracts target functions using the maximum entropy inverse reinforcement learning algorithm. However, this method has a problem of dependence on a large amount of historical data, and challenges in adapting to unknown or abnormal traffic situations. In addition, patent "CN118529081A" proposes an edge scene based self-evolution learning type automatic driving auxiliary system and method, which realizes personalized vehicle automatic driving path recommendation through comprehensive user identity recognition, driving data collection and analysis, self-evolution intelligent learning, and individualized intelligent control module. Although it improves the adaptability of the system through the intelligent learning module, it faces challenges in the comprehensiveness and real-time nature of data collection, and cannot guarantee safety in variable traffic environments and complex driving scenarios. SUMMARY

[0005] In view of the above defects or improvement needs of the prior art, the present application provides a data closed loop based intelligent driving function algorithm safety self-evolution method, which aims to ensure the safety of intelligent driving in a variable traffic environment and a complex driving scene.

[0006] To achieve the above-mentioned purpose, according to one aspect of the present application, a data closed loop based intelligent driving function algorithm safety self-evolution method is provided, comprising:

[0007] Using the current grey box model, the current vehicle state is obtained based on the real-time data obtained by the vehicle sensor; based on the current vehicle state, a preliminary control instruction is generated through MPC optimization;

[0008] The safety of the preliminary control instruction is checked and optimized by using the control barrier function and quadratic programming, to ensure that the control instruction U safe Under the premise of meeting the safety constraint, the initial control instruction U initial is as close as possible; wherein the safety condition required by the vehicle state for the safety state is: h(S) represents the control barrier function, S represents the vehicle state, is the derivative of the control barrier function, and γ is a positive number;

[0009] The control instruction U safe is converted into the actual execution action of the vehicle;

[0010] The grey box model is an addition of a vehicle state physical model not containing system noise and a data driven model for describing the nonlinear characteristics and time-varying characteristics of the vehicle, and the data driven model in the grey box model is updated and optimized after a period of data collection, so as to realize accurate prediction of the vehicle state in a complex and variable environment.

[0011] Further, the generation method of the preliminary control instruction is:

[0012] The real-time data is processed by using a deep learning algorithm to identify traffic environment information, including lane lines, front vehicles, pedestrians and traffic signs; based on the traffic environment information, the target state of the vehicle is determined according to a preset driving strategy; wherein the real-time data includes real-time environment information data and real-time vehicle state data obtained by the vehicle sensor;

[0013] Using the current grey box model, the current vehicle state is obtained based on the real-time data; using the model predictive control (MPC) algorithm, the preliminary control instruction meeting the target state is generated based on the current vehicle state.

[0014] Further, the preliminary control instruction is optimized to obtain the control instruction U safe The quadratic programming optimization model of the control instruction U safe is:

[0015]

[0016] In the formula, U initial represents a preliminary control instruction, U represents a control instruction argument, and f(S, U) represents a vehicle state physical model without system noise.

[0017] Further, the gray-box model is represented as S k+1 = f physical (S k , U k ) + g θ (D key ); wherein f physical (S k , U k ) represents a vehicle state physical model without system noise, g θ (D key ) is a residual function based on a neural network, and theta is a parameter of the neural network; D key is key data obtained after preprocessing of vehicle state, environmental information, control input, and historical driving data.

[0018] The way of updating and optimizing the gray-box model of the vehicle is as follows:

[0019] The vehicle state, environmental information, control input, and historical driving data are used to update the training sample set, and the parameter theta of the neural network is adjusted; wherein, in each training sample, the key data obtained after preprocessing of the vehicle state, environmental information, control input, and historical driving data is used as the input data of the neural network, and the difference between the actual vehicle state and the vehicle state predicted by the physical model is used as the output data of the neural network.

[0020] Further, the implementation of the preprocessing is as follows: an edge computing method is used to filter and denoise the data, extract features, and compress the data.

[0021] According to another aspect of the present application, there is provided a data closed-loop based intelligent driving function algorithm safety self-evolution system, characterized in that it is used to execute the intelligent driving function algorithm safety self-evolution method as described above, and comprises a function algorithm module, a safety supervision module, a data collection and storage module, a model updating module, and an execution module.

[0022] The function algorithm module is used to generate a preliminary control instruction.

[0023] The safety supervision module is used to perform safety verification and optimization on the preliminary control instruction by using a control barrier function and quadratic programming to obtain the control instruction U safe .

[0024] The execution module is configured to convert the control instruction U safe into actual execution actions of the vehicle.

[0025] The data collection and storage module is configured to collect data for model parameter updating of the grey-box model by the model updating module.

[0026] Further, the data collection and storage module comprises a vehicle end and a cloud end, the vehicle end is configured to collect data and pre-process the data and upload the pre-processed data to the cloud end, or the vehicle end is configured to collect data and directly upload the data to the cloud end, and the cloud end is configured to pre-process the data.

[0027] The cloud end is further configured to perform model parameter updating of the grey-box model based on the pre-processed data and transmit the model parameter updating to the functional algorithm module and the safety supervision module, so as to provide current vehicle states for the functional algorithm module and the safety supervision module, respectively.

[0028] According to another aspect of the present application, an electronic device is provided, comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the method when executing the computer program.

[0029] According to another aspect of the present application, a computer readable storage medium is provided, comprising a stored computer program, wherein the computer program controls the device where the storage medium is located to execute the steps of the method when the computer program is run by a processor.

[0030] According to another aspect of the present application, a computer program product is provided, comprising a computer program or instructions, wherein the computer program or instructions implement the steps of the method when executed by a processor.

[0031] Overall, compared with the prior art, the technical scheme provided by the present application mainly has the following beneficial effects:

[0032] 1. This invention proposes a data-closed-loop-based method for the safe self-evolution of intelligent driving function algorithms. This method integrates functional algorithms, safety supervision, data collection and storage, model updates, and command execution to achieve real-time monitoring, dynamic optimization, and intelligent control of vehicle operating status. The data closed-loop mechanism continuously acquires high-quality data, supporting iterative updates of the model algorithm and improving system performance and reliability. The safety supervision mechanism ensures the safety of control commands, the model update operation enables continuous optimization of the vehicle model, and the execution operation guarantees the accurate execution of control commands. The entire method forms a data-driven closed-loop mechanism, enabling the intelligent driving system to adapt and operate reliably in complex environments, laying a solid foundation for the realization of high-level autonomous driving functions.

[0033] 2. This invention also proposes a gray-box model and a method for updating model parameters. The gray-box model combines the interpretability of a physical model with the adaptability of a data-driven model, enabling it to compensate for the shortcomings of the physical model and improve prediction accuracy and system robustness. It can optimize in real time, adapt to complex, nonlinear, and time-varying environments, and improve the safety and decision-making capabilities of autonomous driving systems. By dynamically correcting the errors of the physical model, the gray-box model can effectively cope with various driving scenarios, providing more precise control and decision support. Attached Figure Description

[0034] Figure 1 This is a flowchart of a self-evolution method for the safety of intelligent driving function algorithms based on data closed loop, provided by an embodiment of the present invention.

[0035] Figure 2 This is a flowchart illustrating the offline design and online implementation of the security monitoring mechanism provided in this embodiment of the invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0037] Example 1

[0038] A data-closed-loop-based method for the safe self-evolution of intelligent driving function algorithms, such as... Figure 1 As shown, it includes:

[0039] Using the current gray box model, the current vehicle state is obtained based on real-time data acquired by vehicle sensors; based on the current vehicle state, preliminary control commands are generated through MPC optimization.

[0040] By utilizing control barrier functions and quadratic programming, the initial control commands are subjected to safety verification and optimization to ensure the safety of control command U. safe While meeting safety constraints, it should be as close as possible to the initial control command U. initial The safety conditions that a vehicle must meet to be in a safe state are as follows: h(S) represents the obstacle control function, and s represents the vehicle state. The derivative of the control barrier function is given by γ, which is a positive number.

[0041] Control command U safe Transformed into the actual actions performed by the vehicle;

[0042] The gray box model is the sum of a vehicle state physical model that does not contain system noise and a data-driven model that describes the nonlinear and time-varying characteristics of the vehicle. The data-driven model in the gray box model is updated and optimized after a period of data collection, so as to achieve accurate prediction of vehicle state in complex and ever-changing environments.

[0043] It should be noted that the functional algorithm used to execute the above steps—"using the current gray-box model, obtaining the current vehicle state based on real-time data acquired by vehicle sensors; and generating preliminary control commands through MPC optimization based on the current vehicle state"—is responsible for implementing the most basic intelligent driving functions, i.e., the minimum feasible product of the system, such as adaptive cruise control, lane centering, and automatic parking. It receives data from vehicle sensors and generates preliminary control commands. On the one hand, the functional algorithm provides operational data for subsequent operations and analysis; on the other hand, it obtains optimized algorithm and model parameters from the model update algorithm to continuously improve functional performance. The model update algorithm is the algorithm that performs gray-box model parameter update optimization.

[0044] In other words, the functional algorithm is responsible for implementing the basic functions of intelligent driving, such as adaptive cruise control and lane keeping assist. This algorithm mainly includes three sub-algorithms: perception, decision-making, and planning. The perception sub-algorithm uses sensors such as cameras, millimeter-wave radar, and lidar on the vehicle to acquire environmental information and vehicle status data. This data can be used for selection and processing with deep learning algorithms to identify lane lines, vehicles ahead, pedestrians, traffic signs, and other important traffic environmental information. The decision-making sub-algorithm, based on the perception results, determines the vehicle's target state according to a predetermined driving strategy. The planning sub-algorithm uses a model predictive control (MPC) algorithm to generate control commands U that satisfy the target state. initial As a preferred implementation method, the MPC optimization problem can be formalized as follows:

[0045]

[0046] Among them, s k Let U be the vehicle state at time k. k For the control input, q and R are weight matrices, and N is the prediction time domain length. The functional algorithm module outputs the initial control command U. initial It is then sent to the security monitoring algorithm.

[0047] The safety monitoring algorithm, used to perform the step "using control barrier function and quadratic programming to perform safety verification and optimization of preliminary control commands" mentioned above, utilizes a combination of control barrier function (CBF) and quadratic programming (QP) to determine the vehicle's safety boundaries in real time. By transforming safety constraints into an optimization problem, QP can be used to solve for the optimal control input, ensuring that the vehicle achieves best performance while meeting safety conditions. This algorithm monitors the control commands output by the monitoring algorithm; if a potential risk is detected, the safety monitoring algorithm will adjust the control commands to ensure the safety of vehicle operation.

[0048] This safety monitoring algorithm utilizes the Control Barrier Function (CBF) and Quadratic Programming (QP) to perform safety verification and adjustment on the control commands generated by the functional algorithm. A control barrier function h(S) is defined; when h(S) ≥ 0, the vehicle is in a safe state; when h(S) < 0, the vehicle is in an unsafe state. To ensure that the vehicle state always remains within the safe set, the following conditions must be met:

[0049]

[0050] in, γ is the derivative of the control barrier function, and it is a positive number.

[0051] As a preferred implementation, while maintaining the original control commands as much as possible, the control input u is adjusted to meet safety constraints. The optimization problem can be categorized as follows:

[0052]

[0053] The safety monitoring algorithm solves the above QP problem, and serves as a monitoring controller, such as Figure 2 Receive safe control command U safe Ensure control command U safe While satisfying safety constraints, it should be as close as possible to the initial control command U. initial The optimized safety control commands will be sent to the execution module to ensure that vehicle operation complies with safety standards.

[0054] To execute the above steps, "control command U" safeThe execution algorithm, which "converts the actual actions of the vehicle into actions", receives the control command U generated by the safety monitoring algorithm. safe Then, it is responsible for translating these instructions into actual vehicle actions, such as acceleration, braking, and steering. The control signals are precisely executed through the vehicle's actuators (such as the engine control unit, braking system, and steering system). Furthermore, the execution algorithm collects feedback information from the actuators and vehicle status information in real time to support closed-loop optimization of the system, i.e., model parameter updates.

[0055] The model update algorithm combines a physical model and a data-driven learning approach to build a gray-box model of the vehicle. The physical model describes the vehicle's basic dynamic characteristics, while the data-driven learning approach compensates for the nonlinear and time-varying characteristics that the physical model struggles to accurately describe. Through online learning and historical data accumulation, the gray-box model is continuously updated, improving its accuracy and adaptability. The algorithm utilizes real-time and historical data to evaluate and adjust key parameters of the vehicle's gray-box model and functional algorithms. The optimized model parameters and algorithm are then distributed to the functional and safety supervision algorithms, enabling iterative updates of both the algorithm and the model.

[0056] First, a gray-box model of the vehicle is established and updated. The goal of this model update algorithm is to continuously optimize the vehicle's dynamic model, enabling the system to adapt to constantly changing driving environments and complex traffic conditions.

[0057] Based on the vehicle's dynamics and kinematics, a physical model of the vehicle is established:

[0058] S k+1 =f physical (S k U k )+w k

[0059] Among them, w k System noise represents random disturbances or uncertainties that cannot be fully described by a physical model. Data-driven models utilize machine learning algorithms to learn the nonlinear and time-varying characteristics of real vehicles that are difficult for physical models to accurately describe, and establish residual models.

[0060]

[0061] And approximate it using a neural network model:

[0062] r k ≈g θ (D key )

[0063] Where g θ (D keyLet θ be the residual function based on the neural network, and let θ be the parameters of the neural network. This neural network is trained offline using a large amount of historical driving data (including but not limited to past driving behavior, vehicle status, and environmental information) to learn the mapping relationship of the residual model. Furthermore, real-time collected vehicle status, environmental information, and user operation data will continuously optimize the neural network model during the online learning process to adapt to different driving environments.

[0064] By fusing physical and data-driven models, a gray-box model is formed.

[0065] S k+1 =f physical (S k U k )+g θ (D key )

[0066] To optimize the residual model, a deep neural network is used for fitting. The training objective is to minimize the prediction error, and the loss function is defined as:

[0067]

[0068] The loss function is minimized using an optimization algorithm, and the model parameters θ are updated. The updated gray-box model and algorithm parameters are then sent to the functional algorithm module and safety supervision module on the vehicle side, supporting iterative updates of the algorithm.

[0069] Through the collaborative work of the aforementioned modules, this invention achieves safe self-evolution of the intelligent driving function algorithm. Supported by the data collection and storage module and the model update module, the system can continuously acquire high-quality data and optimize the model, constantly improving the performance of the real-time function algorithm. The safety supervision module ensures the safety of the system during the optimization process, while the execution module guarantees the accurate implementation of control commands. The entire framework forms a data-driven closed-loop mechanism, enabling the intelligent driving system to adapt and operate reliably in complex environments.

[0070] As a preferred implementation method, the above-mentioned preliminary control command generation method is as follows:

[0071] Deep learning algorithms are used to process real-time data and identify traffic environment information, including lane lines, vehicles ahead, pedestrians, and traffic signs. Based on the traffic environment information, the target state of the vehicle is determined according to a preset driving strategy. The real-time data includes real-time environmental information data and real-time vehicle state data acquired by vehicle sensors.

[0072] Using the current gray box model and the aforementioned real-time data, the current vehicle state is obtained; using the Model Predictive Control (MPC) algorithm, preliminary control commands that meet the target state are generated based on the current vehicle state.

[0073] This can be considered a preferred implementation method, with the gray box model represented as S. k+1 =f physical (S k U k )+g θ (D key ); where f physical (S k U k ) represents a physical model of the vehicle state that does not include system noise, g θ (D key ) represents the residual function based on a neural network, and θ represents the parameters of the neural network; D key This refers to key data obtained after preprocessing vehicle status, environmental information, control inputs, and historical driving data.

[0074] The method for updating and optimizing the vehicle's gray box model is as follows:

[0075] The training sample set is updated using vehicle status, environmental information, control input, and historical driving data, and the parameters θ of the neural network are adjusted. In each training sample, the key data obtained after preprocessing the vehicle status, environmental information, control input, and historical driving data are used as input data to the neural network. The output data of the neural network is the difference between the actual vehicle status and the vehicle status predicted by the physical model.

[0076] The preferred method for the above preprocessing is to use edge computing to perform data filtering, noise reduction, feature extraction, and data compression.

[0077] This embodiment proposes a data-closed-loop-based self-evolution method for intelligent driving function algorithms, focusing on continuous optimization and safety assurance in real-world driving environments. Compared to existing technologies, this invention collects and utilizes data through a data closed-loop system to continuously optimize and adjust the algorithm, enabling it to have higher adaptability in complex and ever-changing traffic scenarios. This invention prioritizes safety as the core of the evolutionary process, implementing rigorous monitoring and feedback at every step of algorithm updates and model optimization to ensure high reliability in complex environments.

[0078] Example 2

[0079] A data-closed-loop-based intelligent driving function algorithm safety self-evolution system is used to execute the intelligent driving function algorithm safety self-evolution method as described above, including: a function algorithm module, a safety supervision module, a data collection and storage module, a model update module, and an execution module;

[0080] The functional algorithm module generates preliminary control commands; the safety monitoring module uses control obstacle functions and quadratic programming to perform safety verification and optimization on the preliminary control commands, resulting in control command U. safe The execution module is used to process control commands U safe This is transformed into the actual actions performed by the vehicle; the data collection and storage module is used to collect data so that the model update module can update the model parameters of the gray box model.

[0081] In this system, the functional algorithm module is responsible for implementing the most basic intelligent driving functions, i.e., the minimum feasible product of the system, such as adaptive cruise control, lane centering, and automatic parking. It receives data from vehicle sensors and generates initial control commands. On the one hand, the functional algorithm module provides operational data for subsequent operation and analysis; on the other hand, it obtains optimized algorithm and model parameters from the model update module to continuously improve functional performance.

[0082] In other words, this functional algorithm module is responsible for implementing the basic functions of intelligent driving, such as adaptive cruise control and lane keeping assist. This module mainly includes three sub-modules: perception, decision-making, and planning. The perception sub-module uses sensors such as cameras, millimeter-wave radar, and lidar on the vehicle to acquire environmental information and vehicle status data, and uses deep learning algorithms to process the sensor data to identify lane lines, vehicles ahead, pedestrians, traffic signs, and other important traffic environmental information. Based on the perception results, the decision-making sub-module determines the vehicle's target state according to a predetermined driving strategy. The planning sub-module uses a model predictive control (MPC) algorithm to generate control commands U that satisfy the target state. initial As a preferred implementation method, the MPC optimization problem can be formalized as follows:

[0083]

[0084] Among them, S k Let U be the vehicle state at time k. k For the control input, Q and R are weight matrices, and N is the prediction time domain length. The functional algorithm module outputs the initial control command U. initial Send it to the security monitoring module.

[0085] The safety monitoring module utilizes a combination of Control Barrier Function (CBF) and Quadratic Programming (QP) to determine the vehicle's safety boundaries in real time. By transforming safety constraints into an optimization problem, QP is used to solve for the optimal control input, ensuring the vehicle achieves best performance while meeting safety conditions. This module monitors the control commands output by the functional algorithm module; if a potential risk is detected, the safety monitoring module adjusts the control commands to ensure the vehicle's safe operation.

[0086] The safety monitoring module of this invention utilizes the Control Barrier Function (CBF) and Quadratic Programming (QP) methods to perform safety verification and adjustment on the control commands generated by the functional algorithm module. A control barrier function h(S) is set; when h(S) ≥ 0, the vehicle is in a safe state; when h(S) < 0, the vehicle is in an unsafe state. To ensure that the vehicle state always remains within the safe set, the following conditions must be met:

[0087]

[0088] in, Let γ be the derivative of the control obstacle function, where γ is a positive number. As a preferred implementation, while maintaining the original control command as much as possible, the control input U is adjusted to meet safety constraints. The optimization problem takes the form:

[0089]

[0090] The safety monitoring module solves the above QP problem to obtain the safety control command U. safe Ensure control command U safe While satisfying safety constraints, it should be as close as possible to the initial control command U. initial The optimized safety control commands will be sent to the execution module to ensure that vehicle operation complies with safety standards.

[0091] The execution module receives the control command U generated by the safety supervision module. safe Then, it is responsible for translating these instructions into actual vehicle actions, such as acceleration, braking, and steering. This means receiving final control commands from the safety monitoring module and directly controlling the vehicle's movement. Control signals are precisely executed through the vehicle's actuators (such as the engine control unit, braking system, and steering system). The execution module collects feedback information from the actuators and vehicle status information in real time and sends the feedback information to the data collection and storage module, supporting closed-loop optimization of the system.

[0092] Data Collection and Storage Module: The vehicle is equipped with various sensors, including cameras, radar, and lidar, to collect real-time vehicle status, environmental information, and user operation data. This module utilizes edge computing technology to preprocess the raw data, performing processes such as filtering, compression, and feature extraction to reduce data volume and improve data quality. The processed key data is then securely and efficiently uploaded to the cloud via high-speed wireless communication networks such as 5G or LTE-V. The cloud can clean, label, and integrate the collected data, removing anomalies and redundant information. This module provides high-quality data support for the model update module and simultaneously obtains operational data from the functional algorithm module, safety monitoring module, and execution module, forming a complete data loop.

[0093] The model update module combines a physical model with data-driven learning methods to build a gray-box model of the vehicle. The physical model describes the vehicle's basic dynamic characteristics, while the data-driven learning method compensates for the nonlinear and time-varying characteristics that the physical model cannot accurately describe. Through online learning and historical data accumulation, the gray-box model is continuously updated to improve its accuracy and adaptability. This module uses real-time and historical data to evaluate and adjust key parameters of the vehicle model and functional algorithms. The optimized model parameters and algorithms are then distributed to the functional algorithm module and the safety supervision module, enabling iterative updates of the algorithms and models.

[0094] Through the collaborative work of the aforementioned modules, the intelligent driving function algorithm achieves safe self-evolution. Supported by the data collection and storage module and the model update module, the system continuously acquires high-quality data and optimizes the model, constantly improving the performance of the real-time function algorithm. The safety supervision module ensures the safety of the system during the optimization process, while the execution module guarantees the accurate implementation of control commands. The entire framework forms a data-driven closed-loop mechanism, enabling the intelligent driving system to adapt and operate reliably in complex environments.

[0095] As a preferred implementation, the data collection and storage module includes a vehicle terminal and a cloud terminal. The vehicle terminal is used to collect data and preprocess it before uploading it to the cloud terminal, or the vehicle terminal is used to collect data and directly upload it to the cloud terminal. The cloud terminal is used to preprocess the data. The cloud terminal is also used to update the model parameters of the gray box model based on the preprocessed data and transmit them to the functional algorithm module and the safety supervision module, so as to provide the current vehicle status to the functional algorithm module and the safety supervision module, respectively.

[0096] The data collection and storage module is responsible for collecting, preprocessing, transmitting, and storing data both on the vehicle and in the cloud. On the vehicle side, sensors such as cameras, radar, and lidar acquire real-time vehicle status and environmental information. vehicle And obtain the driver's operation behavior data D from the execution module. userAnd vehicle execution feedback data D execution Edge computing technology can be used as a preferred method to filter and denoise the original data, extract features, and compress the data to obtain key data D. key The preprocessed key data D is transmitted via a high-speed wireless communication network. key The data is uploaded securely and efficiently to the cloud. The cloud cleans and labels the uploaded data, removing anomalies and redundant information, and stores the processed data in a cloud database for use by the model update module.

[0097] Other related technical solutions are the same as in Embodiment 1, and will not be repeated here.

[0098] It's worth noting that the data collection and storage module and edge computing technology can address the real-time performance issues in intelligent driving systems. By acquiring vehicle and environmental data in real time through multiple sensors and preprocessing it using edge computing, the amount of data transmitted is reduced, thereby improving the real-time performance of data processing. Furthermore, by combining a gray-box model with a physical model and a data-driven model, the robustness of the system in complex traffic environments is addressed, enabling continuous system optimization and improved performance in changing environments. The safety monitoring module ensures that the vehicle remains within a safe range of motion during optimization, thus improving vehicle control precision. The data collection and storage module supports comprehensive and real-time data acquisition through multiple sensors and high-speed communication technologies, ensuring the system can acquire real-time data to support subsequent model updates and algorithm optimization, thereby enhancing the overall intelligence and execution efficiency of the system.

[0099] In summary, the system framework proposed in this embodiment aims to address the challenges of insufficient data and limited scenarios in the early stages of intelligent driving system development, as well as the difficulty for fixed rules and static parameter models to meet the demands of complex and ever-changing real-world driving environments. By introducing a data closed-loop mechanism and a safety supervision module, continuous optimization and adaptive evolution of functional algorithms are achieved, while ensuring the system's safety and robustness. In other words, under the premise of ensuring safety, continuous optimization and adaptive evolution of the system's functional algorithms are realized, thereby improving the vehicle's adaptability and reliability in complex traffic environments.

[0100] Example 3

[0101] An electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0102] The electronic device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The memory can be used to store computer programs and / or modules. The processor performs various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory.

[0103] The relevant technical solutions are the same as above, and will not be repeated here.

[0104] Example 4

[0105] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0106] Specifically, the memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0107] The relevant technical solutions are the same as above, and will not be repeated here.

[0108] Example 5

[0109] A computer program product or computer program includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the methods described in the embodiments of this application.

[0110] The relevant technical solutions are the same as above, and will not be repeated here.

[0111] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for the safe self-evolution of intelligent driving function algorithms based on data closed-loop, characterized in that, include: Using the current gray box model, the current vehicle status is obtained based on real-time data acquired by vehicle sensors; Based on the current vehicle status, preliminary control commands are generated through MPC optimization. By utilizing control barrier functions and quadratic programming, the initial control commands are subjected to safety verification and optimization to ensure the safety of the control commands. While meeting safety constraints, it should be as close as possible to the initial control command. The safety conditions that a vehicle must meet to be in a safe state are as follows: ; This represents the control barrier function. Indicates vehicle status. The derivative of the control barrier function, It is a positive number; Control commands Transformed into the actual actions performed by the vehicle; The gray box model is the sum of a vehicle state physical model that does not contain system noise and a data-driven model that describes the nonlinear and time-varying characteristics of the vehicle. The data-driven model in the gray box model is updated and optimized after a period of data collection, so as to achieve accurate prediction of vehicle state in complex and ever-changing environments. Wherein, the gray box model is represented as ;in, This represents a physical model of the vehicle state that does not include system noise. For residual functions based on neural networks, These are the parameters of the neural network; This refers to key data obtained after preprocessing vehicle status, environmental information, control inputs, and historical driving data. The method for updating and optimizing the vehicle's gray box model is as follows: The training sample set is updated and the parameters of the neural network are adjusted using vehicle status, environmental information, control inputs, and historical driving data. In each training sample, the input data to the neural network consists of key data obtained after preprocessing vehicle state, environmental information, control input, and historical driving data; the output data of the neural network consists of the difference between the actual vehicle state and the vehicle state predicted by the physical model.

2. The intelligent driving function algorithm safety self-evolution method as described in claim 1, characterized in that, The initial method for generating control commands is as follows: Deep learning algorithms are used to process real-time data and identify traffic environment information, including lane lines, vehicles ahead, pedestrians, and traffic signs. Based on the traffic environment information, the target state of the vehicle is determined according to a preset driving strategy. The real-time data includes real-time environmental information data and real-time vehicle state data acquired by vehicle sensors. Using the current gray box model and based on the real-time data, the current vehicle state is obtained; using the Model Predictive Control (MPC) algorithm, preliminary control commands that meet the target state are generated based on the current vehicle state.

3. The intelligent driving function algorithm safety self-evolution method as described in claim 1, characterized in that, The initial control commands are optimized to obtain the control commands. The quadratic programming optimization model is as follows: In the formula, This indicates a preliminary control command. Indicates the control command argument. This represents a physical model of the vehicle state that does not include system noise.

4. The intelligent driving function algorithm safety self-evolution method as described in claim 1, characterized in that, The preprocessing is implemented by using edge computing methods to filter and reduce noise, extract features, and compress data.

5. A self-evolving safety algorithm system for intelligent driving functions based on data closed-loop, characterized in that, The method for implementing the intelligent driving function algorithm safety self-evolution method as described in any one of claims 1 to 4 includes: a function algorithm module, a safety supervision module, a data collection and storage module, a model update module, and an execution module; The functional algorithm module is used to generate preliminary control commands. The safety monitoring module is used to perform safety verification and optimization on the initial control commands using control barrier functions and quadratic programming, thereby obtaining the control commands. ; The execution module is used to process control commands. Transformed into the actual actions performed by the vehicle; The data collection and storage module is used for data collection so that the model update module can update the model parameters of the gray box model.

6. The intelligent driving function algorithm safety self-evolution system as described in claim 5, characterized in that, The data collection and storage module includes a vehicle terminal and a cloud terminal. The vehicle terminal is used to collect data and upload it to the cloud terminal after preprocessing it, or the vehicle terminal is used to collect data and upload it directly to the cloud terminal, and the cloud terminal is used to preprocess the data. The cloud platform is also used to update the model parameters of the gray box model based on the preprocessed data and transmit them to the functional algorithm module and the safety supervision module, so as to provide the current vehicle status to the functional algorithm module and the safety supervision module respectively.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 4.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein the computer program, when executed by a processor, controls the device on which the storage medium is located to perform the steps of the method as described in any one of claims 1 to 4.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 4.

Images