A chassis drive-by-wire method and system based on SAC reinforcement learning and dynamic boundary constraint

Abstract

This invention provides a chassis drive-by-wire method and system based on SAC reinforcement learning and dynamic boundary constraints. The method first collects vehicle dynamics data and preprocesses it to output a standardized dataset. Then, it constructs a multi-dimensional state space vector, fusing scene identifiers to form a feature input system. This is then input into the SAC reinforcement learning agent, and combined with a dynamic weight reward algorithm, it outputs the optimal control decision. Subsequently, the decision is split into two types of control flows and scheduled for output. Finally, the constraint commands are verified, transformed, and issued, and feedback data is collected to form a complete control closed loop. This invention achieves continuous collaborative control and multi-objective optimization of multiple systems by constructing a multi-dimensional feature system, calibrating actuator safety boundaries, and introducing a reinforcement learning framework with entropy regularization, thereby improving decision accuracy and adaptability to all scenarios. Furthermore, through high- and low-frequency control flow splitting and closed-loop feedback, it accurately distributes commands, reduces computational load, continuously optimizes strategies, strengthens safety redundancy, and improves operational stability and reliability.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and vehicle control technology, and in particular to a chassis drive-by-wire method and system based on SAC reinforcement learning and dynamic boundary constraints. Background Technology

[0002] As the core execution system of intelligent connected vehicles and autonomous vehicles, the drive-by-wire chassis abandons the traditional mechanical transmission method and uses electrical signals to achieve precise control of vehicle steering, braking, drive, and suspension actuators. It is a key foundation for realizing intelligent and automated vehicle driving. With the rapid development of the automotive industry towards intelligence and connectivity, the market's requirements for vehicle driving safety, stability, comfort, and real-time control decisions are constantly increasing. The drive-by-wire chassis needs to adapt to complex and ever-changing driving environments and different operating conditions, while taking into account the working status and service life of each actuator. In the current vehicle driving process, the dynamic state, environmental perception information, and actuator operation data present multi-source and high-dimensional characteristics. How to effectively integrate multi-source data, construct an intelligent control strategy that adapts to all operating conditions, and enable the various actuators of the drive-by-wire chassis to work collaboratively has become an important research direction for the development of drive-by-wire chassis technology, and has also promoted the research and application of drive-by-wire chassis control methods based on artificial intelligence and deep learning.

[0003] Traditional drive-by-wire chassis control methods often employ a subsystem-independent control model, with each actuator's control strategy operating independently. This lack of effective coordination and linkage mechanisms makes it difficult to achieve global optimization of multiple vehicle driving objectives. In complex operating conditions, control decision conflicts easily arise, impacting overall vehicle performance. Furthermore, traditional control methods rely heavily on preset, fixed control rules, failing to adapt to real-time vehicle driving status, environmental changes, and actuator health status. This results in poor adaptability to operating conditions and difficulty in quickly making optimal control decisions in emergency situations. Additionally, traditional control methods often set fixed action boundaries for actuators without dynamically adjusting based on their real-time health and driving conditions. This can easily lead to actuator overload, affecting control accuracy and reducing actuator lifespan. Moreover, the distribution and triggering methods for control commands are relatively simple, unable to match the response characteristics of different actuators, resulting in wasted control resources and reduced execution efficiency.

[0004] Therefore, it is of great significance to design a chassis intelligent collaborative control system with deep scene understanding capabilities, dynamic adaptive trade-offs, and smooth action guidance. Summary of the Invention

[0005] The purpose of this invention is to provide a chassis drive-by-wire method and system based on SAC reinforcement learning and dynamic boundary constraints. The core of the method relies on a five-step process: data acquisition and preprocessing, feature construction and boundary calibration, collaborative decision generation, asynchronous decision distribution, command execution, and closed-loop feedback to achieve intelligent control of the drive-by-wire chassis. The method integrates multi-source data to construct a feature input system, combines SAC reinforcement learning agents and dynamic weight reward algorithms to generate optimal collaborative decisions, and achieves efficient command scheduling through high- and low-frequency control flow decomposition and hybrid triggering algorithms. Simultaneously, a closed-loop feedback mechanism ensures the real-time performance and adaptability of the control. The system is configured with five functional modules to achieve modular execution of each stage. This invention effectively improves the accuracy, coordination, and safety of drive-by-wire chassis control, adapts to complex working conditions and the collaborative needs of multiple actuators, and provides a reliable technical solution for chassis control of intelligent connected and autonomous vehicles.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: On one hand, a chassis intelligent drive-by-wire method based on deep learning, the method comprising the following steps in sequence: S1, Data Acquisition and Preprocessing: Collect vehicle dynamics data, environmental perception data, and state data of each actuator of the drive-by-wire chassis. Perform timestamp alignment, filtering and noise reduction, and state estimation on the collected data to output a standardized dataset. S2, Feature Construction and Boundary Labeling: Extract features from the standardized dataset to construct a multi-dimensional state space vector, generate scene identifiers based on vehicle speed and distance to the target object, calculate a health index to characterize the health of the actuators based on actuator state data, calculate the real-time safety action boundary of each actuator through a dynamic safety boundary algorithm, and integrate the scene identifiers, health index, and real-time safety action boundary into the multi-dimensional state space vector to form a feature input system; S3, Collaborative Decision Generation: The feature input system is input into the pre-trained SAC reinforcement learning agent. The real-time safe action boundary is used as the output constraint of the policy network. The continuous action vector is output through the SAC algorithm. During the agent's reasoning process, a dynamic weight reward algorithm is introduced to calculate the reward weight under different scenarios to guide the policy network to output the optimal control decision. The mathematical expression for the SAC algorithm is: in Let be the optimization objective function of the policy network. For the length of the decision time domain, This indicates the state-action pair in the experience replay pool. Expectations for The multidimensional state space vector at time step. for Continuous action vector at each moment, Calculated for dynamic weighted reward algorithm Real-time reward value This is a temperature parameter used to adjust the relative ratio of the reward to the entropy term. For strategy In state The entropy value is used to improve the strategy exploration capability. When calculating the continuous action vector through the SAC algorithm, the multi-dimensional state space vector in the feature input system is first input into the policy network of the pre-trained SAC reinforcement learning agent. The policy network performs nonlinear transformation and feature extraction on the input multi-dimensional state features layer by layer. At the same time, the real-time safety action boundary is taken as the hard constraint condition of the policy network output. During the feature mapping process, the evaluation result of the action value by the dual value network is combined simultaneously to iteratively correct the output parameters of the policy network. Finally, the policy network outputs the control increment command of the steer-by-wire subsystem, the steer-by-wire drive and braking subsystem, and the active suspension subsystem, as well as the continuous action vector corresponding to the cooperative weight coefficient of the three subsystems. Each parameter output is a continuously changing value, and the value range of all parameters does not exceed the range defined by the real-time safety action boundary. S4, Asynchronous Decision Distribution: The optimal control decision is split into high-frequency base control flow and low-frequency event control flow. The two types of control flow are triggered and the two types of control flow instructions are calculated by a hybrid trigger output algorithm. After constructing a hybrid priority queue to complete instruction scheduling, the effective control instructions are output. The specific steps for splitting the optimal control decision into high-frequency base control flow and low-frequency event control flow are as follows: Based on the response frequency requirements of the actuator, the priority of control functions, and mechanical characteristics, the drive-by-wire control commands and brake-by-wire control commands that are directly related to the longitudinal driving stability of the vehicle are classified as high-frequency base control flow. This type of control flow needs to meet the requirement of short-cycle continuous output to adapt to real-time control requirements such as drive anti-slip and brake anti-lock. The steering-by-wire control commands used to adjust the lateral driving posture of the vehicle and the active suspension control commands used to suppress the vertical vibration of the vehicle body are classified as low-frequency event control flow. The execution effect of this type of control flow is strongly correlated with the response characteristics of the mechanical structure. It does not require high-frequency repeated output. It only needs to update the commands when the driving state or scenario requirements change to meet the control requirements. The splitting of the two types of control flow must ensure that all control commands are complete and non-overlapping, and each corresponds to a clear actuator control requirement. S5, Command Execution and Closed-Loop Feedback: Verify and restrict effective control commands, convert the processed commands into underlying physical control signals and send them to the corresponding actuators, collect the status data of the vehicle and actuators after execution and feed them back to step S1; The verification and limitation of effective control commands are based on the real-time safe action boundary of each actuator. Effective control commands that exceed the real-time safe action boundary are subjected to saturation limiting processing, and effective control commands that are close to the real-time safe action boundary are subjected to gradient correction processing. The underlying physical control signals include the motor target current signal of the steer-by-wire actuator, the hydraulic pump duty cycle adjustment signal of the brake-by-wire actuator, and the proportional solenoid valve control signal of the active suspension actuator. The state data acquisition frequency fed back to step S1 is consistent with the data acquisition frequency in step S1. The acquired state data needs to be timestamped and filtered for noise reduction before being input into the standardized dataset generation stage of step S1.

[0007] The aforementioned deep learning-based intelligent drive-by-wire chassis method collects vehicle dynamics data including longitudinal velocity, lateral velocity, yaw rate, lateral acceleration, vertical acceleration, body roll angle, body pitch angle, steering wheel angle, braking torque, and suspension actuator feedback force. Environmental perception data includes distance to obstacles around the vehicle, obstacle orientation, relative speed of obstacles, vehicle position information relative to the target trajectory, vehicle heading information, road curvature, and road surface adhesion conditions. The status data of each actuator in the drive-by-wire chassis includes actuator operating temperature, drive current, actual stroke, response delay, cumulative number of working cycles, and operating vibration amplitude.

[0008] The aforementioned deep learning-based intelligent drive-by-wire chassis method includes the following steps: First, the timestamp alignment process uses the global clock of the vehicle controller as a reference to assign high-precision timestamps to vehicle dynamics data, environmental perception data, and the state data of each actuator in the drive-by-wire chassis. This calibrates for time deviations caused by different acquisition devices and transmission buses, ensuring consistency of the three types of data in the time dimension. Second, the filtering and denoising process employs corresponding methods based on the noise characteristics of different data types. Vehicle dynamics data undergoes moving average filtering to remove high-frequency vibration interference, environmental perception data undergoes median filtering to eliminate outliers, and actuator state data undergoes low-pass filtering to isolate electromagnetic interference. Third, the state estimation process, based on the filtered effective data and combined with the fundamental relationships of vehicle kinematics, calculates key vehicle state parameters that cannot be directly acquired through multi-source data cross-validation. Fourth, the generation of the standardized dataset requires unifying the dimensions and normalizing the range of all processed data, arranging them according to a preset data format to form a structured dataset file.

[0009] The aforementioned deep learning-based intelligent drive-by-wire method for chassis includes a multi-dimensional state space vector comprising vehicle motion features, driving task and path tracking features, and actuator and operating condition features. The vehicle motion features are extracted from vehicle dynamics data in a standardized dataset. The driving task and path tracking features include the lateral position deviation, heading angle deviation, and deviation between the target speed and the actual speed of the vehicle relative to the target trajectory. The actuator and operating condition features include steering wheel angle sensitivity, braking torque feedback, suspension force feedback, scene identifier, health index, and real-time safety action boundary.

[0010] The aforementioned deep learning-based intelligent drive-by-wire method for chassis involves generating scene identifiers based on a preset scene classification matrix. The row dimension of the scene classification matrix is ​​divided into different vehicle speed ranges, and the column dimension is divided into different target object distance ranges. By discretizing and encoding the measured vehicle speed and the measured target object distance, the corresponding operating condition code is obtained by matching from the scene classification matrix. This operating condition code serves as the scene identifier. The health index ranges from 0 to 1 and is calculated by weighting three indicators: actuator temperature rise status, wear and aging status, and response characteristics.

[0011] The aforementioned deep learning-based intelligent drive-by-wire method for chassis, wherein the mathematical expression of the dynamic safety boundary algorithm is: in For the first Each actuator Real-time safety action boundaries at any given moment. For steering actuator, For brake actuator, For suspension actuators, For the first The factory-set basic physical limits of each actuator For the first The health index of each actuator This is the working condition scenario factor, with a value range of 0.8 to 1.0. For the first The inherent characteristic coefficients of each actuator range from 0.9 to 1.0. When calculating the real-time safe action boundary of each actuator using the dynamic safety boundary algorithm, the factory physical limit of each actuator is first used as the baseline boundary. The baseline boundary is initially adjusted based on the current health status of the actuator as reflected by the health index. Then, according to the driving condition type corresponding to the scene identifier, the condition adaptation coefficient is introduced to further correct the boundary range. At the same time, the inherent characteristic compensation amount is supplemented by taking into account the mechanical structure characteristics and drive method differences of the actuator. Finally, the upper and lower limits of the action range that each actuator can safely perform in the current state are formed. This range needs to cover all reasonable control requirements of the actuator in the corresponding working condition.

[0012] The aforementioned deep learning-based intelligent chassis drive-by-wire method comprises a SAC reinforcement learning agent consisting of a policy network and a dual value network. The policy network is a three-layer fully connected neural network with the number of input layer nodes matching the feature dimension of the multidimensional state space vector, and the number of hidden layer nodes being 256 and 128 respectively. The output layer outputs a continuous action vector. The dual value network consists of two identical three-layer fully connected neural networks. The continuous action vector includes control increment commands for steer-by-wire, drive-by-wire and braking, and active suspension, as well as the collaborative weight coefficients of each subsystem. The collaborative weight coefficients range from 0 to 1, and the sum of all collaborative weight coefficients is 1.

[0013] The aforementioned deep learning-based intelligent drive-by-wire method for chassis, wherein the mathematical expression of the dynamic weight reward algorithm is: in for Total instant reward value at any given moment To optimize the target dimension, For path tracking accuracy, For driving stability, For passenger comfort, for Time of the first The reward weight of each optimization objective, and , For the first The base reward value for each optimization objective The motion smoothing penalty coefficient, for Change in action vector at time step The actuator health protection penalty coefficient, For the first Each actuator The actual action value at any given moment. For the first Each actuator Real-time safety action boundaries at any given moment. This is a safety reward correction term. When calculating the continuous action vector using the SAC algorithm, the multi-dimensional state space vector in the feature input system is used as input. It is received by the input layer of the policy network and converted into a feature format that the network can recognize. Feature extraction and nonlinear mapping are performed through the fully connected operation of the hidden layer. The output layer combines the constraints of the real-time safety action boundary and outputs control increment parameters for steer-by-wire, drive-by-wire and braking, and active suspension, as well as collaborative weight parameters for coordinating the working priorities of each subsystem. All parameters are continuously changing values ​​and are strictly within the preset reasonable value range.

[0014] The aforementioned deep learning-based intelligent drive-by-wire method for chassis, wherein the mathematical expression of the hybrid trigger output algorithm is: ; ,in for Valid control commands output at all times. for The optimal control decision action value output by the SAC algorithm at time [time]. This is the instruction priority preemption coefficient, with a value ranging from 1.0 to 1.2. This is the instruction validity coefficient, with a value ranging from 0.9 to 1.0. This is a time-triggered decision factor, with a value of either 0 or 1. This is the event triggering factor, which can take the value of 0 or 1. for Valid control commands output at all times. For the first Each actuator The system defines real-time safety action boundaries. When triggering and calculating control flow commands for two types of control flows using a hybrid trigger output algorithm, a fixed trigger cycle is set for the high-frequency base control flow. Within each cycle, the corresponding control command is validated and its value updated to ensure the continuity and real-time nature of the command output. For the low-frequency event control flow, preset trigger conditions such as scene switching, excessive vehicle state deviation, and sudden changes in actuator health are established. The command calculation process is only initiated when any trigger condition is met; otherwise, the previous valid command remains unchanged. Command calculation for both types of control flows must be based on the core parameters of the optimal control decision, combined with their respective control characteristics for numerical conversion and adaptation processing to ensure that the command format meets the receiving requirements of the actuator.

[0015] On the other hand, a deep learning-based intelligent drive-by-wire system for chassis includes: The data acquisition and preprocessing module is used to collect vehicle dynamics data, environmental perception data, and state data of various actuators of the drive-by-wire chassis. The collected data is sequentially processed by timestamp alignment, filtering and noise reduction, and state estimation, and a standardized dataset is output. The feature construction and boundary calibration module is communicatively connected to the data acquisition and preprocessing module. It is used to extract features from the standardized dataset to construct a multi-dimensional state space vector, generate scene identifiers based on vehicle speed and distance to the target object, calculate the health index of the actuator and the real-time safety action boundary of each actuator, and integrate the scene identifier, health index and real-time safety action boundary into the multi-dimensional state space vector to form a feature input system. The collaborative decision generation module is communicatively connected to the feature construction and boundary calibration module. It is used to input the feature input system into the pre-trained SAC reinforcement learning agent, use the real-time safe action boundary as the policy network output constraint, output continuous action vector through the SAC algorithm, and introduce a dynamic weight reward algorithm to calculate the reward weight to guide the policy network to output the optimal control decision. The asynchronous decision distribution module is communicatively connected to the collaborative decision generation module. It is used to split the optimal control decision into a high-frequency base control flow and a low-frequency event control flow. It uses a hybrid trigger output algorithm to determine the trigger of the two types of control flows and calculate the control flow instructions. After constructing a hybrid priority queue to complete the instruction scheduling, it outputs the effective control instructions. The instruction execution and closed-loop feedback module is communicatively connected to the asynchronous decision distribution module and the data acquisition and preprocessing module, respectively. It is used to verify and restrict valid control instructions, convert the processed instructions into underlying physical control signals and send them to the corresponding actuators. At the same time, it collects the status data of the vehicle and the actuators after execution, processes them and feeds them back to the data acquisition and preprocessing module.

[0016] The beneficial effects achieved by this invention are as follows: (1) This invention achieves accurate representation of the vehicle's full-dimensional driving state by standardizing the preprocessing of multi-source driving data and integrating the multi-dimensional state space feature system. Combined with the dynamic adaptation of the health status of the actuators and the driving conditions, it completes the real-time calibration of the safety action boundaries of each actuator, and builds a safety constraint base covering all working conditions for chassis drive-by-wire control. By introducing a reinforcement learning decision framework with entropy regularization, and using the real-time safety boundary as the hard constraint of the strategy output, combined with the scenario-based dynamic weight reward mechanism, it realizes the continuous collaborative control of multiple systems such as drive-by-wire steering, drive braking, and active suspension. It takes into account the multi-objective optimization of path tracking, driving stability and ride comfort, effectively solves the problems of insufficient multi-system collaboration and poor working condition adaptability in traditional drive-by-wire chassis control, and greatly improves the decision accuracy and full-scenario adaptability of the chassis drive-by-wire system.

[0017] (2) This invention achieves precise distribution and orderly scheduling of chassis control commands through differentiated splitting of high and low frequency control flows and hybrid triggering output mechanism. Combining the response characteristics of the actuator and the priority of control functions, it achieves high-frequency continuous output of control commands related to longitudinal driving stability and event-driven on-demand updates of commands related to vehicle posture and body vibration control. While ensuring the real-time performance of core driving safety control, it effectively reduces the system's computational load and the loss of ineffective actions of the actuator. At the same time, through the closed-loop feedback mechanism of the whole link, the status data of the execution end is transmitted back to the data preprocessing stage in real time, realizing continuous iterative optimization of the control strategy. Combined with the safety verification and gradient correction of the command output end, it further strengthens the safety redundancy of the system and significantly improves the operational stability and long-term reliability of the drive-by-wire chassis control. Attached Figure Description

[0018] Figure 1 A flowchart of a deep learning-based intelligent chassis drive-by-wire method; Figure 2 This is a framework diagram of a chassis intelligent drive-by-wire system based on deep learning. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention discloses a chassis drive-by-wire method and system based on SAC reinforcement learning and dynamic boundary constraints. The system employs a step-by-step execution approach, including data acquisition and preprocessing, feature construction and boundary calibration, collaborative decision generation, asynchronous decision distribution, command execution, and closed-loop feedback. Figure 1 As shown, by combining the collaborative work of various functional modules, intelligent, precise, and collaborative control of the drive-by-wire chassis is achieved, effectively improving vehicle stability, safety, and ride comfort, and adapting to the dynamic control requirements of different driving scenarios and actuator working states. The execution steps of this system's corresponding method include a data acquisition and preprocessing module, a feature construction and boundary calibration module, a collaborative decision generation module, an asynchronous decision distribution module, and an instruction execution and closed-loop feedback module. These modules are sequentially connected via communication. The instruction execution and closed-loop feedback module and the data acquisition and preprocessing module are bidirectionally connected, forming a complete closed-loop control system. The implementation details of each step and corresponding module are as follows: S1, Data Acquisition Preprocessing, is performed by the Data Acquisition Preprocessing Module: This step utilizes various acquisition devices and sensors to achieve comprehensive data collection of vehicle dynamics, environmental perception, and the status of drive-by-wire chassis actuators. After a series of processing steps, a standardized dataset is obtained, providing a high-quality data source for subsequent feature construction. The specific implementation steps are as follows: Multi-dimensional data acquisition equipment and sensors are adapted to various monitoring points on the vehicle, simultaneously collecting three types of data: vehicle dynamics data, including longitudinal velocity, lateral velocity, yaw rate, lateral acceleration, vertical acceleration, body roll angle, body pitch angle, steering wheel angle, braking torque, and suspension actuator feedback force; environmental perception data, including distance to obstacles around the vehicle, obstacle orientation, relative speed of obstacles, vehicle position information relative to the target trajectory, vehicle heading information, road curvature, and road surface adhesion conditions; and status data of various actuators in the drive-by-wire chassis, including actuator operating temperature, drive current, actual stroke, response delay, cumulative number of working cycles, and operating vibration amplitude.

[0021] The timestamp alignment process uses the global clock of the vehicle controller as a reference to mark the three types of collected data with high-precision timestamps. It calibrates the time deviation caused by different acquisition devices and different transmission buses to ensure the consistency of vehicle dynamics data, environmental perception data and actuator status data in the time dimension.

[0022] The filtering and denoising process employs differentiated filtering methods based on the noise characteristics of different types of data. For vehicle dynamics data, moving average filtering is used to remove high-frequency vibration interference; for environmental perception data, median filtering is used to remove outliers; and for actuator status data, low-pass filtering is used to isolate electromagnetic interference, thus obtaining effective data with various types of interference removed.

[0023] The state estimation process is based on the filtered effective data. Combined with the basic kinematic relationships of the vehicle, it calculates the key vehicle state parameters that cannot be directly collected by using multi-source data cross-validation, thus completing the data dimensions.

[0024] The standardized dataset generation process unifies the dimensions and normalizes the range of all data after timestamp alignment, filtering and denoising, and state estimation. The data is then organized and arranged according to a preset structured data format to form a standardized dataset file, which is output by the data acquisition and preprocessing module to the feature construction and boundary calibration module.

[0025] S2, Feature Construction and Boundary Labeling, is performed by the Feature Construction and Boundary Labeling module: This step, based on a standardized dataset, constructs a multidimensional state space vector to complete scene identification, health calculation, and safety boundary calibration. Relevant parameters are then integrated into the feature vector to form a feature input system, providing a feature foundation for collaborative decision generation. The specific implementation steps are as follows: The multidimensional state space vector is initially constructed by extracting features from the standardized dataset. The multidimensional state space vector contains three types of features: vehicle motion features are extracted from vehicle dynamics data in the standardized dataset; driving task and path tracking features include the lateral position deviation of the vehicle relative to the target trajectory, heading angle deviation, and deviation between the target vehicle speed and the actual vehicle speed; and actuator and operating condition features initially include steering wheel angle sensitivity, braking torque feedback, and suspension force feedback.

[0026] A scene identifier is generated by a preset scene classification matrix. The row dimension of this matrix is ​​divided into different vehicle speed ranges, and the column dimension is divided into different target object distance ranges. The measured vehicle speed and the measured target object distance are discretized and encoded. The corresponding working condition code is obtained by matching from the scene classification matrix, and the working condition code is used as the scene identifier.

[0027] The health index of the actuator is calculated by selecting three types of indicators: temperature rise, wear and aging, and response characteristics. The health index of each actuator is obtained by weighted calculation. The value of the index ranges from 0 to 1. The closer the value is to 1, the better the health status of the actuator.

[0028] Real-time safety action boundary calculation uses a dynamic safety boundary algorithm to calculate the real-time safety action boundary of each actuator. The mathematical expression of the dynamic safety boundary algorithm is: in For the first Each actuator Real-time safety action boundaries at any given moment. For steering actuator, For brake actuator, For suspension actuators, For the first The factory-set basic physical limits of each actuator For the first The health index of each actuator This is the working condition scenario factor, with a value range of 0.8 to 1.0. For the first The inherent characteristic coefficients of each actuator range from 0.9 to 1.0. The calculation first uses the factory physical limits of each actuator as the baseline boundary, then adjusts the baseline boundary based on the health index. Next, a condition adaptation coefficient is introduced based on the driving condition type corresponding to the scenario identifier for further correction. Finally, the inherent characteristic compensation amount is supplemented by considering the differences in the actuator's mechanical structure characteristics and drive method, forming the upper and lower limits of the safe operating range that each actuator can perform under the current state.

[0029] The feature input system integrates scene identifiers, health indices of each actuator, and real-time safety action boundaries of each actuator into a preliminarily constructed multi-dimensional state space vector, thereby improving the feature dimensions of the actuators and operating conditions and forming a complete feature input system. This system is then output from the feature construction and boundary calibration module to the collaborative decision generation module.

[0030] S3, Collaborative Decision Generation, is executed by the Collaborative Decision Generation module: This step inputs the feature input system into the pre-trained SAC reinforcement learning agent. Through algorithmic constraints and reward guidance, it outputs a continuous action vector containing control commands and collaborative weights for each subsystem, thus obtaining the optimal control decision. The specific implementation steps are as follows: The SAC reinforcement learning agent will be constructed using a policy network and a dual-value network. The policy network is a three-layer fully connected neural network with the number of input layer nodes matching the feature dimension of the multidimensional state space vector. The number of hidden layer nodes is 256 and 128 respectively, and the output layer is used to output continuous action vectors. The dual-value network consists of two identical three-layer fully connected neural networks used to evaluate the value of actions. After pre-training, the agent is in a state awaiting inference.

[0031] The policy network input and constraint setting involves inputting the multi-dimensional state space vector from the feature input system into the policy network of the SAC reinforcement learning agent, while retrieving the real-time safety action boundary calculated in step two and using it as a hard constraint condition for the policy network output, thus limiting the range of values ​​for the network output parameters.

[0032] Dynamic weighted reward calculation is introduced during agent inference to calculate reward weights under different scenarios. The mathematical expression of the dynamic weighted reward algorithm is: in for Total instant reward value at any given moment To optimize the target dimension, For path tracking accuracy, For driving stability, For passenger comfort, for Time of the first The reward weight of each optimization objective, and , For the first The base reward value for each optimization objective The motion smoothing penalty coefficient, for Change in action vector at time step The actuator health protection penalty coefficient, For the first Each actuator The actual action value at any given moment. For the first Each actuator Real-time safety action boundaries at any given moment. This is a safety reward correction term. The reward weights calculated by this algorithm guide the policy network towards the optimal control decision direction.

[0033] SAC Algorithm Inference and Action Vector Output: The SAC algorithm is used for agent inference. The mathematical expression of the SAC algorithm is: in Let be the optimization objective function of the policy network. For the length of the decision time domain, This indicates the state-action pair in the experience replay pool. Expectations for The multidimensional state space vector at time step. for Continuous action vector at each moment, Calculated for dynamic weighted reward algorithm Real-time reward value This is a temperature parameter used to adjust the relative ratio of the reward to the entropy term. For strategy In state The entropy value is used to enhance the strategy exploration capability. During the inference process, the strategy network performs layer-by-layer nonlinear transformation and feature extraction on the multidimensional state features of the input. At the same time, combined with the evaluation results of the action value by the dual value network, the output parameters of the strategy network are iteratively corrected. Finally, the strategy network outputs a continuous action vector. This vector contains the control increment commands of the steer-by-wire subsystem, the drive-by-wire and braking subsystem, and the active suspension subsystem, as well as the corresponding cooperative weight coefficients of the three subsystems (with values ​​ranging from 0 to 1 and a sum of 1). All parameters are continuously changing values ​​and their values ​​do not exceed the range defined by the real-time safety action boundary. This continuous action vector is the optimal control decision.

[0034] S4, asynchronous decision dispatch, is executed by the asynchronous decision dispatch module: This step decomposes the optimal control decision into two types of control flows, completes instruction judgment and scheduling through a hybrid triggering algorithm, and outputs effective control instructions to adapt to the response characteristics and control requirements of different actuators. The specific implementation steps are as follows: The control flow is decomposed based on the actuator's response frequency requirements, control function priorities, and mechanical characteristics. The optimal control decision is divided into a high-frequency base control flow and a low-frequency event control flow, ensuring that all control commands are complete and non-overlapping, corresponding to specific actuator control requirements. Specifically, steer-by-wire control commands and brake-by-wire control commands, directly related to the vehicle's longitudinal stability, are classified as high-frequency base control flows, adapting to real-time control requirements such as traction control and anti-lock braking. Steer-by-wire control commands used to adjust the vehicle's lateral attitude and active suspension control commands used to suppress vertical vibrations are classified as low-frequency event control flows, which only need to be updated when driving conditions or scenario requirements change to meet control requirements.

[0035] The hybrid trigger output algorithm is used to determine the trigger for both types of control flow and calculate the control flow command. The mathematical expression of the hybrid trigger output algorithm is: ; ,in for Valid control commands output at all times. for The optimal control decision action value output by the SAC algorithm at time [time]. This is the instruction priority preemption coefficient, with a value ranging from 1.0 to 1.2. This is the instruction validity coefficient, with a value ranging from 0.9 to 1.0. This is a time-triggered decision factor, with a value of either 0 or 1. This is the event triggering factor, which can take the value of 0 or 1. for Valid control commands output at all times. For the first Each actuator Real-time safety action boundaries at all times. For high-frequency base control flow, a fixed trigger cycle is set, and the corresponding control command is validated and updated in each cycle. For low-frequency event control flow, preset trigger conditions such as scene switching, vehicle state deviation exceeding the standard, and sudden change in actuator health are set. The command calculation process is only initiated when any of the trigger conditions are met. If no trigger condition is met, the last valid command remains unchanged. The command calculation for both types of control flow is based on the core parameters of optimal control decision, and numerical conversion and adaptation are performed in combination with their respective control characteristics to ensure that the command format meets the requirements of the actuator.

[0036] The instruction scheduling and output constructs a hybrid priority queue. Based on the functional priority and triggering results of the two types of control flows, the control flow instructions calculated by the hybrid triggering output algorithm are scheduled. The high-frequency base control flow has a higher priority than the low-frequency event control flow. After scheduling is completed, the integrated effective control instructions are output to the instruction execution and closed-loop feedback module.

[0037] S5, instruction execution and closed-loop feedback, is executed by the instruction execution and closed-loop feedback module: This step verifies and processes valid control commands, converts them into physical signals for execution, and simultaneously collects post-execution status data to complete closed-loop feedback, forming continuous dynamic control. The specific implementation steps are as follows: Effective control command verification and limitation are based on the real-time safe action boundaries of each actuator calculated in step two. The received effective control commands are verified and processed a second time. Commands that exceed the real-time safe action boundaries are saturated and limited, and commands that are close to the real-time safe action boundaries are graded and corrected to ensure that the command parameters are within the safe operating range of the actuator.

[0038] The underlying physical control signal conversion and transmission process converts valid control commands, after verification and limitation, into underlying physical control signals that can be recognized by the actuators. Among them, the steer-by-wire actuator corresponds to the motor target current signal, the brake-by-wire actuator corresponds to the hydraulic pump duty cycle adjustment signal, and the active suspension actuator corresponds to the proportional solenoid valve control signal. These various physical control signals are then transmitted to the corresponding steer-by-wire chassis actuators to execute control actions.

[0039] After the actuator completes the control action, the vehicle dynamics data, environmental perception data, and status data of each actuator of the drive-by-wire chassis are collected synchronously using the same acquisition equipment and sensors as in step one. The acquisition frequency is consistent with the data acquisition frequency in step one.

[0040] Data processing and closed-loop feedback perform timestamp alignment and filtering / denoising on the acquired post-execution status data. The processed dataset is then fed back to the data acquisition and preprocessing module in step one as the raw data for the next round of control flow, completing the closed-loop control of the entire chassis intelligent drive-by-wire system and enabling dynamic iteration and optimization of the control strategy. Figure 2 As shown.

[0041] This invention discloses a chassis drive-by-wire method and system based on SAC reinforcement learning and dynamic boundary constraints. Through a five-step closed-loop process and the collaborative operation of five major modules, it achieves collaborative intelligent control of multiple subsystems in a drive-by-wire chassis. Combining the SAC reinforcement learning algorithm with a dynamic weighted reward mechanism, it makes control decisions more adaptable to different driving scenarios and actuator states. Asynchronous decomposition and hybrid triggering of the control flow improve the real-time performance and adaptability of command execution. Dynamic calibration of real-time safety action boundaries and multiple verification of commands ensure the safe operation of the actuators and the driving safety of the vehicle. This method and system are widely adaptable to various drive-by-wire chassis vehicles, effectively improving the control accuracy and intelligence level of drive-by-wire chassis, and reducing vehicle driving risks caused by actuator failures and insufficient scenario adaptability.

[0042] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0043] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in the claims, they should all fall within the protection scope of the present invention.

Claims

1. A chassis intelligent drive-by-wire method based on deep learning, characterized in that, The method includes the following steps in sequence: S1, Data Acquisition and Preprocessing: Collect vehicle dynamics data, environmental perception data, and state data of each actuator of the drive-by-wire chassis. Perform timestamp alignment, filtering and noise reduction, and state estimation on the collected data to output a standardized dataset. S2, Feature Construction and Boundary Labeling: Extract features from the standardized dataset to construct a multi-dimensional state space vector, generate scene identifiers based on vehicle speed and distance to the target object, calculate a health index to characterize the health of the actuators based on actuator state data, calculate the real-time safety action boundary of each actuator through a dynamic safety boundary algorithm, and integrate the scene identifiers, health index, and real-time safety action boundary into the multi-dimensional state space vector to form a feature input system; S3, Collaborative Decision Generation: The feature input system is input into the pre-trained SAC reinforcement learning agent. The real-time safe action boundary is used as the output constraint of the policy network. The continuous action vector is output through the SAC algorithm. During the agent's reasoning process, a dynamic weight reward algorithm is introduced to calculate the reward weight under different scenarios to guide the policy network to output the optimal control decision. S4, Asynchronous Decision Distribution: The optimal control decision is split into high-frequency base control flow and low-frequency event control flow. The two types of control flow are triggered and the two types of control flow instructions are calculated by a hybrid trigger output algorithm. After constructing a hybrid priority queue to complete instruction scheduling, the effective control instructions are output. S5, Command Execution and Closed-Loop Feedback: Verify and restrict valid control commands, convert the processed commands into underlying physical control signals and send them to the corresponding actuators, collect the status data of the vehicle and actuators after execution and feed them back to step S1.

2. The chassis intelligent drive-by-wire method based on deep learning according to claim 1, characterized in that, In step S1, the collected vehicle dynamics data includes the vehicle's longitudinal velocity, lateral velocity, yaw rate, lateral acceleration, vertical acceleration, body roll angle, body pitch angle, steering wheel angle, braking torque, and suspension actuator feedback force. The environmental perception data includes the distance to obstacles around the vehicle, obstacle orientation, obstacle relative velocity, vehicle position information relative to the target's trajectory, vehicle heading information, road curvature, and road surface adhesion conditions. The status data of each actuator of the drive-by-wire chassis includes the actuator's operating temperature, drive current, actual stroke, response delay, cumulative number of working cycles, and operating vibration amplitude.

3. The chassis intelligent drive-by-wire method based on deep learning according to claim 1, characterized in that, In step S1, the timestamp alignment process uses the global clock of the vehicle controller as a reference to add high-precision timestamps to the vehicle dynamics data, environmental perception data, and the status data of each actuator of the drive-by-wire chassis. It calibrates the time deviation caused by different acquisition devices and different transmission buses to ensure the consistency of the three types of data in the time dimension. The filtering and denoising process adopts corresponding methods for the noise characteristics of different types of data. The vehicle dynamics data is filtered by moving average filtering to remove high-frequency vibration interference, the environmental perception data is filtered by median filtering to remove outliers, and the actuator status data is filtered by low-pass filtering to isolate electromagnetic interference. The state estimation process calculates the key vehicle state parameters that cannot be directly collected based on the filtered effective data and combined with the basic kinematic relationships of the vehicle. The generation of standardized datasets requires unifying the dimensions and normalizing the range of all processed data, and arranging them according to a preset data format to form a structured dataset file.

4. The intelligent chassis drive-by-wire method based on deep learning according to claim 1, characterized in that, In step S2, the multidimensional state space vector includes vehicle motion features, driving task and path tracking features, actuator and operating condition features. The vehicle motion features are extracted from vehicle dynamics data in a standardized dataset. The driving task and path tracking features include the lateral position deviation of the vehicle relative to the target trajectory, the heading angle deviation, and the deviation between the target vehicle speed and the actual vehicle speed. The actuator and operating condition features include steering wheel angle sensitivity, braking torque feedback, suspension force feedback, scene identifier, health index, and real-time safety action boundary.

5. The chassis intelligent drive-by-wire method based on deep learning according to claim 1, characterized in that, In step S2, the scene identifier is generated based on a preset scene classification matrix. The row dimension of the scene classification matrix is ​​divided into different vehicle speed ranges, and the column dimension is divided into different target distance ranges. By discretizing and encoding the measured vehicle speed and the measured target distance, the corresponding working condition code is obtained from the scene classification matrix. This working condition code serves as the scene identifier. The health index ranges from 0 to 1. The health index is calculated by weighting three indicators: the temperature rise status of the actuator, the wear and aging status, and the response characteristics.

6. The intelligent chassis drive-by-wire method based on deep learning according to claim 1, characterized in that, In step S2, the mathematical expression of the dynamic safety boundary algorithm is: in For the first One actuator in Real-time safety action boundaries at all times. For steering actuator, For brake actuator, For suspension actuators, For the first The factory-set physical limits of each actuator For the first The health index of the actuator This is the working condition scenario factor, with a value range of 0.8 to 1.

0. For the first The inherent characteristic coefficient of each actuator ranges from 0.9 to 1.

0.

7. The intelligent chassis drive-by-wire method based on deep learning according to claim 1, characterized in that, In step S3, the SAC reinforcement learning agent consists of a policy network and a dual value network. The policy network is a three-layer fully connected neural network with the number of input layer nodes matching the feature dimension of the multidimensional state space vector. The number of hidden layer nodes is 256 and 128 respectively, and the output layer outputs a continuous action vector. The dual value network consists of two identical three-layer fully connected neural networks. The continuous action vector includes control increment commands for steer-by-wire, drive-by-wire and braking, and active suspension, as well as the collaborative weight coefficients of each subsystem. The collaborative weight coefficients range from 0 to 1, and the sum of all collaborative weight coefficients is 1.

8. The intelligent chassis drive-by-wire method based on deep learning according to claim 1, characterized in that, In step S3, the mathematical expression of the dynamic weighted reward algorithm is: in for Total instant reward value at any given moment To optimize the target dimension, For path tracking accuracy, For driving stability, For passenger comfort, for Time of the first The reward weight of each optimization objective, and , For the first The base reward value for each optimization objective The motion smoothing penalty coefficient, for Change in action vector at time step The actuator health protection penalty coefficient, For the first The actual action value of each actuator at time t. For the first One actuator in Real-time safety action boundaries at all times. This is a safety reward modification.

9. The intelligent chassis drive-by-wire method based on deep learning according to claim 1, characterized in that, In step S4, the mathematical expression of the hybrid trigger output algorithm is: ; ,in for Valid control commands output at all times. for The optimal control decision action value output by the SAC algorithm at time [time]. This is the instruction priority preemption coefficient, with a value ranging from 1.0 to 1.

2. This is the instruction validity coefficient, with a value ranging from 0.9 to 1.

0. This is a time-triggered decision factor, with a value of either 0 or 1. This is the event triggering factor, which can take the value 0 or 1. Valid control commands that are output at all times. For the first One actuator in Real-time safety action boundaries at every moment.

10. A chassis intelligent drive-by-wire system based on deep learning, the system being applicable to the chassis intelligent drive-by-wire method based on deep learning as described in any one of claims 1-9, characterized in that, The system includes: The data acquisition and preprocessing module is used to collect vehicle dynamics data, environmental perception data, and state data of various actuators of the drive-by-wire chassis. The collected data is sequentially processed by timestamp alignment, filtering and noise reduction, and state estimation, and a standardized dataset is output. The feature construction and boundary calibration module is communicatively connected to the data acquisition and preprocessing module. It is used to extract features from the standardized dataset to construct a multi-dimensional state space vector, generate scene identifiers based on vehicle speed and distance to the target object, calculate the health index of the actuator and the real-time safety action boundary of each actuator, and integrate the scene identifier, health index and real-time safety action boundary into the multi-dimensional state space vector to form a feature input system. The collaborative decision generation module is communicatively connected to the feature construction and boundary calibration module. It is used to input the feature input system into the pre-trained SAC reinforcement learning agent, use the real-time safe action boundary as the policy network output constraint, output continuous action vector through the SAC algorithm, and introduce a dynamic weight reward algorithm to calculate the reward weight to guide the policy network to output the optimal control decision. The asynchronous decision distribution module is communicatively connected to the collaborative decision generation module. It is used to split the optimal control decision into a high-frequency base control flow and a low-frequency event control flow. It uses a hybrid trigger output algorithm to determine the trigger of the two types of control flows and calculate the control flow instructions. After constructing a hybrid priority queue to complete the instruction scheduling, it outputs the effective control instructions. The instruction execution and closed-loop feedback module is communicatively connected to the asynchronous decision distribution module and the data acquisition and preprocessing module, respectively. It is used to verify and restrict valid control instructions, convert the processed instructions into underlying physical control signals and send them to the corresponding actuators. At the same time, it collects the status data of the vehicle and the actuators after execution, processes them and feeds them back to the data acquisition and preprocessing module.

Images