Vehicle steering control method, vehicle and storage medium

Abstract

This application provides a vehicle steering control method, a vehicle, and a storage medium, relating to the field of vehicle technology. The method includes: acquiring vehicle state data and driving state data; determining multi-dimensional state parameters based on the vehicle state data and driving state data, wherein the multi-dimensional state parameters include: driver state score, driving style coefficient, and environmental risk value; determining weight parameters based on the multi-dimensional state parameters, preset rules, and a reinforcement learning model, wherein the preset rules are used to allocate weight parameters under preset scenarios; determining a target steering command based on the weight parameters, driving style coefficient, environmental risk value, system-planned steering angle, and steering wheel angle; and performing vehicle steering control based on the target steering command, driving style coefficient, and the rate of change of the weight parameters. This application solves the technical problems of poor adaptability, insufficient coordination, and scenario mismatch in human-machine co-driving steering control in traditional solutions.

Description

Technical Field

[0001] This application relates to the field of vehicle technology, and more specifically, to a vehicle steering control method, a vehicle, and a storage medium. Background Technology

[0002] In intelligent vehicle steer-by-wire systems, users increasingly demand that the system can actively recognize driving intentions, adapt to personalized styles, and cope with complex environments.

[0003] However, traditional solutions rely solely on a single sensor threshold for static weight switching, and the allocation of control rights depends on only a single control condition, resulting in technical problems such as poor adaptability, insufficient coordination, and scenario mismatch in human-machine co-driving steering control.

[0004] There is currently no good solution to the above problems. Summary of the Invention

[0005] This application provides a vehicle steering control method, a vehicle, and a storage medium to at least solve the technical problems of poor adaptability, insufficient coordination, and scenario mismatch in human-machine co-driving steering control in traditional solutions.

[0006] According to one aspect of the embodiments of this application, a vehicle steering control method is provided, comprising: acquiring vehicle state data and driving state data, wherein the vehicle state data is used to characterize the vehicle's driving state and the environmental state of the vehicle during driving, and the driving state data is used to characterize the driver's operating intention and attention level; determining multi-dimensional state parameters based on the vehicle state data and driving state data, wherein the multi-dimensional state parameters include: driver state score, driving style coefficient, and environmental risk value; determining weight parameters based on the multi-dimensional state parameters, preset rules, and a reinforcement learning model, wherein the preset rules are used to allocate the weight parameters under preset scenarios, and the weight parameters include driver control weight and system control weight; determining a target steering command based on the weight parameters, the driving style coefficient, the environmental risk value, the system-planned steering angle, and the steering wheel angle; and performing vehicle steering control based on the target steering command, the driving style coefficient, and the rate of change of the weight parameters.

[0007] Furthermore, based on the vehicle status data and the driving status data, the multi-dimensional status parameters are determined as follows: a driver status score is determined based on steering torque, steering angular velocity, the percentage of time the driver's gaze deviates from the center of the lane, the driver's blink frequency, and the driver's head pitch angle; a driving style coefficient is determined based on the standard deviation of steering amplitude, following distance, and the rate of change of accelerator pedal opening; and an environmental risk value is determined based on road curvature, road slope, road adhesion coefficient, relative distance to obstacles, relative speed of obstacles, and weather conditions.

[0008] Furthermore, the driver's state score is determined based on steering torque, steering angular velocity, the percentage of time the driver's gaze is off-center from the lane center, the driver's blinking frequency, and the driver's head pitch angle. This includes: determining the driver's operational intention score based on steering torque and steering angular velocity; determining the driver's attention score based on the percentage of time the driver's gaze is off-center from the lane center, the driver's blinking frequency, and the driver's head pitch angle; and determining the driver's state score based on the operational intention score and the attention score.

[0009] Furthermore, determining the driving style coefficient based on the standard deviation of steering amplitude, following distance, and accelerator pedal opening change rate includes: determining the historical standard deviation of steering amplitude, historical following distance, and historical accelerator pedal opening change rate based on historical vehicle status data; using a clustering algorithm to cluster the historical standard deviation of steering amplitude, historical following distance, and historical accelerator pedal opening change rate to obtain multiple clustering results; determining a preset driving style type based on the multiple clustering results, and establishing a style coefficient mapping table based on the preset driving style type, wherein the style coefficient mapping table is used to record the mapping relationship between the preset driving style type and the driving style coefficient; using a clustering algorithm to cluster the standard deviation of steering amplitude, following distance, and accelerator pedal opening change rate to obtain at least one target clustering result; and determining the driving style coefficient corresponding to at least one target clustering result based on the style coefficient mapping table.

[0010] Furthermore, the environmental risk value is determined based on road curvature, road slope, road adhesion coefficient, relative distance to obstacles, relative speed of obstacles, and weather conditions. This includes: determining the static risk coefficient based on road curvature, road slope, and road adhesion coefficient; determining the collision time based on the relative distance to obstacles and the relative speed of obstacles, and determining the dynamic risk coefficient based on the collision time; determining the weather conditions based on onboard meteorological sensors and onboard cameras, and determining the meteorological risk coefficient based on the weather conditions; and determining the environmental risk value based on the static risk coefficient, dynamic risk coefficient, and meteorological risk coefficient.

[0011] Furthermore, based on the multidimensional state parameters, preset rules, and reinforcement learning model, the weight parameters are determined as follows: determining whether the vehicle is in a preset scenario based on the multidimensional state parameters; in response to the vehicle being in a preset scenario, determining the weight parameters based on the multidimensional state parameters and preset rules; in response to the vehicle not being in a preset scenario, processing the multidimensional state parameters using the reinforcement learning model to obtain initial weight parameters; and correcting the initial weight parameters based on the driving style coefficient to obtain the final weight parameters.

[0012] Furthermore, the target steering instruction is determined based on the weight parameters, driving style coefficient, environmental risk value, system planning steering angle, and steering wheel steering angle. This includes: determining multiple target weight coefficients based on the driver control weight, environmental risk value, and driving style coefficient; and determining the target steering instruction based on the multiple target weight coefficients, system planning steering angle, steering wheel steering angle, and constraints. The constraints include steering angle constraints and steering rate constraints.

[0013] Furthermore, the vehicle steering control based on the target steering command, driving style coefficient, and weight parameter change rate includes: performing closed-loop control on the target steering command using a preset closed-loop control strategy to obtain the initial execution torque, wherein the preset closed-loop control strategy includes a position closed-loop control strategy, a torque closed-loop control strategy, and a current closed-loop control strategy; determining a transition function based on the weight parameter change rate and the initial execution torque, and determining a transition torque based on the transition function and the initial execution torque; determining the road feel compensation torque of the road feel feedback motor based on the transition torque, driving style coefficient, and road feel simulation rules, wherein the road feel simulation rules are used to determine the corresponding road feel compensation torque based on different driving style coefficients; determining the target execution torque based on the road feel compensation torque and the transition torque, and performing vehicle steering control based on the target execution torque.

[0014] According to another aspect of the embodiments of this application, a vehicle steering control device is also provided, comprising: an acquisition module, configured to acquire vehicle state data and driving state data, wherein the vehicle state data is used to characterize the vehicle's driving state and the environmental state of the vehicle during driving, and the driving state data is used to characterize the driver's operating intention and attention level; a state parameter determination module, configured to determine multi-dimensional state parameters based on the vehicle state data and the driver's driving state data, wherein the vehicle state data is used to characterize the vehicle's driving state and the driving environment state, and the driving state data is used to characterize the driver's operating intention and attention level, and the multi-dimensional state parameters include: driver state score, driving style coefficient, and environmental risk value; a weight parameter determination module, configured to determine weight parameters based on the multi-dimensional state parameters, preset rules, and a reinforcement learning model, wherein the preset rules are used to allocate weight parameters under preset scenarios, and the weight parameters include driver control weight and system control weight; an instruction determination module, configured to determine a target steering instruction based on the weight parameters, driving style coefficient, environmental risk value, system planned steering angle, and steering wheel steering angle; and a control module, configured to perform steering control on the vehicle based on the target steering instruction, driving style coefficient, and the rate of change of the weight parameters.

[0015] According to another aspect of the embodiments of this application, a vehicle is also provided, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods in various embodiments of this application when it runs.

[0016] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of this application.

[0017] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the methods of various embodiments of this application.

[0018] In this embodiment, vehicle state data and driving state data are acquired. Vehicle state data characterizes the vehicle's driving state and the environmental state during driving, while driving state data characterizes the driver's operational intention and attention level. Based on the vehicle state data and driver's driving state data, multi-dimensional state parameters are determined. These parameters include driver state score, driving style coefficient, and environmental risk value. Weight parameters are determined based on the multi-dimensional state parameters, preset rules, and a reinforcement learning model. The preset rules allocate weight parameters in preset scenarios, and the weight parameters include driver control weight and system control weight. A target steering command is determined based on the weight parameters, driving style coefficient, environmental risk value, system-planned steering angle, and steering wheel angle. Finally, vehicle steering control is performed based on the target steering command, driving style coefficient, and the rate of change of the weight parameters. This achieves the goal of accurately matching the driver's intentions with the safety requirements of the scenario, thereby realizing the technical effects of personalized control allocation and adaptive execution response. It also solves the technical problems of poor adaptability, insufficient coordination, and scenario mismatch in human-machine co-driving steering control in traditional solutions. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0020] Figure 1 This is a flowchart of a vehicle steering control method according to one embodiment of this application;

[0021] Figure 2 This is a structural block diagram of a vehicle steering control system according to one embodiment of this application;

[0022] Figure 3 This is a structural block diagram of a vehicle steering control device according to one embodiment of this application. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] Human-machine co-driving technology, as a crucial transitional stage before achieving fully autonomous driving, hinges on realizing a safe, smooth, and efficient allocation and collaboration of control between the driver and the autonomous driving system. While steer-by-wire systems, by eliminating mechanical connections, provide the hardware foundation for flexible control allocation, traditional solutions still suffer from the following significant drawbacks:

[0026] 1. Static and singular allocation of control: dependent on curve curvature threshold ( ) or steering wheel torque ( The switching is triggered by a single parameter, without taking into account driver fatigue, driving style, and weather conditions; linear weighting cannot adapt to the different needs of aggressive / conservative drivers, resulting in a disconnect between control and intention, which can easily lead to problems such as "system control grabbing" or "driver misoperation".

[0027] 2. Simplified instruction conflict resolution: Employing either "hard switching" or "fixed weighting" ( The logic, failing to consider multiple objectives such as safety, intent, and comfort, is prone to sudden torque changes. The above leads to "thugs" and in low-risk scenarios, the forced system dominance causes a sense of operational deprivation.

[0028] 3. Fixed control parameters: The proportional-integral-derivative control (PID) parameters and torque transition function of the steer-by-wire motor are fixed and not tied to scenario risks or driving styles: emergency scenarios ( Response lag (delay exceeding) The transitions between normal scenarios are abrupt, and the road feel is not differentiated by style (e.g., aggressive drivers need a strong road feel, while conservative drivers need a weak resistance), resulting in inconsistent operating experiences.

[0029] In summary, traditional solutions cannot achieve dynamic adaptation across the entire "perception-decision-optimization-execution" chain, and there is an urgent need for a human-machine collaborative control technology that deeply integrates multi-dimensional states, dynamically adjusts strategies, and finely coordinates conflicts.

[0030] According to an embodiment of this application, a method embodiment for vehicle steering control is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0031] This embodiment provides a vehicle steering control method. Figure 1 This is a flowchart of a vehicle steering control method according to one embodiment of this application, such as... Figure 1 As shown, the process includes the following steps:

[0032] Step S11: Obtain vehicle status data and driving status data. The vehicle status data is used to characterize the vehicle's driving status and the environmental status of the vehicle during driving. The driving status data is used to characterize the driver's operating intention and attention level.

[0033] In this embodiment, the vehicle steering control system first acquires vehicle state data and driving state data. The vehicle state data characterizes the vehicle's driving state and the environmental state during driving, reflecting the constraints of the external environment on steering control. Exemplarily, this includes the standard deviation of the steering amplitude. Following distance accelerator pedal opening change rate Road curvature ,slope Adhesion coefficient derived from road surface texture characteristics Relative distances to surrounding obstacles (vehicles / pedestrians) Relative velocity Weather conditions, etc.

[0034] Driving state data is used to characterize the driver's operational intentions and attention levels, reflecting the driver's current willingness to actively control and cognitive state. Examples include steering torque T_δ and steering angular velocity. Percentage of time spent with eyes off the center of the lane blinking frequency and head tilt angle wait.

[0035] Therefore, the vehicle steering control system acquires vehicle status data and driving status data. By synchronously acquiring vehicle and driving status data, this application dynamically acquires dual information of "human" and "environment", avoiding the crude mode of traditional systems that rely on a single sensor (such as steering wheel torque) or fixed threshold (such as curvature > 0.01 rad / m) to determine control authority, and providing data support for subsequent data analysis.

[0036] Step S12: Determine multi-dimensional state parameters based on vehicle state data and driving state data. The multi-dimensional state parameters include: driver state score, driving style coefficient, and environmental risk value.

[0037] In this embodiment of the application, the vehicle steering control system determines multi-dimensional state parameters based on vehicle state data and driving state data. The multi-dimensional state parameters include driver state score, driving style coefficient and environmental risk value, which are three quantitative indicators that characterize the collaborative interaction state of "human-vehicle-environment" and constitute the input for the system's dynamic decision-making.

[0038] Driver Status Rating It comprehensively reflects the driver's current operational intentions and level of attention. For example, a scoring range is set. A higher value indicates a more stable driver state and a stronger willingness to participate.

[0039] Driving Style Coefficient Based on long-term driving behavior data, this is used to characterize the driver's personalized behavioral preferences and inertial patterns during steering operations. For example, this can be achieved by analyzing nearly 50km of historical data (steering amplitude standard deviation). Following distance accelerator pedal opening change rate The scores are divided into three types: radical (1.0), standard (0.7), and conservative (0.4).

[0040] Environmental risk value This is used to comprehensively, dynamically, and quantitatively characterize the overall hazard level of the vehicle's current driving environment. For example, the environmental risk value can integrate comprehensive assessment indicators of static road risk (curvature, slope, coefficient of friction), dynamic collision risk (TTC), and meteorological risk (rain, snow, fog, etc.), with a set value range of [missing information]. A higher value indicates a more dangerous driving environment, and the system needs to increase its control.

[0041] Therefore, vehicle steering control determines multi-dimensional state parameters based on vehicle state data and driving state data, transforming the original heterogeneous sensor data into semantically clear, computable, and comparable high-level state features, providing a scientific and stable data foundation for subsequent dynamic weight allocation, multi-objective optimization control, and adaptive execution.

[0042] Step S13: Determine the weight parameters based on the multidimensional state parameters, preset rules, and reinforcement learning model. The preset rules are used to allocate the weight parameters under preset scenarios. The weight parameters include driver control weights and system control weights.

[0043] In this embodiment, the vehicle steering control system determines weight parameters based on multi-dimensional state parameters, preset rules, and a reinforcement learning model. The preset rules are used to allocate weight parameters under preset scenarios with different risk levels. For example, extreme scenarios (such as high-risk or clearly incapacitated scenarios) can be handled by referring to an expert rule base to ensure that the system has safety constraints under special circumstances.

[0044] The weighting parameters include driver control weight and system control weight. These weighting parameters refer to two complementary coefficients used to dynamically allocate driver control and autonomous driving system control during human-machine co-driving steering control. Specifically: driver control weight ( `_driver` represents the relative dominance of the driver's intention in the final output of the current steering command, with a value range of [0,1]. System control weights ( ) represents the relative dominance of the desired steering commands generated by the autonomous driving system in the final output, where .

[0045] For example, reinforcement learning models can use deep reinforcement learning networks. It is responsible for making decisions in routine scenarios. Network (3 fully connected layers, 3 input dimensions, 1 output dimension, ReLU activation function) output The network calculates the Q-value update strategy based on the reward function and iterates once every 20ms. It is suitable for personalized and adaptive decision-making in normal scenarios, and realizes nonlinear, personalized and multi-objective optimization allocation of steering control in human-machine co-driving.

[0046] Therefore, the vehicle steering control system determines the weight parameters based on multi-dimensional state parameters, preset rules and reinforcement learning models, and generates the driver and system control weights in real time based on the multi-dimensional state parameters. This solves the rigid switching problem of traditional human-machine co-driving systems and also realizes dynamic allocation of rights and responsibilities.

[0047] Step S14: Determine the target steering command based on the weight parameters, driving style coefficient, environmental risk value, system planned steering angle, and steering wheel steering angle.

[0048] In this embodiment, the vehicle steering control system determines the target steering command based on weight parameters, driving style coefficient, environmental risk value, system-planned steering angle, and steering wheel steering angle.

[0049] Among them, the system planning turning angle This is used to provide the steering system with optimal safe steering reference commands generated by autonomous driving functions, based on environmental perception and path planning. For example, it can be generated by the autonomous driving system using a Model Predictive Control (MPC) algorithm based on high-precision maps, Global Positioning System (GPS), and vehicle dynamics models, representing the optimal steering angle trajectory the vehicle is expected to follow within the next 3 seconds. This reflects the system's globally optimal decisions regarding safe path maintenance, obstacle avoidance, and lane centering, and is the system's "ideal control target."

[0050] Steering wheel angle The steering angle refers to the physical steering angle directly applied by the driver through the steering wheel. For example, it can be collected in real time by a high-resolution steering wheel angle sensor and then personalized after being compensated by driving style correction (aggressive +1°, standard 0°, conservative -1°), representing the driver's direct control intention.

[0051] The target steering command is the final output fused steering command, which is the result of multi-objective optimization of the system-planned steering angle and the driver's steering angle under dynamic weight adjustment, taking into account safety, intention compliance and operation smoothness.

[0052] Therefore, the vehicle steering control system intelligently integrates the "safe path" of the autonomous driving system with the "true intention" of the driver by constructing a multi-objective optimization function with different dynamic weights as coefficients, and generates precise, smooth and personalized target steering commands while meeting mechanical limits and comfort constraints.

[0053] Step S15: Perform steering control on the vehicle based on the target steering command, driving style coefficient, and rate of change of weight parameters.

[0054] In this embodiment, the vehicle steering control performs steering control on the vehicle based on the target steering command, driving style coefficient, and the rate of change of weight parameters.

[0055] Among them, the rate of change of the weight parameters ( This is used to characterize the intensity of human-machine control transitions in real time, determine the scenario type, and select the corresponding transition function. For example, it can be based on... Size determination is a normal transition ( (e.g., style switching) or emergency transition ( (e.g., emergency takeover).

[0056] Therefore, the vehicle steering control system controls the vehicle's steering based on the target steering command, driving style coefficient, and the rate of change of weight parameters. It uses the target steering command, driving style coefficient, and rate of change of weight parameters as the control input for steering execution. By dynamically sensing the intensity of the control switch, it intelligently selects the transition strategy and synchronously links the road feel feedback parameters to achieve adaptive and coordinated control, ensuring rapid response and smooth transition. This constructs a human-machine co-driving closed loop with full-link style perception of "perception-decision-execution-feedback".

[0057] In summary, this application first achieves dynamic allocation of control by fusing driver state scores, environmental risk values, and driving style coefficients through multimodal perception, thereby improving accident avoidance rates in extreme scenarios. Secondly, it employs a hybrid reinforcement learning model and pre-defined rules for collaborative decision-making based on multi-dimensional state parameters, ensuring that control allocation is both personalized and safe. Then, a multi-objective optimization function based on dynamic weights collaboratively integrates the system-planned steering angle with the style-corrected manual steering angle, maintaining stable lane accuracy and low latency in intent response. Subsequently, optimized commands ensure smooth steering trajectories and avoid overly aggressive maneuvers; a piecewise transition function driven by weight change rate ensures smooth transitions during normal switching and responds with feedforward compensation during emergency takeovers, resolving the "hand-hitting" problem. Furthermore, this application incorporates a driving style-adaptive road feel feedback mechanism, providing aggressive drivers with clear road feel and conservative drivers with gentle resistance, contributing to improved human-machine trust and operational satisfaction.

[0058] The above steps of this application involve acquiring vehicle state data and driving state data. Vehicle state data characterizes the vehicle's driving state and the environmental state during driving, while driving state data characterizes the driver's operational intentions and attention levels. Based on the vehicle state data and driver's driving state data, multi-dimensional state parameters are determined. These parameters include driver state score, driving style coefficient, and environmental risk value. Weight parameters are determined based on the multi-dimensional state parameters, preset rules, and a reinforcement learning model. The preset rules allocate weight parameters under preset scenarios, and these weight parameters include driver control weight and system control weight. A target steering command is determined based on the weight parameters, driving style coefficient, environmental risk value, system-planned steering angle, and steering wheel angle. Finally, vehicle steering control is performed based on the target steering command, driving style coefficient, and the rate of change of the weight parameters. This achieves the goal of accurately matching the driver's intentions with the safety requirements of the scenario, thereby realizing the technical effects of personalized control allocation and adaptive execution response. It also solves the technical problems of poor adaptability, insufficient coordination, and scenario mismatch in human-machine co-driving steering control in traditional solutions.

[0059] Optionally, in step S12, determining the multidimensional state parameters based on the vehicle state data and driving state data may include the following execution steps:

[0060] Step S121: Determine the driver's state score based on steering torque, steering angular velocity, the percentage of time the driver's gaze is off-center from the lane center, the driver's blinking frequency, and the driver's head pitch angle.

[0061] Step S122: Determine the driving style coefficient based on the standard deviation of steering amplitude, following distance, and accelerator pedal opening change rate.

[0062] Step S123: Determine the environmental risk value based on road curvature, road slope, road adhesion coefficient, relative distance to obstacles, relative speed of obstacles, and weather conditions.

[0063] In this embodiment, the vehicle steering control system first determines the driver state score based on steering torque, steering angular velocity, the percentage of time the driver's gaze is off the center of the lane, the driver's blinking frequency, and the driver's head pitch angle.

[0064] Specifically, the vehicle steering control system collects five types of multimodal data in real time: steering torque, steering angular velocity, the percentage of time the driver's gaze deviates from the center of the lane, blink frequency, and head pitch angle. It then calculates a driver state score between 0 and 1, which represents the driver's current willingness to control the vehicle and the degree of concentration. Steering torque and angular velocity are used to quantify the intention to operate, the percentage of time the gaze deviates from the center of the lane reflects visual focus, and blink frequency and head pitch angle indicate physiological fatigue and the trend of distraction, respectively. The system integrates the above indicators into a single continuous value based on a preset weighted model, avoiding the risk of misjudgment from traditional single threshold judgments.

[0065] For example, in a real-world application scenario, when a driver is cruising at high speed and fatigue causes steering torque to drop below 2.5 N·m, angular velocity to drop below 60° / s, lane departure to reach 28%, blinking frequency to rise to 22 times / minute, and head tilt to 12°, the system calculates a state score of 0.544, indicating a moderate-risk state of attention deficit. Through cross-validation of physical operation signals and physiological behavioral signals, a highly robust assessment system for the driver's true state is constructed. This system not only identifies "whether the driver is operating" but also judges "whether the driver is operating attentively," thereby achieving a shift from "passive response" to "active prediction."

[0066] Therefore, by continuously, multidimensionally, and non-intrusively quantifying the driver's state, the system significantly improves the accurate identification of fatigue and distraction, avoids the lag that traditional systems only trigger when the driver is completely incapacitated, and avoids frequent takeover interference caused by misjudgment, making the system intervention more intelligent.

[0067] Then, the vehicle steering control system determines the driving style coefficient based on the standard deviation of steering amplitude, following distance, and rate of change of accelerator pedal opening.

[0068] For example, the vehicle steering control system analyzes three core behavioral features of the driver—steering amplitude standard deviation, following distance, and accelerator pedal opening change rate—from nearly 50 kilometers of historical driving data. Based on the K-means++ clustering algorithm, it automatically identifies and quantifies the driver's driving style and outputs a continuous driving style coefficient. The value ranges from 0.4 to 1.0, corresponding to three typical driving behavior modes: conservative, standard, and aggressive.

[0069] Among them, the standard deviation of steering amplitude is used to measure the frequency and intensity of steering wheel operation, the following distance reflects the driver's preference for the safety margin of the vehicle in front, and the rate of change of accelerator pedal opening characterizes the decisiveness and aggressiveness of the acceleration decision. The system normalizes the above three features and inputs them into the clustering model, automatically dividing them into three statistically significant driving behavior groups, and assigning corresponding style weights to each group, thereby realizing unsupervised modeling of the driver's long-term behavior patterns.

[0070] For example, in a real-world application scenario, if a driver drives continuously for 50km with a steering angle standard deviation of 6.2°, an average following distance of 15m, and an accelerator pedal opening change rate of 0.85% / s, the system will determine through clustering that the driver belongs to the aggressive style. Assigned a value of 1.0, the system then prioritizes the user's control intentions in subsequent human-machine collaborations. For example, in low- to medium-risk curves, the system proactively increases the control weight. The torque has been increased from 0.6 to 0.8, and the road feel stiffness and damping have been enhanced, making the steering response more direct and the steering wheel feedback clearer.

[0071] This enables adaptive recognition of driving style, overcoming the rigidity of traditional systems that use default styles or manual calibration. The system can automatically adapt without reconfiguration when different drivers switch, improving operational consistency and psychological safety.

[0072] Finally, the vehicle steering control system determines the environmental risk value based on road curvature, road slope, road adhesion coefficient, relative distance to obstacles, relative speed of obstacles, and weather conditions.

[0073] Specifically, the vehicle steering control system integrates six types of dynamic and static environmental parameters, including road curvature, road slope, road adhesion coefficient, relative distance to obstacles, relative speed of obstacles, and weather conditions, to construct a global, continuous, and multi-dimensional environmental risk assessment model. The final output is a comprehensive environmental risk between 0 and 1, which is used to characterize the potential threat level to driving safety posed by the current driving scenario.

[0074] Among these factors, road curvature reflects the sharpness of curves, slope affects vehicle longitudinal stability and braking performance, and the coefficient of adhesion is derived from road surface texture and wet / slip conditions, directly related to the tire's grip limit. Relative distance and relative speed to obstacles are used to calculate Time to Collision (TTC) to determine the existence of collision risk. Weather conditions are assessed through a combination of meteorological sensors and image recognition, quantifying the combined weakening effect of weather factors such as rain, fog, and snow on visibility and adhesion. The system organically integrates static road features, dynamic obstacle situations, and weather interference using a weighted fusion model trained by gradient descent, avoiding misjudgments based on a single dimension.

[0075] For example, when a vehicle is traveling at 30 km / h on a slippery urban road, and a pedestrian is crossing laterally at 5 m / s 20 meters ahead, with a road curvature of 0.08 rad / m, a slope of 3°, a road surface adhesion coefficient of 0.5, and moderate rain, the system calculates a static risk of 0.662 and a dynamic risk of 0.2. However, since the actual time-to-market (TTC) is 4 seconds, the system only saturates to 1 when the TTC is ≤ 1.5 seconds, resulting in a value of 0.2 in this case. The meteorological risk is 0.6R, and the final comprehensive environmental risk value Q_{\text{risk}}k = 0.4186. Based on this, the system determines the risk to be moderate and initiates assisted steering intervention. The technical logic of this step is to abandon the traditional "threshold-triggered" risk assessment and use a machine learning model to perform nonlinear weighted fusion of multi-source heterogeneous environmental parameters. This allows the risk assessment to not only reflect the "current danger" but also predict the "risk evolution trend," achieving a leap from "event response" to "situational awareness."

[0076] This enables dynamic quantification of three-dimensional environmental risks, including road, obstacle, and weather conditions, improving the robustness and accuracy of risk identification in complex scenarios. It ensures that the system respects the driver's control when the risk is low and intervenes promptly and smoothly when the risk is medium to high, avoiding "premature intervention" that leads to a sense of operational deprivation or "late response" that leads to accidents. This provides a reliable and quantifiable environmental basis for the dynamic allocation of control.

[0077] Optionally, in step S121, determining the driver state score based on steering torque, steering angular velocity, the percentage of time the driver's gaze deviates from the lane center, the driver's blink frequency, and the driver's head pitch angle may include the following steps:

[0078] Step S1211: Determine the driver's intention score based on steering torque and steering angular velocity.

[0079] Step S1212: Determine the driver's attention score based on the percentage of time the driver's gaze is off-center from the lane, the driver's blinking frequency, and the driver's head tilt angle.

[0080] Step S1213: Determine the driver's state score based on the operation intention score and attention score.

[0081] In this embodiment of the application, firstly, the vehicle steering control system determines the driver's intention score based on the steering torque and steering angular velocity. ).

[0082] For example, the vehicle steering control system collects the steering torque output by the steering wheel torque sensor and the steering angular velocity output by the steering wheel angle sensor in real time. Based on the strong correlation between physical action and intention expression, it quantifies the driver's current active control intention and defines it as an operation intention score. This score is a dimensionless index in the continuous interval [0,1], used to characterize whether the driver is actively initiating steering intention rather than passively following or relaxing.

[0083] For example, in a driving scenario, such as when a driver intends to change lanes at the entrance of a highway ramp, the steering wheel is quickly turned to 40° with an angular velocity of 72° / s, while a torque of 3.1 N·m is applied. The system immediately determines the score of the intention to change lanes. A score of 1 indicates a clear and explicit intention, and the system will accordingly increase the driver's control weight. When the driver makes slight directional corrections while cruising at a constant speed, applying only 1.2 N·m of torque and an angular velocity of 30° / s, the system will recognize the intended action as having a lower score. =0.36, which is considered a low-intensity correction. The system can moderately enhance assistance to reduce the operational burden, rather than actively taking over.

[0084] Therefore, this application upgrades the traditional "whether to turn" judgment to "whether there is an active and forceful intention to turn". Through the dual coupling verification of torque and angular velocity, it effectively distinguishes between "driver's active operation" and "system compensation disturbance" or "passive rotation caused by road interference", improves the anti-interference ability and semantic accuracy of intention recognition, and avoids false triggering caused by road bumps, wind resistance or EPS compensation.

[0085] Then, the vehicle steering control system determines the driver's attention score based on the percentage of time the driver's gaze is off-center from the lane, the driver's blinking frequency, and the driver's head pitch angle. ).

[0086] For example, the vehicle steering control system uses three non-invasive physiological behavioral characteristics collected by the driver monitoring system (DMS): the percentage of time the driver's gaze deviates from the center of the lane, the frequency of blinking, and the head pitch angle, to construct the driver's attention score, which is used to quantify the degree of cognitive focus during driving. This score is a dimensionless index in the continuous interval [0,1].

[0087] Specifically, the percentage of time the driver's gaze deviates from the center of the lane reflects whether their visual focus remains on the road, providing the most direct visual basis for judging distracted behavior. Blinking frequency serves as a reliable biosignal of physiological fatigue and inattention; in a normal waking state, the driver blinks approximately 15–20 times per minute, with more than 30 blinks often associated with fatigue or mental relaxation. Head pitch angle indicates whether the driver has deviated from their normal driving posture due to drowsiness, looking down at the road, or shifting their gaze; the system considers ±30° as the normal range, and anything exceeding this is considered abnormal posture. These three factors are comprehensively evaluated using a weighted fusion formula to ensure a balanced contribution from each dimension and prevent misjudgment dominated by a single feature.

[0088] For example, during long-distance driving at night, if a driver's gaze deviates from the center of the lane for 12 consecutive seconds, their blinking frequency increases to 36 times per minute, and their head tilts forward by 22°, the system calculates the driver's attention score. =0.324, indicating a severely impaired level of attention. The system could activate voice prompts and moderately enhance steering assist, but did not immediately take over. In congested urban areas, although the driver occasionally glanced down at the central control screen, the blinking frequency remained stable at 18 times per minute, and head posture was normal. The system calculated the driver's attention score accordingly. =0.71, which is considered moderate focus, and the system maintains normal assistance.

[0089] Therefore, this application models the physiological-behavioral consistency of three heterogeneous behavioral signals: gaze, physiological rhythm, and head posture. This overcomes the one-sidedness of traditional solutions that rely solely on gaze duration or the duration of eye closure. By dynamically weighted fusion, it eliminates misjudgments caused by single-point false detections (such as blinking in bright light or briefly looking down), achieving stable identification of "attention decay trends" and significantly improving the robustness and accuracy of fatigue and distraction detection.

[0090] Finally, the vehicle steering control system determines the driver's state score based on the intention score and attention score. ).

[0091] For example, the vehicle steering control system ultimately integrates the driver's operational intention score and attention score through linear weighting to generate a comprehensive score reflecting the driver's overall control state. This index is a continuous value in the range [0,1], used to quantify the driver's coordinated ability in two dimensions: "whether they want to take control" and "whether they can take control." The operational intention score represents the intensity of the driver's willingness to actively control, while the attention score represents the physiological basis for maintaining control. Both are indispensable—even if the driver's intention is strong, if their attention is severely scattered, they still lack the ability to safely take over. Conversely, if attention is good but there is no steering input, it indicates that the driver is in a passive following or over-dependent state, and the system needs to intervene appropriately.

[0092] For example, on a highway, a sober and focused driver prepares to change lanes; their intention score is... =0.95 (both torque and angular velocity exceed the threshold), attention score =0.88 (stable gaze, normal blinking, head upright), the system calculates the driver's condition score. =0.915, judged as "high-state priority control", the system immediately releases more than 90% of the control, only for safety redundancy monitoring; in congested urban areas, if the driver's attention has decreased to 0.35 due to prolonged driving (frequent blinking, gaze deviation), but the intention to actively lightly depress the steering wheel due to the vehicle in front slowing down is obvious, the system calculates the driver's state score. =0.525, which is judged as "moderate fluctuation". The system does not take over immediately, but guides it to regain focus through slight steering damping enhancement and auxiliary prompts, while being ready to intervene at any time.

[0093] Therefore, this application integrates the two dimensions of "subjective will" and "objective ability" into a unified decision variable, avoiding the illusion of "mistaking the person for driving" caused by relying solely on torque judgment, and also preventing "over-intervention" caused by relying solely on DMS fatigue detection, thus improving the scientific and humanistic level of control allocation.

[0094] Optionally, in step S122, determining the driving style coefficient based on the standard deviation of steering amplitude, following distance, and accelerator pedal opening change rate may include the following steps:

[0095] Step S1221: Determine the historical steering amplitude standard deviation, historical following distance, and historical accelerator pedal opening change rate based on the vehicle's historical status data.

[0096] Step S1222: Clustering algorithm is used to cluster the historical steering amplitude standard deviation, historical following distance, and historical accelerator pedal opening change rate to obtain multiple clustering results.

[0097] Step S1223: Determine the preset driving style type based on multiple clustering results, and establish a style coefficient mapping table based on the preset driving style type. The style coefficient mapping table is used to record the mapping relationship between the preset driving style type and the driving style coefficient.

[0098] Step S1224: Use a clustering algorithm to cluster the standard deviation of steering amplitude, following distance and accelerator pedal opening change rate to obtain at least one target clustering result.

[0099] Step S1225: Determine the driving style coefficient corresponding to at least one target clustering result based on the style coefficient mapping table.

[0100] In this embodiment, the vehicle steering control system first determines the historical steering amplitude standard deviation, historical following distance, and historical accelerator pedal opening change rate based on the vehicle's historical state data.

[0101] For example, the vehicle steering control system first continuously collects and stores historical state data of the vehicle during the most recent 50 kilometers of driving through the on-board domain controller, and extracts three key driving behavior characteristics from it: historical steering amplitude standard deviation. Historical following distance ( Historical accelerator pedal opening change rate ( These three indicators are used to characterize the driver's volatility in steering operations, spatial interaction habits with the vehicle in front, and the aggressiveness of power response. These three indicators form the basis for the quantitative expression of driving style.

[0102] Among them, the standard deviation of steering amplitude reflects the frequency and intensity of the driver's steering input between straight lines and curves. A high value indicates frequent lane changes or large-angle corrections, while a low value reflects smooth tracking. Following distance reflects the driver's preference for safety margins with the vehicle in front. Short distances are usually accompanied by high risk tolerance and a tendency to actively overtake, while long distances reflect caution and conservatism. The rate of change of accelerator pedal opening quantifies the driver's control rhythm of the accelerator. Steep changes represent frequent acceleration and deceleration, which is a typical characteristic of an aggressive style, while smooth changes correspond to gentle driving.

[0103] For example, the standard deviation of the steering amplitude of an aggressive driver who continuously changes lanes and frequently overtakes on a highway is... 6.8°, following distance remained stable within 15 meters, accelerator pedal opening increased from 30% to 85% within 1 second ( =55% / s); Conversely, a conservative driver in urban areas maintains a stable lane, a distance of more than 40 meters from the vehicle in front, steering input fluctuations of less than 2.5°, and smooth accelerator pedal changes ( =0.8% / s).

[0104] Therefore, the traditional crude method of "user preset style" or "fixed template matching" is abandoned, and unsupervised automatic recognition of driving style based on long-term behavioral statistics is realized in mass-produced steer-by-wire systems.

[0105] Secondly, the vehicle steering control system uses a clustering algorithm to cluster the historical steering amplitude standard deviation, historical following distance, and historical accelerator pedal opening change rate, resulting in multiple clustering results.

[0106] For example, the vehicle steering control system first continuously collects and stores historical state data of the vehicle during the most recent 50 kilometers of driving through the on-board domain controller, and extracts three key driving behavior characteristics from them: historical steering amplitude standard deviation, historical following distance, and historical accelerator pedal opening change rate. These are used to characterize the driver's volatility in steering operations, spatial interaction habits with the vehicle in front, and the aggressiveness of power response, respectively. These three indicators constitute the basis for the quantitative expression of driving style.

[0107] Therefore, by uncovering the implicit patterns in real driving behavior data, a data foundation was provided for the subsequent establishment of a style coefficient mapping table.

[0108] Then, the vehicle steering control system determines the preset driving style type based on multiple clustering results, and establishes a style coefficient mapping table based on the preset driving style type. The style coefficient mapping table is used to record the mapping relationship between the preset driving style type and the driving style coefficient.

[0109] For example, the vehicle steering control system collects continuous driving behavior data (including steering amplitude standard deviation, following distance, accelerator pedal opening change rate, and other multi-dimensional features) from the driver over approximately 50 kilometers. Using the K-means++ clustering algorithm, it automatically identifies driving pattern groups with significant behavioral differences. These data clusters are then categorized into three preset driving style types—aggressive, standard, and conservative—rather than relying on subjective labels or manual settings, thus achieving objective quantification and automated classification of driving styles. Subsequently, the system assigns a corresponding style coefficient to each clustered style type, which can be directly invoked by the control algorithm. (e.g., aggressive type = 1.0, standard type = 0.7, conservative type = 0.4), and construct this into a structured, storable, and updatable mapping table, which serves as the input basis for subsequent dynamic weight allocation and road feel adjustment modules. This process transforms fuzzy human driving preferences into calculable and reusable mathematical parameters, enabling the system to "remember" the unique style of each driver and intelligently adapt to different scenarios.

[0110] For example, when the system identifies that a driver has a long-term preference for large-angle and rapid lane changes and close following distances, it automatically classifies the driver as aggressive and enables a high weighting coefficient, making the system more inclined to relinquish control in cooperative control. When the system identifies that the driver is accustomed to smooth deceleration and maintaining a safe distance, it classifies the driver as conservative and reduces its control weight, while the road feel feedback system provides a softer steering resistance.

[0111] This breaks away from the rigidity of traditional "one-size-fits-all" style settings, improves the personalization and accuracy of control strategies, and avoids control conflicts or operational discomfort caused by style misjudgment.

[0112] Subsequently, the vehicle steering control system uses a clustering algorithm to cluster the standard deviation of steering amplitude, following distance, and accelerator pedal opening change rate to obtain at least one target clustering result.

[0113] For example, the vehicle steering control system's standard deviation of steering amplitude ( Following distance () ) and accelerator pedal opening change rate ( These three continuous behavioral parameters reflecting driving style are subjected to K-means++ clustering to automatically identify and classify driver groups with similar driving patterns, thus achieving objective quantification of driving style without the need for manual rule setting.

[0114] This allows for precise identification of driver style (aggressive / standard / conservative), enabling personalized weighting and road feel adjustment, thereby improving the matching degree of human-machine collaboration intent and operational comfort.

[0115] Finally, the vehicle steering control system determines the driving style coefficient corresponding to at least one target clustering result based on the style coefficient mapping table.

[0116] Thus, through multimodal perception and dynamic weight adjustment, the human-machine co-driving steering control is adapted to the three objectives of safety, intention compliance and comfort, improving the system response accuracy, operation smoothness and driver satisfaction, and solving the problems of poor adaptability and insufficient coordination of traditional static control strategies.

[0117] Optionally, in step S123, determining the driving style coefficient based on the standard deviation of steering amplitude, following distance, and accelerator pedal opening change rate may include the following steps:

[0118] Step S1231: Determine the static risk coefficient based on road curvature, road slope, and road adhesion coefficient.

[0119] Step S1232: Determine the collision time based on the relative distance and relative speed of the obstacles, and determine the dynamic risk coefficient based on the collision time.

[0120] Step S1233: Determine the weather conditions based on the vehicle-mounted meteorological sensors and vehicle-mounted cameras, and determine the meteorological risk coefficient based on the weather conditions.

[0121] Step S1234: Determine the environmental risk value based on the static risk coefficient, dynamic risk coefficient, and meteorological risk coefficient.

[0122] In this embodiment, the vehicle steering control system first determines the static risk coefficient based on the road curvature, road slope, and road adhesion coefficient.

[0123] Static risk coefficient These are core parameters used by vehicle steering control systems to assess the inherent hazard level of a road; for example, they can be identified and fused from the road curvature by a forward-facing camera. ,slope Adhesion coefficient derived from road surface texture characteristics The comprehensive calculation of the three types of static environmental characteristics reflects the potential risk of loss of control, which is determined solely by road geometry and physical conditions and does not depend on dynamic obstacles.

[0124] For example, on a mountain road section with a curve curvature of 0.03 rad / m, a slope of 5°, and a road surface adhesion coefficient reduced to 0.3 due to snow accumulation, the static risk coefficient is calculated to be 1.612. After normalization, it is mapped to the interval [0,1] to obtain... The value is approximately 0.87, indicating that this road section has an extremely high static risk.

[0125] Therefore, without relying on real-time radar or lidar, the inherent risks of the road can be obtained at low cost and with low latency using only existing sensors, effectively making up for the blind spot of traditional systems that only focus on dynamic obstacles and ignore the "danger of the terrain itself".

[0126] Secondly, the vehicle steering control system determines the collision time based on the relative distance and relative speed of the obstacle, and then determines the dynamic risk coefficient based on the collision time. .

[0127] For example, dynamic risk It can be based on lidar The relative distance to surrounding obstacles (vehicles / pedestrians) identified by millimeter-wave radar. Relative velocity Calculate collision time ,formula ( ).

[0128] Time-to-Collision (TTC) refers to the time required for a collision to occur between the current vehicle and an obstacle ahead (such as another vehicle or pedestrian) while maintaining their current relative motion. TTC directly reflects the urgency of a potential collision and is a core input for dynamic risk assessment. As a physical quantity with a time dimension, TTC more closely resembles a human driver's intuitive judgment of an impending collision, enabling the system to identify high-risk situations in advance and initiate strong intervention when TTC is ≤1.5 seconds. This effectively reduces decision-making delays and avoids collisions caused by delayed reactions.

[0129] For example, in a congested urban area, when a vehicle is traveling at 30 km / h and a pedestrian suddenly crosses the road ahead, the relative distance is measured to be 20 meters and the relative speed is 5 m / s (the pedestrian is moving towards the vehicle), then TTC = 4.0 seconds, at which point the system determines it to be a low to medium risk; if the pedestrian accelerates and approaches, the relative distance decreases to 10 meters and the relative speed increases to 10 m / s, then TTC = 1.0 seconds, the system immediately determines it to be an extremely high risk, and triggers the emergency takeover mechanism.

[0130] Therefore, by capturing the position and speed of obstacles in real time and performing dynamic calculations in conjunction with the vehicle's own motion state, the system avoids misjudgments that rely solely on static distance or speed thresholds, significantly improving the accuracy of response to sudden dangers.

[0131] Then, the vehicle steering control system determines the weather conditions based on the onboard weather sensors and onboard cameras, and determines the weather risk factor (R_weather) based on the weather conditions.

[0132] The vehicle steering control system uses an onboard meteorological sensor and an onboard camera to collaboratively perceive the current weather conditions. For example, the onboard meteorological sensor collects physical parameters such as ambient humidity, rainfall, and temperature in real time (humidity accuracy ±2%, rainfall resolution 0.1 mm / h), while the onboard camera uses an image recognition algorithm to semantically classify the visibility and precipitation patterns (such as raindrops, snowflakes, and fog) of the current environment. Combining the data from both, the system uses a pre-trained gradient descent model to map the weather conditions into a continuous meteorological risk coefficient, with a value ranging from 0.1 (sunny) to 0.9 (heavy snow).

[0133] For example, under light rain conditions, the rain sensor detects a precipitation intensity of 0.5 mm / h, while the camera identifies increased road surface reflectivity and continuous wiper operation. The system comprehensively determines this as a "light rain" state, corresponding to R_weather=0.3. However, in moderate rain or heavy fog conditions, if the rainfall exceeds 2 mm / h and the image recognition determines that the visibility is less than 50 meters, the system automatically increases the risk coefficient to 0.6, thereby significantly increasing the overall environmental risk value and triggering the system's enhanced intervention strategy.

[0134] Thus, by using multimodal data fusion, the shortcomings of a single camera failing to recognize objects at night or in strong light, or the inability of weather sensors to distinguish between rain and fog types, are effectively avoided. This achieves a highly robust quantitative assessment of weather risks and provides key support for safe human-machine collaborative driving in complex weather conditions.

[0135] Finally, the vehicle steering control system determines the environmental risk value based on the static risk coefficient, dynamic risk coefficient, and meteorological risk coefficient.

[0136] The environmental risk value (Q_risk) is a comprehensive assessment index obtained by weighted fusion of three types of risk factors in the vehicle's driving environment: static, dynamic, and meteorological. It is used to dynamically reflect the degree of safety threat in the current driving scenario. For example, the three factors are weighted and fused by a gradient descent training model to form the final environmental risk value Q_risk.

[0137] For example, on a straight road in clear weather, even with slow-moving vehicles, if the curvature is extremely low and there is no precipitation, the overall risk remains at a low to medium level of Q_risk=0.35, ensuring that the system does not over-intervene. Conversely, on sharp curves in rainy or foggy weather, the combination of these three factors can cause Q_risk to soar to 0.85, triggering the system to take over control and avoiding misjudgment based on a single dimension.

[0138] Therefore, by collaborative modeling of multidimensional physical quantities, risk assessment is made adaptable to the environment and generalizable to different scenarios, which improves the predictability and robustness of control strategies. It avoids the crude judgment that relies solely on distance or curvature in the past, enabling the system to make better decisions in advance in complex urban scenarios and effectively reduce the probability of false triggering and missed triggering.

[0139] Optionally, in step S13, determining the weight parameters based on the multidimensional state parameters, preset rules, and reinforcement learning model may include the following steps:

[0140] Step S131: Determine whether the vehicle is in a preset scenario based on the multi-dimensional state parameters.

[0141] Step S132: In response to the vehicle being in a preset scenario, determine the weight parameters based on multi-dimensional state parameters and preset rules.

[0142] Step S133: In response to the vehicle not being in the preset scenario, the multi-dimensional state parameters are processed using a reinforcement learning model to obtain the initial weight parameters.

[0143] Step S134: Correct the initial weight parameters according to the driving style coefficient to obtain the weight parameters.

[0144] In this embodiment, the vehicle steering control system first determines whether the vehicle is in a preset scenario based on multi-dimensional state parameters.

[0145] For example, the vehicle steering control system first constructs a dynamic perception scenario judgment mechanism by integrating three types of multi-dimensional state parameters: driver state score, comprehensive environmental risk value, and driving style weighting coefficient. This mechanism is used to identify whether the current driving situation falls into a preset extreme or normal scenario category. For instance, when the system detects that the driver's gaze is off-center for more than 70% of the time, blinks 40 times per minute, and has a head pitch angle exceeding 25°, combined with a pedestrian collision time (TTC) of 1.0 second measured by lidar, moderate rain detected by weather sensors, and historical driving data clustering indicating that the driver has a conservative driving style, P_driver=0.25 and Q_risk=0.91. The system then determines this to be a preset extreme scenario of "extremely high risk + severe driver inattention".

[0146] As a result, the static judgment that originally relied on a single threshold has been upgraded to a dynamic contextual understanding based on multimodal perception, spatiotemporal coupling and style adaptation, which improves the safety redundancy in complex and sudden scenarios.

[0147] Secondly, when the vehicle is in a preset scenario, the vehicle steering control system determines the weight parameters based on multi-dimensional state parameters and preset rules.

[0148] For example, based on real-time perception of multi-dimensional dynamic information such as driver status (e.g., attention level and operational intention), environmental risks (e.g., road curvature, weather conditions, obstacle distance), and driving style (aggressive, standard, or conservative), the system intelligently calculates the responsibility allocation ratio between the driver and the autonomous driving system in steering control, i.e., the driver's control weight, by integrating expert rules and reinforcement learning models. With system control weights And the sum of the two is always 1.

[0149] For example, on a highway, an aggressive driver is traveling at a constant speed of 100 km / h with a slight intention to change lanes. The environment is light rain with no sudden obstacles. The system first determines that the scenario does not belong to an extreme risk or disability state, enters the DDPG network decision-making process, inputs the state vector, and outputs the basic weights through the Actor network. =0.6 =0.4. This achieves a dynamic balance between fully respecting the driver's intentions and ensuring system safety. Compared to traditional fixed rights or single threshold triggering mechanisms, it effectively avoids the sense of control deprivation caused by "forced system takeover" or the safety hazards caused by "driver misoperation without intervention".

[0150] Then, when the vehicle is not in the preset scenario, the vehicle steering control system uses a reinforcement learning model to process the multi-dimensional state parameters to obtain the initial weight parameters.

[0151] For example, when the vehicle is not in a preset extreme or expert rule-mandated intervention scenario (such as a risk value Q_risk below 0.8 and a driver state score P_driver above 0.3), the steer-by-wire cooperative control system will activate an autonomous decision-making mechanism based on deep reinforcement learning. At this time, the system no longer relies on hard threshold judgments, but instead integrates the driver state score. Comprehensive environmental risks And driving style weight These three types of continuously and dynamically changing multidimensional state parameters are input into a pre-trained DDPG (Deep Deterministic Policy Gradient) neural network, from which its Actor network outputs a continuous initial driver control weight. The weight changes smoothly between 0 and 1, reflecting the system's tendency to judge whether the current driving situation should be "human-led" or "system-assisted".

[0152] For example, in a typical scenario involving driving on a highway curve, light rain, focused driver attention, moderate risk, and an aggressive driving style, the network might output an initial... =0.65 indicates that the system tends to allow the driver to take the lead in steering operations, but retains 35% of the system's intervention capability to deal with potential risks. Thus, through long-term online training, the reinforcement learning model learns how to balance the reward functions of four categories of objectives: "safety," "intentional compliance," "comfort," and "personalization," thereby autonomously evolving the optimal control allocation strategy, rather than relying on static formulas preset by engineers.

[0153] Finally, the vehicle steering control system corrects the initial weight parameters based on the driving style coefficient to obtain the weight parameters.

[0154] In this application, the driving style coefficient ( This is a quantitative representation of a driver's long-term driving behavior characteristics. For example, by clustering analysis of historical steering amplitude, following distance, and acceleration characteristics, drivers are categorized into aggressive, standard, or conservative types. This coefficient reflects their preference for control—aggressive drivers tend to be more proactive in their driving, while conservative drivers rely more on system assistance. The basic control weights are output in the hybrid reinforcement learning dynamic weight allocation layer. Then, the system will adjust the weights based on the matching relationship between real-time environmental risk and driving style: when an aggressive driver is detected in a low-to-medium risk scenario (0.2≤Q_risk<0.5), the system will adjust the weights in the DDPG output. An offset of +0.1 to +0.2 is added on top of the existing offset to enhance the driver's control and match their psychological expectation of "actively intervening and controlling the road feel".

[0155] Conversely, when a conservative driver is in a medium-to-high risk scenario (Q_risk≥0.5), the system proactively reduces [the risk level]. The value is increased from 0.1 to 0.2 to enhance system dominance and reduce the risks caused by hesitation or misoperation. For example, in light rain on a highway, an aggressive driver intends to change lanes, and the DDPG system initially outputs... =0.6, which was increased to 0.7 after style correction.

[0156] Therefore, by dynamically adjusting the control weights using a driving style coefficient, the system can accurately identify and adapt to the behavioral inertia and psychological expectations of different drivers.

[0157] Optionally, in step S14, determining the target steering command based on the weighting parameters, driving style coefficient, environmental risk value, system-planned steering angle, and steering wheel angle may include the following execution steps:

[0158] Step S141: Determine multiple target weight coefficients based on driver control weights, environmental risk values, and driving style coefficients.

[0159] Step S142: Determine the target steering command based on multiple target weight coefficients, system planning steering angle, steering wheel steering angle, and constraints. The constraints include steering angle constraints and steering rate constraints.

[0160] In this embodiment of the application, firstly, the vehicle steering control system determines multiple target weight coefficients based on the driver control weight, environmental risk value, and driving style coefficient.

[0161] For example, the system first integrates the dynamic state space composed of driver control weights, comprehensive environmental risk values, and driving style coefficients, and then calculates the three key weight coefficients in the multi-objective optimization function through preset mathematical relationships: .in, It reflects the priority of system control commands, and its value increases as environmental risks increase and driver control decreases, ensuring that the system takes the lead in steering to ensure safety in high-risk scenarios; It reflects the degree of compliance with the driver's operating intentions, which increases with the driver's weight and style becoming more aggressive, enabling the system to prioritize responding to the driver's steering needs when there is low risk and the driver's intentions are clear; The smoothness of the steering process is adjusted and significantly increased only in low-risk scenarios to suppress sudden changes in steering angle rate and improve ride comfort.

[0162] For example, in light rain on a highway, an aggressive driver changes lanes proactively, and the system detects Q_risk=0.4. =0.7, calculated as follows =0.38 (medium system priority) =0.79 (Highly Compliant Driver) =0.3 (moderate smoothing), the optimizer generates steering commands that are biased towards the driver's intention rather than the system's recommended path, while ensuring that the steering rate changes smoothly without any jerking.

[0163] Therefore, by dynamically adjusting the weights of the objectives through multi-dimensional parameter linkage, the control strategy is upgraded from an "either / or" switching logic to a flexible decision-making mechanism of "multi-objective collaborative trade-offs," effectively solving the problems of "system power grab" or "driving deprivation" caused by fixed weights.

[0164] Then, the vehicle steering control system determines the target steering command based on multiple target weight coefficients, the system-planned steering angle, the steering wheel angle, and constraints. The constraints include steering angle constraints and steering rate constraints.

[0165] For example, the vehicle steering control system uses dynamically calculated target weighting coefficients—namely, safety priority weights. Driver's intention to comply weight Weight of steering smoothness — Combining the desired steering angle planned by the autonomous driving system and the actual steering angle input by the driver after driving style correction, the optimal target steering command is solved by constructing a constrained multi-objective optimization function. This process is completed by an interior point method solver every 10 milliseconds to ensure real-time performance and accuracy.

[0166] Among them, the steering angle constraint condition (Steering-by-wire mechanical limits), ensuring that the mechanical limits of the steering-by-wire system are not exceeded, steering rate constraint conditions. (Comfort-type constraints) limit the steering response rate to avoid passenger motion sickness or system overshoot caused by excessively fast steering, thereby establishing a physical boundary between safety and comfort.

[0167] Therefore, through mathematical modeling of dynamic weights, the three mutually constraining goals of "safety", "intent" and "comfort" are unified into a quantifiable and solvable optimization problem.

[0168] Optionally, steering control of the vehicle based on the target steering command, driving style coefficient, and rate of change of weight parameters may include the following execution steps:

[0169] Step S16: Use a preset closed-loop control strategy to perform closed-loop control on the target steering command to obtain the initial execution torque. The preset closed-loop control strategy includes a position closed-loop control strategy, a torque closed-loop control strategy, and a current closed-loop control strategy.

[0170] Step S17: Determine the transition function based on the rate of change of the weight parameters and the initial execution torque, and determine the transition torque based on the transition function and the initial execution torque.

[0171] Step S18: Determine the road feel compensation torque of the road feel feedback motor based on the transition torque, driving style coefficient, and road feel simulation rules. The road feel simulation rules are used to determine the corresponding road feel compensation torque based on different driving style coefficients.

[0172] Step S19: Determine the target execution torque based on the road feel compensation torque and the transition torque, and perform steering control on the vehicle based on the target execution torque.

[0173] In this embodiment, the vehicle steering control system first uses a preset closed-loop control strategy to perform closed-loop control on the target steering command to obtain the initial execution torque. The preset closed-loop control strategy includes a position closed-loop control strategy, a torque closed-loop control strategy, and a current closed-loop control strategy.

[0174] For example, the vehicle steering control system uses a three-stage cascade control structure of "position closed loop—torque closed loop—current closed loop" to execute the target steering command δ_cmd output by the multi-objective optimization module with high precision and low latency. First, the position closed loop uses a sliding mode controller, with δ_cmd as the reference input, to calculate the required target output torque T_pos in real time, effectively dealing with uneven road surfaces and wind disturbances. Then, the torque closed loop uses a high-precision torque sensor (2kHz sampling) to feed back the actual motor output torque, and dynamically corrects the deviation between T_pos and the measured value through a proportional-integral controller, ensuring that the control error is stable within ±0.3N. Within a range of m, precise mapping from command to torque is achieved. Finally, the current closed loop employs space vector pulse width modulation technology to adjust the three-phase current waveform of the permanent magnet synchronous motor in real time based on the current command output by the torque loop, compressing the steering angle tracking error to within 0.05°, thus completing the final conversion from digital command to physical action.

[0175] Therefore, this pre-set closed-loop control strategy, through structured hierarchical control and high-frequency response mechanism, efficiently, smoothly and reliably transforms high-level decision-making instructions into physical steering actions. It is the key guarantee for realizing the entire closed-loop of human-machine collaborative steering "perception-decision-execution". It solves the engineering bottlenecks of traditional drive-by-wire systems such as response lag, poor anti-interference, and control drift, and provides a solid and reliable underlying execution foundation for advanced intelligent driving.

[0176] Secondly, the vehicle steering control system determines the transition function based on the rate of change of the weight parameters and the initial execution torque, and determines the transition torque based on the transition function and the initial execution torque.

[0177] The transition function is a torque smoothing transition strategy that is adaptively selected based on the scenario type determined in real time according to the rate of change of control weight (dω / dt). It is a nonlinear function designed to deal with torque mutations that may be caused during the dynamic handover of control between the driver and the system, aiming to eliminate the "hand-kickback" phenomenon and improve the comfort and safety of human-machine collaboration.

[0178] For example, the rate of change of the weight parameter dω / dt reflects the severity of the control switch. If the change is slow (such as a driver style switch or a gradual takeover), the system judges it as a normal transition scenario and adopts an S-curve transition function. This function is based on an error function and has the characteristics of being gentle at the beginning and end and smooth in the middle, so that the torque smoothly transitions from the initial value T0 to the target value within 0.4 seconds, with the abrupt change controlled within 0.8 N·m. If the change is drastic (such as an emergency takeover by the system or driver incapacitation), the system judges it as an emergency transition scenario and activates a linear transition function with feedforward compensation. This function actively predicts and compensates for the dynamic demand brought about by the sudden change in control by introducing a feedforward term of the weight change rate, so that the torque completes a rapid response within 0.2 seconds, with the maximum abrupt change controlled within 1.2 N·m. This satisfies the timeliness of emergency avoidance and avoids violent steering wheel vibration caused by applying torque too quickly.

[0179] Therefore, by taking the "dynamic intensity" of the control handover as the core criterion for the selection of the transition function, intelligent adaptive smooth control of "gradual change and rapid response" is realized. This not only reduces torque mutations in normal scenarios and compresses response delays in emergency scenarios, but also realizes intelligent smooth adjustment of torque changes during the handover of control. While ensuring the speed of emergency response, it eliminates "hitting" and jerking, making the system takeover no longer abrupt.

[0180] Then, the vehicle steering control system determines the road feel compensation torque of the road feel feedback motor based on the transition torque, driving style coefficient, and road feel simulation rules. The road feel simulation rules are used to determine the corresponding road feel compensation torque based on different driving style coefficients.

[0181] Among them, the road feel compensation torque is generated by combining the transition torque with the driving style coefficient and the road feel simulation rules to dynamically generate the feedback torque acting on the steering wheel. This is used to simulate the "road feel" experience of a real mechanical steering system, allowing the driver to perceive changes in vehicle status and environment through force feedback even in a drive-by-wire system without mechanical connection, thereby establishing trust and control over vehicle behavior.

[0182] For example, the road feel simulation rules set the reference ratios of the damping coefficient and stiffness coefficient according to the driving style coefficients (aggressive type 1.0, standard type 0.7, conservative type 0.4). The damping coefficient affects the response intensity of the steering resistance as speed changes, while the stiffness coefficient determines the linear relationship between the steering angle and the feedback torque. After obtaining the transition torque T_trans, the system adjusts the output parameters of the feedback motor in real time according to the driving style coefficient.

[0183] For example, in the aggressive style, the damping coefficient is increased to 1.5 times the baseline value and the stiffness coefficient is increased to 1.2 times, making the steering wheel more resistant and providing clearer feedback during turns. The driver can clearly perceive the road texture and steering load, matching their psychological expectation of pursuing driving pleasure. In the standard style, both damping and stiffness remain at the baseline values, providing balanced force feedback while balancing comfort and information transmission. In the conservative style, both damping and stiffness are reduced to 0.8 times, resulting in light and soft steering, reducing physical exertion and enhancing driving safety.

[0184] Thus, by coupling the driving style coefficient with the transition torque, a differentiated, personalized, and dynamically responsive road feel compensation mechanism is constructed.

[0185] Finally, the vehicle steering control system determines the target execution torque based on the road feel compensation torque and the transition torque, and then performs steering control on the vehicle based on the target execution torque.

[0186] For example, the vehicle steering control system vector-superimposes the dynamically smoothed transition torque with the road feel compensation torque customized based on driving style to form the final target torque.

[0187] Thus, by integrating transition torque and stylized road feel compensation torque to generate target execution torque, the unity of control logic and human-machine experience is achieved.

[0188] This application aims to specifically address the following three core pain points in human-machine co-driving steering control in traditional solutions:

[0189] 1. Break through the limitations of static rules to achieve precise matching of control with the driver's real-time state (attention, intention), style (aggressive / standard / conservative) and environmental risks (weather, traffic).

[0190] 2. Construct a multi-objective optimization framework that takes into account "driving safety (lane keeping), compliance with operating intentions (driver instruction tracking), and ride comfort (smooth steering)" to eliminate jerking and "hand-hitting" phenomena caused by instruction conflicts.

[0191] 3. Implement dynamic binding of motor control parameters (closed-loop gain, transient function) with scenario risks and driving style, taking into account both emergency scenario response and smooth operation in normal scenarios.

[0192] Therefore, another implementation of this application proposes a "four-layer progressive" system architecture, the core of which includes a multimodal perception and feature extraction layer, a hybrid reinforcement learning dynamic weight allocation layer, a multi-objective collaborative optimization control layer, and a scene adaptive execution control layer.

[0193] I. Multimodal Perception and Feature Extraction Layer (Input Layer)

[0194] By fusing multiple sensors, driver status scores, environmental risk values, and style coefficients are output, providing accurate input for subsequent decision-making.

[0195] ① Driver status:

[0196] 1. Operational Intent ( Based on torque sensor (accuracy) The steering torque T_δ collected by the sensor and the steering wheel angle sensor (resolution) The collected steering angular velocity Quantifying the current state through fuzzy logic and hour, (The intention is clear); conversely, reduce proportionally ( ).

[0197] 2. Attention ( Based on driver monitoring system camera ( ), monitoring the percentage of time the line of sight deviates from the center of the lane blinking frequency and head tilt angle ,formula: .

[0198] 3. Driver's overall score : ( A higher value indicates a better driver condition.

[0199] ② Driving style :

[0200] Based on nearly 50km of historical data (steering amplitude standard deviation) Following distance accelerator pedal opening change rate K-means++ clustering was used to classify it as radical ( ), standard type ( ), conservative type ( ), assign style weight coefficients

[0201] ③ Environmental risks :

[0202] A three-dimensional risk fusion approach combining static, dynamic, and meteorological data is employed, and a gradient descent model trained on 100,000 sets of scene data is used to calculate the comprehensive risk.

[0203] 1. Static Risks : by forward-facing camera Identify and integrate road curvature ,slope Adhesion coefficient derived from road surface texture characteristics ,formula: .

[0204] 2. Dynamic Risk Based on lidar The relative distance to surrounding obstacles (vehicles / pedestrians) identified by millimeter-wave radar. Relative velocity Calculate collision time ,formula: ( ).

[0205] 3. Weather risk (R_weather): Weather is identified by vehicle-mounted weather sensors (humidity accuracy ±2%, rainfall resolution 0.1mm / h) and camera images (sunny=0.1, light rain=0.3, moderate rain / fog=0.6, heavy snow=0.9).

[0206] 4. Overall Risk (Q_risk): The overall risk value is calculated using a weighted model trained with gradient descent. ( (The higher the value, the higher the risk).

[0207] II. Hybrid Reinforcement Learning Dynamic Weight Allocation Layer (Decision Layer)

[0208] It adopts a hybrid architecture of "reinforcement learning adaptive decision-making + expert rule safety backup" to dynamically output the driver control weight ω_driver and the system control weight. Balancing personalization in regular scenarios with security in extreme scenarios:

[0209] 1. State Space Input status .

[0210] 2. Motion space Output action (Discretization values, range 0~1, step size 0.05) .

[0211] 3. Reward Function Design A multi-objective reward function that balances safety, comfort, and personalization:

[0212]

[0213] in: Lane departure amount ( ); ΔT is the magnitude of the sudden change in steering torque (N) m, target ≤ 1N m); Steering smoothness (based on the second derivative of the steering angle, formula) ), weighting coefficient (Safety), (Comfort) (Personalization) (Smoothed) (Optimized by reinforcement learning training).

[0214] Hybrid decision-making algorithm:

[0215] 1. Deep reinforcement learning networks: responsible for decision-making in routine scenarios. Network (3 fully connected layers, 3 input dimensions, 1 output dimension, ReLU activation function) output The network updates its Q-value strategy based on a reward function, iterating every 20ms.

[0216] 2. Expert rule base handles extreme scenarios:

[0217] a. Extremely high risk: (like → Forced (Full system control).

[0218] b. Driver incapacity: (like (i.e., severe distraction) → forced .

[0219] c. Low risk + favorable condition: →Forced (The system is only for auxiliary monitoring).

[0220] d. Style adaptation and adjustment:

[0221] radical style Low to medium risk →For reinforcement learning output Offset.

[0222] conservative style Medium to high risk → For reinforcement learning output Offset.

[0223] III. Multi-objective collaborative optimization control layer (optimization layer)

[0224] Based on control weights and scene parameters, the optimal instruction δ_cmd is solved using a constrained multi-objective optimization solver.

[0225] Command acquisition and correction:

[0226] 1. System Commands Based on the Model Predictive Control (MPC) algorithm, the expected path in the next 3 seconds (high-precision map and GPS positioning, accuracy ±0.1m) is used as a reference, and the model is generated in combination with the vehicle's three degrees of freedom model.

[0227] 2. Driver's instructions Data collected by the steering wheel angle sensor, plus style correction amount. It adapts to different user habits.

[0228] Dynamic weighted multi-objective optimization function:

[0229] With the goal of "minimizing system command deviation, driver command deviation, and sudden changes in steering rate", the formula is: in, To dynamically adjust the weights, the calculation logic is as follows:

[0230] (The higher the risk, the greater the system weight) The larger the size, the stronger the security.

[0231] (The higher the driver's weight and the more aggressive their style,) The larger it is, the more it conforms to the intention.

[0232] 3) (In low-risk scenarios, comfort is the priority) (Increases, suppresses sudden changes in steering rate).

[0233] Constraints:

[0234] 1> Steering angle limits: (Steering-by-wire mechanical limits).

[0235] 2> Steering Rate Limit: (Comfort-type constraints).

[0236] Solver: Employs interior-point method to solve constrained optimization problems, outputting optimal steering angle command every 10ms. Compared with reference torque .

[0237] IV. Scenario-Adaptive Execution Control Layer (Execution Layer)

[0238] Based on optimized instructions, precise and smooth steering is achieved through three-loop control and scenario-based strategies, while adapting to the road feel requirements of different driving styles.

[0239] Three-loop control:

[0240] 1. Position loop: A sliding mode controller is used to... To achieve the target, output the target torque. Its anti-interference performance is 40% better than that of PID.

[0241] 2. Torque loop: Based on feedback from the steering wheel torque sensor (response frequency 2kHz), the torque deviation is adjusted via a proportional-integral controller to minimize error. .

[0242] 3. Current Loop: Employs Space Vector Pulse Width Modulation (SVPWM) technology to adjust the phase current of the permanent magnet synchronous motor in real time, reducing steering angle tracking error. .

[0243] Piecewise smooth torque transition: Determine the scene type based on the weight change rate dω / dt and select the corresponding transition function:

[0244] 1. Regular transition ( (e.g., style switching): Use an S-curve transition function. The transition time is 0.4s, and the torque surge is ≤0.8N. m.

[0245] 2. Emergency transition ( (e.g., emergency takeover): employs a linear transition function with feedforward compensation. ( (feedforward coefficients) Transition is completed internally, mutation There was no "thugs" involved.

[0246] Stylized road feel simulation: via a road feel feedback motor (rated torque 5N) m, precision 0.1N m) Dynamically adjust parameters:

[0247] 1. Aggressive Style (S1): Damping Coefficient = 1.5 × Reference Value (2.0 N) m / (° / s)), stiffness coefficient = 1.2 × reference value (1.5N) (m / °), clear road feel and strong feedback.

[0248] 2. Standard Style (S2): Damping coefficient = 1.0 × reference value, stiffness coefficient = 1.0 × reference value, balancing feedback and comfort.

[0249] 3. Conservative style (S3): Damping coefficient = 0.8 × reference value, stiffness coefficient = 0.8 × reference value, soft road feel and low resistance.

[0250] An embodiment of this application can achieve the following beneficial effects:

[0251] 1. More precise and personalized weight allocation: By integrating three-dimensional data of driver status, environment and driving style, the fixed weights improve adaptability. Experiments have shown that the accident avoidance rate is increased by 50% and driver operation satisfaction is increased by 45% in extreme scenarios.

[0252] 2. Command fusion with multi-objective balance: Innovative multi-objective optimization function avoids the limitations of single priority judgment, controls lane deviation within ±0.1m, and has an operation intention response delay of <10ms, balancing safety and driving experience.

[0253] 3. Differences in motor control adaptation scenarios: The combination of three closed-loop control and piecewise transient function compresses the steering angle response delay to 12ms in emergency scenarios, while the torque mutation amplitude is ≤1N in normal scenarios. m, reduces steering wheel vibration by 70%, solving the "kickback" problem.

[0254] 4. Enhanced compatibility and scalability: Reuses existing sensors, adds style recognition and weather perception, supports OTA upgrades of reinforcement learning models, adapts to different brands of steer-by-wire systems, and reduces mass production adaptation costs by 35%.

[0255] The following is an implementation example of this application:

[0256] (a) Hardware Architecture

[0257] The hardware of this system consists of a sensing unit, a decision-making unit, an execution unit, an interaction unit, and a communication unit, as shown in Table 1.

[0258] Table 1

[0259]

[0260] (II) Software Flow

[0261] 1. System initialization (completed within 1 second after power-on, ready before vehicle starts):

[0262] Sensor calibration: Zero-position calibration of torque sensor / steering wheel angle sensor.

[0263] Model loading: Load the pre-trained DDPG model (Actor / Critic network weights) and expert rule base.

[0264] Style recognition: Read nearly 50km of historical driving data and generate λ_style using the K-means++ algorithm.

[0265] 2. Real-time perception and feature extraction (cycle time 5ms, data synchronization error ≤1ms):

[0266] Multi-source data acquisition: Synchronous acquisition of DMS data ( Torque / steering angle data Radar / camera data Meteorological data.

[0267] Feature calculation: Output It is stored in shared memory for use by the decision-making level.

[0268] 3. Hybrid decision-making and weight allocation (20ms cycle, matching the reinforcement learning iteration cycle):

[0269] Expert rule base priority matching: if extreme scenario conditions are met... "or" , and directly output ω_driver.

[0270] Decision-making in typical scenarios: Using the DDPG network, input S=[P_driver,Q_risk,λ_style], iteratively calculate ω_driver, and add offsets based on λ_style (e.g., S1 plus 0.1) to output the final ω_driver / ω_system.

[0271] 4. Multi-objective optimization and instruction generation (cycle 10ms, matching the optimization solver cycle):

[0272] Command acquisition: Read δ_sys output from Model Predictive Control (MPC) and δ_driver acquired and corrected by sensors.

[0273] Weight calculation: Based on Q_risk and ω_driver, calculate γ1, γ2, and γ3.

[0274] Optimization solution: Call the interior point method solver, input the optimization function and constraints, output δ_cmd and T_ref, and send them to the execution layer.

[0275] 5. Scene-adaptive execution control (1ms cycle, matched with motor control cycle):

[0276] Three-loop control: Position loop outputs T_pos → Torque loop adjusts deviation → Current loop drives the motor, tracking δ_cmd.

[0277] Transition strategy selection: Calculate dω / dt, select an S-curve or a linear transition with feedforward, and suppress sudden torque changes.

[0278] Road feel adjustment: Adjust the damping / stiffness coefficient of the road feel feedback motor according to λ_style to output road feel torque.

[0279] (III) Verification in Typical Scenarios

[0280] The effectiveness of this invention was verified through real-vehicle testing. Two typical scenarios were selected, and the test conditions and results are as follows:

[0281] Scenario 1: Aggressive driver + highway cruising in light rain

[0282] Test conditions: Straight section of highway, vehicle speed 100km / h, light rain (μ=0.6), no sudden obstacles; .

[0283] Decision-making process: Expert rule matching for "aggressive style + low to medium risk", DDPG output. .

[0284] Optimized integration: ; (Driver's intention to change lanes) (System recommended path), after optimization (Preferring the driver's intention).

[0285] Execution result: Drivers reported that they had "clear road feel, no delay in operation, and no jerking."

[0286] Scenario 2: Conservative driver + pedestrians crossing the street in a congested city.

[0287] Test conditions: Urban intersection, vehicle speed 30km / h, pedestrians crossing ( (TTC decreased from 4s to 1.0s). .

[0288] Decision-making process: Expert rules trigger "extremely high risk" "forced" .

[0289] Optimized integration: ; (System risk avoidance and steering command) (Driver reaction delay), after optimization (Completely trace system commands).

[0290] Execution result: Employing a linear transition with feedforward, the torque increased from 3N within 0.2 seconds. m rises to 12N m, mutation amplitude = 1.1N m; Steering response delay = 12ms, successfully avoided pedestrian (at TTC = 1.0s, the vehicle deviated 0.8m from its original path, avoiding the pedestrian); No steering wheel kickback, road feel stiffness = 1.2N. m / °, conservative drivers reported that "the operation is smooth and gives a strong sense of security".

[0291] Figure 2 This is a structural block diagram of a vehicle steering control system according to one embodiment of this application, such as... Figure 2 As shown, the system includes a multimodal perception and feature extraction module 201, used to collect driver status data, environmental data, and vehicle operation data, and quantitatively output a comprehensive driver status score. Comprehensive environmental risk value and driving style weighting coefficient ,in:

[0292] , This is a quantification of the driving intention based on steering torque and steering angular velocity. The attention concentration coefficient is based on the proportion of eye deviation, blink frequency, and head tilt angle.

[0293] This is a static risk index based on road curvature, slope, and adhesion coefficient. This is a dynamic risk index based on collision time. This is a meteorological risk index based on weather type.

[0294] The driving style weight coefficients (aggressive type = 1.0, standard type = 0.7, conservative type = 0.4) are obtained by clustering historical driving data using the K-means++ algorithm.

[0295] The hybrid reinforcement learning dynamic weight allocation module 202 is used to output driver control weights. With system control weights ( The decision-making architecture employs a combination of a Deep Deterministic Policy Gradient (DDPG) network and an expert rule base, with the input state... .

[0296] The multi-objective collaborative optimization control module 203 is used to acquire the desired steering angle of the system. vs. driver's desired steering angle after style modification Through dynamic weighted multi-objective optimization function Solving for the optimal steering command ,in , .

[0297] The scene adaptive execution control module 204 adopts a three-closed-loop control architecture of "position loop + torque loop + current loop". It selects an S-curve or a linear transition function with feedforward based on the weight change rate dω / dt, and adjusts the damping and stiffness coefficients of the road feel feedback motor based on λ_style.

[0298] Furthermore, the driver state monitoring module in the multimodal perception and feature extraction module includes a torque sensor (accuracy 0.01N). The system includes a steering wheel angle sensor (0.01° resolution), a driver monitoring system camera (30fps frame rate), a 16-line lidar (150m detection range), an onboard weather sensor (humidity accuracy ±2%, rainfall resolution 0.1mm / h), and a high-precision map module (positioning accuracy ±0.1m).

[0299] Furthermore, the expert rule base of the hybrid reinforcement learning dynamic weight assignment module includes at least: ; ; For the output of the DDPG network ; For the output of the DDPG network .

[0300] Furthermore, the optimization constraints of the multi-objective collaborative optimization control module are as follows: The optimal steering command is determined every 10ms using the interior point method. .

[0301] Furthermore, the segmented transition function of the scene adaptive execution control module includes: At that time, an S-shaped curve transition function is used. The transition time is 0.4s, and the torque surge is ≤0.8N. m; At that time, a linear function with feedforward compensation is used. The transition time is 0.2s. Mutation ≤ 1.2N m.

[0302] This application also provides a vehicle steering control method, including the following steps:

[0303] Step 1: System initialization: Calibrate the sensors, load the pre-trained DDPG model and expert rule base, and generate the driving style weight coefficient λ_style using nearly 50km of historical driving data.

[0304] Step 2: Real-time perception and feature extraction: Multi-source data are collected synchronously every 5ms to calculate the comprehensive driver status score P_driver and the comprehensive environmental risk value Q_risk.

[0305] Step 3: Hybrid decision-making and weight allocation: Expert rules are matched first every 20ms. In extreme scenarios, ω_driver is directly output. In normal scenarios, ω_driver is calculated iteratively through the DDPG network, and the offset is added according to λ_style.

[0306] Step 4: Multi-objective optimization and instruction generation: Collect δ_sys and δ_driver every 10ms, calculate dynamic weights γ1~γ3, solve the multi-objective optimization function using the interior point method, and output the optimal steering instruction δ_cmd.

[0307] Step 5: Scene Adaptive Execution Control: Drive the steering motor through three closed-loop control every 1ms, select the transition function according to dω / dt, and adjust the road feel feedback parameters based on λ_style.

[0308] Furthermore, the DDPG network in step 3 includes an Actor network and a Critic network. The Actor network is a 3-layer fully connected structure (with the ReLU activation function), and the Critic network is based on the reward function. Update strategy, where ΔY is the lane departure amount (target ≤ ±0.1m) and ΔT is the torque change amplitude (target ≤ 1N). m), .

[0309] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements all the functions of the system in the various embodiments of this application.

[0310] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements all the steps of the methods in the various embodiments of this application.

[0311] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0312] According to an embodiment of this application, a vehicle steering control device is provided. It should be noted that the device can be used to execute the above-described vehicle steering control method.

[0313] Figure 3 This is a structural block diagram of a vehicle steering control device according to one embodiment of this application, such as... Figure 3 As shown, taking a vehicle steering control device 300 as an example, the device includes:

[0314] The acquisition module 301 is used to acquire vehicle state data and driving state data. The vehicle state data characterizes the vehicle's driving state and the environmental state during driving, while the driving state data characterizes the driver's operational intention and attention level. The state parameter determination module 302 is used to determine multi-dimensional state parameters based on the vehicle state data and the driver's driving state data. The vehicle state data characterizes the vehicle's driving state and the driving environment state, while the driving state data characterizes the driver's operational intention and attention level. The multi-dimensional state parameters include: driver state score, driving style coefficient, and environmental risk value. The weight parameter determination module 303 is used to determine weight parameters based on the multi-dimensional state parameters, preset rules, and a reinforcement learning model. The preset rules are used to allocate weight parameters under preset scenarios. The weight parameters include driver control weight and system control weight. The instruction determination module 304 is used to determine the target steering instruction based on the weight parameters, driving style coefficient, environmental risk value, system-planned steering angle, and steering wheel angle. The control module 305 is used to perform steering control on the vehicle based on the target steering instruction, driving style coefficient, and the rate of change of the weight parameters.

[0315] Embodiments of this application also provide a vehicle, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods described in various embodiments of this application when it runs.

[0316] Embodiments of this application also provide a computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of this application.

[0317] Embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the methods of various embodiments of this application.

[0318] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0319] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0320] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0321] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0322] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0323] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A vehicle steering control method, characterized in that, The method includes: Acquire vehicle status data and driving status data, wherein the vehicle status data is used to characterize the vehicle's driving status and the environmental status of the vehicle during driving, and the driving status data is used to characterize the driver's operating intention and attention level. Based on the vehicle status data and the driving status data, multi-dimensional status parameters are determined, wherein the multi-dimensional status parameters include: driver status score, driving style coefficient and environmental risk value; Based on the multidimensional state parameters, preset rules, and reinforcement learning model, weight parameters are determined, wherein the preset rules are used to allocate the weight parameters under preset scenarios, and the weight parameters include driver control weights and system control weights; The target steering command is determined based on the weighting parameters, the driving style coefficient, the environmental risk value, the system-planned steering angle, and the steering wheel angle. The vehicle is steering controlled based on the target steering command, the driving style coefficient, and the rate of change of the weighting parameters.

2. The method according to claim 1, characterized in that, The step of determining the multidimensional state parameters based on the vehicle state data and the driving state data includes: The driver state score is determined based on steering torque, steering angular velocity, the percentage of time the driver's gaze is off-center from the lane, the driver's blinking frequency, and the driver's head pitch angle. The driving style coefficient is determined based on the standard deviation of steering radius, following distance, and rate of change of accelerator pedal opening. The environmental risk value is determined based on road curvature, road slope, road adhesion coefficient, relative distance to obstacles, relative speed of the obstacles, and weather conditions.

3. The method according to claim 2, characterized in that, The driver state score, determined based on steering torque, steering angular velocity, the percentage of time the driver's gaze deviates from the lane center, the driver's blink frequency, and the driver's head pitch angle, includes: The driver's operational intent score is determined based on the steering torque and the steering angular velocity. The driver's attention score is determined based on the percentage of time the driver's gaze is off-center from the lane, the driver's blinking frequency, and the driver's head tilt angle. The driver's state score is determined based on the operational intent score and the attention score.

4. The method according to claim 2, characterized in that, The determination of the driving style coefficient based on the standard deviation of steering radius, following distance, and accelerator pedal opening rate includes: Based on the vehicle's historical status data, determine the historical steering amplitude standard deviation, historical following distance, and historical accelerator pedal opening change rate; Clustering algorithms were used to cluster the historical steering amplitude standard deviation, the historical following distance, and the historical accelerator pedal opening change rate to obtain multiple clustering results; A preset driving style type is determined based on the multiple clustering results, and a style coefficient mapping table is established based on the preset driving style type, wherein the style coefficient mapping table is used to record the mapping relationship between the preset driving style type and the driving style coefficient; The clustering algorithm is used to cluster the standard deviation of the steering amplitude, the following distance, and the rate of change of the accelerator pedal opening to obtain at least one target clustering result. The driving style coefficient corresponding to the at least one target clustering result is determined according to the style coefficient mapping table.

5. The method according to claim 2, characterized in that, The determination of the environmental risk value based on road curvature, road slope, road adhesion coefficient, relative distance to obstacles, relative speed of obstacles, and weather conditions includes: The static risk coefficient is determined based on the road curvature, the road slope, and the road adhesion coefficient. The collision time is determined based on the relative distance and relative speed of the obstacles, and the dynamic risk coefficient is determined based on the collision time. The weather conditions are determined based on the vehicle's weather sensors and camera, and the weather risk coefficient is determined based on the weather conditions. The environmental risk value is determined based on the static risk coefficient, the dynamic risk coefficient, and the meteorological risk coefficient.

6. The method according to claim 1, characterized in that, The determination of weight parameters based on the multidimensional state parameters, preset rules, and reinforcement learning model includes: Determine whether the vehicle is in the preset scenario based on the multidimensional state parameters; In response to the vehicle being in the preset scenario, the weight parameters are determined according to the multi-dimensional state parameters and the preset rules; In response to the vehicle not being in the preset scenario, the reinforcement learning model is used to process the multidimensional state parameters to obtain initial weight parameters; The initial weight parameters are corrected based on the driving style coefficient to obtain the weight parameters.

7. The method according to claim 1, characterized in that, The step of determining the target steering command based on the weighting parameter, the driving style coefficient, the environmental risk value, the system-planned steering angle, and the steering wheel angle includes: Based on the driver control weight, the environmental risk value, and the driving style coefficient, multiple target weight coefficients are determined; The target steering command is determined based on the multiple target weight coefficients, the system planning steering angle, the steering wheel steering angle, and the constraints, wherein the constraints include steering angle constraints and steering rate constraints.

8. The method according to any one of claims 1-7, characterized in that, The step of controlling the vehicle's steering based on the target steering command, the driving style coefficient, and the rate of change of the weighting parameters includes: The target steering command is controlled in a closed loop using a preset closed-loop control strategy to obtain the initial execution torque. The preset closed-loop control strategy includes a position closed-loop control strategy, a torque closed-loop control strategy, and a current closed-loop control strategy. The transition function is determined based on the rate of change of the weight parameters and the initial execution torque, and the transition torque is determined based on the transition function and the initial execution torque. The road feel compensation torque of the road feel feedback motor is determined based on the transition torque, the driving style coefficient, and the road feel simulation rule, wherein the road feel simulation rule is used to determine the corresponding road feel compensation torque according to different driving style coefficients; The target execution torque is determined based on the road feel compensation torque and the transition torque, and the vehicle is steering controlled based on the target execution torque.

9. A vehicle, characterized in that, include: Memory, which stores executable programs; A processor for running the program, wherein the program, when running, performs the vehicle steering control as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program is configured to perform the vehicle steering control as described in any one of claims 1 to 8 when executed on a computer or processor.

Images