A self-turning brake safety control method for a four-wheel independent steering-by-wire chassis after a collision

Abstract

The application discloses a kind of four-wheel independent steering by-wire chassis collision after self-rotation brake safety control method, belong to intelligent vehicle active safety and by-wire chassis control technical field.For the problem that direct braking under the condition of large side slip, high yaw and limited tire adhesion resource after collision is easy to exacerbate instability, three controllable margins are constructed by dynamic two-dimensional side slip angle phase plane state, tire adhesion margin and vehicle yaw recovery ability, and the control is divided into three stages of spin suppression, self-rotation braking and safe parking by using phase supervisor with hysteresis and dwell time constraint;Each stage generates low-dimensional candidate generalized control target, and four-wheel independent rotation angle and brake force command are obtained by four-wheel constraint optimization distribution after safety action projection.The application can determine the braking intervention time according to the tire lateral recovery ability, and reconstruct the brake force direction by four-wheel large-angle steering, considering the post-collision attitude recovery, kinetic energy dissipation, hard constraint safety and real-time deployment of vehicle end.

Description

Technical Field

[0001] This invention belongs to the field of intelligent vehicle active safety and drive-by-wire chassis control technology, and specifically relates to a phased safety reinforcement learning control method for four-wheel independent steering drive-by-wire chassis with post-collision self-rotation braking based on three layers of controllable margin constraints: state layer, tire layer, and vehicle layer. It is applicable to the attitude recovery, energy dissipation, and safe stopping control of passenger cars, autonomous vehicles, and special vehicles with four-wheel independent steering and four-wheel independent braking capabilities after side or rear-side collisions. Background Technology

[0002] When a vehicle is subjected to a lateral or rear-side collision at high speed, the impact impulse causes a sudden change in longitudinal velocity, lateral velocity, yaw rate, and vertical load within a very short time, causing the vehicle to rapidly enter a state of high center of gravity sideslip angle, high yaw rate, and tire force saturation. At this point, the vehicle's original stability constraints are weakened, and the rear axle tires usually enter the lateral saturation zone first, making the vehicle prone to fishtailing, spinning, lateral drift, or even secondary collisions. Traditional electronic stability control systems are mainly designed for normal extreme handling conditions, and their control logic is mostly based on yaw rate deviation or center of gravity sideslip angle thresholds. Under conditions of strong nonlinearity, strong coupling, and severely limited tire adhesion resources after a collision, the control effect is significantly limited.

[0003] Existing post-collision stability control methods mainly include differential braking, active front wheel steering, four-wheel steering, and coordinated steering and braking control. These methods can reduce yaw response in minor collisions or when tires are not yet severely saturated, but they still have the following problems: (1) Existing methods usually start braking control directly based on the vehicle's yaw rate or center of gravity sideslip angle, without fully considering the occupation of tire lateral adhesion resources by longitudinal braking force. When the vehicle is in a large sideslip state, directly applying forced braking will cause the longitudinal force of the tire to be built up quickly, further compressing the available capacity of lateral force, resulting in a decrease in the lateral recovery ability of the rear axle, and even exacerbating the vehicle's tail-swing.

[0004] (2) Existing stability domain criteria mostly use fixed boundaries or single state thresholds, which are difficult to adapt to changes in vehicle speed, road surface adhesion coefficient, four-wheel steering angle, collision intensity, and tire longitudinal slip state. Whether the vehicle state has exceeded the boundary, whether the tire still has lateral recovery space, and whether the current tire resultant torque can suppress yaw are controllability issues at different levels. It is difficult to accurately determine the timing of controller intervention and control stage by relying on a single indicator.

[0005] (3) Traditional model control methods require the establishment of accurate vehicle and tire models and rely on manual setting of multi-objective weights. When the nonlinearity of the vehicle after the collision, the initial state of the collision, and the tire saturation change significantly, fixed parameter controllers have difficulty coordinating conflicting objectives such as yaw stability, lateral displacement, tire margin, deceleration efficiency, and actuator ride comfort.

[0006] (4) Reinforcement learning can learn complex nonlinear control laws through interaction with the environment, but existing reinforcement learning post-collision control usually adopts a single-stage end-to-end strategy, directly outputting the front and rear wheel steering angles or wheel-end control quantities, lacking an explicit expression of the physical process of "first restoring lateral capability, then implementing constrained braking, and finally stopping safely" after a collision. At the same time, the black-box strategy may output control commands that violate the constraints of friction ellipse, steering angle range, or rate of change of braking force, making it difficult to meet the interpretability and verifiability requirements of vehicle safety control.

[0007] (5) The four-wheel independent steering steer-by-wire chassis can control the turning angle of each wheel separately. By turning at a large angle, it changes the direction of the longitudinal braking force of the tire in the vehicle coordinate system, thereby forming an additional lateral component and restoring yaw moment. Existing research has not fully utilized this capability to establish a phased self-rotation braking control mechanism that matches the tire controllability margin, nor has it formed a complete technical solution that combines physical criteria, reinforcement learning coordination and safety constraint allocation.

[0008] In summary, accurately determining when a vehicle should prioritize tail-wagging and when braking force should intervene, while preserving the interpretability of collision dynamics and tire mechanics, and how to utilize large-angle independent steering of the four wheels to reconstruct the direction of braking force and safely coordinate the braking force of the four wheels, are technical problems that urgently need to be solved in the field of post-collision active safety control. Summary of the Invention

[0009] The purpose of this invention is to address the problems of large center of gravity deviation, high yaw rate, tire compound slip, and rapid saturation of adhesion resources that occur after a lateral or rear-side collision. It provides a self-rotation braking safety control method for a four-wheel independent steering steer-by-wire chassis after a collision. While maintaining the interpretability of collision dynamics and tire mechanics, the method dynamically determines the control phase and braking intervention timing based on the vehicle state, remaining tire adhesion capacity, and overall vehicle yaw recovery capacity. Furthermore, it utilizes safety reinforcement learning and four-wheel constraint optimization to achieve attitude recovery, kinetic energy dissipation, and safe stopping.

[0010] To achieve the above objectives, the present invention adopts the following technical solution. The method includes the following steps: S1 acquires the vehicle motion state, four-wheel tire state, collision triggering and working condition information, road adhesion information and actuator state after the collision, and constructs a two-dimensional side slip angle phase plane using the front axle equivalent side slip angle and the rear axle equivalent side slip angle as two coordinate variables.

[0011] S2, based on one or more operating parameters among longitudinal speed, road surface adhesion coefficient, equivalent front and rear axle rotation angles, longitudinal braking force occupancy, and vehicle structural parameters, the dynamic boundary of the two-dimensional sideslip angle phase plane is generated online, and the controllability margin of the state layer, the controllability margin of the tire layer, and the controllability margin of the whole vehicle layer are calculated respectively.

[0012] S3, based on the three-layer controllable margin, through a stage supervisor with hysteresis threshold, continuous satisfaction time and minimum dwell time constraints, the post-collision control process is divided into a tail-slip suppression stage, a self-rotation braking stage and a safe stopping stage, and allows reversal according to preset conditions when the vehicle becomes unstable again or the stability margin decreases.

[0013] S4, during the drift suppression phase, uses a basic stability controller and residual reinforcement learning strategy to generate candidate equivalent steering angles for the front and rear axles, and restricts or prohibits forced movement, prioritizing the reduction of yaw rate and center of gravity sideslip angle to release the tire's lateral available capacity.

[0014] S5, during the self-rotation braking phase, adopts a continuous action reinforcement learning strategy to generate or incrementally update the candidate equivalent rotation angles of the front and rear axles, the candidate longitudinal resultant forces of the whole vehicle, and the candidate additional yaw moments. It uses the independent large-angle steering of the four wheels to change the direction of the wheel-end braking force in the vehicle coordinate system, so that the braking force generates a favorable lateral component and recovers the yaw moment while dissipating the vehicle's kinetic energy.

[0015] S6 generates smooth deceleration and attitude maintenance targets based on the remaining vehicle speed, road boundaries, and safe parking area during the safe parking phase, enabling the vehicle to gradually stop under yaw stability and road boundary constraints.

[0016] S7 projects the candidate generalized control objective generated in the current stage to the safe and feasible domain that satisfies the vehicle state boundary, tire adhesion constraint, road boundary and actuator constraint to obtain the safe generalized control objective; then solves the four-wheel independent steering angle and four-wheel braking force commands respectively through four-wheel constraint optimization allocation, and applies them to the vehicle through steering and braking actuators to form closed-loop control.

[0017] The state layer controllable margin is used to characterize the degree of deviation of the current vehicle state from the dynamic stability boundary and its divergence trend; the tire layer controllable margin is used to characterize the four-wheel adhesion utilization rate, the rear axle lateral residual margin, and the rear axle lateral force saturation; the vehicle layer controllable margin is used to characterize the ability of the maximum recoverable yaw moment achievable under the current tire and actuator constraints to recover the required yaw moment.

[0018] The dynamic boundary can be established by using a radial basis function neural network optimized by a genetic algorithm to establish the mapping between operating parameters and stable equilibrium points and boundary parameters at all levels, or it can be generated by Gaussian process regression, support vector regression, multilayer perceptron or lookup table interpolation.

[0019] The stage supervisor uses different entry and exit thresholds to form a hysteresis loop, and frequently switches between stages by combining continuous satisfaction time and minimum stage dwell time limits. When the severe instability entry condition is met again in the spin braking stage or safe stopping stage, it reverts to the fishtail suppression stage. When the stability margin decreases in the safe stopping stage but has not yet reached the severe instability condition, it reverts to the spin braking stage. The stage switching uses the current vehicle state of the subsequent stage and the control quantity of the previous control cycle as the initial conditions to maintain the continuity of the control state.

[0020] The equivalent steering angles of the front and rear axles are generalized control objectives at the upper axle level, and do not limit the left and right wheels of the same axle to using the same steering angle. The lower-level four-wheel constraint optimization allocation uses the independent steering angles of the four wheels and the longitudinal tire forces of the four wheels as decision variables. While tracking the equivalent steering angles of the front and rear axles, the longitudinal resultant force of the whole vehicle, and the additional yaw moment, it satisfies the constraints of steering angle, steering rate, braking force, braking force change rate, and tire friction, thereby obtaining different steering angles and braking forces of the four wheels.

[0021] The present invention also provides a four-wheel independent steering steer-by-wire chassis post-collision stability control system, including a state acquisition unit, a dynamic phase plane boundary generation unit, a three-layer controllability margin evaluation unit, a stage supervision unit, a reinforcement learning strategy inference unit, a safety action projection unit, a four-wheel constraint optimization allocation unit, and a four-wheel steering and braking execution unit. The system is used to execute the above-mentioned methods.

[0022] The present invention also provides a vehicle having a four-wheel independent steering actuator and a four-wheel independent braking actuator, and equipped with the post-collision stability control system.

[0023] Compared with the prior art, the present invention has at least the following beneficial effects: 1. The timing of braking intervention is determined based on the lateral margin of the rear axle and the adhesion utilization rate. During severe tail-swing, longitudinal force is unloaded first and four-wheel steering is used to restore lateral capability, which can avoid premature forced braking and further compression of tire lateral adhesion resources.

[0024] 2. By utilizing the dynamic two-dimensional sideslip angle phase plane boundary that varies with vehicle speed, adhesion, turning angle and longitudinal slip state, and integrating the three controllable margins of the state layer, tire layer and whole vehicle layer, it is possible to simultaneously determine whether the vehicle has crossed the boundary, whether the tire still has room for recovery, and whether the current yaw recovery capability is sufficient.

[0025] 3. The post-collision control process is divided into three physically distinct stages: tail-flick suppression, self-rotation braking, and safe stopping. Smooth switching between stages is achieved through hysteresis, dwell time, and backoff mechanisms, thereby improving the control adaptability and interpretability under strongly nonlinear post-collision conditions.

[0026] 4. By utilizing four-wheel independent steering to change the direction of braking force in the vehicle coordinate system, the braking force is not only used for deceleration, but also to generate a favorable lateral component and restore yaw moment, giving full play to the multi-degree-of-freedom control capabilities of the large-angle four-wheel independent steering and independent braking steer-by-wire chassis.

[0027] 5. By using low-dimensional candidate generalized control targets, safe action projection, and real-time optimization allocation of four wheels, the training difficulties and motion incoordination caused by the reinforcement learning network directly outputting a large number of wheel-end commands are avoided, and the tire friction and actuator hard constraints are explicitly satisfied.

[0028] 6. By randomizing collision conditions, implementing phased learning courses, verifying high-fidelity joint simulations, and freezing deployment on the vehicle side, the strategy can adapt to different collision intensities, adhesion conditions, vehicle parameters, and actuator states, demonstrating good potential for engineering deployment. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall framework of the phased security reinforcement learning control method of the present invention; Figure 2 This is a schematic diagram of the three-layer controllability margin assessment and stage state machine of the present invention; Figure 3 This is a schematic diagram of the constrained Markov decision process and reinforcement learning interaction flow of the present invention. Figure 4 This is a schematic diagram of the tail-flick suppression residual reinforcement learning control process in the first stage of this invention; Figure 5 This is a schematic diagram of the reinforcement learning coordination process for the four-wheel independent steering, self-rotation, and braking in stage two of this invention. Figure 6 This is a schematic diagram of the safety action projection and four-wheel constraint optimization allocation process of the present invention; Figure 7 This is a schematic diagram of the offline training and vehicle-side deployment process of the present invention. Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are used to illustrate the technical solution of the present invention, and are not intended to limit the scope of protection of the present invention.

[0031] To address the aforementioned technical problems, this invention provides a phased safety reinforcement learning control method for a four-wheel independent steering steer-by-wire chassis based on three layers of controllable margin constraints, involving post-collision self-rotation braking. For example... Figure 1 As shown, the present invention includes a post-collision vehicle-tire state acquisition module, a dynamic phase plane boundary generation module, a three-layer controllable margin assessment module, a stage supervision module, a stage one tail-slip suppression control module, a stage two self-rotation braking coordination module, a stage three safe stopping module, a safe action projection module, and a four-wheel constraint optimization allocation module.

[0032] This invention first constructs three layers of controllable margin indices—state layer, tire layer, and vehicle layer—based on the post-collision vehicle state, tire compound slip state, and overall vehicle yaw moment recovery capability. Then, a stage supervisor with hysteresis and dwell time constraints determines whether the vehicle is in a drift suppression, spin braking, or safe stopping phase. "Spinning" in this paper refers to the undesired yaw rotation of the vehicle around its center of mass vertical axis under collision disturbance, i.e., the overall yaw rotation of the vehicle, not the rotation of the wheels around their axles. In the drift suppression phase, a combination of model control and residual reinforcement learning is used to prioritize reducing yaw and sideslip, releasing the available lateral capacity of the tires. In the spin braking phase, a continuous motion reinforcement learning strategy is used to generate generalized control targets such as the equivalent front and rear axle steering angles, total braking force, and additional yaw moment, reconstructing the braking force direction using large-angle steering of the four wheels. In the safe stopping phase, constrained deceleration control is used to keep the vehicle stable and gradually bring it to a stop. The outputs of each stage are converted into steering angle and braking force commands for the four wheels after being projected into safe motion and optimized by four-wheel constraints.

[0033] The specific process of this invention will be described below.

[0034] 1. Problem Modeling This invention models the post-collision phased control process as a switching-constrained Markov decision process with phase variables and safety cost constraints, which can be expressed as:

[0035] in, For state space, For continuous action space, Let be the vehicle state transition probability. For phased reward functions, For the safety cost function, As a discount factor, For stage set, , and These correspond to the drift suppression stage, the spin braking stage, and the safe stopping stage, respectively.

[0036] time The vehicle's basic state vector includes longitudinal velocity. Lateral velocity yaw rate yaw acceleration , centroid side slip angle lateral angle Lateral deviation relative to the safety reference line Four-wheel vertical load Four-wheel side slip angle Four-wheel slip ratio and the executor instructions from the previous moment .

[0037] The equivalent sideslip angles of the front and rear axles are obtained by averaging the sideslip angles of the left and right wheels, respectively.

[0038] In one implementation, the vehicle's basic state is obtained by an onboard inertial measurement unit, wheel speed sensors, steering angle sensors, brake pressure sensors, and a combined navigation system; lateral velocity, center of gravity sideslip angle, and tire force, which are difficult to measure directly, can be estimated by an unscented Kalman filter, an extended Kalman filter, or other state observers. Road surface adhesion coefficient. The adhesion coefficient can be obtained from the online adhesion estimation module. When the accurate adhesion coefficient cannot be obtained, a conservative lower bound or adhesion coefficient range can be set according to the road type. The lower bound of the adhesion coefficient range is used to calculate the tire adhesion margin and safety constraints to avoid the tire force demand exceeding the actual available adhesion range due to overestimation of the road surface adhesion capability.

[0039] The reinforcement learning state of this invention includes not only the vehicle's motion state, but also three layers of controllable margin and stage variables, constituting:

[0040] in, For phase plane instability, The rate of change of instability To achieve the maximum adhesion utilization rate for all four wheels, This represents the minimum lateral margin for the rear axle. The degree of saturation of lateral forces on the rear axle. This represents the vehicle's yaw recovery coefficient.

[0041] Different action definitions are used for each stage. Candidate actions for the drift suppression stage are the equivalent steering angle residuals of the front and rear axles and their rates of change; candidate actions for the spin braking stage are the equivalent steering angle increments of the front and rear axles, the total longitudinal braking force increment, and the additional yaw moment increment; actions for the safe stopping stage are the target deceleration, the target yaw moment, and the path-keeping correction. None of the candidate actions are directly sent to the actuators; instead, executable control commands are first obtained through safe action projection and four-wheel constraint optimization allocation.

[0042] 2. Module Function Introduction (1) Post-collision vehicle-tire status acquisition module The post-collision vehicle-tire status acquisition module is used to collect or estimate vehicle motion status, tire status, collision trigger information, and road boundary information. Collision trigger information includes at least one of the following: collision time, collision location, collision direction, collision duration, and collision intensity. After a collision, the system freezes any drastic erroneous inputs from the driver or limits them according to access control logic, and switches control authority to the post-collision stability controller.

[0043] (2) Dynamic phase plane boundary generation module The dynamic phase plane boundary generation module, for a two-dimensional sideslip angle phase plane constructed using the equivalent sideslip angles of the front and rear axles as two coordinate axes, dynamically determines the stable equilibrium point, inner stable boundary, transition boundary, and severe instability boundary of this two-dimensional sideslip angle phase plane by using one or more operating parameters selected from longitudinal velocity, road adhesion coefficient, equivalent front and rear axle rotation angles, longitudinal braking force occupancy, and vehicle structural parameters as boundary adjustment inputs. In one implementation, a radial basis function neural network optimized by a genetic algorithm is used to establish the mapping between operating parameters and phase plane boundary parameters.

[0044] in, Including the coordinates of the stable equilibrium point and boundary radii or boundary function parameters at each level, This is a statistical measure of the four-wheel slip ratio. Besides GA-RBF, dynamic boundaries can also be generated using Gaussian process regression, support vector regression, multilayer perceptron, or lookup table interpolation.

[0045] (3) Three-layer controllability margin assessment module The three-layer controllability margin assessment module is used to determine whether the vehicle is currently controllable from three levels: vehicle state deviation, tire adhesion resources, and vehicle yaw recovery torque, and outputs the basis for stage switching to the stage supervision module.

[0046] (4) Stage Supervision Module The phase monitoring module determines phase variables based on the three-layer controllability margin index. Furthermore, frequent switching between control phases is avoided by setting hysteresis thresholds, continuous satisfaction time, and minimum dwell time limits. When the vehicle re-enters a severely unstable state, the phase supervision module has the authority to revert from the self-rotation braking phase or the safe stopping phase to the fishtail suppression phase.

[0047] (5) Stage 1 tail-flick suppression control module The Stage 1 drift suppression control module employs a parallel structure of a base stability controller and a residual reinforcement learning strategy. The base stability controller provides interpretable baseline front and rear axle steering angles, while the residual reinforcement learning strategy superimposes a finite-amplitude correction on the baseline control values ​​to adapt to different collision intensities, adhesion conditions, and tire nonlinear states. This stage restricts or prohibits forced movement, prioritizing the convergence of the drive longitudinal force towards zero, reducing the rear axle sideslip angle, and restoring lateral capacity.

[0048] (6) Stage 2 Self-rotation Braking Coordination Module Phase 2 self-rotation braking coordination module uses reinforcement learning strategy to output low-dimensional generalized control target, and changes the direction of braking force in the vehicle coordinate system by large-angle steering of four wheels, so that the braking force generates a beneficial component to suppress yaw and lateral movement while completing the dissipation of vehicle kinetic energy.

[0049] (7) Phase 3 Safe Parking Module The Phase 3 safe parking module, after controlling vehicle yaw and side slip, generates target deceleration and path-keeping targets to gradually bring the vehicle to a stop within the road boundary or safe area. This module can be implemented using model predictive control, sliding mode control, linear quadratic control, or rule-based control.

[0050] (8) Safety Action Projection and Four-Wheel Constraint Optimization Allocation Module The safety action projection module projects reinforcement learning candidate actions to the feasible domain that satisfies the state safety boundary, tire friction constraints, and actuator constraints; the four-wheel constraint optimization allocation module converts generalized objectives such as the front and rear axle equivalent rotation angles, total longitudinal force, and additional yaw moment into four-wheel rotation angle and four-wheel braking force commands.

[0051] 3. Three-tiered controllable margin assessment and phase switching 3.1 State-level controllability margin Let the equivalent sideslip angles of the front and rear axles of the current vehicle be respectively... and The coordinates of the stable equilibrium point under the current operating condition, determined by the dynamic phase plane boundary generation module, are: , and These represent the stable equilibrium values ​​of the equivalent sideslip angles of the front and rear axles under the current operating conditions. The radial distance between the current state point and the stable equilibrium point is defined as:

[0052] Let the equivalent radius of the dynamic severe instability boundary be... Define the normalized phase plane instability:

[0053] in, To prevent positive numbers with a denominator of zero. When At that time, the vehicle entered a state of severe instability; when At that time, the vehicle is located within the severe instability boundary. To reflect the divergence trend of the vehicle's state, the following calculations are made:

[0054] when Not yet crossed the boundary but When the value remains positive and exceeds the preset threshold, the tail-flick suppression phase can be initiated earlier.

[0055] In implementations of non-circular stable boundaries, implicit boundary functions can be used. The principle for calculating the signed instability is to take a negative value or less than a preset threshold when the vehicle is on the safe side, and a positive value or greater than a preset threshold when it is on the dangerous side.

[0056] 3.2 Tire ply controllability margin For the For each wheel, define the adhesion utilization rate:

[0057] The maximum adhesion utilization rate of the four wheels is:

[0058] Under the frictional elliptic approximation, the first... The available lateral force capacity of each wheel under the current longitudinal force occupancy condition is:

[0059] The corresponding signed lateral residual margin is:

[0060] The minimum lateral margin of the rear axle is defined as:

[0061] The degree of saturation of the rear axle lateral force is defined as follows:

[0062] when Approaching 1 Close to 0 or negative, and When the value is high, it indicates that the rear axle tires lack lateral recovery space. At this time, the monitoring module limits the intervention of braking force and prioritizes reducing the slip angle and longitudinal force occupation through four-wheel steering.

[0063] 3.3 Controllable Margin of the Entire Vehicle Layer No. The tire forces of each wheel in the vehicle body coordinate system are obtained by transforming the longitudinal and lateral tire forces in the wheel coordinate system according to the wheel rotation angle:

[0064] in, and These are the first in the wheel coordinate system. The longitudinal and lateral forces of each wheel, For the first The turning angle of each wheel; and These are the longitudinal force and lateral force after transformation to the vehicle body coordinate system.

[0065] No. The yaw moment generated by each wheel relative to the vehicle's center of gravity is:

[0066] in, and The first The longitudinal and lateral positions of each wheel relative to the vehicle's center of gravity.

[0067] Therefore, changing the steering angle of the four wheels Able to change longitudinal braking force The direction of action in the vehicle coordinate system and its contribution of yaw moment enable the braking force to dissipate the vehicle's kinetic energy while generating a force component that helps to suppress yaw and lateral movement. This characteristic constitutes the mechanical basis of the self-rotation braking control of this invention.

[0068] Let the desired yaw recovery direction be determined by the yaw error: , in, This refers to the vehicle's actual yaw rate. This represents the target yaw rate corresponding to the current control phase; during the post-collision attitude recovery process, It can be set to zero or determined based on the safety reference motion state.

[0069] The required yaw moment is calculated based on the vehicle's yaw rate error, center of gravity sideslip angle, and their changing states. In one implementation:

[0070] in, , and These are non-negative control parameters. This represents the rate of change of the yaw rate error.

[0071] Furthermore, under the current constraints of tire adhesion, four-wheel steering angle, and wheel-end braking force, the maximum achievable recoverable yaw moment that can be generated along the direction of the required recoverable yaw moment is defined as:

[0072] This represents the candidate control variables comprised of four-wheel steering angle and tire force. This represents the current feasible control set, which consists of tire friction constraints, wheel angle and angular rate constraints, braking force and braking force change rate constraints.

[0073] Therefore, the yaw recovery capability coefficient of the whole vehicle is defined as follows:

[0074] in, This represents the maximum recovering yaw moment achievable along the demand recovery direction under the current constraints. To restore the yaw moment to demand, To prevent positive numbers with a denominator of zero. When When this occurs, it indicates that the maximum recovering yaw moment achievable under the current tire and actuator constraints is less than the vehicle's current required recovering yaw moment, indicating insufficient yaw recovery capability; when This indicates that the tires and actuators still have the control capability to meet the vehicle's yaw recovery requirements.

[0075] 3.4 Stage Switching Logic The stage variable is set to the tail-suppression stage when any of the following conditions are met. The following conditions are met: Phase plane instability exceeds the entry threshold; Phase plane instability change rate continuously exceeds the divergence threshold; Rear axle lateral margin is lower than the entry threshold; Maximum four-wheel adhesion utilization exceeds the entry threshold; Vehicle yaw recovery capability is lower than the entry threshold.

[0076] When the vehicle is in continuous time When the following conditions are met simultaneously: phase plane instability is below the exit threshold, instability rate of change no longer diverges, rear axle lateral margin is above the recovery threshold, maximum adhesion utilization is below the allowable threshold, and yaw recovery capability is above the allowable threshold, the stage variable is changed from... Switching to the self-rotation braking phase .

[0077] When the vehicle is in continuous time When the yaw rate, center of gravity sideslip angle, phase plane instability, maximum four-wheel adhesion utilization, and lateral deviation are all below the corresponding thresholds, and the vehicle does not show any tendency to re-enter the severe instability zone, the stage variables are changed from... Switch to safe parking phase .

[0078] The entry and exit thresholds are set with different values ​​to form a hysteresis loop, and a minimum stage dwell time is set. This is to avoid frequent switching caused by noise and fluctuations near the boundary. When the vehicle is in the self-rotation braking phase... or safe parking phase If the vehicle regains the drift suppression phase... If any of the following entry conditions occur, including severe phase plane incursion, persistent instability divergence, rear axle lateral margin below the entry threshold, maximum four-wheel adhesion utilization exceeding the entry threshold, or yaw recovery capability below the entry threshold, the stage monitoring module will directly switch to the drift suppression stage. When the vehicle is in the safe parking phase The vehicle stability margin has decreased but has not yet reached the stage of fishtail suppression. If the entry conditions are not met and the conditions for maintaining the safe parking phase are no longer satisfied, the phase monitoring module will revert to the self-rotation braking phase. .

[0079] 4. Stage 1 Tail-Flip Suppression Residual Reinforcement Learning Control The primary objective of Phase One is to rapidly reduce the vehicle's yaw rate and sideslip angle, bringing the rear axle tires back from severe lateral saturation to a controllable range. This phase does not primarily aim for rapid deceleration; instead, it sets a lower upper limit on braking force or temporarily refrains from applying forced braking, prioritizing the removal of driving force to bring the tire's longitudinal force closer to zero, thereby releasing lateral force capacity.

[0080] The basic stability controller generates the front and rear axle reference equivalent rotation angles based on yaw error, centroid sideslip angle, and phase plane state:

[0081] The basic stability controller can employ sliding mode control, model predictive control, linear quadratic control, backstepping control, or other robust control methods. The residual reinforcement learning strategy depends on the state. Output:

[0082] The target control quantity for Phase 1 is:

[0083] in, The residual magnitude scaling matrix, The function defines the steering angle and the rate of change of steering angle. A stable direction is provided by the base controller, and a reinforcement learning strategy is used to adapt to unknown collision intensity, adhesion coefficient, load transfer, and tire saturation nonlinearity, reducing the probability of dangerous maneuvers during severe instability phases in a purely black-box strategy.

[0084] The Stage 1 reward function includes sub-items such as yaw suppression, sideslip suppression, state return to bounds, tire margin recovery, yaw recovery capability improvement, and control ride comfort, which can be expressed as:

[0085] in, , , , , , , , and Non-negative weights Used to limit the scale of rewards. The last item is used to apply the braking force too early or too strongly in the first phase of punishment.

[0086] 5. Phase Two: Four-Wheel Independent Steering, Self-Rotation, Braking, Enhanced Learning and Coordination Once the vehicle's condition returns to the permissible range and tire lateral margin is restored, the vehicle enters the self-rotation braking phase. The goal of this phase is to dissipate vehicle kinetic energy and limit lateral displacement and the risk of secondary collisions by coordinating large-angle steering of all four wheels and wheel-end braking forces while maintaining yaw and side slip reduction.

[0087] Phase 2 reinforcement learning strategy outputs low-dimensional generalized actions:

[0088] in, and This represents the equivalent steering angle increment for the front and rear axles. This represents the increase in total longitudinal braking force. This adds an increment to the yaw moment. Using incremental action instead of absolute action reduces abrupt changes in output and makes it easier to set rate of change constraints.

[0089] Based on the previous moment Control target formation at the current moment Generalized goal:

[0090]

[0091]

[0092]

[0093] This results in a low-dimensional generalized control objective for the output of the reinforcement learning strategy:

[0094] The generalized control target does not directly correspond to a single wheel-end actuator, but rather describes the overall steering requirements of the front and rear axles, the vehicle's deceleration requirements, and the yaw recovery requirements. After being projected onto a safety action, the generalized control target is sent to the four-wheel constraint optimization and allocation module, which further determines the steering angle and braking force of each of the four wheels.

[0095] By employing independent four-wheel steering, the direction of braking force on each wheel is reconstructed based on the vehicle's sideslip direction, the side of impact, and the yaw recovery direction. For example, when the vehicle spins clockwise, the braking force of some wheels can generate a counterclockwise restoring torque in the vehicle's coordinate system; when the lateral displacement approaches the road boundary, the combination of front and rear wheel steering angles can be adjusted so that the longitudinal braking force simultaneously generates a lateral component pointing towards the safe area. The steering angle combination is not limited to traditional in-phase or out-of-phase four-wheel steering; different steering angles for the four wheels can be formed based on reinforcement learning strategies and constraint allocation results.

[0096] The Phase 2 reward function includes sub-items such as yaw and sideslip convergence, lateral displacement limitation, effective deceleration, tire grip margin, yaw recovery moment, and ride comfort.

[0097] in, This refers to the deviation between the vehicle's heading and the safety reference direction. The threshold for safe adhesion utilization is set. This reward ensures that the strategy only yields a high return when it achieves effective deceleration without significantly compromising tire lateral margin, thereby learning the dynamic coordination rules of "when to start braking, how quickly to increase braking force, and how steering and braking should be coordinated".

[0098] 6. Phase Three Safe Parking Control Once vehicle yaw, side slip, and tire load have all decreased, the Stage 3 controller generates a smooth deceleration target based on remaining vehicle speed, road boundaries, safe stopping area, and actuator capabilities. The target longitudinal velocity can gradually decrease using an exponential or polynomial form.

[0099] in, The moment to enter the safe parking phase, The deceleration coefficient is used. The Stage 3 controller simultaneously limits yaw rate, sideslip angle, lateral deviation, and longitudinal jerk; its reward or cost function can be expressed as:

[0100] in, This refers to longitudinal jerkiness. When the vehicle speed is below the stopping threshold, the yaw rate and the center of gravity sideslip angle are below the stability threshold, and the vehicle is within the safe zone, the control round ends and a safe stopping terminal reward is given.

[0101] Phase 3 may not employ reinforcement learning, but instead use constraint model predictive control or rule control; alternatively, the Phase 2 strategy may be reused and the strategy may be automatically switched to a behavior mode primarily based on deceleration and attitude maintenance by inputting phase variables.

[0102] 7. Optimized Allocation of Safety Motion Projection and Four-Wheel Constraints The equivalent steering angles of the front and rear axles are axle-level generalized control objectives generated by the upper-level controller, and do not imply that the left and right wheels of the same axle adopt the same steering angle. The four-wheel constraint optimization allocation module uses the independent steering angles of the four wheels as optimization decision variables. While satisfying the front and rear axle equivalent steering angle tracking requirements, it combines the target longitudinal resultant force, target yaw moment, vertical load of each wheel, tire adhesion margin, and actuator constraints to solve for the target steering angles of the left front, right front, left rear, and right rear wheels respectively. Therefore, the left and right wheels of the same axle are allowed to have different steering angles.

[0103] To enable the three control stages to share the safe motion projection and four-wheel constraint optimization allocation module, the outputs of each stage are first uniformly converted into candidate generalized control objectives:

[0104] In the tail-flip suppression phase, the equivalent front and rear axle rotation angles obtained from the basic stability controller and residual reinforcement learning are respectively used as... and This phase prioritizes restoring the tire's lateral capacity. Set to zero, minimal braking force requirement, or determined based on tire longitudinal force unloading requirements, and generate according to yaw stability control requirements. .

[0105] During the self-rotation braking phase, the system is directly generated or incrementally updated by the continuous action reinforcement learning strategy. It is used to coordinate four-wheel steering, vehicle deceleration, and yaw recovery.

[0106] During the safe stopping phase, based on the marked target longitudinal acceleration Calculate the resultant longitudinal force of the target vehicle:

[0107] And generated by the path-keeping controller or the stability controller. , as well as At least one of them is used to form a unified candidate generalized control target and input it into the safety action projection module; after obtaining the safety generalized control target that satisfies the safety constraints through safety action projection, it is then input into the four-wheel constraint optimization allocation module.

[0108] 7.1 Safety Action Projection The candidate generalized control objectives generated in the current control phase are first evaluated for safety using a one-step state prediction model. Safety functions are defined for phase plane boundaries, tire margin, yaw rate, center of gravity sideslip angle, roll index, and road boundaries. The safe set is represented as:

[0109] The safety action projection obtains the safety action that is closest to the candidate action by solving the following formula:

[0110] in, This is the predicted vehicle state for the next moment. For safe convergence coefficients, Let be the action deviation weight matrix. If the candidate generalized control objective has satisfied all constraints, then the safe generalized control objective is equal to the candidate generalized control objective; otherwise, it is projected to the safe feasible region with the minimum correction amount.

[0111] 7.2 Four-wheel constraint optimization allocation The generalized control objective, after projection of safety actions, is represented as:

[0112] in, and These are the equivalent rotation angles of the front and rear axle targets, respectively. To achieve the overall longitudinal force of the vehicle, A yaw moment is added to the target. The four-wheel constraint optimization allocation module, under the conditions of satisfying four-wheel steering angle constraints, steering rate constraints, tire adhesion constraints, braking force constraints, and braking force change rate constraints, solves for the independent steering angles of the four wheels and the longitudinal tire forces of the four wheels, so that the actual combined effect of the four wheels tracks the safety generalized control target as closely as possible. .

[0113] The decision variables for the four-wheel constraint optimization allocation module include the four-wheel steering angle and the four-wheel longitudinal tire force:

[0114] The actual equivalent rotation angles of the front and rear axles are calculated as follows:

[0115]

[0116] When there is a large difference in the vertical load between the left and right wheels, the equivalent rotation angle of the front and rear axles can also be calculated by weighting the vertical load.

[0117] For a given candidate four-wheel steering angle and longitudinal tire force, the lateral force of each wheel is not treated as an independent optimization variable, but is calculated based on the current vehicle state and the compound slip tire model:

[0118] in, The first wheel angle is calculated from the vehicle's motion state and the candidate wheel angle. Each wheel slip angle To correspond to the longitudinal sliding state, For vertical loads, This is the road surface adhesion coefficient.

[0119] In one implementation using friction ellipse correction, the lateral force without considering longitudinal force occupancy is first obtained based on a pure sideslip tire model:

[0120] Then, a composite slip correction is performed on the lateral force based on the current longitudinal force:

[0121] This ensures that when the longitudinal braking force increases, the available lateral tire force decreases accordingly, thus reflecting the relationship between the longitudinal force of the tire and the occupancy of lateral adhesion resources.

[0122] Convert the forces of each wheel tire to the vehicle body coordinate system:

[0123]

[0124] The actual longitudinal force of the entire vehicle is:

[0125] The actual yaw moment of the whole vehicle is:

[0126] Therefore, the four-round constraint optimization assignment problem is defined as:

[0127] in, , , , , and Non-negative weights For the first The adhesion utilization rate of each wheel For the current optimization decision variables, This represents the optimization result from the previous control cycle.

[0128] The constraints must include at least:

[0129]

[0130]

[0131]

[0132]

[0133] The optimized solution yields the target steering angles for all four wheels. , , , And the longitudinal tire force of the four-wheel target , , , .

[0134] Wheel-end braking torque is calculated based on the tire's longitudinal force and effective rolling radius:

[0135] in, For the first The target braking torque amplitude of each wheel For the first The effective rolling radius of each wheel For the first The target longitudinal tire force of each wheel in the wheel coordinate system.

[0136] Ultimately, the target steering angle and wheel-end braking torque of the four wheels are tracked by the steering servo controller and the brake servo controller and then applied to the vehicle. If a single wheel steering or braking actuator fails, the optimizer modifies its available control range according to the failure efficiency factor of the corresponding actuator and uses the remaining actuators for fault-tolerant redistribution.

[0137] 8. Strengthen learning and training and vehicle-side deployment 8.1 Rapid Training Environment To improve training efficiency, this invention employs a vehicle dynamics model with collision force input, a composite slip tire model, a vertical load transfer model, and a 4WIS / WIB actuator model to construct a rapid training environment. The collision force can be generated using a triangular pulse, a half-sine pulse, a Kelvin contact model, or a finite contact time impulse model based on momentum conservation.

[0138] During training, the following parameters are randomized: initial velocity of the impacted vehicle, velocity of the colliding vehicle, collision location, collision angle, peak collision force, collision duration, road adhesion coefficient, vehicle mass, center of gravity position, yaw moment of inertia, tire parameters, steering delay, braking delay, sensor noise, and single-wheel actuator efficiency. This parameter randomization generates diverse post-collision states ranging from mild to severe and from high to low adhesion, improving the strategy's robustness to unknown conditions.

[0139] 8.2 Phased Course Learning The training process employs a phased learning approach. First, phase variables are fixed, and the tail-flick suppression strategy and the spin-braking coordination strategy are trained separately. Then, a phase supervisor is activated to perform joint training or fine-tuning during the complete post-collision process. In the early stages of training, safe samples generated by the base controller can be used to fill the experience replay buffer, gradually increasing the collision intensity and the random range of parameters to prevent the strategy from frequently entering unrecoverable states before it has achieved basic stability.

[0140] Continuous action reinforcement learning algorithms can employ dual-delay deep deterministic policy gradient, soft actor commentator, proximal policy optimization, distributed actor commentator, or other reinforcement learning algorithms suitable for continuous control. Taking soft actor commentator as an example, the policy is trained by maximizing the discounted cumulative reward and policy entropy; the value network estimates the long-term rewards of different state-action pairs. Security costs can be introduced into the training through Lagrange multipliers, penalty functions, feasible region projection, or security experience screening.

[0141] 8.3 Termination Conditions and Security Costs The training round terminates and a large negative reward is given when a secondary collision occurs, the vehicle crosses the road boundary, the roll index exceeds the threshold, the tire adhesion utilization rate remains significantly greater than 1, the turning angle or braking force exceeds the hardware limit, or the vehicle remains in an irreversible spin state.

[0142] Each cost corresponds to road boundary, tire adhesion, roll, actuator, and secondary collision constraints, respectively.

[0143] 8.4 High-fidelity verification and deployment After the rapid model training is completed, the strategy is connected to the CarSim and MATLAB / Simulink co-simulation platform for high-fidelity verification. It is compared with no control, differential braking, traditional four-wheel steering, fixed weight coordinated control and reinforcement learning strategy without safety projection under different collision positions, vehicle speeds, adhesion coefficients, sensor noise and actuator delay conditions.

[0144] During vehicle-side deployment, the policy network parameters and state normalization parameters are frozen. Each control cycle sequentially executes state acquisition, three-layer controllability margin calculation, stage determination, policy network forward inference, safety action projection, four-wheel optimization allocation, and actuator servo control. The policy network outputs only low-dimensional generalized actions; the four-wheel-end commands are calculated in real-time by the optimizer, reducing the network output dimensionality and improving verifiability and real-time performance.

[0145] During the aforementioned phase switching process, in order to avoid sudden changes in control commands caused by phase switching, the next phase uses the vehicle state at the time of switching and the generalized control target or actuator command of the previous control cycle as initial conditions to achieve continuous transition of control state between phases.

[0146] The technical features described in the above embodiments can be combined without contradicting each other. For those skilled in the art, any equivalent substitutions, modifications, or improvements made without departing from the inventive concept should be included within the scope of protection of this invention, which is defined by the claims.

Claims

1. A method for self-rotation braking safety control after a collision of a four-wheel independent steering steer-by-wire chassis, characterized in that, Includes the following steps: S1, obtain the vehicle motion state, four-wheel tire state, collision triggering and working condition information, road adhesion information and actuator state after the collision, and construct a two-dimensional side slip angle phase plane using the front axle equivalent side slip angle and the rear axle equivalent side slip angle as two coordinate variables; S2, Based on the vehicle operating conditions, generate the dynamic boundary of the two-dimensional sideslip angle phase plane online, and calculate the controllability margin of the state layer, the controllability margin of the tire layer, and the controllability margin of the whole vehicle layer respectively. S3, based on the three-layer controllable margin, through a stage supervisor with hysteresis threshold, continuous satisfaction time and minimum dwell time constraints, the post-collision control process is divided into a tail-slip suppression stage, a spin-braking stage and a safe stopping stage, and stage forward switching or backward switching is executed according to the changes in the three-layer controllable margin. S4, in the tail-slip suppression stage, a basic stability controller is used to generate the reference front and rear axle equivalent steering angles, and a residual reinforcement learning strategy is used to generate a restricted steering angle correction amount, which limits the intervention of braking force to reduce vehicle yaw and sideslip and restore the tire's lateral available capacity, forming the candidate generalized control target for this stage. S5, during the self-rotation braking phase, a continuous action reinforcement learning strategy is used to generate or incrementally update at least two of the following: front axle candidate equivalent steering angle, rear axle candidate equivalent steering angle, candidate vehicle longitudinal resultant force, and candidate additional yaw moment, including at least one steering control objective and one braking or yaw control objective, forming the candidate generalized control objective for this phase. S6, during the safe parking phase, a smooth deceleration and attitude maintenance target is generated based on the remaining vehicle speed, road boundaries, and safe parking area, forming a candidate generalized control target for this phase; S7 projects the candidate generalized control objective output from the current stage onto the safe and feasible domain that satisfies the vehicle state boundary, tire adhesion constraint, road boundary and actuator constraint to obtain the safe generalized control objective, and obtains the four-wheel independent steering angle and four-wheel braking force commands through four-wheel constraint optimization allocation.

2. The method according to claim 1, characterized in that, The controllable margin of the state layer includes at least the phase plane instability degree and the phase plane instability change rate; the controllable margin of the tire layer includes at least the maximum adhesion utilization rate of the four wheels, the minimum lateral residual margin of the rear axle, and the saturation degree of the lateral force of the rear axle; the controllable margin of the whole vehicle layer includes the yaw recovery capability coefficient of the maximum achievable recoverable yaw moment relative to the required recoverable yaw moment under the current tire adhesion constraints and actuator constraints; the dynamic boundary is determined by one or more of the following: longitudinal speed, road adhesion coefficient, equivalent rotation angle of the front and rear axles, longitudinal braking force occupation degree, and vehicle structural parameters, and is generated online using a radial basis function neural network optimized by a genetic algorithm, a regression network, or a lookup table interpolation method.

3. The method according to claim 1, characterized in that, The stage supervisor uses different entry and exit thresholds to form a hysteresis loop, and combines continuous satisfaction time and minimum stage dwell time limit stage frequent switching; when the vehicle is in the self-rotation braking stage or the safe stopping stage and meets any entry condition of the tail-slip suppression stage again, it reverts to the tail-slip suppression stage. When the vehicle is in the safe stopping phase, the stability margin decreases but the entry conditions for the fishtail suppression phase are not yet met and the conditions for maintaining the safe stopping phase are no longer met, it reverts to the self-rotation braking phase.

4. The method according to claim 1, characterized in that, The basic stability controller in the tail-slip suppression stage is a sliding mode controller, model prediction controller, linear quadratic controller, backstepping controller, or robust controller. The output of the residual reinforcement learning strategy is superimposed on the basic stability control quantity after amplitude scaling, angle limiting, and angle rate limiting. In this stage, the candidate vehicle longitudinal resultant force is set to zero, the braking force requirement is small, or it is determined according to the tire longitudinal force unloading requirement, so as to prioritize the recovery of tire lateral capacity.

5. The method according to claim 1, characterized in that, The self-rotation braking phase changes the rotation angle of at least one wheel, so that the longitudinal braking force in the wheel coordinate system forms a lateral component in the vehicle coordinate system to suppress yaw error and / or restore yaw moment; the candidate equivalent rotation angles of the front axle and the candidate equivalent rotation angles of the rear axle are axle-level generalized control targets, and do not limit the left and right wheels of the same axle to the same rotation angle. The target rotation angles of each of the four wheels are determined by the four-wheel constraint optimization allocation.

6. The method according to claim 1, characterized in that, The safety action projection is obtained by minimizing the weighted distance between the safety generalized control target and the candidate generalized control target, and is constrained to ensure that the predicted vehicle state at the next moment is within the phase plane safety boundary, the tire force satisfies the friction ellipse, the vehicle does not cross the road boundary, and the actuator does not exceed the amplitude and rate of change limits; when the candidate generalized control target satisfies all constraints, the safety generalized control target is equal to the candidate generalized control target.

7. The method according to claim 1, characterized in that, The four-wheel constraint optimization allocation uses the independent steering angles of the four wheels and the longitudinal tire forces of the four wheels as decision variables. The objective function is composed of at least two of the following: the equivalent steering angle tracking error of the front axle, the equivalent steering angle tracking error of the rear axle, the longitudinal resultant force tracking error of the whole vehicle, the yaw moment tracking error, the four-wheel adhesion utilization rate, and the control command change rate. The four-wheel steering angle, steering rate, braking force, braking force change rate, and tire adhesion constraints are used as constraints. After optimization, the target longitudinal tire force is converted into wheel-end braking torque.

8. The method according to claim 1, characterized in that, The reinforcement learning strategy is trained offline in a vehicle dynamics training environment with collision force input. During the training process, at least three of the following are randomly changed: collision position, collision direction, collision intensity, collision duration, initial vehicle speed, road surface adhesion coefficient, vehicle parameters, tire parameters, sensor noise, actuator delay, and actuator failure efficiency. The training adopts a phased course learning approach, and after high-fidelity co-simulation verification, the strategy network parameters and state normalization parameters are frozen for vehicle-side deployment.

9. A four-wheel independent steering steer-by-wire chassis rear-collision stability control system, characterized in that, The system includes a state acquisition unit, a dynamic phase plane boundary generation unit, a three-layer controllability margin evaluation unit, a stage supervision unit, a reinforcement learning policy inference unit, a safety action projection unit, a four-wheel constraint optimization allocation unit, and a four-wheel steering and braking execution unit, and is configured to perform the method described in any one of claims 1 to 8.

10. A vehicle, characterized in that, The vehicle has a four-wheel independent steering actuator and a four-wheel independent braking actuator, and is equipped with the post-collision stability control system as described in claim 9.

Images