Self-evolution method and system for vehicle stability control based on self-closed loop continuous learning

Abstract

The application provides a vehicle stability control self-evolution method and system based on a self-closed loop continuous learning, and relates to the technical field of vehicle active safety control. In the simulation stage, the method uses a teacher optimizer to generate a recovery reference control amount, and distills the reference control amount into a residual model relative to a basic controller; in the real vehicle stage, the teacher optimizer is not deployed, and the complete chassis controller is not retrained, but the residual model parameters, stability error weights and residual participation ratios are continuously corrected under safety constraints and version management, so that the real vehicle differences are absorbed while the basic function of the basic controller is maintained. The application is not limited to a single yaw moment control, but the target yaw moment, the rear wheel steering correction angle and the active front wheel steering fine adjustment amount constitute a generalized stability control instruction, and the target yaw moment is further distributed to the wheel end longitudinal force, and the continuous improvement of the control effect is realized through residual compensation, safety projection and closed loop version management.

Description

Technical Field

[0001] This invention relates to the field of vehicle active safety control technology, and in particular to a self-evolving method and system for vehicle stability control based on self-closed-loop continuous learning. Background Technology

[0002] Existing vehicle stability control typically focuses on the target yaw moment, suppressing yaw rate errors through differential braking or drive torque distribution. While this approach has high engineering applicability on traditional vehicle platforms, for intelligent chassis with rear-wheel steering, active front-wheel steering, or multi-actuator coordination capabilities, constructing the controller solely around the yaw moment weakens the control strategy's comprehensive ability to adjust for factors such as sideslip angle, high-speed stability, low-adhesion response, and actuator margin.

[0003] Model prediction optimization, reinforcement learning, or offline policy search can find more restorative control actions in simulation environments. However, if such optimizers are directly used as online stability controllers for real vehicles, the requirements for prediction model accuracy, real-time computing power, and constraint feasibility are high. When the vehicle has tire wear, load changes, road adhesion changes, actuator hysteresis, and sensor errors, the online optimization output is easily affected by model mismatch.

[0004] The learned controller obtained through simulation training will also be affected by the aforementioned differences in a real vehicle. If the learner directly outputs complete chassis control commands, the safety risk is high; if it only describes the continuous learning process without specific control formulas and version constraints, it is difficult to explain its specific technical improvements compared to existing controllers, and it is also difficult to guarantee that the learning process will not disrupt existing stable control boundaries.

[0005] Therefore, a vehicle stability control scheme that can distinguish between offline teacher optimization and real-vehicle online execution is needed. Summary of the Invention

[0006] Purpose of the Invention: This invention aims to provide a self-evolving method and system for vehicle stability control based on self-closed-loop continuous learning. In the simulation phase, a teacher optimizer is used to generate a restorative reference control quantity, which is then distilled into a residual model relative to the base controller. In the real vehicle phase, the teacher optimizer is not deployed, nor is the complete chassis controller retrained. Instead, the residual model parameters, stability error weights, and residual participation ratios are continuously corrected under safety constraints and version management, thereby absorbing the differences of real vehicles while maintaining the baseline function of the base controller.

[0007] To achieve the above objectives, this invention provides a self-evolving method for vehicle stability control based on self-closed-loop continuous learning, comprising the following steps:

[0008] S1: Construct a set of simulation training scenarios, run the basic generalized stable controller in the set of simulation training scenarios to obtain the basic generalized stable control quantity; use the teacher / reference optimizer to generate a restorative reference control quantity, and use the difference between the restorative reference control quantity and the basic generalized stable control quantity as the simulation residual label; train the initial residual model based on the simulation residual label. S2: Collect vehicle state data to form a vehicle state vector, calculate yaw error and sideslip error, and calculate composite stability error based on the yaw error and sideslip error; S3: The basic generalized stability controller outputs a baseline generalized stability control quantity based on the composite stability error; the initial residual model outputs a generalized residual compensation quantity based on the vehicle state vector and the composite stability error; S4: Construct difference samples based on the actual vehicle closed-loop response, update the generalized residual learning parameter matrix in the candidate policy replica based on the difference samples, and correct the error weights and residual participation ratios; S5: Perform a safe projection on the generalized residual compensation to obtain restricted residual compensation; superimpose the basic generalized stability control quantity and the restricted residual compensation according to the residual participation ratio to obtain the final generalized stability control instruction; allocate multiple actuators according to the final generalized stability control instruction.

[0009] As a preferred embodiment, in step S1, the basic generalized stability control quantity and the restorative reference control quantity both include yaw moment, rear wheel steering correction angle and active front wheel steering fine-tuning quantity.

[0010] As a preferred embodiment, the generation of restorative reference control quantities using a teacher / reference optimizer specifically includes: A finite prediction window is established, and the impact of candidate control variables on the stability state is predicted using a simulated vehicle dynamics model. The teacher / reference optimizer constructs a recovery reference optimization objective function based on the recovery stability error weight matrix, the control deviation weight matrix, and the control rate of change weight matrix, obtains the control sequence within the safe and feasible control set, and takes the first step of the control sequence as the recovery reference control quantity.

[0011] As a preferred embodiment, in step S2, the yaw error Lateral deviation error and composite stability error The calculation formula is:

[0012]

[0013]

[0014] In the formula, r represents the yaw rate; This represents the desired yaw rate; Indicates the centroid sideslip angle; Indicates the expected centroid sideslip angle; Indicates the weight of the yaw rate error; This represents the weight of the sideslip angle error.

[0015] As a preferred embodiment, in step S3, the basic generalized stability controller calculates the basic generalized stability control quantity based on the composite stability error s. The basic generalized stability control quantity Including the basic target yaw moment Basic rear wheel steering correction angle and basic active front wheel steering fine-tuning The calculation formula is:

[0016]

[0017]

[0018]

[0019] In the formula, This indicates the yaw rate error feedback gain; This indicates the centroid sideslip angle error feedback gain; Indicates the composite stability error feedback gain; and These represent the yaw rate error gain and sideslip angle error gain of the rear wheel steering channel, respectively. and These represent the yaw rate error gain and sideslip angle error gain of the active front wheel steering channel, respectively. Indicates the upper limit of the rear wheel steering correction angle; This indicates the upper limit of the active front wheel steering fine-tuning amount; This indicates that a boundary layer saturation function is applied to the composite stability error s. , This represents the boundary layer width, and -1 and 1 represent the lower and upper bounds of the normalized output.

[0020] As a preferred embodiment, in step S4, the product of the feature vector and the generalized residual learning parameter matrix is ​​used as the generalized residual compensation amount; In the candidate policy replicas, the generalized residual learning parameter matrix is ​​updated using recursive least squares with a forgetting factor, specifically including:

[0021]

[0022]

[0023]

[0024] In the formula, This indicates a recursive update of the gain; and These represent the parameter covariance matrices before and after the update, respectively. Indicates the forgetting factor; This represents the prediction error of the generalized control residual; and These represent the candidate parameter matrices before and after the update, respectively. Indicates the safety parameter field The projection operator; wherein the feature vector is the feature vector of the t-th difference sample, and the generalized residual learning objective is the difference between the recovered generalized control quantity and the basic generalized stability control quantity corresponding to the t-th difference sample.

[0025] As a preferred embodiment, step S5 involves performing a secure projection on the generalized residual compensation amount, specifically including: Calculate the residual reserve of yaw moment Rear wheel steering residual reserve and active front steering residual reserve :

[0026]

[0027]

[0028] In the formula, y represents the lateral position vector of the four wheels relative to the vehicle's center of gravity; R represents the available longitudinal force margin vector of the four wheels. and These represent the upper limit of the rear wheel steering angle rate and the upper limit of the active front wheel steering fine-tuning rate, respectively. Indicates the control cycle; Indicates the base rear wheel steering correction angle; This indicates the basic active front wheel steering fine-tuning amount; For the candidate yaw moment residuals provided by the generalized residual compensation amount Candidate rear wheel steering residual and candidate active front wheel steering residual Saturation projections were performed separately to obtain the safety residual compensation:

[0029]

[0030]

[0031]

[0032] In the formula, , and These represent the corresponding residuals after safe projection; This represents the generalized residual compensation after safe projection.

[0033] As a preferred embodiment, in step S5, the basic generalized stability control quantity is... With the aforementioned restricted residual compensation According to the aforementioned residual participation ratio Superposition yields the final generalized stability control command. :

[0034] In the formula, , , and These represent the residual participation ratios of yaw moment, rear-wheel steering, and active front-wheel steering, respectively.

[0035] As a preferred embodiment, step S5, which involves allocating multiple actuators according to the final generalized stability control instruction, specifically includes: The yaw moment component is extracted from the final generalized stability control command. A four-wheel longitudinal force distribution formula is established based on the lateral position of the wheels. The four-wheel longitudinal force correction vector is calculated and converted into the drive torque correction amount and brake pressure correction amount sent to the vehicle actuator. The extracted steering correction component is sent to the corresponding steering actuator. After allocating multiple executors according to the final generalized stability control instruction, the process further includes: performing closed-loop version evaluation on candidate strategies that include the updated generalized residual learning parameter matrix, error weights, and residual participation ratios; and making the candidate strategy a new stable version when the preset safety and performance upgrade conditions are met, otherwise rolling back to the current stable version.

[0036] Furthermore, this invention proposes a self-evolving vehicle stability control system based on self-closed-loop continuous learning. This system is used to execute the aforementioned self-evolving vehicle stability control method based on self-closed-loop continuous learning. Specifically, this system includes: The simulation pre-training unit is used to construct a set of simulation training scenarios, run a basic generalized stable controller in the set of simulation training scenarios to obtain basic generalized stable control quantities; generate restorative reference control quantities using a teacher / reference optimizer, and use the difference between the restorative reference control quantities and the basic generalized stable control quantities as simulation residual labels; train an initial residual model based on the simulation residual labels. The basic generalized stability control unit is used to collect vehicle state data to form a vehicle state vector, calculate yaw error and sideslip error, calculate composite stability error based on the yaw error and sideslip error, and output a basic generalized stability control quantity with a minimum guarantee based on the composite stability error. The candidate residual learning unit is used to output a generalized residual compensation amount from the initial residual model based on the vehicle state vector and the composite stability error. The real vehicle difference sample construction unit is used to construct difference samples based on the real vehicle closed-loop response, update the generalized residual learning parameter matrix in the candidate policy replica based on the difference samples, and correct the error weights and residual participation ratios. A safety projection unit is used to perform a safety projection on the generalized residual compensation amount to obtain a restricted residual compensation; the basic generalized stability control amount and the restricted residual compensation are superimposed according to the residual participation ratio to obtain the final generalized stability control command; A multi-actuator allocation unit is used to allocate multiple actuators according to the final generalized stability control instruction; The closed-loop version management unit is used to perform closed-loop version evaluation on candidate strategies that include the updated generalized residual learning parameter matrix, error weights, and residual participation ratios; when the preset security and performance upgrade conditions are met, the candidate strategy becomes a new stable version; otherwise, it is rolled back to the current stable version.

[0037] Beneficial Effects: This invention differs from direct online model predictive control schemes and direct end-to-end online learning control schemes. The teacher / reference optimizer is only responsible for providing more restorative training labels in the offline phase; the actual vehicle executes a combination of control quantities from the basic controller and limited residual compensation. The learner only learns the residual compensation and does not directly generate complete chassis control commands. Self-evolution is not an arbitrary modification of the controller, but is limited to limited updates of the residual parameter matrix, yaw rate error weight, center of gravity sideslip angle error weight, and residual participation ratio. This invention is applicable to vehicle platforms with differential braking, distributed drive, rear-wheel steering, active front-wheel steering, or a combination thereof. The stability control proposed in this invention is not limited to single yaw moment control, but constitutes a generalized stability control command by combining the target yaw moment, rear-wheel steering correction angle, and active front-wheel steering fine-tuning, further distributing the target yaw moment as wheel-end longitudinal forces, and achieving continuous improvement of control performance through residual compensation, safety projection, and closed-loop version management. Attached Figure Description

[0038] Figure 1 This is a flowchart of the self-evolutionary method for vehicle stability control based on self-closed-loop continuous learning, as described in this invention. Detailed Implementation

[0039] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described to avoid obscuring the invention. The term "comprising" as used herein indicates non-exclusive inclusion.

[0040] This invention discloses a self-evolving method and system for vehicle stability control based on self-closed-loop continuous learning, applicable to vehicle platforms with differential braking, distributed drive, rear-wheel steering, active front-wheel steering, or a combination thereof. The stability control is not limited to single yaw moment control, but rather constructs a generalized stability control command from the target yaw moment, rear-wheel steering correction angle, and active front-wheel steering fine-tuning. The target yaw moment is further distributed as wheel-end longitudinal forces, and continuous improvement of the control effect is achieved through residual compensation, safety projection, and closed-loop version management.

[0041] In one embodiment, the system of the present invention includes a basic generalized stability control unit, a simulation pre-training unit, a real-vehicle difference sample construction unit, a candidate residual learning unit, a safety projection unit, a multi-actuator allocation unit, and a closed-loop version management unit. The basic generalized stability control unit is always reserved as a fallback option; the simulation pre-training unit is used to form initial residual model parameters; the real-vehicle difference sample construction unit is used to extract learning samples from the actual closed-loop control effect; the candidate residual learning unit only updates the policy copy; the safety projection unit determines whether candidate compensations can participate in the final control; and the closed-loop version management unit determines version advancement or rollback based on shadow verification and grayscale takeover results.

[0042] The following embodiments use vehicles with rear-wheel steering, active front-wheel steering, four-wheel braking, and distributed drive capabilities as examples. For vehicles without a particular actuator, the corresponding channel can be set to zero or closed, but the main lines of generalized stability control commands, residual learning, safety projection, and execution allocation remain unchanged.

[0043] The overall technical concept of this invention is: instead of directly transferring the simulation optimizer to the real vehicle for online solution, and instead of letting the learner directly replace the basic stability controller, the restorative control capability obtained by offline optimization is converted into an interpretable residual compensation model, and then the residual compensation model is subjected to restricted evolution by the closed-loop evaluation of the real vehicle.

[0044] This concept forms a two-stage closed loop. The first stage is the offline teacher closed loop, where the teacher / reference optimizer generates restorative reference control quantities based on stability error, attachment utilization, and actuator constraints in the simulation scenario. The difference between the reference control quantity and the basic control quantity is used as the residual label to train the initial residual model. The second stage is the real-vehicle self-evolution closed loop, where difference samples are constructed based on the closed-loop stability effect during real vehicle operation. The residual learning parameters, error weights, and residual participation ratios are adjusted in candidate versions. Only after safe projection, shadow verification, and grayscale takeover are new stable versions allowed to be formed.

[0045] First, a generalized stability control command is defined that simultaneously includes the target yaw moment, the rear wheel steering correction angle, and the active front wheel steering fine-tuning, and this command is represented using control variable symbols commonly used in the control field.

[0046]

[0047] In the formula, Indicates the first Vehicle state vector for each control cycle; Indicates longitudinal vehicle speed; Indicates lateral vehicle speed; Indicates yaw rate; Indicates the centroid sideslip angle; Indicates lateral acceleration; This represents the estimated value of road adhesion. Indicates the actual steering angle of the front wheels; Indicates the actual steering angle of the rear wheels; to This indicates the wheel speed of the four wheels; Represents the generalized stability control quantity; Indicates the target yaw moment; This indicates a command to correct the rear wheel steering angle. This indicates a command for active front wheel steering fine-tuning. This represents the final generalized stability control quantity; Represents the basic generalized stability control quantity; This represents the residual participation ratio matrix; The generalized residual compensation amount after safe projection is shown in equations (20) and (21).

[0048] To ensure that the base controller and learner share the same stability objective, the desired yaw rate is calculated based on the steady-state steering response and adhesion constraints.

[0049]

[0050] In the formula, This represents the desired yaw rate; Indicates the vehicle's wheelbase; Indicates the steady-state understeering coefficient; This indicates the upper limit of yaw rate obtained from road adhesion limitations; Represents gravitational acceleration; This indicates the speed quantity used to prevent division by zero at low speeds; This indicates the preset minimum calculated vehicle speed; Represents a saturation function; This indicates a function that takes the larger value.

[0051] To simultaneously reflect the yaw response deviation and the risk of centroid lateral deviation, a composite stability error is defined.

[0052]

[0053] In the formula, Indicates the yaw rate error; Indicates the error of the centroid sideslip angle; This represents the desired centroid sideslip angle, which can be zero in one implementation. Indicates the composite stability error; Indicates the weight of the yaw rate error; This represents the weighting of the sideslip angle error. This is achieved by setting the weighting separately. and It can adjust the relative importance of yaw response and sideslip risk for different vehicle speeds, adhesion, and overall vehicle configurations.

[0054] To avoid discontinuous output from the basic controller in the critical steady state, a boundary layer saturation function is used for the composite error.

[0055]

[0056] In the formula, Represents the boundary layer saturation function; The composite stability error is defined by expression (3); Indicates the width of the boundary layer; and This represents the lower and upper bounds of the normalized output.

[0057] To provide a baseline yaw stability control, the base controller calculates the target yaw moment of the base based on the composite stability error.

[0058]

[0059] In the formula, Indicates the basic target yaw moment; This indicates the yaw rate error feedback gain; This indicates the centroid sideslip angle error feedback gain; Indicates the composite stability error feedback gain; The boundary layer saturation function is defined by expression (4).

[0060] To improve sideslip angle and yaw response by utilizing both rear-wheel steering and active front-wheel steering, the base controller simultaneously calculates two steering correction channels.

[0061]

[0062] In the formula, Indicates the base rear wheel steering correction angle; This indicates the basic active front wheel steering fine-tuning amount; and These represent the yaw rate error gain and sideslip angle error gain of the rear wheel steering channel, respectively. and These represent the yaw rate error gain and sideslip angle error gain of the active front wheel steering channel, respectively. Indicates the upper limit of the rear wheel steering correction angle; This indicates the upper limit of the active front wheel steering fine-tuning. When the vehicle does not have rear wheel steering or an active front wheel steering mechanism, the corresponding channel is set to zero.

[0063] To enable unified processing of multiple types of stability control variables in subsequent learning, projection, and allocation, the basic yaw moment and steering correction are combined into a basic generalized stability control command.

[0064]

[0065] In the formula, Represents the basic generalized stability control quantity; The basic target yaw moment in expression (5); and The basic rear wheel steering correction angle and the basic active front wheel steering fine adjustment amount in equation (6) are respectively expressed.

[0066] During the simulation phase, this invention preferably employs an interpretable feature-parameterized residual model, rather than directly using an end-to-end black-box strategy as the actual vehicle controller. This residual model uses stability error, adhesion margin, actuator margin, steering input, and vehicle speed as input features, and compensation amounts relative to the basic generalized stable control variables as outputs. Reinforcement learning, offline policy optimization, or model prediction optimization are only used as simulation teacher tools to generate the recovery control objective; the initial model deployed on the actual vehicle is still organized into a residual compensation model in parameter matrix form.

[0067] In this embodiment, the present invention does not use online model predictive control as the final controller for the actual vehicle. Model predictive control relies on a high-precision vehicle prediction model and real-time rolling optimization solution. When tire wear, load changes, actuator hysteresis, or sensor noise deviates from the calibration model, problems such as increased solution time, infeasible constraints, or output sensitivity to model mismatch can easily occur. The present invention utilizes its ability to find the recovery direction in the simulation or offline stage, and then distills the recovery direction into an initial residual parameter matrix and a projectable residual compensation channel; in the actual vehicle stage, the control quantity is executed after constraints from the basic controller, residual compensation, safety projection, and version management. The calculation form is simple and easy to calibrate and rollback.

[0068] To ensure that the initial residual model covers the typical stability differences that a real vehicle may encounter, a multi-condition training scenario set is first constructed in the simulation environment.

[0069]

[0070] In the formula, This represents a set of simulation training scenarios; Indicates the speed range; Indicates the actual steering angle of the front wheels; Indicates the extent of road attachment; Indicates the first The load variation parameter vector for each simulation scenario Indicates the overall vehicle weight. This indicates the distance from the center of gravity to the front axle. This represents the distance from the center of mass to the rear axle. Indicates the height of the center of mass; Indicates the first The equivalent tire parameter vector for each simulation scenario. This indicates the equivalent lateral stiffness of the front axle. This indicates the equivalent lateral stiffness of the rear axle. This represents the equivalent longitudinal slip stiffness used to characterize the response to braking or driving longitudinal forces. This represents the adhesion correction factor used to characterize tire wear, temperature, or local road surface differences; Represents the dynamic parameter vector of the actuator. This represents the braking pressure response time constant. Indicates the drive torque response time constant. Represents the rear wheel steering response time constant. Indicates the active front wheel steering response time constant. Indicates the upper limit of the rear wheel steering angular velocity. This indicates the upper limit of the active front wheel micro-adjustment angular velocity; Represents the sensor noise parameter vector. This indicates the standard deviation of noise in vehicle speed measurement. This indicates the standard deviation of noise in yaw rate measurement. This indicates the standard deviation of noise in lateral acceleration measurements. This indicates the standard deviation of noise in wheel speed measurements; Indicates the first The combination of vehicle, tire, actuator, and sensor parameters corresponding to each scenario; Indicates the first The simulation scenario Vehicle status for each cycle; Represents a simulated vehicle dynamics model; This represents the generalized stability control variable applied during the simulation.

[0071] To illustrate that the teacher / reference optimizer is not arbitrarily given, normalized weights are first constructed based on the vehicle stability safety boundary and actuator capabilities, so that yaw, sideslip, adhesion, and control deviation are under comparable dimensions.

[0072]

[0073] In the formula, This represents the error weight matrix for restoring stability. This represents the control deviation weight matrix; Represents the control rate of change weight matrix; , and These represent the basic weights for yaw rate error, centroid sideslip angle error, and adhesion utilization risk, respectively. Indicates the yaw rate safety boundary; Indicates the safety boundary of the centroid sideslip angle; Indicates the maximum upper limit of adhesion utilization; , and These represent the basic weights for yaw moment, rear-wheel steering, and deviation from active front-wheel steering control, respectively. , and These represent the upper limit of the target yaw moment, the upper limit of the rear wheel steering correction angle, and the upper limit of the active front wheel fine-tuning angle, respectively. , and These represent the basic weights of the rate of change of the three control channels mentioned above; , and These represent the upper limits of allowable yaw moment variation, rear wheel steering correction angle variation, and active front wheel fine-tuning angle variation for adjacent control cycles, respectively.

[0074] To obtain a more restorative reference control during the simulation phase, the teacher / reference optimizer establishes a finite prediction window on each simulation sample and uses the simulation model to predict the impact of candidate control variables on the stability state.

[0075]

[0076] In the formula, This represents the initial state of the prediction window, which is taken from the current simulation state. ; Indicates the first [number]th ... Predicting vehicle status step by step; Indicates the first [number]th ... Candidate generalized stability control variables for each step; Indicates the prediction step size; This indicates the recovery of the optimization error vector; Indicates the first [number]th ... The yaw rate of the step; Indicates the first [number]th ... The expected yaw rate of the step; Indicates the first [number]th ... The centroid sideslip angle of the step; Indicates the first [number]th ... The expected centroid sideslip angle of the step can be taken as zero in one implementation. Indicates the first [number]th ... Maximum adhesion utilization rate of the step; This represents the control sequence consisting of all candidate control variables within the prediction window.

[0077] After the prediction model and weights are determined, the teacher / reference optimizer obtains the control sequence through constrained optimization, and only takes the first step of the sequence as the simulation teacher control quantity to generate residual labels.

[0078]

[0079] In the formula, This indicates the change in the candidate control quantity between two adjacent steps within the prediction window; Let h be the candidate control variable; Candidate control variables at time h-1; This represents the basic control quantity given by the basic generalized stable controller within the prediction window; This indicates the objective function for restoring the reference optimization. This represents the set of feasible control sequences for restoring the reference optimization; and These represent the lower and upper limits of the generalized stability control quantity, respectively. The upper and lower limits simultaneously constrain the target yaw moment, the rear wheel steering correction angle, and the active front wheel fine-tuning amount. This represents the upper limit of the rate of change of the generalized stability control variable; Indicates the first The simulation scenario The recovery reference control sequence obtained from each cycle; This represents the feasible control sequence that minimizes the objective function; This represents the selection matrix for extracting the first control quantity from the control sequence. This represents the simulated teacher's control parameters; Represents the residual labels used for supervised training during the simulation phase; This represents the output of the basic generalized stable controller under the same simulation conditions.

[0080] The aforementioned teacher / reference optimizer is constructed based on vehicle stability objectives, actuator capabilities, and safety constraints. It can be solved offline using methods such as sequential quadratic programming, dynamic programming, batch model prediction optimization, or simulation-based reinforcement learning search. Regardless of the solver used, the output must satisfy... The constraints in the text, and only used for training. The teacher sample is used. The teacher optimizer is not run during actual vehicle deployment, nor is it... Send it directly to the implementing agency.

[0081] When the optimization solution for a certain simulation sample fails to converge or is infeasible, that sample is not included in the residual training set; alternatively, a high recovery gain control law can be used to first generate candidate yaw moment, rear wheel steering correction angle, and active front wheel fine-tuning, and then project them onto the residual training set. A backup teacher control variable is formed. Therefore, the role of the teacher / reference optimizer is to provide more restorative training labels relative to the base controller, rather than to replace the actual vehicle self-evolutionary control link of this invention.

[0082] To ensure that the initial residual model has a clear input-output relationship, the residual model is written as the product of the eigenvectors and the parameter matrix.

[0083]

[0084] In the formula, Represents the residual learning feature vector during the simulation phase; Indicates the composite stability error; Indicates attachment margin; Indicates actuator margin; This represents the residual learning parameter matrix to be trained; Represents the parameter matrix The generalized residual control quantity output on the simulation sample.

[0085] To obtain the initial residual model parameters before deployment to the actual vehicle, the initial parameter matrix is ​​trained based on the simulation residual labels.

[0086]

[0087] In the formula, This represents the simulation residual prediction error; This represents the set of simulation training sample indices, determined by the simulation scene number. and time index composition; Indicates the first A set of training time indices in a simulation scenario; Indicates the first The single-sample training loss corresponding to each simulated sample; This represents the simulation training objective function; Represents a set Sum of all sample indices in the dataset; This represents a matrix used for dimension normalization and weight allocation of the yaw moment, rear wheel steering, and active front wheel steering residuals; Indicates the regularization weights of the parameters; Denotes the Frobenius norm of the parameter matrix; Indicates the weight of residual output smoothness; This represents the parameter matrix that minimizes the simulation training objective function; Indicates the parameter security domain The projection operator; This represents the initial residual model parameter matrix obtained from simulation training.

[0088] After simulation training is completed, the vehicle controller loads the initial residual model parameter matrix and the initial calibration evolution parameters. In the real vehicle phase, the complete chassis controller is not retrained, nor is the basic stability controller directly replaced. Instead, the residual learning parameter matrix, error weights, and residual participation ratio are continuously corrected within the safety parameter domain.

[0089] To ensure that continuous learning is not an unconstrained replacement of the basic controller, but rather a learning of the shortcomings of the basic controller exposed in real vehicles, a more restorative generalized control objective is first generated and a residual learning objective is formed.

[0090]

[0091] In the formula, Indicates the restoration of the target yaw moment; This indicates that the rear wheel steering correction angle has been restored. Indicates the restoration of active front wheel steering fine-tuning; superscript This indicates that the corresponding control quantity or feedback gain belongs to the restoring control law; This represents the generalized residual learning objective, which is equal to the difference between the recovered generalized control variable and the basic generalized stable control variable.

[0092] The above It is only used for calculating parameter update errors in the candidate residual learner and is not directly sent to the actuator. The final control quantity of the actual vehicle is still formed by the basic control quantity, the candidate residual, the constrained residual after safe projection, and the residual participation ratio; therefore, even if the recovery target is obtained using the offline optimization approach, the actual vehicle control link of this invention is not directly online model predictive control.

[0093] To ensure that the real vehicle continuously learns from stability difference samples rather than ordinary driving logs, event filtering conditions are set for the data entering the learning cache.

[0094]

[0095] In the formula, Indicates the first Closed-loop difference samples formed by each control cycle; This indicates that the actual vehicle is continuously learning from the cache. This indicates the maximum adhesion utilization rate of the four wheels; This represents the minimum stability error threshold required to enter the learning cache; Indicates tire adhesion margin; This represents the minimum attachment margin that allows learning; This indicates the number of times the safety constraint was violated. This condition ensures that only data with valid stability differences that do not trigger safety violations are used for learning.

[0096] To clarify the source of the calculation of attachment margin and actuator margin, the margin metric used in the learning features is defined.

[0097]

[0098] In the formula, Indicates the first The adhesion utilization rate of each wheel; Indicates longitudinal tire force; Indicates lateral tire force; Indicates vertical load; Indicates the first Estimated adhesion values ​​at each wheel; Indicates tire adhesion margin; Indicates actuator margin; and These represent the current driving torque and braking pressure, respectively. and These represent the upper limit of driving torque and the upper limit of braking pressure, respectively.

[0099] To ensure that the learner has a clear input-output relationship, feature vectors are constructed using state error, attachment margin, and actuator margin, and candidate generalized residual compensations are output.

[0100]

[0101] In the formula, Represents the residual learning feature vector; This represents the characteristic term used to enhance the nonlinear stability error response; Represents the candidate generalized residual learning parameter matrix; The candidate generalized residual compensation includes, in order, the candidate yaw moment residual, the candidate rear wheel steering residual, and the candidate active front wheel steering residual.

[0102] To ensure that the real vehicle has a complete parameter update chain for continuous learning, a recursive least squares update with a forgetting factor is used in the candidate policy replicas, and the covariance matrix is ​​updated simultaneously.

[0103]

[0104] In the formula, This indicates a recursive update of the gain; and These represent the parameter covariance matrices before and after the update, respectively. Indicates the forgetting factor; This represents the prediction error of the generalized control residual; and These represent the candidate parameter matrices before and after the update, respectively. Indicates the safety parameter field The projection operator. This update is only executed in candidate policy replicas and does not directly overwrite the current stable version.

[0105] To prevent the candidate residuals from exceeding the capabilities of the tires and steering actuators, the yaw moment reserve, rear wheel steering residual reserve, and active front wheel steering residual reserve are calculated first.

[0106]

[0107] In the formula, Indicates the first Available longitudinal force margin for each wheel; This indicates the longitudinal force used by the basic controller; This represents the lateral position vector of the four wheels relative to the vehicle's center of mass; This represents the available longitudinal force margin vector for the four wheels; Indicates the residual reserve of yaw moment; Indicates rear wheel steering residual reserve; Indicates active front wheel steering residual reserve; and These represent the upper limit of the rear wheel steering angle rate and the upper limit of the active front wheel steering fine-tuning rate, respectively. Indicates the control cycle.

[0108] To obtain executable safety residual compensation, saturation projections are performed on the candidate yaw moment residual, candidate rear wheel steering residual, and candidate active front wheel steering residual, respectively.

[0109]

[0110] In the formula, , and These represent the candidate yaw moment residual, the candidate rear wheel steering residual, and the candidate active front wheel steering residual, respectively. , and These represent the corresponding residuals after safe projection; This represents the generalized residual compensation after safe projection.

[0111] To ensure that the learning residuals can only participate in control at a limited proportion, the final generalized stability control instruction is written as a combination of the basic instruction and the limited residuals.

[0112]

[0113] In the formula, This represents the residual participation ratio matrix; Indicates the proportion of the yaw moment residual involved; Indicates the proportion of rear wheel steering residuals involved; This indicates the proportion of active front wheel steering residuals involved. This proportion is determined based on stability quality assessment, candidate strategy confidence, and event labeling, decreasing to zero or near zero at high risk or low confidence levels.

[0114] To ensure that the final generalized stability control command is applied to the specific actuator, three control components are first separated from the generalized command: yaw moment, rear wheel steering, and active front wheel steering.

[0115]

[0116] In the formula, Indicates the final target yaw moment; This indicates the final rear wheel steering correction angle command; This indicates the final active front wheel steering adjustment command; , and This represents the row vector used to select the corresponding component from the generalized stability control command.

[0117] To achieve the target yaw moment through differential braking or drive torque, a formula for the longitudinal force distribution of the four wheels is established based on the lateral position of the wheels.

[0118]

[0119] In the formula, The mapping matrix representing the longitudinal force to the yaw moment of the four wheels; This represents the longitudinal force correction vector for the four wheels; This represents the weight matrix for the force distribution at the wheel ends; , , and These represent the lateral positions of the left front, right front, left rear, and right rear wheels relative to the vehicle's center of gravity, respectively. This represents the available longitudinal force margin vector for the four wheels.

[0120] In order to generate commands that the vehicle actuators can receive, the longitudinal force correction of the four wheels is converted into drive torque correction and braking pressure correction, and the steering correction is sent to the corresponding steering actuator.

[0121]

[0122] In the formula, Indicates the first The amount of drive torque correction for each wheel; Indicates the effective rolling radius of the wheel; Indicates the first Longitudinal force correction for each wheel; Indicates the first Braking pressure correction for each wheel. Indicates the brake pressure conversion factor; This indicates the steering angle command sent to the rear wheel steering actuator; This indicates the angle command sent to the active front wheel steering actuator.

[0123] To clarify which parameters the self-evolution of this invention specifically affects, a set of evolvable parameters is defined for each stable version. This set does not directly rewrite the hard limits of the actuator and the basic safety threshold, but rather rewrites the residual learning parameters, error weights, and residual participation ratios.

[0124]

[0125] In the formula, Indicates the first A stable set of evolvable parameters; Indicates the first A stable version of the generalized residual learning parameter matrix; Indicates the first A stable version of the calibrated evolutionary parameter vector; and They represent the first The yaw rate error weight and sideslip error weight in each stable version; , and These represent the participation ratios of the yaw moment residual, the rear wheel steering residual, and the active front wheel steering residual, respectively. This indicates the allowed version parameter security domain.

[0126] To illustrate how the candidate versions are driven by real-vehicle closed-loop evaluation, finite difference evaluation gradient and projection update are used for the calibration evolution parameters.

[0127]

[0128] In the formula, The first term in the calibration evolution parameter vector represents the first term. Evaluation gradient estimation of each parameter; Indicates the closed-loop evaluation index of the shadow model; This represents the generalized residual learning parameter matrix obtained by recursively updating the candidate policy; Represents finite difference perturbation; Indicates the first Unit direction vectors; Indicates by each The gradient estimation vector is formed; Indicates the parameter evolution step size; Indicates the safety domain of calibration parameters The projection operator; This represents the calibrated evolutionary parameter vector of the candidate versions; This represents the set of evolvable parameters for candidate versions.

[0129] To clarify how the evolved parameters actually affect the control output, the error weights of the candidate versions, the residual learning parameter matrix, and the residual participation ratio are substituted into the control law.

[0130]

[0131] In the formula, This represents the composite stability error under the influence of candidate version parameters; and These represent the yaw rate error weight and sideslip angle error weight in the candidate versions, respectively. This represents the residual participation ratio matrix in the candidate versions; This represents the generalized residual compensation of the candidate version output; This represents the feature vector constructed from the composite stability error of the candidate versions; This represents the basic generalized stability control quantity obtained under the influence of candidate version error weights; This represents the generalized residual compensation after the candidate version has been securely projected. This represents the final generalized stability control quantity corresponding to the candidate version. Therefore, parameter evolution influences the final control output by altering the composite error construction, the residual compensation direction, and the residual participation ratio.

[0132] To avoid the candidate version replacing the current stable version all at once, a takeover coefficient is used during the gray-scale takeover phase to merge the output of the current stable version and the output of the candidate version.

[0133]

[0134] In the formula, Indicates the first The takeover coefficient of candidate versions in each gray-scale takeover stage; This indicates the percentage of takeover that can be increased at each grayscale stage; Indicates the maximum allowable takeover ratio during the gray-scale takeover phase; This represents the generalized stability control quantity output by the current stable version; This represents the generalized stability control quantity output by the candidate version; This represents the generalized stability control quantity actually sent to the execution allocation unit during the gray-scale takeover phase.

[0135] To ensure that the advancement of a candidate strategy is determined by its actual closed-loop stability, a closed-loop evaluation index is defined that simultaneously considers yaw error, lateral risk, attachment utilization, and residual smoothness.

[0136]

[0137] In the formula, Represents the set of evolvable parameters Corresponding closed-loop evaluation indicators; Indicates the evaluation time window; This indicates that the average value of the evaluation window is calculated. , , , and These represent the weights of yaw rate error, sideslip angle error, adhesion utilization, residual amplitude, and residual rate of change, respectively. Represents the L2 norm; This represents the generalized residual compensation amount allowed to be incorporated into vehicle control commands after the candidate residual model output has undergone safety constraint processing in the t-th control cycle. This closed-loop evaluation index is used to evaluate the impact of different sets of evolvable parameters on the closed-loop stability of the real vehicle.

[0138] To prevent candidate policies from taking over the vehicle based solely on a decrease in training error, candidate policies are first evaluated in shadow mode based on three categories: effectiveness, safety, and output difference.

[0139]

[0140] In the formula, Indicates the margin for improvement in shadow evaluation; This represents the stability evaluation index in shadow mode; This represents the set of evolvable parameters for the current stable version; This represents the set of evolvable parameters for candidate versions; Indicates the minimum improvement amount; Indicates the difference margin of the candidate policy output; This indicates the degree of difference between the candidate version and the current stable version's output; Indicates the maximum allowed output difference; Indicates the number of times a safety constraint has been violated; This indicates whether the candidate version passed shadow verification.

[0141] To ensure that the strategy version upgrade is determined by the actual vehicle closed-loop control effect, the stable version is updated using the actual vehicle evaluation conditions after the gray-scale takeover; if the conditions are not met, it is rolled back to the current stable version.

[0142]

[0143] In the formula, This represents the set of evolutionary parameters for the new stable version after the upgrade; The closed-loop evaluation index of the actual vehicle during the gray-scale takeover period; Indicates the safety margin required for version upgrade; This indicates the number of security constraint violations during the gray-scale takeover period; Indicates the number of valid events that have been covered; This represents the minimum event coverage threshold. When the above conditions are met, the candidate version... , , and Together they become parameters for the next stable version; if not satisfied, they will be maintained. constant.

[0144] In one complete embodiment, the method of the present invention includes the following steps: 1. Construct a training scenario set in the simulation environment that includes different vehicle speeds, adhesion, loads, equivalent tire parameters, actuator hysteresis, and sensor noise.

[0145] 2. Run the basic generalized stable controller in each simulation scenario to obtain the basic target yaw moment, the basic rear wheel steering correction angle, and the basic active front wheel steering fine-tuning.

[0146] 3. The teacher / reference optimizer generates a restorative reference control quantity within the safe and feasible control set, and uses the difference between the reference control quantity and the basic control quantity as the simulation residual label.

[0147] 4. The initial residual learning parameter matrix is ​​trained using a feature-parameterized residual model, and the resulting initial residual model is used as the initial candidate model for the actual vehicle stage; no teacher / reference optimizer is deployed on the actual vehicle stage.

[0148] 5. During actual vehicle operation, the vehicle status, front and rear wheel steering angles, four-wheel wheel speeds, braking pressure, driving torque, and road adhesion estimates are collected to form a vehicle state vector.

[0149] 6. Calculate the desired yaw rate based on the steady-state steering response and adhesion limit, and calculate the yaw error, sideslip error, composite stability error and boundary layer saturation output based on the desired yaw rate and vehicle state.

[0150] 7. The basic generalized stability controller outputs the minimum generalized stability control quantity, and the initial residual model or candidate residual model outputs the generalized residual compensation quantity.

[0151] 8. Construct difference samples based on the closed-loop response of the real vehicle, and add a learning cache when the stability error, adhesion margin and safety constraints are met.

[0152] 9. In the candidate strategy replica, update the generalized residual learning parameter matrix based on the difference samples, and correct the yaw rate error weight, side slip error weight and residual participation ratio based on the shadow evaluation, while keeping the current stable version from being overwritten.

[0153] 10. Perform a safety projection on the candidate residuals under constraints of tire adhesion margin, steering actuator margin, output amplitude, and output rate.

[0154] 11. The basic generalized stability control command and the constrained residual are superimposed on each other in a finite proportion and distributed to the rear wheel steering, active front wheel steering, differential braking and drive torque channels.

[0155] 12. Perform shadow verification, parameter grayscale takeover, real vehicle closed-loop evaluation, version upgrade or rollback on candidate strategies to make the new residual parameters, error weights and residual participation ratios become stable versions after meeting safety conditions.

[0156] Through the aforementioned process, this invention forms a complete closed loop, encompassing simulation scenario construction, teacher reference optimization, initial residual model training, real vehicle state acquisition, basic generalized control, real vehicle difference samples, candidate residual learning, safety projection, multi-actuator allocation, and closed-loop version updates. Specifically, self-evolution does not involve abstractly updating the controller name, nor does it involve running the teacher optimizer online on the real vehicle. Instead, it first learns the initial residual model parameter matrix during the simulation phase, and then updates the generalized residual learning parameter matrix, yaw rate error weight, sideslip error weight, and residual participation ratio under constraints during the real vehicle phase. These parameters are then incorporated into the final control input through composite stability error, residual compensation, and residual superposition ratio. The learning strategy does not directly replace the basic stability controller but gradually participates in control within an interpretable, projectable, and rollback-capable residual channel, thus balancing the continuous improvement capability of the real vehicle with the chassis stability safety boundary.

[0157] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A self-evolutionary method for vehicle stability control based on self-closed-loop continuous learning, characterized in that, Includes the following steps: S1: Construct a set of simulation training scenarios, run the basic generalized stable controller in the set of simulation training scenarios to obtain the basic generalized stable control quantity; use the teacher / reference optimizer to generate a restorative reference control quantity, and use the difference between the restorative reference control quantity and the basic generalized stable control quantity as the simulation residual label; Train an initial residual model based on the simulation residual labels; S2: Collect vehicle state data to form a vehicle state vector, calculate yaw error and sideslip error, and calculate composite stability error based on the yaw error and sideslip error; S3: The basic generalized stability controller outputs a baseline generalized stability control quantity based on the composite stability error; the initial residual model outputs a generalized residual compensation quantity based on the vehicle state vector and the composite stability error; S4: Construct difference samples based on the actual vehicle closed-loop response, update the generalized residual learning parameter matrix in the candidate policy replica based on the difference samples, and correct the error weights and residual participation ratios; S5: Perform a safe projection on the generalized residual compensation to obtain restricted residual compensation; superimpose the basic generalized stability control quantity and the restricted residual compensation according to the residual participation ratio to obtain the final generalized stability control instruction; allocate multiple actuators according to the final generalized stability control instruction.

2. The self-evolving method for vehicle stability control based on self-closed-loop continuous learning according to claim 1, characterized in that, In step S1, the basic generalized stability control quantity and the restorative reference control quantity both include yaw moment, rear wheel steering correction angle and active front wheel steering fine-tuning quantity.

3. The self-evolutionary method for vehicle stability control based on self-closed-loop continuous learning according to claim 1, characterized in that, The generation of restorative reference control quantities using a teacher / reference optimizer specifically includes: A finite prediction window is established, and the impact of candidate control variables on the stability state is predicted using a simulated vehicle dynamics model. The teacher / reference optimizer constructs a recovery reference optimization objective function based on the recovery stability error weight matrix, the control deviation weight matrix, and the control rate of change weight matrix, obtains the control sequence within the safe and feasible control set, and takes the first step of the control sequence as the recovery reference control quantity.

4. The self-evolving method for vehicle stability control based on self-closed-loop continuous learning according to claim 1, characterized in that, In step S2, the yaw error Lateral deviation error and composite stability error The calculation formula is: In the formula, r represents the yaw rate; This represents the desired yaw rate; Indicates the centroid sideslip angle; Indicates the expected centroid sideslip angle; Indicates the weight of the yaw rate error; This represents the weight of the sideslip angle error.

5. The self-evolutionary method for vehicle stability control based on self-closed-loop continuous learning according to claim 4, characterized in that, In step S3, the basic generalized stability controller calculates the basic generalized stability control quantity based on the composite stability error s. The basic generalized stability control quantity Including the basic target yaw moment Basic rear wheel steering correction angle and basic active front wheel steering fine-tuning The calculation formula is: In the formula, This indicates the yaw rate error feedback gain; This indicates the centroid sideslip angle error feedback gain; Indicates the composite stability error feedback gain; and These represent the yaw rate error gain and sideslip angle error gain of the rear wheel steering channel, respectively. and These represent the yaw rate error gain and sideslip angle error gain of the active front wheel steering channel, respectively. Indicates the upper limit of the rear wheel steering correction angle; This indicates the upper limit of the active front wheel steering fine-tuning amount; This indicates that a boundary layer saturation function is applied to the composite stability error s. , This represents the boundary layer width, and -1 and 1 represent the lower and upper bounds of the normalized output.

6. The self-evolutionary method for vehicle stability control based on self-closed-loop continuous learning according to claim 1, characterized in that, In step S4, the product of the feature vector and the generalized residual learning parameter matrix is ​​used as the generalized residual compensation amount; In the candidate policy replicas, the generalized residual learning parameter matrix is ​​updated using recursive least squares with a forgetting factor, specifically including: In the formula, This indicates a recursive update of the gain; and These represent the parameter covariance matrices before and after the update, respectively. Indicates the forgetting factor; This represents the prediction error of the generalized control residual; and These represent the candidate parameter matrices before and after the update, respectively. Indicates the safety parameter field The projection operator; The feature vector is the feature vector of the t-th difference sample, and the generalized residual learning objective is the difference between the recovered generalized control quantity and the basic generalized stability control quantity corresponding to the t-th difference sample.

7. The self-evolutionary method for vehicle stability control based on self-closed-loop continuous learning according to claim 1, characterized in that, In step S5, a safe projection is performed on the generalized residual compensation amount, specifically including: Calculate the residual reserve of yaw moment Rear wheel steering residual reserve and active front steering residual reserve : In the formula, y represents the lateral position vector of the four wheels relative to the vehicle's center of gravity; R represents the available longitudinal force margin vector of the four wheels. and These represent the upper limit of the rear wheel steering angle rate and the upper limit of the active front wheel steering fine-tuning rate, respectively. Indicates the control cycle; Indicates the base rear wheel steering correction angle; This indicates the basic active front wheel steering fine-tuning amount; For the candidate yaw moment residuals provided by the generalized residual compensation amount Candidate rear wheel steering residual and candidate active front wheel steering residual Saturated projections were performed separately to obtain the safety residual compensation: In the formula, , and These represent the corresponding residuals after safe projection; This represents the generalized residual compensation after safe projection.

8. The self-evolutionary method for vehicle stability control based on self-closed-loop continuous learning according to claim 1, characterized in that, In step S5, the basic generalized stability control quantity is... With the aforementioned restricted residual compensation According to the aforementioned residual participation ratio Superposition yields the final generalized stability control command. : In the formula, , , and These represent the residual participation ratios of yaw moment, rear-wheel steering, and active front-wheel steering, respectively.

9. A self-evolving method for vehicle stability control based on self-closed-loop continuous learning according to claim 1 or 8, characterized in that, In step S5, multiple actuators are allocated according to the final generalized stability control instruction, specifically including: The yaw moment component is extracted from the final generalized stability control command. A four-wheel longitudinal force distribution formula is established based on the lateral position of the wheels. The four-wheel longitudinal force correction vector is calculated and converted into the drive torque correction amount and brake pressure correction amount sent to the vehicle actuator. The extracted steering correction component is sent to the corresponding steering actuator. After allocating multiple executors according to the final generalized stability control instruction, the method further includes: performing closed-loop version evaluation on candidate strategies that include the updated generalized residual learning parameter matrix, error weights, and residual participation ratios; and making the candidate strategy a new stable version when the preset safety and performance upgrade conditions are met, otherwise rolling back to the current stable version.

10. A self-evolving vehicle stability control system based on self-closed-loop continuous learning, used to execute the self-evolving vehicle stability control method based on self-closed-loop continuous learning as described in any one of claims 1 to 9, characterized in that, The system includes: The simulation pre-training unit is used to construct a set of simulation training scenarios, run a basic generalized stable controller in the set of simulation training scenarios to obtain basic generalized stable control quantities; generate restorative reference control quantities using a teacher / reference optimizer, and use the difference between the restorative reference control quantities and the basic generalized stable control quantities as simulation residual labels; train an initial residual model based on the simulation residual labels. The basic generalized stability control unit is used to collect vehicle state data to form a vehicle state vector, calculate yaw error and sideslip error, calculate composite stability error based on the yaw error and sideslip error, and output a basic generalized stability control quantity with a minimum guarantee based on the composite stability error. The candidate residual learning unit is used to output a generalized residual compensation amount from the initial residual model based on the vehicle state vector and the composite stability error. The real vehicle difference sample construction unit is used to construct difference samples based on the real vehicle closed-loop response, update the generalized residual learning parameter matrix in the candidate policy replica based on the difference samples, and correct the error weights and residual participation ratios. A safety projection unit is used to perform a safety projection on the generalized residual compensation amount to obtain a restricted residual compensation; the basic generalized stability control amount and the restricted residual compensation are superimposed according to the residual participation ratio to obtain the final generalized stability control command; A multi-actuator allocation unit is used to allocate multiple actuators according to the final generalized stability control instruction; The closed-loop version management unit is used to perform closed-loop version evaluation on candidate strategies that include the updated generalized residual learning parameter matrix, error weights, and residual participation ratios; when the preset security and performance upgrade conditions are met, the candidate strategy becomes a new stable version; otherwise, it is rolled back to the current stable version.

Images