A vehicle limit stability control method, system, and medium based on operating condition perception

Abstract

The application discloses a vehicle limit stability control method and system based on working condition perception and a medium, belongs to the technical field of vehicle stability control, and is based on a DYC subsystem, an ARS subsystem and an ASS subsystem. The method comprises the following steps: acquiring a vehicle motion state, a road adhesion estimation value and control inputs of each subsystem, and constructing working condition context features; utilizing an Actor network to acquire cooperative parameters of each subsystem, constructing a cooperative game optimization problem, and solving the problem through a differentiable ADMM layer cascaded at the back end of the Actor network to obtain output control quantities of each subsystem and a vehicle predicted state trajectory; and the differentiable ADMM layer is used as a differentiable gradient propagation path in offline training, and gradient information of the output control quantities and the predicted state trajectory is reversely propagated to the Actor network. The application can dynamically adjust the control sharing relationship among the subsystems under the limit working condition, and improve the vehicle yaw stability, side deflection convergence and anti-rolling capability.

Description

Technical Field

[0001] This invention belongs to the field of vehicle stability control technology, specifically relating to a vehicle limit stability control method, system, and medium based on operating condition perception. Background Technology

[0002] Angular modular vehicles integrate drive, braking, steering, and suspension actuators at the wheel ends, allowing each wheel to independently adjust longitudinal force, steering angle, and vertical force. This structure expands the controllable boundaries of the vehicle under extreme conditions such as low adhesion, high-speed obstacle avoidance, continuous slalom, and double lane change, but also significantly enhances the coupling between yaw, sideslip, roll, and tire adhesion utilization. Once the vehicle enters the tire nonlinear region, a single actuator or fixed-parameter control strategy struggles to simultaneously achieve yaw stability, sideslip suppression, and roll suppression.

[0003] Existing vehicle stability control methods typically employ a direct yaw moment control subsystem (DYC subsystem), an active rear-wheel steering subsystem (ARS subsystem), an active suspension control subsystem (ASS subsystem), or a combination thereof, and describe the collaborative relationships of multiple actuators through cooperative game theory and ADMM consistency coordination. However, the task weights, actuator priorities, and control allocation relationships in existing methods often rely on fixed calibrations or segmented rules, making it difficult to continuously adjust according to road adhesion, vehicle state evolution, and actuator intervention. Furthermore, when a certain operating condition simultaneously presents insufficient yaw, increased lateral slip, and roll risks, fixed weights may lead to mutual cancellation or excessive intervention among the actuators.

[0004] In addition, there are also methods that extract working condition features from historical states through reinforcement learning and attention mechanisms and provide a basis for adaptive control parameters. However, the high-level policy only outputs control variables or coarse weights and does not form an interpretable, constrainable and differentiable feedback loop with the underlying physical constraint optimization. This may lead to problems such as constraint violation, unstable convergence or insufficient real-time performance of the trained policy under extreme working conditions. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a vehicle extreme stability control method, system, and medium based on operating condition perception. It can dynamically adjust the control sharing relationship between the DYC, ARS, and ASS subsystems under extreme operating conditions such as low-adhesion continuous steering, high-speed double lane change, and emergency obstacle avoidance, thereby improving the vehicle's yaw stability, lateral convergence, and anti-roll capability.

[0006] This invention provides the following technical solution:

[0007] Firstly, a vehicle limit stability control method based on operating condition perception is provided, employing a collaborative control system integrating a DYC subsystem, an ARS subsystem, and an ASS subsystem. The method includes: The vehicle motion state, road surface adhesion estimate, and control input of each subsystem are obtained. The vehicle state vector and single-step working condition observation vector are constructed, and the working condition features and control channel features of each subsystem are constructed and then concatenated to obtain the combined features. Based on the single-step working condition observation vectors of the current time and its preceding time, extract the time context vectors, and combine them with the combined features to perform cross-attention fusion to obtain the working condition context features; Based on the characteristics of the working condition context, the Actor network using the Actor-Critic algorithm outputs continuous actions and maps them to the cooperative parameters of each subsystem. Given the cooperative parameters of each subsystem and the vehicle state prediction model, a cooperative game optimization problem is constructed with each subsystem as a participant. The cooperative game optimization problem is solved by cascading a differentiable ADMM layer at the back end of the Actor network to obtain the output control variables of each subsystem and the predicted state trajectory of the vehicle. The differentiable ADMM layer is composed of an ADMM expansion with a fixed number of iterations, and provides a gradient backpropagation path for Actor network updates through the auxiliary loss of the predicted state trajectory during offline training.

[0008] Optionally, the vehicle state vector is: ; In the formula, Indicates the current sampling period The vehicle state vector; Indicates the longitudinal speed of the vehicle; Indicates the lateral speed of the vehicle; Indicates yaw rate; Indicates the centroid sideslip angle; Indicates the vehicle body roll angle; Indicates the roll rate; This represents the estimated value of road surface adhesion. This represents the control inputs of each subsystem in the previous sampling period, including the DYC additional yaw moment, the ARS rear wheel steering angle, and the ASS active suspension anti-roll force. This indicates the concatenation of vectors or matrices; Indicates transpose; The single-step operating condition observation vector is: ; In the formula, Indicates the current sampling period The single-step working condition observation vector; This indicates the error of the centroid sideslip angle relative to the expected value; This indicates the error of the yaw rate relative to the expected value; This indicates the error of the vehicle body roll angle relative to the expected value; Indicates the equivalent steering angle or steering input of the front wheels; This indicates the lateral acceleration of the vehicle.

[0009] Optionally, the step of extracting the time context vector and combining it with the combined features to perform cross-attention fusion to obtain the working condition context features is as follows: extracting the time context vector using a time feature extraction network, projecting the time context vector to obtain the query vector, projecting the combined features to obtain the key feature matrix and the value feature matrix, and fusing them through a multi-head attention mechanism to obtain the working condition context features.

[0010] Optionally, the Actor network based on the Actor-Critic algorithm, which uses the working condition context features to output continuous actions and maps them to the cooperative parameters of each subsystem, specifically: Input the working condition context features into the Actor network, and output continuous actions; ; In the formula, This indicates that the Actor network is in the current sampling period. The continuous output action; The parameter is Actor network; Indicates the contextual features of the operating condition; This indicates the distribution trend of lateral stabilization tasks between the DYC subsystem and the ARS subsystem; This indicates the importance of lateral stability objectives; This underscores the importance of a tilt-stabilized target; The continuous actions output by the Actor network are used to construct the cooperative parameters of each subsystem: ; ; ; ; In the formula, Represents the set of collaborative parameters; , , These represent the DYC subsystem, ARS subsystem, and ASS subsystem respectively during the current sampling period. The following are the collaborative parameters.

[0011] Optionally, given the cooperative parameters of each subsystem and the vehicle state prediction model, constructing a cooperative game optimization problem by treating each subsystem as a participant specifically involves: Establish a vehicle state prediction model that includes actuator delay; Transform the cooperative parameters of the subsystem into state tracking weights; ; in, For the first Each subsystem in the current sampling period State tracking weight matrix; For the first Each subsystem in the current sampling period The collaborative parameters; Let be the semi-positive definite nominal state tracking weight matrix. It is a collection of subsystems, including the DYC subsystem, the ARS subsystem, and the ASS subsystem; Constructing a cooperative game optimization problem : , ; ; in, Indicates the first Subsystems in local decision variables The cost of local decision variables Including the The vehicle state sequence, its own control sequence, and local estimates of the control sequences of other subsystems are obtained from the prediction of each subsystem. Representing local decision variables Belongs to the Local feasible region of each subsystem ; Representing local decision variables With shared variables Maintain consistency and share variables This includes globally consistent predicted state trajectories and globally consistent control sequences for each subsystem. To predict the step size, For the first The subsystem in the first The stable target tracking error for each prediction step is obtained based on a pre-calculated stable reference target; Indicates the first The subsystem in the first Control input for each prediction step; Indicates the first The subsystem in the first The control input increment for each prediction step; and These represent the input consumption weight and the input increment weight, respectively. This indicates transpose.

[0012] Optionally, the first The subsystem in the first Stable target tracking error in each prediction step Specifically: ; ; ; ; In the formula, Indicates the first Each prediction step stability error vector This represents the error selection matrix, with values... , or , For yaw and lateral deviation errors, For lateral and yaw errors, This refers to the error between the roll angle and the roll rate. This indicates that a diagonal matrix is ​​constructed using the elements within the parentheses as diagonal elements. , , and They represent the first The actual centroid sideslip angle, actual yaw rate, actual body roll angle, and actual body roll rate are predicted in each prediction step; , , and These represent the stability reference center of gravity sideslip angle, stability reference yaw rate, stability reference body roll angle, and stability reference body roll rate, respectively.

[0013] Optionally, the solution of the cooperative game optimization problem by cascading a differentiable ADMM layer at the back end of the Actor network specifically involves: merging the local decision variables, dual variables, and shared variables of each iteration step of the differentiable ADMM layer into a joint internal state; performing a preset fixed number of iterations for iterative updates; and extracting the output control variables of each subsystem and the predicted vehicle state trajectory from the joint internal state of the last iteration step. The local decision variables include the vehicle state sequence predicted by the current subsystem, its own control sequence, and local estimates of the control sequences of other subsystems; the shared variables include the globally consistent predicted state trajectory and the globally consistent control sequence of each subsystem; and the dual variables are used to accumulate and correct the consistency deviation between the local decision variables and the shared variables.

[0014] Alternatively, during offline training, the objective function can be used as follows: Update the parameters of the Actor network: ; ; in, For strategy benefit items, For the parameters of the Actor network, To assist in loss weighting, To mitigate losses, Indicates the prediction step size; , , and These represent the auxiliary loss weights for the center of gravity sideslip angle deviation, yaw rate deviation, body roll angle deviation, and control increment, respectively. , and These represent the deviations between the predicted state and the corresponding stability reference centroid sideslip angle, stability reference yaw rate, and stability reference vehicle roll angle, respectively. This indicates the change in control quantity between adjacent prediction steps; In offline training, the differentiable ADMM layer serves as a differentiable gradient propagation path, backpropagating the gradient information of the output control quantity and the predicted state trajectory to the Actor network. Specifically, when updating the Actor network, the gradients of the output control quantity and the predicted state trajectory with respect to the cooperative parameter set and the continuous actions of the Actor are first calculated based on the Jacobian recursion relationship of each iteration mapping in the differentiable ADMM layer. Then, the objective function is obtained. Regarding Actor network parameters The gradient is then calculated, and the optimizer performs gradient descent to update the Actor network.

[0015] Secondly, a vehicle limit stability control system based on operating condition perception is provided, comprising: The data acquisition module is used to acquire vehicle motion state, road surface adhesion estimation value, and control input of DYC subsystem, ARS subsystem, and ASS subsystem. It constructs vehicle state vector and single-step working condition observation vector, and then constructs working condition features and control channel features of each subsystem and splices them together to obtain combined features. The spatiotemporal attention working condition perception module is used to extract the temporal context vector based on the single-step working condition observation vector of the current time and its previous time, and combine it with the combined features to perform cross-attention fusion to obtain the working condition context features. The collaborative parameter generation module is used to continuously output actions based on the working condition context features using the Actor-Critic algorithm's Actor network, and map these actions to the collaborative parameters of each subsystem. The cooperative game modeling module is used to construct a cooperative game optimization problem with each subsystem as a participant, given the cooperative parameters of each subsystem and the vehicle state prediction model. The control output module is used to solve the cooperative game optimization problem by cascading a differentiable ADMM layer at the back end of the Actor network, and obtain the output control quantity and vehicle predicted state trajectory of each subsystem. The differentiable ADMM layer is composed of ADMM expansion with a fixed number of iterations, and serves as a differentiable gradient propagation path in offline training to backpropagate the gradient information of the output control quantity and the predicted state trajectory to the Actor network. The DYC subsystem, ARS subsystem, and ASS subsystem are used to output the output control quantities of each subsystem obtained by the control output module to the corresponding actuators.

[0016] Thirdly, a computer-readable storage medium is provided for storing a computer program; when the computer program is executed by a processor, it implements the steps of the vehicle limit stability control method based on operating condition perception as described in any one of the first aspects.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention identifies stability risks such as road adhesion, vehicle speed, yaw, roll, and body roll through condition perception. It unifies and coordinates direct yaw moment, active rear-wheel steering, and active suspension anti-roll control, enabling different actuators to dynamically share control tasks according to the current risk, reducing control conflicts caused by fixed weights or fixed channel priorities. This invention generates cooperative parameters through the Actor-Critic algorithm's Actor network, and the differentiable ADMM layer cascaded at the back end of the Actor network solves for the actual control quantities within vehicle dynamics, adhesion, and actuator constraints. This improves yaw stability, yaw convergence, and anti-roll capability while ensuring control feasibility. The differentiable ADMM layer of this invention can utilize bottom-level optimization feedback during the training phase and maintain stable output during the online control phase under sensor noise, low adhesion, or actuator limitations. This solves the problems of constraint violation, unstable convergence, or insufficient real-time performance that may occur in existing technologies under extreme conditions, thus improving vehicle safety, ride comfort, real-time performance, and engineering robustness under extreme conditions. Attached Figure Description

[0018] Figure 1 This is a flowchart of the vehicle limit stability control method based on operating condition perception according to the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be used to limit the scope of protection of the present invention. It should be noted that the term "comprising" and any variations thereof in the specification, claims and the above-mentioned drawings of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or devices.

[0020] Example 1: like Figure 1 As shown, a vehicle limit stability control method based on operating condition perception employs a collaborative control system integrating a DYC (Direct Yaw Moment Control) subsystem, an ARS (Active Rear Steering) subsystem, and an ASS (Active Suspension System) subsystem. The stability control method includes the following steps: Step S1: Obtain the vehicle motion state, road surface adhesion estimate, and control input of each subsystem. Construct the vehicle state vector and single-step working condition observation vector, and then construct the working condition features and control channel features of each subsystem and concatenate them to obtain the combined features.

[0021] Step S1 specifically includes: Step S11: Obtain the vehicle motion state, road surface adhesion estimate, and control input of each subsystem, and construct the vehicle state vector and single-step working condition observation vector.

[0022] The control inputs for each subsystem are the control inputs from the previous cycle of the DYC, ARS, and ASS subsystems. Vehicle motion states include: longitudinal velocity, lateral velocity, yaw rate, sideslip angle, roll angle, and roll rate. Road surface adhesion estimates can be obtained based on wheel speed, longitudinal / lateral acceleration, yaw rate, and tire force response, or through a fusion estimation method based on vehicle dynamics residuals and tire adhesion utilization.

[0023] The constructed vehicle state vector is: ; In the formula, Indicates the current sampling period The vehicle state vector; Indicates the longitudinal speed of the vehicle; Indicates the lateral speed of the vehicle; Indicates yaw rate; Indicates the centroid sideslip angle; Indicates the vehicle body roll angle; Indicates the roll rate; This represents the estimated value of road surface adhesion. Indicates the previous sampling period The control inputs for each subsystem include DYC additional yaw moment, ARS rear wheel steering angle, and ASS active suspension anti-roll force; This indicates the concatenation of vectors or matrices; This indicates transpose.

[0024] The constructed single-step working condition observation vector is as follows: ; In the formula, Indicates the current sampling period The single-step working condition observation vector; This indicates the error of the centroid sideslip angle relative to the expected value; This indicates the error of the yaw rate relative to the expected value; This indicates the error of the vehicle body roll angle relative to the expected value; Indicates the equivalent steering angle or steering input of the front wheels; This represents the vehicle's lateral acceleration. The constructed single-step observation vector. Used for operating condition identification, not directly as actuator control commands.

[0025] Step S12: Construct the operating condition characteristics and the control channel characteristics of each subsystem, and then splice them together to obtain the combined characteristics. .

[0026] ; ;

[0027]

[0028] ; in, , and The DYC subsystem, ARS subsystem, and ASS subsystem are respectively in the current sampling period. The control channel characteristics, In the current sampling period The operating conditions and characteristics , and These represent the additional yaw moment of DYC, the rear wheel steering angle of ARS, and the anti-roll force of ASS active suspension in the previous sampling period, respectively. The ARS rear wheel steering angle change rate in the previous sampling period; , and These represent the saturation levels of the DYC subsystem, ARS subsystem, and ASS subsystem, respectively. , and These represent the remaining control margins of the DYC subsystem, ARS subsystem, and ASS subsystem, respectively. For the lateral acceleration of the vehicle, This indicates the confidence level of the attachment estimate.

[0029] Step S2: Based on the single-step working condition observation vectors of the current time and its preceding time, extract the time context vectors, and combine them with the combined features to perform cross-attention fusion to obtain the working condition context features.

[0030] Step S2 specifically includes: Step S21: Based on the single-step operating condition observation vectors of the current time and its preceding time, construct the historical operating condition observation sequence and extract the time context vector.

[0031] To describe the temporal evolution of the extreme conditions, the current time is taken as the end point and the structural length is... The historical observation window covers vehicle state deviations, changes in road surface adhesion, and the history of execution system interventions, and outputs a time context vector.

[0032] ; In the formula, Indicates the current sampling period Historical operating condition observation sequence; Indicates the current sampling period The single-step working condition observation vector; This represents the single-step operating condition observation vector of the previous sampling period; Indicates the length of the history window; This represents a sequence assembled chronologically. The length of the historical observation window. It is determined by the vehicle control cycle, the response delay of each actuator, and the time constant of the extreme operating condition state change.

[0033] ; In the formula, This represents a time context vector extracted from historical operating condition observation sequences; The temporal feature extraction network can be represented by one-dimensional convolution, gated recurrent network, temporal attention network or a combination thereof; This represents the network parameters for extracting time features.

[0034] Step S22: Perform cross-attention fusion on the time context vector and combined features to obtain the working condition context features.

[0035] ; ; ; ; ; In the formula, Indicates the first One point of attention; Represented by the time context vector The query vector obtained by projection; and They represent the combined features respectively The key feature matrix and value feature matrix obtained by projection; , and They represent the first The query, key, and value projection matrix of each attention head; Indicates the first The output of each attention head; This represents the normalized exponential function, used to convert relevance scores into attention weights; This indicates that scaling is performed using the square root of the feature dimension; Indicates the first The feature dimension of each attention head; Indicates transpose; Indicates the contextual features of the operating condition; Represents a multilayer perceptron mapping; Indicates the first The output of each attention head; Indicates the total number of attention heads; This means concatenating all attention head outputs according to the feature dimension.

[0036] Step S3: Based on the working condition context features, the Actor network of the Actor-Critic algorithm is used to output continuous actions and map them to the cooperative parameters of each subsystem.

[0037] Step S3 specifically includes: Step S31: The Actor network outputs continuous actions using the Actor-Critic algorithm (a reinforcement learning algorithm).

[0038] The Actor network outputs continuous actions based on the characteristics of the operating context. These continuous actions are then mapped to form the collaborative parameters of the DYC subsystem, ARS subsystem, and ASS subsystem. These collaborative parameters are used to adjust the state tracking weights in the underlying cooperative game optimization problem, rather than being directly used as vehicle actuator commands.

[0039] ; In the formula, This indicates that the Actor network is in the current sampling period. The continuous output action; The parameter is Actor policy network; Indicates the contextual features of the operating condition; This indicates the distribution trend of lateral stabilization tasks between the DYC subsystem and the ARS subsystem; This indicates the importance of lateral stability objectives; This indicates the importance of the tilt-stabilized target.

[0040] Step S32: Construct the continuous actions output by the Actor network into the cooperative parameters of each subsystem: ; ; ; ; In the formula, Represents the set of collaborative parameters; , , These represent the DYC subsystem, ARS subsystem, and ASS subsystem respectively during the current sampling period. The following are the collaborative parameters.

[0041] Step S4: Given the cooperative parameters of each subsystem and the vehicle state prediction model, construct a cooperative game optimization problem with each subsystem as a participant.

[0042] Step S4 specifically includes: Step S41: Establish a vehicle state prediction model including actuator delay.

[0043] A vehicle lateral, yaw, and roll coupling prediction model is established, and actuator delay, input amplitude, and input increment constraints are incorporated into the prediction model and feasible region.

[0044] ; In the formula, This represents the predicted vehicle state vector for the next sampling period; Indicates the current sampling period The vehicle state vector; Represents the discrete state matrix, with subscripts Indicates discretization; , and These represent the control input matrices of the DYC subsystem, ARS subsystem, and ASS subsystem, respectively. , and These represent the additional yaw moment input of the DYC subsystem, the rear wheel steering angle input of the ARS subsystem, and the active suspension anti-roll force input of the ASS subsystem, respectively. This represents unmodeled disturbances or parameter uncertainties.

[0045] The input constraints for each subsystem are: ; ; In the formula, The index represents the subsystem, corresponding to the DYC subsystem, ARS subsystem, or ASS subsystem; Indicates the first Control inputs for each subsystem; and They represent the first Control input of each subsystem The lower and upper limits allowed under the current operating conditions; Indicates the first The increment of the control input in adjacent sampling periods of each subsystem; and These represent the lower and upper bounds of the input increment under the current operating conditions, respectively.

[0046] Step S42: Convert the cooperative parameters of the subsystem into state tracking weights.

[0047] ; in, For the first Each subsystem in the current sampling period State tracking weight matrix; For the first Each subsystem in the current sampling period The collaborative parameters; Let be the semi-positive definite nominal state tracking weight matrix. It is a collection of subsystems.

[0048] Step S43: Calculate the stability reference target and calculate the stability target tracking error of each subsystem at each prediction step.

[0049] Stability reference targets include the stability reference centroid sideslip angle. Stability reference yaw rate Stability reference vehicle body roll angle Stability reference vehicle body roll rate The upper and lower bounds of the stability reference target are determined by the vehicle wheelbase, mass range, axle load range, tire lateral stiffness range, adhesion envelope, and roll stability margin, and are not arbitrary preset constants. The stability reference center of gravity sideslip angle, stability reference body roll angle, and stability reference body roll angular velocity can be taken as 0 under the linear stability target, or given according to the roll stability margin and vehicle attitude constraints.

[0050] ; In the formula, The stability reference is the yaw rate; Represents a saturation function, when Less than the lower bound Time to take ,when Greater than the upper bound Time to take Otherwise take ; Indicates the equivalent steering angle or steering input of the front wheels; Indicates the vehicle's wheelbase; The vehicle stability factor is calculated from the vehicle mass, the distance from the center of gravity to the front axle, the distance from the center of gravity to the rear axle, and the lateral stiffness of the front and rear axle tires. Represents gravitational acceleration; Indicates the longitudinal speed of the vehicle The square of; and The directions are opposite. The upper and lower limits of saturation are determined by both adhesion and vehicle speed, and are used to prevent the reference yaw response from exceeding the road surface's achievable capabilities.

[0051] Calculate the stable target tracking error of each subsystem at each prediction step: ; ; ; ; In the formula, For the first The subsystem in the first The stable target tracking error in each prediction step, Indicates the first Each prediction step stability error vector This represents the error selection matrix, with values... , or , For yaw and lateral deviation errors, For lateral and yaw errors, This refers to the error between the roll angle and the roll rate. This indicates that a diagonal matrix is ​​constructed using the elements within the parentheses as diagonal elements. , , and They represent the first The actual centroid sideslip angle, actual yaw rate, actual body roll angle, and actual body roll rate are predicted in each prediction step; , , and These represent the stability reference center of gravity sideslip angle, stability reference yaw rate, stability reference body roll angle, and stability reference body roll rate, respectively.

[0052] Step S44: Construct a cooperative game optimization problem by including the DYC subsystem, ARS subsystem, and ASS subsystem as participants. .

[0053] Given the cooperative parameters and prediction model, DYC, ARS, and ASS are considered as cooperative game participants. Each participant minimizes its local prediction cost under its own input constraints, while maintaining consistency with other participants through shared state and shared control estimation. In this embodiment, the local decision variables, shared variables, and feasible region are written in the following specific form:

[0054]

[0055]

[0056] in, Indicates the first Local decision variables of each subsystem; Indicates stacking by column; Indicates the first Each subsystem in the current sampling period The predicted vehicle state sequence; and Indicates the first The subsystem for the first Local estimation of the control sequence of each subsystem and Indicates the first Each subsystem's own control sequence For the first The subsystem in the first Predicting vehicle state in each prediction step. Indicates the prediction step size. For the first The subsystem in the first The control input for each control step This represents the control step size. Therefore, the three participants, DYC, ARS, and ASS, all hold state and control sequence estimates of the same dimension, enabling them to impose consensus constraints.

[0057] The cooperative game optimization problem is as follows: , ; ; in, Indicates the first Subsystems in local decision variables The cost of local decision variables Including the The vehicle state sequence, its own control sequence, and local estimates of the control sequences of other subsystems are obtained from the prediction of each subsystem. This indicates that the constraints are met; Representing local decision variables Belongs to the Local feasible region of each subsystem The determination is based on the vehicle state prediction model in step S41 and the input constraints of each subsystem; Representing local decision variables With shared variables Maintain consistency and share variables This includes globally consistent predicted state trajectories and globally consistent control sequences for each subsystem. To predict the step size, For the first The subsystem in the first The stable target tracking error for each prediction step is obtained based on a pre-calculated stable reference target; Indicates the first The subsystem in the first Control input for each prediction step; Indicates the first The subsystem in the first The control input increment for each prediction step; For the first Each subsystem in the current sampling period State tracking weight matrix, and These represent the input consumption weight and the input increment weight, respectively, and are calibrated and kept positive definite or semi-positive definite according to the requirements of actuator energy consumption, response speed and ride comfort. This indicates transpose.

[0058] Step S5: Solve the cooperative game optimization problem by cascading a differentiable ADMM layer at the back end of the Actor network to obtain the output control quantities of each subsystem and the predicted state trajectory of the vehicle.

[0059] The differentiable ADMM layer is constructed by expanding ADMM (Alternating Direction Method of Multipliers) with a fixed number of iterations. During offline training, it serves as the differentiable gradient propagation path, backpropagating the gradient information of the output control variable and the predicted state trajectory to the Actor network. In this embodiment, when solving the cooperative game optimization problem using the differentiable ADMM layer, the local decision variables, dual variables, and shared variables of each iteration step are merged into a joint internal state, and a preset fixed number of iterations is performed. The iterative update is performed, and the output control variables of each subsystem and the vehicle's predicted state trajectory are extracted from the joint internal state of the last iteration step.

[0060] Step S5 specifically includes: Step S51: Define the joint internal state of the differentiable ADMM layer at each iteration step.

[0061] To facilitate backpropagation, the first The local decision variables, dual variables, and shared variables of the iterative step (expansion layer) are merged into a joint internal state: ; In the formula, The first layer of the differentiable ADMM layer The joint internal state of the iteration step; Indicates the first The subsystem in the first Local decision variables for each iteration step; Indicates the first The subsystem in the first The dual variable of the iteration step; Indicates the first Shared variables for iteration steps.

[0062] Step S52: Perform a preset fixed number of iterations. Iterative updates.

[0063] No. Joint internal state of the iteration step for: ; In the formula, The first layer of the differentiable ADMM layer Iterative mapping of iteration steps; This indicates the order of the expansion steps, which is the order of the fixed iteration steps (expansion layers) of the differential ADMM layer.

[0064] In each iteration of the joint internal state, the local decision variables, dual variables, and shared variables are continuously updated, as expressed by the following formula:

[0065]

[0066]

[0067] In the formula, , and They represent the updated (number) Local decision variables, shared variables, and dual variables of the iterative step. Indicates the local feasible region Find the local decision variable that minimizes the objective within the brackets. ; Indicates the first The local prediction cost of each participant under the influence of the collaborative parameter set; This represents the ADMM quadratic penalty coefficient; Represents the Euclidean norm; Represents a shared feasible set The projection operator; This indicates that the average is calculated over all participants.

[0068] Step S53: Read the underlying output from the joint internal state of the last iteration step.

[0069] After the fixed-step expansion is completed, instead of solving a new optimization problem, the underlying output is read from the joint internal state of the last iteration step:

[0070] ; In the formula, This indicates the function for reading control values; This represents the function for reading the predicted state trajectory. Indicates the joint internal state of the last iteration step. and The current sampling period Below, the output control quantities of each subsystem and the predicted state trajectory of the vehicle are obtained through the differentiable ADMM layer.

[0071] Before performing step S5, it is necessary to construct a differentiable ADMM layer and train the Actor network and Critic network offline. The differentiable ADMM layer participates in backpropagation as a fixed-step differentiable solution mapping.

[0072] Constructing a differentiable ADMM layer is specifically as follows: During the training phase, the ADMM with a fixed number of iterations is written as an unfolded mapping to construct a differentiable ADMM layer. The differentiable ADMM layer receives the current vehicle state vector, cooperative parameters, and unfolded steps, and outputs the output control quantities of each subsystem and the predicted vehicle state trajectory.

[0073] ; In the formula, The fixed number of iterations for a differentiable ADMM layer; and All reads are extracted from the joint internal state in the final step. The specific composition is defined as follows: ; In the formula, Indicates the current sampling period The following is the result of optimization using a differentiable ADMM layer: Predict the vehicle status step ; Indicates the prediction step size; This indicates stacking by column. This represents the fixed-step ADMM expansion mapping of a differentiable ADMM layer.

[0074] Offline training of Actor and Critic networks includes: The Critic network updates its parameters by minimizing its own temporal difference error loss, as described in existing techniques. Meanwhile, the Actor network employs the following objective function. Update the parameters.

[0075] ; ; in, The strategy payoff term is obtained through the Q-value or advantage function output by the Critic network, as detailed in existing technologies. For the parameters of the Actor network, To aid in loss weighting, non-negative values ​​are used. To mitigate losses, Indicates the prediction step size; , , and These represent the auxiliary loss weights for the center of gravity sideslip angle deviation, yaw rate deviation, body roll angle deviation, and control increment, respectively. , and These represent the deviations between the predicted state and the corresponding stability reference centroid sideslip angle, stability reference yaw rate, and stability reference vehicle roll angle, respectively. This represents the change in control input between adjacent prediction steps. An auxiliary loss constructed from the lower-level prediction trajectory is introduced, which constrains the updating of higher-level collaborative parameters to the stability of the lower-level predictions and the smoothness of control.

[0076] When updating the Actor network in reverse, the gradient of the bottom-level optimized output with respect to the cooperative parameter set and the continuous actions of the Actor is first calculated based on the Jacobian recursion of each iterative mapping in the differentiable ADMM layer. Then, the objective function is obtained. Regarding Actor network parameters The gradient is then calculated, and the optimizer performs gradient descent to update the Actor network.

[0077] The gradient of the underlying optimization output with respect to the set of cooperative parameters and the continuous actions of the Actor is calculated as follows: Since the updates of local decision variables, shared variable projections, and dual variable updates all consist of a finite number of deterministic computational steps, the gradient relationship between the bottom-level outputs (predicted state trajectories and output control variables) and the set of cooperative parameters can be calculated during the training phase based on the Jacobian recursion of the expanded layers. First, the Jacobian matrix of the joint internal states with respect to the cooperative parameters is defined: ; In the formula, Indicates the first The joint internal state of the iterative step on the set of cooperative parameters The Jacobian matrix is ​​used to characterize the gradient relationship between the internal state of the fixed-step ADMM unfolded layer and the cooperative parameters. The symbol represents the partial derivative, used to describe the rate of change of one variable with respect to another.

[0078]

[0079] In the formula, Indicates the first Iterative mapping of iteration steps Jacobian matrix for the joint internal state; Indicates the first Iterative mapping of iteration steps The direct impact on the set of collaborative parameters; through As can be seen from the formula, the gradient of the joint internal state is obtained by superimposing the propagation term of the previous iteration step and the direct action term of the parameter of the current iteration step.

[0080]

[0081]

[0082] In the formula, The output control quantity of the differentiable ADMM layer is expressed in relation to the set of cooperative parameters. Jacobian matrix; This represents the predicted state trajectory output of a differentiable ADMM layer with respect to the set of cooperative parameters. Jacobian matrix; and These represent two read functions. and The Jacobian matrix of the joint internal states in the last iteration step, Indicates the first The joint internal state of the iterative step on the set of cooperative parameters The Jacobian matrix.

[0083] ; ; In the formula, The output control quantity of the differentiable ADMM layer continuously affects the Actor output. Jacobian matrix, Represents the continuous action of the cooperative parameter set on the Actor's output. Jacobian matrix, The predicted state trajectory of the differentiable ADMM layer output is a continuous action of the Actor output. The gradient. The above chain relationship is used to represent the optimization results of the lower-level constraints of the differentiable ADMM layer via the collaborative parameter set. The partial derivative is passed to the Actor network output, thereby participating in the Actor network parameter update. In this embodiment, when the partial derivative object is a vector or trajectory variable, the corresponding partial derivative represents the Jacobian matrix; the gradient information or gradient relationship of this application includes the local change relationship represented by the aforementioned Jacobian matrix.

[0084] Find the objective function Regarding Actor network parameters The gradient is as follows: ; In the formula, This indicates the parameters of the Actor network. Find the gradient; This represents the continuous action of the auxiliary loss on the Actor's output. The gradient is obtained based on the Jacobian recursive relationship of each iteration mapping in the fixed-step ADMM expansion layer; Represents the continuous actions output by the Actor. For Actor network parameters The gradient of the underlying predictive control is shown in the equation. This equation indicates that the underlying predictive control performance can participate in the Actor parameter update.

[0085] In this embodiment, the fixed-step ADMM expansion (differentiable ADMM layer) is retained as the constraint optimization solver during the online control phase, and the control quantities of each subsystem and the predicted state trajectory of the vehicle in the current sampling period are output using a rolling optimization method; gradient backpropagation and parameter update are only used to establish parameter update channels during the offline training phase.

[0086] Example 2: A vehicle limit stability control system based on operating condition perception includes: The data acquisition module is used to acquire vehicle motion state, road surface adhesion estimation value, and control input of DYC subsystem, ARS subsystem, and ASS subsystem. It constructs vehicle state vector and single-step working condition observation vector, and then constructs working condition features and control channel features of each subsystem and splices them together to obtain combined features. The spatiotemporal attention working condition perception module is used to extract the temporal context vector based on the single-step working condition observation vector of the current time and its previous time, and combine it with the combined features to perform cross-attention fusion to obtain the working condition context features. The collaborative parameter generation module is used to continuously output actions based on the working condition context features using the Actor-Critic algorithm's Actor network, and map these actions to the collaborative parameters of each subsystem. The cooperative game modeling module is used to construct a cooperative game optimization problem with each subsystem as a participant, given the cooperative parameters of each subsystem and the vehicle state prediction model. The control output module is used to solve the cooperative game optimization problem by cascading a differentiable ADMM layer at the back end of the Actor network, and obtain the output control quantity and vehicle predicted state trajectory of each subsystem. The differentiable ADMM layer is composed of ADMM expansion with a fixed number of iterations, and serves as a differentiable gradient propagation path in offline training to backpropagate the gradient information of the output control quantity and the predicted state trajectory to the Actor network. The DYC subsystem, ARS subsystem, and ASS subsystem are used to output the output control quantities of each subsystem obtained by the control output module to the corresponding actuators.

[0087] For more detailed information on the above modules, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0088] Example 3: The present invention provides a computer device, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the steps of the above-described vehicle limit stability control method based on operating condition perception.

[0089] For more detailed information on the above methods, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0090] Example 4: The present invention provides a computer-readable storage medium for storing a computer program; when the computer program is executed by a processor, it implements the steps of the above-described vehicle limit stability control method based on operating condition perception.

[0091] For more detailed information on the above methods, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0092] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The systems and storage media disclosed in the embodiments are described simply because they correspond to the methods disclosed in the embodiments; relevant details can be found in the method section.

[0093] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0094] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A vehicle limit stability control method based on operating condition perception, employing a collaborative control system integrating a DYC subsystem, an ARS subsystem, and an ASS subsystem, characterized in that, The method includes: The vehicle motion state, road surface adhesion estimate, and control input of each subsystem are obtained. The vehicle state vector and single-step working condition observation vector are constructed, and the working condition features and control channel features of each subsystem are constructed and then concatenated to obtain the combined features. Based on the single-step working condition observation vectors of the current time and its preceding time, extract the time context vectors, and combine them with the combined features to perform cross-attention fusion to obtain the working condition context features; Based on the characteristics of the working condition context, the Actor network using the Actor-Critic algorithm outputs continuous actions and maps them to the cooperative parameters of each subsystem. Given the cooperative parameters of each subsystem and the vehicle state prediction model, a cooperative game optimization problem is constructed with each subsystem as a participant. The cooperative game optimization problem is solved by cascading a differentiable ADMM layer at the back end of the Actor network to obtain the output control quantity and vehicle predicted state trajectory of each subsystem. The differentiable ADMM layer is composed of ADMM expansion with a fixed number of iterations and serves as a differentiable gradient propagation path in offline training, propagating the gradient information of the output control quantity and predicted state trajectory back to the Actor network.

2. The vehicle limit stability control method based on operating condition perception according to claim 1, characterized in that, The vehicle state vector is: ； In the formula, Indicates the current sampling period The vehicle state vector; Indicates the longitudinal speed of the vehicle; Indicates the lateral speed of the vehicle; Indicates yaw rate; Indicates the centroid sideslip angle; Indicates the vehicle body roll angle; Indicates the roll rate; This represents the estimated value of road surface adhesion. This represents the control inputs of each subsystem in the previous sampling period, including the DYC additional yaw moment, the ARS rear wheel steering angle, and the ASS active suspension anti-roll force. This indicates the concatenation of vectors or matrices; Indicates transpose; The single-step operating condition observation vector is: ； In the formula, Indicates the current sampling period The single-step working condition observation vector; This indicates the error of the centroid sideslip angle relative to the expected value; This indicates the error of the yaw rate relative to the expected value; This indicates the error of the vehicle body roll angle relative to the expected value; Indicates the equivalent steering angle or steering input of the front wheels; This indicates the lateral acceleration of the vehicle.

3. The vehicle limit stability control method based on operating condition perception according to claim 1, characterized in that, The extraction of time context vectors and the fusion of combined features with cross-attention to obtain working condition context features are specifically as follows: time context vectors are extracted using a time feature extraction network, the time context vectors are projected to obtain query vectors, the combined features are projected to obtain key feature matrices and value feature matrices, and the working condition context features are obtained by fusing them through a multi-head attention mechanism.

4. The vehicle limit stability control method based on operating condition perception according to claim 1, characterized in that, The Actor network, based on the working condition context features and utilizing the Actor-Critic algorithm, outputs continuous actions and maps them to the collaborative parameters of each subsystem, specifically as follows: Input the working condition context features into the Actor network, and output continuous actions; ； In the formula, This indicates that the Actor network is in the current sampling period. The continuous output action; The parameter is Actor network; Indicates the contextual features of the operating condition; This indicates the distribution trend of lateral stabilization tasks between the DYC subsystem and the ARS subsystem; This indicates the importance of lateral stability objectives; This underscores the importance of a tilt-stabilized target; The continuous actions output by the Actor network are used to construct the cooperative parameters of each subsystem: ； ； ； ； In the formula, Represents the set of collaborative parameters; , , These represent the DYC subsystem, ARS subsystem, and ASS subsystem respectively during the current sampling period. The following are the collaborative parameters.

5. The vehicle limit stability control method based on operating condition perception according to claim 1, characterized in that, Given the cooperative parameters of each subsystem and the vehicle state prediction model, a cooperative game optimization problem is constructed by treating each subsystem as a participant, specifically as follows: Establish a vehicle state prediction model that includes actuator delay; Transform the cooperative parameters of the subsystem into state tracking weights; ； in, For the first Each subsystem in the current sampling period State tracking weight matrix; For the first Each subsystem in the current sampling period The collaborative parameters; Let be the semi-positive definite nominal state tracking weight matrix. It is a collection of subsystems, including the DYC subsystem, the ARS subsystem, and the ASS subsystem; Constructing a cooperative game optimization problem : ， ； ； in, Indicates the first Subsystems in local decision variables The cost of local decision variables Including the The vehicle state sequence, its own control sequence, and local estimates of the control sequences of other subsystems are obtained from the prediction of each subsystem. Representing local decision variables Belongs to the Local feasible region of each subsystem ; Representing local decision variables With shared variables Maintain consistency and share variables This includes globally consistent predicted state trajectories and globally consistent control sequences for each subsystem. This indicates that the constraints are met. To predict the step size, For the first The subsystem in the first The stable target tracking error for each prediction step is obtained based on a pre-calculated stable reference target; Indicates the first The subsystem in the first Control input for each prediction step; Indicates the first The subsystem in the first The control input increment for each prediction step; and These represent the input consumption weight and the input increment weight, respectively. This indicates transpose.

6. The vehicle limit stability control method based on operating condition perception according to claim 5, characterized in that, The first The subsystem in the first Stable target tracking error in each prediction step Specifically: ； ； ； ； In the formula, Indicates the first Each prediction step stability error vector This represents the error selection matrix, with values... , or , For yaw and lateral deviation errors, For lateral and yaw errors, This refers to the error between the roll angle and the roll rate. This indicates that a diagonal matrix is ​​constructed using the elements within the parentheses as diagonal elements. , , and They represent the first The actual centroid sideslip angle, actual yaw rate, actual body roll angle, and actual body roll rate are predicted in each prediction step; , , and These represent the stability reference center of gravity sideslip angle, stability reference yaw rate, stability reference body roll angle, and stability reference body roll rate, respectively.

7. The vehicle limit stability control method based on operating condition perception according to claim 1, characterized in that, The method of solving the cooperative game optimization problem by cascading a differentiable ADMM layer at the back end of the Actor network involves: merging the local decision variables, dual variables, and shared variables of each iteration step of the differentiable ADMM layer into a joint internal state; performing a preset fixed number of iterations for iterative updates; and extracting the output control variables of each subsystem and the predicted vehicle state trajectory from the joint internal state of the last iteration step. The local decision variables include the vehicle state sequence predicted by the current subsystem, its own control sequence, and local estimates of the control sequences of other subsystems. The shared variables include the globally consistent predicted state trajectory and the globally consistent control sequence of each subsystem. The dual variables are used to accumulate and correct the consistency deviation between the local decision variables and the shared variables.

8. The vehicle limit stability control method based on operating condition perception according to claim 1, characterized in that, During offline training, the objective function is as follows: Update the parameters of the Actor network: ； ； in, For strategy benefit items, For the parameters of the Actor network, To assist in loss weighting, To mitigate losses, Indicates the prediction step size; , , and These represent the auxiliary loss weights for the center of gravity sideslip angle deviation, yaw rate deviation, body roll angle deviation, and control increment, respectively. , and These represent the deviations between the predicted state and the corresponding stability reference centroid sideslip angle, stability reference yaw rate, and stability reference vehicle roll angle, respectively. This indicates the change in control quantity between adjacent prediction steps; In offline training, the differentiable ADMM layer serves as a differentiable gradient propagation path, backpropagating the gradient information of the output control quantity and the predicted state trajectory to the Actor network. Specifically, when updating the Actor network, the gradients of the output control quantity and the predicted state trajectory with respect to the cooperative parameter set and the continuous actions of the Actor are first calculated based on the Jacobian recursion relationship of each iteration mapping in the differentiable ADMM layer. Then, the objective function is obtained. Regarding Actor network parameters The gradient is then calculated, and the optimizer performs gradient descent to update the Actor network.

9. A vehicle limit stability control system based on operating condition perception, characterized in that, include: The data acquisition module is used to acquire vehicle motion state, road surface adhesion estimation value, and control input of DYC subsystem, ARS subsystem, and ASS subsystem. It constructs vehicle state vector and single-step working condition observation vector, and then constructs working condition features and control channel features of each subsystem and splices them together to obtain combined features. The spatiotemporal attention working condition perception module is used to extract the temporal context vector based on the single-step working condition observation vector of the current time and its previous time, and combine it with the combined features to perform cross-attention fusion to obtain the working condition context features. The collaborative parameter generation module is used to continuously output actions based on the working condition context features using the Actor-Critic algorithm's Actor network, and map these actions to the collaborative parameters of each subsystem. The cooperative game modeling module is used to construct a cooperative game optimization problem with each subsystem as a participant, given the cooperative parameters of each subsystem and the vehicle state prediction model. The control output module is used to solve the cooperative game optimization problem by cascading a differentiable ADMM layer at the back end of the Actor network, and obtain the output control quantity and vehicle predicted state trajectory of each subsystem. The differentiable ADMM layer is composed of ADMM expansion with a fixed number of iterations, and serves as a differentiable gradient propagation path in offline training to backpropagate the gradient information of the output control quantity and the predicted state trajectory to the Actor network. The DYC subsystem, ARS subsystem, and ASS subsystem are used to output the output control quantities of each subsystem obtained by the control output module to the corresponding actuators.

10. A computer-readable storage medium, characterized in that, Used to store computer programs; when executed by a processor, the computer programs implement the steps of the vehicle limit stability control method based on operating condition perception as described in any one of claims 1-8.

Images