Angle module induced effect compensation system and method based on physical-data fusion

Abstract

The invention relates to an angle module induced effect compensation system and method based on physical-data fusion, and the system comprises a whole vehicle motion planning module which obtains a whole vehicle expected attitude and a rotation angle through calculation; the transfer effect feedforward compensation module is used for constructing a physical embedded feedforward model and estimating transfer disturbance torque in real time; the data driving prediction module is used for constructing a dual-scale space-time fusion network model, extracting long-time-sequence inertial features and short-time-sequence sudden change features of vehicle dynamics, and predicting a vehicle prediction state in a rolling manner; the multi-constraint optimization control module is used for solving the optimal residual compensation torque; the feed-forward parameter online self-adaptive calibration module is used for designing a disturbance observer based on inverse dynamics, extracting real physical disturbance in real time, and correcting feed-forward model parameters online by using a recursive least square algorithm; and an actuator actuation module. Non-linear interference of a turning effect is overcome through a feed-forward-optimization-self-adaptive closed loop mechanism, and high-precision tracking of a vehicle turning angle and coordinated stable control of a whole vehicle yaw attitude are achieved.

Description

Technical Field

[0001] This invention belongs to the field of vehicle drive-by-wire intelligent chassis control, specifically relating to a corner module-induced rotation effect compensation system and method based on physical-data fusion. Background Technology

[0002] With the development of vehicle electrification and intelligence, modular chassis capable of high-mobility maneuvers such as 90° steering and lateral translation have become a key technological direction. The split-type modular vehicle suspension system retains the traditional connection method with the vehicle body. The suspension does not rotate synchronously with the wheels, has high lateral stiffness, and can withstand impact loads at high speeds, making it suitable for off-road driving and high-speed maneuvering conditions of special vehicles.

[0003] In traditional vehicles, the driving and braking forces acting on the tires, applied to the kingpin offset, generate a torque that causes the wheel to turn, known as the "rotational effect." This torque is typically counteracted by the mechanical connection of the steering tie rod in conventional designs. However, in modular steering vehicles, the elimination of the steering tie rod means that the rotational effect resulting from the coupling of the tire's longitudinal driving / braking forces with the suspension's dynamic vibrations cannot be passively counteracted by the mechanical structure. Instead, it transforms into a strong nonlinear disturbance directly acting on the steering actuator. This makes it difficult for existing steering controllers to accurately track the target steering angle, especially during vehicle acceleration / deceleration or when driving on uneven surfaces, leading to decreased steering angle control accuracy and vehicle deviation from its trajectory.

[0004] In the field of corner module vehicle steering control, for example, Chinese invention patent application No. CN202410342841.4, entitled "An Integrated Control Method for Drive, Braking and Steering of a Fully Electromagnetic Intelligent Corner Module", calculates the driving and braking forces of each wheel based on the desired steering angle and drive and braking model, compensating for path tracking deviation caused by steering delay and improving the real-time path tracking performance of the vehicle; Chinese invention patent application No. CN202411072093.9, entitled "Real-time Motion Control System and Method for Four-Wheel Steering Vehicle with Corner Module Configuration", proposes a method based on a combination of model predictive control and sliding mode control. The hierarchical motion control architecture solves the problem of coordinated control of path tracking accuracy and lateral stability of corner module vehicles at different speeds by planning the optimal front wheel steering angle and additional yaw moment in the upper layer and distributing the front and rear wheel steering angles in the middle layer using a feedforward-feedback strategy. In Chinese invention patent application number CN202510475133.2, entitled "A method for suppressing yaw in a four-wheel independent steer-by-wire system", a wheel yaw dynamic model of a four-wheel independent steer-by-wire system is proposed. Through the design of an adaptive backstepping sliding mode controller with yaw amplitude-frequency characteristics, active adaptive control is performed on yaw phenomena of different frequencies and amplitudes.

[0005] However, existing corner module steering control methods have the following two problems:

[0006] First, existing traditional control algorithms (sliding mode control, PID, etc.) have control lag problems, while model predictive control (traditional MPC) has limitations in computing power. It lacks the ability to actively compensate for such high-frequency, strongly coupled "turnover effect" disturbances, and has problems of response lag and insufficient robustness.

[0007] Second, existing tracking algorithms often simplify the model linearly, ignoring the nonlinear coupling relationship between the dynamic motion of the vehicle suspension and the steering disturbance torque, and thus cannot describe the mechanism between suspension, steering and drive.

[0008] Third, therefore, how to design a control system that can actively, in real time and accurately compensate for the interference of the "steering effect" and thus avoid the problem of low steering control accuracy or even instability of the corner module vehicle is a key issue that needs to be solved to achieve high-precision control of the split corner module vehicle. Summary of the Invention

[0009] The purpose of this invention is to provide a corner module-induced rotation effect compensation system and method based on physical-data fusion.

[0010] The technical solution to achieve the purpose of this invention is: a corner module rotation effect compensation system based on physical-data fusion, including a vehicle motion planning module, a rotation effect feedforward compensation module, a data-driven prediction module, a multi-constraint optimization control module, a feedforward parameter adaptive calibration module, and an actuator actuation module;

[0011] Vehicle motion planning module: Receives steering wheel angle signals from the driver. and the vehicle's current longitudinal speed signal Based on a two-degree-of-freedom vehicle dynamics model and a four-wheel steering Ackerman geometric model, the desired attitude of the whole vehicle is calculated and output in real time. ) and expected turning angle ;

[0012] The turn-induced effect feedforward compensation module establishes a split-type angle module physically embedded turn-induced coupling characteristic equation based on the principles of Lagrange dynamics. It then uses the full dataset obtained from offline identification to solve for the weight coefficients of the turn-induced coupling characteristic equation. Based on the calculated turn-induced coupling characteristic equation, it calculates the feedforward value of the turn-induced disturbance torque using real-time vehicle longitudinal force, suspension dynamic deflection, and wheel rotation angle signals. ;

[0013] The data-driven prediction module constructs a dual-scale spatiotemporal fusion network model to extract long-term inertial features and short-term abrupt change features of vehicle dynamics in parallel, constrained by the physical envelope projection operator. Based on the candidate compensation torque sequence output by the multi-constraint optimization control module, it predicts the vehicle's turning angle and yaw attitude in the future time domain. ;

[0014] Multi-constraint optimization control module: An improved particle swarm optimization algorithm based on dynamic trust region and time series hot start is designed. It iterates under the constraints of vehicle stability and actuator physical saturation, based on the feedforward value of the rotation disturbance torque. The desired attitude target and desired rotation angle are used to solve the dual-scale spatiotemporal fusion network model and obtain the optimal residual compensation torque sequence that minimizes the tracking error.

[0015] The feedforward parameter adaptive calibration module constructs a steer-by-wire inverse dynamics observer, calculates the true observation value based on the electromagnetic torque of the steering motor and the vehicle motion state, calculates the deviation between the true observation value and the feedforward value using the recursive least squares algorithm, and outputs the corrected characteristic equation weight coefficients online.

[0016] Actuator module: Receives the total torque resulting from the superposition of the feedforward value of the torque-induced disturbance and the optimal residual compensation torque, and outputs the final target torque command;

[0017] A method for compensating for the angular module-induced ...

[0018] Step (1): Establish a two-degree-of-freedom vehicle dynamics model and a four-wheel steering Ackerman geometric model;

[0019] Step (2): Construct the physical embedded rotation coupling characteristic equation of the split angle module based on the Lagrange dynamics principle, use offline identification to fit the weight coefficient of the characteristic equation, and output the feedforward compensation torque;

[0020] Step (3): Construct a dual-scale spatiotemporal fusion network with physical envelope projection operators to predict future states in a rolling manner;

[0021] Step (4): Design an improved particle swarm optimization algorithm based on dynamic trust region and time series hot start, and transform vehicle phase plane stability and actuator physical saturation into constraints to solve for the optimal residual compensation torque;

[0022] Step (5): Adaptively calibrate the weight coefficients of the characteristic equation based on steering inverse dynamics and recursive least squares strategy;

[0023] Step (6): The feedforward compensation torque and the optimal residual torque are superimposed and output to achieve active compensation for the rotation disturbance.

[0024] Compared with the prior art, the significant advantages of this invention are:

[0025] 1. This invention designs an improved particle swarm optimization algorithm based on dynamic trust region and time series hot start. It maintains wide-area search when the model accuracy is low and automatically shrinks the trust region boundary when the model accuracy improves. This can significantly compress the iteration convergence time, so that the optimization process does not need to blindly iterate in the invalid solution space, avoiding the risk of real-time control failure caused by the excessive computation time of complex nonlinear control algorithms.

[0026] 2. This invention constructs a dual-scale spatiotemporal fusion prediction model (LSTM+CNN) under physical envelope constraints, which can capture transient change features when facing road impacts or high-frequency interference. At the same time, it forces the output of the neural network to always be within the feasible domain of vehicle dynamics, avoiding control divergence and vehicle instability caused by the pure data-driven model under unknown conditions, and promoting the corner tracking accuracy and robustness of the system under high-frequency nonlinear conditions.

[0027] 3. This invention constructs a parameter adaptive calibration method based on inverse dynamics, which solves physical disturbances and corrects the parameters of the feedforward model to continuously reduce the prediction bias of the model. This enables the system to update and compress the solution range of the optimization algorithm in real time based on higher model confidence, avoiding the waste of computing power that would otherwise have to reserve too wide a search margin due to excessive model uncertainty.

[0028] 4. Based on the collaborative control architecture of mechanism feedforward and data prediction, this invention realizes the estimation and active cancellation of nonlinear rotation interference caused by split-type corner modules; while effectively eliminating the corner tracking deviation caused by the coupling of driving and braking loads and suspension motion, it significantly improves the trajectory tracking accuracy and driving stability of the vehicle under complex dynamic conditions. Attached Figure Description

[0029] Figure 1 This is a flowchart of the control module of the angle module-induced rotation effect compensation system in this application;

[0030] Figure 2 This is a general block diagram of the corner module-induced transfer effect compensation method of this application;

[0031] Figure 3 This is a schematic diagram illustrating the mechanism of the transduction effect in this application. Detailed Implementation

[0032] The present invention will now be described in further detail with reference to the accompanying drawings.

[0033] A corner module-induced rotation effect compensation system based on physical-data fusion includes a vehicle motion planning module, a rotation effect feedforward compensation module, a data-driven prediction module, a multi-constraint optimization control module, a feedforward parameter adaptive calibration module, and an actuator actuation module.

[0034] Vehicle motion planning module: Receives steering wheel angle signals from the driver. and the vehicle's current longitudinal speed signal Based on a two-degree-of-freedom vehicle dynamics model and a four-wheel steering Ackerman geometric model, the desired attitude of the whole vehicle is calculated and output in real time. ) and expected turning angle ;

[0035] The turn-induced effect feedforward compensation module establishes a split-type angle module physically embedded turn-induced coupling characteristic equation based on the principles of Lagrange dynamics. It then uses the full dataset obtained from offline identification to solve for the weight coefficients of the turn-induced coupling characteristic equation. Based on the calculated turn-induced coupling characteristic equation, it calculates the feedforward value of the turn-induced disturbance torque using real-time vehicle longitudinal force, suspension dynamic deflection, and wheel rotation angle signals. ; The data-driven prediction module constructs a dual-scale spatiotemporal fusion network model to extract long-term inertial features and short-term abrupt change features of vehicle dynamics in parallel, constrained by the physical envelope projection operator. Based on the candidate compensation torque sequence output by the multi-constraint optimization control module, it predicts the vehicle's turning angle and yaw attitude in the future time domain. ;

[0036] Multi-constraint optimization control module: An improved particle swarm optimization algorithm based on dynamic trust region and time series hot start is designed. It iterates under the constraints of vehicle stability and actuator physical saturation, based on the feedforward value of the rotation disturbance torque. The desired attitude target and desired rotation angle are used to solve the dual-scale spatiotemporal fusion network model and obtain the optimal residual compensation torque sequence that minimizes the tracking error.

[0037] The feedforward parameter adaptive calibration module constructs a steer-by-wire inverse dynamics observer, calculates the true observation value based on the electromagnetic torque of the steering motor and the vehicle motion state, calculates the deviation between the true observation value and the feedforward value using the recursive least squares algorithm, and outputs the corrected characteristic equation weight coefficients online.

[0038] Actuator module: Receives the total torque resulting from the superposition of the feedforward value of the torque-induced disturbance and the optimal residual compensation torque, and outputs the final target torque command;

[0039] A method for compensating for the angular module-induced ...

[0040] Step (1): Establish a two-degree-of-freedom vehicle dynamics model and a four-wheel steering Ackerman geometric model;

[0041] Step (2): Construct the physical embedded rotation coupling characteristic equation of the split angle module based on the Lagrange dynamics principle, use offline identification to fit the weight coefficient of the characteristic equation, and output the feedforward compensation torque;

[0042] Step (3): Construct a dual-scale spatiotemporal fusion network with physical envelope projection operators to predict future states in a rolling manner;

[0043] Step (4): Design an improved particle swarm optimization algorithm based on dynamic trust region and time series hot start, and transform vehicle phase plane stability and actuator physical saturation into constraints to solve for the optimal residual compensation torque;

[0044] Step (5): Adaptively calibrate the weight coefficients of the characteristic equation based on steering inverse dynamics and recursive least squares strategy;

[0045] Step (6): The feedforward compensation torque and the optimal residual torque are superimposed and output to achieve active compensation for the rotation disturbance;

[0046] Furthermore, step (1) specifically includes:

[0047] Step (11): Receive the steering wheel angle signal from the driver. and the vehicle's current longitudinal speed signal ;

[0048] Step (12): Calculate the equivalent front wheel steering angle based on the fixed steering gear ratio. ;

[0049]

[0050] In the formula, This is the equivalent front wheel steering angle; Steering wheel angle; A fixed steering wheel gear ratio;

[0051] Based on a zero-center-of-gravity sideslip angle control strategy and a two-degree-of-freedom vehicle dynamics model, the current vehicle speed is calculated. Front and rear wheel steering ratio :

[0052]

[0053] In the formula: This is the distance from the center of mass to the front and rear axles; For the overall vehicle weight; For the front and rear axle lateral stiffness; Wheelbase; Current vehicle speed;

[0054] Step (13): Based on the two-degree-of-freedom dynamic model, the equivalent front wheel steering angle is calculated. and equivalent rear wheel steering angle As a system control input:

[0055] =0

[0056]

[0057] In the formula, This refers to the yaw rate; It is the centroid sideslip angle; For vehicle stability factors;

[0058] Step (14): Establish a four-wheel steering Ackerman steering geometry model and calculate the expected steering angles of the four wheels based on the current vehicle state:

[0059]

[0060] In the formula, For the turning angle of the wheel, , For the front axle, For the rear axle, , On the right side, Left side; The lateral position coefficient is defined as -1 on the left and 1 on the right; L is the wheelbase; B is the track width.

[0061] Furthermore, step (2) specifically includes:

[0062] Step (21): Establish a split-type angular module-induced rotation effect model based on the Lagrange dynamics equation, and select the wheel rotation angle. Vertical deflection of the suspension As the generalized coordinate vector of the system Based on the principles of Lagrange dynamics Define the system energy term:

[0063] Total kinetic energy of the system T:

[0064]

[0065] In the formula, The equivalent moment of inertia of the steering system (including the motor rotor and reduction gear); unsprung mass

[0066] Total potential energy of the system U:

[0067]

[0068] In the formula, This refers to the vertical stiffness of the suspension. , is the coupling potential energy of the positioning parameters; Based on the base lift stiffness coefficient, This is the geometric coupling sensitivity coefficient;

[0069] Relationship between suspension dynamic deflection and kingpin offset:

[0070]

[0071] In the formula, This represents the second-order nonlinear curvature characteristic of the suspension motion; is the first-order linear rate of change of the suspension motion; This refers to the static kingpin offset.

[0072] Using the principle of virtual work Solving for the generalized forces in the steering degree of freedom yields the second-order dynamic equations of the rotation-induced disturbance mechanism:

[0073]

[0074] In the formula, The equivalent viscous damping coefficient of the steering system; This refers to the electromagnetic torque of the motor. This is the restoring torque; This is the dry friction torque;

[0075] Based on the disturbance torque term of the second-order equation, construct the physical embedded characteristic equation:

[0076]

[0077] In the formula, The coefficient of sensitivity for induced coupling; This is the equivalent normalizing stiffness coefficient; It is the equivalent viscous damping coefficient;

[0078] Step (22): Input the longitudinal drive / braking load covering the entire range, targeting Features, application Stepped longitudinal load; for Features, application Vertical excitation of the road surface;

[0079] Generate decoupled feature vectors Construct a full dataset containing input features and output labels:

[0080]

[0081] Step (23): Construct the regression equation and the observation vector Y and feature matrix X:

[0082]

[0083] Define the weight coefficient vector The optimal weighting coefficients are solved using the least squares method:

[0084]

[0085] Step (24): Define the residual The root mean square error formula is used to quantify the absolute error:

[0086]

[0087] Furthermore, step (3) specifically includes:

[0088] Step (31): Utilize the dataset from the feedforward compensation module for the induction effect. A time series dataset is constructed using the sliding window method, and an input tensor containing historical states and future predictions is designed. :

[0089]

[0090] In the formula, , which is a sequence of historical states; For future control state sequences; For historical steps; To predict the step size;

[0091] Output tensor corresponding to the future prediction step size :

[0092]

[0093] Step (32): Design a dual-scale spatiotemporal fusion network to capture the long-term inertial characteristics and short-term abrupt change characteristics of the vehicle, respectively;

[0094] Using recursive operators Capture long-term time dependencies and simulate the steady-state steering response of the vehicle:

[0095]

[0096] In the formula: The extracted low-frequency inertial feature vector; Characterizing the nonlinear recursive mapping of stacked LSTM cells; For input tensors; For sequential memory weight set;

[0097] Using local convolution operator Capture transient changes and simulate the direct excitation of the actuator by road impact:

[0098]

[0099] In the formula: The extracted high-frequency transient feature vector; Characterizes one-dimensional temporal convolution operations; For short-term receptive field convolution kernel weights; The convolution bias vector;

[0100] To address the alignment issue of features across different frequency bands, a weighted fusion mechanism is constructed to map inertial and transient features to the same latent space.

[0101]

[0102] In the formula: This is a global dynamical hidden vector; , These are the low-frequency and high-frequency adaptive fusion matrices, respectively. This is the global feature fusion bias vector;

[0103] Design a physical envelope projection operator Force the predicted value to fall within the physical feasible region of the corner module:

[0104]

[0105] In the formula: Using the hyperbolic tangent activation function, the hidden layer features are nonlinearly compressed to... interval; The physical limit scaling tensor is defined as a diagonal matrix. ; The projection matrix of the fully connected output layer;

[0106] Step (33): Using a dual averaging strategy, design a multi-objective weighted loss function:

[0107]

[0108] In the formula, It is the number of samples; To predict the step size; Weights for the execution layer; For trajectory layer weights; Here are the weights for the stable layer; where the weights satisfy the following conditions: ;

[0109] Step (34): Perform training iterations, feeding the input tensor The superposition amplitude is Gaussian white noise perturbation; define the optimal time. Termination conditions :

[0110]

[0111]

[0112] In the formula, To validate the set loss; The current moment; The threshold number of steps;

[0113] Furthermore, step (4) specifically includes:

[0114] Step (41): Design a dynamic trust region based on tracking error, define the population size as P, and define each particle j as a... The residual compensation torque sequence for predicting the step size:

[0115]

[0116] Real-time calculation of the comprehensive tracking error norm at the current moment Construct a nonlinear scaling factor :

[0117]

[0118] In the formula, These are the upper and lower bounds of the trust region scaling factor, respectively; This is the error sensitivity coefficient;

[0119] Uncertainty based on feedforward model Error, combined with a dynamic scaling factor, defines the search boundary at the current moment. :

[0120]

[0121] Step (42): Design a hot-start particle swarm optimization algorithm based on time series inheritance; extract the previous control cycle. Output of the globally optimal residual torque sequence The current time step is generated by the left shift operator. High confidence seeds :

[0122]

[0123] In the formula, This is the optimal solution vector after optimization at the previous time step; 0 represents the optimal solution vector. The subsequent control variables are assumed to have zero null.

[0124] The first particle is directly assigned a value. The remaining particles are The mean center is used to generate the data according to a Gaussian distribution within the search boundary.

[0125] Step (43): Construct the total torque: Call the network trained in step (3) to predict the future state tensor. And calculate the fitness function. ;

[0126] Computational tracking accuracy versus energy consumption cost:

[0127]

[0128] In the formula, This refers to the angle error; This refers to the yaw error; This refers to the centroid lateral deviation error. This represents the term for minimizing the residual torque and the constraint on control energy consumption;

[0129] Define a stability index based on a rhombus envelope. With phase plane stability penalty function :

[0130]

[0131]

[0132] In the formula, A high penalty factor; , For dynamic coefficients, take respectively ,

[0133] and ;

[0134] Define the saturation penalty function for the executor:

[0135]

[0136] In the formula, For candidate total command torque; This represents the maximum physical torque of the steering motor; As a penalty factor;

[0137] Calculate the fitness function :

[0138] Step (44): Perform evolutionary update under the dynamic trust region, and perform projection operation on the updated particle positions:

[0139]

[0140]

[0141] In the formula, This refers to the residual torque after the update. This is a transient position; For projection operators;

[0142] Step (45): After the iteration terminates, output the first element of the optimal sequence as the optimal residual compensation torque. ;

[0143] Furthermore, step (5) specifically includes:

[0144] Step (51): Construct an inverse dynamics observation model of the steering system based on the dynamic model of the feedforward compensation module for the steering effect, and extract the actual disturbance observation values. :

[0145]

[0146] Step (52): Extract the real interference from the observer Compared with the reference torque calculated by the current feedforward model Obtain the prediction error of the feedforward model :

[0147]

[0148] Step (53): Based on the current error The recursive least squares strategy is used to correct the weight coefficient vector of the physical characteristic equation of the feedforward compensation module for the inversion effect in real time.

[0149]

[0150] In the formula, This is the gain matrix;

[0151] Step (54): Weight coefficient vector The input is sent to the feedforward compensation module and takes effect in the next control cycle.

Claims

1. A corner module-induced rotation effect compensation system based on physical-data fusion, characterized in that, It includes a vehicle motion planning module, a rotation effect feedforward compensation module, a data-driven prediction module, a multi-constraint optimization control module, a feedforward parameter adaptive calibration module, and an actuator actuation module; Vehicle motion planning module: Receives steering wheel angle signals from the driver. and the vehicle's current longitudinal speed signal Based on a two-degree-of-freedom vehicle dynamics model and a four-wheel steering Ackerman geometric model, the desired attitude of the whole vehicle is calculated and output in real time. ) and expected turning angle ; The turn-induced effect feedforward compensation module establishes a split-type angle module physically embedded turn-induced coupling characteristic equation based on the principles of Lagrange dynamics. It then uses the full dataset obtained from offline identification to solve for the weight coefficients of the turn-induced coupling characteristic equation. Based on the calculated turn-induced coupling characteristic equation, it calculates the feedforward value of the turn-induced disturbance torque using real-time vehicle longitudinal force, suspension dynamic deflection, and wheel rotation angle signals. ; The data-driven prediction module constructs a dual-scale spatiotemporal fusion network model to extract long-term inertial features and short-term abrupt change features of vehicle dynamics in parallel, constrained by the physical envelope projection operator. Based on the candidate compensation torque sequence output by the multi-constraint optimization control module, it predicts the vehicle's turning angle and yaw attitude in the future time domain. ; Multi-constraint optimization control module: An improved particle swarm optimization algorithm based on dynamic trust region and time series hot start is designed. It iterates under the constraints of vehicle stability and actuator physical saturation, based on the feedforward value of the rotation disturbance torque. The desired attitude target and desired rotation angle are used to solve the dual-scale spatiotemporal fusion network model and obtain the optimal residual compensation torque sequence that minimizes the tracking error. The feedforward parameter adaptive calibration module constructs a steer-by-wire inverse dynamics observer, calculates the true observation value based on the electromagnetic torque of the steering motor and the vehicle motion state, calculates the deviation between the true observation value and the feedforward value using the recursive least squares algorithm, and outputs the corrected characteristic equation weight coefficients online. Actuator module: Receives the total torque resulting from the superposition of the feedforward value of the torque causing the rotation disturbance and the optimal residual compensation torque, and outputs the final target torque command.

2. A method for compensating for the angle-induced rotation effect using the angle module rotation effect compensation system according to claim 1, characterized in that, Includes the following steps: Step (1): Establish a two-degree-of-freedom vehicle dynamics model and a four-wheel steering Ackerman geometric model; Step (2): Construct the physical embedded rotation coupling characteristic equation of the split angle module based on the principle of Lagrange dynamics, use offline identification to fit the weight coefficient of the characteristic equation, and output the feedforward compensation torque; Step (3): Construct a dual-scale spatiotemporal fusion network with physical envelope projection operators to predict future states in a rolling manner; Step (4): Design an improved particle swarm optimization algorithm based on dynamic trust region and time series hot start, and transform vehicle phase plane stability and actuator physical saturation into constraints to solve for the optimal residual compensation torque; Step (5): Adaptively calibrate the weight coefficients of the characteristic equation based on steering inverse dynamics and recursive least squares strategy; Step (6): The feedforward compensation torque and the optimal residual torque are superimposed and output to achieve active compensation for the rotation disturbance.

3. The method according to claim 2, characterized in that, Step (1) specifically includes: Step (11): Receive the steering wheel angle signal from the driver. and the vehicle's current longitudinal speed signal ; Step (12): Calculate the equivalent front wheel steering angle based on the fixed steering gear ratio. ; ， In the formula, This is the equivalent front wheel steering angle; Steering wheel angle; A fixed steering wheel gear ratio; Based on a zero-center-of-gravity sideslip angle control strategy and a two-degree-of-freedom vehicle dynamics model, the current vehicle speed is calculated. Front and rear wheel steering ratio : ， In the formula: This represents the distance from the center of mass to the front and rear axles, respectively. For the overall vehicle weight; For the front and rear axle lateral stiffness; Wheelbase; Current vehicle speed; Step (13): Based on the two-degree-of-freedom dynamic model, the equivalent front wheel steering angle is calculated. and equivalent rear wheel steering angle As a system control input: =0， ， In the formula, This refers to the yaw rate; It is the centroid sideslip angle; For vehicle stability factors; Step (14): Establish a four-wheel steering Ackerman steering geometry model and calculate the expected steering angles of the four wheels based on the current vehicle state: ， In the formula, For the turning angle of the wheel, , For the front axle, For the rear axle, , On the right side, Left side; is the lateral position coefficient, defined as -1 on the left and 1 on the right; B is the wheel track.

4. The method according to claim 3, characterized in that, Step (2) specifically includes: Step (21): Establish a split-type angular module-induced rotation effect model based on the Lagrange dynamics equation, and select the wheel rotation angle. Vertical deflection of the suspension As the generalized coordinate vector of the system Based on the principles of Lagrange dynamics Define the system energy term: Total kinetic energy of the system T: ， In the formula, The equivalent rotational inertia of the steering system; Unsprung mass; Total potential energy of the system U: ， In the formula, This refers to the vertical stiffness of the suspension. To locate the coupling potential energy of the parameters, ; Based on the lifting stiffness coefficient, This is the geometric coupling sensitivity coefficient; Relationship between suspension dynamic deflection and kingpin offset: ， In the formula, This represents the second-order nonlinear curvature characteristic of the suspension motion; is the first-order linear rate of change of the suspension motion; This refers to the static kingpin offset. Using the principle of virtual work Solving for the generalized forces in the steering degree of freedom yields the second-order dynamic equations of the rotation-induced disturbance mechanism: ， In the formula, The equivalent viscous damping coefficient of the steering system; This refers to the electromagnetic torque of the motor. This is the restoring torque; This is the dry friction torque; Based on the disturbance torque term of the second-order equation, construct the physical embedded rotation coupling characteristic equation: ， In the formula, The coefficient of sensitivity for induced coupling; This is the equivalent normalizing stiffness coefficient; It is the equivalent viscous damping coefficient; Step (22): Input the longitudinal drive / braking load covering the entire range, targeting Features, application Stepped longitudinal loads; for Features, application Vertical excitation of the road surface; Generate decoupled feature vectors Construct a full dataset containing input features and output labels: ， Step (23): Construct the regression equation and the observation vector Y and feature matrix X: ， ， Define the weight coefficient vector The optimal weighting coefficients are solved using the least squares method: ， Step (24): Define the residual The root mean square error formula is used to quantify the absolute error: 。 5. The method according to claim 4, characterized in that, Step (3) specifically includes: Step (31): Utilize the dataset from the feedforward compensation module for the induction effect. A time series dataset is constructed using the sliding window method, and an input tensor containing historical states and future predictions is designed. : ， In the formula, , represents a historical state sequence; , is the sequence of future control states; For historical steps; To predict the step size; Output tensor corresponding to the future prediction step size : ， Step (32): Design a dual-scale spatiotemporal fusion network to capture the long-term inertial characteristics and short-term abrupt change characteristics of the vehicle, respectively; Using recursive operators Capture long-term time dependencies and simulate the steady-state steering response of the vehicle: ， In the formula: The extracted low-frequency inertial feature vector; Characterizing the nonlinear recursive mapping of stacked LSTM cells; For input tensors; For sequential memory weight set; Using local convolution operator Capture transient changes and simulate the direct excitation of the actuator by road impact: ， In the formula: The extracted high-frequency transient feature vector; Characterizes one-dimensional temporal convolution operations; For short-term receptive field convolution kernel weights; The convolution bias vector; A weighted fusion mechanism is constructed to map inertial features and transient features to the same latent space: ， In the formula: This is a global dynamical hidden vector; , These are the low-frequency and high-frequency adaptive fusion matrices, respectively. This is the global feature fusion bias vector; Design a physical envelope projection operator Force the predicted value to fall within the physical feasible region of the corner module: ， In the formula: Using the hyperbolic tangent activation function, the hidden layer features are nonlinearly compressed to... interval; The physical limit scaling tensor is defined as a diagonal matrix. ; The projection matrix of the fully connected output layer; Step (33): Using a dual averaging strategy, design a multi-objective weighted loss function: ， In the formula, It is the number of samples; To predict the step size; Weights for the execution layer; For trajectory layer weights; Here, the weights are the weights of the stabilizing layer; where the weights satisfy the following conditions: ; Step (34): Perform training iterations, feeding the input tensor The superposition amplitude is Gaussian white noise perturbation; define the optimal time. Termination conditions : ， ， In the formula, To validate the set loss; The current moment; The threshold number of steps.

6. The method according to claim 5, characterized in that, Step (4) specifically includes: Step (41): Design a dynamic trust region based on tracking error, define the population size as P, and define each particle j as a... The residual compensation torque sequence for predicting the step size: ， Real-time calculation of the comprehensive tracking error norm at the current moment Construct a nonlinear scaling factor : ， In the formula, These are the upper and lower bounds of the trust region scaling factor, respectively; This is the error sensitivity coefficient; Uncertainty based on feedforward model Error, combined with a dynamic scaling factor, defines the search boundary at the current moment. : ； Step (42): Design a hot-start particle swarm optimization algorithm based on time series inheritance; extract the previous control cycle. Output of the globally optimal residual torque sequence The current time step is generated by the left shift operator. High confidence seeds : ， In the formula, This is the optimal solution vector after optimization at the previous time step; 0 represents the optimal solution vector. The subsequent control variables are assumed to have zero null. The first particle is directly assigned a value. The remaining particles are The mean center is used to generate the data according to a Gaussian distribution within the search boundary. Step (43): Construct the total torque: Call the network trained in step (3) to predict the future state tensor And calculate the fitness function. ; Computational tracking accuracy versus energy consumption cost: ， In the formula, This refers to the angle error; This refers to the yaw error; This refers to the centroid lateral deviation error. This represents the term for minimizing the residual torque and the constraint on control energy consumption; Define a stability index based on a rhombus envelope. With phase plane stability penalty function : ， ， In the formula, A high penalty factor; , For dynamic coefficients, take respectively , and ; Define the saturation penalty function for the executor: ， In the formula, For candidate total command torque; This represents the maximum physical torque of the steering motor; As a penalty factor; Calculate the fitness function ; Step (44): Perform evolutionary update under the dynamic trust region, and perform projection operation on the updated particle positions: ， ， In the formula, This refers to the residual torque after the update. This is a transient position; For projection operators; Step (45): After the iteration terminates, output the first element of the optimal sequence as the optimal residual compensation torque. .

7. The method according to claim 6, characterized in that, Step (5) specifically includes: Step (51): Construct the inverse dynamics observation model of the steering system based on the dynamic model of the feedforward compensation module for the steering effect, and extract the actual disturbance observation values. : ， Step (52): Extract the real interference from the observer Compared with the reference torque calculated by the current feedforward model Obtain the prediction error of the feedforward model : ， Step (53): Based on the current error The recursive least squares strategy is used to correct the weight coefficient vector of the physical characteristic equation of the feedforward compensation module for the inversion effect in real time. ， In the formula, This is the gain matrix; Step (54): Weight coefficient vector The input is sent to the feedforward compensation module and takes effect in the next control cycle.

Images