Motion control method and system for distributed ultra-flat motion control platform

Abstract

The application provides a motion control method and system for a distributed super-flat motion control platform, and belongs to the technical field of automatic driving. The application analyzes the wheel-ground interaction relationship of the distributed super-flat motion control platform, constructs an integrated model of the wheel, the whole control platform and the electric drive module, designs a method of integrating the road adhesion coefficient estimation and its uncertainty into the path tracking control law, analyzes the high-speed stability influence mechanism of the motion platform control and constructs a robust stability control strategy, and can effectively improve the robustness and tracking accuracy of the distributed super-flat platform motion controller.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically relating to a motion control method and system for a distributed ultra-flat motion control platform. Background Technology

[0002] As an important direction for the future development of intelligent electric vehicles, the distributed ultra-flat motion control platform has technical advantages such as low center of gravity, strong handling, and highly controllable electric drive. Compared with traditional centralized power platforms, its wheel-side drive architecture provides more space for maximizing vehicle adhesion utilization and high dynamic response control.

[0003] However, achieving precise control under extreme conditions remains a significant challenge for this platform. Traditional filtering estimation methods often rely on sufficient time-domain signal excitation, making them unsuitable for conditions with weak motion excitation. Furthermore, many current studies have not adequately assessed the error range and distribution characteristics of the estimation results, leading to insufficient robustness of the path tracking controller, especially when the adhesion coefficient undergoes abrupt or rapid changes. In addition, measurement noise inevitably causes errors in the adhesion coefficient estimation, increasing the uncertainty of the system model and controller constraints. Both model inaccuracies and conservative constraints negatively impact vehicle control performance. Therefore, conducting research on adaptive motion control for road adhesion on distributed ultra-flat platforms, analyzing the wheel-ground interaction of small-diameter wheels, constructing an integrated model of wheels, vehicle body, and electric drive modules, designing a method to integrate adhesion coefficient estimation and its uncertainties into the path tracking control law, analyzing the mechanism affecting the high-speed stability of the motion platform control, and constructing a robust and stable control strategy are crucial for improving the robustness and tracking accuracy of the distributed ultra-flat platform motion controller.

[0004] Therefore, it is necessary to provide a motion control method and system for a distributed ultra-flat motion control platform to solve the above problems. Summary of the Invention

[0005] This invention provides a motion control method and system for a distributed ultra-flat motion control platform. Addressing the influence of uncertain road surfaces on vehicle dynamics models and probabilistic state constraints, a learning-based stochastic model predictive control path tracking method is designed to resolve the uncertainty in road surface adhesion coefficient estimation during predictive model and chance constraint design. This solves at least one of the technical problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the present invention is implemented as follows: A motion control method for a distributed ultra-flat motion control platform includes the following steps: Step S1: Based on the dynamic model of the distributed ultra-flat motion control platform and the dynamic model of the wheel, derive the expression function of the road surface adhesion coefficient, and determine the state parameters related to the road surface adhesion coefficient under the condition that the expression function has a unique real solution. Step S2: Construct a spatiotemporal convolutional network. A set of data is formed by different combinations of multiple state parameters. The road surface adhesion coefficient under this set of data is used as the label. Multiple sets of data and corresponding labels are combined to form a dataset. The dataset is input into the spatiotemporal convolutional network for learning. High-dimensional features are extracted from the time domain signal and the time-frequency domain signal respectively. The output high-dimensional feature vector is used as the input vector of the sparse variational Gaussian process. The mapping relationship between the state parameters and the road surface adhesion coefficient is learned by using the sparse variational strategy. After training is completed, the state parameters are combined and input into the spatiotemporal convolutional network to output the estimated value of the road surface adhesion coefficient and the estimated variance. Step S3: Based on the estimated road surface adhesion coefficient and its variance, design an adaptive path tracking controller to regulate the motion state of the distributed ultra-flat motion control platform.

[0007] As a preferred improvement, the dynamic model of the seven-degree-of-freedom distributed ultra-flat motion control platform is expressed as follows: In the formula, Indicates quality; Indicates longitudinal velocity; Indicates longitudinal acceleration; , These represent the longitudinal forces on the left and right front wheels, respectively. Indicates the front wheel steering angle; , These represent the lateral forces on the left and right front wheels, respectively. , These represent the longitudinal forces on the left and right rear wheels, respectively. Indicates lateral velocity. Indicates yaw rate; , , These are air resistance, wheel rolling resistance, and gradient resistance, respectively. Indicates lateral acceleration; , These represent the lateral forces on the left and right rear wheels, respectively. Indicates yaw acceleration; , These represent the distances from the front and rear axles to the center of mass, respectively. Indicates the wheelbase (left and right); This indicates the yaw moment.

[0008] As a preferred improvement, the wheel dynamics model is expressed as: In the formula, Indicates the moment of inertia of the wheel. This indicates the angular velocity of the wheel's rotation; This represents the angular acceleration of the wheel's rotation; This indicates the wheel drive torque output by the hub motor drive system; Indicates the wheel radius; Wheel drive torque Composed of high-frequency and low-frequency signals, it is represented as: In the formula, , These represent low-frequency signals and high-frequency signals, respectively. Indicates the amplitude of the signal; Indicates the frequency of the signal. Indicates time; Therefore, the wheel dynamics model in driving mode is represented as: In the formula, Indicates the magnitude of the torque; express ; Represents the imaginary unit; Indicates the relaxation coefficient; Indicates the longitudinal stiffness of the tire; The wheel dynamics model in braking mode can be represented as: As a preferred improvement, under unsteady driving conditions, the coefficient of adhesion to the road surface... Closely related parameters include: mass Wheel slip ratio Longitudinal wheel stiffness Longitudinal acceleration Vertical force of wheels , ,in This indicates the longitudinal force on the wheel; coefficient of adhesion to road surface under steady-state driving conditions Closely related parameters include: longitudinal velocity Wheel rotational angular velocity Vertical force of wheels and .

[0009] As a preferred improvement, the spatiotemporal convolutional network includes a non-steady-state maneuvering spatiotemporal convolutional sub-network and a steady-state maneuvering spatiotemporal convolutional sub-network. The non-steady-state maneuvering spatiotemporal convolutional sub-network takes the state parameters and labels under the non-steady-state maneuvering condition as input, while the steady-state maneuvering spatiotemporal convolutional sub-network takes the state parameters and labels under the steady-state maneuvering condition as input.

[0010] As a preferred improvement, the non-steady-state maneuvering spatiotemporal convolutional subnetwork includes three convolutional units connected in sequence. Each convolutional unit includes a convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence. The kernel sizes of the three convolutional layers are 1×4×8, 1×4×16, and 1×4×32, respectively. The steady-state maneuvering spatiotemporal convolutional subnetwork includes a time-domain feature extraction branch and a time-frequency domain feature extraction branch set in parallel. The time-domain feature extraction branch is used to directly process training data segments and extract time-domain features through convolution operations. The time-domain feature extraction branch includes three convolutional units connected in sequence. Each convolutional unit includes a convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence. The activation layer consists of three convolutional layers with kernel sizes of 1×4×8, 1×4×16, and 1×4×32, respectively. The time-frequency domain feature extraction branch obtains two-dimensional time-frequency domain features through continuous wavelet transform. The time-frequency domain feature extraction branch includes four convolutional units connected in sequence. Each convolutional unit includes a convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence. The kernel sizes of the three convolutional layers are 4×2×4, 3×4×8, 2×2×16, and 2×2×32, respectively. The time-domain features and the time-frequency domain features are fused together in the Flatten layer to form the input of the sparse variational Gaussian process.

[0011] As a preferred improvement, the sparse variational Gaussian process specifically includes the following steps: Step S21: Select representative sample points from the input vector as induction points, and regard the function values ​​at the induction points as a set of induction variables, so as to represent the key information of the entire input vector. Step S22: In variational inference, a variational distribution is introduced to approximate the true posterior distribution; Step S23: With the goal of maximizing the lower bound of the marginal log-likelihood as the optimization objective, the standard stochastic gradient descent method is used for optimization to make the variational distribution as close as possible to the true posterior distribution.

[0012] As a preferred improvement, the lateral dynamic equation of the distributed ultra-flat motion control platform is: In the formula, This represents the total mass of the distributed ultra-flat motion control platform; Represents longitudinal velocity; Indicates lateral velocity; The first derivative of the centroid sideslip angle; Indicates yaw rate; , These represent the lateral forces on the front and rear wheels, respectively. This represents the moment of inertia of the entire vehicle about the z-axis; , These represent the distances from the front axle and rear axle to the center of mass, respectively. The wheel model is represented as: In the formula, , These represent the lateral affine forces of the front and rear wheels in the previous control cycle; , These represent the equivalent lateral stiffness of the front and rear wheels, respectively. This represents the estimated value of the road surface adhesion coefficient; , These represent the front and rear wheel slip angles at the current moment; , These represent the front and rear wheel slip angles respectively in the previous control cycle; Substituting the wheel model into the lateral dynamics equation, the lateral dynamics equation can be further expressed as: ; In the formula, , , , , , , , All represent intermediate calculation parameters; where: ; ; ; ; ; ; ; ; The differential equation of the path tracking model is expressed as: ; In the formula, This indicates the state of the distributed ultra-flat motion control platform. , Indicates lateral error; Indicates the heading angle error; Represents the transpose matrix; This represents the control input of the distributed ultra-flat motion control platform. , Indicates system disturbance; , , Both represent weight matrices, where: ; ; ; In the formula, Path curvature; Discretize the differential equations of the path tracking model as follows: ; In the formula, , They represent and The status of the distributed ultra-flat motion control platform at all times; express Control input for the real-time distributed ultra-flat motion control platform; express System disturbances at any given time; , , express The weight matrix at time step; where: In the formula, Represents the identity matrix; Indicates the sampling time.

[0013] As a preferred improvement, the control prediction model of the adaptive path tracking controller is expressed as: ; st ; ; ; ; In the formula, , , Represents the weight matrix; Represents the state matrix; Represents the reference state matrix; Indicates the increment of the control quantity; This represents the terminal state matrix; k represents the time node; N represents the total number of time nodes; ; , , , , These represent the maximum lateral error, the maximum heading angle error, the maximum yaw rate, and the maximum sideslip angle of the center of mass, respectively. Represents a probability function; Indicates a specific numerical value for the probability; , These represent the front wheel steering angles. Minimum and maximum limits; , These represent the front wheel steering angle increments. Minimum and maximum limits; Solve the constructed control prediction model to obtain the optimal control sequence. Take the first control variable of the optimal control sequence as the current expected front wheel steering angle and send it to the underlying steering actuator to realize the path tracking control of the distributed ultra-flat motion control platform.

[0014] A system for executing the above-described motion control method for a distributed ultra-flat motion control platform includes: The parameter determination module is used to derive the expression function of the road surface adhesion coefficient based on the dynamic model of the distributed ultra-flat motion control platform and the dynamic model of the wheel, and determine the state parameters related to the road surface adhesion coefficient under the condition that the expression function has a unique real solution. Spatiotemporal convolutional networks are used to form a set of data with different combinations of multiple state parameters. The road surface adhesion coefficient under this set of data is used as the label. Multiple sets of data and their corresponding labels are combined to form a dataset. The dataset is input into the spatiotemporal convolutional network for learning. High-dimensional features are extracted from the time domain signal and the time-frequency domain signal respectively. The output high-dimensional feature vector is used as the input vector of a sparse variational Gaussian process. The sparse variational strategy is used to learn the mapping relationship between the state parameters and the road surface adhesion coefficient. After training, the state parameters are combined and input into the spatiotemporal convolutional network to output the estimated value of the road surface adhesion coefficient and the estimated variance. The controller design module is used to design an adaptive path tracking controller based on the estimated road surface adhesion coefficient and its variance, thereby regulating the motion state of the distributed ultra-flat motion control platform.

[0015] The beneficial effects of this invention are as follows: (1) By using the random variational deep kernel learning estimation strategy, the road surface adhesion coefficient under the complex motion excitation condition including the weak motion excitation condition is estimated, and the uncertainty assessment of the road surface adhesion coefficient estimation is realized. (2) The input parameter types of the stochastic variational deep kernel learning network were determined based on the different states of unsteady-state maneuvering conditions and steady-state maneuvering conditions in the time and frequency domains. (3) A linear path tracking method based on affine force input model. In view of the influence of uncertain road surface on vehicle dynamics model and probabilistic state constraints, a learning-type stochastic model predictive control path tracking method is designed to solve the influence of uncertainty in road surface adhesion coefficient estimation in prediction model and chance constraint design. (4) By analyzing the wheel-ground interaction relationship of the distributed ultra-flat motion control platform, an integrated model of the wheel, control platform and electric drive modules is constructed. A method is designed to integrate the road surface adhesion coefficient estimation and its uncertainty into the path tracking control law. The mechanism of the influence of high-speed stability of the motion platform control is analyzed and its robust and stable control strategy is constructed. This can effectively improve the robustness and tracking accuracy of the distributed ultra-flat platform motion controller. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart illustrating the motion control method for a distributed ultra-flat motion control platform provided by the present invention. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This embodiment provides a motion control method for a distributed ultra-flat motion control platform, including the following steps: Step S1: Based on the dynamic model of the distributed ultra-flat motion control platform and the dynamic model of the wheel, derive the expression function of the road surface adhesion coefficient. Determine the state parameters related to the road surface adhesion coefficient, assuming that the expression function has a unique real solution.

[0019] The dynamic model of the seven-degree-of-freedom distributed ultraflat motion control platform is expressed as follows: In the formula, Indicates quality; Indicates longitudinal velocity; Indicates longitudinal acceleration; , These represent the longitudinal forces on the left and right front wheels, respectively. Indicates the front wheel steering angle; , These represent the lateral forces on the left and right front wheels, respectively. , These represent the longitudinal forces on the left and right rear wheels, respectively. Indicates lateral velocity. Indicates yaw rate; , , These are air resistance, wheel rolling resistance, and gradient resistance, respectively. Indicates lateral acceleration; , These represent the lateral forces on the left and right rear wheels, respectively. Indicates yaw acceleration; , These represent the distances from the front and rear axles to the center of mass, respectively. Indicates the wheelbase (left and right); This indicates the yaw moment.

[0020] Based on the seven-degree-of-freedom distributed ultra-flat motion control platform model, we can obtain: In the formula, , These represent the longitudinal wheel stiffness and the lateral wheel stiffness, respectively. , , , These represent the slip ratios of the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively. Indicates the front wheel steering angle; , These represent the wheel slip angles of the left and right front wheels, respectively. , , , These represent the status judgment parameters for the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively. This indicates that the wheel is in the linear zone; This indicates that the wheel is in the nonlinear region. ; The intermediate parameter is represented as: In the formula, Indicates the road surface adhesion coefficient; This represents the vertical force on the wheel; This indicates the slip ratio of the wheel; Indicates the wheel slip angle; The above formula can be expressed as the coefficient of adhesion to the road surface. Related quadratic functions , is represented as: quadratic function The purpose of the derivation is mainly to establish the relationship between state parameters and road adhesion coefficient. The relationship between them, and in the subsequent learning process of the neural network model, usually only the state parameters and the road adhesion coefficient are considered. The relative relationship between them, rather than depending on the absolute value, is used in this application to treat quadratic functions. constant term in The method of removal for quadratic functions Simplifying the process will not affect the feature learning and training performance of the neural network model. This simplification method can reduce the amount of data processing required for the neural network model and improve training efficiency. The simplified quadratic function... Represented as: quadratic function There are two possible solutions: one is that there are no real roots, and the other is that there is a unique real root. Since the distributed ultra-flat motion control platform is a cyber-physical system, the road surface adhesion coefficient... And all state parameters have actual physical meaning, therefore A unique real solution must exist. Based on this property, the quadratic function can be analyzed... The slope of the curve is used to indirectly determine the road surface adhesion coefficient. .

[0021] The distributed ultra-flat motion control platform includes steady-state and non-steady-state maneuvering conditions. The steady-state maneuvering condition refers to the longitudinal acceleration of the distributed ultra-flat motion control platform. The longitudinal driving conditions are defined as follows: the non-steady-state driving maneuver conditions are defined as all longitudinal driving conditions other than the steady-state maneuver conditions.

[0022] Under unsteady driving conditions, the coefficient of adhesion with the road surface The closely related parameter is quality. Wheel slip ratio Longitudinal wheel stiffness Longitudinal acceleration Vertical force of wheels , .

[0023] Under steady-state maneuvering conditions, all wheels operate within a linear range. The saturation value of wheel friction is positively correlated with the slope of the normalized wheel force-slip ratio curve, which reflects the dynamic response characteristics of the wheel under different road surface adhesion conditions. Different road surface adhesion conditions alter the slope of the normalized wheel force-slip ratio curve. When the wheel slip ratio is within the linear range, the wheel force and slip ratio exhibit a linear relationship. Therefore, by analyzing the slope of the linear segment of the normalized wheel force-slip ratio curve, the relationship with the road surface adhesion coefficient can be extracted. Relevant feature information.

[0024] The wheel dynamics model is represented as follows: In the formula, Indicates the moment of inertia of the wheel. This indicates the angular velocity of the wheel's rotation; This represents the angular acceleration of the wheel's rotation; This indicates the wheel drive torque output by the hub motor drive system; Indicates the radius of the wheel.

[0025] Wheel drive torque Composed of high-frequency and low-frequency signals, it is represented as: In the formula, , These represent low-frequency signals and high-frequency signals, respectively. Indicates the amplitude of the signal; Indicates the frequency of the signal. Indicates time.

[0026] Therefore, the wheel dynamics model in driving mode is represented as: In the formula, Indicates the magnitude of the torque; express ; Represents the imaginary unit; Indicates the relaxation coefficient; Indicates the longitudinal stiffness of the tire; The wheel dynamics model in braking mode can be represented as: Analysis of the wheel dynamics models under driving and braking modes revealed different mapping relationships between the state parameters in the amplitude-frequency function and the slope of the linear portion of the normalized wheel force-slip ratio curve. Therefore, the time-varying characteristic parameters involved in the amplitude-frequency function are related to the road adhesion coefficient. The relevant state parameters include: longitudinal velocity Wheel rotational angular velocity Vertical force of wheels and .

[0027] Based on the above analysis, the road surface adhesion coefficient It can be represented as: In the formula, Indicates parameters With adhesion coefficient The mapping function between them.

[0028] Step S2: Construct a spatiotemporal convolutional network. A set of data is formed by different combinations of multiple state parameters. The road surface adhesion coefficient under this set of data is used as the label. Multiple sets of data and corresponding labels are combined to form a dataset. The dataset is input into the spatiotemporal convolutional network for learning. High-dimensional features are extracted from the time domain signal and the time-frequency domain signal respectively. The output high-dimensional feature vector is used as the input vector of the sparse variational Gaussian process. The mapping relationship between the state parameters and the road surface adhesion coefficient is learned by using the sparse variational strategy. After training, the state parameters are combined and input into the spatiotemporal convolutional network to output the estimated value of the road surface adhesion coefficient and the estimated variance.

[0029] Spatiotemporal convolutional networks include non-steady-state maneuvering spatiotemporal convolutional subnetworks and steady-state maneuvering spatiotemporal convolutional subnetworks. The non-steady-state maneuvering spatiotemporal convolutional subnetwork takes the state parameters and labels under non-steady-state maneuvering conditions as input, while the steady-state maneuvering spatiotemporal convolutional subnetwork takes the state parameters and labels under steady-state maneuvering conditions as input.

[0030] The non-steady-state maneuvering spatiotemporal convolutional subnetwork comprises three convolutional units connected in sequence. Each convolutional unit includes a convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence. The kernel sizes of the three convolutional layers are 1×4×8, 1×4×16, and 1×4×32, respectively. The steady-state maneuvering spatiotemporal convolutional subnetwork includes a parallel temporal feature extraction branch and a time-frequency domain feature extraction branch. The temporal feature extraction branch directly processes training data segments, extracting temporal features through convolutional operations. This branch comprises three convolutional units connected in sequence, each including a convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence. The kernel sizes of the three convolutional layers are 1×4×8, 1×4×16, and 1×4×32, respectively. The time-frequency domain feature extraction branch obtains two-dimensional time-frequency domain features through continuous wavelet transform. This branch comprises four convolutional units connected in sequence, each including a convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence. The activation layer consists of three convolutional layers with kernel sizes of 4×2×4, 3×4×8, 2×2×16, and 2×2×32, respectively.

[0031] The time-domain features and time-frequency domain features are fused together in the Flatten layer to form the input of a sparse variational Gaussian process, which specifically includes the following steps: Step S21: Select representative sample points from the input vector as induction points, and regard the function values ​​at the induction points as a set of induction variables, so as to represent the key information of the entire input vector. Step S22: In variational inference, a variational distribution is introduced to approximate the true posterior distribution; Step S23: With the goal of maximizing the lower bound of the marginal log-likelihood as the optimization objective, the standard stochastic gradient descent method is used for optimization to make the variational distribution as close as possible to the true posterior distribution.

[0032] After training, output the estimated value of the road surface adhesion coefficient. And the estimation of uncertainty, i.e., the estimation of variance. Estimated value With Estimated Variance Design of an adaptive path tracking controller for a distributed ultra-flat motion control platform.

[0033] Step S3: Considering the uncertainties in the adhesion coefficient estimation process, design an adaptive path tracking controller to regulate the motion state of the distributed ultra-flat motion control platform.

[0034] The lateral dynamic equation of the distributed ultra-flat motion control platform is: In the formula, This represents the total mass of the distributed ultra-flat motion control platform; Represents longitudinal velocity; Indicates lateral velocity; The first derivative of the centroid sideslip angle; Indicates yaw rate; , These represent the lateral forces on the front and rear wheels, respectively. This represents the moment of inertia of the entire vehicle about the z-axis; , These represent the distances from the front axle and rear axle to the center of mass, respectively.

[0035] The wheel model is represented as: In the formula, , These represent the lateral affine forces of the front and rear wheels in the previous control cycle; , These represent the equivalent lateral stiffness of the front and rear wheels, respectively. This represents the estimated value of the road surface adhesion coefficient; , These represent the front and rear wheel slip angles at the current moment; , These represent the front and rear wheel slip angles of the previous control cycle, respectively.

[0036] Substituting the wheel model into the lateral dynamics equation, the lateral dynamics equation can be further expressed as: ; In the formula, , , , , , , , All represent intermediate calculation parameters; where: ; ; ; ; ; ; ; .

[0037] The differential equation of the path tracking model is expressed as: ; In the formula, This indicates the state of the distributed ultra-flat motion control platform. , Indicates lateral error; Indicates the heading angle error; Represents the transpose matrix; This represents the control input of the distributed ultra-flat motion control platform. , Indicates system disturbance; , , Both represent weight matrices, where: ; ; ; In the formula, Path curvature.

[0038] Discretize the differential equations of the path tracking model as follows: ; In the formula, , They represent and The status of the distributed ultra-flat motion control platform at all times; express Control input for the real-time distributed ultra-flat motion control platform; express System disturbances at any given time; , , express The weight matrix at time step; where: In the formula, Represents the identity matrix; Indicates the sampling time.

[0039] The motion of the distributed ultra-flat motion control platform needs to satisfy the following constraints: (1) The lateral stability constraint is expressed as: ; In the formula, ; , , , , These represent the maximum lateral error, the maximum heading angle error, the maximum yaw rate, and the maximum sideslip angle of the center of mass, respectively.

[0040] (2) The opportunity constraint is expressed as: ; In the formula, Represents a probability function; Indicates a specific numerical value for the probability; (3) The physical constraints of the actuator are expressed as: ; In the formula, , These represent the front wheel steering angles. Minimum and maximum limits; , These represent the front wheel steering angle increments. The minimum and maximum limits.

[0041] Based on the path tracking model and constraints described above, a control prediction model is constructed, which is expressed as: ; st ; ; ; ; In the formula, , , Represents the weight matrix; Represents the state matrix; Represents the reference state matrix; Indicates the increment of the control quantity; This represents the terminal state matrix; k represents the time node; and N represents the total number of time nodes.

[0042] Solve the constructed control prediction model to obtain the optimal control sequence. Take the first control variable of the optimal control sequence as the current expected front wheel steering angle and send it to the underlying steering actuator to realize the path tracking control of the distributed ultra-flat motion control platform.

[0043] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A motion control method for a distributed ultra-flat motion control platform, characterized in that, Includes the following steps: Step S1: Based on the dynamic model of the distributed ultra-flat motion control platform and the dynamic model of the wheel, derive the expression function of the road surface adhesion coefficient, and determine the state parameters related to the road surface adhesion coefficient under the condition that the expression function has a unique real solution. Step S2: Construct a spatiotemporal convolutional network. A set of data is formed by different combinations of multiple state parameters. The road surface adhesion coefficient under this set of data is used as the label. Multiple sets of data and corresponding labels are combined to form a dataset. The dataset is input into the spatiotemporal convolutional network for learning. High-dimensional features are extracted from the time domain signal and the time-frequency domain signal respectively. The output high-dimensional feature vector is used as the input vector of the sparse variational Gaussian process. The mapping relationship between the state parameters and the road surface adhesion coefficient is learned by using the sparse variational strategy. After training is completed, the state parameters are combined and input into the spatiotemporal convolutional network to output the estimated value of the road surface adhesion coefficient and the estimated variance. Step S3: Based on the estimated road surface adhesion coefficient and its variance, design an adaptive path tracking controller to regulate the motion state of the distributed ultra-flat motion control platform. The wheel dynamics model is represented as follows: In the formula, Indicates the moment of inertia of the wheel. This indicates the angular velocity of the wheel's rotation; This represents the angular acceleration of the wheel's rotation; This indicates the wheel drive torque output by the hub motor drive system; Indicates the wheel radius; This indicates the longitudinal force on the wheel; Wheel drive torque Composed of high-frequency and low-frequency signals, it is represented as: In the formula, , These represent low-frequency signals and high-frequency signals, respectively. Indicates the amplitude of the signal; Indicates the frequency of the signal. Indicates time; Therefore, the wheel dynamics model in driving mode is represented as: In the formula, Indicates the magnitude of the torque; express ; Represents the imaginary unit; Indicates the relaxation coefficient; Indicates the longitudinal stiffness of the tire; This represents the vertical force on the wheel; The wheel dynamics model in braking mode can be represented as: Under unsteady driving conditions, the coefficient of adhesion with the road surface Closely related parameters include: mass Wheel slip ratio Longitudinal wheel stiffness Longitudinal acceleration Vertical force of wheels , ; coefficient of adhesion to road surface under steady-state driving conditions Closely related parameters include: longitudinal velocity Wheel rotational angular velocity Vertical force of wheels and ; The spatiotemporal convolutional network includes an unsteady maneuvering spatiotemporal convolutional subnetwork and a steady-state maneuvering spatiotemporal convolutional subnetwork. The unsteady maneuvering spatiotemporal convolutional subnetwork takes the state parameters and labels under the unsteady maneuvering condition as input, while the steady-state maneuvering spatiotemporal convolutional network takes the state parameters and labels under the steady-state maneuvering condition as input. The sparse variational Gaussian process specifically includes the following steps: Step S21: Select representative sample points from the input vector as induction points, and regard the function values ​​at the induction points as a set of induction variables, so as to represent the key information of the entire input vector. Step S22: In variational inference, a variational distribution is introduced to approximate the true posterior distribution; Step S23: With the goal of maximizing the lower bound of the marginal log-likelihood as the optimization objective, the standard stochastic gradient descent method is used for optimization to make the variational distribution as close as possible to the true posterior distribution.

2. The motion control method for a distributed ultra-flat motion control platform according to claim 1, characterized in that, The dynamic model of the seven-degree-of-freedom distributed ultraflat motion control platform is expressed as follows: In the formula, Indicates quality; Indicates longitudinal velocity; Indicates longitudinal acceleration; , These represent the longitudinal forces on the left and right front wheels, respectively. Indicates the front wheel steering angle; , These represent the lateral forces on the left and right front wheels, respectively. , These represent the longitudinal forces on the left and right rear wheels, respectively. Indicates lateral velocity. Indicates yaw rate; , , These are air resistance, wheel rolling resistance, and gradient resistance, respectively. Indicates lateral acceleration; , These represent the lateral forces on the left and right rear wheels, respectively. Indicates yaw acceleration; , These represent the distances from the front and rear axles to the center of mass, respectively. Indicates the wheelbase (left and right); Indicates yaw moment; This represents the moment of inertia of the entire vehicle about the z-axis.

3. The motion control method for a distributed ultra-flat motion control platform according to claim 1, characterized in that, The non-steady-state maneuvering spatiotemporal convolutional subnetwork comprises three sequentially connected convolutional units. Each convolutional unit includes a sequentially connected convolutional layer, a batch normalization layer, and a ReLU activation layer. The kernel sizes of the three convolutional layers are 1×4×8, 1×4×16, and 1×4×32, respectively. The steady-state maneuvering spatiotemporal convolutional subnetwork includes parallel temporal feature extraction branches and time-frequency domain feature extraction branches. The temporal feature extraction branch directly processes training data segments, extracting temporal features through convolutional operations. This branch comprises three sequentially connected convolutional units, each including a sequentially connected convolutional layer, a batch normalization layer, and a ReLU activation layer. The kernel sizes of the three convolutional layers are 1×4×8, 1×4×16, and 1×4×32, respectively. The time-frequency domain feature extraction branch obtains two-dimensional time-frequency domain features through continuous wavelet transform. This branch comprises four sequentially connected convolutional units, each including a sequentially connected convolutional layer, a batch normalization layer, and a ReLU activation layer. The activation layer consists of four convolutional layers with kernel sizes of 4×2×4, 3×4×8, 2×2×16, and 2×2×32, respectively. Temporal and temporal-frequency features are fused together in the Flatten layer to form the input of a sparse variational Gaussian process.

4. The motion control method for a distributed ultra-flat motion control platform according to claim 1, characterized in that, The lateral dynamic equation of the distributed ultra-flat motion control platform is: In the formula, This represents the total mass of the distributed ultra-flat motion control platform; Represents longitudinal velocity; Indicates lateral velocity; The first derivative of the centroid sideslip angle; Indicates yaw rate; , These represent the lateral forces on the front and rear wheels, respectively. This represents the moment of inertia of the entire vehicle about the z-axis; , These represent the distances from the front axle and rear axle to the center of mass, respectively. The wheel model is represented as: In the formula, , These represent the lateral affine forces of the front and rear wheels in the previous control cycle; , These represent the equivalent lateral stiffness of the front and rear wheels, respectively. This represents the estimated value of the road surface adhesion coefficient; , These represent the front and rear wheel slip angles at the current moment; , These represent the front and rear wheel slip angles respectively in the previous control cycle; Substituting the wheel model into the lateral dynamics equation, the lateral dynamics equation can be further expressed as: ； In the formula, , , , , , , , All of these represent intermediate calculation parameters; Indicates the front wheel steering angle; where: ； ； ； ； ； ； ； ； The differential equation of the path tracking model is expressed as: ； In the formula, This indicates the state of the distributed ultra-flat motion control platform. , Indicates lateral error; Indicates the heading angle error; Represents the transpose matrix; This represents the control input of the distributed ultra-flat motion control platform. , Indicates system disturbance; , , Both represent weight matrices, where: ； ； ； In the formula, Path curvature; Discretize the differential equations of the path tracking model as follows: ； In the formula, , They represent and The status of the distributed ultra-flat motion control platform at all times; express Control input of the real-time distributed ultra-flat motion control platform; express System disturbances at any given time; , , express The weight matrix at time step; where: In the formula, Represents the identity matrix; Indicates the sampling time.

5. The motion control method for a distributed ultra-flat motion control platform according to claim 1, characterized in that, The control prediction model of the adaptive path tracking controller is expressed as follows: ； s.t ; ； ； ； In the formula, , , Represents the weight matrix; Represents the state matrix; Represents the reference state matrix; Indicates the increment of the control quantity; This represents the terminal state matrix; k represents the time node; n represents the total number of time nodes; ; , , , , These represent the maximum lateral error, the maximum heading angle error, the maximum yaw rate, and the maximum sideslip angle of the center of mass, respectively. Represents a probability function; Indicates a specific numerical value for the probability; , These represent the front wheel steering angles. Minimum and maximum limits; , These represent the front wheel steering angle increments. Minimum and maximum limits; Solve the constructed control prediction model to obtain the optimal control sequence. Take the first control variable of the optimal control sequence as the current expected front wheel steering angle and send it to the underlying steering actuator to realize the path tracking control of the distributed ultra-flat motion control platform.

6. A system for executing the motion control method for a distributed ultra-flat motion control platform according to any one of claims 1-5, characterized in that, include: The parameter determination module is used to derive the expression function of the road surface adhesion coefficient based on the dynamic model of the distributed ultra-flat motion control platform and the dynamic model of the wheel, and determine the state parameters related to the road surface adhesion coefficient under the condition that the expression function has a unique real solution. Spatiotemporal convolutional networks are used to form a set of data with different combinations of multiple state parameters. The road surface adhesion coefficient under this set of data is used as the label. Multiple sets of data and their corresponding labels are combined to form a dataset. The dataset is input into the spatiotemporal convolutional network for learning. High-dimensional features are extracted from the time domain signal and the time-frequency domain signal respectively. The output high-dimensional feature vector is used as the input vector of a sparse variational Gaussian process. The sparse variational strategy is used to learn the mapping relationship between the state parameters and the road surface adhesion coefficient. After training, the state parameters are combined and input into the spatiotemporal convolutional network to output the estimated value of the road surface adhesion coefficient and the estimated variance. The controller design module is used to design an adaptive path tracking controller based on the estimated road surface adhesion coefficient and its variance, thereby regulating the motion state of the distributed ultra-flat motion control platform.

Images