An adaptive collaborative control system and method for commercial vehicle semi-automatic suspension based on data-driven model

Abstract

The present invention discloses an adaptive collaborative control system and method for a commercial vehicle semi-automatic suspension based on a data-driven model. The physical model is combined with a neural network to construct a Lagrangian mechanical model of the vehicle's pitching motion, effectively improving the accuracy of the modeling. The semi-active suspension is mainly controlled by adjusting a limited damping change, so an accurate damping coefficient will help improve the control effect. At the same time, due to the fast computing speed of the neural network, the response speed of the suspension system is improved, and the pitch and vertical motion control effects of the vehicle are improved. The desired control force and torque are solved by the upper-level roll controller, pitch and vertical controller, and the force of each suspension is solved by the lower-level optimal distribution controller, so that collaborative control between the four suspensions is achieved, effectively suppressing the vehicle's roll, improving the vehicle's stability and safety, and effectively suppressing the vehicle's pitching motion, improving comfort. The adaptive weight matrix is ​​constructed to achieve the adaptability of the suspension under different working conditions.

Description

Technical Field

[0001] The present invention relates to a commercial vehicle intelligent chassis control technology, and in particular to a commercial vehicle semi-automatic suspension adaptive collaborative control system and method based on a data-driven model. Background Art

[0002] With the continuous development of autonomous driving technology, especially the increasingly mature application and implementation of autonomous driving in commercial vehicles, the handling stability and comfort of commercial vehicles are lower than those of passenger cars due to their heavy load, high center of mass and slow response speed. Therefore, the research on multi-objective intelligent collaborative control of commercial vehicle stability, safety and comfort is particularly important.

[0003] Autonomous driving for commercial vehicles has gradually been implemented in relatively simple or closed environments, such as highways, ports, industrial parks, and mining areas. However, as commercial vehicles' cargo capacity and transport speeds increase, load transfer in complex and extreme operating conditions, such as cornering and emergency obstacle avoidance, will increase significantly, significantly increasing the risk of rollover. Traditional, simplified vehicle-based monorail bicycle models fail to consider factors such as load transfer between the left and right wheels and the roll of the vehicle body. Therefore, applying this simplified model-based control method to commercial vehicles will reduce control effectiveness. Furthermore, during emergency braking, commercial vehicles' high center of mass will increase their pitch motion, hindering cargo transportation and worsening passenger comfort. Therefore, suppressing pitch motion in commercial vehicles will significantly improve vehicle stability and ride comfort.

[0004] Semi-active suspensions are widely used in passenger cars and commercial vehicles due to their simple structure and low cost. However, most current semi-active suspensions are based on a simplified quarter-vehicle model, establishing the relationship between unsprung mass, suspension, and sprung mass to reduce sprung mass acceleration, suspension travel, and tire dynamic loads. Independent control of the four suspensions in this manner cannot coordinate with the other suspensions to control vehicle roll and pitch, nor can the control strategy be dynamically adjusted based on the vehicle's driving state. This is detrimental to the safety and comfort of commercial vehicles and does not fully utilize the capabilities of semi-active suspension for coordinated vehicle control.

[0005] On the other hand, traditional vehicle modeling techniques, in order to simplify the model and reduce computational complexity, often simplify the model or make it difficult to capture the system's variable parameters. They often treat parameters as constant, such as the damping and stiffness of the suspension system. In particular, damping varies under different compression displacements and compression speeds, which is a problem that traditional modeling methods find difficult to address. With the rapid development and widespread application of deep learning, its excellent nonlinear fitting capabilities and extremely high computational speed have shown significant advantages in modeling. However, the lack of physical interpretability of neural networks can reduce the reliability of the model. Therefore, combining neural networks with physical models can effectively improve modeling accuracy, stability, and generalization capabilities. Summary of the Invention

[0006] To address the above problems, the present invention proposes an adaptive collaborative control system and method for a commercial vehicle semi-automatic suspension based on a data-driven model. The system mainly includes a part that constructs a dual-track vehicle roll motion model based on the Newton-Euler method, a part that constructs a vehicle pitch and vertical motion model based on Lagrangian mechanics, a part that constructs a pitch and vertical motion model based on the fusion of neural networks and physical models, an upper-level controller, and an optimal allocation and adaptive controller.

[0007] The Newton-Euler method is used to construct a roll motion model for a dual-track vehicle, which can construct the relationship between the roll moment when the vehicle rotates around the roll axis and the suspension force and other external forces; the vehicle pitch and vertical motion model based on Lagrangian mechanics can clarify the relationship between the kinetic energy generated when the vehicle pitches, the potential energy generated by spring deformation and tire deformation, and the dissipated energy generated by the damper and the external forces of the system; the pitch and vertical motion model is constructed based on the fusion of neural networks and physical models, which mainly integrates the neural network and the pitch and vertical motion model constructed based on Lagrangian mechanics to construct a more accurate model for improving the control effect; the upper-level controller is used to generate the desired pitch and roll control torque; the optimal distribution and adaptive controller is used to optimally distribute the desired pitch and roll control torque generated by the upper-level controller, that is, to solve the optimal vertical force of the four suspensions.

[0008] First, a double-track vehicle roll motion model is constructed based on the Newton-Euler method. Before constructing the vehicle model, the present invention first clarifies the conversion relationship between different coordinate systems to facilitate complex vehicle modeling. First, define S V The vehicle coordinate system is fixed to the center of mass of the vehicle, which satisfies the right-hand rule, with the positive direction of the Z axis pointing upwards. At the same time, define S B is the body coordinate system, which also satisfies the right-hand rule. The x-axis of the body coordinate system is the same as the vehicle coordinate system S V The X-axis of the body and the body can be rotated around it, and the rotation angle is defined as the roll angle φ of the vehicle. Therefore, the rotation matrix of the body coordinate system is converted to the vehicle coordinate system It can be expressed as:

[0009]

[0010] Among them, c φ and s φ represent cosφ and sinφ respectively.

[0011] The vehicle model established in this invention is a dual-track vehicle model that includes roll motion. It consists of two parts: the chassis and the body. The chassis has longitudinal, lateral, and yaw degrees of freedom, while the body has roll motion around the roll axis. The suspension part connecting the chassis and the body is characterized by stiffness and damping around the roll axis. For the proposed dual-track model, based on the Euler equation, the sum of the external moments acting on the vehicle system, τ, is constructed. V The relationship between the rate of change of angular velocity is:

[0012]

[0013]

[0014]

[0015] Among them, I V represents the vehicle moment of inertia matrix in the vehicle coordinate system, ω represents the angular velocity matrix in the vehicle body coordinate system, ω V Represents the angular velocity matrix in the vehicle coordinate system, I B Represents the vehicle moment of inertia matrix in the body coordinate system, I xx , I yy , I zz Respectively represent the moment of inertia around the x, y, and z axes of the vehicle body coordinate system. V It can be expressed as the following formula:

[0016]

[0017] Using the aforementioned Euler equation, the vehicle's roll moment τ can be derived: x Roll angular velocity change rate The relationship between them is:

[0018]

[0019] Among them, it means Yaw angular velocity.

[0020] At the same time, the roll moment applied to the vehicle comes from the stiffness and damping of the suspension and the force of the tires. Therefore, the roll moment τ x It can also be expressed as the following formula:

[0021]

[0022] Among them, c φ represents the damping coefficient of the suspension system, k φ Indicates the roll stiffness of the suspension system, m s Indicates sprung mass, F y Indicates the lateral force on the vehicle. In the present invention, F y =m s a y , h represents the height of the center of mass, M x Indicates the roll control torque.

[0023] Therefore, τ x Substitute M x We can get:

[0024]

[0025] Therefore, the Newton-Euler method is used to construct the vehicle roll model.

[0026] The present invention then constructs a model for the vehicle's pitch and vertical motion based on Lagrangian mechanics. Unlike roll motion, pitch dynamics are not significantly related to vehicle stability, but rather, like vertical motion, are closely related to ride comfort. Therefore, it is crucial to employ a model that can simultaneously describe the pitch and vertical motion of the vehicle's sprung mass. Therefore, the present invention establishes a model for the pitch and vertical motion of the half-vehicle based on Lagrangian mechanics.

[0027] First construct the vehicle's kinetic energy T:

[0028]

[0029] in, is the vertical velocity at the center of mass of the sprung mass, m u1 and m u2 denote the unsprung masses of the front and rear suspensions, respectively. and denote the vertical velocities of the front and rear unsprung masses, Indicates the vehicle body pitch angular velocity.

[0030] The potential energy V of the vehicle is mainly generated by the deformation of the springs and tires, which can be expressed as:

[0031]

[0032] Among them, k1 and k2 represent the suspension stiffness of the front and rear suspension of the vehicle respectively, k t1 and k t2 Denote the radial stiffness of the front and rear tires, z srepresents the vertical displacement of the sprung mass center of mass, a and b represent the distance from the center of mass to the front and rear axles, respectively, and z u1 and z u2 are the vertical displacements of the front and rear unsprung masses, and θ is the vehicle body pitch angle.

[0033] The dissipated energy D of the vehicle is mainly generated by the damper. Since the damping value of the tire is very small, the damping characteristics of the tire are not considered in this invention. Therefore, the dissipated energy can be expressed as:

[0034]

[0035] Among them, c1 and c2 represent the suspension damping coefficients of the front and rear suspensions respectively.

[0036] The present invention defines the state quantity q=[z s θz u1 z u2 ] T , therefore, the vehicle's equation of motion can be expressed by the Lagrange equation as:

[0037]

[0038] Where Q represents the generalized force, Represents state quantity The first-order derivative with respect to time. Q is solved by the principle of virtual work, as shown below:

[0039]

[0040]

[0041] M y =aF zf -bF zr

[0042] Among them, δW represents virtual work, δq represents virtual displacement, F zf and F zr Denote the forces acting on the front and rear suspensions, δz s and δθ represent the virtual displacement at the center of mass of the sprung mass and the virtual displacement of the pitch angle, respectively, δz uf and δz ur Denote the virtual displacement of the front and rear suspension unsprung masses, M y represents the pitching moment.

[0043] Through the above derivation, we can obtain the ordinary differential equations of vehicle pitch and vertical motion based on Lagrangian dynamics:

[0044]

[0045]

[0046]

[0047]

[0048] in, It represents the second-order derivative of the state quantity with respect to time, that is, the acceleration state vector, M represents the mass matrix, C represents the damping matrix, and K represents the stiffness matrix.

[0049] The aforementioned vehicle pitch and vertical motion model is constructed based on Lagrangian mechanics. In order to improve the modeling accuracy, the present invention constructs the pitch and vertical motion model based on the fusion of neural network and physical model. Since the mass, stiffness and damping matrix in the aforementioned vehicle pitch and vertical motion model are fixed values, the mass, suspension stiffness and damping of the system will change with the change of their own state, especially the damping coefficient will change significantly with the change of compression and rebound speed of the damper. Therefore, it is necessary to establish a mapping relationship between the damper speed and the corresponding damping coefficient value through the neural network. At the same time, the relationship between the vehicle mass and suspension stiffness and the state quantity will also be constructed through the neural network. Therefore, the fusion of neural network and physical model will improve the accuracy of the model.

[0050] First, the neural network uses the vehicle state q as the input of the neural network. The number of hidden layers of the neural network is 2, the number of neurons in the hidden layer is 64, and the number of neurons in the output layer of the neural network is 30. The calculation process is shown in the following formula.

[0051] z i =W i h i-1 +b i

[0052] h i =g(z i )

[0053] Among them, W i represents the weight matrix of the i-th layer neural network, b i represents the bias matrix of the i-th layer neural network, h i-1 represents the output matrix of the i-1th layer neural network, z i represents the weighted input of the i-th layer, g(·) represents the activation function, and h i Represents the output of the i-th layer neural network.

[0054] Among the 30 neurons in the output layer, the 1st to 10th output neurons constitute l M , the outputs of neurons 11 to 20 constitute l C , the outputs of neurons 21 to 30 constitute lK , the hidden activation function of the neural network is softplus, and the activation function of the output layer is l M 、l C 、l K l in d The activation function of the element is softplus, l M 、l C 、l K l in o The activation function of the element is a linear activation function. As shown in the following formula:

[0055] l M =(l dM (1),l oM (1),l dM (2),l oM (2),l oM (3),l dM (3),l oM (4),l oM (5),l oM (6),l dM (4) T

[0056] l C =(l dC (1),l oC (1),l dC (2),l oC (2),l oC (3),l dC (3),l oC (4),l oC (5),l oC (6),l dC (4) T

[0057] l K =(l dK (1),l oK (1),l dK (2),l oK (2),l oK (3),l dK (3),l oK (4),l oK (5),l oK (6),l dK (4) T

[0058] l M 、l C 、l K After output, it will be transformed into the lower triangular matrix L M、L C 、L K , l M 、l C 、l K l in d are the diagonal elements of the matrix, l o is a non-diagonal element, as shown in the formula, L M 、L C 、L K The transpose of L M 、L C 、L K The product of can get the symmetric positive definite matrix They are called predicted mass, predicted damping and predicted stiffness matrices respectively. The predicted mass matrix and acceleration state Damping Matrix and speed state Stiffness matrix Multiplying by q and summing them up will give the system external force predicted by the neural network. Instead of this As shown in the following formula:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065] Therefore, when training a neural network, the external force is minimized by The difference between the actual suspension force Q is used to optimize the weights of the neural network, as shown in the following formula.

[0066]

[0067]

[0068] in, Represents the inverse function, that is, constructing the state quantity q, The mapping relationship with the generalized force Q, ζ represents the parameters in the neural network, such as weight W and bias b.

[0069] The above content establishes a vehicle roll and pitch model. Based on the above model, the present invention first designs an upper-layer controller, which includes a roll controller and a pitch and vertical controller.

[0070] The present invention first designs a roll controller. In order to realize the roll control of the vehicle, the roll controller is designed by using sliding mode control based on the double-track vehicle roll model derived from the Newton-Euler method.

[0071] First, the control goal of the present invention is to reduce the roll motion of the vehicle so that the vehicle roll angle is 0, and to use the generalized moment M x The sliding mode controller is designed as the control variable, where the roll angular acceleration can be expressed as:

[0072]

[0073]

[0074] in, represents the roll angle, represents the roll angular velocity, k φ represents the equivalent roll stiffness of the suspension, c φ represents the equivalent damping of the suspension, a y represents lateral acceleration and g represents acceleration due to gravity.

[0075] Therefore, the present invention first designs the sliding surface s shown below based on the sliding mode control theory:

[0076]

[0077] Among them, λ is greater than 0. So The design of the reaching law in the present invention is Where ε>0, sgn() represents the sign function, so the sliding mode control rate for vehicle roll control can be expressed as:

[0078]

[0079] Among them, M xdes represents the desired roll control torque solved by the sliding mode controller.

[0080] Next, the present invention designs pitch and vertical controllers. In order to achieve pitch and vertical control of the vehicle, the present invention designs the controller based on the aforementioned vehicle Lagrangian dynamics model constructed using a neural network, and the control method also adopts sliding mode control theory.

[0081] The goal of pitch and vertical motion control is to reduce the pitch motion and vertical oscillation of the vehicle so that the pitch acceleration and vertical acceleration of the vehicle are 0. Control objective The vehicle pitch and vertical motion can be expressed by the ordinary differential equations mentioned above, where the vehicle's mass matrix, stiffness matrix, and damping matrix are replaced by the predicted mass matrix, predicted stiffness matrix, and predicted damping matrix established by the neural network. The expression is as follows:

[0082]

[0083] Therefore, the present invention defines the following sliding surface σ:

[0084]

[0085] Among them, Λ is greater than 0. So The design of the reaching law in the present invention is Where k>0, so the sliding mode control rate for pitch and vertical control can be expressed as:

[0086]

[0087] Among them, F zdes represents the desired vertical control force solved by the sliding mode controller, M ydes represents the desired pitch control torque solved by the sliding mode controller.

[0088] Based on the above-mentioned sliding mode control rate for reducing vehicle roll motion and the sliding mode control rate for reducing vehicle pitch and vertical motion, the desired virtual control amount, that is, the desired roll control torque M, can be obtained. xdes , desired pitch control torque M ydes and the desired vertical control force F zdes .

[0089] Based on the desired virtual control variable solved by the sliding mode controller, in actual vehicle control, the desired virtual control variable needs to be converted and distributed to the four suspension dampers. The dampers generate actual forces to control the vehicle's roll and pitch motion. Therefore, the present invention designs an optimal distribution controller for the virtual control variable. The optimal distribution problem needs to be solved using quadratic programming, and its expression is as follows:

[0090]

[0091]

[0092]

[0093] Where γ>0 and is large enough, v represents the desired control quantity solved by the sliding mode controller, B represents the control conversion matrix, which is used to construct the mapping relationship between the suspension force and the virtual control signal, u represents the actual control force of the suspension, and Wv (α) and W u Both represent weight matrices, and They represent the upper and lower limit constraints of the control quantity, u0 represents the damping force solved by the ceiling damping control method, F sky The skyhook control theory is based on the vertical velocity of the sprung mass and the relative velocity between the sprung and unsprung masses The specific expressions of the above-mentioned quantities for the solved equilibrium damping force are as follows:

[0094]

[0095] V=[M xdes M ydes F zdes ] T

[0096] u=[F zfl F zfr F zrl F zrr ] T

[0097] u0=[u 1,eq u 2,eq u 3,eq u 4,eq ] T

[0098]

[0099]

[0100] Among them, H represents the left and right wheelbase of the vehicle, F zfl and F zfr Represents the forces acting on the left and right suspensions of the front suspension, F zrl 、F zrr Respectively represent the forces acting on the left and right suspensions of the rear suspension, u i,eq represents the equilibrium damping force of each suspension based on the skyhook control theory, where i = 1, 2, 3, 4, representing the front left, front right, rear left, and rear right, respectively, and c sky represents the ceiling damping coefficient.

[0101] Therefore, the roll control moment, pitch control moment and vertical force generated by the forces generated by the four suspensions of the front and rear suspensions can be expressed as follows:

[0102]

[0103] M y =a(F zfl+F zfr )-b(F zrl +F zrr )

[0104] F z =F zfl +F zfr +F zrl +F zrr

[0105] Among them, F z It represents the vertical resultant force generated by the four suspensions.

[0106] The purpose of the first term of the above cost function is to minimize the distribution error, that is, to minimize the error between the roll moment and pitch moment formed by the actual force generated by the suspension and the expected value solved by the sliding mode control. The u0 in the second term of the cost function is the equilibrium point constructed based on the skyhook control theory. When the weight of the first term of the cost function is very small and can be ignored, the cost function only has the second term. At this time, the entire control system is simplified to the independent control of the four suspensions based on the skyhook control theory.

[0107] During vehicle driving, when the vehicle makes an emergency turn or avoids an obstacle, the vehicle may experience severe roll or even rollover. Therefore, when the vehicle roll angle or lateral acceleration is too large, the suspension system should increase the weight of roll control to improve safety. Similarly, when the vehicle brakes suddenly, the pitch angle increases or the pitch acceleration is large, and the weight of pitch control should be increased to improve comfort. Therefore, the present invention adjusts the control focus according to different working conditions, and thus designs an adaptive control weight, the adaptive weight W v The expression of (α) is as follows:

[0108]

[0109] in. represents the variable weight of roll control, represents the variable weight of pitch control, Represents the vertical control variable weight.

[0110] The value of is determined by the following formula:

[0111]

[0112] in, express and The maximum of the two.

[0113] The value of is determined by the roll angle and lateral acceleration of the vehicle. The thresholds of the roll angle and lateral acceleration are selected first, where the threshold of the roll angle is is the critical rollover angle φ c The threshold value of lateral acceleration is half of Normalize the roll angle and lateral acceleration, and map the normalized signal to the interval [0,1] after passing the function h. Then sum the two signals and limit them between 0 and 1. The summed signal can be obtained by smoothing and filtering. The expression of function h is:

[0114]

[0115] Similarly, The value of is determined by the pitch angle and longitudinal acceleration of the vehicle. The thresholds of pitch angle and longitudinal acceleration are selected first, where the threshold of pitch angle is The threshold value of longitudinal acceleration is Normalize the pitch angle and longitudinal acceleration, and map the normalized signal to the interval [0,1] after passing the function h. Then sum the two signals and limit them between 0 and 1. The summed signal can be obtained by smoothing and filtering. The function h is expressed as:

[0116]

[0117] Beneficial effects of the present invention:

[0118] 1. Compared with the traditional independent suspension control based on a quarter-vehicle model, the present invention uses the upper-level roll controller, pitch and vertical controller to solve the desired control force and torque, and then uses the lower-level optimal distribution control algorithm to solve the forces acting on each suspension. This achieves coordinated control among the four suspensions, which can effectively suppress vehicle roll, improve vehicle stability and safety, and effectively suppress vehicle pitch motion, improving comfort.

[0119] 2. The present invention achieves the adaptability of the suspension under different working conditions by constructing an adaptive weight matrix. When the vehicle is turning or avoiding obstacles in an emergency, the risk of vehicle rollover increases. At this time, the weight of roll control is increased to suppress vehicle roll and reduce the risk of rollover. On the other hand, when the vehicle is emergency braked, the weight of pitch control is increased to suppress vehicle pitch and improve passenger comfort. When there is no turning or emergency braking, the suspension is approximately independently controlled, mainly used to reduce the vertical acceleration of the center of mass, improve passenger comfort and protect cargo from excessive vertical acceleration.

[0120] 3. The present invention constructs a Lagrangian mechanical model of vehicle pitch motion by combining a physical model with a neural network. This will effectively improve the accuracy of the modeling, especially the modeling accuracy of the damping coefficient of the system. Since the semi-active suspension is mainly controlled by adjusting a limited damping variation range, an accurate damping coefficient will help improve the control effect. At the same time, due to the fast computing speed of the neural network, this will improve the response speed of the suspension system and further enhance the control effect of vehicle pitch and vertical motion. BRIEF DESCRIPTION OF THE DRAWINGS

[0121] Figure 1 It is the rolling dynamics model of double-track vehicle;

[0122] Figure 2 is the pitch dynamics model;

[0123] Figure 3 A pitch and vertical motion model that integrates neural networks and physical models;

[0124] Figure 4 Schematic diagram of the principle of semi-active suspension parameter adaptive cooperative control system. DETAILED DESCRIPTION

[0125] The present invention will be further described below with reference to the accompanying drawings.

[0126] Figure 1 For the roll dynamics model of a dual-track vehicle, before constructing the vehicle model, the present invention first clarifies the conversion relationship between different coordinate systems to facilitate complex modeling of the vehicle. First, define S V The vehicle coordinate system is fixed to the center of mass of the vehicle, which satisfies the right-hand rule, with the positive direction of the Z axis pointing upwards. At the same time, define S B is the body coordinate system, which also satisfies the right-hand rule. The x-axis of the body coordinate system is the same as the vehicle coordinate system S V The X-axis of the body and the body can be rotated around it, and the rotation angle is defined as the roll angle φ of the vehicle. Therefore, the rotation matrix of the body coordinate system is converted to the vehicle coordinate system It can be expressed as:

[0127]

[0128] Among them, c φ and s φ represent cosφ and sinφ respectively.

[0129] The established dual-track vehicle model with roll motion consists of two parts: the chassis and the body. The chassis has longitudinal, lateral, and yaw degrees of freedom, while the body has roll motion around the roll axis. The suspension connecting the chassis and the body is represented by the stiffness and damping around the roll axis. For the proposed dual-track model, based on the Euler equation, the sum of the external moments acting on the vehicle system, τ, is constructed. V The relationship between the rate of change of angular velocity is:

[0130]

[0131]

[0132]

[0133] Among them, I V represents the vehicle moment of inertia matrix in the vehicle coordinate system, ω represents the angular velocity matrix in the vehicle body coordinate system, ω V Represents the angular velocity matrix in the vehicle coordinate system, I B Represents the vehicle moment of inertia matrix in the body coordinate system, I xx , I yy , I zz Respectively represent the moment of inertia around the x, y, and z axes of the vehicle body coordinate system. V It can be expressed as the following formula:

[0134]

[0135] Using the aforementioned Euler equation, the vehicle's roll moment τ can be derived: x Roll angular velocity change rate The relationship between them is:

[0136]

[0137] Among them, it means Yaw angular velocity.

[0138] At the same time, the roll moment applied to the vehicle comes from the stiffness and damping of the suspension and the force of the tires. Therefore, the roll moment τ x It can also be expressed as the following formula:

[0139]

[0140] Among them, c φ represents the damping coefficient of the suspension system, k φ Indicates the roll stiffness of the suspension system, m s Indicates sprung mass, F y Indicates the lateral force on the vehicle. In the present invention, F y =ms a y , h represents the height of the center of mass, M x Indicates the roll control torque.

[0141] Therefore, τ x Substitute M x We can get:

[0142]

[0143] Figure 2 The pitch dynamics model is a Lagrangian-based model that describes both the pitch and vertical motion of a vehicle. Unlike roll motion, pitch dynamics are not particularly relevant to vehicle stability, but rather, like vertical motion, are closely related to ride comfort. Therefore, it is crucial to have a model that can simultaneously describe the pitch and vertical motion of a vehicle's sprung mass.

[0144] First construct the vehicle's kinetic energy T:

[0145]

[0146] in, is the vertical velocity at the center of mass of the sprung mass, m u1 and m u2 denote the unsprung masses of the front and rear suspensions, respectively. and denote the vertical velocities of the front and rear unsprung masses, Indicates the vehicle body pitch angular velocity.

[0147] The potential energy V of the vehicle is mainly generated by the deformation of the springs and tires, which can be expressed as:

[0148]

[0149] Among them, k1 and k2 represent the suspension stiffness of the front and rear suspension of the vehicle respectively, k t1 and k t2 Denote the radial stiffness of the front and rear tires, z s represents the vertical displacement of the sprung mass center of mass, a and b represent the distance from the center of mass to the front and rear axles, respectively, and z u1 and z u2 are the vertical displacements of the front and rear unsprung masses, and θ is the vehicle body pitch angle.

[0150] The dissipated energy D of the vehicle is mainly generated by the damper. Since the damping value of the tire is very small, the damping characteristics of the tire are not considered in this invention. Therefore, the dissipated energy can be expressed as:

[0151]

[0152] Among them, c1 and c2 represent the suspension damping coefficients of the front and rear suspensions respectively.

[0153] The present invention defines the state quantity q=[z s θz u1 z u2 ] T , therefore, the vehicle's equation of motion can be expressed by the Lagrange equation as:

[0154]

[0155] Where Q represents the generalized force, Represents state quantity The first derivative with respect to time. The solution of Q is solved by the principle of virtual work, as shown below:

[0156]

[0157]

[0158] M y =aF zf -bF zr

[0159] Among them, δW represents virtual work, δq represents virtual displacement, F zf and F zr Represent the forces acting on the front and rear suspensions, and δθ represent the virtual displacement at the sprung mass center and the pitch angle virtual displacement, respectively. and Denote the virtual displacement of the front and rear suspension unsprung masses, M y represents the pitching moment.

[0160] Through the above derivation, we can obtain the ordinary differential equations of vehicle pitch and vertical motion based on Lagrangian dynamics:

[0161]

[0162]

[0163]

[0164]

[0165] in, It represents the second-order derivative of the state quantity with respect to time, that is, the acceleration state vector, M represents the mass matrix, C represents the damping matrix, and K represents the stiffness matrix.

[0166] Figure 3This is a pitch and vertical motion model that integrates neural networks and physical models. Since the mass, stiffness and damping matrices in the aforementioned vehicle pitch and vertical motion model are fixed values, the system's mass, suspension stiffness and damping of the damper will change with changes in their own states, especially the damping coefficient will change significantly with changes in the compression and rebound speeds of the damper. Therefore, it is necessary to establish a mapping relationship between the damper speed and the corresponding damping coefficient value through a neural network. At the same time, the relationship between the vehicle mass and suspension stiffness and the state quantity will also be constructed through a neural network. Therefore, the integration of neural networks and physical models will improve the accuracy of the model.

[0167] First, the neural network uses the vehicle state q as the input of the neural network. The number of hidden layers of the neural network is 2, the number of neurons in the hidden layer is 64, and the number of neurons in the output layer of the neural network is 30. The calculation process is shown in the following formula.

[0168] z i =W i h i-1 +b i

[0169] h i =g(z i )

[0170] Among them, W i represents the weight matrix of the i-th layer neural network, b i represents the bias matrix of the i-th layer neural network, h i-1 represents the output matrix of the i-1th layer neural network, z i represents the weighted input of the i-th layer, g represents the activation function, h i Represents the output of the i-th layer neural network.

[0171] Among the 30 neurons in the output layer, the 1st to 10th output neurons constitute l M , the outputs of neurons 11 to 20 constitute l C , the outputs of neurons 21 to 30 constitute l K , the hidden activation function of the neural network is softplus, and the activation function of the output layer is l M 、l C 、l K l in d The activation function of the element is softplus, l M 、l C 、l K l in o The activation function of the element is a linear activation function. As shown in the following formula:

[0172] lM =(l dM (1),l oM (1),l dM (2),l oM (2),l oM (3),l dM (3),l oM (4),l oM (5),l oM (6),l dM (4) T

[0173] l C =(l dC (1),l oC (1),l dC (2),l oC (2),l oC (3),l dC (3),l oC (4),l oC (5),l oC (6),l dC (4) T

[0174] l K =(l dK (1),l oK (1),l dK (2),l oK (2),l oK (3),l dK (3),l oK (4),l oK (5),l oK (6),l dK (4) T

[0175] l M 、l C 、l K After output, it will be transformed into the lower triangular matrix L M 、L C 、L K , l M 、l C 、l K l in d are the diagonal elements of the matrix, l o is a non-diagonal element, as shown in the formula, L M 、L C 、L K The transpose of L M 、L C 、L KThe product of can get the symmetric positive definite matrix They are called predicted mass, predicted damping and predicted stiffness matrices respectively. The predicted mass matrix and acceleration state Damping Matrix and speed state Stiffness matrix Multiplying by q and summing them up will give the system external force predicted by the neural network. Instead of this As shown in the following formula:

[0176]

[0177]

[0178]

[0179]

[0180]

[0181]

[0182] Therefore, when training a neural network, the external force is minimized by The difference between the actual suspension force Q is used to optimize the weights of the neural network, as shown in the following formula.

[0183]

[0184]

[0185] in, Represents the inverse function, that is, constructing the state quantity q, The mapping relationship with the generalized force Q, ζ represents the parameters in the neural network, such as weight W and bias b.

[0186] Figure 4 This is a semi-active suspension parameter adaptive collaborative control system. The system uses the roll model and the pitch and vertical motion models of the fusion neural network as the controlled objects, and uses a sliding mode controller as the upper-level controller to solve the desired roll control torque and pitch control torque. These are then distributed through an optimal distribution control algorithm, and the weights are adjusted according to changes in vehicle operating conditions to achieve adaptive control. Finally, the forces acting on the four suspensions are solved to achieve collaborative control. The vehicle then feeds back its own state quantity to the controller to achieve closed-loop control.

[0187] For the upper-level controller, the pitch controller is designed first. In order to realize the roll control of the vehicle, the roll controller is designed using sliding mode control based on the double-track vehicle roll model derived from the Newton-Euler method.

[0188] First, the control goal of the pitch controller is to reduce the roll motion of the vehicle so that the vehicle roll angle is 0, and the generalized moment M x The sliding mode controller is designed as the control variable, where the roll angular acceleration can be expressed as:

[0189]

[0190]

[0191] in, represents the roll angle, represents the roll angular velocity, k φ represents the equivalent roll stiffness of the suspension, c φ represents the equivalent damping of the suspension, a y represents lateral acceleration and g represents acceleration due to gravity.

[0192] Therefore, the present invention first designs the sliding surface s shown below based on the sliding mode control theory:

[0193]

[0194] Among them, λ is greater than 0. So The design of the reaching law in the present invention is Where ε>0, therefore, the sliding mode control rate for vehicle roll control can be expressed as:

[0195]

[0196] Among them, M xdes represents the desired roll control torque solved by the sliding mode controller.

[0197] Next, the present invention designs pitch and vertical controllers. In order to achieve pitch and vertical control of the vehicle, the present invention designs the controllers based on the pitch and vertical motion model that integrates the neural network and the physical model. The control method also adopts the sliding mode control theory.

[0198] The goal of pitch and vertical motion control is to reduce the pitch motion and vertical oscillation of the vehicle so that the pitch acceleration and vertical acceleration of the vehicle are 0. Control objective The vehicle pitch and vertical motion can be expressed by the ordinary differential equations mentioned above, where the vehicle's mass matrix, stiffness matrix, and damping matrix are replaced by the predicted mass matrix, predicted stiffness matrix, and predicted damping matrix established by the neural network. The expression is as follows:

[0199]

[0200] Therefore, the present invention defines the following sliding surface σ:

[0201]

[0202] Among them, Λ is greater than 0. So The design of the reaching law in the present invention is Where k>0, so the sliding mode control rate for pitch and vertical control can be expressed as:

[0203]

[0204] Among them, F zdes represents the desired vertical control force solved by the sliding mode controller, M ydes represents the desired pitch control torque solved by the sliding mode controller.

[0205] Based on the above-mentioned sliding mode control rate for reducing vehicle roll motion and the sliding mode control rate for reducing vehicle pitch and vertical motion, the desired virtual control amount, that is, the desired roll control torque M, can be obtained. xdes , desired pitch control torque M ydes and the desired vertical control force F zdes .

[0206] Based on the desired virtual control variable solved by the sliding mode controller, in actual vehicle control, the desired virtual control variable needs to be converted and distributed to the four suspension dampers. The dampers generate actual forces to control the vehicle's roll and pitch motion. Therefore, the present invention designs an optimal distribution controller for the virtual control variable. The optimal distribution problem needs to be solved using quadratic programming, and its expression is as follows:

[0207]

[0208]

[0209]

[0210] Where γ>0 and is large enough, v represents the desired control quantity solved by the sliding mode controller, B represents the control conversion matrix, which is used to construct the mapping relationship between the suspension force and the virtual control signal, u represents the actual control force of the suspension, and W v (α) and W u Both represent weight matrices, and They represent the upper and lower limit constraints of the control quantity, u0 represents the damping force solved by the ceiling damping control method, F skyThe skyhook control theory is based on the vertical velocity of the sprung mass and the relative velocity between the sprung and unsprung masses The specific expressions of the above-mentioned quantities for the solved equilibrium damping force are as follows:

[0211]

[0212] V=[M xdes M ydes F zdes ] T

[0213] u=[F zfl F zfr F zrl F zrr ] T

[0214] u0=[u 1,eq u 2,eq u 3,eq u 4,eq ] T

[0215]

[0216]

[0217] Among them, H represents the left and right wheelbase of the vehicle, F zfl and F zfr Represents the forces acting on the left and right suspensions of the front suspension, F zrl 、F zrr Respectively represent the forces acting on the left and right suspensions of the rear suspension, u i,eq represents the equilibrium damping force of each suspension based on the skyhook control theory, where i = 1, 2, 3, 4, representing the front left, front right, rear left, and rear right, respectively, and c sky represents the ceiling damping coefficient.

[0218] Therefore, the roll control moment, pitch control moment and vertical force generated by the forces generated by the four suspensions of the front and rear suspensions can be expressed as follows:

[0219]

[0220] M y =a(F zfl +F zfr )-b(F zrl +F zrr )

[0221] F z =F zfl +Fzfr +F zrl +F zrr

[0222] Wherein, represents the vertical force generated by the four suspensions.

[0223] The purpose of the first term of the above cost function is to minimize the distribution error, that is, to minimize the error between the roll moment and pitch moment formed by the actual force generated by the suspension and the expected value solved by the sliding mode control. The u0 in the second term of the cost function is the equilibrium point constructed based on the skyhook control theory. When the weight of the first term of the cost function is very small and can be ignored, the cost function only has the second term. At this time, the entire control system is simplified to the independent control of the four suspensions based on the skyhook control theory.

[0224] During vehicle driving, when the vehicle makes an emergency turn or avoids an obstacle, the vehicle may experience severe roll or even rollover. Therefore, when the vehicle roll angle or lateral acceleration is too large, the suspension system should increase the weight of roll control to improve safety. Similarly, when the vehicle brakes suddenly, the pitch angle increases or the pitch acceleration is large, and the weight of pitch control should be increased to improve comfort. Therefore, the present invention adjusts the control focus according to different working conditions, and thus designs an adaptive control weight, the adaptive weight W v The expression of (α) is as follows:

[0225]

[0226] in. represents the variable weight of roll control, represents the variable weight of pitch control, Represents the vertical control variable weight.

[0227] The value of is determined by the following formula:

[0228]

[0229] Among them, ∨ represents and The maximum of the two.

[0230] The value of is determined by the roll angle and lateral acceleration of the vehicle. The thresholds of the roll angle and lateral acceleration are selected first, where the threshold of the roll angle is is the critical rollover angle φ c The threshold value of lateral acceleration is half of Normalize the roll angle and lateral acceleration, and map the normalized signal to the interval [0,1] after passing the function h. Then sum the two signals and limit them between 0 and 1. The summed signal can be obtained by smoothing and filtering. The expression of function h is:

[0231]

[0232] Similarly, The value of is determined by the pitch angle and longitudinal acceleration of the vehicle. The thresholds of pitch angle and longitudinal acceleration are selected first, where the threshold of pitch angle is The threshold value of longitudinal acceleration is Normalize the pitch angle and longitudinal acceleration, and map the normalized signal to the interval [0,1] after passing the function h. Then sum the two signals and limit them between 0 and 1. The summed signal can be obtained by smoothing and filtering. The expression of function h is.

[0233]

[0234] The series of detailed descriptions listed above are only specific descriptions of feasible implementation methods of the present invention. They are not intended to limit the scope of protection of the present invention. Any equivalent methods or changes that do not deviate from the technology of the present invention should be included in the scope of protection of the present invention.

Claims

1. A commercial vehicle semi-automatic suspension adaptive cooperative control system based on a data-driven model, characterized by: include: A dual-track vehicle roll motion model constructed based on the Newton-Euler method, a pitch and vertical motion model constructed based on the fusion of neural networks and physical models, an upper-level controller, and an optimal allocation and adaptive controller; the vehicle roll motion model can construct the relationship between the roll moment and the suspension force and other external forces when the vehicle rotates around the roll axis; the vehicle pitch and vertical motion model can clarify the relationship between the kinetic energy generated when the vehicle pitches, the potential energy generated by spring deformation and tire deformation, and the dissipated energy generated by the damper and the external forces of the system; the pitch and vertical motion model constructed based on the fusion of neural networks and physical models is mainly to integrate the neural network and the pitch and vertical motion model constructed based on Lagrangian mechanics to construct a more accurate model for improving the control effect; the upper-level controller is used to generate the desired pitch and roll control torque; the optimal allocation and adaptive controller is used to optimally distribute the desired pitch and roll control torque generated by the upper-level controller, that is, to solve the optimal vertical force of the four suspensions; The upper controller includes the roll controller and the pitch and vertical controller; Roll controller, in order to achieve vehicle roll control, a roll controller is designed using sliding mode control based on the double-track vehicle roll motion model derived from the Newton-Euler method; Pitch and vertical controller, in order to realize the pitch and vertical control of the vehicle, the controller is designed based on the pitch and vertical motion model constructed by the fusion of neural network and physical model, and the pitch and vertical controller is designed using sliding mode control.

2. The data-driven model-based commercial vehicle semi-automatic suspension adaptive cooperative control system according to claim 1, characterized in that: The double-track vehicle roll motion model constructed based on the Newton-Euler method consists of two parts: the chassis and the body. The chassis has longitudinal, lateral, and yaw degrees of freedom, while the body has roll motion around the roll axis. The suspension part connecting the chassis and the body is expressed as stiffness and damping around the roll axis. Based on the Euler equation, the sum of the external moments acting on the vehicle system, τ, is constructed. V The relationship between the rate of change of angular velocity is: Among them, I V represents the vehicle moment of inertia matrix in the vehicle coordinate system, ω represents the angular velocity matrix in the vehicle body coordinate system, ω V Represents the angular velocity matrix in the vehicle coordinate system, I B Represents the vehicle moment of inertia matrix in the body coordinate system, I xx , I yy , I zz Respectively represent the moment of inertia around the x, y, and z axes of the vehicle body coordinate system; I V It can be expressed as the following formula: Using the aforementioned Euler equation, the vehicle's roll moment τ can be derived: x Roll angular velocity change rate The relationship between them is: Among them, it means yaw rate; At the same time, the roll moment applied to the vehicle comes from the stiffness and damping of the suspension and the force of the tires. Therefore, the roll moment τ x It can also be expressed as the following formula: Among them, c φ represents the damping coefficient of the suspension system, k φ Indicates the roll stiffness of the suspension system, m s Indicates sprung mass, F y Indicates the lateral force on the vehicle, F y =m s a y , h represents the height of the center of mass, M x represents the roll control torque; τ x Substitute M x We can get:

3. The data-driven model-based commercial vehicle semi-automatic suspension adaptive cooperative control system according to claim 2, characterized in that: It also includes the transformation coordinate system, defining S V The vehicle coordinate system is fixed to the center of mass of the vehicle, which satisfies the right-hand rule, with the positive direction of the Z axis pointing upwards. At the same time, define S B is the body coordinate system, which also satisfies the right-hand rule. The x-axis of the body coordinate system is the same as the vehicle coordinate system S V The X-axis of the vehicle is coincident and can rotate around it. The rotation angle is defined as the roll angle φ of the vehicle. Therefore, the rotation matrix of the vehicle body coordinate system is converted to the vehicle coordinate system It can be expressed as: Among them, c φ and s φ represent cosφ and sinφ respectively.

4. The data-driven model-based commercial vehicle semi-automatic suspension adaptive cooperative control system according to claim 1, characterized in that: The pitch and vertical motion model constructed based on the fusion of neural network and physical model is based on the pitch and vertical motion model constructed based on Lagrangian mechanics. The pitch and vertical motion model constructed based on Lagrangian mechanics is obtained as follows: First construct the vehicle's kinetic energy T: in, is the vertical velocity at the center of mass of the sprung mass, m u1 and m u2 denote the unsprung masses of the front and rear suspensions, respectively. and denote the vertical velocities of the front and rear unsprung masses, Indicates the vehicle body pitch angular velocity; The potential energy V of the vehicle is mainly generated by the deformation of the springs and tires, which can be expressed as: Among them, k1 and k2 represent the suspension stiffness of the front and rear suspension of the vehicle respectively, k t1 and k t2 Denote the radial stiffness of the front and rear tires, z s represents the vertical displacement of the sprung mass center of mass, a and b represent the distance from the center of mass to the front and rear axles, respectively, and z u1 and z u2 They represent the vertical displacement of the front and rear unsprung masses, and θ represents the vehicle body pitch angle; The dissipated energy D of the vehicle is mainly generated by the damper. Since the damping value of the tire is very small, the dissipated energy can be expressed as: Where c1 and c2 represent the suspension damping coefficients of the front and rear suspensions, respectively; Define the state quantity q=[z s θz u1 z u2 ] T , the vehicle's motion equation is expressed through the Lagrange equation as: Where Q represents the generalized force, Represents state quantity The solution to the first-order derivative of time, Q, is obtained by the principle of virtual work, as shown below: M y =aF zf -bF zr Among them, δW represents virtual work, δq represents virtual displacement, F zf and F zr Denote the forces acting on the front and rear suspensions, δz s and δθ represent the virtual displacement at the center of mass of the sprung mass and the virtual displacement of the pitch angle, respectively, δz uf and δz ur Denote the virtual displacement of the front and rear suspension unsprung masses, M y represents the pitching moment; By derivation, the ordinary differential equations for the vehicle pitch and vertical motion based on Lagrangian dynamics are obtained: in, It represents the second-order derivative of the state quantity with respect to time, that is, the acceleration state vector, M represents the mass matrix, C represents the damping matrix, and K represents the stiffness matrix.

5. The data-driven model-based commercial vehicle semi-automatic suspension adaptive cooperative control system according to claim 4, characterized in that: The pitch and vertical motion models constructed based on the fusion of neural networks and physical models are obtained as follows: First, the neural network uses the vehicle state q as the input of the neural network. The number of hidden layers of the neural network is 2, the number of neurons in the hidden layer is 64, and the number of neurons in the output layer of the neural network is 30. The operation process is shown in the following formula: z i =W i h i-1 +b i h i =g(z i ) Among them, W i represents the weight matrix of the i-th layer neural network, b i represents the bias matrix of the i-th layer neural network, h i-1 represents the output matrix of the i-1th layer neural network, z i represents the weighted input of the i-th layer, g represents the activation function, h i Represents the output of the i-th layer neural network; Among the 30 neurons in the output layer, the 1st to 10th output neurons constitute l M , the outputs of neurons 11 to 20 constitute l C , the outputs of neurons 21 to 30 constitute l K , the hidden activation function of the neural network is softplus, and the activation function of the output layer is l M 、l C 、l K l in d The activation function of the element is softplus, l M 、l C 、l K l in o The activation function of the element is a linear activation function, as shown below: l M ＝(l dM (1),l oM (1),l dM (2),l oM (2),l oM (3),l dM (3),l oM (4),l oM (5),l oM (6),l dM (4)) T l C ＝(l dC (1),l oC (1),l dC (2),l oC (2),l oC (3),l dC (3),l oC (4),l oC (5),l oC (6),l dC (4)) T l K ＝(l dK (1),l oK (1),l dK (2),l oK (2),l oK (3),l dK (3),l oK (4),l oK (5),l oK (6),l dK (4)) T l M 、l C 、l K After output, it will be transformed into the lower triangular matrix L M 、L C 、L K , l M 、l C 、l K l in d are the diagonal elements of the matrix, l o are off-diagonal elements, L M 、L C 、L K The transpose of L M 、L C 、L K The product of can get the symmetric positive definite matrix They are called predicted mass, predicted damping and predicted stiffness matrices respectively. The predicted mass matrix and acceleration state Damping Matrix and speed state Stiffness matrix Multiplying by q and summing them up will give the system external force predicted by the neural network. Instead of this As shown in the following formula: When training a neural network, the external force is predicted by minimizing The difference between the actual suspension force Q and the actual suspension force Q is used to optimize the weight of the neural network, as shown in the following formula in, Represents the inverse function, that is, constructing the state quantity q, The mapping relationship with the generalized force Q, ζ represents the parameters in the neural network, such as weight W and bias b.

6. The data-driven model-based commercial vehicle semi-automatic suspension adaptive cooperative control system according to claim 1, characterized in that: The control goal of the roll controller is to reduce the roll motion of the vehicle so that the vehicle roll angle is 0, and the generalized moment M x The sliding mode controller is designed as the control variable, where the roll angle acceleration is expressed as: in, represents the roll angle, represents the roll angular velocity, k φ represents the equivalent roll stiffness of the suspension, c φ represents the equivalent damping of the suspension, a y represents lateral acceleration, g represents acceleration due to gravity; Based on the sliding mode control theory, the sliding surface s is first designed: Among them, λ is greater than 0, so The reaching law design is Where ε>0, so the sliding mode control rate for vehicle roll control is expressed as: Among them, M xdes represents the desired roll control torque solved by the sliding mode controller.

7. The data-driven model-based commercial vehicle semi-automatic suspension adaptive cooperative control system according to claim 1, characterized in that: The control goal of the pitch and vertical controller is to reduce the pitch motion and vertical oscillation of the vehicle so that the pitch acceleration and vertical acceleration of the vehicle are 0. It is expressed by ordinary differential equations of vehicle pitch and vertical motion, in which the vehicle's mass matrix, stiffness matrix and damping matrix will be replaced by the predicted mass matrix, predicted stiffness matrix and predicted damping matrix established by the neural network. The expression is as follows: Define the sliding surface σ: Where Λ is greater than 0, The reaching law design is Where k>0, so the sliding mode control rate for pitch and vertical control can be expressed as: Among them, F zdes represents the desired vertical control force solved by the sliding mode controller, M ydes represents the desired pitch control torque solved by the sliding mode controller.

8. The data-driven model-based commercial vehicle semi-automatic suspension adaptive cooperative control system according to claim 1, characterized in that: The optimal allocation and adaptive controller are designed for the desired virtual control quantity output by the upper controller. The optimal allocation problem is solved using quadratic programming, and its expression is as follows: Where γ>0 and is large enough, v represents the desired control quantity solved by the sliding mode controller, B represents the control conversion matrix, which is used to construct the mapping relationship between the suspension force and the virtual control signal, u represents the actual control force of the suspension, and W v (α) and W u Both represent weight matrices, and They represent the upper and lower limit constraints of the control quantity, u0 represents the damping force solved by the ceiling damping control method, F sky The skyhook control theory is based on the vertical velocity of the sprung mass and the relative velocity between the sprung and unsprung masses The solved equilibrium damping force, where the specific expressions of each quantity are as follows: V=[M xdes M ydes F zdes ] T u=[F zfl F zfr F zrl F zrr ] T u0=[u 1,eq u 2,eq u 3,eq u 4,eq ] T Among them, H represents the left and right wheelbase of the vehicle, F zfl and F zfr Represents the forces acting on the left and right suspensions of the front suspension, F zrl 、F zrr Respectively represent the forces acting on the left and right suspensions of the rear suspension, u i,eq represents the equilibrium damping force of each suspension based on the skyhook control theory, where i = 1, 2, 3, 4, representing the front left, front right, rear left, and rear right, respectively, and c sky represents the ceiling damping coefficient; The roll control moment, pitch control moment, and vertical force generated by the forces generated by the four front and rear suspensions can be expressed as follows: M y =a(F zfl +F zfr )-b(F zrl +F zrr ) F z =F zfl +F zfr +F zrl +F zrr Among them, F z It represents the vertical resultant force generated by the four suspensions; The purpose of the first term in the cost function is to minimize the distribution error, that is, to minimize the error between the roll moment and pitch moment generated by the actual force generated by the suspension and the solved expected value. The u0 in the second term of the cost function is the equilibrium point constructed based on the skyhook control theory. When the weight of the first term in the cost function is very small and can be ignored, the cost function only has the second term, and the entire control system is simplified to the independent control of the four suspensions based on the skyhook control theory. Adaptive control weights are designed for different working conditions, and the adaptive weight W v The expression of (α) is as follows: in, represents the variable weight of roll control, represents the variable weight of pitch control, represents the vertical control variable weight; The value of is determined by the following formula: Among them, ∨ represents and The maximum of the two; The value of is determined by the roll angle and lateral acceleration of the vehicle. The thresholds of the roll angle and lateral acceleration are selected first, where the threshold of the roll angle is is the critical rollover angle φ c The threshold value of lateral acceleration is half of Normalize the roll angle and lateral acceleration, and map the normalized signal to the interval [0,1] after passing the function h. Then sum the two signals and limit them between 0 and 1. The summed signal can be obtained by smoothing and filtering. The expression is: Similarly, The value of is determined by the pitch angle and longitudinal acceleration of the vehicle. The thresholds of pitch angle and longitudinal acceleration are selected first, where the threshold of pitch angle is The threshold value of longitudinal acceleration is Normalize the pitch angle and longitudinal acceleration, and map the normalized signal to the interval [0,1] after passing the function h. Then sum the two signals and limit them between 0 and 1. The summed signal can be obtained by smoothing and filtering. Expressed as:

9. A method for adaptive collaborative control of a commercial vehicle semi-automatic suspension based on a data-driven model, characterized in that: These include: S1. Construct a rolling motion model of a dual-track vehicle based on the Newton-Euler method; It consists of two parts: the chassis and the body. The chassis has longitudinal, lateral and yaw degrees of freedom, while the body has roll motion around the roll axis. The suspension part connecting the chassis and the body is expressed as stiffness and damping around the roll axis. Based on the Euler equation, the sum of the external torques acting on the vehicle system is constructed as τ V The relationship between the rate of change of angular velocity is: Among them, I V represents the vehicle moment of inertia matrix in the vehicle coordinate system, ω represents the angular velocity matrix in the vehicle body coordinate system, ω V Represents the angular velocity matrix in the vehicle coordinate system, I B Represents the vehicle moment of inertia matrix in the body coordinate system, I xx , I yy , I zz Respectively represent the moment of inertia around the x, y, and z axes of the vehicle body coordinate system; I V It can be expressed as the following formula: Derive the vehicle's rolling moment τ x Roll angular velocity change rate The relationship between them is: Among them, it means yaw rate; At the same time, the roll moment applied to the vehicle comes from the stiffness and damping of the suspension and the force of the tires. Therefore, the roll moment τ x It can also be expressed as the following formula: Among them, c φ represents the damping coefficient of the suspension system, k φ Indicates the roll stiffness of the suspension system, m s Indicates sprung mass, F y Indicates the lateral force on the vehicle, F y =m s a y , h represents the height of the center of mass, M x represents the roll control torque; τ x Substitute M x We can get: S2. Construct vehicle pitch and vertical motion models based on Lagrangian mechanics; First construct the vehicle's kinetic energy T: in, is the vertical velocity at the center of mass of the sprung mass, m u1 and m u2 denote the unsprung masses of the front and rear suspensions, respectively. and denote the vertical velocities of the front and rear unsprung masses, Indicates the vehicle body pitch angular velocity; The potential energy V of the vehicle is generated by the deformation of the springs and tires and is expressed as: Among them, k1 and k2 represent the suspension stiffness of the front and rear suspension of the vehicle respectively, k t1 and k t2 Denote the radial stiffness of the front and rear tires, z s represents the vertical displacement of the sprung mass center of mass, a and b represent the distance from the center of mass to the front and rear axles, respectively, and z u1 and z u2 They represent the vertical displacement of the front and rear unsprung masses, and θ represents the vehicle body pitch angle; The dissipated energy D of the vehicle is generated by the damper. Since the damping value of the tire is very small, the damping characteristics of the tire are not considered. Therefore, the dissipated energy is expressed as: Where c1 and c2 represent the suspension damping coefficients of the front and rear suspensions, respectively; Define the state quantity q=[z s θz u1 z u2 ] T , the vehicle's motion equation can be expressed by the Lagrange equation as: Where Q represents the generalized force, Represents state quantity The first-order derivative of time, Q, is solved by the principle of virtual work as shown below: M y =aF zf -bF zr Among them, δW represents virtual work, δq represents virtual displacement, F zf and F zr Denote the forces acting on the front and rear suspensions, δz s and δθ represent the virtual displacement at the center of mass of the sprung mass and the virtual displacement of the pitch angle, respectively, δz uf and δz ur Denote the virtual displacement of the front and rear suspension unsprung masses, M y represents the pitching moment; Through the above derivation, the ordinary differential equations of vehicle pitch and vertical motion based on Lagrangian dynamics are obtained: in, Represents the second-order derivative of the state quantity with respect to time, that is, the acceleration state vector, M represents the mass matrix, C represents the damping matrix, and K represents the stiffness matrix; S3, constructing pitch and vertical motion models based on the fusion of neural networks and physical models; First, the neural network uses the vehicle state q as the input of the neural network. The number of hidden layers of the neural network is 2, the number of neurons in the hidden layer is 64, and the number of neurons in the output layer of the neural network is 30. The operation process is shown in the following formula: z i =W i h i-1 +b i h i =g(z i ) Among them, W i represents the weight matrix of the i-th layer neural network, b i represents the bias matrix of the i-th layer neural network, h i-1 represents the output matrix of the i-1th layer neural network, z i represents the weighted input of the i-th layer, g represents the activation function, h i Represents the output of the i-th layer neural network; Among the 30 neurons in the output layer, the 1st to 10th output neurons constitute l M , the outputs of neurons 11 to 20 constitute l C , the outputs of neurons 21 to 30 constitute l K , the hidden activation function of the neural network is softplus, and the activation function of the output layer is l M 、l C 、l K l in d The activation function of the element is softplus, l M 、l C 、l K l in o The activation function of the element is a linear activation function, as shown below: l M ＝(l dM (1),l oM (1),l dM (2),l oM (2),l oM (3),l dM (3),l oM (4),l oM (5),l oM (6),l dM (4)) T l C ＝(l dC (1),l oC (1),l dC (2),l oC (2),l oC (3),l dC (3),l oC (4),l oC (5),l oC (6),l dC (4)) T l K ＝(l dK (1),l oK (1),l dK (2),l oK (2),l oK (3),l dK (3),l oK (4),l oK (5),l oK (6),l dK (4)) T l M 、l C 、l K After output, it will be transformed into the lower triangular matrix L M 、L C 、L K , l M 、l C 、l K l in d are the diagonal elements of the matrix, l o are off-diagonal elements, L M 、L C 、L K The transpose of L M 、L C 、L K The product of can get the symmetric positive definite matrix They are called predicted mass, predicted damping and predicted stiffness matrices respectively. The predicted mass matrix and acceleration state Damping Matrix and speed state Stiffness matrix Multiplying by q and summing them up will give the system external force predicted by the neural network. Instead of this As shown in the following formula: When training a neural network, the external force is predicted by minimizing The difference between the actual suspension force Q and the actual suspension force Q is used to optimize the weight of the neural network, as shown in the following formula in, Represents the inverse function, that is, constructing the state quantity q, The mapping relationship with the generalized force Q, ζ represents the parameters in the neural network, such as weight W and bias b; S4. Design the upper controller, including the roll controller and the pitch and vertical controller; Design a roll controller based on the double-track vehicle roll model derived from the Newton-Euler method described above, and use sliding mode control to design the roll controller; First, the control goal is to reduce the vehicle's roll motion so that the vehicle roll angle is 0, and the generalized moment M x The sliding mode controller is designed as the control variable, where the roll angular acceleration can be expressed as: in, represents the roll angle, represents the roll angular velocity, k φ represents the equivalent roll stiffness of the suspension, c φ represents the equivalent damping of the suspension, a y represents lateral acceleration, g represents acceleration due to gravity; Design sliding surface s: Where λ is greater than 0, The reaching law design is Where ε>0, so the sliding mode control rate for vehicle roll control can be expressed as: Among them, M xdes represents the desired roll control torque solved by the sliding mode controller; Design pitch and vertical controllers. The goal of the pitch and vertical motion controllers is to reduce the pitch motion and vertical oscillation of the vehicle so that the pitch acceleration and vertical acceleration of the vehicle are 0. The control target The vehicle pitch and vertical motion can be expressed by the ordinary differential equations mentioned above, where the vehicle's mass matrix, stiffness matrix, and damping matrix are replaced by the predicted mass matrix, predicted stiffness matrix, and predicted damping matrix established by the neural network. The expression is as follows: Define the sliding surface σ: Among them, Λ is greater than 0, so The reaching law is calculated as Where k>0, so the sliding mode control rate for pitch and vertical control is expressed as: Among them, F zdes represents the desired vertical control force solved by the sliding mode controller, M ydes represents the desired pitch control torque solved by the sliding mode controller; S5. Design optimal allocation and adaptive controller; The optimal allocation problem is solved using quadratic programming, and its expression is as follows: Where γ>0 and is large enough, v represents the desired control quantity solved by the sliding mode controller, B represents the control conversion matrix, which is used to construct the mapping relationship between the suspension force and the virtual control signal, u represents the actual control force of the suspension, and W v (α) and W u Both represent weight matrices, and They represent the upper and lower limit constraints of the control quantity, u0 represents the damping force solved by the ceiling damping control method, F sky The skyhook control theory is based on the vertical velocity of the sprung mass and the relative velocity between the sprung and unsprung masses The solved equilibrium damping force, where the specific expressions of each quantity are as follows: V=[M xdes M ydes F zdes ] T u=[F zfl F zfr F zrl F zrr ] T u0=[u 1,eq u 2,eq u 3,eq u 4,eq ] T Among them, H represents the left and right wheelbase of the vehicle, F zfl and F zfr Represents the forces acting on the left and right suspensions of the front suspension, F zrl 、F zrr Respectively represent the forces acting on the left and right suspensions of the rear suspension, u i,eq represents the equilibrium damping force of each suspension based on the skyhook control theory, where i = 1, 2, 3, 4, representing the front left, front right, rear left, and rear right, respectively, and c sky represents the ceiling damping coefficient; The roll control moment, pitch control moment, and vertical force generated by the forces generated by the four front and rear suspensions can be expressed as follows: M y =a(F zfl +F zfr )-b(F zrl +F zrr ) F z =F zfl +F zfr +F zrl +F zrr Among them, represents the vertical force generated by the four suspensions; The purpose of the first term in the cost function is to minimize the distribution error, that is, to minimize the error between the roll moment and pitch moment generated by the actual force generated by the suspension and the expected value solved by the sliding mode control. The u0 in the second term of the cost function is the equilibrium point constructed based on the skyhook control theory. When the weight of the first term in the cost function is very small and can be ignored, the cost function only has the second term, and the entire control system is simplified to the independent control of the four suspensions based on the skyhook control theory. Design the adaptive control weight, adaptive weight W v The expression of (α) is as follows: in, represents the variable weight of roll control, represents the variable weight of pitch control, represents the vertical control variable weight; The value of is determined by the following formula: Among them, ∨ represents and The maximum of the two; The value of is determined by the roll angle and lateral acceleration of the vehicle. The thresholds of the roll angle and lateral acceleration are selected first, where the threshold of the roll angle is is the critical rollover angle φ c The threshold value of lateral acceleration is half of Normalize the roll angle and lateral acceleration, and map the normalized signal to the interval [0,1] after passing the function h. Then sum the two signals and limit them between 0 and 1. The summed signal can be obtained by smoothing and filtering. The expression is: Similarly, The value of is determined by the pitch angle and longitudinal acceleration of the vehicle. The thresholds of pitch angle and longitudinal acceleration are selected first, where the threshold of pitch angle is The threshold value of longitudinal acceleration is Normalize the pitch angle and longitudinal acceleration, and map the normalized signal to the interval [0,1] after passing the function h. Then sum the two signals and limit them between 0 and 1. The summed signal can be obtained by smoothing and filtering. Expressed as:

Images