Automobile suspension pre-control method and system based on LQR and artificial neural network

Abstract

The invention provides an automobile suspension pre-control method and system based on an LQR and an artificial neural network, and the method comprises the steps: obtaining the road surface undulation degree information in advance through a binocular camera, building a dynamic model of a suspension system through the combination of vehicle attitude parameters and motion parameters obtained by a vehicle-mounted sensor, and carrying out the expression through a linear space equation. Thirdly, designing an artificial neural network, and calculating a weighting coefficient matrix according to the vehicle parameters and the road surface height; and optimal suspension parameters are solved based on a linear quadratic regulator (LQR), and pre-control of the suspension is achieved. And calculating the posture of the suspension after deformation through a precise discretization method, and carrying out online supervision training on the artificial neural network. According to the method, the suspension deformation can be calculated in advance, the road impact is effectively absorbed, and the driving safety and comfort are improved. Compared with traditional real-time feedback control, the method has higher environmental adaptability and reliability.

Description

Technical Field

[0001] The invention relates to the technical field of automobiles, and in particular to an automobile suspension pre-control method based on LQR and artificial neural network. Background Art

[0002] As an important part of the vehicle, the performance of the suspension system of the vehicle is directly related to the driving performance of the vehicle and the comfort of the passengers. Excellent suspension control can effectively absorb road impact and reduce the impact of vibration on the vehicle body, thereby improving driving safety and comfort. With the continuous advancement of automobile technology, the intelligence and automation of the suspension system has gradually become a development trend. The realization of dynamic adjustment of suspension characteristics according to complex road conditions to achieve the best driving effect has become the focus of current research.

[0003] However, most current suspension control methods are limited to real-time feedback control, that is, the suspension control strategy is determined only based on the data collected by the sensors at the current moment. This method may significantly reduce the riding experience when facing a bumpy road because the suspension cannot deform in time. In order to overcome this limitation, the industry urgently needs a method that can calculate the suspension deformation in advance and realize suspension pre-control.

[0004] Although some related patents have proposed improved suspension control methods, these methods still have shortcomings. For example, the semi-active suspension control method of the vehicle proposed in patent document CN114919365B (application number: CN202210741404.0) can determine whether it is entering a special working condition by obtaining the vertical acceleration of the wheel, and control the working state of the wheel shock absorber accordingly, but it does not know the road surface information in advance and cannot make pre-control of the suspension posture. The active suspension control method proposed in another patent document CN110901325B (application number: CN201911197599.1) can observe the unevenness curve of the road ahead and estimate the expected value of the body posture that the vehicle should maintain, but its subsequent control is based on the classic PID algorithm, resulting in large differences in control performance under different road environments.

[0005] Therefore, it is necessary to propose a new technical solution to improve the above technical problems. Summary of the invention

[0006] In view of the defects in the prior art, the object of the present invention is to provide a vehicle suspension pre-control method and system based on LQR and artificial neural network.

[0007] According to the present invention, a vehicle suspension pre-control method based on LQR and artificial neural network is provided, and the method comprises the following steps:

[0008] Step S1: Based on binocular depth estimation technology, the undulation information of the road ahead is obtained in advance, and all posture parameters and motion parameters of the vehicle and suspension system are obtained in real time by the on-board sensors;

[0009] Step S2: construct a dynamic model of the suspension system according to the acquired parameters and describe it through a state space expression;

[0010] Step S3: designing an artificial neural network, which takes vehicle parameters and road height as input, calculates and outputs a weight coefficient matrix for LQR control by learning;

[0011] Step S4: using the LQR method, combined with the weighted coefficient matrix output by the artificial neural network, to solve the feedback gain matrix, thereby determining the magnitude of the active control external force, and predicting the posture of the suspension after deformation based on the precise discretization method, which is used for online supervised training of the artificial neural network;

[0012] Step S5: According to the LQR solution result, dynamically adjust the air spring stiffness of the suspension, and repeat steps S1 to S5 for continuous pre-control.

[0013] Preferably, the road surface undulation information in step S1 is calculated based on the depth image acquired by the binocular camera, while taking into account the current vehicle speed to determine the time it takes for the tire to reach each road surface undulation position.

[0014] Preferably, when constructing the dynamic model of the suspension system in step S2, the sprung mass, unsprung mass, air spring equivalent stiffness, damper, active control external force, tire equivalent stiffness, and sprung mass displacement and unsprung mass displacement parameters are considered.

[0015] Preferably, the artificial neural network designed in step S3 has a multi-layer fully connected structure, including an input layer, at least three hidden layers and an output layer, wherein the hidden layers are activated using a ReLU function, and the output layer is designed according to requirements to generate a specific dimension of a weight coefficient matrix;

[0016] The artificial neural network output layer includes two sub-output layers, one for generating a symmetric semi-positive definite weighted sparse matrix, and the other for outputting a symmetric positive definite weighted coefficient matrix after being activated by a Softplus function.

[0017] Preferably, when solving the Riccati equation to obtain a matrix in step S4, the weight coefficient matrix output by the artificial neural network is used as a part of the parameters of the equation to perform LQR parameter adjustment based on learning;

[0018] When predicting the posture of the suspension after deformation based on the precise discretization method, the matrix, state variables and optimal control external force in the current state are considered, and the Taylor series expansion is used for approximate calculation;

[0019] When performing online supervised training of the artificial neural network, the smooth L1 loss function is used to minimize the vertical posture change of the vehicle between the previous and next moments.

[0020] Preferably, when dynamically adjusting the stiffness of the air spring of the suspension in step S5, the changes in the air cavity volume and the effective force-bearing area of ​​the air spring after the suspension posture changes are taken into consideration to calculate a new equivalent stiffness.

[0021] The present invention also provides a vehicle suspension pre-control system based on LQR and artificial neural network, the system comprising the following modules:

[0022] Module M1: Based on binocular depth estimation technology, it obtains the undulation information of the road ahead in advance, and combines all the posture parameters and motion parameters of the vehicle and suspension system obtained in real time by the on-board sensors;

[0023] Module M2: Based on the acquired parameters, the dynamic model of the suspension system is constructed and described by the state space expression;

[0024] Module M3: Design an artificial neural network that takes vehicle parameters and road height as input, calculates and outputs the weighting coefficient matrix for LQR control through learning;

[0025] Module M4: Using the LQR system, combined with the weighted coefficient matrix output by the artificial neural network, the feedback gain matrix is ​​solved to determine the magnitude of the active control external force, and the posture of the suspension after deformation is predicted based on the precise discretization system, which is used for online supervised training of the artificial neural network;

[0026] Module M5: Dynamically adjust the air spring stiffness of the suspension according to the LQR solution results, and repeat modules M1 to S5 for continuous pre-control.

[0027] Preferably, the road surface undulation information in the module M1 is calculated based on the depth image acquired by the binocular camera, while taking into account the current vehicle speed to determine the time when the tire reaches each road surface undulation position;

[0028] When constructing the dynamic model of the suspension system in the module M2, the sprung mass, unsprung mass, air spring equivalent stiffness, damper, active control external force, tire equivalent stiffness, and sprung mass displacement and unsprung mass displacement parameters are considered.

[0029] Preferably, the artificial neural network designed in the module M3 has a multi-layer fully connected structure, including an input layer, at least three hidden layers and an output layer, wherein the hidden layers are activated using a ReLU function, and the output layer is designed according to requirements to generate a specific dimension of a weighted coefficient matrix;

[0030] The artificial neural network output layer includes two sub-output layers, one for generating a symmetric semi-positive definite weighted sparse matrix, and the other for outputting a symmetric positive definite weighted coefficient matrix after being activated by a Softplus function.

[0031] Preferably, when solving the Riccati equation to obtain a matrix in the module M4, the weight coefficient matrix output by the artificial neural network is used as a part of the parameters of the equation to perform LQR parameter adjustment based on learning;

[0032] When predicting the posture of the suspension after deformation based on the accurate discretization system, the matrix, state variables and optimal control external force in the current state are considered, and the Taylor series expansion is used for approximate calculation;

[0033] When performing online supervised training of artificial neural networks, the smooth L1 loss function is used to minimize the vertical posture change of the vehicle between the previous and next moments.

[0034] When dynamically adjusting the stiffness of the air spring of the suspension in the module M5, the changes in the air cavity volume and the effective force-bearing area of ​​the air spring after the suspension posture changes are considered, and a new equivalent stiffness is calculated.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The present invention proposes a new automobile suspension pre-control method. By obtaining the road undulation information in advance and combining the current vehicle and suspension parameters, the optimal suspension posture at the future moment can be calculated in advance, so that the suspension can be deformed earlier, thereby more effectively absorbing the road impact and reducing the impact of vibration on the vehicle body, significantly improving driving safety and passenger comfort, especially in the face of complex and changeable road environments. The effect is particularly obvious;

[0037] 2. The present invention adopts an LQR parameter matrix adjustment method based on artificial neural network. By learning, the value of the weighting coefficient matrix can be adaptively adjusted according to the current parameters of the vehicle and suspension, so that the suspension system can better adapt to different road conditions and driving conditions, thereby improving the environmental adaptability and reliability of the suspension;

[0038] 3. After experimental testing, the suspension pre-control method of the present invention has demonstrated significant control effects. When the pre-control method is used, the overall control effect of the suspension is significantly stronger than that without pre-control. This not only verifies the effectiveness and feasibility of the present invention, but also further proves its great potential in improving the overall performance of the automobile suspension system. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0040] Figure 1 This is a flowchart of the implementation of the method described in the embodiment of the present invention;

[0041] Figure 2 is a diagram of an equivalent model of a suspension according to an embodiment of the present invention;

[0042] Figure 3 A graph showing the variation of road surface undulation in a control experiment according to an embodiment of the present invention;

[0043] Figure 4 4 is a comparison diagram of suspension control results of an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0045] Embodiment 1:

[0046] Reference Figure 1 According to the present invention, a vehicle suspension pre-control method based on LQR and artificial neural network is provided, and the method comprises the following steps:

[0047] Step S1: Based on the binocular depth estimation technology, the undulation information of the road ahead is obtained in advance, and combined with all the posture parameters and motion parameters of the vehicle and suspension system obtained in real time by the on-board sensors; the road undulation information is calculated based on the depth image obtained by the binocular camera, while taking into account the current vehicle speed to determine the time it takes for the tire to reach each road undulation position.

[0048] Step S2: construct a dynamic model of the suspension system based on the acquired parameters and describe it through a state space expression; when constructing the dynamic model of the suspension system, the sprung mass, unsprung mass, air spring equivalent stiffness, damper, active control external force, tire equivalent stiffness, and sprung mass displacement and unsprung mass displacement parameters are considered.

[0049] Step S3: Design an artificial neural network, which takes vehicle parameters and road height as input, calculates and outputs a weight coefficient matrix for LQR control by learning; the designed artificial neural network has a multi-layer fully connected structure, including an input layer, at least three hidden layers and an output layer, wherein the hidden layers are activated by ReLU functions, and the output layer is designed according to requirements to generate a specific dimension of the weight coefficient matrix; the output layer of the artificial neural network includes two sub-output layers, one for generating a symmetric semi-positive definite weighted sparse matrix, and the other outputs a symmetric positive definite weight coefficient matrix after being activated by a Softplus function.

[0050] Step S4: using the LQR method, combined with the weighted coefficient matrix output by the artificial neural network, the feedback gain matrix is ​​solved to determine the size of the active control external force, and the posture of the suspension after deformation is predicted based on the precise discretization method, which is used for online supervised training of the artificial neural network; when solving the Riccati equation to obtain the matrix, the weighted coefficient matrix output by the artificial neural network is used as part of the parameters of the equation to perform learning-based LQR parameter adjustment; when predicting the posture of the suspension after deformation based on the precise discretization method, the matrix, state variables and optimal control external force in the current state are considered, and Taylor series expansion is used for approximate calculation; when performing online supervised training of the artificial neural network, the smooth L1loss function is used to minimize the change in the vehicle's posture in the vertical direction between the previous and next moments.

[0051] Step S5: According to the LQR solution, dynamically adjust the air spring stiffness of the suspension, and repeat steps S1 to S5 for continuous pre-control; when dynamically adjusting the air spring stiffness of the suspension, consider the changes in the air cavity volume and effective force area of ​​the air spring after the suspension posture changes, and calculate the new equivalent stiffness.

[0052] The present invention also provides a vehicle suspension pre-control system based on LQR and artificial neural network. The vehicle suspension pre-control system based on LQR and artificial neural network can be realized by executing the process steps of the vehicle suspension pre-control method based on LQR and artificial neural network, that is, those skilled in the art can understand the vehicle suspension pre-control method based on LQR and artificial neural network as a preferred implementation of the vehicle suspension pre-control system based on LQR and artificial neural network.

[0053] Embodiment 2:

[0054] The present invention also provides a vehicle suspension pre-control system based on LQR and artificial neural network, the system comprising the following modules:

[0055] Module M1: Based on binocular depth estimation technology, the undulation information of the road ahead is acquired in advance, and all the attitude parameters and motion parameters of the vehicle and suspension system acquired in real time by the on-board sensors are combined; the road undulation information is calculated based on the depth image acquired by the binocular camera, and the current vehicle speed is taken into account to determine the time it takes for the tire to reach each road undulation position;

[0056] Module M2: Based on the acquired parameters, a dynamic model of the suspension system is constructed and described through a state-space expression; when constructing the dynamic model of the suspension system, the sprung mass, unsprung mass, air spring equivalent stiffness, damper, active control external force, tire equivalent stiffness, and sprung mass displacement and unsprung mass displacement parameters are considered.

[0057] Module M3: Design an artificial neural network that takes vehicle parameters and road height as input, calculates and outputs a weight coefficient matrix for LQR control by learning; the designed artificial neural network has a multi-layer fully connected structure, including an input layer, at least three hidden layers and an output layer, wherein the hidden layers are activated using the ReLU function, and the output layer is designed according to requirements to generate a specific dimension of the weight coefficient matrix; the output layer of the artificial neural network includes two sub-output layers, one for generating a symmetric semi-positive definite weighted sparse matrix, and the other outputs a symmetric positive definite weight coefficient matrix after being activated by the Softplus function.

[0058] Module M4: Using the LQR system, combined with the weighted coefficient matrix output by the artificial neural network, the feedback gain matrix is ​​solved to determine the size of the active control external force, and the posture of the suspension after deformation is predicted based on the precise discretization system, which is used for online supervised training of the artificial neural network; when solving the Riccati equation to obtain the matrix, the weighted coefficient matrix output by the artificial neural network is used as part of the equation parameters to perform learning-based LQR parameter adjustment; when predicting the posture of the suspension after deformation based on the precise discretization system, the matrix, state variables and optimal control external force in the current state are considered, and Taylor series expansion is used for approximate calculation; when conducting online supervised training of the artificial neural network, the smooth L1loss function is used to minimize the change in the vehicle's posture in the vertical direction between the previous and next moments;

[0059] Module M5: According to the LQR solution results, dynamically adjust the air spring stiffness of the suspension, and repeat modules M1 to S5 for continuous pre-control; when dynamically adjusting the air spring stiffness of the suspension, consider the changes in the air cavity volume and effective force area of ​​the air spring after the suspension posture changes, and calculate the new equivalent stiffness.

[0060] Embodiment 3:

[0061] The present invention relates to the field of automobile technology, and provides a method for automobile suspension pre-control based on LQR and artificial neural network. The method comprises: obtaining road surface undulation information in advance based on a binocular camera, and then obtaining the posture parameters and motion parameters of the vehicle from the vehicle-mounted sensor, jointly establishing a dynamic model of the vehicle suspension system, and expressing it using a linear space equation. Subsequently, an artificial neural network is designed, and a weighted coefficient matrix is ​​calculated in a learning manner according to the vehicle parameters and the road surface height, and the optimal suspension parameters are solved based on a linear quadratic regulator (LQR) to achieve suspension pre-control, and then the posture of the suspension after deformation is calculated by an accurate discretization method, thereby achieving online supervised training of the artificial neural network.

[0062] The present invention provides a vehicle suspension pre-control method based on LQR and artificial neural network, which aims to obtain the undulation information of the road surface in advance, and then combine all the parameters of the vehicle and the suspension system at the current moment, use the artificial neural network to calculate the weighted coefficient matrix based on the learning method, and then optimally control the suspension system according to the linear quadratic regulator to ensure driving safety and comfort.

[0063] The embodiment of the present invention provides a vehicle suspension pre-control method based on LQR and artificial neural network, such as Figure 1 As shown, the method comprises the following steps:

[0064] S1: Based on the binocular depth estimation method, the road surface undulation information is obtained, and based on the on-board sensors, all the posture parameters and motion parameters of the vehicle and suspension system are obtained.

[0065] As a preferred embodiment of the present invention, the road surface undulation information obtained in step S1 is denoted as z r , at the current speed, the tire travels to a road surface with a roughness of z r The time required to reach the position is recorded as t; all the attitude parameters and motion parameters of the vehicle and suspension system including the sprung mass m s , unsprung mass m u , air spring equivalent stiffness k s , damper b s , actively control the external force F a , tire equivalent stiffness k t , sprung mass displacement zs , unsprung mass displacement z u .

[0066] S2: Construct a dynamic model of the suspension system and establish a state space expression by selecting appropriate state variables.

[0067] As a preferred embodiment of the present invention, Figure 2 As shown, the dynamic model of the suspension system described in step S2 is constructed, and the state space expression is established by selecting appropriate state variables. The specific implementation method is:

[0068] S21: Based on all the parameters obtained in step S1, according to Newton's kinematics principle, the dynamic expression of the suspension system can be obtained:

[0069]

[0070] S22: Since the system has a total of 4 energy storage elements, the number of state variables is also 4, and x1=z s -z u , x3=z u -z r , The dynamic expression is thus transformed into a state space expression: Where x represents a vector consisting of four state variables, the external force F a That is, it represents the control quantity of the system, A, B, and M represent 4×4, 4×1, and 4×1 matrices respectively. The specific expressions of each part are as follows:

[0071]

[0072] S3: Design an artificial neural network to calculate the weighted coefficient matrix in a learning manner based on vehicle parameters and road height.

[0073] As a preferred embodiment of the present invention, the artificial neural network designed in step S3 calculates the weight coefficient matrix in a learning manner according to the vehicle parameters and the road surface height, and its specific implementation method is:

[0074] S31: The input information of the network is a 10-dimensional vector, which contains x1, x2, x3, x4, m s ,m u ,k s ,k t ,b s ,z r , then construct three fully connected layers with hidden layer dimensions of 64, 256, and 512 respectively, and use the ReLU function to activate between every two fully connected layers, and finally obtain a 512-dimensional intermediate output vector

[0075] S32: Design fully connected layers with output dimensions of 16 and 1 respectively. The former is used to generate a 4×4 matrix Q tmp After all, one step through The weighted sparse matrix Q is calculated, and after being activated by the Softplus function, the weighted coefficient matrix R is output.

[0076] More specifically, the Q matrix outputted in step S32 is a symmetric positive semi-definite matrix, and the R matrix is ​​a symmetric positive definite matrix.

[0077] S4: Based on the LQR method, the feedback gain matrix is ​​solved to obtain the size of the active control external force, and based on the precise discretization method, the posture of the suspension after deformation is calculated for supervised training of the artificial neural network.

[0078] As a preferred embodiment of the present invention, the feedback gain matrix is ​​obtained by solving the LQR method described in step S4, thereby obtaining the magnitude of the active control external force, and the posture of the suspension after deformation is calculated based on the accurate discretization method for supervised training of the artificial neural network. The specific implementation method is as follows:

[0079] S41: Solve the P matrix in the Riccati equation according to the weighted coefficient matrices Q and R obtained in step S3. The specific equation expression is as follows:

[0080] PA+A T P+Q-PBR -1 B T P=0

[0081] S42: Based on the obtained P matrix, further obtain the feedback gain matrix K=R -1 B T P, and the optimal external force F under the current state a =-Kx.

[0082] S43: Based on the A, B matrix, state variable x and optimal control external force F in the current state a Based on the idea of ​​precise discretization, the new state variable x of the suspension after time t is further derived. next , the specific calculation method is as follows:

[0083] G=e At , H = ∫0te Aτ dτ·B

[0084] x next =Gx+HF a

[0085] More specifically, step S43 calculates eAt Taylor series expansion can be used to select the first several items for approximate calculation according to the actual accuracy requirements.

[0086] S44: In order to minimize the change in the vertical posture of the vehicle between the previous and the next moments, the smooth L1 loss is used to supervise the training of the designed artificial neural network. The specific loss function expression is as follows:

[0087]

[0088] S5: According to the solution results, the spring stiffness of the air spring of the suspension is updated, and all the posture parameters and motion parameters of the vehicle and suspension system are re-acquired for the next round of pre-control.

[0089] As a preferred embodiment of the present invention, the specific method of updating the spring stiffness of the air spring of the suspension according to the solution result in step S5 is: when the suspension posture changes, according to the current air cavity volume V of the air spring, e and the current effective force area A e , calculate the new equivalent stiffness k at the current moment t , the specific calculation formula is as follows:

[0090]

[0091] The present invention realizes a vehicle suspension pre-control method based on LQR and artificial neural network, and the test shows that when the pre-control method is used, the overall control effect of the suspension is significantly stronger than that without pre-control. The specific experimental road environment and test results are as follows: Figure 3 and Figure 4 In view of the great significance of active suspension control in ensuring driving safety and passenger comfort, the present invention has important practical value, and is particularly suitable for improving the safety and efficiency of automobile driving systems and improving vehicle driving comfort.

[0092] Those skilled in the art may understand this embodiment as a more specific description of Embodiment 1 and Embodiment 2.

[0093] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0094] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A vehicle suspension pre-control method based on LQR and artificial neural network, characterized in that: The method comprises the following steps: Step S1: Based on binocular depth estimation technology, the undulation information of the road ahead is obtained in advance, and all posture parameters and motion parameters of the vehicle and suspension system are obtained in real time by the on-board sensors; Step S2: construct a dynamic model of the suspension system according to the acquired parameters and describe it through a state space expression; Step S3: designing an artificial neural network, which takes vehicle parameters and road height as input, calculates and outputs a weight coefficient matrix for LQR control by learning; Step S4: using the LQR method, combined with the weighted coefficient matrix output by the artificial neural network, to solve the feedback gain matrix, thereby determining the magnitude of the active control external force, and predicting the posture of the suspension after deformation based on the precise discretization method, which is used for online supervised training of the artificial neural network; Step S5: According to the LQR solution result, dynamically adjust the air spring stiffness of the suspension, and repeat steps S1 to S5 for continuous pre-control.

2. The automobile suspension pre-control method based on LQR and artificial neural network according to claim 1 is characterized in that: The road surface undulation information in step S1 is calculated based on the depth image acquired by the binocular camera, while taking the current vehicle speed into consideration to determine the time when the tire reaches each road surface undulation position.

3. The automobile suspension pre-control method based on LQR and artificial neural network according to claim 1 is characterized in that: When constructing the dynamic model of the suspension system in step S2, the sprung mass, unsprung mass, air spring equivalent stiffness, damper, active control external force, tire equivalent stiffness, and sprung mass displacement and unsprung mass displacement parameters are considered.

4. The automobile suspension pre-control method based on LQR and artificial neural network according to claim 1 is characterized in that: The artificial neural network designed in step S3 has a multi-layer fully connected structure, including an input layer, at least three hidden layers and an output layer, wherein the hidden layers are activated using a ReLU function, and the output layer is designed according to requirements to generate a specific dimension of a weighted coefficient matrix; The artificial neural network output layer includes two sub-output layers, one for generating a symmetric semi-positive definite weighted sparse matrix, and the other for outputting a symmetric positive definite weighted coefficient matrix after being activated by a Softplus function.

5. The automobile suspension pre-control method based on LQR and artificial neural network according to claim 1, characterized in that: When solving the Riccati equation to obtain a matrix in step S4, the weighted coefficient matrix output by the artificial neural network is used as a part of the parameters of the equation to perform LQR parameter adjustment based on learning; When predicting the posture of the suspension after deformation based on the precise discretization method, the matrix, state variables and optimal control external force in the current state are considered, and the Taylor series expansion is used for approximate calculation; When performing online supervised training of the artificial neural network, the smooth L1 loss function is used to minimize the vertical posture change of the vehicle between the previous and next moments.

6. The automobile suspension pre-control method based on LQR and artificial neural network according to claim 1, characterized in that: When dynamically adjusting the stiffness of the air spring of the suspension in step S5, the changes in the air cavity volume and the effective force-bearing area of ​​the air spring after the suspension posture changes are considered to calculate a new equivalent stiffness.

7. A vehicle suspension pre-control system based on LQR and artificial neural network, characterized in that: The system includes the following modules: Module M1: Based on binocular depth estimation technology, it obtains the undulation information of the road ahead in advance, and combines all the posture parameters and motion parameters of the vehicle and suspension system obtained in real time by the on-board sensors; Module M2: Based on the acquired parameters, the dynamic model of the suspension system is constructed and described by the state space expression; Module M3: Design an artificial neural network that takes vehicle parameters and road height as input, calculates and outputs the weighting coefficient matrix for LQR control through learning; Module M4: Using the LQR system, combined with the weighted coefficient matrix output by the artificial neural network, the feedback gain matrix is ​​solved to determine the magnitude of the active control external force, and the posture of the suspension after deformation is predicted based on the precise discretization system, which is used for online supervised training of the artificial neural network; Module M5: Dynamically adjust the air spring stiffness of the suspension according to the LQR solution results, and repeat modules M1 to S5 for continuous pre-control.

8. The automobile suspension pre-control system based on LQR and artificial neural network according to claim 7, characterized in that: The road surface undulation information described in the module M1 is calculated based on the depth image obtained by the binocular camera, while taking into account the current vehicle speed to determine the time when the tire reaches each road surface undulation position; When constructing the dynamic model of the suspension system in the module M2, the sprung mass, unsprung mass, air spring equivalent stiffness, damper, active control external force, tire equivalent stiffness, and sprung mass displacement and unsprung mass displacement parameters are considered.

9. The automobile suspension pre-control system based on LQR and artificial neural network according to claim 7, characterized in that: The artificial neural network designed in the module M3 has a multi-layer fully connected structure, including an input layer, at least three hidden layers and an output layer, wherein the hidden layers are activated using a ReLU function, and the output layer is designed according to requirements to generate a specific dimension of a weighted coefficient matrix; The artificial neural network output layer includes two sub-output layers, one for generating a symmetric semi-positive definite weighted sparse matrix, and the other for outputting a symmetric positive definite weighted coefficient matrix after being activated by a Softplus function.

10. The automobile suspension pre-control system based on LQR and artificial neural network according to claim 7, characterized in that: When solving the Riccati equation to obtain a matrix in the module M4, the weighted coefficient matrix output by the artificial neural network is used as a part of the parameters of the equation to perform LQR parameter adjustment based on learning; When predicting the posture of the suspension after deformation based on the accurate discretization system, the matrix, state variables and optimal control external force in the current state are considered, and the Taylor series expansion is used for approximate calculation; When performing online supervised training of artificial neural networks, the smooth L1 loss function is used to minimize the vertical posture change of the vehicle between the previous and next moments. When dynamically adjusting the stiffness of the air spring of the suspension in the module M5, the changes in the air cavity volume and the effective force-bearing area of ​​the air spring after the suspension posture changes are considered, and a new equivalent stiffness is calculated.

Images