A full-band vibration suppression method for active suspension

Abstract

This invention relates to a method for full-frequency vibration suppression of active suspension, comprising: constructing a high-frequency dynamic model of the vehicle to obtain vehicle state information; based on the vehicle state information, establishing a model predictive controller and a reinforcement learning controller respectively, and fusing the reinforcement learning controller and the model predictive controller to obtain the active suspension control input; establishing a safety supervision mechanism to perform constraint monitoring on the active suspension control input, and switching to independent model predictive control when a risk is triggered, using only the output of the model predictive controller as the active suspension control input. Compared with the prior art, this invention combines the predictive optimization and constraint handling advantages of model predictive control with the adaptive capability of reinforcement learning, enabling coordinated unification of low-frequency vibration suppression and mid-to-high-frequency compensation control, thereby improving the comprehensive vibration control performance of the vehicle's active suspension across the entire frequency band.

Description

Technical Field

[0001] This invention relates to the field of active suspension control technology for vehicles, and in particular to a method for suppressing vibration across the entire frequency band of an active suspension. Background Technology

[0002] With the rapid development of the automotive industry towards electrification and intelligence, steer-by-wire chassis technologies, including brake steer-by-wire, steering steer-by-wire, and electronically controlled suspension systems, have become key technologies for improving vehicle dynamics and driving safety. Among them, electronically controlled suspension (ECS) has received widespread attention due to its ability to significantly improve vehicle ride comfort and handling stability. ECS systems play a central role in vehicle vibration control, road feel optimization, and intelligent driving assistance, and market demand continues to grow.

[0003] Most current suspension control research focuses on suppressing low-frequency vibrations (0–20 Hz) to improve comfort and handling. However, recent studies have found that while ECS (Electronic Control System) suppresses low-frequency vibrations, it easily introduces mid-to-high-frequency vibrations (>20 Hz). For example, the stepped control signal generated by discontinuities in the control signal, sampling delays, and actuator bandwidth limitations can excite high-frequency resonances in the unsprung mass and vehicle structure. These high-frequency vibrations not only affect the vehicle's handling stability but also transmit them to the cabin through the body structure, increasing in-cabin noise and thus reducing overall ride comfort.

[0004] Furthermore, existing technologies apply reinforcement learning (RL) to suspension control. Reinforcement learning optimizes policies by interacting with the environment and can improve control performance without the need for an accurate model. However, in practical applications, it is easy to generate control forces that change at high frequencies, thereby triggering mid-to-high frequency vibration problems in the suspension system.

[0005] In summary, how to effectively suppress mid-to-high frequency vibrations while suppressing low-frequency vibrations, i.e., to achieve full-frequency vibration suppression, is a problem that urgently needs to be solved in the current active suspension control of vehicles. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art by providing an active suspension full-frequency vibration suppression method that can achieve synergistic optimization of low-frequency ride comfort and mid-to-high-frequency vibration suppression, thereby improving the overall vibration control performance of the active suspension across the entire frequency range.

[0007] The objective of this invention can be achieved through the following technical solution: a method for suppressing vibration across the entire frequency band of an active suspension, comprising the following steps:

[0008] S1. Construct a high-frequency dynamic model of the vehicle to obtain vehicle state information;

[0009] S2. Based on vehicle state information, establish a model predictive controller and a reinforcement learning controller respectively, and fuse the reinforcement learning controller and the model predictive controller to output the active suspension control input of the vehicle.

[0010] S3. Establish a safety supervision mechanism to monitor the constraints of the vehicle's active suspension control input, and switch to separate model predictive control when a risk is triggered, using only the output of the model predictive controller as the vehicle's active suspension control input.

[0011] Furthermore, step S1 specifically involves constructing a 1 / 4 vehicle high-frequency dynamic active suspension control model to obtain vehicle state information including the vertical acceleration of the sprung mass, the vertical velocity of the sprung mass, the suspension dynamic deflection, and the suspension speed.

[0012] Furthermore, the process of constructing the high-frequency dynamics model of the vehicle in step S1 is as follows:

[0013] Considering the damper tower bushing and replacing the simplified tire model with a rigid ring tire model in the vertical direction, the vibration differential equation expression of the vehicle's high-frequency dynamics model is obtained as follows:

[0014]

[0015] in, m s For the sprung mass, m tu For the unsprung mass of the section without rigid rings, m tb For the mass of a rigid ring, k b and c b These are the wheel-ring stiffness and damping, respectively, which connect the rim and the rigid ring. k ct To ensure the grounding stiffness connecting the rigid ring and the ground, k sb and c sb These are the stiffness and damping of the damper tower top bushing, respectively. k sr ,and c sr These represent the remaining stiffness and damping of the suspension after the bushings are removed. z s , z 0 , z u and zb These represent the displacements of the spring-loaded mass, the lower end of the damper tower bushing, the unspring-loaded mass, and the rigid ring, respectively. q For road surface excitation displacement, F a This is for active suspension control.

[0016] Furthermore, the process of establishing the model predictive controller in step S2 is as follows: design a model predictive controller based on the high-frequency dynamics model of the vehicle, which is used to solve the optimal control force that satisfies the constraints according to the system state evolution in the prediction time domain, and introduce the value function output of the reinforcement learning controller as the terminal cost term of the model predictive control.

[0017] Furthermore, the operation of the model prediction controller in step S2 includes:

[0018] At each control sampling time, a model predictive control optimization problem with a finite prediction time domain is constructed based on the vehicle's high-frequency dynamics model and the current system state.

[0019] In the prediction time domain, a phase performance index function is constructed based on the system state and control input to constrain vehicle body acceleration and suspension dynamic deflection;

[0020] The value function output of the reinforcement learning controller is introduced as the terminal cost term of the model predictive control, which is used to evaluate the long-term performance of predicting the terminal state in the time domain.

[0021] Solve the model predictive control optimization problem to obtain the predictive control sequence, and predict the state variables in all prediction steps. Use the first control variable in the predictive control sequence as the active suspension control input of the vehicle at the current moment.

[0022] Furthermore, the solution to the model predictive control optimization problem specifically uses minimizing the stage performance index function and the terminal cost term as the optimization objective.

[0023] Furthermore, the process of establishing a reinforcement learning controller in step S2 is as follows: the state variables predicted by the model predictive controller are used as additional observation information input, and a compensation control force is generated through the policy network to correct the control results output by the model predictive controller.

[0024] Furthermore, the operation of the reinforcement learning controller in step S2 includes:

[0025] Define the observation as s k =[ , - , - , , - , - ],in, s k for k State observations at time, including k Vertical acceleration of the sprung mass at time t Vertical velocity of the sprung mass Suspension dynamic deflection - Suspension speed - and model predictive control prediction in the time domain k + N The predicted value for each terminal step, i.e. k + N The state observations at any given time include the vertical acceleration of the sprung mass. Vertical velocity of the sprung mass Suspension dynamic deflection - and suspension speed - , N This refers to the number of prediction steps in model predictive control.

[0026] Define the action quantity and reward function, where the action quantity... a t = u RL , u RL To enhance the compensating control force output by the learning controller, u RL ∈[ u min , u max ];

[0027] reward function R ( s t , a t )= R time ( s t , a t )+ R freq ( st , a t ), R time ( s t , a t () is the time-domain reward function. R freq ( s t , a t Let be the frequency domain reward function, and its expressions are as follows:

[0028]

[0029]

[0030] In the formula, α 1. α 2 and α 3 is a positive weighting coefficient. ω 1. ω 2. ω 3 is the normalization coefficient. Let be the vertical acceleration of the sprung mass. z 0– z u For suspension dynamic deflection, The vertical jerk of the sprung mass. RMS 0-20Hz , RMS 20-50Hz and RMS 50-500Hz These represent the root mean square values ​​of the sprung mass acceleration in the frequency bands of 0-20 Hz, 20-50 Hz, and 50-500 Hz, respectively. β 1. β 2. β 3 represents the positive weighting coefficient of the frequency domain reward function;

[0031] Construct an agent based on the Deep Deterministic Policy Gradient (DDPG) algorithm, and design the agent's network structure, including: Critic network, target Critic network, Actor network, and target Actor network;

[0032] The Critic network evaluates the value of the current action and state, and uses the evaluation result as the terminal cost term of the model predictive controller to guide the optimization process of the model predictive control.

[0033] By employing experience replay and target network soft update mechanisms, the reinforcement learning agent is trained, enabling the value evaluation results to gradually converge and achieving adaptive optimization of the model's prediction of the control terminal cost.

[0034] Furthermore, in step S2, the reinforcement learning controller and the model prediction controller are fused. Specifically, the compensation control force output by the reinforcement learning controller and the control quantity output by the model prediction controller are superimposed and fused to output the active suspension control input of the vehicle.

[0035] Furthermore, the safety monitoring mechanism in step S3 is as follows: when the suspension travel exceeds the travel limit or the tire dynamic load exceeds the safety constraint, a risk is detected and the system switches to individual model predictive control.

[0036] Compared with the prior art, the present invention has the following advantages:

[0037] This invention constructs a high-frequency dynamic model of a vehicle to obtain vehicle state information. Then, based on the vehicle state information, a model predictive controller and a reinforcement learning controller are established respectively. The reinforcement learning controller and the model predictive controller are fused together to output the active suspension control input of the vehicle. This combines the predictive optimization and constraint handling advantages of model predictive control with the adaptive capability of reinforcement learning, which can achieve the coordinated unity of low-frequency vibration suppression and mid-to-high-frequency compensation control, thereby improving the comprehensive vibration control performance of the vehicle's active suspension across the entire frequency band.

[0038] In this invention, the model predictive controller generates basic control force based on the system prediction model and explicit constraints, which can ensure the stability and safety of the suspension system under constraints such as suspension travel and tire dynamic load; the reinforcement learning controller can compensate and correct the model predictive control output through adaptive learning of vehicle state and operating condition information, thereby effectively dealing with the nonlinear dynamic characteristics of the vehicle and complex operating condition disturbances.

[0039] This invention incorporates the vertical jerkiness of the vehicle body into the time-domain reward function and designs a frequency-domain reward function during the training process of the reinforcement learning controller, which can achieve vibration suppression across the entire frequency band from 0 to 500 Hz and effectively suppress high-frequency vibrations.

[0040] This invention employs a control mechanism that uses offline training of reinforcement learning agents and online collaboration with model predictive control. This avoids complex online learning and parameter switching processes, reduces the online computational burden while ensuring control performance, and improves the real-time performance and engineering feasibility of the control system. It provides a feasible solution for the comprehensive vibration optimization of the entire frequency band of intelligent vehicle active suspension systems.

[0041] This invention also establishes a safety monitoring mechanism to monitor the constraints of the vehicle's active suspension control input. When the suspension travel exceeds the travel limit or the tire dynamic load exceeds the safety constraint, a risk trigger is determined, and the system switches to a separate model predictive control. That is, the output of the model predictive controller is used only as the vehicle's active suspension control input, which can fully ensure the safety of system operation. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0043] Figure 2 This is a schematic diagram of the high-frequency dynamic active suspension control model of a 1 / 4 vehicle in the embodiment;

[0044] Figure 3 This is a schematic diagram of the framework of the active suspension full-frequency vibration control method in the embodiment;

[0045] Figure 4 This is a network architecture diagram of the reinforcement learning agent in the embodiment;

[0046] Figure 5 The time-domain results of the vehicle body vertical acceleration under different suspension control strategies for a 1 / 4 vehicle high-frequency dynamic active suspension control model on a Class C random road surface are presented.

[0047] Figure 6 The frequency domain results of the vehicle's vertical acceleration under different suspension control strategies for a 1 / 4 vehicle high-frequency dynamic active suspension control model on a Class C random road surface are presented. Detailed Implementation

[0048] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0049] Example

[0050] This scheme takes into account that Model Predictive Control (MPC), by explicitly introducing the system prediction model and constraints, can achieve multi-objective optimization within a finite prediction time domain and has good stability and safety constraint handling capabilities. Therefore, the design combines the adaptive capabilities of reinforcement learning with the multi-constraint optimization capabilities of MPC to achieve synergistic optimization of low-frequency ride comfort and mid-to-high-frequency vibration suppression while ensuring system stability and safety.

[0051] This solution proposes an active suspension method for full-frequency vibration suppression, such as... Figure 1 As shown, it includes the following steps:

[0052] S1. Construct a high-frequency dynamic model of the vehicle to obtain vehicle state information;

[0053] S2. Based on vehicle state information, establish a model predictive controller and a reinforcement learning controller respectively, and fuse the reinforcement learning controller and the model predictive controller to output the active suspension control input of the vehicle.

[0054] S3. Establish a safety supervision mechanism to monitor the constraints of the vehicle's active suspension control input, and switch to separate model predictive control when a risk is triggered, using only the output of the model predictive controller as the vehicle's active suspension control input.

[0055] This embodiment applies the above-mentioned scheme to suppress vibrations across the entire frequency band of the active suspension (including the 0-20Hz suspension vibration main frequency band, the 20-50Hz vibration and sound transition main frequency band, and the 50-500Hz in-vehicle noise main frequency band). First, a 1 / 4-inch vehicle active suspension model incorporating high-frequency dynamic characteristics is constructed to obtain system state information such as vehicle acceleration, suspension travel, and tire dynamic load. Figure 2 As shown, this embodiment considers the damper tower top bushing and replaces the simplified tire model with a rigid ring tire model in the vertical direction. The vibration differential equation expression of this high-frequency dynamic model is:

[0056]

[0057] In the formula, m s Represents the sprung mass. m tu Represents the unsprung mass of the section without a rigid ring. m tb Represents the mass of the rigid ring. k b and c b These represent the wheel-ring stiffness and damping, respectively, connecting the rim and the rigid ring. k ct This represents the grounding stiffness connecting the rigid ring and the ground. k sb and c sb These represent the stiffness and damping of the damper tower top bushing, respectively. k sr ,and c sr These represent the remaining stiffness and damping of the suspension after the bushings are removed, respectively. z s , z 0 , z u and z b These represent the displacements of the sprung mass, the lower end of the damper tower bushing, the unsprung mass, and the rigid ring, respectively. qRepresents the road surface excitation displacement. F a This represents the active suspension control force.

[0058] Subsequently, by applying road surface excitations for simulation, vehicle state information, including the vertical acceleration of the sprung mass, can be obtained based on the above 1 / 4 vehicle active suspension model. Vertical velocity of the sprung mass Suspension dynamic deflection z0 -z u and suspension speed - .

[0059] This embodiment combines reinforcement learning with model predictive control, such as Figure 3 As shown, on the one hand, a model predictive controller is designed to solve for the optimal control force that satisfies the constraints based on the system state evolution in the prediction time domain, and the output of the reinforcement learning value function is introduced as the terminal cost term of the model predictive control. On the other hand, a reinforcement learning controller is constructed, which takes the terminal state predicted by the model predictive controller as additional observation information input, and generates a compensating control force through a policy network to correct the control result output by the model predictive controller. u MPC With reinforcement learning to compensate for control u RL The inputs are superimposed and fused to obtain the final active suspension control input. u total = u RL + u MPC It is applied to the 1 / 4 vehicle active suspension model.

[0060] Specifically, in the model predictive control section, at each control sampling time... k A predictive model for model predictive control is constructed based on a vehicle suspension dynamics model. Its discrete state-space form can be expressed as follows:

[0061]

[0062] in, x ( k ) represents the system state vector. u ( k ) represents the active suspension control input, and A and B are the system state matrix and control matrix, respectively. The prediction time domain length is... N In this case, the future state of the system can be obtained recursively from the current state.

[0063] Furthermore, a performance index function is constructed in the prediction time domain to simultaneously constrain vehicle body vibration and suspension safety, and its form can be expressed as:

[0064]

[0065] Where Q and R are the system state weighting matrix and the control input weighting matrix, respectively, used to balance vehicle acceleration suppression and control energy consumption.

[0066] The state value evaluation model trained by reinforcement learning is introduced as the terminal cost term of the model predictive control, used to evaluate the long-term performance of the predicted terminal state in the time domain. This terminal cost can be expressed as:

[0067]

[0068] in, T (·) represents the nonlinear state-value mapping function learned by the value evaluation network of reinforcement learning, which is used to replace the quadratic terminal cost in traditional model predictive control.

[0069] Based on the above-mentioned stage performance indicators and terminal costs, the objective function for model predictive control is constructed as follows:

[0070]

[0071] Under the premise of satisfying the constraints of suspension travel, tire dynamic load, and control input, the predictive control sequence can be obtained by solving the optimization problem: u ( k )={ u ( k | k ), u ( k +1| k ), …, u ( N | k )}, and predict the state variables within all prediction steps, and predict the first control variable in the control sequence. u ( k | k The active suspension control input is applied to the system at the current moment.

[0072] In the reinforcement learning control section, a reinforcement learning controller is constructed, and the observables of the reinforcement learning agent are defined as follows: s k =[ , - , - , , - , - ],in s k express k State observations at time, including k Vertical acceleration of the sprung mass at time t Vertical velocity of the sprung mass Suspension dynamic deflection - and suspension speed - And model predictive control prediction in the time domain. k + N The predicted value for each terminal step, i.e. k + N The state observations at any given time include the vertical acceleration of the sprung mass. Vertical velocity of the sprung mass Suspension dynamic deflection - and suspension speed - , N This indicates the number of prediction steps in model predictive control.

[0073] Define the action quantity of a reinforcement learning agent. a t = u RL ,in u RL This represents the compensation control force of the reinforcement learning controller. u RL ∈[ u min , u max ],in, u min and u max It should be set according to the maximum capacity of the actuator.

[0074] Define the reward function for the reinforcement learning agent as follows: R ( s t , a t )= R time (s t , a t )+ R freq ( s t , a t ),in R time ( s t , a t ) represents the time-domain reward function. R freq ( s t , a t () represents the frequency domain reward function, expressed as:

[0075]

[0076]

[0077] In the formula, α 1. α 2 and α 3 represents the positive weighting coefficient of the time-domain reward function; in this embodiment, the values ​​are 2, 0.1, and 4 respectively. ω 1. ω 2. ω 3 is the normalization coefficient, which is taken as 5, 0.02, and 3500 in this embodiment. For the acceleration of the sprung mass, z 0– z u For suspension dynamic deflection, The jerk of the sprung mass. RMS 0-20Hz , RMS 20-50Hz and RMS 50-500Hz These represent the root mean square values ​​of vehicle body acceleration in the 0-20Hz, 20-50Hz, and 50-500Hz frequency bands, respectively. β 1. β 2. β 3 represents the positive weighting coefficient of the frequency domain reward function, which is taken as 2.5, 2.5, and 5 in this embodiment;

[0078] Construct a DDPG-based intelligent agent and design its network structure (e.g., Figure 4As shown in the figure, the network layers used in this embodiment and their output dimensions are labeled, including: a Critic network, a target Critic network, an Actor network, and a target Actor network. The Critic network, as a value evaluation network, is used to fit the action value function corresponding to the environment state and the agent's actions in real time. The target Critic network has the same structure as the Critic network, and its network parameters are synchronized from the Critic network through periodic soft updates to generate relatively stable target action values, avoiding oscillations or non-convergence during training due to real-time changes in the Critic network parameters. The Actor network, as a policy decision network, outputs the agent's action policy based on the current environment state. The target Actor network has the same structure as the Actor network, and its parameters are also synchronized from the Actor network through a soft update mechanism to generate target actions. It works in conjunction with the target Critic network to calculate the target value, ensuring the training stability and convergence efficiency of the reinforcement learning algorithm.

[0079] The Critic network evaluates the value of the current action and state, and its output serves as the terminal cost term of the model predictive control, guiding the optimization process of the model predictive control. In this embodiment, the reinforcement learning agent is trained using an experience replay and target network soft update mechanism, so that the value evaluation results gradually converge as the system runs (and the reward function also converges), thus obtaining the optimal agent and achieving adaptive optimization of the terminal cost of the model predictive control.

[0080] In practical applications, this embodiment also establishes a safety monitoring mechanism. When the suspension travel exceeds the travel limit (i.e., |z0( t ) -z u ( t (≥ 0.05m) or the tire dynamic load exceeds the safety constraint (i.e. k ct ( z b ( t ) -q ( t )) / (( m s +m u ) g If ) > 1), then the control command at the current moment will be switched to the output of the model predictive controller. u total = u MPC .

[0081] To verify the effectiveness of this solution, a simulation comparison experiment was conducted in this embodiment. The vehicle parameters were set as follows: sprung mass ms = 300, unsprung mass of the section without rigid rings m tu = 30, mass of rigid ring m tb = 6.9, Wheel-ring stiffness connecting the rim and the rigid ring k b = 1804000, Wheel-ring damping connecting the rim and the rigid ring. c b = 324, Grounding stiffness connecting the rigid ring and the ground. k ct = 282160, Stiffness of the damper tower top bushing k sb = 750000, Damping of the damper tower top bushing c sb = 6000, residual stiffness of the suspension after removing the bushings k sr = 20548, Residual damping of the suspension after removing the bushings. c sr = 2000, u min =-1000, u max =1000.

[0082] In this embodiment, the vehicle speed is 20 m / s, and the road excitation is a Class C random road surface. Simulations are used to verify the vibration suppression effect across the entire frequency band. The proposed active suspension full-frequency vibration control method (RL-MPC), which integrates reinforcement learning and model predictive control, is compared with several control strategies, including passive suspension, a linear quadratic regulator (LQR) with full-state feedback, a single reinforcement learning control method (RL) based on DDPG, and a model predictive controller (MPC). The results are as follows: Figure 5 and Figure 6 As shown.

[0083] like Figure 5 As shown, compared to other control strategies, this scheme effectively suppresses the sprung mass vertical acceleration in the time domain. Specifically, in the 0-20Hz frequency band, compared to passive suspension, the sprung mass vertical acceleration of this scheme is reduced by 10.42%.

[0084] like Figure 6 As shown, compared with other control strategies, this scheme has the best overall performance in the frequency domain, effectively suppressing the high-frequency resonance problem caused by the beat frequency generated by the discrete control signal and the unsprung natural frequency. Specifically, in the 20-50 Hz frequency band, compared with passive suspension, the vertical acceleration of the sprung mass corresponding to this scheme is reduced by 4.18%. In the 50-500 Hz frequency band, compared with the single reinforcement learning control method, the high-frequency vibration problem corresponding to this scheme is improved by 20.1%. This shows that this scheme can effectively alleviate the high-frequency vibration problem caused by the discontinuity of control signals, sampling delay and actuator bandwidth limitation in existing electronic suspension control technology.

[0085] In summary, this solution effectively combines the adaptive capabilities of reinforcement learning with the multi-constraint optimization capabilities of model predictive control. Under the premise of ensuring system stability and safety, it can achieve synergistic optimization of low-frequency ride comfort and mid-to-high-frequency vibration suppression, thereby effectively improving the overall vibration control performance of the active suspension system in the 0–500 Hz frequency band.

Claims

1. A method for suppressing vibration across the entire frequency band of an active suspension system, characterized in that, Includes the following steps: S1. Construct a high-frequency dynamics model of the vehicle to obtain vehicle state information. Specifically, construct a 1 / 4 vehicle high-frequency dynamics active suspension control model to obtain vehicle state information including the vertical acceleration of the sprung mass, the vertical velocity of the sprung mass, the suspension dynamic deflection, and the suspension velocity. The process of constructing a high-frequency dynamic model of a vehicle is as follows: Considering the damper tower bushing and replacing the simplified tire model with a rigid ring tire model in the vertical direction, the vibration differential equation expression of the vehicle's high-frequency dynamics model is obtained as follows: in, m s For the sprung mass, m tu For the unsprung mass of the section without rigid rings, m tb For the mass of a rigid ring, k b and c b These are the wheel-ring stiffness and damping, respectively, which connect the rim and the rigid ring. k ct To ensure the grounding stiffness connecting the rigid ring and the ground, k sb and c sb These are the stiffness and damping of the damper tower top bushing, respectively. k sr ,and c sr These represent the remaining stiffness and damping of the suspension after the bushings are removed. z s , z 0 , z u and z b These represent the displacements of the spring-loaded mass, the lower end of the damper tower bushing, the unspring-loaded mass, and the rigid ring, respectively. q For road surface excitation displacement, F a For active suspension control force; S2. Based on vehicle state information, establish a model predictive controller and a reinforcement learning controller respectively, and fuse the reinforcement learning controller and the model predictive controller to output the active suspension control input of the vehicle. S3. Establish a safety supervision mechanism to monitor the constraints of the vehicle's active suspension control input, and switch to separate model predictive control when a risk is triggered, using only the output of the model predictive controller as the vehicle's active suspension control input.

2. The active suspension full-frequency vibration suppression method according to claim 1, characterized in that, The process of establishing the model predictive controller in S2 is as follows: design a model predictive controller based on the high-frequency dynamics model of the vehicle, which is used to solve the optimal control force that satisfies the constraints according to the system state evolution in the prediction time domain, and introduce the value function output of the reinforcement learning controller as the terminal cost term of the model predictive control.

3. The active suspension full-frequency vibration suppression method according to claim 2, characterized in that, The working process of the model prediction controller in S2 includes: At each control sampling time, a model predictive control optimization problem with a finite prediction time domain is constructed based on the vehicle's high-frequency dynamics model and the current system state. In the prediction time domain, a phase performance index function is constructed based on the system state and control input to constrain vehicle body acceleration and suspension dynamic deflection; The value function output of the reinforcement learning controller is introduced as the terminal cost term of the model predictive control, which is used to evaluate the long-term performance of predicting the terminal state in the time domain. Solve the model predictive control optimization problem to obtain the predictive control sequence, and predict the state variables in all prediction steps. Use the first control variable in the predictive control sequence as the active suspension control input of the vehicle at the current moment.

4. The active suspension full-frequency vibration suppression method according to claim 3, characterized in that, The specific objective of solving the model predictive control optimization problem is to minimize the stage performance index function and the terminal cost term.

5. The active suspension full-frequency vibration suppression method according to claim 3, characterized in that, The process of establishing a reinforcement learning controller in S2 is as follows: the state variables predicted by the model predictive controller are used as additional observation information inputs, and a compensation control force is generated through the policy network to correct the control results output by the model predictive controller.

6. The active suspension full-frequency vibration suppression method according to claim 5, characterized in that, The operation of the reinforcement learning controller in S2 includes: Define the observation as s k =[ , - , - , , - , - ],in, s k for k State observations at time, including k Vertical acceleration of the sprung mass at time t Vertical velocity of the sprung mass Suspension dynamic deflection - Suspension speed - and model predictive control prediction in the time domain k + N The predicted value for each terminal step, i.e. k + N The state observations at any given time include the vertical acceleration of the sprung mass. Vertical velocity of the sprung mass Suspension dynamic deflection - and suspension speed - , N The number of prediction steps in model predictive control; Define the action quantity and reward function, where the action quantity... a t = u RL , u RL To enhance the compensating control force output by the learning controller, u RL ∈[ u min , u max ]; reward function R ( s t , a t )= R time ( s t , a t )+ R freq ( s t , a t ), R time ( s t , a t ) is the time-domain reward function. R freq ( s t , a t Let be the frequency domain reward function, and its expressions are as follows: In the formula, α 1. α 2 and α 3 is a positive weighting coefficient. ω 1. ω 2. ω 3 is the normalization coefficient. Let be the vertical acceleration of the sprung mass. z 0– z u For suspension dynamic deflection, The vertical jerk of the sprung mass. RMS 0-20Hz , RMS 20-50Hz and RMS 50-500Hz These represent the root mean square values ​​of the sprung mass acceleration in the frequency bands of 0-20 Hz, 20-50 Hz, and 50-500 Hz, respectively. β 1. β 2. β 3 represents the positive weighting coefficient of the frequency domain reward function; Construct an agent based on the Deep Deterministic Policy Gradient Algorithm (DDPG), and design the agent's network structure, including: Critic network, target Critic network, Actor network, and target Actor network; The Critic network evaluates the value of the current action and state, and uses the evaluation result as the terminal cost term of the model predictive controller to guide the optimization process of the model predictive control. By employing experience replay and target network soft update mechanisms, the reinforcement learning agent is trained, enabling the value evaluation results to gradually converge and achieving adaptive optimization of the model's prediction of the control terminal cost.

7. The active suspension full-frequency vibration suppression method according to claim 6, characterized in that, In step S2, the reinforcement learning controller and the model prediction controller are fused. Specifically, the compensation control force output by the reinforcement learning controller and the control quantity output by the model prediction controller are superimposed and fused to output the active suspension control input of the vehicle.

8. A method for suppressing vibration across the entire frequency band of an active suspension according to any one of claims 1 to 7, characterized in that, The safety monitoring mechanism in S3 is as follows: when the suspension travel exceeds the travel limit or the tire dynamic load exceeds the safety constraint, a risk is detected and the system switches to individual model predictive control.

Images