Human-like feedback learning control method for spring damping system for electric vehicle
Through human-like feedback learning methods, Transformer classification model and reinforcement learning technology are used to solve the shortcomings of traditional spring damping system control in road surface perception and zero-sample data control, and accurate system control and efficient road surface perception are achieved.
Patent Information
- Application Number
- CN202510156968.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-13
- Publication Date
- 2025-03-14
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The spring damping system control for traditional electric vehicles has shortcomings in road perception, nonlinear dynamic behavior processing and zero sample data control, resulting in weak road perception ability and poor control effect.
Using human-like feedback learning method, by measuring the center of mass position and mass changes of the spring damping system, the Transformer classification model is used to output the quantized value of road surface unevenness, construct a physical model and establish an overall reward function for human-like feedback and reinforcement learning, and optimize the objective function to update the strategy parameters to obtain the final system control instructions.
Accurate control under zero sample data is achieved, the coordination between the controller and the driver is improved, and the technical gap in the long-term prediction of the automotive spring damping system is filled.
Smart Images
Figure CN119610980A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of new energy vehicles, and in particular to a spring damping system control method for electric vehicles with human-like feedback learning. Background Art
[0002] At present, the self-control of spring damping systems for electric vehicles faces multiple challenges. First, the spring damping control of traditional electric vehicles requires accurate information on road roughness. Existing methods rely on simple models and empirical rules and do not consider changes in mass and center of mass position, resulting in weak road perception capabilities. Second, the traditional vehicle spring damping system control does not take into account nonlinear dynamic behaviors, and the control effect is difficult to match human experience control. Finally, learning-based control relies on a large amount of experimental data and cannot achieve precise control under zero-sample data.
[0003] It can be seen that how to create a new spring damping system control method has become a goal that the industry urgently needs to improve. Summary of the invention
[0004] In view of this, an embodiment of the present disclosure provides a spring damping system control method for an electric vehicle with human-like feedback learning to solve the problems existing in the prior art.
[0005] In a first aspect, an embodiment of the present disclosure provides a method for controlling a spring damping system for an electric vehicle using human-like feedback learning, the method comprising the following steps: Measure the change in the center of mass position of the spring-damper system and the change in the mass of the spring-damper system; Based on the mass of the spring-damper system and the change in the center of mass position of the spring-damper system, the Transformer classification model is used to output the quantitative value of road roughness. Construct a physical model of the spring-damping system, establish a system state vector that takes into account road roughness, and establish a human-like feedback and reinforcement learning overall reward function; by optimizing the objective function and updating the strategy parameters, the final system control instructions are obtained.
[0006] According to a specific implementation of the embodiment of the present disclosure, the method of outputting a quantized value of road surface roughness using a Transformer classification model includes: Input time series data and define a time series feature matrix; wherein the time series data includes: mass change, time series data of dynamic mass change, center of mass position change, dynamic change of center of mass position, and system response characteristics; wherein the system response characteristics include: dynamic acceleration, system natural frequency and energy; Normalizing the feature matrix; Process the normalized feature matrix based on Transformer; Use the self-attention mechanism to calculate global feature relationships; The method of calculating the global feature relationship by using the self-attention mechanism includes: The calculation is based on the following formula The feature representation matrix of the layer: =LayerNorm( +Attention ); in, For the The feature representation matrix of the layer; For the The feature representation matrix of the layer; LayerNorm is a function that normalizes the feature dimension of each sample; is the query matrix; is the key matrix; is the value matrix; The feed-forward network update is performed based on the following formula: =LayerNorm( +FFN( )); in, For the The feature representation matrix of the layer; FFN is a two-layer fully connected network; ; in, , are the weights of different network layers; , is the bias of different layers; Z is the feature representation matrix; The final output of the feedforward network update is classified through a pooling operation and a fully connected layer: ; in, is the classification probability vector; Wo is the weight of the output classification; is the bias of the initial layer; The road roughness is calculated based on the following formula: ; in, is the quantified value of road roughness; is the probability vector output by the model; is the roughness grade number; Indicates that the road belongs to The probability of level unevenness; is a specific constant indicating the severity of the unevenness.
[0007] According to a specific implementation of the embodiment of the present disclosure, the step of constructing a physical model of a spring damping system includes: The dynamic equation of the spring-damper system is constructed based on the following formula: ; ; in, is the sprung mass; is the vertical acceleration of the sprung mass; is the wheel mass; is the spring stiffness; is the damping coefficient of the spring damping system; is the stiffness of the tire; is the vertical displacement of the unsprung mass; is the vertical velocity of the unsprung mass; is the vertical acceleration of the unsprung mass; is the vertical displacement of the sprung mass; Input for road surface; is the control input; The step of establishing a system state vector taking into account the road surface roughness comprises: ; Where S is the system state vector; is the vertical velocity of the sprung mass; is the vertical displacement of the sprung mass; is the vertical displacement of the unsprung mass; To measure the roughness of the road surface; is the vehicle pitch angle; is the vehicle pitch angular velocity; is the vehicle yaw angle; is the vehicle yaw angular velocity; The method of establishing a human-like feedback and reinforcement learning overall reward function includes the following steps: The overall reward function is calculated based on the following formula : ; in, s is the state space; is the action space; is the weight coefficient of the human-like feedback and target task reward function; is the reward function based on the target task; Provide a reward function for human-like feedback; Construct a human-like feedback reward function based on the following formula : ; in, is the comfort reward function; for The weight of is the manipulative reward function; for The weight of is the energy consumption reward function; for The weight of ; ; ; Wherein, T is the upper limit of the integral time; for The second derivative of is the power consumption of the actuator of the active spring damping system; The method of obtaining the final system control instruction by optimizing the objective function and updating the strategy parameters includes: Using candidate control inputs Boot strategy update: ; in, To control the trade-off between human-like priors and policy learning; is the action space; is the strategy function; Optimize the objective function based on the following formula and update the strategy parameters : ; in, is the cost function; is the average reward expectation; Q The law of strategy execution; The final system control instructions are generated based on the following formula: .
[0008] The electric vehicle spring damping system control method of the human-like feedback learning in the disclosed embodiment has the following beneficial effects compared with the prior art: 1. This paper designs a Transformer-based road roughness classifier, which can fill the technical gap of long-term prediction of current automotive spring damping systems; 2. The present invention establishes a control optimization method with human-like feedback, improves the degree of coordination between the controller and the driver, and realizes the control of the suspension damping system under zero sample data. BRIEF DESCRIPTION OF THE DRAWINGS
[0009] The above is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0010] Figure 1 A schematic flow chart of a method for controlling a spring damping system for an electric vehicle with human-like feedback learning provided in an embodiment of the present disclosure. DETAILED DESCRIPTION
[0011] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.
[0012] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.
[0013] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement a device and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this device and / or practice this method.
[0014] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects described may be practiced without these specific details.
[0015] The embodiment of the present invention provides a human-like feedback learning control method for a spring damping system for an electric vehicle. The method measures the information such as the center of mass position and system mass change of the spring damping system through a displacement sensor, and outputs a quantitative value of road surface roughness using a Transformer classification model. A physical model of the spring damping system is constructed, a system state vector considering road surface roughness is established, and an overall reward function of human-like feedback and reinforcement learning is established. The method optimizes the objective function, updates the strategy parameters, and obtains the final system control instructions, thereby achieving precise control of the spring damping system for the vehicle under zero data samples and solving the problems existing in the prior art.
[0016] Figure 1 A schematic diagram of the process of a spring damping system control method for an electric vehicle with human-like feedback learning provided in an embodiment of the present disclosure.
[0017] like Figure 1 As shown, at step S110, the change in the center of mass position of the spring damping system and the mass change of the spring damping system are measured.
[0018] More specifically, the displacement sensor is used to measure information such as the center of mass position of the spring damping system and the change in system mass.
[0019] More specifically, the process proceeds to step S120.
[0020] At step S120 , based on the mass of the spring damping system and the change in the center of mass position of the spring damping system, a quantitative value of the road surface roughness is output using a Transformer classification model.
[0021] More specifically, the Transformer classification model is used to output the quantitative value of road roughness through the changes in the mass of the spring-damper system and the position of the center of mass of the spring-damper system.
[0022] In an embodiment of the present invention, the method of outputting a quantitative value of road roughness using a Transformer classification model includes: inputting time series data and defining a time series feature matrix; normalizing the feature matrix; processing the normalized feature matrix based on the Transformer; and calculating global feature relationships using a self-attention mechanism.
[0023] In an embodiment of the present invention, the time series data includes: mass change, time series data of dynamic mass change, center of mass position change, dynamic change of center of mass position, and system response characteristics; wherein the system response characteristics include: dynamic acceleration, system natural frequency and energy.
[0024] In an embodiment of the present invention, the method of calculating the global feature relationship using the self-attention mechanism includes: The calculation is based on the following formula The feature representation matrix of the layer: =LayerNorm( +Attention ); in, For the The feature representation matrix of the layer; For the The feature representation matrix of the layer; LayerNorm is a function that normalizes the feature dimension of each sample; is the query matrix; is the key matrix; is the value matrix; The feed-forward network update is performed based on the following formula: =LayerNorm( +FFN( )); in, For the The feature representation matrix of the layer; FFN is a two-layer fully connected network; ; in, , are the weights of different network layers; , is the bias of different layers; Z is the feature representation matrix; The final output of the feedforward network update is classified through a pooling operation and a fully connected layer: ; in, is the classification probability vector; Wo is the weight of the output classification; is the bias of the initial layer.
[0025] The road roughness is calculated based on the following formula: ; in, is the quantified value of road roughness; is the probability vector output by the model; is the roughness grade number; Indicates that the road belongs to i The probability of level unevenness; is a specific constant indicating the severity of the unevenness.
[0026] More specifically, the Transformer classification model is used to output the quantitative value of road roughness as follows: Create input time series data. The specific input features include: 1. Mass change (Δm): Time series data of dynamic mass change.
[0027] 2. Change in center of mass position (Δx): The dynamic change in the center of mass position.
[0028] 3. System response characteristics: including dynamic acceleration ; System natural frequency ;energy: .
[0029] Define the time series feature matrix: ……Formula 1 Where T is the number of sampling time steps.
[0030] The output is the quantified value of road roughness ( ), according to ISO standards The value range of is [1,10].
[0031] The input feature matrix is further normalized, and the enhanced feature matrix is in the form of: X_enhanced=[X_normalized,FFT(x(t))]……Equation 2 Among them, FFT(x(t)) is the frequency component obtained by performing fast Fourier transform on the signal.
[0032] A Transformer-based time series model is used to process input features as follows: Input matrix: X_enhanced; Through the linear transformation of the embedding layer, the feature embedding is obtained: Z 0 =Embedding(X_enhanced)……Equation 3 Among them, Z 0 is the initial input embedding matrix (generated by the Embedding layer).
[0033] Use the self-attention mechanism to calculate the global feature relationship: ……Formula 4 in, is the query matrix; is the key matrix; is the value matrix; Represents the Transformer The feature representation matrix of the layer; , , : The mapping weight matrix used to calculate the attention.
[0034] The calculated attention result and the input are updated into new features through residual connection: =LayerNorm( +Attention )……Formula 5 Among them, LayerNorm is a normalization function for the feature dimension of each sample (that is, the neuron output of each layer).
[0035] Feedforward network update: =LayerNorm( +FFN( ))……Formula 6 FFN is a two-layer fully connected network: ……Formula 7 in, , are the weights of different network layers, , is the bias of different layers.
[0036] The final output is classified through a pooling operation and a fully connected layer: ……Formula 8 Among them, Wo is the weight of the output classification, is the classification probability vector.
[0037] According to the classification probability Calculate the quantified value of road roughness The process can be further refined, and the specific steps are as follows: Use a classification model to predict the level of road roughness, and the classification model outputs the predicted probability of each roughness level. is the probability vector output by the model, where is the number of classifications (i.e., the number of roughness levels), each Indicates that the road belongs to The probability of level unevenness.
[0038] The unevenness is divided into and assign a quantitative value to each level , these quantitative values can be defined according to specific standards (such as ISO standards), usually a continuous or discrete value, indicating the severity of the unevenness. The quantitative value corresponding to each level , the final road roughness quantification value is calculated by weighted average .
[0039] ……Formula 9 Next, go to step S130.
[0040] In step S130, a physical model of the spring damping system is constructed, a system state vector considering the road roughness is established, and a human-like feedback and reinforcement learning overall reward function is established; by optimizing the objective function, the strategy parameters are updated to obtain the final system control instructions.
[0041] In an embodiment of the present invention, the step of constructing a physical model of a spring damping system includes: The dynamic equation of the spring-damper system is constructed based on the following formula: ; ; in, is the sprung mass; is the vertical acceleration of the sprung mass; is the wheel mass; is the spring stiffness; is the damping coefficient of the spring damping system; is the stiffness of the tire; is the vertical displacement of the unsprung mass; is the vertical velocity of the unsprung mass; is the vertical acceleration of the unsprung mass; is the vertical displacement of the sprung mass; Input for road surface; For control input.
[0042] In the embodiment of the present invention, the establishment of a system state vector taking into account the road surface roughness includes: ; Where S is the system state vector; is the vertical velocity of the sprung mass; is the vertical displacement of the sprung mass; is the vertical displacement of the unsprung mass; To measure the roughness of the road surface; is the vehicle pitch angle; is the vehicle pitch angular velocity; is the vehicle yaw angle; is the vehicle yaw angular velocity.
[0043] In an embodiment of the present invention, the process of establishing a human-like feedback and reinforcement learning overall reward function includes the following steps: The overall reward function is calculated based on the following formula : ; in, s is the state space; is the action space; is the weight coefficient of the human-like feedback and target task reward function; is the reward function based on the target task; Reward function for human-like feedback.
[0044] In this embodiment of the present invention, a human-like feedback reward function is constructed based on the following formula: : ; in, is the comfort reward function; for The weight of is the manipulative reward function; for The weight of is the energy consumption reward function; for The weight of ; ; ; Wherein, T is the upper limit of the integral time; for The second derivative of is the power consumption of the actuator of the active spring damping system.
[0045] In the embodiment of the present invention, the step of optimizing the objective function and updating the strategy parameters to obtain the final system control instruction includes: Using candidate control inputs Boot strategy update: ; in, To control the trade-off between human-like priors and policy learning; is the action space; is the strategy function; Optimize the objective function based on the following formula and update the strategy parameters : ; in, is the cost function; is the average reward expectation; Q The law of strategy execution; The final system control instructions are generated based on the following formula: .
[0046] More specifically, the steps of human-like feedback control are as follows: Construct the dynamic equation of the spring damping system for the vehicle: ……Formula 10 ……Formula 11 in, is the sprung mass, is the wheel mass, is the spring stiffness; is the damping coefficient of the spring damping system; is the stiffness of the tire; is the vertical displacement of the unsprung mass; is the vertical displacement of the sprung mass; For road surface input, For control input.
[0047] The state vector of the spring-damper system is defined as: ……Formula 12 in, is the vertical displacement and velocity of the sprung mass of the vehicle body, are the wheel positions and velocities, is the vehicle pitch angle and its pitch velocity, are the vehicle yaw angle and yaw velocity.
[0048] Define the action vector as , represents the external input force of the system, and constructs a human-like feedback reward function , including the following parts: ……Formula 13 in, is the comfort reward function, is the manipulative reward function, is the energy consumption reward function, is the weight of the corresponding function. The specific formula is as follows: ……Formula 14 ……Formula 15 ……Formula 16 is the power consumption of the actuator of the active spring damping system, combined with the state S and the human-like reward , the control input is generated through the denoising process: ……Formula 17 in, is the probability of action execution; Y is normal distribution; is the action at time t; is the action at time t-1; is the mean value of the reward.
[0049] Finally, the candidate control input is output through iteration: ……Formula 18 in, is the control input; To integrate the action execution probability of human-like experience; Further establish human-like feedback and reinforcement learning overall reward function: ……Formula 19 in, is a reward function based on the target task, is the weight coefficient of the human-like feedback and the target task reward function.
[0050] Calculate the cumulative reward function for the control input: ……Formula 20 in, is the discount factor.
[0051] Using candidate control inputs Boot strategy update: ……Formula 21 in, To control the trade-off between human-like priors and policy learning.
[0052] Update the strategy parameters by optimizing the objective function : ……Formula 22 The final control input generated is: ……Formula 23 The electric vehicle spring damping system control method of human-like feedback learning proposed in the present invention has the following beneficial effects compared with the prior art: 1. This paper designs a Transformer-based road roughness classifier, which can fill the technical gap of long-term prediction of current automotive spring damping systems; 2. The present invention establishes a control optimization method with human-like feedback, improves the degree of coordination between the controller and the driver, and realizes the control of the suspension damping system under zero sample data.
[0053] The above is only a specific implementation of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present disclosure should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be based on the protection scope of the claims.
Claims
1. A human-like feedback learning spring damping system control method for an electric vehicle, characterized in that: The method comprises the following steps: Measure the change in the center of mass position of the spring-damper system and the change in the mass of the spring-damper system; Based on the mass of the spring-damper system and the change in the center of mass position of the spring-damper system, the Transformer classification model is used to output the quantitative value of road roughness. Construct a physical model of the spring-damping system, establish a system state vector that takes into account road roughness, and establish a human-like feedback and reinforcement learning overall reward function; by optimizing the objective function and updating the strategy parameters, the final system control instructions are obtained.
2. The method for controlling a spring damping system for an electric vehicle using human-like feedback learning according to claim 1, characterized in that: The method of using the Transformer classification model to output a quantitative value of road surface roughness includes: Input time series data and define a time series feature matrix; wherein the time series data includes: mass change, time series data of dynamic mass change, center of mass position change, dynamic change of center of mass position, and system response characteristics; wherein the system response characteristics include: dynamic acceleration, system natural frequency and energy; Normalizing the feature matrix; Process the normalized feature matrix based on Transformer; Use the self-attention mechanism to calculate global feature relationships; The method of calculating the global feature relationship by using the self-attention mechanism includes: The calculation is based on the following formula The feature representation matrix of the layer: =LayerNorm( +Attention ); in, For the The feature representation matrix of the layer; For the The feature representation matrix of the layer; LayerNorm is a function that normalizes the feature dimension of each sample; is the query matrix; is the key matrix; is the value matrix; Feedforward network updates are performed based on the following formula: =LayerNorm( +FFN( )); in, For the The feature representation matrix of the layer; FFN is a two-layer fully connected network; ; in, , are the weights of different network layers; , is the bias of different layers; Z is the feature representation matrix; The final output of the feedforward network update is classified through a pooling operation and a fully connected layer: ； in, is the classification probability vector; Wo is the weight of the output classification; is the bias of the initial layer; The road roughness is calculated based on the following formula: ； in, is the quantified value of road roughness; is the probability vector output by the model; is the roughness grade number; Indicates that the road belongs to The probability of level unevenness; is a specific constant indicating the severity of the unevenness.
3. The method for controlling a spring damping system for an electric vehicle using human-like feedback learning according to claim 1, characterized in that: The construction of the spring damping system physical model includes: The dynamic equation of the spring-damper system is constructed based on the following formula: ； ； in, is the sprung mass; is the vertical acceleration of the sprung mass; is the vertical velocity of the sprung mass; is the wheel mass; is the spring stiffness; is the damping coefficient of the spring damping system; is the stiffness of the tire; is the vertical displacement of the unsprung mass; is the vertical velocity of the unsprung mass; is the vertical acceleration of the unsprung mass; is the vertical displacement of the sprung mass; Input for road surface; is the control input; The step of establishing a system state vector taking into account the road surface roughness comprises: ； Where S is the system state vector; is the vertical velocity of the sprung mass; is the vertical displacement of the sprung mass; is the vertical displacement of the unsprung mass; To measure the roughness of the road surface; is the vehicle pitch angle; is the vehicle pitch angular velocity; is the vehicle yaw angle; is the vehicle yaw angular velocity; The method of establishing a human-like feedback and reinforcement learning overall reward function includes the following steps: The overall reward function is calculated based on the following formula : ； in, is the state space; is the action space; is the weight coefficient of the human-like feedback and target task reward function; is the reward function based on the target task; Provide a reward function for human-like feedback; Construct a human-like feedback reward function based on the following formula : ； in, is the comfort reward function; for The weight of is the manipulative reward function; for The weight of is the energy consumption reward function; for The weight of ； ； ； Wherein, T is the upper limit of the integral time; for The second derivative of is the power consumption of the actuator of the active spring damping system; The method of obtaining the final system control instruction by optimizing the objective function and updating the strategy parameters includes: Using candidate control inputs Boot strategy update: ； in, To control the trade-off between human-like priors and policy learning; is the action space; is the strategy function; Optimize the objective function based on the following formula and update the strategy parameters : ； in, is the cost function; is the average reward expectation; Q The law of strategy execution; The final system control instructions are generated based on the following formula: 。