A semi-active suspension control system based on Transformer architecture

Abstract

The present invention belongs to the technical field of vehicle suspension system control. The present invention discloses a semi-active suspension control system based on Transformer architecture, comprising the following steps: arranging a sensor network to collect vehicle dynamic data; using a preset algorithm to align asynchronous data from different sensors to maintain the temporal consistency of heterogeneous data; the measurement value of sensor i at time t is x i (t), in the present invention, the Transformer mechanism is used to reduce the weight distribution of state quantities, avoid the erroneous control quantity caused by sensor errors, improve the robustness of the suspension control system to data anomalies, and maintain stable control in the event of sensor failure, injection of false sensor data by external attacks, or extreme road conditions. The DDPG reinforcement learning control algorithm is used to enhance the system's adaptability in complex and changing environments, and it can quickly adapt to different road conditions, loads, and driving styles, improving the control effect while ensuring computational efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle suspension system control, and in particular relates to a semi-active suspension control system based on a Transformer architecture. Background Art

[0002] As a core component of automotive engineering, the vehicle suspension system is crucial for mitigating vibrations and impacts during driving, and is crucial for enhancing ride comfort and driving stability. Advanced strategies are crucial for controlling the controllable damping of the suspension to optimize these properties. Currently, vehicle suspension control strategies are undergoing a transition from traditional rule-based approaches to intelligent control. Passive, semi-active, and active suspension systems each have their own unique characteristics, but also limitations. While passive suspension offers a simple structure, it struggles to adapt to complex road conditions and provide flexible vibration damping. Semi-active suspension, while capable of a degree of control by adjusting the damping force of the shock absorbers, lacks the ability to actively apply force. Active suspension, while offering superior control performance, is prohibitively expensive and complex, hindering widespread adoption in standard vehicles.

[0003] The integration of intelligent control technologies, particularly fuzzy control, adaptive control, neural networks, and genetic algorithms, has brought new vitality to suspension system control, improving effectiveness and flexibility. In the process of intelligent vehicle networking, the injection of false sensor data via the CAN bus poses a threat to vehicle safety. However, these methods still face challenges in scalability and real-time performance when dealing with complex real-world operating conditions, particularly their inability to handle abnormal sensor data. Furthermore, while reinforcement learning has shown great potential in the control field, enabling optimization of control strategies through continuous interaction with the environment, its practical application in suspension systems is limited by issues such as training instability, insensitivity to FDI attacks, low sample efficiency, and limited policy generalization. The Transformer model, with its superior sequence modeling capabilities, has achieved significant success in fields such as natural language processing. Its core self-attention mechanism effectively captures long-term and short-term dependencies, a feature that holds great potential for processing time-series sensor data. However, the application of the Transformer model to suspension system control remains innovative and requires overcoming numerous technical challenges, including high computational complexity, optimization to meet real-time control requirements, and the design of Transformer variants suitable for control tasks. Therefore, a semi-active suspension control system based on Transformer architecture is proposed to solve the above problems. Summary of the Invention

[0004] The object of the present invention is to provide a semi-active suspension control system based on a Transformer architecture to solve the problems raised in the above background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a semi-active suspension control system based on a Transformer architecture, comprising the following steps:

[0006] S1: Deploy sensor networks to collect vehicle dynamic data;

[0007] S2: Use a preset algorithm to align asynchronous data from different sensors to maintain temporal consistency of heterogeneous data;

[0008] The measurement value of sensor i at time t is x i (t), align all sensor data to the same time point t k : where t j ≤t k <t j+1 ;

[0009] S3: Build a quarter vehicle two-degree-of-freedom model

[0010]

[0011] where m s is the vehicle body mass, m u is the unsprung mass, F is the damping force of the magnetorheological damper, x s is the vehicle body displacement, x u is the unsprung displacement, x r is the wheel displacement, c is the uncontrollable damping of the damper, k s is the spring stiffness, k t is the wheel stiffness;

[0012] S4: Modeling of Magnetorheological Damper via Neural Networks:

[0013]

[0014] The damper experimental data is obtained through the preset equipment, and the data is fitted through the neural network to obtain the weight matrix w I , w H and w O ;

[0015] Combining the damper neural network model with the vehicle suspension dynamics model to derive an integrated model;

[0016] Interact the reinforcement learning structure with the integrated model to train the neural network parameters;

[0017] S5: Apply DDPG reinforcement learning algorithm to calculate the control quantity of the state quantity. The DDPG network structure is Actor-Crit ic structure, in which the Actor network generates actions: a t =μ(s|θ μ ), the Critic network evaluates the state-action value Q(s,a|θ Q );

[0018] S6: The state quantity is passed into the Transformer architecture to replace the traditional Actor network, and the correlation of the state quantity is analyzed to distinguish between valid information and state quantity errors. i (t k ) is passed into the DDPG Transformer architecture;

[0019] S7: Combine the encoded high-dimensional data X with the weights Point multiplication to obtain a weighted matrix Each row of V represents the high-dimensional representation of sensor data;

[0020] S8: Perform transpose multiplication on the Q and K matrices to obtain QK T matrix;

[0021] S9: Divide the K matrix dimension by the obtained attention weight result Scaling is performed to calculate the attention weight matrix Perform softmax activation function transformation and convert the result into percentage;

[0022] S10: The transformed attention weights Multiplying with matrix V, we get Each element in the result is the sum of the attention weights of each sample data corresponding to all other sample data in the 64-dimensional Q,K matrix. Adjust V matrix weights Get the result Attention(Q,K,V);

[0023] S11: Self-attention matrix head i =Attention(Q,K,V) performs multiple parallel trainings to obtain the multi-head attention mechanism result Concat(head1,…,head h ), the multi-head attention mechanism result is combined with the multi-head weight matrix W O Multiply them together to get the multi-head attention result MultiHead(Q,K,V)=Concat(head1,…,head h )W O , each head iIt is the result of the attention mechanism in different spaces;

[0024] S12: Replace the traditional reinforcement learning actor network architecture with the Transformer architecture and generate a parameterized strategy μ(s|θ μ ) to output the optimal action a under a given state t =μ(s|θ μ ), whose goal is to maximize the expected cumulative reward: The Critic network is used to evaluate the state-action value function Q(s,a|θ Q ), and optimize its parameters by minimizing the following loss function: L(θ Q )=E (s,a,r,s′)～D [(r+γQ(s′,μ(s′|θ μ )|θ Q )-Q(s,a|θ Q )) 2 ] Where r is the immediate reward, γ is the discount factor, and D is the data in the experience replay buffer;

[0025] S13: Design reinforcement learning reward function;

[0026] S14: uses a dual network architecture, including two independent Actor networks and With two independent Cr it ic networks and The minimum action amount and state-action value are calculated using the following formula: The loss function is defined as:

[0027]

[0028] S15: Store the data obtained during the interaction with the environment in the playback buffer D, and randomly sample batches of data from the buffer for training to improve sample utilization. For the dual-Actor and Cr it ic network update mechanisms, the target network parameters θ are maintained and delayed for update. μ ' and θ Q ', each time the network parameters are updated, the target network uses the soft update rule: Actor network update rule: θ μ '←τθ μ +(1-τ)θ μ ', Cr it ic network update rule: θ Q '←τθ Q +(1-τ)θ Q ', where τ is the soft update parameter, which balances the current network parameters with the parameters to be updated, adjusts the network parameters, and obtains the control signal I.

[0029] Preferably, in S1, the sensor network includes an accelerometer, a gyroscope, a displacement sensor and a pressure sensor.

[0030] Preferably, in S1, the vehicle dynamic data includes vehicle body acceleration, suspension travel, and tire deformation.

[0031] Preferably, in S4, the preset equipment includes a suspension test bench.

[0032] Preferably, in S4, the output value of the neural network is the controllable damping force F of the magnetorheological damper.

[0033] Preferably, the state quantity is passed into the Transformer architecture to replace the traditional Actor network, analyze the correlation of the state quantity, distinguish between valid information and state quantity error, and convert the data x i (t k ) is passed into the DDPG Transformer architecture, which includes the following steps:

[0034] First, the data x i (t k ) Upgraded to 64-dimensional S t , using learnable Fourier feature pairs S t Position code: PE (pos,2i) =sin(pos·w i ),PE (pos,2i+1) =cos(pos·w i ), where w i is a learnable parameter

[0035] Preferably, in S4, the integrated model is:

[0036]

[0037] In the formula, m s is the vehicle body mass, m u is the unsprung mass, F is the damping force of the magnetorheological damper, x s is the vehicle body displacement, x u is the unsprung displacement, x r is the wheel displacement, c is the uncontrollable damping of the damper, k s is the spring stiffness, k t is the wheel stiffness.

[0038] Preferably, in S2, the preset algorithm includes an interpolation algorithm.

[0039] Preferably, in S13, the reinforcement learning reward function is The reward function is composed of four parts: vehicle acceleration, suspension travel, tire deformation, and suspension dissipative constraint. The vehicle acceleration and suspension travel are multiplied by weight coefficients a1 and a2 respectively to distribute the weights. In addition, obstacle functions P, Q, and N are added. The obstacle function P constrains the suspension travel and keeps the suspension travel within the limit range, so that (x s -x u ) at the maximum dynamic travel of the suspension max Within the range, the obstacle function Q constrains the suspension wheel deformation, and the obstacle function N constrains the magnetorheological semi-active suspension input.

[0040] Preferably, the barrier function The barrier function The barrier function

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] In the present invention, the Transformer mechanism is used to reduce the weight distribution of state quantities, avoid the erroneous control caused by sensor errors, improve the robustness of the suspension control system to data anomalies, and maintain stable control in the event of sensor failure, injection of false sensor data by external attacks, or extreme road conditions.

[0043] The DDPG reinforcement learning control algorithm enhances the system's adaptability in complex and changing environments. It can quickly adapt to different road conditions, loads, and driving styles, improving control effects while ensuring computational efficiency.

[0044] Compared with traditional methods, the present invention improves ride comfort by an average of 36.7% and handling stability by 30.45% under various road conditions. Through lightweight design and optimization algorithms, it ensures the real-time performance of the control system and improves the interpretability of the system. The basis for control decisions can be understood by analyzing the attention weights of the Transformer, reducing the system's dependence on precise physical models and enhancing its generalization capabilities in practical applications.

[0045] The present invention is different from existing traditional control methods and traditional intelligent algorithms. Traditional algorithms cannot accurately characterize the current state when sensors have errors, which in turn affects the precise and stable control of the vehicle body. The method described in the present invention is a deep deterministic policy gradient (DDPG) reinforcement learning control method based on a data-driven Transformer architecture. The Transformer architecture is used to perform correlation analysis on abnormal data to obtain the importance weights between the data. The influence of abnormal data on the control quantity is reduced by a smaller weight coefficient, and the weighted data is subjected to reinforcement learning training to obtain a control method that meets the control requirements. The method is robust to sensor failures, injection of false sensor data by external attacks, or extreme road conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is the overall control flow chart of the control method of the present invention;

[0047] Figure 2 This is a diagram showing the internal structure of a specific controller of the present invention;

[0048] Figure 3 A diagram of a two-degree-of-freedom suspension structure in an environment that interacts with a control algorithm in the present invention;

[0049] Figure 4 A schematic diagram of vehicle body acceleration when performing a convex road simulation experiment using the suspension control system of the present invention;

[0050] Figure 5 A schematic diagram of the suspension dynamic travel when using the present invention to control the suspension in a convex road simulation experiment;

[0051] Figure 6 Schematic diagram of the dynamic and static load ratio of tires when conducting a vehicle convex road simulation experiment using the control method of the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] See also Figures 1-6 As shown, the present invention provides the following technical solutions:

[0054] The present invention provides a system for processing data anomalies and optimizing suspension control using a Transformer architecture and reinforcement learning, characterized by comprising the following steps:

[0055] S1: Design a comprehensive sensor network using accelerometers, gyroscopes, displacement sensors, and pressure sensors to collect vehicle dynamic data, including body acceleration, suspension travel, and tire deformation. The sensor-collected state data can be subject to two types of data anomalies: point errors caused by electromagnetic interference and interval errors caused by unstable power supply. This invention can mitigate these potential impacts.

[0056] S2: Use interpolation algorithms to align asynchronous data from different sensors to ensure the timing consistency of heterogeneous data. The measurement value of sensor i at time t is x i (t), the present invention aligns all sensor data to a unified time point t through interpolation method k : where t j ≤t k <t j+1 .

[0057] S3: For the establishment of the reinforcement learning interactive environment, the suspension system is modeled using the vehicle dynamics model. A quarter vehicle two-degree-of-freedom model is established. where m s is the vehicle body mass, m u is the unsprung mass, F is the damping force of the magnetorheological damper, x s is the vehicle body displacement, x u is the unsprung displacement, x r is the wheel displacement, c is the uncontrollable damping of the damper, k s is the spring stiffness, k t Wheel stiffness.

[0058] S4: Magnetorheological damper modeling using neural networks: The damper experimental data is obtained through the suspension test bench, and the data is fitted through the neural network to obtain the weight matrix w I， w H , w O The output value of the neural network is the controllable damping force F of the magnetorheological damper. Combining the damper neural network model with the vehicle suspension dynamics model, an integrated model is obtained: The reinforcement learning structure interacts with the environment model to train the neural network parameters to achieve the best control effect.

[0059] S5: Apply the DDPG reinforcement learning algorithm to the state quantity to calculate the control quantity. The DDPG algorithm is a model-free reinforcement learning algorithm suitable for continuous action space. The DDPG network structure is an Actor-Critic structure. The Actor network generates actions: a t =μ(s|θ μ), the Critic network evaluates the state-action value Q(s,a|θ Q ).

[0060] S6: The state quantity is transferred to the improved Actor network of this patent - Transformer architecture to replace the traditional Actor network, analyze the correlation of the state quantity, and distinguish between valid information and state quantity errors. i (t k ) is passed into the DDPG Transformer architecture. First, the data x i (t k ) Upgraded to 64-dimensional S t , using learnable Fourier feature pairs S t Position code: PE (pos,2i) =sin(pos·w i ),PE (pos,2i+1) =cos(pos·w i ), where w i are learnable parameters.

[0061] S7: Combine the encoded high-dimensional data X with the weights Point multiplication to obtain a weighted matrix Each row of V represents a high-dimensional representation of sensor data.

[0062] S8: Perform transpose multiplication on the Q and K matrices to obtain QK T The matrix is ​​used to determine the relative relationship between data at different locations. The magnitude of the resulting number reflects the degree of relationship between the data, which is the attention weight. Larger numbers indicate a close relationship between the data, while smaller numbers indicate a distant relationship. This attention weight reflects the relationship between the focused data and other data. For sensor-measured anomalous data, since the anomalous data deviates significantly from the true data and is located farther away in high-dimensional space, meaning that its relationship with other data is relatively weak, the resulting attention weight is smaller, minimizing the impact of the anomalous data on control effectiveness.

[0063] S9: Divide the K matrix dimension by the obtained attention weight result Scale the calculated attention weight matrix Perform a softmax activation function transformation and convert the result into a percentage. The row and column represent the correlation between each high-dimensional data and other data, and the sum of each row is 1.

[0064] S10: The transformed attention weights Multiplying with matrix V, we get Each element in the result is the sum of the attention weights of each sample data corresponding to all other sample data in the 64-dimensional Q,K matrix. Adjust V matrix weights Get the accurate result Attention(Q,K,V).

[0065] S11: Self-attention matrix head i =Attention(Q,K,V) performs multiple parallel trainings to obtain the multi-head attention mechanism result Concat(head1,…,head h ). Combine the multi-head attention mechanism result with the multi-head weight matrix W O Multiply them together to get the multi-head attention result MultiHead(Q,K,V)=Concat(head1,…,head h )W O Each head i It is the result of the attention mechanism in different spaces. It can capture multiple characteristics of the input data in parallel, analyze abnormal data by comparing weight results from multiple spatial angles, and minimize the impact of abnormal data by reducing the weight.

[0066] S12: This paper adopts the Actor-critic architecture of the DDPG algorithm. It mainly consists of two parts: the Actor network and the Critic network. In this paper, the Transformer architecture replaces the traditional reinforcement learning Actor network architecture, and the parameterized strategy μ(s|θ μ ) to output the optimal action a under a given state t =μ(s|θ μ ). Its goal is to maximize the expected cumulative reward: The Critic network is used to evaluate the state-action value function Q(s,a|θ Q ), and optimize its parameters by minimizing the following loss function: L(θ Q )=E (s,a,r,s′)～D [(r+γQ(s′,μ(s′|θ μ )|θ Q )-Q(s,a|θ Q )) 2 ]Where r is the immediate reward, γ is the discount factor, and D is the data in the experience replay buffer.

[0067] S13: The present invention designs an optimal reinforcement learning reward function for the magnetorheological semi-active suspension system control objective. The suspension system has three goals: reducing vehicle acceleration, maintaining handling stability, and maintaining safe driving conditions. To address these three requirements, a reward function is constructed using vehicle acceleration, suspension travel, tire deformation, and suspension dissipative constraints. The vehicle acceleration and suspension travel are weighted by weight coefficients a1 and a2, respectively. Furthermore, obstacle functions P, Q, and N are added. The obstacle function Constrain the suspension travel and keep it within the limit range so that (x s -x u ) at the maximum dynamic travel of the suspension max Within the specified range, the suspension system is prevented from being damaged and the comfort is destroyed. Constrain the deformation of the suspension wheel to maintain a firm contact between the wheel and the road surface, keep it within the allowable range, and make the vehicle reach a safe driving condition. The input of the magnetorheological semi-active suspension is constrained to meet the dissipative constraint requirements.

[0068] S14: To improve system stability and control accuracy, a dual network architecture is adopted. In this architecture, there are two independent Actor networks. and With two independent Cr it ic networks and The minimum action amount and state-action value are calculated using the following formula: The loss function is defined as: The dual Q network can effectively reduce the deviation in Q value estimation, thereby improving the stability of the training process and avoiding strategy oscillation and divergence.

[0069] S15: The present invention also uses the experience replay mechanism and the target network update mechanism to further improve the stability and training efficiency of the system. By storing the data obtained during the interaction with the environment in the replay buffer D, the Transformer mechanism can break the time correlation and randomly sample batches of data from the buffer for training, thereby improving sample utilization. For the dual actor and critical network update mechanisms, the target network parameters θ are maintained and updated with delay. μ ' and θ Q '. Each time the network parameters are updated, the target network uses the soft update rule: Actor network update rule: θ μ '←τθ μ +(1-τ)θ μ ', Crit ic network update rule: θ Q '←τθ Q +(1-τ)θQ '. Where τ is the soft update parameter, which balances the current network parameters with the parameters to be updated. The network parameters are adjusted to obtain the optimal control signal I.

[0070] In addition, in the present invention, in S12, a reinforcement learning control strategy based on the Transformer architecture is used to output a control quantity I in real time in response to the road surface excitation applied to the vehicle, thereby achieving adaptive control of the vehicle suspension.

[0071] In summary, based on the above disclosed method, a suspension convex road simulation experiment is carried out in the present invention.

[0072] Assume that the height of the convex road surface in this experiment is h = 0.065m, the length of the convex road surface is l = 4m, and the vehicle travels through the road surface at a speed of v = 36km / h.

[0073] It is expected that the vehicle body acceleration will be effectively reduced when driving on bumpy roads, the suspension travel will not exceed 0.07, and the dynamic and static load ratio will not exceed the rated limit of 1.

[0074] Based on the above parameters, a lane keeping simulation experiment was conducted and the results were as follows: Figure 4-Figure 6 The simulation results shown are Figure 4 It can be seen that the above control method can effectively realize that the suspension system can effectively reduce the acceleration of the vehicle body when it is excited by the road surface. Figure 5 and Figure 6 It can be seen that the above control method can effectively achieve safe driving and operation of the vehicle.

[0075] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A semi-active suspension control system based on a Transformer architecture, characterized by: The steps include: S1: Deploy sensor networks to collect vehicle dynamic data; S2: Use a preset algorithm to align asynchronous data from different sensors to maintain temporal consistency of heterogeneous data; The measurement value of sensor i at time t is x i (t), align all sensor data to the same time point t k : where t j ≤t k <t j+1 ; S3: Build a quarter vehicle two-degree-of-freedom model where m s is the vehicle body mass, m u is the unsprung mass, F is the damping force of the magnetorheological damper, x s is the vehicle body displacement, x u is the unsprung displacement, x r is the wheel displacement, c is the uncontrollable damping of the damper, k s is the spring stiffness, k t is the wheel stiffness; S4: Modeling of Magnetorheological Damper via Neural Networks: The damper experimental data is obtained through the preset equipment, and the data is fitted through the neural network to obtain the weight matrix w I , w H and w O ; Combining the damper neural network model with the vehicle suspension dynamics model to derive an integrated model; Interact the reinforcement learning structure with the integrated model to train the neural network parameters; S5: Apply the DDPG reinforcement learning algorithm to the state quantity to calculate the control quantity. The DDPG network structure is an Actor-Critic structure, where the Actor network generates actions: a t =μ(s|θ μ ), the Critic network evaluates the state-action value Q(s,a|θ Q ); S6: The state quantity is passed into the Transformer architecture to replace the traditional Actor network, and the correlation of the state quantity is analyzed to distinguish between valid information and state quantity errors. i (t k ) is passed into the DDPG Transformer architecture; S7: Combine the encoded high-dimensional data X with the weights Point multiplication to obtain a weighted matrix Each row of V represents the high-dimensional representation of sensor data; S8: Perform transpose multiplication on the Q and K matrices to obtain QK T matrix; S9: Divide the K matrix dimension by the obtained attention weight result Scaling is performed to calculate the attention weight matrix Perform softmax activation function transformation and convert the result into percentage; S10: The transformed attention weights Multiplying with matrix V, we get Each element in the result is the sum of the attention weights of each sample data corresponding to all other sample data in the 64-dimensional Q,K matrix. Adjust V matrix weights Get the result Attention(Q,K,V); S11: Self-attention matrix head i =Attention(Q,K,V) performs multiple parallel trainings to obtain the multi-head attention mechanism result Concat(head1,…,head h ), the multi-head attention mechanism result is combined with the multi-head weight matrix W O Multiply them together to get the multi-head attention result MultiHead(Q,K,V)=Concat(head1,…,head h )W O , each head i It is the result of the attention mechanism in different spaces; S12: Replace the traditional reinforcement learning actor network architecture with the Transformer architecture and generate a parameterized strategy μ(s|θ μ ) to output the optimal action a under a given state t =μ(s|θ μ ), whose goal is to maximize the expected cumulative reward: Critic network is used to evaluate the state-action value function Q(s,a|θ Q ), and optimize its parameters by minimizing the following loss function: L(θ Q )=E (s,a,r,s′)～D [(r+γQ(s′,μ(s′|θ μ )|θ Q )-Q(s,a|θ Q )) 2 ] Where r is the immediate reward, γ is the discount factor, and D is the data in the experience replay buffer; S13: Design reinforcement learning reward function; S14: uses a dual network architecture, including two independent Actor networks and With two independent critic networks and The minimum action amount and state-action value are calculated using the following formula: The loss function is defined as: S15: The data obtained during the interaction with the environment is stored in the playback buffer D. Randomly sample batches of data from the buffer for training to improve sample utilization. For the dual actor and critic network update mechanism, the target network parameters θ are maintained and updated separately. μ ' and θ Q ', each time the network parameters are updated, the target network uses the soft update rule: Actor network update rule: θ μ '←τθ μ +(1-τ)θ μ ', Critic network update rule: θ Q '←τθ Q +(1-τ)θ Q ', where τ is the soft update parameter, which balances the current network parameters with the parameters to be updated, adjusts the network parameters, and obtains the control signal I.

2. The semi-active suspension control system based on the Transformer architecture according to claim 1, characterized in that: In S1, the sensor network includes an accelerometer, a gyroscope, a displacement sensor, and a pressure sensor.

3. The semi-active suspension control system based on the Transformer architecture according to claim 1, characterized in that: In S1, the vehicle dynamic data includes vehicle body acceleration, suspension travel, and tire deformation.

4. The semi-active suspension control system based on the Transformer architecture according to claim 1, characterized in that: In S4, the preset equipment includes a suspension test bench.

5. The semi-active suspension control system based on the Transformer architecture according to claim 1, characterized in that: In S4, the output value of the neural network is the controllable damping force F of the magnetorheological damper.

6. The semi-active suspension control system based on the Transformer architecture according to claim 1, characterized in that: The state quantity is passed into the Transformer architecture to replace the traditional Actor network, analyze the correlation of the state quantity, distinguish between valid information and state quantity error, and convert the data x i (t k ) is passed into the DDPG Transformer architecture, which includes the following steps: First, the data x i (t k ) Upgraded to 64-dimensional S t , using learnable Fourier feature pairs S t Position code: PE (pos,2i) =sin(pos·w i ),PE (pos,2i+1) =cos(pos·w i ), where w i are learnable parameters.

7. The semi-active suspension control system based on the Transformer architecture according to claim 1, characterized in that: In S4, the integrated model is: In the formula, m s is the vehicle body mass, m u is the unsprung mass, F is the damping force of the magnetorheological damper, x s is the vehicle body displacement, x u is the unsprung displacement, x r is the wheel displacement, c is the uncontrollable damping of the damper, k s is the spring stiffness, k t is the wheel stiffness.

8. The semi-active suspension control system based on the Transformer architecture according to claim 1, characterized in that: In S2, the preset algorithm includes an interpolation algorithm.

9. The semi-active suspension control system based on the Transformer architecture according to claim 1, characterized in that: In S13, the reinforcement learning reward function is The reward function is composed of four parts: vehicle acceleration, suspension travel, tire deformation, and suspension dissipative constraint. The vehicle acceleration and suspension travel are multiplied by weight coefficients a1 and a2 respectively to distribute the weights. In addition, obstacle functions P, Q, and N are added. The obstacle function P constrains the suspension travel and keeps the suspension travel within the limit range, so that (x s -x u ) at the maximum dynamic travel of the suspension max Within the range, the obstacle function Q constrains the suspension wheel deformation, and the obstacle function N constrains the magnetorheological semi-active suspension input.

10. The semi-active suspension control system based on the Transformer architecture according to claim 9, characterized in that: The barrier function The barrier function The barrier function

Images