Air suspension and ESC coordinated control method and system based on deep learning

Abstract

The present invention discloses a method and system for coordinated control of air suspension and ESC based on deep learning, which belongs to the field of automobile braking and includes the following steps: S1, collecting historical data of vehicle driving, and using vehicle dynamics simulation software to generate extreme working condition data to form a training data set; S2, building a dual-branch fusion control model and training it; S3, collecting vehicle state parameters and road condition data in real time; S4, extracting and fusing the processed multi-source parameters, and inputting the trained dual-branch fusion control model to generate an optimal control strategy; S5, ESC executing the optimal control strategy and collecting the vehicle state data after execution to form a control closed loop. The above-mentioned method and system for coordinated control of air suspension and ESC based on deep learning, combined with deep learning and vehicle dynamics model, can monitor the vehicle motion state in real time and make precise manipulations based on real-time data to avoid potential loss of control.

Description

Technical Field

[0001] The present invention relates to the field of automobile braking technology, and in particular to a deep learning-based air suspension and ESC coordinated control method and system. Background Art

[0002] Amid the wave of intelligent and electrified vehicles, chassis control technology is undergoing profound changes, gradually evolving from the traditional independent control of single subsystems to the coordinated optimization of multiple systems. In this evolution, air suspension and electronic stability control (ESC) have emerged as key components enhancing vehicle safety, comfort, and handling. The integration and intelligent upgrade of these technologies has become a research focus in the current automotive industry.

[0003] In recent years, both academia and industry have actively explored the coordinated control of air suspension and ESC. Numerous studies have attempted to optimize suspension stiffness and ESC braking force distribution by leveraging model predictive control (MPC) or fuzzy logic algorithms, combining vehicle posture and tire force information. These efforts have yielded some results, providing new ideas and methods for improving vehicle performance.

[0004] However, it cannot be ignored that existing collaborative control methods still face significant challenges. For one thing, the complexity of data fusion is a prominent issue. During vehicle operation, data from multiple, heterogeneous sensors, such as inertial measurement units, wheel speed sensors, and pressure sensors, is complex and has diverse characteristics. Real-time processing of this data requires not only efficient data processing algorithms but also high-precision dynamic state estimation. However, there is still significant room for improvement in this area, and achieving both the real-time performance of data processing and the accuracy of dynamic state estimation remains difficult to achieve.

[0005] On the other hand, model adaptability is limited. Traditional cooperative control is often based on physical models, which exhibit significant shortcomings when faced with complex and changing real-world operating conditions. This is particularly true in extreme nonlinear and highly coupled scenarios, such as high-speed cornering, emergency braking, and slippery roads. Traditional physical models often struggle to accurately describe the vehicle's dynamic characteristics, leading to the failure of cooperative control strategies and an inability to effectively ensure vehicle safety and stability. Summary of the Invention

[0006] The purpose of the present invention is to provide an air suspension and ESC coordinated control method and system based on deep learning to solve the above technical problems.

[0007] To achieve the above objectives, the present invention provides a method for coordinated control of air suspension and ESC based on deep learning, comprising the following steps:

[0008] S1. Collect historical vehicle driving data and use vehicle dynamics simulation software to generate extreme operating condition data to form a training data set;

[0009] S2. Build a dual-branch fusion control model based on the deep learning model, define the control targets of the air suspension and ESC actuators, and then use the training dataset to train the dual-branch fusion control model.

[0010] S3, Multi-source parameter acquisition: Real-time acquisition of vehicle status parameters and road condition data, followed by data cleaning and time alignment using linear interpolation;

[0011] S4. After feature extraction and fusion of the processed multi-source parameters, the parameters are input into the trained dual-branch fusion control model to generate the optimal control strategy;

[0012] S5. ESC executes the optimal control strategy and collects vehicle status data after execution to form a control closed loop.

[0013] Preferably, the vehicle travel history data in step S1 includes historical vehicle state parameters and road condition data;

[0014] The historical vehicle state parameters described in step S1 and the vehicle state parameters described in step S3 both include vehicle dynamics parameters, air suspension and ESC actuator state parameters; wherein, the vehicle dynamics parameters include vehicle body three-axis acceleration, vehicle speed, steering wheel angle, yaw angular velocity, pitch angle, roll angle and wheel speed; the air suspension and ESC actuator state parameters include air suspension stiffness and execution feedback signals, and the execution feedback signals include air suspension height adjustment amount, braking torque distribution result, and engine torque; the road condition data includes road friction coefficient estimation and road roughness.

[0015] Preferably, an LSTM model is used to learn temporal features and static features to estimate the real-time road friction coefficient;

[0016] The input layer of the LSTM model is used to input time series features and static features. The time series features include slip rate, lateral force, and vertical load, and the static features include road surface type code and road surface moisture. The LSTM layer of the LSTM model is used to capture time series dependencies. The output layer of the LSTM model is used to output the current friction coefficient estimate.

[0017] Moreover, the training objective of the LSTM model is set to minimize the mean square error between the estimated value of the predicted friction coefficient and the true value of the friction coefficient:

[0018] (1);

[0019] Where, represents the mean square error loss function; Indicates the number of samples; Indicates the Estimated friction coefficient for samples; Indicates the The true value of the friction coefficient of the samples.

[0020] Preferably, the dual-branch fusion control model in step S2 includes an ESC control sub-model and an air suspension control sub-model, and the ESC control sub-model and the air suspension control sub-model communicate with each other via a bus to form a joint observation , represents the state space of the air suspension control submodel, Represents the state space of the ESC control submodel;

[0021] Among them, the state space of the air suspension control sub-model is , Indicates the vertical acceleration of the vehicle body; Indicates the dynamic deflection of the air suspension; Indicates the vertical displacement of the vehicle body; Represents the road roughness spectrum; the action space is the main force of the suspension Or damping adjustment coefficient , and its action value , Represents an array;

[0022] The state space of the ESC control submodel , Indicates the steering wheel angle; represents the yaw angular velocity, and ; represents the lateral acceleration; Represents the wheel slip rate; the action space is the braking torque distribution coefficient or stability control torque , and its action value ;

[0023] The multi-objective weighted loss function of the ESC control sub-model is set as follows:

[0024] (2);

[0025] Where, represents the total loss of the ESC control sub-model; 、 and Both represent weight coefficients; Represents the safety loss, and its optimization goal is to minimize the sideslip angle and yaw rate error ; Represents the first comfort loss, and its optimization goal is to penalize the brake pressure change rate under the premise of safety. Smooth brake pressure changes, Indicates the time period Pressure change value inside; represents a loss of compliance;

[0026] and minimize the sideslip angle and yaw rate error The expression is as follows:

[0027] (3);

[0028] (4);

[0029] Where, Indicates vehicle speed, Indicates wheelbase; represents the reference yaw rate;

[0030] The multi-objective weighted loss function of the air suspension control sub-model is set as follows:

[0031] (5);

[0032] Where, represents the total loss of the air suspension control sub-model; Represents the second comfort loss, and its optimization goal is to minimize the vertical acceleration of the vehicle body and pitch angular velocity The root mean square value of ; Represents the loss of missing out, and its penalty term is , represents the vertical displacement of the wheel, Indicates the vertical displacement of the road surface; 、 and Both represent weight coefficients; Indicates energy loss;

[0033] And minimize the vertical acceleration of the vehicle body and pitch angular velocity The root mean square value of The expression is as follows:

[0034] (6);

[0035] Where, Indicates the The vertical acceleration of the vehicle body of each sample; Indicates the The pitch angular velocity of samples; Indicates the sample size.

[0036] Preferably, in step S2, the dual-branch fusion control model is set to adopt a centralized training-distributed execution strategy: using the MADDPG framework, the central critic network collects the state space information and action space information of the ESC control sub-model and the air suspension control sub-model; the actor network generates actions based on the state space information of the ESC control sub-model or the air suspension control sub-model;

[0037] Set the optimizer to AdamW or RMSprop, and use a sliding window or sequence generator to implement real-time data stream training.

[0038] Preferably, the linear interpolation expression in step S3 is as follows:

[0039] (7);

[0040] Where, Represents the time-aligned data; and Respectively represent two original data at adjacent moments; and Represent two timestamps at adjacent moments; Indicates the target timestamp.

[0041] Preferably, step S4 specifically includes the following steps:

[0042] S41. Feature extraction: Based on the vehicle state parameters, use LSTM or Transformer to perform time series modeling and extract vehicle state time series features;

[0043] For road condition data, local features are extracted through the CNN network;

[0044] S42, inputting the vehicle state time series features and local features into the Transformer, and using the Transformer to capture the global time series association;

[0045] S421, set the feature sequence composed of time series features and local features , Represents the feature sequence Middle The features at the moment will be input into the feature sequence of Transformer Linear transformation to query matrix , key matrix Sum Matrix :

[0046] (8);

[0047] Where, 、 and denotes attention weights, and , Represents the model dimension; represents the attention head dimension;

[0048] S422. Dynamically assign attention weights by scaling dot product attention:

[0049] (9);

[0050] S423. Use the multi-head attention mechanism to concatenate the results of multiple attention heads and perform linear transformation:

[0051] (10);

[0052] Where, Represents the output of the multi-head attention mechanism after concatenation and linear transformation; Respectively represent The output of the attention head is Indicates the number of attention heads; represents the output weight, and , The dimension representing the value;

[0053] S43, feature fusion: The global temporal association captured by Transformer is combined with the vehicle state time series features and local features to obtain fused features;

[0054] S44. Input the fusion features into the dual-branch fusion control model to generate the optimal control strategy.

[0055] Preferably, in step S41, for the road condition data, firstly, a plurality of convolution kernels of a CNN network are used to extract local features:

[0056] (11);

[0057] Where, Represents the local features of the output; represents the activation function; Represents the convolution kernel weight; Indicates that the input sequence is data points; Represents the convolution kernel size; represents the bias term;

[0058] Then the dimension is reduced by the maximum pooling:

[0059] (12);

[0060] Where, Represents the result after maximum pooling dimensionality reduction; Respectively represent the pooling window the characteristics of the moment; Indicates the pooling window size.

[0061] Preferably, the optimal control strategy generated in step S44 includes a suspension control strategy and an ESC control strategy, wherein the suspension control strategy includes a vehicle height control strategy and a suspension stiffness adjustment strategy. The vehicle height control strategy is expressed as follows:

[0062] (13);

[0063] Where, Indicates driving mode, and ,when It indicates comfort mode when When it is in sports mode, : indicates off-road mode; Indicates unequal levels of road surface; 、 、 、 、 and All represent fitting coefficients; represents the dynamic compensation term, and , and denote the lateral acceleration and longitudinal acceleration at the center of mass, respectively. and Represents wheelbase and track width respectively. represents the acceleration due to gravity;

[0064] The suspension stiffness adjustment strategy expression is as follows:

[0065] (14);

[0066] Where, Indicates the adjusted suspension stiffness; represents the foundation stiffness; and All represent empirical coefficients;

[0067] The ESC control strategy is to adjust the braking force distribution and torque, where

[0068] The braking force distribution adjustment formula is as follows:

[0069] (15);

[0070] Where, Indicates the front axle braking force; and are the partition coefficients, and ; Indicates the total braking force; Indicates the rear axle braking force;

[0071] The torque adjustment formula is as follows:

[0072] (16);

[0073] Where, Indicates the torque transmitted to the drive wheels; It means controlling the engine torque by adjusting the throttle opening; Indicates the transmission ratio; Indicates efficiency.

[0074] A system for executing a deep learning-based coordinated control method for air suspension and ESC, comprising:

[0075] The historical data collection module is used to collect historical vehicle driving data and use vehicle dynamics simulation software to generate extreme working condition data to form a training data set;

[0076] A model building module is used to build a dual-branch fusion control model based on the deep learning model, define the control objectives of the air suspension and ESC actuators, and then use the training dataset to train the dual-branch fusion control model;

[0077] The real-time data acquisition module is used to collect vehicle status parameters and road condition data in real time, and perform data cleaning and time alignment processing using linear interpolation. It includes a brake pressure sensor, wheel speed sensor, inertial measurement unit, gyroscope, yaw rate sensor, road surface perception sensor and suspension status sensor;

[0078] The optimal strategy generation module is used to extract and fuse the processed multi-source parameters, input them into the trained dual-branch fusion control model, and generate the optimal control strategy;

[0079] In the execution module, ESC executes the optimal control strategy and collects vehicle status data after execution to form a control closed loop.

[0080] Therefore, the present invention adopts the above-mentioned deep learning-based air suspension and ESC coordinated control method and system, which has the following beneficial effects:

[0081] 1. In terms of comfort, the system automatically adjusts suspension stiffness and damping based on road stimuli and driving mode. Combined with ESC's braking torque optimization, it significantly reduces body vibration and enhances steering sensitivity.

[0082] 2. In terms of safety, the system achieves rapid response to extreme conditions such as skidding and tire blowouts through active suspension stiffness and damping adjustment and ESC predictive braking force distribution. It also supports cross-vehicle compatibility (sedans, SUVs, trucks) and cross-environmental compatibility (deserts, snow, mud).

[0083] 3. Breaking through the limitations of traditional single-subsystem control, this system achieves deep collaboration between the air suspension (mechanical system) and the ESC (electronic control system) for the first time. This system resolves performance conflicts under complex operating conditions through multi-physics coupling control (such as combined optimization of stiffness, damping, and braking torque).

[0084] 4. The system's online incremental learning and reinforcement learning capabilities based on deep learning give it the potential for lifelong optimization and can dynamically adapt to driver habits and environmental changes. Its performance and scalability lay the foundation for future intelligent chassis technologies (such as autonomous driving scenarios).

[0085] In summary, the present invention achieves a comprehensive improvement in vehicle performance through real-time data fusion and dynamic decision-making.

[0086] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 This is a flowchart of the deep learning-based air suspension and ESC coordinated control method of the present invention. DETAILED DESCRIPTION

[0088] In order to make the purposes, technical solutions and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, where the same or similar numbers throughout represent the same or similar elements or elements with the same or similar functions.

[0089] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or devices.

[0090] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0091] like Figure 1 As shown, a deep learning-based coordinated control method for air suspension and ESC includes the following steps:

[0092] S1. Collect historical vehicle driving data and use vehicle dynamics simulation software (such as CarSim / Simulink co-simulation) to generate extreme operating condition data to form a training dataset;

[0093] The vehicle travel history data in step S1 includes historical vehicle state parameters and road condition data;

[0094] The historical vehicle state parameters described in step S1 and the vehicle state parameters described in step S3 both include vehicle dynamics parameters, air suspension and ESC actuator state parameters; wherein, the vehicle dynamics parameters include vehicle body three-axis acceleration, vehicle speed, steering wheel angle, yaw angular velocity, pitch angle, roll angle and wheel speed; the air suspension and ESC actuator state parameters include air suspension stiffness and execution feedback signals, and the execution feedback signals include air suspension height adjustment amount, braking torque distribution result, and engine torque; the road condition data includes road friction coefficient estimation and road roughness.

[0095] The LSTM model is used to learn temporal features and static features to estimate the real-time road friction coefficient;

[0096] The input layer of the LSTM model is used to input time series features and static features. The time series features include slip rate, lateral force, and vertical load, and the static features include road surface type code and road surface moisture. The LSTM layer of the LSTM model is used to capture time series dependencies. The output layer of the LSTM model is used to output the current friction coefficient estimate.

[0097] Moreover, the training objective of the LSTM model is set to minimize the mean square error between the estimated value of the predicted friction coefficient and the true value of the friction coefficient:

[0098] (1);

[0099] Where, represents the mean square error loss function; Indicates the number of samples; Indicates the Estimated friction coefficient for samples; Indicates the The true value of the friction coefficient of the samples.

[0100] S2. Build a dual-branch fusion control model based on the deep learning model, define the control targets of the air suspension and ESC actuators, and then use the training dataset to train the dual-branch fusion control model.

[0101] The dual-branch fusion control model described in step S2 includes an ESC control sub-model and an air suspension control sub-model, and the ESC control sub-model and the air suspension control sub-model communicate via a bus to form a joint observation , represents the state space of the air suspension control submodel, Represents the state space of the ESC control submodel;

[0102] Among them, the state space of the air suspension control sub-model is , Indicates the vertical acceleration of the vehicle body; Indicates the dynamic deflection of the air suspension; Indicates the vertical displacement of the vehicle body; Represents the road roughness spectrum; the action space is the main force of the suspension Or damping adjustment coefficient , and its action value , Represents an array;

[0103] The state space of the ESC control submodel , Indicates the steering wheel angle; represents the yaw angular velocity, and ; represents the lateral acceleration; Represents the wheel slip rate; the action space is the braking torque distribution coefficient or stability control torque , and its action value ;

[0104] The multi-objective weighted loss function of the ESC control sub-model is set as follows:

[0105] (2);

[0106] Where, represents the total loss of the ESC control sub-model; 、 and Both represent weight coefficients; Represents the safety loss, and its optimization goal is to minimize the sideslip angle and yaw rate error ; Represents the first comfort loss, and its optimization goal is to penalize the brake pressure change rate under the premise of safety. Smooth brake pressure changes, Indicates the time period Pressure change value inside; represents a loss of compliance;

[0107] In this embodiment, the weight coefficients of the multi-objective weighted loss function are dynamically adjusted according to the driving mode or the real-time state of the vehicle. ( represents the maximum sideslip angle value), the safety weight Automatically increases to above 0.8; when in comfort mode, the comfort weight Prioritize allocation to the 0.6-0.7 range, and receive driving mode signals and side slip detection flags in real time through the vehicle bus to trigger weight adjustment.

[0108] and minimize the sideslip angle and yaw rate error The expression is as follows:

[0109] (3);

[0110] (4);

[0111] Where, Indicates vehicle speed, Indicates wheelbase; represents the reference yaw rate;

[0112] The multi-objective weighted loss function of the air suspension control sub-model is set as follows:

[0113] (5);

[0114] Where, represents the total loss of the air suspension control sub-model; Represents the second comfort loss, and its optimization goal is to minimize the vertical acceleration of the vehicle body and pitch angular velocity The root mean square value of ; Represents the loss of missing out, and its penalty term is , represents the vertical displacement of the wheel, Indicates the vertical displacement of the road surface; 、 and Both represent weight coefficients; Indicates energy loss;

[0115] And minimize the vertical acceleration of the vehicle body and pitch angular velocity The root mean square value of The expression is as follows:

[0116] (6);

[0117] Where, Indicates the The vertical acceleration of the vehicle body of each sample; Indicates the The pitch angular velocity of samples; Indicates the sample size.

[0118] In step S2, the dual-branch fusion control model adopts a centralized training-distributed execution strategy: using the MADDPG framework, the central critic network collects state space information and action space information of the ESC control sub-model and the air suspension control sub-model; the actor network generates actions based on the state space information of the ESC control sub-model or the air suspension control sub-model;

[0119] Set the optimizer to AdamW or RMSprop, and use a sliding window or sequence generator to implement real-time data stream training.

[0120] In this embodiment, the trained model is deployed to the vehicle computing unit (DCU), and low-latency inference (≤10ms) is achieved through TensorRT quantization, pruning and other technologies.

[0121] S3, Multi-source parameter acquisition: Real-time acquisition of vehicle status parameters and road condition data, followed by data cleaning and time alignment using linear interpolation;

[0122] The linear interpolation expression in step S3 is as follows:

[0123] (7);

[0124] Where, Represents the time-aligned data; and Respectively represent two original data at adjacent moments; and Represent two timestamps at adjacent moments; Indicates the target timestamp.

[0125] In this embodiment, the data cleaning includes using Kalman filtering to eliminate drift of the inertial measurement unit (IMU) and using moving average filtering to smooth high-frequency vibration signals. The time alignment processing ensures that the synchronization accuracy of multi-sensor data is within ±1ms, and generates virtual frame data that matches the target timestamp through linear interpolation.

[0126] S4. After feature extraction and fusion of the processed multi-source parameters, the parameters are input into the trained dual-branch fusion control model to generate the optimal control strategy;

[0127] Step S4 specifically includes the following steps:

[0128] S41. Feature extraction: Based on the vehicle state parameters, use LSTM or Transformer to perform time series modeling and extract vehicle state time series features;

[0129] For road condition data, local features are extracted through the CNN network;

[0130] S42, inputting the vehicle state time series features and local features into the Transformer, and using the Transformer to capture the global time series association;

[0131] S421, set the feature sequence composed of time series features and local features , Represents the feature sequence Middle The features at the moment will be input into the feature sequence of Transformer Linear transformation to query matrix , key matrix Sum Matrix :

[0132] (8);

[0133] Where, 、 and denotes attention weights, and , Represents the model dimension; represents the attention head dimension;

[0134] S422. Dynamically assign attention weights by scaling dot product attention:

[0135] (9);

[0136] S423. Use the multi-head attention mechanism to concatenate the results of multiple attention heads and perform linear transformation:

[0137] (10);

[0138] Where, Represents the output of the multi-head attention mechanism after concatenation and linear transformation; Respectively represent The output of the attention head is Indicates the number of attention heads; represents the output weight, and , The dimension representing the value;

[0139] S43, feature fusion: The global temporal association captured by Transformer is combined with the vehicle state time series features and local features to obtain fused features;

[0140] S44. Input the fusion features into the dual-branch fusion control model to generate the optimal control strategy.

[0141] Preferably, in step S41, for the road condition data, firstly, a plurality of convolution kernels of a CNN network are used to extract local features:

[0142] (11);

[0143] Where, Represents the local features of the output; represents the activation function; Represents the convolution kernel weight; Indicates that the input sequence is data points; Represents the convolution kernel size; represents the bias term;

[0144] Then the dimension is reduced by the maximum pooling:

[0145] (12);

[0146] Where, Represents the result after maximum pooling dimensionality reduction; Respectively represent the pooling window the characteristics of the moment; Indicates the pooling window size.

[0147] Preferably, the optimal control strategy generated in step S44 includes a suspension control strategy and an ESC control strategy, wherein the suspension control strategy includes a vehicle height control strategy and a suspension stiffness adjustment strategy. The vehicle height control strategy is expressed as follows:

[0148] (13);

[0149] Where, Indicates driving mode, and ,when It indicates comfort mode when When it is in sports mode, : indicates off-road mode; Indicates unequal levels of road surface; 、 、 、 、 and All represent fitting coefficients; represents the dynamic compensation term, and , and denote the lateral acceleration and longitudinal acceleration at the center of mass, respectively. and Represents wheelbase and track width respectively. represents the acceleration due to gravity;

[0150] The suspension stiffness adjustment strategy expression is as follows:

[0151] (14);

[0152] Where, represents the adjusted suspension stiffness, and ; represents the foundation stiffness; and All represent empirical coefficients;

[0153] The ESC control strategy is to adjust the braking force distribution and torque adjustment, where

[0154] The braking force distribution adjustment formula is as follows:

[0155] (15);

[0156] Where, Indicates the front axle braking force; and are the partition coefficients, and , ; Indicates the total braking force; Indicates the rear axle braking force;

[0157] The torque adjustment formula is as follows:

[0158] (16);

[0159] Where, Indicates the torque transmitted to the drive wheels; It means controlling the engine torque by adjusting the throttle opening; Indicates the transmission ratio; represents efficiency, and .

[0160] In this embodiment, safety redundancy is also set in the rule layer of the deep learning model. The safety redundancy includes ESC control redundancy and suspension control redundancy. The ESC control redundancy adjusts the braking pressure of a single wheel through PID so that the actual yaw rate tracks the reference value:

[0161] (17);

[0162] Where, represents the additional yaw moment; 、 and They represent the proportional, integral, and differential coefficients respectively.

[0163] Suspension control redundancy includes passive suspension control and semi-active suspension control. Passive suspension control adjusts the shock absorber damping based on the PID algorithm (such as formula (15)), and the response delay is relatively large (>50ms). Active suspension control adjusts the damping through linear mapping of the solenoid valve opening, which includes the following steps:

[0164] Step 1: Determine the damping adjustment variable: According to the system requirements, select the physical quantity directly related to the damping (such as solenoid valve opening, coil current, valve core displacement, etc.) as the control variable. For example, establish the solenoid valve opening through experiments or simulations. and damping force The linear relationship:

[0165] (18);

[0166] Where, represents the proportionality coefficient, Indicates the offset.

[0167] Step 2: Calibrate the opening-damping curve: On the test bench, change the opening of the solenoid valve (such as adjusting the valve core displacement or coil current) and record the corresponding damping force data.

[0168] Step 3: Use the least squares method to fit the data and obtain the linear mapping relationship between opening and damping.

[0169] S5. ESC executes the optimal control strategy and collects vehicle status data after execution to form a control closed loop.

[0170] A system for executing a deep learning-based coordinated control method for air suspension and ESC, comprising:

[0171] The historical data collection module is used to collect historical vehicle driving data and use vehicle dynamics simulation software to generate extreme working condition data to form a training data set;

[0172] A model building module is used to build a dual-branch fusion control model based on the deep learning model, define the control objectives of the air suspension and ESC actuators, and then use the training dataset to train the dual-branch fusion control model;

[0173] The real-time data acquisition module is used to collect vehicle status parameters and road condition data in real time, and perform data cleaning and time alignment processing using linear interpolation. It includes a brake pressure sensor, a wheel speed sensor, an inertial measurement unit, a gyroscope, a yaw rate sensor, a road surface perception sensor, and a suspension status sensor. The road surface perception sensor includes a lidar for identifying road roughness and a camera for visually perceiving road texture / water accumulation. The suspension status sensor includes an air spring pressure sensor, a solenoid valve displacement sensor, and a damper current feedback sensor for monitoring damper action.

[0174] The optimal strategy generation module is used to extract and fuse the processed multi-source parameters, input them into the trained dual-branch fusion control model, and generate the optimal control strategy;

[0175] In the execution module, ESC executes the optimal control strategy and collects vehicle status data after execution to form a control closed loop.

[0176] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based coordinated control method for air suspension and ESC, characterized by: The following steps are involved: S1. Collect historical vehicle driving data and use vehicle dynamics simulation software to generate extreme operating condition data to form a training data set; S2. Build a dual-branch fusion control model based on the deep learning model, define the control targets of the air suspension and ESC actuators, and then use the training dataset to train the dual-branch fusion control model. The dual-branch fusion control model described in step S2 includes an ESC control sub-model and an air suspension control sub-model, and the ESC control sub-model and the air suspension control sub-model communicate via a bus to form a joint observation , represents the state space of the air suspension control submodel, Represents the state space of the ESC control submodel; Among them, the state space of the air suspension control sub-model is , Indicates the vertical acceleration of the vehicle body; Indicates the dynamic deflection of the air suspension; Indicates the vertical displacement of the vehicle body; Represents the road roughness spectrum; the action space is the main force of the suspension Or damping adjustment coefficient , and its action value , Represents an array; The state space of the ESC control submodel , Indicates the steering wheel angle; represents the yaw angular velocity, and ; represents the lateral acceleration; Represents the wheel slip rate; the action space is the braking torque distribution coefficient or stability control torque , and its action value ; The multi-objective weighted loss function of the ESC control sub-model is set as follows: （2）； Where, represents the total loss of the ESC control sub-model; 、 and Both represent weight coefficients; Represents the safety loss, and its optimization goal is to minimize the sideslip angle and yaw rate error ; Represents the first comfort loss, and its optimization goal is to penalize the brake pressure change rate under the premise of safety. Smooth brake pressure changes, Indicates the time period Pressure change value inside; represents a loss of compliance; and minimize the sideslip angle and yaw rate error The expression is as follows: （3）； （4）； Where, Indicates vehicle speed, Indicates wheelbase; represents the reference yaw rate; The multi-objective weighted loss function of the air suspension control sub-model is set as follows: （5）； Where, represents the total loss of the air suspension control sub-model; Represents the second comfort loss, and its optimization goal is to minimize the vertical acceleration of the vehicle body and pitch angular velocity The root mean square value of ; Represents the loss of missing out, and its penalty term is , represents the vertical displacement of the wheel, Indicates the vertical displacement of the road surface; 、 and Both represent weight coefficients; Indicates energy loss; And minimize the vertical acceleration of the vehicle body and pitch angular velocity The root mean square value of The expression is as follows: （6）； Where, Indicates the The vertical acceleration of the vehicle body of each sample; Indicates the The pitch angular velocity of samples; represents the number of samples; S3, Multi-source parameter acquisition: Real-time acquisition of vehicle status parameters and road condition data, followed by data cleaning and time alignment using linear interpolation; S4. After feature extraction and fusion of the processed multi-source parameters, the parameters are input into the trained dual-branch fusion control model to generate the optimal control strategy; Step S4 specifically includes the following steps: S41. Feature extraction: Based on the vehicle state parameters, use LSTM or Transformer to perform time series modeling and extract vehicle state time series features; For road condition data, local features are extracted through the CNN network; S42, inputting the vehicle state time series features and local features into the Transformer, and using the Transformer to capture the global time series association; S421, set the feature sequence composed of time series features and local features , Represents the feature sequence Middle The features at the moment will be input into the feature sequence of Transformer Linear transformation to query matrix , key matrix Sum Matrix : （8）； Where, 、 and denotes attention weights, and , Represents the model dimension; represents the attention head dimension; S422. Dynamically assign attention weights by scaling dot product attention: （9）； S423. Use the multi-head attention mechanism to concatenate the results of multiple attention heads and perform linear transformation: （10）； Where, Represents the output of the multi-head attention mechanism after concatenation and linear transformation; Respectively represent The output of the attention head is Indicates the number of attention heads; represents the output weight, and , The dimension representing the value; S43, feature fusion: The global temporal association captured by Transformer is combined with the vehicle state time series features and local features to obtain fused features; S44, inputting the fusion features into a dual-branch fusion control model to generate an optimal control strategy; The optimal control strategy generated in step S44 includes a suspension control strategy and an ESC control strategy. The suspension control strategy includes a vehicle height control strategy and a suspension stiffness adjustment strategy. The vehicle height control strategy is expressed as follows: （13）； Where, Indicates driving mode, and ,when It indicates comfort mode when Indicates sports mode when : indicates off-road mode; Indicates unequal levels of road surface; 、 、 、 、 and All represent fitting coefficients; represents the dynamic compensation term, and , and denote the lateral acceleration and longitudinal acceleration at the center of mass, respectively. and Represents wheelbase and track width respectively. represents the acceleration due to gravity; The suspension stiffness adjustment strategy expression is as follows: （14）； Where, Indicates the adjusted suspension stiffness; represents the foundation stiffness; and All represent empirical coefficients; The ESC control strategy is to adjust the braking force distribution and torque, where The braking force distribution adjustment formula is as follows: （15）； Where, Indicates the front axle braking force; and are the partition coefficients, and ; Indicates the total braking force; Indicates the rear axle braking force; The torque adjustment formula is as follows: （16）； Where, Indicates the torque transmitted to the drive wheels; It means controlling the engine torque by adjusting the throttle opening; Indicates the transmission ratio; Indicates efficiency; S5. ESC executes the optimal control strategy and collects vehicle status data after execution to form a control closed loop.

2. The deep learning-based coordinated control method for air suspension and ESC according to claim 1, characterized in that: The vehicle travel history data in step S1 includes historical vehicle state parameters and road condition data; The historical vehicle state parameters described in step S1 and the vehicle state parameters described in step S3 both include vehicle dynamics parameters, air suspension and ESC actuator state parameters; wherein, the vehicle dynamics parameters include vehicle body three-axis acceleration, vehicle speed, steering wheel angle, yaw angular velocity, pitch angle, roll angle and wheel speed; the air suspension and ESC actuator state parameters include air suspension stiffness and execution feedback signals, and the execution feedback signals include air suspension height adjustment amount, braking torque distribution result, and engine torque; the road condition data includes road friction coefficient estimation and road roughness.

3. The deep learning-based coordinated control method for air suspension and ESC according to claim 2, characterized in that: The LSTM model is used to learn temporal features and static features to estimate the real-time road friction coefficient; The input layer of the LSTM model is used to input time series features and static features. The time series features include slip rate, lateral force, and vertical load, and the static features include road surface type code and road surface moisture. The LSTM layer of the LSTM model is used to capture time series dependencies. The output layer of the LSTM model is used to output the current friction coefficient estimate. Moreover, the training objective of the LSTM model is set to minimize the mean square error between the estimated value of the predicted friction coefficient and the true value of the friction coefficient: （1）； Where, represents the mean square error loss function; Indicates the number of samples; Indicates the Estimated friction coefficient for samples; Indicates the The true value of the friction coefficient of the samples.

4. The deep learning-based coordinated control method for air suspension and ESC according to claim 3, characterized in that: In step S2, the dual-branch fusion control model adopts a centralized training-distributed execution strategy: using the MADDPG framework, the central critic network collects state space information and action space information of the ESC control sub-model and the air suspension control sub-model; the actor network generates actions based on the state space information of the ESC control sub-model or the air suspension control sub-model; Set the optimizer to AdamW or RMSprop, and use a sliding window or sequence generator to implement real-time data stream training.

5. The deep learning-based coordinated control method for air suspension and ESC according to claim 3, characterized in that: The linear interpolation expression in step S3 is as follows: （7）； Where, Represents the time-aligned data; and Respectively represent two original data at adjacent moments; and Represent two timestamps at adjacent moments; Indicates the target timestamp.

6. The deep learning-based coordinated control method for air suspension and ESC according to claim 5, characterized in that: In step S41, for the road condition data, local features are first extracted using multiple convolution kernels of the CNN network: （11）； Where, Represents the local features of the output; represents the activation function; Represents the convolution kernel weight; Indicates that the input sequence is data points; Represents the convolution kernel size; represents the bias term; Then the dimension is reduced by the maximum pooling: （12）； Where, Represents the result after maximum pooling dimensionality reduction; Respectively represent the pooling window the characteristics of the moment; Indicates the pooling window size.

7. A system for executing the deep learning-based coordinated control method for air suspension and ESC according to any one of claims 1 to 6, characterized in that: include: The historical data collection module is used to collect historical vehicle driving data and use vehicle dynamics simulation software to generate extreme working condition data to form a training data set; A model building module is used to build a dual-branch fusion control model based on the deep learning model, define the control objectives of the air suspension and ESC actuators, and then use the training dataset to train the dual-branch fusion control model; The real-time data acquisition module is used to collect vehicle status parameters and road condition data in real time, and perform data cleaning and time alignment processing using linear interpolation. It includes a brake pressure sensor, wheel speed sensor, inertial measurement unit, gyroscope, yaw rate sensor, road surface perception sensor and suspension status sensor; The optimal strategy generation module is used to extract and fuse the processed multi-source parameters, input them into the trained dual-branch fusion control model, and generate the optimal control strategy; In the execution module, ESC executes the optimal control strategy and collects vehicle status data after execution to form a control closed loop.

Images