An adaptive height control method for electronically controlled air suspension based on an independent suspension lateral coupling model

By constructing a multi-timescale neural network adaptive control method based on a double wishbone independent suspension lateral coupling model, the problems of insufficient suspension height adjustment accuracy and large air pressure fluctuation in the existing technology are solved, and rapid and accurate adjustment and stability improvement are achieved under multiple working conditions.

CN123104301APending Publication Date: 2026-11-17ZHENGZHOU UNIVERSITY OF LIGHT INDUSTRY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611201348.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing electronically controlled air suspension height control methods neglect the lateral kinematic coupling effect of double wishbone independent suspension, resulting in insufficient height adjustment accuracy, large air pressure fluctuations, and poor adaptability to operating conditions. Furthermore, the controller parameters are difficult to tune, making it difficult to achieve effective adjustment in multi-timescale systems.

Method used

A height-adaptive control method based on a lateral coupling model of a double wishbone independent suspension is constructed. A multi-time-scale neural network and reinforcement learning mechanism are adopted. The system is decomposed into slow-variable subsystems and fast-variable subsystems through singular perturbation theory. An adaptive control architecture is designed, and parameter calibration and online identification are performed in combination with real vehicle test data to achieve real-time compensation of the lateral coupling coefficient and optimization of control parameters.

Benefits of technology

It significantly improves the accuracy and robustness of suspension height control, enabling rapid and precise height adjustment under various operating conditions, suppressing height adjustment deviations caused by lateral coupling effects, and improving air pressure regulation quality and system stability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123104301A_ABST
    Figure CN123104301A_ABST
Patent Text Reader

Abstract

This invention discloses an adaptive height control method for electronically controlled air suspension based on a lateral coupling model of an independent suspension, belonging to the field of automotive electronically controlled suspension technology. The method first establishes a lateral coupling kinematic-dynamic mathematical model of a double wishbone independent suspension, quantifying the coupling relationship between wheel center lateral displacement and suspension vertical bounce. Secondly, considering the differences between fast and slow states, a multi-timescale dynamic neural network dynamic equation is designed, with the slow network responsible for learning the long-term laws of vehicle height changes and the fast network responsible for learning the suspension vibration and transient air pressure response. Thirdly, the performance objective function is set as a weighted sum of vehicle height tracking error, wheel center lateral displacement, air pressure fluctuation, and solenoid valve control flow. Finally, an adaptive reinforcement learning architecture for electronically controlled air suspension is constructed, and multi-timescale neural network adaptive architecture training is performed. The evaluator network is then approximated to the optimal performance index function, and time-varying parameters such as the lateral coupling coefficient are identified online to calculate the optimal control input. This method deploys offline pre-training and online adaptive training to the vehicle's electronic control unit under multiple operating conditions, enabling precise, rapid, and adaptive adjustment of the height of the electronically controlled air suspension under all operating conditions. It effectively suppresses the height deviation caused by the lateral coupling effect of the independent suspension under uneven road surface excitation, and significantly improves control accuracy and system robustness.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of automotive electronic suspension technology, specifically relating to a height adaptive control method for electronically controlled air suspension, which is particularly suitable for electric vehicles using double wishbone independent suspension. This invention integrates suspension kinematic modeling, singular perturbation multi-timescale dynamic analysis, and adaptive control technology, and can be widely applied to active chassis height adjustment and driving posture control in passenger cars, commercial vehicles, and new energy vehicles. Background Technology

[0002] Electronically controlled air suspension systems actively control vehicle height by adjusting air spring pressure, serving as a key system for improving vehicle passability, ride comfort, and driving stability. Double wishbone independent suspensions are widely used in high-end passenger cars and electric vehicles due to their excellent kinematic characteristics. However, when the wheels move vertically, the wheel center experiences a non-negligible lateral displacement due to the difference in length between the upper and lower control arms and the geometric constraints of the hinge point arrangement. This creates an inherent kinematic coupling effect between lateral and vertical motion. This coupling effect directly alters the suspension's stress state and air spring volume during steering, roll, or single-wheel obstacle crossing, interfering with the accuracy of air pressure regulation and causing deviations in vehicle height control from the desired value.

[0003] Existing electronically controlled air suspension height control methods generally simplify the suspension system into a single-degree-of-freedom vertical model, ignoring or using fixed compensation coefficients to handle lateral coupling effects. These methods have the following main shortcomings in vehicle height adjustment: First, traditional control models lack a quantitative representation of lateral coupling effects. Existing models treat the wheel center motion of a double wishbone suspension as a pure vertical bounce, failing to establish the kinematic relationship between wheel center lateral displacement and suspension vertical displacement, and also neglecting to incorporate suspension geometric parameters into the control model. When the vehicle undergoes steering or road roll excitation, the equivalent spring force change and air pressure volume disturbance caused by wheel center lateral displacement cannot be perceived by the model. A structural deviation exists between the controller's decision-making basis and the actual physical state of the system, leading to a significant increase in height tracking error under conditions of high lateral acceleration.

[0004] Secondly, the existing control architecture struggles to match the inherent multi-timescale characteristics of the suspension system. During air suspension height adjustment, vehicle height changes are low-frequency, slow-changing dynamics, while suspension vibrations and air pressure transient responses are high-frequency, fast-changing dynamics, with time constants differing by tens of times. When dealing with such multi-scale systems, traditional single-rate controllers, if using a fast control cycle, experience excessively frequent slow-loop parameter updates, leading to unnecessary actuator actions; conversely, if a slow control cycle is used, air pressure fluctuations and suspension vibrations cannot be suppressed in time. This timescale mismatch makes controller parameter tuning difficult, with fast and slow dynamics mutually constraining each other, making it difficult to simultaneously guarantee air pressure regulation quality and height adjustment stability.

[0005] Third, the control parameters rely on fixed calibration and lack the ability to adapt to time-varying uncertainties. The lateral coupling coefficient is not a constant value; it changes dynamically with operating conditions such as suspension travel position, tire load, and road adhesion coefficient. Existing methods typically calibrate fixed control parameters offline based on nominal operating conditions, failing to perceive real-time changes in the coupling coefficient online. Once actual operating conditions deviate from the calibration conditions, the controller's model fit decreases, and control quality deteriorates. Especially in typical scenarios such as load variations and road surface undulations, fixed-parameter controllers struggle to maintain the desired height adjustment performance.

[0006] Fourth, the application of intelligent methods such as neural networks in suspension control still suffers from problems such as a single network structure and an inability to adapt to differences in fast and slow dynamics. Some studies have attempted to use neural networks for suspension control parameter optimization, but the single-time-scale networks used process all state variables at the same frequency. Fast state details are easily overwhelmed by slow states, or slow state trends are interfered with by fast state noise. The network learning efficiency is low, and the state estimation accuracy is poor, making it difficult to achieve effective deployment under the condition of limited on-board computing resources in actual vehicles.

[0007] Fifth, controller parameter tuning relies on human experience and lacks an autonomous optimization mechanism based on global performance indicators. In existing methods, the selection of slow-loop altitude tracking parameters and fast-loop pressure regulation parameters often depends on trial and error by engineers or simple empirical formulas. The trade-offs between various control objectives are difficult to quantify, and the performance index function does not systematically incorporate lateral displacement constraints and energy consumption costs. Multiple objectives, such as altitude regulation accuracy, pressure stability, lateral offset suppression, and energy economy, are difficult to coordinate and optimize, and the optimality of the control strategy lacks theoretical guarantees.

[0008] Sixth, existing control strategies lack adaptability and ease of deployment. On the one hand, the training data collection conditions are limited, failing to systematically cover typical scenarios such as static commands, different road surface excitations, and various load states. The insufficient representativeness of the training samples results in limited generalization ability of the control strategies. On the other hand, most intelligent control methods adopt a purely online learning mode, leading to poor control quality in the initial deployment stage due to a lack of prior knowledge. Furthermore, existing solutions do not fully consider hardware compatibility with the solenoid valve drive interface of actual vehicles. In addition, the lack of a smooth transition mechanism when switching between different operating conditions can easily cause abrupt changes in control inputs, thereby triggering high oscillations and affecting system stability and ride comfort.

[0009] In summary, there is an urgent need to establish an electronically controlled air suspension height adaptive control method that can accurately characterize the lateral coupling effect of double wishbone independent suspension, adapt to the multi-time-scale characteristics of the system, have the ability to identify time-varying parameters online, achieve autonomous optimization of control parameters, and is easy to deploy in engineering, so as to overcome the shortcomings of the existing technologies. Summary of the Invention

[0010] This invention addresses the problems of insufficient height adjustment accuracy, large air pressure fluctuations, and poor adaptability in existing electronically controlled air suspension height control methods due to neglecting the lateral kinematic coupling effect of the double wishbone independent suspension. It provides an adaptive height control method for electronically controlled air suspension based on an independent suspension lateral coupling model. This method constructs a high-fidelity coupled mathematical model, designs an adaptive control architecture, and combines multi-timescale neural networks and reinforcement learning mechanisms to achieve precise, rapid, and robust suspension height adjustment under various operating conditions.

[0011] To achieve the above objectives, the technical solution adopted by the present invention mainly includes the following four steps: S1: Construct a mathematical model of motion-dynamic coupling for air suspension height adjustment based on the lateral coupling effect of a double wishbone independent suspension. This step integrates suspension kinematic constraints, solenoid valve flow characteristics, air spring thermodynamic equations, and suspension dynamic equations to establish a system mathematical model that can accurately characterize the coupling relationship between wheel center lateral displacement and suspension vertical runout. Parameter calibration is performed using real vehicle test data, providing a high-fidelity object model for subsequent controller design.

[0012] S2: Establishing dynamic equations for multi-timescale neural networks containing fast and slow states based on singular perturbations. Based on the system dynamic characteristics revealed by the model established in step S1, and taking the vehicle height tracking error, air spring pressure fluctuation and wheel center lateral displacement constraint as the basis for regulation, the system is decomposed into a slow-variable subsystem and a fast-variable subsystem based on the singular perturbation fast decomposition theory. The slow network is responsible for learning the long-term law of vehicle height change; the fast network is responsible for learning the suspension vibration and air pressure transient response, and a multi-timescale dynamic neural network dynamic equation is designed.

[0013] S3: Construct an adaptive reinforcement learning architecture for electronically controlled air suspension, identify time-varying parameters such as lateral coupling coefficient online, and calculate the optimal control input based on the evaluator network.

[0014] To achieve adaptive online optimization of controller parameters, an adaptive control architecture based on an actor network, an evaluator network, and an identifier network is constructed. A dual-time-scale neural network is introduced to adapt to the different dynamic characteristics of fast and slow subsystems. The network weights are updated online through a reinforcement learning mechanism to approximate the optimal control strategy.

[0015] S4: Multi-condition training and real-vehicle deployment application The system collects data under typical operating conditions, completes offline pre-training and online adaptive training, and deploys the trained control strategy to the vehicle electronic control unit to achieve precise, rapid and adaptive adjustment of suspension height under all operating conditions.

[0016] Furthermore, step S1 specifically includes: S1.1: Based on the kinematic constraints of the double wishbone independent suspension, a quantitative relationship between the lateral displacement of the wheel center and the vertical relative displacement of the suspension is established. The proportional coefficient is defined as the lateral coupling coefficient, which is determined by the length of the upper and lower control arms and the spatial coordinates of the hinge point. It is used to quantify the degree of interference of the lateral movement of the wheel center on the vertical height adjustment of the suspension under vehicle steering or roll conditions.

[0017] S1.2: Establish a flow characteristic model for the proportional switching solenoid valve. Based on the relationship between the pressure ratio at both ends of the solenoid valve and the critical pressure ratio, describe the calculation method of gas mass flow rate in segments to accurately characterize the nonlinear flow characteristics of the solenoid valve during the charging and discharging process.

[0018] S1.3: Based on the first law of thermodynamics, an equation for air pressure change of air spring under three working conditions of inflation, deflation and pressure holding is established. The functional relationship between the rate of change of air pressure over time and the instantaneous change of airbag volume, gas mass flow rate and thermodynamic index is quantified, reflecting the bidirectional coupling effect between air pressure dynamics and suspension motion.

[0019] S1.4: Based on Newton's second law, establish the air suspension dynamics equations that retain only the damping and air pressure terms, and take the product of the air spring pressure and the effective area as the only source of elastic restoring force, and construct the vertical motion differential equations for the unsprung mass and the sprung mass respectively.

[0020] S1.5: Integrate the above sub-models to form a complete mathematical model of lateral coupling motion-dynamic coupling of air suspension. Use test data of actual vehicle steering and height adjustment conditions to calibrate and correct the lateral coupling coefficients to ensure that the model can accurately characterize the actual response of lateral coupling effect and air pressure regulation characteristics.

[0021] Furthermore, step S2 specifically includes: Based on the kinematic and dynamic response of the air suspension under the influence of lateral coupling, and taking the vehicle height tracking error, air spring pressure fluctuation and wheel center lateral displacement constraint as the control basis, the system is decomposed into a slow variable subsystem and a fast variable subsystem based on the singular perturbation fast decomposition theory. The slow network is responsible for learning the long-term law of vehicle height change, and the fast network is responsible for learning the suspension vibration and air pressure transient response. Multi-timescale dynamic neural network dynamic equations are designed.

[0022] Furthermore, step S3 specifically includes: S3.1: Reorganize the coupled mathematical model from step S1 into a standard matrix form suitable for singular perturbation analysis, clearly distinguishing between two types of state variables. Slow state variables include the vehicle's vertical displacement and vertical velocity, representing the low-frequency, slowly varying vehicle height dynamics; fast state variables include suspension travel, relative suspension velocity, and air spring pressure, representing the high-frequency transient suspension vibration and air pressure response. The time scale factor represents the ratio of the time constants between the fast and slow dynamics.

[0023] S3.2: To address the significant differences in the time scales between the fast and slow dynamics mentioned above, a multi-time-scale dynamic neural network is designed. This network comprises two parallel sub-networks. The slow network uses a larger time constant and a lower update frequency, specifically responsible for learning the long-term evolution of vehicle height changes; the fast network uses a smaller time constant and a higher update frequency, specifically responsible for learning the rapid response characteristics of suspension vibration and transient air pressure changes. Both sub-networks contain linear components, nonlinear activation function components, and road input components, and are interconnected through weight matrices to jointly form a complete system state estimator.

[0024] S3.3: Set the comprehensive performance objective function, which consists of four weighted factors: the weighted average of the square of the vehicle height tracking error, the weighted average of the square of the air spring pressure fluctuation amplitude, the weighted average of the square of the wheel center lateral displacement deviation, and the weighted average of the square of the solenoid valve control energy consumption. By reasonably allocating the weight coefficients, this performance index takes into account multiple control objectives such as height adjustment accuracy, air pressure stability, lateral offset suppression, and energy economy during the optimization process.

[0025] S3.4: Design an identifier network to identify uncertain dynamic parameters in the system in real-time online, including time-varying parameters that are difficult to model precisely, such as lateral coupling coefficients and the effective area of ​​air springs. The identifier drives the online updating of network weights by minimizing the deviation between its own state estimate and the actual system output. Its learning rate controls the step size of the weight adjustment, enabling the identifier to continuously track parameter changes and provide the controller with accurate object dynamic information. Simultaneously, design an evaluator network. This network takes fast and slow state vectors as input and outputs an approximate estimate of the optimal performance index function, i.e., the state value function, through a combination of a nonlinear excitation function and a weight matrix, to evaluate the merits of the current control strategy.

[0026] S3.5: Design the actuator network, which uses fast and slow state vectors to form the input space. It directly outputs the optimal control quantity, which is the target gas mass flow rate of the solenoid valve, through a weight matrix and activation function. The objective of the actuator network's weight update is to minimize the performance index function output by the evaluator network, i.e., to reduce the long-term cumulative cost by continuously adjusting the actuator network weights. The weight update law is determined by the actuator network's learning rate, the partial derivative of the performance index with respect to the control input, and the partial derivative of the control input with respect to the actuator network weights, thereby achieving continuous evolution of the control strategy towards the optimal direction.

[0027] Furthermore, step S4 specifically includes: S4.1: Design a system for collecting training data under four typical operating conditions. Condition 1 is a static manual height command condition, where the chassis is adjusted to three target heights (lowest, standard, and highest) by operating the height control switch while the vehicle is stationary, with the travel set to ±60 mm. Condition 2 is a constant-speed straight-line driving condition on a level road, where the vehicle travels at different speeds and random uneven road surface excitations are applied, collecting state data during the dynamic height adjustment process. Condition 3 is a constant-speed straight-line driving condition on a potholed road surface, where the vehicle travels through potholed areas at different speeds, collecting system response data under pulse road impacts. Condition 4 is a load variation condition, covering three load states: unloaded, half-loaded, and fully loaded. The collected state variables include vehicle height, air spring pressure, suspension travel, and vertical acceleration. Simultaneously, actual control outputs and performance index values ​​are recorded. No fewer than thirty sets of independent test data are collected for each condition.

[0028] S4.2: First, offline pre-training is performed using the collected experimental data. The training data is input into the fast and slow networks, and all network parameters are iteratively optimized according to the weight update laws of the evaluator network, the identifier network, and the executor network. After offline pre-training is completed, the trained network weights are deployed to the on-board electronic control unit. During actual vehicle operation, the system continuously performs online adaptive training. The identifier network estimates the lateral coupling coefficient in real time and feeds the identification results back to the controller parameter adjustment module, realizing dynamic adaptive compensation for system uncertainties.

[0029] S4.3: Execute a complete closed-loop calculation process within each fast-loop control cycle. First, read the signals from each sensor and update the fast and slow state vectors; the slow-loop controller outputs height tracking parameters based on slow network dynamics and calculates the target air pressure by combining the vehicle height tracking error; the fast-loop controller outputs air pressure adjustment parameters based on fast network dynamics and calculates the optimal mass flow command by combining the actuator network output; this mass flow command is converted into a pulse width modulation duty cycle signal to drive the solenoid valve to achieve precise adjustment of the air spring pressure; simultaneously, the evaluator network output is updated according to the evaluator network update law, the identifier network state estimate is updated according to the identifier network update law, and the weights of each network are updated synchronously according to the weight update law of each network, forming a continuous optimization closed loop of control-while-learning.

[0030] S4.4: The system automatically identifies the current operating condition based on real-time collected acceleration, vehicle speed, load, and air pressure signals, and retrieves the corresponding optimal controller parameters from the offline-trained parameter library. When the operating condition changes, a smooth transition strategy is adopted to gradually transition the controller parameters from the current value to the target value through linear interpolation, effectively avoiding height oscillations that may be caused by sudden parameter changes. Through the above process, precise, rapid, and adaptive adjustment of the air suspension height under different dynamic operating conditions is achieved, effectively suppressing height adjustment deviations caused by lateral coupling effects and significantly improving the system's control quality and robustness.

[0031] Compared with the prior art, this application has the following significant advantages: 1. Systematically fills the gap in lateral coupling modeling, fundamentally improving high-precision control. This application is the first to systematically introduce lateral kinematic coupling constraints of a double wishbone independent suspension into the height control of an electronically controlled air suspension, constructing a complete technical closed loop of modeling, decomposition, intelligent optimization, and deployment. By directly linking the suspension geometric parameters to the control model through lateral coupling coefficients, the controller can anticipate differences in the degree of lateral interference under different suspension structural parameters. This feedforward incorporation of the coupling effect into the control model fundamentally solves the height control deviation problem caused by the traditional method of simplifying the suspension as a purely vertical system and ignoring lateral interference. The height tracking accuracy under vehicle steering or roll conditions is significantly improved. The air pressure change equation simultaneously considers airbag volume changes and gas mass exchange factors, fully characterizing the bidirectional coupling relationship between air pressure dynamics and suspension motion, laying a physical foundation for precise air pressure regulation. Calibration and correction of the coupling coefficients using real vehicle test data gives the mathematical model both physical interpretability and data-driven accuracy compensation capabilities, significantly enhancing the model's engineering confidence.

[0032] 2. Dual-loop hierarchical decoupling control significantly optimizes both fast and slow dynamic performance and air pressure regulation quality. This application designs a dual-loop control structure based on the singular perturbation fast-slow separation theory. It treats vehicle height dynamics as a slow-variable subsystem and suspension vibration and air pressure dynamics as fast-variable subsystems, allowing the slow-loop and fast-loop controllers to be designed and tuned independently, significantly reducing the control design complexity of multivariable coupled systems. The slow-loop controller uses height tracking error as its main input, has a larger time constant, and focuses on steady-state accuracy and anti-interference capability. The fast-loop controller uses air pressure deviation and suspension travel as its main inputs, has a fast response speed, and is mainly responsible for suppressing transient fluctuations. The two loops are cascaded through the target air pressure value, with clear division of labor and close coordination. This structure effectively avoids the problem of mutual constraints between fast and slow loops in controller parameter tuning. The high-frequency response capability of the fast loop effectively suppresses transient height fluctuations caused by lateral coupling, while the low-frequency adjustment of the slow loop ensures the steady-state accuracy of height regulation. The quality of air pressure regulation and the stability of the overall vehicle height are significantly improved.

[0033] 3. Dual-time-scale network design: synergistic improvement in fast and slow dynamic learning efficiency and state estimation accuracy. This application designs a dual-timescale dynamic neural network to address the significant differences between fast and slow dynamic timescales. The slow network employs a large time constant and a low update frequency, focusing on learning the long-term evolution of vehicle height changes and possessing the ability to recognize the trend characteristics of height changes under different operating conditions. The fast network employs a small time constant and a high update frequency, focusing on learning the transient response of suspension vibration and air pressure, and possessing the agile characteristic of quickly capturing road impacts and sudden changes in air pressure. The combination of the two networks balances the system's ability to optimize long-term performance with its response speed to transient disturbances. Both sub-networks include linear components, nonlinear activation function components, and road input components, combining the stability of linear systems with the approximation ability of nonlinear systems, significantly improving state estimation accuracy and learning efficiency.

[0034] 4. The identifier continuously learns online, substantially enhancing its time-varying parameter adaptation capability and system robustness. This application introduces an identifier network to continuously identify time-varying parameters such as the lateral coupling coefficient and the effective area of ​​the air spring online. The lateral coupling coefficient dynamically changes with factors such as suspension travel position, tire load, and road adhesion coefficient during actual driving, making it difficult for traditional fixed-parameter control methods to handle this type of model mismatch. The identifier drives online weight updates by minimizing state estimation bias, tracking the trend of time-varying parameter changes in real time and updating the internal model parameters of the controller accordingly, ensuring that the control strategy always makes decisions based on accurate system dynamic information. The evaluator network approximates the optimal performance index function, providing a quantitative evaluation standard for improving the executor network strategy. The executor network uses fast and slow state vectors as input space and directly maps the state space to the optimal mass flow command through end-to-end learning, avoiding the accumulation of errors in traditional two-step calculations. The executor network weight update aims to minimize the evaluator network output. The update law is jointly determined by the learning rate, the sensitivity of the performance index to control, and the sensitivity of control to weights, combining the stability of gradient descent with the exploratory nature of reinforcement learning. The identification, evaluation, and execution architecture forms a continuous optimization loop, substantially enhancing the system's adaptability to time-varying parameters and its overall robustness.

[0035] 5. Multi-condition system training and two-stage deployment comprehensively guarantee strategy generalization capability and engineering practicality. This application designs four typical operating conditions for system data acquisition: static height command, random uneven road surface, pulse pothole road surface, and load variations under no-load, half-load, and full-load conditions. At least thirty independent data sets are collected for each condition to ensure the sufficiency and statistical representativeness of the training samples. A two-stage strategy of offline pre-training and online adaptive training is adopted. Offline pre-training uses batch data to complete the initial parameter tuning of the network, enabling the controller to have basic control capabilities from the outset and avoiding the risk of poor initial control quality. Online adaptive training continuously refines the network parameters during actual vehicle operation, providing personalized adaptation for specific vehicles and operating environments. The method of directly driving the solenoid valve with pulse width modulation duty cycle signals is compatible with existing electronically controlled air suspension hardware interfaces, allowing for deployment without additional modifications and demonstrating strong engineering practicality.

[0036] 6. Real-time closed-loop control and smooth switching of operating conditions achieve both real-time control performance and stability across all operating conditions. This application synchronously completes the entire process of sensor reading, status update, dual-loop calculation, pulse width modulation output, and network weight update within each fast-loop control cycle, fully meeting the real-time requirements of actual vehicle control and realizing an intelligent control paradigm of learning while controlling. The system automatically identifies the current operating condition based on acceleration, vehicle speed, load, and air pressure signals, and calls the corresponding optimal controller parameters from the parameter library. A linear interpolation smooth transition strategy is used during operating condition switching to avoid height oscillations caused by parameter abrupt changes. Through the above-mentioned full-process design, this application ultimately achieves high-precision, fast-response, and robust adaptive adjustment of the electronically controlled air suspension height under the influence of lateral coupling effects, achieving a good balance between multi-condition optimization and smooth switching. Attached Figure Description

[0037] Figure 1 The flowchart of air suspension height adaptive control based on the independent suspension lateral coupling model used in the embodiment; Figure 2 This is a block diagram of the air suspension lateral coupling motion-dynamics model used in the embodiment; Figure 3 The block diagram of the fast and slow subsystems and performance target structure based on singular perturbation theory used in the embodiments; Figure 4 This is a block diagram of the multi-timescale neural network adaptive control architecture used in the embodiment; Figure 5 The training process for adaptive control parameters under four typical operating conditions used in the implementation example is as follows. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described examples are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0039] An adaptive height control method for electronically controlled air suspension based on an independent suspension lateral coupling model is illustrated in the flowchart below. Figure 1As shown, firstly, a kinematic model of electric vehicle air suspension height adjustment motion-dynamic coupling considering the lateral coupling kinematic constraints of the double wishbone independent suspension is constructed to quantify the coupling relationship between wheel center lateral displacement and suspension vertical bounce and the dynamic response under different operating conditions. Then, based on the vehicle height tracking error, air pressure fluctuation, and wheel center lateral displacement constraints as the adjustment basis, a dual-loop adaptive controller is designed based on the singular perturbation fast-slow separation theory. Precise control of air spring pressure is achieved by adjusting the mass flow rate of the solenoid valve. Finally, using a multi-timescale neural network and an actor-evaluator-identifier architecture, the wheel center lateral displacement is incorporated into the state space, and the system's uncertain dynamics, such as the lateral coupling coefficient, are identified online. The control parameters and optimal control law of the adaptive controller are then intelligently optimized. This height control method specifically includes: S1: Construct a mathematical model of motion-dynamic coupling for air suspension height adjustment based on the lateral coupling effect of a double wishbone independent suspension. like Figure 2 As shown, the air spring pressure is used as the sole elastic restoring force. The coupling relationship between the lateral and vertical displacements of the wheel center is quantified through suspension kinematic constraints. A comprehensive model is constructed for the air spring, double wishbone independent suspension, solenoid valve, air tank, and the sprung / unsprung mass of the entire vehicle. The influence of lateral coupling on the air spring volume and air pressure dynamics, as well as the coupling response of air pressure regulation and suspension motion under different operating conditions, are analyzed. The model is compared with simulation and real-vehicle test data to ensure accurate characterization of the lateral coupling effect and air pressure regulation characteristics. The specific process is as follows: S1.1: Establish the coupling relationship between wheel center lateral displacement and suspension vertical runout. Based on the kinematic constraints of the double wishbone independent suspension, the coupling relationship between the wheel center lateral displacement and the suspension vertical runout is established, and the influence of lateral motion on vertical height adjustment is quantified: (1) In the formula, This is the lateral displacement of the wheel center. The lateral coupling coefficient is determined by the lengths of the upper and lower control arms of the suspension and the coordinates of the hinge points. This refers to the vertical displacement of the vehicle body. This represents the vertical displacement of the wheel.

[0040] S1.2: Establish the flow characteristic model of the proportional switching solenoid valve The flow characteristics of a proportional switching solenoid valve can be approximated as follows: (2) In the formula: For gas mass flow rate, This is the critical pressure ratio. This is the equivalent cross-sectional area of ​​the solenoid valve. Thermodynamic temperature pu and These represent the air pressure at both ends of the solenoid valve.

[0041] S1.3: Establish the air pressure change equation for the air spring charging, discharging, and pressure holding conditions. Based on the first law of thermodynamics, the equation for the pressure change during the inflation and deflation of an air spring is as follows: (3) In the formula, The gas thermodynamic index is ; The airbag pressure at a certain moment. The rate of change of airbag pressure over time.

[0042] S1.4: Establish the dynamic equations of the air suspension Based on Newton's second law, we construct the dynamic equations for an air suspension system where air pressure provides the elastic force. We omit the traditional suspension spring stiffness and lateral coupling force terms, retaining only the damping and air pressure terms: (4) In the formula, For unsprung mass, For the sprung mass, For suspension damping, This refers to the vertical stiffness of the tire.

[0043] S1.5: Model Integration and Parameter Calibration By combining equations (1)-(4), a mathematical model of lateral coupling motion-dynamics of the air suspension is obtained, and the lateral coupling coefficient is corrected through actual vehicle steering and height adjustment tests. This improves model accuracy.

[0044] S2: Establishing dynamic equations for multi-timescale neural networks containing fast and slow states based on singular perturbations. Based on the air suspension kinematics response under lateral coupling, and using vehicle height tracking error, air spring pressure fluctuation, and wheel center lateral displacement constraints as the control basis, the system is decomposed into a slow-variable subsystem and a fast-variable subsystem based on the singular perturbation fast decomposition theory. The slow network is responsible for learning the long-term law of vehicle height change; the fast network is responsible for learning the suspension vibration and air pressure transient response. Multi-timescale dynamic neural network dynamic equations are designed. Figure 3 As shown, the slow-loop height tracking parameters and fast-loop air pressure adjustment parameters are optimized, and the flow rate of the solenoid valve is adjusted to use the active control force output by the air spring to precisely adjust the vehicle height. At the same time, the height adjustment deviation caused by lateral coupling is suppressed by air pressure control.

[0045] The vehicle height tracking error is expressed as: (5) In the formula, This represents the desired vertical height of the vehicle body. This is the actual value of the vehicle's vertical height.

[0046] S3: Construct an adaptive reinforcement learning architecture for electronically controlled air suspension, identify time-varying parameters such as lateral coupling coefficient online, and calculate the optimal control input based on the evaluator network.

[0047] Based on the air suspension height adjustment requirements, a dual-loop adaptive controller parameter optimization model is trained using a multi-timescale neural network combined with an actor-evaluator-identifier architecture, such as... Figure 4 As shown, the design includes fast and slow state spaces, action spaces, and performance index functions. The slow state space consists of the vehicle's vertical displacement and velocity; the fast state space consists of suspension travel, relative velocity, air spring pressure, and wheel center lateral displacement; the action space consists of the dual-loop adaptive controller parameters; and the performance index functions include vehicle height tracking error, air pressure fluctuation value, wheel center lateral displacement deviation, and solenoid valve control loss. The fast and slow state information, desired vehicle height and air pressure values, and cumulative performance index values ​​obtained during training are stored in an experience pool. Randomly collected sample data is used for neural network training, thereby adaptively optimizing the slow-loop height tracking parameters and the fast-loop air pressure adjustment parameters. The specific process is as follows: S3.1: Matrix Form of Singular Perturbation Fast and Slow Subsystems The laterally coupled motion-dynamic coupling mathematical model is rewritten in the matrix form of singularly perturbated fast and slow subsystems, yielding the fast-slow dynamic coupling relationship under pure air pressure control as follows: (6) In the formula: >0 indicates a time scale factor. To control the input, Characterizes the vehicle body displacement and speed at a slow state. Characterizes fast-moving states such as suspension dynamic travel, relative speed, and air pressure; This is the system's slow state vector, representing low-frequency, slowly varying dynamics such as vehicle height and sprung displacement. The table shows the fast state vector of the system, which characterizes high-frequency transient dynamics such as suspension dynamic travel, airbag pressure, and unsprung vibration. For road surface excitation.

[0048] S3.2: Fast and Slow State Matrix and Dual-Time-Scale Neural Network Design The fast and slow state matrix is ​​defined as follows: (7) To address the difference between fast and slow dynamics, a dual-time-scale dynamic neural network is designed, comprising two sub-networks: a slow network (long-time memory layer) with a large time constant and low update frequency, responsible for learning the long-term patterns of vehicle height changes; and a fast network (short-time memory layer) with a small time constant and high update frequency, responsible for learning the transient responses of suspension vibration and air pressure. The corresponding network dynamic equations satisfy: (8) In the formula: and This is the linear part matrix of the neural network; and These are the weights of the output layer; and These are network activation functions; and The weights are the output layers of the road surface.

[0049] S3.3: Setting the performance target function The performance objective function is set as a weighted sum of vehicle height tracking error, air pressure fluctuation, and solenoid valve control loss, taking into account both height adjustment accuracy and air pressure stability. (9) In the formula: This is the air pressure fluctuation vector of the air spring. , , , These are the weight matrices; This is the vehicle height tracking error vector.

[0050] S3.4: Multi-timescale Neural Network Adaptive Architecture Training and Evaluator Network Design A multi-timescale adaptive neural network architecture is trained to estimate the real-time dynamics of the system online, and the evaluator network is approximated to the optimal performance metric function. The evaluator network output is: (10) In the formula: To evaluate the weight matrix of the network, It is a nonlinear excitation function. To identify errors.

[0051] The identifier network is used for online identification of uncertain dynamics in the system, including parameters such as the lateral coupling coefficient and the effective area of ​​the air spring. Its identification error drives the online updating of the weights. The identifier weights are updated online by minimizing the L2 norm of the identification error, and the learner rate controls the step size of the weight adjustment. Through continuous online learning by the identifier, the system can track changes in time-varying parameters such as the lateral coupling coefficient in real time, providing the controller with accurate dynamic information about the object.

[0052] S3.5: Actor Networks and Optimal Control Laws Based on the feedback of altitude tracking error and air pressure fluctuation, the weights are updated. The actuator network approximates the optimal control strategy, and the evaluator network calculates the optimal control input. (11) In the formula: for The estimated value, The state equation is the control gain matrix, and the optimal control input is... That is, the gas mass flow rate of the solenoid valve .

[0053] The executor network directly outputs the optimal control input. Its network structure consists of the executor network weight matrix and activation function, with the fast and slow state vectors forming the input space. The goal of updating the executor network weights is to minimize the performance index function output by the evaluator network, i.e., to reduce the long-term cumulative cost by adjusting the executor network weights. The executor network weight update law is jointly determined by the executor network learning rate, the partial derivative of the performance index with respect to the control input, and the partial derivative of the control input with respect to the executor network weights.

[0054] S4: Multi-condition training and real-vehicle deployment application like Figure 5 As shown, adaptive control parameters are trained under different operating conditions. The optimal solenoid valve mass flow control signal is output to the air suspension electronic control system, and the suspension height is quickly and accurately adjusted by precisely controlling the air spring pressure. The specific implementation method is as follows: S4.1: Typical Operating Condition Design and Data Acquisition Four typical operating conditions were selected for training data collection. Condition 1 involved static manual height control. With the vehicle stationary, the vehicle's chassis height was adjusted to three states—lowest, standard, and highest—by operating the vehicle's electronically controlled air suspension height control switch. The chassis height (measured near the center of the vehicle's side) was measured at each height, with a travel range of ±60mm. Condition 2 involved constant-speed straight-line driving at different speeds on a level road surface, collecting data on height increases and decreases under random uneven road surface excitation. Condition 3 involved constant-speed straight-line driving at different speeds on a bumpy road surface, collecting data on pulsed road surface conditions. Condition 4 involved load variations, covering unloaded, half-loaded, and fully loaded states. State variables included vehicle height, air pressure, suspension travel, and vertical acceleration, with actual control outputs and performance index values ​​recorded simultaneously. At least thirty sets of independent test data were collected for each condition.

[0055] S4.2: Offline Pre-training and Online Adaptive Training First, offline pre-training is performed using the collected experimental data. The training data is input into the fast and slow networks, and the network parameters are updated according to the update laws of the evaluator network weights, the identifier network weights, and the executor network weights. After offline pre-training, the network weights are deployed to the onboard electronic control unit. During actual vehicle operation, continuous online adaptive training is conducted. Simultaneously, the identifier network continuously estimates the lateral coupling coefficient online and feeds the identification results back to the controller parameter adjustment module, achieving real-time adaptive compensation for system uncertainties.

[0056] S4.3: Control Signal Generation and Execution Within each fast-loop control cycle, the following calculation process is executed: Sensor signals are read, and the fast and slow state vectors are updated. The slow-loop controller outputs height tracking parameters based on slow network dynamics and calculates the target air pressure by combining this with the vehicle height tracking error. The fast-loop controller outputs air pressure regulation parameters based on fast network dynamics and calculates the optimal mass flow command by combining this with the actuator network output. The mass flow command is converted into a pulse-width modulation duty cycle signal to drive the solenoid valve and adjust the air spring pressure. The evaluator network output is updated according to the evaluator network update law, the identifier network state estimate is updated according to the identifier network update law, and the weights of each network are updated according to the respective network weight update laws.

[0057] S4.4: Operating Condition Switching and Parameter Calling Based on acceleration, vehicle speed, load, and air pressure signals under different operating conditions, the system automatically identifies the current operating condition and retrieves the corresponding optimal controller parameters from the trained parameter library. When the operating condition changes, a smooth transition strategy is used to linearly interpolate the controller parameters from the current value to the target value, avoiding height oscillations caused by sudden parameter changes. Through this process, precise, rapid, and adaptive adjustment of the air suspension height under different dynamic operating conditions is achieved, effectively suppressing height adjustment deviations caused by lateral coupling effects and significantly improving the system's control quality and robustness.

[0058] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. An electronically controlled air suspension height self-adaptive control method based on an independent suspension lateral coupling model, characterized in that, Includes the following steps: Step S1: Construct a mathematical model of motion-dynamic coupling of air suspension height adjustment based on the lateral coupling effect of double wishbone independent suspension; Step S2: Establish the dynamic equations of a multi-timescale neural network containing fast and slow states based on singular perturbations; Step S3: Construct an adaptive reinforcement learning architecture for electronically controlled air suspension, identify time-varying parameters such as lateral coupling coefficient online, and calculate the optimal control input based on the evaluator network; Step S4: Conduct multi-condition training and real-vehicle deployment application.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: S1.1, Based on the kinematic constraints of the double wishbone independent suspension, a quantitative relationship between the lateral displacement of the wheel center and the vertical relative displacement of the suspension is established, where the proportional coefficient is defined as the lateral coupling coefficient, which is determined by the length of the upper and lower control arms and the spatial coordinates of the hinge point. S1.2, Establish a flow characteristic model for a proportional switching solenoid valve, and describe the calculation method of gas mass flow rate in segments based on the relationship between the pressure ratio at both ends of the solenoid valve and the critical pressure ratio. S1.3, Based on the first law of thermodynamics, establish the air pressure change equation of the air spring under the conditions of inflation, deflation and pressure holding; S1.4, based on Newton's second law, establishes an air suspension dynamics equation that retains only the damping term and the air pressure term, and takes the product of the air spring pressure and the effective area as the only source of elastic restoring force; S1.5, integrate the above sub-models to form a complete mathematical model of lateral coupling motion-dynamics of the air suspension, and use real vehicle test data to calibrate the lateral coupling coefficient.

3. The method according to claim 1, characterized in that, In step S2, the system is decomposed into a slow-variable subsystem and a fast-variable subsystem based on the control basis of vehicle height tracking error, air spring pressure fluctuation and wheel center lateral displacement constraint, according to the singular perturbation fast decomposition theory. The slow network is responsible for learning the long-term law of vehicle height change; the fast network is responsible for learning the suspension vibration and air pressure transient response, and a multi-timescale dynamic neural network dynamic equation is designed.

4. The method according to claim 1, characterized in that, Step S3 specifically includes: S3.1, The coupled mathematical model in step S1 is organized into a singular perturbation fast and slow subsystem matrix form. The slow state variables include the vertical displacement and vertical velocity of the vehicle body, and the fast state variables include the suspension travel, the relative motion speed of the suspension, and the air spring pressure. S3.2, design a multi-timescale dynamic neural network. The slow network uses a larger time constant and a lower update frequency to learn the long-term evolution law of vehicle height change; the fast network uses a smaller time constant and a higher update frequency to learn the fast response characteristics of suspension vibration and transient air pressure changes. S3.3 sets the comprehensive performance objective function, which consists of four weighted components: vehicle height tracking error, air spring pressure fluctuation, wheel center lateral displacement deviation, and solenoid valve control loss. S3.4, Design an identifier network to identify the time-varying parameters of the lateral coupling coefficient and the effective area of ​​the air spring online; Design an evaluator network to approximate the optimal performance index function; S3.5 Design the executor network, which uses fast and slow state vectors to form the input space and directly outputs the target gas mass flow rate of the solenoid valve. Its weight update aims to minimize the output of the evaluator network.

5. The method according to claim 4, characterized in that, The slow and fast networks of the dual-timescale dynamic neural network each contain a linear part, a nonlinear activation function part, and a road input part, and are interconnected through a weight matrix to jointly constitute a system state estimator.

6. The method according to claim 4, characterized in that, The identifier network drives the online updating of network weights by minimizing the deviation between its own state estimate and the actual system output, and the learner learning rate controls the step size of the weight adjustment.

7. The method according to claim 4, characterized in that, The weight update law of the executor network is jointly determined by the executor network learning rate, the partial derivative of the performance index with respect to the control input, and the partial derivative of the control input with respect to the weights of the executor network.

8. The method according to claim 1, characterized in that, Step S4 specifically includes: S4.1 Design four typical operating conditions for training data collection, including static manual height command condition, uniform speed straight driving on a level road condition, uniform speed straight driving on a potholed road condition, and load change condition. Collect no less than thirty sets of independent test data for each condition. S4.2, offline pre-training is performed using the collected test data, and the trained network weights are deployed to the vehicle electronic control unit for continuous online adaptive training during actual vehicle operation; S4.3 In each fast loop control cycle, the sensor signal is read to update the fast and slow state vectors, the slow loop controller calculates the target air pressure, the fast loop controller calculates the optimal mass flow command and converts it into a pulse width modulation duty cycle signal to drive the solenoid valve to act, and the weights of each network are updated synchronously. S4.4 The system automatically identifies the current operating condition based on acceleration signal, vehicle speed signal, load signal and air pressure signal and calls the corresponding optimal controller parameters. Linear interpolation is used for smooth transition when switching operating conditions.

9. The method according to claim 8, characterized in that, In step S4.1, the static manual height command travel is set to ±60 mm; the load change conditions cover no-load, half-load and full-load states; the collected state variables include vehicle height, air spring pressure, suspension travel and vertical acceleration.

10. The method according to any one of claims 1-9, characterized in that, The method achieves rapid and precise suspension height adjustment by accurately controlling the air spring pressure, effectively suppressing height adjustment deviation caused by lateral coupling effect.