A Distributed Wire-Controlled Chassis Personalized Stability Control Method and System

Abstract

This invention relates to the field of vehicle active safety control technology, specifically a distributed drive-by-wire chassis personalized stability control method and system. The personalized stability collaborative control method addresses the shortcomings of existing technologies, such as ignoring individual driver differences, fixed control thresholds, and insufficient multi-actuator coordination. It utilizes factor analysis for dimensionality reduction, K-means++ clustering, and LSTM neural networks to identify driving styles in real time, matching personalized stability energy thresholds with initial differential braking weights. A two-stage dynamic roll axis method is used to calculate normalized stability energy, and lateral acceleration is combined to divide the vehicle state into three control zones. A composite controller combining MPC feedforward and LQR disturbance feedback correction is constructed to execute differentiated multi-actuator collaborative control strategies in each zone. This solution enables personalized adaptation of vehicle active safety interventions, effectively improving stability control accuracy and robustness under operating conditions, balancing vehicle driving safety and comfort, and is compatible with distributed drive chassis architectures, possessing significant engineering application value.

Description

Technical Field

[0001] This invention relates to the field of vehicle active safety control technology, specifically a distributed drive-by-wire chassis personalized stability control method and a distributed drive-by-wire chassis personalized stability control system. Background Technology

[0002] Some vehicles have a high center of gravity, large load capacity, and narrow wheelbase. These vehicles are prone to skidding / rollover accidents under conditions such as high-speed turning and emergency obstacle avoidance. The fatality rate of rollover accidents is much higher than that of ordinary vehicle collision accidents. Therefore, vehicle stability control has always been a key research direction in the field of active safety.

[0003] With the continuous development of drive-by-wire chassis and distributed drive technology (such as existing technologies CN110606079B and CN114357631B), chassis architectures integrating drive, steering, braking, and suspension systems at the wheel ends (such as existing technology CN118387190B) are gradually being implemented. This chassis architecture can achieve independent steering, independent drive / braking, and independent suspension operation for all four wheels, possessing the inherent advantage of multi-actuator coordinated control, and providing new hardware support for upgrading vehicle stability performance.

[0004] Currently, the mainstream stability control solution in the industry is a combined control of active suspension and differential braking, such as the existing technology CN113147735B. In this system, active suspension suppresses vehicle roll by adjusting the vertical force of the wheels, while differential braking generates a yaw moment to counteract instability by utilizing the difference in braking force between the left and right wheels. The two work together to simultaneously ensure vehicle ride comfort and safety. In this collaborative control system, the industry generally uses a weighted summation method to allocate the control proportion of the two types of actuators and integrate multiple control commands. Current weight designs are mainly divided into two forms: fixed weight (such as the existing technology CN106970524B) and linear dynamic weight (such as the existing technology CN120245947B).

[0005] Fixed-weight schemes pre-define the weight coefficients for suspension and differential braking, keeping parameters constant across all vehicle operating conditions. Due to their simple structure and low computational cost, they are widely used in mass-produced vehicles and basic research. Linear dynamic weights, on the other hand, adjust braking weights in real time using a linear function based on stability risk indicators such as lateral acceleration and rollover energy, offering a degree of risk self-adaptation compared to fixed-weight schemes.

[0006] Meanwhile, existing technologies have publicly disclosed mature technologies such as energy-based stability risk assessment models, driving style recognition algorithms, and personalized stability energy thresholds set in conjunction with driving style. Based on these technologies, vehicle stability risks can be assessed in real time, driver operating characteristics can be differentiated, and the vehicle's operating state can be divided into three zones—safe zone, preventative zone, and danger zone—based on lateral acceleration. The preventative zone generally employs a collaborative control logic combining active suspension and differential braking, thus establishing a comprehensive framework for zoned stability control.

[0007] However, the existing weight design method has revealed many inherent defects in the prevention zone operation of distributed drive vehicles, and cannot simultaneously meet the multiple requirements of driving safety, driving smoothness and personalized driver adaptation: When using fixed weights, the weight parameters do not dynamically change with the vehicle's real-time stability risk. When the vehicle is in the low-risk zone of the prevention zone, the differential braking weight is too large, which can easily lead to excessive braking intervention, interfering with the driver's normal operation and disrupting ride smoothness. When the vehicle approaches the end of the prevention zone and the stability risk increases sharply, the braking weight cannot increase synchronously, resulting in insufficient braking force and a significant reduction in stability protection capabilities, posing a safety hazard. Furthermore, fixed weights cannot be differentiated according to the operating characteristics of cautious, moderate, and aggressive drivers, resulting in poor human-machine adaptation capabilities.

[0008] While linear dynamic weights can adjust their coefficients according to stability risk indicators, the linear nature of these weights can cause step fluctuations in braking torque. The distributed drive motors and braking mechanisms of the wheel units are highly sensitive to sudden changes in control inputs. Torque steps not only cause chassis vibration and transmission system shocks, reducing ride comfort, but also accelerate actuator wear and shorten hardware lifespan. Furthermore, the range of linear weight changes is difficult to correlate with the driver's personalized stability energy threshold, making it difficult to match the safety intervention boundaries for different types of drivers.

[0009] In addition, all existing weighting schemes share a common problem: they do not set initial weights to differentiate between different driving styles. Aggressive drivers make quick steering maneuvers and are extremely sensitive to braking intervention; moderate and cautious drivers operate smoothly and have a relatively high tolerance for control intervention. A uniform initial weight cannot adapt to the driving habits of these three types of drivers.

[0010] In summary, traditional fixed-weight and linear dynamic-weight schemes are difficult to adapt to the complex working conditions and personalized driving scenarios of the prevention zone of distributed drive vehicles, and have become a technical bottleneck restricting the improvement of the overall stability performance of vehicles. Summary of the Invention

[0011] To address the technical problems of weak adaptive capability, poor control smoothness, and difficulty in matching different driver operating styles in traditional fixed-weight and linear dynamic weight allocation schemes when distributed drive vehicles are in the prevention zone, this invention provides a distributed drive-by-wire chassis personalized stability control method and a distributed drive-by-wire chassis personalized stability control system.

[0012] To achieve the above objectives, the present invention provides the following technical solution: A distributed drive-by-wire chassis personalized stability control method is applied to a vehicle stability control system that divides the vehicle's operating state into a safe zone, a prevention zone, and a danger zone. This vehicle stability control system pre-identifies driving styles and obtains the corresponding stability energy threshold E for each driving style. type Furthermore, the normalized vehicle stability energy E is calculated in real time. When the vehicle is in the prevention zone, the following control steps are executed: The initial differential braking weight k is set according to the driving style. type The real-time differential braking weight μ is calculated according to the following exponential formula: ; The active suspension power F of the i-th wheel unit i With the torque T of the wheel-side motor i ,according to Calculate the overall control output J of the wheel unit. i Where k1 is the active suspension weight; According to each J i The active suspension and distributed drive motor of the corresponding wheel unit are synchronously controlled, and no additional wheel angle is output in the prevention zone. Vehicle stability control is achieved solely through the coordinated action of active suspension and differential braking.

[0013] As a further improvement to the above scheme: the three-zone division of vehicle operating status is based on the vehicle's lateral acceleration a. y Determination of stability based on two indicators: a y <0.4g and E<0.3, is the safe zone; 0.4g≤a y <0.85g and 0.3≤E <E type E is the prevention zone. type ≤E represents the danger zone; where g is the acceleration due to gravity.

[0014] As a further improvement to the above scheme: a two-stage dynamic roll axis is used to solve for the normalized stability energy E. The first stage is the initial roll stage. The line connecting the instantaneous rotation centers of the front and rear axles is taken as the first roll axis. Energy is calculated only for the sprung mass. The total sprung energy E1 is: ; Where, m s v is the sprung mass. y I represents the lateral speed of the vehicle. x1 Let be the moment of inertia of the sprung mass about the first tilt axis; The roll angle is... ω is the roll angular velocity; g is the acceleration due to gravity; h1 is the distance from the center of mass of the sprung mass to the first roll axis; cos is the cosine function; The second stage is the extreme roll stage. Taking the line connecting the outer wheel to the ground as the second roll axis, the total energy of the entire vehicle's sprung and unsprung masses is calculated uniformly. The total energy of the entire vehicle, E2, is: ; Where m is the total vehicle mass; I x2 d is the moment of inertia of the vehicle about the second roll axis; d is the wheelbase of the vehicle. h is the angle between the Y-axis of the vehicle coordinate system and the horizontal plane; h0 is the vertical height of the first roll axis from the vehicle reference bearing plane; h s The height of the center of mass of the sprung mass; m u h is the unsprung mass. u The height of the center of mass of the unsprung mass; sin is the sine function; The total energy of vehicle real-time stability E total =E1 or E total =E2, combined with the vehicle's limit stability reference energy E max After normalization, we obtain the normalized vehicle stability energy E=E total / E max .

[0015] As a further improvement to the above scheme: a model predictive control algorithm is adopted, with the additional steering angle of each wheel, the torque of the wheel-side motor, and the active suspension as the power as the collaborative optimization control variables. The reference feedforward control quantity of each wheel is obtained by rolling time domain optimization solution, and then the final wheel control command is obtained by vector superposition of disturbance feedback compensation quantity.

[0016] As a further improvement to the above scheme, the process for obtaining the reference feedforward control values ​​for each wheel is as follows: Construct a nominal dynamic model of the vehicle without external disturbances, neglecting the influence of ground input disturbances, and establish the discrete state equations of the nominal dynamic model of the vehicle: ; Where x(k) and x(k+1) are the original state vectors of the vehicle nominal dynamics model at times k and k+1, respectively; u(k) is the original control input vector of the vehicle nominal dynamics model at time k; y(k) is the original output vector of the vehicle nominal dynamics model at time k; A, B, C and D are the original state matrix, original input matrix, original output matrix and original direct feed matrix of the vehicle nominal dynamics model, respectively. Based on x(k), construct the augmented state vector of the vehicle's nominal dynamics model at time k. Where u(k-1) is the original control input vector of the vehicle's nominal dynamics model at time k-1, and the superscript T denotes vector transpose; and the augmented discrete state equations of the vehicle's nominal dynamics model are obtained: ; Where ξ(k+1) is the augmented state vector of the vehicle's nominal dynamics model at time k+1; Augmented state matrix Augmented input matrix Augmented output matrix 0 m×n Let I be an m-row, n-column matrix containing all zeros. m Let be an m-dimensional identity matrix; η(k) be the augmented output vector of the vehicle's nominal dynamics model at time k; Δu(k) be the control increment vector of the vehicle's nominal dynamics model at time k. Δδ1, Δδ2, Δδ3 and Δδ4 are the increments of the additional steering angles δ1, δ2, δ3 and δ4 of the 1st, 2nd, 3rd and 4th wheels, respectively; T1, T2, T3 and T4 are the wheel-side motor torques of the 1st, 2nd, 3rd and 4th wheels, respectively; and F1, F2, F3 and F4 are the active suspension power sources of the 1st, 2nd, 3rd and 4th wheels, respectively. The predicted output equation of the vehicle's nominal dynamics model is obtained based on rolling time-domain prediction: ; in, This is the multi-step prediction output vector of the vehicle's nominal dynamics model; N c To control the time domain; N p For prediction in the time domain; η(k+1), η(k+2), η(k+N) c ) and η(k+N p The numbers ) represent the nominal dynamics models of the vehicle at k+1, k+2, and k+N, respectively. c and k+N p The augmented output vector at time step T; the superscript T denotes matrix transpose. , These are the corresponding prediction coefficient matrices; ; , , and These are the 1st, 2nd, and Nth digits in the prediction time domain, respectively. c and N p Moment and The product; This represents the predicted state vector of the vehicle's nominal dynamics model at time k. This is the predicted increment vector of the vehicle's nominal dynamics model at time k; ;Δu(k), Δu(k+1), Δu(k+N c The numbers ) represent the nominal dynamics models of the vehicle at k+1, k+2, and k+N, respectively. c Control increment vector at any given time; based on The objective optimization function J for constructing the nominal dynamics model of the whole vehicle nom : ; in, As a reference output vector; The prediction error vector; the expanded output weight matrix. , For N p An identity matrix of order 1. The Kronecker product is used, and Q is the single-step output weight matrix; the extended control weight matrix is ​​used. R is the single-step control weight matrix; Solving the objective function J under the physical boundary constraints of the actuator nom This yields the prediction increment vector that satisfies the optimal objective. , will This is denoted as the reference feedforward control increment.

[0017] As a further improvement to the above scheme, the process for obtaining the disturbance feedback compensation amount for each wheel is as follows: Construct a full-vehicle dynamics model that includes external road surface and lateral disturbances: ; in, , and These are the original state vector, original control vector, and original output vector of the actual vehicle dynamics model at time k, respectively. This represents the original state vector of the actual vehicle dynamics model at time k+1. This represents the disturbance input vector of the actual vehicle dynamics model at time k; Error state vector between the nominal dynamics model and the actual dynamics model of the vehicle Establish an error dynamics model: ; in, The first derivative of e(k); For the error dynamics model, an error optimization objective function J is constructed. real : ; Where e(N) is the terminal error state vector, that is, the error state vector at time N in the control time domain; The optimal feedback gain matrix K is solved recursively using the discrete algebraic Riccati equation, and the following recursive formula is defined: ; Where K(Nk) is the optimal feedback gain matrix K at time Nk in the control time domain; B(Nk) is the original input matrix at time Nk in the control time domain; P(k-1) is the solution to the Riccati equation at time k-1 in the control time domain; R R (Nk) is the extended control weight matrix at time Nk in the control time domain; A(Nk) is the original state matrix at time Nk in the control time domain; the superscript -1 indicates the inverse matrix; The solution P(k) of the Riccati equation at time k in the control time domain is calculated as follows: ; Among them, Q Q (Nk) is the extended output weight matrix at time Nk in the control time domain; The optimal feedback gain matrix K at each time point in the control time domain is obtained by recursive calculation using the recursive formula. Based on the obtained feedback gain matrix K, design a discrete LQR feedback control law: ; The disturbance feedback compensation increment at the current moment is calculated. .

[0018] As a further improvement to the above scheme, the physical boundary constraints of the actuator are specifically represented as follows: .

[0019] As a further improvement to the above scheme: identify driving style and obtain the corresponding stability energy threshold E. type The specific steps are as follows: The system collects driving feature samples within a preset time period, consisting of the maximum steering wheel angle, maximum steering wheel speed, vehicle speed, yaw rate, lateral acceleration, longitudinal acceleration, and lateral load transfer rate. The system then uses the K-means++ clustering algorithm to cluster the driving feature samples, classifying driving styles into three categories: cautious, moderate, and aggressive. Pre-calibrate the stability energy threshold E corresponding to each driving style type Initial weight k of differential braking type And the E of cautious drivers type >E for general-type drivers type >E of aggressive drivers type ; After the vehicle is powered on, it continuously collects real-time driving characteristics and updates the driving style recognition results based on these characteristics. It also updates the stability energy threshold E corresponding to the current vehicle in real time based on these recognition results. type With differential braking initial weight k type .

[0020] As a further improvement to the above scheme: before clustering the driving feature samples, factor dimensionality reduction is performed on the driving feature parameter matrix: the driving feature parameter matrix is ​​standardized to eliminate the influence of dimensions, resulting in a standardized matrix; the correlation coefficient matrix of the standardized matrix is ​​calculated and the eigenvalues ​​and corresponding eigenvectors are solved; principal factors are extracted according to the rule that the eigenvalue is greater than 1, and factor rotation is performed using the maximum variance method. The rotated factor loading matrix is ​​multiplied by the standardized matrix to obtain the score of each driving feature sample under each factor; the top three factors with a cumulative variance contribution rate greater than 85% are selected as the clustering input vectors.

[0021] A distributed drive-by-wire chassis personalized stability control system is characterized by executing a distributed drive-by-wire chassis personalized stability control method, comprising a data acquisition module, a driving style identification module, a stability energy solution module, a model predictive control feedforward module, an LQR disturbance feedback correction module, and an actuator drive module. The data acquisition module is used to collect raw vehicle operation data in real time; the driving style identification module completes feature dimensionality reduction, clustering and classification, and outputs the matching stability energy threshold and differential braking initial weight; the stability energy solution module uses the two-stage dynamic roll axis method to calculate the normalized vehicle stability energy; the model predictive control feedforward module is used to solve the baseline feedforward control increment; the LQR disturbance feedback correction module is used to solve the disturbance feedback compensation increment; the actuator drive module receives the superimposed complete control increment and controls the four-wheel additional steering angle, wheel-side drive motor, and active suspension to complete the chassis stability coordinated control action.

[0022] Compared with the prior art, the beneficial effects of the present invention are: 1. In the cooperative control method, the initial weights k of the differential braking are first matched based on the pre-identified driver style. type The system adapts to the operating characteristics of different drivers from the perspective of control benchmark, using the stability energy threshold E as the independent variable. Secondly, it constructs an exponential real-time differential braking weight calculation model with normalized vehicle stability energy E as the independent variable. This model ensures that the weights continuously and monotonically increase with rising stability risk. In the low-risk phase, active suspension intervention is dominant with low braking intervention, while braking intensity gradually increases in the high-risk phase, achieving adaptive matching between control intensity and operating condition risk. Simultaneously, the exponential function is continuously differentiable across the entire domain, ensuring that the weights precisely correspond to the initial values ​​at the start of the prevention zone and smoothly transition to full braking intervention at the end. This eliminates the rigidity of fixed weight control and avoids the abrupt changes of linear weights, guaranteeing smooth and vibration-free control throughout the entire process. Finally, through J... i = k1F i +μT i The weighted fusion mechanism combines the power output of the active suspension with the torque output of the wheel-side motors, and does not activate the additional wheel steering angle within the prevention zone. Under the premise of matching the risk level of the prevention zone, it systematically solves the three types of technical defects of the traditional weight allocation scheme, and takes into account both stability control effect and ride comfort.

[0023] 2. In the initial roll phase, the line connecting the instantaneous rotation centers of the front and rear axles is used as the first roll axis. Translational kinetic energy, rotational kinetic energy, and gravitational potential energy are calculated only for the sprung mass. This accurately reflects the actual physical form of the vehicle body roll around the suspension roll center under small roll conditions, eliminating redundant interference from unsprung mass in energy calculations during small rolls. It can sensitively capture risk changes in the early stages of roll, providing accurate quantitative basis for early smooth intervention in the prevention zone and avoiding inaccurate intervention timing due to energy calculation deviations. In the extreme roll phase, it automatically switches to using the line connecting the outer wheels to the ground as the second roll axis. The total energy is calculated uniformly for both sprung and unsprung mass, fully incorporating multiple dimensions such as vehicle coordinate system tilt angle, vertical height of the roll axis, changes in center of gravity height, and unsprung mass potential energy. By accurately reproducing the true mechanical characteristics of the vehicle's rollover around the wheel-side ground connection line under critical stability conditions, this system solves the core defects of a fixed rolling center axis that deviates significantly from the actual rotation fulcrum under large roll conditions and large errors in the estimation of critical stability energy. This significantly improves the accuracy and reliability of the danger zone trigger judgment. Finally, by normalizing the limit stability benchmark energy, the stability energy of two different calculation dimensions is unified to a standardized dimension, achieving consistency and comparability of risk quantification during full roll. It can be directly matched with personalized thresholds and three-zone division rules that are seamlessly integrated with driving style, providing a precise judgment basis for subsequent dynamic weight allocation and multi-actuator collaborative control, fundamentally improving the control accuracy and condition adaptability of the entire stability system. Attached Figure Description

[0024] Figure 1 This is a flowchart of the collaborative control method.

[0025] Figure 2 A block diagram for integrating stability control into the vehicle chassis.

[0026] Figure 3 This is a schematic diagram of a nine-degree-of-freedom dynamics model of a vehicle.

[0027] Figure 4 This is a schematic diagram of the vehicle rotating around the first tilt axis.

[0028] Figure 5 This is a schematic diagram of the vehicle rotating around the second tilt axis.

[0029] Figure 6 This is a timing diagram of mode switching under experimental conditions for fishhooks.

[0030] Figure 7 This is a comparison curve of vehicle roll angle between the pipe model predictive control and the traditional model predictive control for a cautious driver under double lane change conditions.

[0031] Figure 8 This is a comparison curve of vehicle roll angle between the pipe model predictive control and the traditional model predictive control for a typical driver under double lane change conditions.

[0032] Figure 9 This is a comparison curve of vehicle roll angle between the pipe model predictive control and the traditional model predictive control for an aggressive driver under double lane change conditions. Detailed Implementation

[0033] 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 embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] like Figure 1 and Figure 2As shown, this invention provides a distributed drive-by-wire chassis personalized stability control method, applied to vehicles equipped with a distributed drive system. The vehicle is equipped with four-wheel independent steering actuators, four-wheel independent wheel-side drive motors, four-wheel independent active suspension units, and data acquisition devices such as vehicle state sensors, steering wheel angle sensors, vehicle speed sensors, and inertial measurement units. The control method acquires driver behavior data by constructing a driving acquisition scenario, and uses factor analysis, K-means++ clustering, and LSTM neural networks to identify driving styles in real time and match personalized stability energy thresholds. Based on this, a nine-degree-of-freedom vehicle dynamics model is constructed, and a chassis integrated stability controller based on Tube Model Predictive Control (Tube MPC) is designed. The vehicle state is divided into safe, preventative, and dangerous zones according to lateral acceleration and stability energy. Differentiated distributed drive control strategies are adopted in each zone, and an event triggering mechanism is introduced to save communication resources, ultimately achieving personalized, high-precision collaborative vehicle stability control.

[0035] I. Driving Data Collection and Driving Style Recognition

[0036] (I) Driving scenario setup and raw data collection

[0037] A driver driving scenario was constructed, using an arc-shaped loop with different curvatures and a double lane-changing road as test roads. In this embodiment, PreScan software and Simulink were used for joint simulation, and the test conditions were constructed in conjunction with the Logitech G29 driving simulator. Multiple drivers were recruited to conduct driving tests on the simulator, and raw data was collected during the driving process. The raw data included the maximum steering wheel angle, the maximum steering wheel speed, vehicle speed, yaw rate, lateral acceleration, longitudinal acceleration, and lateral load transfer rate.

[0038] (II) Raw Data Preprocessing

[0039] The collected raw data were processed sequentially, including outlier removal, filtering and noise reduction, and effective segment extraction. The specific steps are as follows: 1. Outlier detection based on interquartile range is used to remove outliers from the original data: the upper quartile Q3 and the lower quartile Q1 of each data sequence are calculated, and data points that exceed the range of [Q1-1.5(Q3-Q1), Q3+1.5(Q3-Q1)] are identified as outliers and removed. The missing positions after removal are filled by cubic spline interpolation.

[0040] 2. The completed data is passed through a Butterworth low-pass filter with a cutoff frequency of 5Hz to filter out high-frequency noise from the sensor.

[0041] 3. Extract continuous segments with an absolute value of steering wheel angle change greater than 10° / s or vehicle speed change greater than 0.5m / s² as valid driving intention data for subsequent calculation of driving characteristic parameters.

[0042] (III) Extraction of driving style feature parameters

[0043] Statistical feature parameters are extracted from valid driving intention data to obtain driving feature parameters, specifically including: 1. The peak value of the rate of change of steering wheel angle, i.e., the maximum value of the time derivative of the steering wheel angle during the data collection period; 2. The standard deviation of the steering wheel angle during turning; 3. The root mean square value of the yaw rate; 4. Average turning speed, which is the average longitudinal speed of the vehicle under turning conditions; 5. The coefficient of variation of vehicle speed; 6. The mean of the absolute values ​​of the yaw rate; 7. The mean value of the absolute value of lateral acceleration; 8. Standard deviation of the absolute value of lateral acceleration; 9. Peak value of lateral load transfer rate.

[0044] (iv) Factor analysis dimensionality reduction

[0045] The above driving feature parameter matrix is ​​subjected to factorization dimensionality reduction to eliminate data redundancy and the influence of dimensions, resulting in low-dimensional input indicators for clustering. The specific process is as follows: The original driving feature parameter matrix is ​​standardized to eliminate the influence of dimensions, resulting in a standardized matrix. The correlation coefficient matrix of this standardized matrix is ​​calculated, and its eigenvalues ​​and corresponding eigenvectors are solved. Principal factors are extracted based on the criterion that the eigenvalue is greater than 1, and factor rotation is performed using the maximum variance method to simplify the factor loading matrix structure. The score coefficient matrix of each factor is calculated, and the rotated factor loading matrix is ​​multiplied by the standardized matrix to obtain the score of each sample on each factor. The top three factors with a cumulative variance contribution rate exceeding 85% are selected as the principal factors after dimensionality reduction, and their score coefficient matrix is ​​the coefficient matrix of the three principal factors. The coefficient matrix is ​​multiplied by the normalized standardized matrix to obtain the final input index for driving style identification, and the corresponding input index dataset is constructed.

[0046] (v) K-means++ clustering and driving style classification

[0047] The top three indicators in terms of cumulative contribution rate from factor analysis were selected as the clustering input vector. The K-means++ classification algorithm was used to cluster the driving feature parameters, resulting in several clusters, each representing a different driving style. The specific steps are as follows: 1. Initial cluster center selection: Randomly select the first cluster center from the input index dataset; then calculate the shortest distance between each sample point and the currently selected cluster center, and select the next cluster center according to the probability distribution proportional to the square of the distance. Repeat this process until M initial cluster centers are selected. In this invention, 3 are selected (M=3). 2. Iterative Clustering: Perform standard K-means++ iterations, assign each sample point to the nearest cluster center, and update the cluster center to the mean of all samples in the current cluster; repeat the assignment and update steps until the cluster centers no longer change or the preset maximum number of iterations is reached; 3. Style tag mapping: The final result is three clusters, which correspond to the driver's cautious, normal and aggressive driving styles respectively.

[0048] (vi) Training and online recognition of LSTM driving style recognition model

[0049] A real-time driving style identification model is built using a long short-term memory (LSTM) neural network. The driving style labels obtained by clustering are used as the model output, and the input indicators after factor dimensionality reduction are used to construct sequential sample data according to time windows.

[0050] 70% of the total data was selected as the training set for training and optimizing the LSTM model; the remaining 30% of the data was used as the validation and test set for validating and evaluating the model performance.

[0051] The LSTM output layer uses a softmax function to map the hidden states to probability distributions of various driving styles, and the category corresponding to the maximum probability is used as the driver style recognition result. Model training uses the cross-entropy loss function, the Adam optimizer, an initial learning rate of 0.001, a batch size of 32, and 200 iterations. An early stopping mechanism is set to monitor the validation set loss; training stops if the loss does not decrease for 10 consecutive iterations. Finally, the model with the highest validation accuracy is saved as the driver style classifier.

[0052] After the vehicle is powered on, real-time driving features are continuously collected. The driving style recognition results are updated continuously using a trained LSTM model, and the stability energy threshold E corresponding to the current vehicle is updated in real time based on the recognition results. type With differential braking initial weight k type Pre-calibrate the stability energy threshold E corresponding to each driving style. type Initial weight k of differential braking type And the E of cautious drivers type >E for general-type drivers type >E of aggressive drivers type In this embodiment, the initial differential braking weight k for cautious and normal drivers is...type =0.2, the initial weight k of differential braking for aggressive drivers type =0.1.

[0053] II. Construction of a Nine-DOF Vehicle Dynamics Model

[0054] like Figure 3 As shown, a nine-degree-of-freedom (DOF) dynamic model of the vehicle is established. This model consists of a seven-DOF suspension model and a two-DOF model. The seven-DOF suspension model considers pitch, roll, and ground disturbance; the two-DOF model describes the vehicle's yaw and lateral motion, and its input is provided by a distributed drive system that supports independent output of additional steering angle, drive / braking torque, and suspension dynamics for each wheel.

[0055] Figure 3 This is the core component of the nine-degree-of-freedom vehicle dynamics model used in this invention. The diagram uses oblique projection to represent the vehicle suspension system structure, establishing a vehicle dynamics coordinate system. The x-axis extends forward longitudinally, the y-axis extends laterally to the right, and the z-axis extends vertically upward. The origin of the coordinate system is located at the center of mass of the sprung mass. The cuboid structure in the diagram represents the sprung mass of the vehicle, i.e., the vehicle body, and the lateral velocity v is marked at the center of mass. y The sprung mass has three rotational degrees of freedom: roll motion about the x-axis corresponding to the roll angle. The pitch motion around the y-axis corresponds to the pitch angle θ, and the yaw motion around the z-axis is also labeled with the angle Ψ. The geometric parameters of the model include: a is the longitudinal distance from the vehicle's center of gravity to the front axle, b is the longitudinal distance from the vehicle's center of gravity to the rear axle, and d is the vehicle's track width. The suspension system comprises four independent wheel units, each consisting of parallel suspension actuation branches, unsprung mass, and tire elasticity branches connected in series. The parameter symbols for each unit are as follows: the active suspension actuation forces F1~F4 for the first to fourth wheel units, and the suspension spring stiffness k. s1 ~k s4 Suspension damping coefficients c1~c4, additional suspension forces f1~f4, and vertical displacement z of the sprung mass at that position. s1 ~z s4 Unsprung mass m u1 ~m u4 Vertical displacement z of unsprung mass u1 ~z u4 Vertical excitation displacement z of the road surface g1 ~z g4The active suspension is arranged in parallel with the suspension springs, dampers, and additional forces, acting together between the sprung and unsprung masses to actively output vertical forces to adjust the vehicle's roll attitude. The unsprung mass is connected to the road surface through the tire elastic branch, and the vertical excitation displacement of the road surface is used to simulate the impact of ground unevenness disturbances on the vehicle's roll state. This seven-degree-of-freedom suspension model is coupled with the two-degree-of-freedom lateral vehicle model to form the nine-degree-of-freedom vehicle dynamics model of this invention, providing a precise dynamic basis for subsequent stability energy calculations and stability controller design.

[0056] A nine-degree-of-freedom (DOF) dynamic model of the vehicle is established, which is composed of a seven-DOF suspension model and a two-DOF vehicle model. Seven-DOF suspension model: Considering the vehicle's pitch, roll and ground disturbances, it includes the vertical hop freedom of the four wheels, the vertical motion freedom of the sprung mass, and the pitch and roll motion freedom of the vehicle body. Two-degree-of-freedom vehicle model: used to describe the lateral and yaw motion of the vehicle. It is composed of four independent wheel units integrating steering, drive / braking and suspension sub-models. Each wheel unit independently outputs the additional wheel angle, driving torque or braking torque of the corresponding wheel, as well as the active suspension power.

[0057] Ignoring the changes in tire elastic potential energy and suspension system potential energy during vehicle dynamic stability, and only considering the changes in translational kinetic energy, rotational kinetic energy, and gravitational potential energy, the vehicle stability energy is calculated based on the above model.

[0058] III. Solving the Normalized Vehicle Stability Energy Using the Two-Segment Dynamic Roll Axis Method

[0059] The normalized vehicle stability energy E is calculated using the two-stage dynamic roll axis method. Based on the change of the roll axis during the vehicle roll process, the roll process is divided into two characteristic stages: the initial roll stage and the extreme roll stage, which correspond to different roll axes and energy calculation models.

[0060] (I) First stage: Initial roll stage

[0061] like Figure 4 As shown, the first stage is the initial tilting stage. The line connecting the instantaneous rotation centers of the front and rear axles is the first tilting axis. The sprung mass rotates around this axis. The energy is calculated only for the sprung mass. The total sprung energy E1 is the sum of translational kinetic energy, rotational kinetic energy, and gravitational potential energy. ; Where, m s v is the sprung mass. y I represents the lateral speed of the vehicle. x1 Let be the moment of inertia of the sprung mass about the first tilt axis; The roll angle is... ω is the roll angular velocity; g is the gravitational acceleration; h1 is the distance from the center of mass of the spring mass to the first roll axis; cos is the cosine function.

[0062] (II) Second stage: Extreme roll stage

[0063] After the vehicle completes the first stage around the first roll axis, if the steering angle continues to increase, when Greater than the set tilt angle threshold At this time, the lateral tilting motion enters the second stage; such as Figure 5 As shown, the second stage is the extreme roll stage. Taking the line connecting the outer wheel to the ground as the second roll axis, the entire vehicle rotates around this axis. The total energy of the entire vehicle's sprung and unsprung masses is calculated uniformly. The total energy E2 of the entire vehicle is the sum of translational kinetic energy, rotational kinetic energy, gravitational potential energy of sprung mass, and gravitational potential energy of unsprung mass. ; Where m is the total vehicle mass; I x2 Let be the moment of inertia of the vehicle about the second roll axis; h is the angle between the Y-axis of the vehicle coordinate system and the horizontal plane; h0 is the vertical height of the first roll axis from the vehicle reference bearing plane; h s The height of the center of mass of the sprung mass; m u h is the unsprung mass. u denoted as , where is the height of the center of mass of the unsprung mass; sin is the sine function.

[0064] (III) Normalization Processing

[0065] The energy value for the corresponding stage is selected based on the vehicle's real-time roll state, i.e., the total energy for real-time vehicle stability, E. total =E1 or E total =E2. Combined with the vehicle's limit stability reference energy E max Normalizing the total energy value under the critical stability state, we obtain the normalized vehicle stability energy E=E total / E max .

[0066] IV. Three-zone division of vehicle operating status and event triggering mechanism

[0067] (I) Rules for dividing the three zones

[0068] The three-zone division of vehicle operating status is based on the vehicle's lateral acceleration a. y Determination based on two indicators: normalized vehicle stability energy E. when a y When the concentration is less than 0.4g and E is less than 0.3, it is considered a safe zone; When 0.4g≤ay <0.85g and 0.3≤E <E type At that time, it was determined to be a prevention zone; When E type When the value is less than or equal to E, it is considered a danger zone. Where g is the acceleration due to gravity, E type Stability control trigger thresholds are set for different driving styles.

[0069] (ii) Event Triggering Control Mechanism

[0070] A preset control trigger update rule is used to determine whether a control command update should be triggered. When the vehicle state change meets the trigger update rule, the controller updates and issues a control command. If no trigger is triggered, the control command from the previous moment remains unchanged, thereby reducing communication and computing resource consumption. The trigger rule is bound to the stability energy change rate and state partition switching conditions. When the stability energy change exceeds a set threshold or the vehicle is running across partitions, a control update is triggered immediately.

[0071] V. Constructing a Distributed Drive Chassis Integrated Stability Controller

[0072] A distributed drive chassis integrated stability controller is built using Tube Model Predictive Control (Tube MPC). The additional steering angle of each wheel, the torque of the wheel-side motor, and the active suspension are used as the power sources for collaborative optimization control variables. The baseline feedforward control quantities for each wheel are obtained through rolling time-domain optimization, and then the final wheel control commands are obtained by vector superposition of disturbance feedback compensation quantities. The controller is composed of a combined Model Predictive Control feedforward module and an LQR disturbance feedback correction module (MPC feedforward + LQR feedback correction).

[0073] (I) Incremental solution of reference feedforward control

[0074] 1. Discrete state equations of the vehicle's nominal dynamics model

[0075] Construct a nominal dynamic model of the vehicle without external disturbances, neglecting the influence of ground input disturbances, and establish the discrete state equations of the nominal dynamic model of the vehicle: ; Where x(k) and x(k+1) are the original state vectors of the vehicle nominal dynamics model at times k and k+1, respectively; u(k) is the original control input vector of the vehicle nominal dynamics model at time k; y(k) is the original output vector of the vehicle nominal dynamics model at time k; A, B, C and D are the original state matrix, original input matrix, original output matrix and original direct feed matrix of the vehicle nominal dynamics model, respectively.

[0076] 2. Constructing an augmented state space

[0077] To avoid oscillations caused by the controller directly outputting the steering angles of each wheel, the input variables are converted into incremental form, and the augmented state vector of the vehicle's nominal dynamics model at time k is constructed based on x(k): , where u(k-1) is the original control input vector of the vehicle's nominal dynamics model at time k-1, and the superscript T indicates vector transpose.

[0078] The augmented discrete state equations of the vehicle's nominal dynamics model are obtained as follows: ; Where ξ(k+1) is the augmented state vector of the vehicle's nominal dynamics model at time k+1.

[0079] The augmented state matrix is: ; The augmented input matrix is: ; The augmented output matrix is: ; Among them, 0 m×n Let I be an m-row, n-column matrix containing all zeros. m Let be an m-dimensional identity matrix; η(k) be the augmented output vector of the vehicle's nominal dynamics model at time k; Δu(k) be the control increment vector of the vehicle's nominal dynamics model at time k.

[0080] The control increment vector of the vehicle's nominal dynamics model at time k is specifically represented as follows: Δδ1, Δδ2, Δδ3 and Δδ4 are the increments of the additional steering angles δ1, δ2, δ3 and δ4 of the 1st, 2nd, 3rd and 4th wheels, respectively; T1, T2, T3 and T4 are the wheel-side motor torques of the 1st, 2nd, 3rd and 4th wheels (positive values ​​indicate driving and negative values ​​indicate braking); and F1, F2, F3 and F4 are the active suspension power sources of the 1st, 2nd, 3rd and 4th wheels, respectively.

[0081] 3. Rolling Time Domain Prediction Output Equation

[0082] The predicted output equation of the vehicle's nominal dynamics model is obtained based on rolling time-domain prediction: ; in, The multi-step prediction output vector of the vehicle's nominal dynamics model is represented as follows: ; The output vector includes the sideslip angle β, yaw rate ω, and vertical velocity of the vehicle's center of gravity. Pitch angle θ, roll angle The dynamic deflection z of the suspension of the four wheels s1 -z u1 z s2 -z u2 z s3 -z u3 z s4 -z u4 There are 9 states in total; N c To control the time domain; N p For prediction in the time domain; η(k+1), η(k+2), η(k+N) c ) and η(k+N p The numbers ) represent the nominal dynamics models of the vehicle at k+1, k+2, and k+N, respectively. c and k+N p The augmented output vector at time step.

[0083] , These are the corresponding prediction coefficient matrices, specifically the state prediction coefficient matrix. Specifically, it is expressed as follows: , , , and These are the 1st, 2nd, and Nth digits in the prediction time domain, respectively. c and N p Moment and The product of.

[0084] This is the predicted state vector of the vehicle's nominal dynamics model at time k.

[0085] This is the predicted increment vector of the vehicle's nominal dynamics model at time k; ;Δu(k), Δu(k+1), Δu(k+N c The numbers ) represent the nominal dynamics models of the vehicle at k+1, k+2, and k+N, respectively. c Control increment vector at time step.

[0086] 4. Solving the objective optimization function and the baseline feedforward control increment

[0087] Based on the predicted output vector The objective optimization function J for constructing the nominal dynamics model of the whole vehicle nom : ; in, As a reference output vector; The prediction error vector; the expanded output weight matrix. , For N p An identity matrix of order 1. The Kronecker product is used, and Q is the single-step output weight matrix; the extended control weight matrix is ​​used. R is the single-step control weight matrix.

[0088] Solving the objective function J under the physical boundary constraints of the actuator nom The predicted increment vector that satisfies the optimal objective is obtained, and this predicted increment vector is denoted as the reference feedforward control increment.

[0089] The physical boundary constraints of the actuator are specifically represented as follows: .

[0090] (II) Solution of Incremental Disturbance Feedback Compensation

[0091] 1. Actual dynamics model and error dynamics model of the whole vehicle

[0092] Construct a full-vehicle dynamics model that includes external road surface and lateral disturbances: ; in, , and These are the original state vector, original control vector, and original output vector of the actual vehicle dynamics model at time k, respectively. This represents the original state vector of the actual vehicle dynamics model at time k+1. This is the disturbance input vector of the actual vehicle dynamics model at time k.

[0093] Error state vector between the nominal dynamics model and the actual dynamics model of the vehicle Establish an error dynamics model: ; in, It is the first derivative of the error state vector e(k).

[0094] 2. LQR Error Optimization Objective Function

[0095] For the error dynamics model, an error optimization objective function J is constructed. real : ; Where e(N) is the terminal error state vector, i.e., the error state vector at time N in the control time domain; N is the LQR finite optimization time domain, and the prediction time domain N of the model predictive control feedforward module. p Control time domain N c They are independent of each other and their values ​​are unrelated.

[0096] 3. Recursive Solution of Discrete Algebraic Riccati Equations

[0097] The optimal feedback gain matrix K is solved recursively using the discrete algebraic Riccati equation, and the following recursive formula is defined: ; Where K(Nk) is the optimal feedback gain matrix K at time Nk in the control time domain; B(Nk) is the original input matrix at time Nk in the control time domain; P(k-1) is the solution to the Riccati equation at time k-1 in the control time domain; R R (Nk) is the extended control weight matrix at time Nk in the control time domain; A(Nk) is the original state matrix at time Nk in the control time domain.

[0098] The solution P(k) of the Riccati equation at time k in the control time domain is calculated as follows: ; Among them, Q Q (Nk) is the extended output weight matrix at time Nk in the control time domain.

[0099] The optimal feedback gain matrix K at each time point in the control time domain is obtained by recursive calculation using a recursive formula.

[0100] Based on the obtained feedback gain matrix K, design a discrete LQR feedback control law: ; The disturbance feedback compensation increment at the current moment is calculated. .

[0101] The reference feedforward control increment and the disturbance feedback compensation increment vector are superimposed to obtain the final complete control increment sent to each wheel unit.

[0102] VI. Differentiated Collaborative Control Strategy by Region

[0103] When the event-triggered mechanism determines that the control needs to be updated, the distributed drive chassis integrated stability controller sends corresponding control commands to each wheel according to the control area where the vehicle is located, thereby achieving zoned differentiated collaborative stability control.

[0104] (a) Safe Zone Control Strategy

[0105] When the vehicle is in a safe zone, suspension power commands are only sent to the active suspension units of each wheel, and the active suspension outputs power according to the basic parameters; no additional wheel steering angle is output, and the wheel-side drive motors do not output additional torque, maintaining the normal driving state of the vehicle.

[0106] (II) Prevention Zone Control Strategy

[0107] When the vehicle is in the prevention zone, wheel-side motor torque commands and active suspension power commands are sent to the drive motors and active suspension units of each wheel. At this time, the steering motors of each wheel are not working, and vehicle stability control is achieved solely through the coordinated action of the active suspension and differential braking. The specific steps are as follows: 1. Set the corresponding initial weight k for differential braking based on the currently identified driving style. type The real-time differential braking weight μ is calculated according to the following exponential formula: ; The real-time differential braking weight μ increases exponentially from its initial value with the normalized vehicle stability energy E within the prevention zone, achieving a smooth transition from suspension dominance to gradual intervention of differential braking.

[0108] 2. Derivation of the differential braking weight formula

[0109] The real-time differential braking weights are constructed using an exponential function, and the derivation process is as follows: (1) Design objectives The design of differential braking weights must meet the following boundary conditions and performance requirements: ① Starting boundary: When the normalized vehicle stability energy E = 0.3 (starting point of the prevention zone), the differential braking weight takes the initial value μ = k. type Where k type The initial weight for differential braking is determined by driver style: k is chosen for cautious and moderate drivers. type =0.2, aggressive drivers take k type =0.1; At this point, the braking intervention is minimal, with active suspension as the primary means of stability control, avoiding premature intervention that could negatively impact the driving experience.

[0110] ②Endpoint Boundary: When the normalized vehicle stability energy E=Etype (danger zone trigger threshold), the differential braking weight μ=1. At this time, the differential braking fully intervenes, providing a smooth connection for the subsequent multi-actuator coordinated control entering the danger zone and avoiding sudden changes in control quantity when switching zones.

[0111] ③ Smoothness requirement: μ increases monotonically with the normalized stability energy E, and the function is continuously differentiable over the entire domain to eliminate control steps and avoid vehicle vibration and driving discomfort.

[0112] ④ Adaptive requirement: The growth rate of the differential braking weight can vary with the initial weight k type Adaptive adjustment, with different weight growth curves corresponding to different driving styles, to achieve personalized intervention intensity matching.

[0113] (2) Function form selection

[0114] While using a linear function μ=aE+b (where a and b are the corresponding proportionality coefficient and intercept) to construct the weights is simple, it cannot simultaneously accommodate differentiated initial weights for different driving styles and a uniform endpoint threshold. Furthermore, the fixed growth rate fails to meet the intervention rhythm requirements of different driving styles. Therefore, an exponential function is chosen to construct the weights. (1) In the formula, λ is the weight growth coefficient, which is used to control the growth rate of the weight with the stability energy.

[0115] This exponential form naturally satisfies the initial boundary condition: when E = 0.3, the exponential term is 0, and at this time μ = k type It perfectly matches the initial braking weight; and the function is continuously differentiable over the entire domain, with its monotonicity guaranteed by the λ sign, which meets the design requirements for smooth control.

[0116] (3) Solving for the growth coefficient

[0117] Substitute the endpoint boundary conditions into the above exponential function to solve for the growth coefficient λ.

[0118] When E=E type When μ is required to be 1, substituting it into the exponential function yields: (2) Simplify both sides of the equation, placing the term containing the exponent separately on one side: (3) Take the natural logarithm of both sides of the equation and use the properties of logarithms. and Simplify: (4) After simplification, the analytical expression for the growth coefficient is obtained: (5) (4) Final Formula Substituting the growth coefficient λ obtained from the solution back into the initial exponential function (Equation (1)), we obtain the final formula for calculating the real-time differential braking weight: (6) The differential braking weight calculated by formula (6) increases exponentially and smoothly from the initial value with the normalized stability energy in the prevention zone, realizing a continuous transition from "suspension-dominated" to "gradual braking intervention". At the same time, different driving styles correspond to different initial weights and growth rates. Aggressive drivers have lower initial weights and faster growth, while cautious drivers have higher initial weights and slower growth, which are precisely matched with the driving characteristics of the driver.

[0119] 3. Based on the active suspension force F of the i-th wheel uniti With the torque T of the wheel-side motor i The overall control output J of the wheel unit is calculated according to the following formula. i : ; Wherein, k1 is the active suspension weight, and in this embodiment, k1=1; the wheel-side motor torque is output in the form of negative torque to achieve differential braking effect.

[0120] 4. Based on the comprehensive control output J of each wheel unit i The active suspension and distributed drive motor of the corresponding wheel unit are synchronously controlled, and no additional wheel steering angle is output within the prevention zone.

[0121] (III) Hazardous Area Control Strategy

[0122] When the vehicle is in a danger zone, additional steering angle commands, drive / braking torque commands, and suspension action commands are simultaneously sent to the steering motors, drive motors, and active suspension units of each wheel to execute a full-domain multi-actuator collaborative control strategy. If the danger zone does indeed increase energy deviation terms, corresponding terms and weights need to be added to the objective function. If only the original output weights are adjusted, it is necessary to specify which weight to adjust and the adjustment coefficient to ensure consistency between the context and the control variables. The three types of control variables—four-wheel additional steering angle, wheel-side motor torque, and active suspension action—are optimized synchronously. After superimposing the baseline feedforward control increment and the disturbance feedback compensation increment, a complete control command is output. The vehicle's roll tendency is suppressed by the combined action of four-wheel steering, differential motor torque, and active suspension, achieving full-actuator collaborative stability control.

[0123] VII. Verification of Implementation Results

[0124] To verify the effectiveness of this method, comparative tests were conducted in the TruckSim / Simulink co-simulation environment. The hook condition and the double line-shifting condition were selected as typical test scenarios. Traditional model predictive control (MPC) was used as a control to verify the stability control effect and partition switching characteristics of the tube model predictive control (Tube MPC) strategy adopted in this invention.

[0125] (a) Features of switching between different working conditions for fishing hooks

[0126] Figure 6 This is a comparison diagram of the control interval switching timing between the traditional MPC and the Tube MPC of this invention under the experimental conditions of a fishhook. Figure 6 The horizontal axis represents time in seconds, ranging from 0 to 12 seconds; the vertical axis represents the control zone level, with values ​​1, 2, and 3 corresponding to the safe zone, prevention zone, and danger zone, respectively. Figure 6As can be seen, under the fishhook steering condition, as the lateral excitation increases and the stability energy rises, both control strategies sequentially move from the safe zone to the prevention zone and then to the danger zone according to the risk level. After the steering condition ends and the stability risk decreases, they gradually return to the safe zone. Compared with the traditional MPC, the Tube MPC strategy of this invention enters the high-level control zone later and has a shorter duration in the prevention and danger zones, indicating that it can more effectively suppress stability risks and allow the vehicle to return to a safe driving state more quickly. At the same time, it reduces unnecessary strong control interventions, avoids the impact of premature and excessive intervention on the driver's operating experience, and effectively improves driving smoothness and ride comfort.

[0127] (II) Side tilt control effect under dual lane change conditions

[0128] Simulation tests were conducted under double lane change conditions for three driving styles: cautious, normal, and aggressive. The body roll angle control effect under different styles was compared to verify the effectiveness of personalized stability control.

[0129] Figure 7 Comparison curves of vehicle roll angle under double lane change conditions for cautious drivers. Figure 7 The horizontal axis represents time in seconds, ranging from 0 to 15 seconds; the vertical axis represents the vehicle body roll angle in degrees. Figure 7 As can be seen, under the double lane change steering condition, the roll angle of both control strategies exhibits the typical double lane change response characteristics of two opposing fluctuations. Compared with the traditional MPC, the Tube MPC strategy of this invention has a smaller peak roll angle, faster roll fluctuation convergence speed, and a smoother curve transition without obvious high-frequency oscillations. These results indicate that, for cautious drivers, this method can achieve stability control with a gentler intervention while ensuring driving safety, adapting to the smooth driving needs of cautious drivers and avoiding excessive intervention that disrupts the driving experience.

[0130] Figure 8 The curves show the vehicle roll angle comparison under typical driver double lane change conditions. Figure 8 The horizontal axis represents time in seconds, ranging from 0 to 15 seconds; the vertical axis represents the vehicle body roll angle in degrees. Figure 8 As can be seen, compared with traditional MPC, the Tube MPC strategy of this invention can effectively reduce the peak roll angle, suppress the roll fluctuation amplitude, and accelerate the vehicle's attitude recovery speed. Considering the driving characteristics of typical drivers, this method achieves a balanced control effect between stability and driving comfort, ensuring stability control under extreme conditions without abruptly affecting the driving experience, and is highly compatible with the driving habits of typical drivers.

[0131] Figure 9Comparison curves of vehicle roll angle under double lane change conditions for aggressive drivers. Figure 9 The horizontal axis represents time in seconds, ranging from 0 to 15 seconds; the vertical axis represents the vehicle body roll angle in degrees. Figure 9 As can be seen, vehicle roll fluctuations are more severe under aggressive driving conditions. Traditional MPC control results in a larger peak roll angle and significant oscillations during the recovery process, indicating insufficient disturbance resistance. The Tube MPC strategy of this invention can significantly suppress the peak roll angle, weaken secondary oscillations, and quickly stabilize the vehicle's attitude. Addressing the higher instability risk associated with aggressive driving styles, this method demonstrates superior stability control and robustness, effectively improving vehicle driving safety under extreme conditions and reducing the probability of instability accidents.

[0132] In summary, the distributed drive-by-wire chassis personalized stability control method proposed in this invention can dynamically adjust the timing and intensity of intervention according to the driver's style. It can achieve smooth switching between zone modes under the "fishhook" condition and effectively reduce the vehicle roll angle and improve posture stability under the "double lane change" condition. Compared with the traditional MPC, the Tube MPC control strategy of this solution has significant advantages in disturbance suppression, response smoothness and risk control efficiency, taking into account both vehicle driving safety and driving comfort.

[0133] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A distributed drive-by-wire chassis personalized stability control method, applied to a vehicle stability control system that divides the vehicle's operating state into a safe zone, a prevention zone, and a danger zone. This vehicle stability control system pre-identifies driving styles and obtains the corresponding stability energy threshold E for each driving style. type Furthermore, the normalized vehicle stability energy E is calculated in real time, characterized in that... When the vehicle is in the prevention zone, the following control steps are executed: The initial differential braking weight k is set according to the driving style. type The real-time differential braking weight μ is calculated according to the following exponential formula: ； The active suspension power F of the i-th wheel unit i With the torque T of the wheel-side motor i ,according to Calculate the overall control output J of the wheel unit. i Where k1 is the active suspension weight; According to each J i The active suspension and distributed drive motor of the corresponding wheel unit are synchronously controlled, and no additional wheel angle is output in the prevention zone. Vehicle stability control is achieved only by the active suspension and differential braking working together. The normalized stability energy E is solved using a two-stage dynamic roll axis: The first stage is the initial roll stage. The line connecting the instantaneous rotation centers of the front and rear axles is taken as the first roll axis. Energy is calculated only for the sprung mass. The total sprung energy E1 is: ； Where, m s v is the sprung mass. y I represents the lateral speed of the vehicle. x1 Let be the moment of inertia of the sprung mass about the first tilt axis; The roll angle is... ω is the roll angular velocity; g is the acceleration due to gravity; h1 is the distance from the center of mass of the sprung mass to the first roll axis; cos is the cosine function; The second stage is the extreme roll stage. Using the outer wheel contact line as the second roll axis, the total energy is calculated for both sprung and unsprung masses of the vehicle. The total energy of the vehicle, E2, is: ； Where m is the total vehicle mass; I x2 d is the moment of inertia of the vehicle about the second roll axis; d is the wheelbase of the vehicle. h is the angle between the Y-axis of the vehicle coordinate system and the horizontal plane; h0 is the vertical height of the first roll axis from the vehicle reference bearing plane; h s The height of the center of mass of the sprung mass; m u h is the unsprung mass. u The height of the center of mass of the unsprung mass; sin is the sine function; The total energy of vehicle real-time stability E total =E1 or E total =E2, combined with the vehicle's limit stability reference energy E max After normalization, we obtain the normalized vehicle stability energy E=E total / E max ; A model predictive control algorithm is adopted, using the additional steering angle of each wheel, the torque of the wheel-side motor, and the active suspension as power as the co-optimized control variables. The reference feedforward control quantity of each wheel is obtained through rolling time-domain optimization, and then the final wheel control command is obtained by vector superposition of disturbance feedback compensation quantity. The process of obtaining the reference feedforward control quantity of each wheel is as follows: Construct a nominal dynamic model of the vehicle without external disturbances, neglecting the influence of ground input disturbances, and establish the discrete state equations of the nominal dynamic model of the vehicle: ； Where x(k) and x(k+1) are the original state vectors of the vehicle nominal dynamics model at times k and k+1, respectively; u(k) is the original control input vector of the vehicle nominal dynamics model at time k; y(k) is the original output vector of the vehicle nominal dynamics model at time k; A, B, C and D are the original state matrix, original input matrix, original output matrix and original direct feed matrix of the vehicle nominal dynamics model, respectively. Based on x(k), construct the augmented state vector of the vehicle's nominal dynamics model at time k. Where u(k-1) is the original control input vector of the vehicle's nominal dynamics model at time k-1, and the superscript T denotes vector transpose; and the augmented discrete state equations of the vehicle's nominal dynamics model are obtained: ； Where ξ(k+1) is the augmented state vector of the vehicle's nominal dynamics model at time k+1; Augmented state matrix Augmented input matrix Augmented output matrix 0 m×n Let I be an m-row, n-column matrix containing all zeros. m Let be an m-dimensional identity matrix; η(k) be the augmented output vector of the vehicle's nominal dynamics model at time k; Δu(k) be the control increment vector of the vehicle's nominal dynamics model at time k. Δδ1, Δδ2, Δδ3 and Δδ4 are the increments of the additional steering angles δ1, δ2, δ3 and δ4 of the 1st, 2nd, 3rd and 4th wheels, respectively; T1, T2, T3 and T4 are the wheel-side motor torques of the 1st, 2nd, 3rd and 4th wheels, respectively; and F1, F2, F3 and F4 are the active suspension power sources of the 1st, 2nd, 3rd and 4th wheels, respectively. The predicted output equation of the vehicle's nominal dynamics model is obtained based on rolling time-domain prediction: ； in, This is the multi-step prediction output vector of the vehicle's nominal dynamics model; N c To control the time domain; N p For prediction in the time domain; η(k+1), η(k+2), η(k+N) c ) and η(k+N p The numbers ) represent the nominal dynamics models of the vehicle at k+1, k+2, and k+N, respectively. c and k+N p The augmented output vector at time step; , These are the corresponding prediction coefficient matrices; ; , , and These are the 1st, 2nd, and Nth digits in the prediction time domain, respectively. c and N p Moment and The product; This represents the predicted state vector of the vehicle's nominal dynamics model at time k. This is the predicted increment vector of the vehicle's nominal dynamics model at time k; ;Δu(k), Δu(k+1), Δu(k+N c The numbers ) represent the nominal dynamics models of the vehicle at k+1, k+2, and k+N, respectively. c Control increment vector at any given time; based on The objective optimization function J for constructing the nominal dynamics model of the whole vehicle nom : ； in, Use the output vector as a reference. The prediction error vector; the expanded output weight matrix. , For N p An identity matrix of order 1. The Kronecker product is used, and Q is the single-step output weight matrix; the extended control weight matrix is ​​used. R is the single-step control weight matrix; the superscript T indicates matrix transpose; Solving the objective function J under the physical boundary constraints of the actuator nom This yields the prediction increment vector that satisfies the optimal objective. , will This is denoted as the reference feedforward control increment.

2. The method for personalized stability control of a distributed drive-by-wire chassis according to claim 1, characterized in that, The three-zone division of vehicle operating status is based on the vehicle's lateral acceleration a. y Determination of stability based on two indicators: a y <0.4g and E<0.3, is the safe zone; 0.4g≤a y <0.85g and 0.3≤E <E type E is the prevention zone. type ≤E represents the danger zone; where g is the acceleration due to gravity.

3. The method for personalized stability control of a distributed drive-by-wire chassis according to claim 1, characterized in that, The process for obtaining the disturbance feedback compensation amount for each wheel is as follows: Construct a full-vehicle dynamics model that includes external road surface and lateral disturbances: ； in, , and These are the original state vector, original control vector, and original output vector of the actual vehicle dynamics model at time k, respectively. This represents the original state vector of the actual vehicle dynamics model at time k+1. This represents the disturbance input vector of the actual vehicle dynamics model at time k; Error state vector between the nominal dynamics model and the actual dynamics model of the vehicle Establish an error dynamics model: ； in, The first derivative of e(k); For the error dynamics model, an error optimization objective function J is constructed. real : ； Where e(N) is the terminal error state vector, that is, the error state vector at time N in the control time domain; The optimal feedback gain matrix K is solved recursively using the discrete algebraic Riccati equation, and the following recursive formula is defined: ； Where K(Nk) is the optimal feedback gain matrix K at time Nk in the control time domain; B(Nk) is the original input matrix at time Nk in the control time domain; P(k-1) is the solution to the Riccati equation at time k-1 in the control time domain; R R (Nk) is the extended control weight matrix at time Nk in the control time domain; A(Nk) is the original state matrix at time Nk in the control time domain; the superscript -1 indicates the inverse matrix; The solution P(k) of the Riccati equation at time k in the control time domain is calculated as follows: ； Among them, Q Q (Nk) is the extended output weight matrix at time Nk in the control time domain; The optimal feedback gain matrix K at each time point in the control time domain is obtained by recursive calculation using the recursive formula. Based on the obtained feedback gain matrix K, design a discrete LQR feedback control law: ； The disturbance feedback compensation increment at the current moment is calculated. .

4. A distributed wire-controlled chassis personalized stability control method according to claim 1 or 3, characterized in that, The physical boundary constraints of the actuator are specifically represented as follows: 。 5. The method for personalized stability control of a distributed drive-by-wire chassis according to claim 1, characterized in that, Identify driving styles and obtain the corresponding stability energy threshold E. type The specific steps are as follows: Collect driving feature samples consisting of the maximum steering wheel angle, maximum steering wheel speed, vehicle speed, yaw rate, lateral acceleration, longitudinal acceleration, and lateral load transfer rate within a preset time period; The K-means++ clustering algorithm was used to cluster driving feature samples, and driving styles were divided into three categories: cautious, moderate, and aggressive driving styles. Pre-calibrate the stability energy threshold E corresponding to each driving style type Initial weight k of differential braking type And the E of cautious drivers type >E for general-type drivers type >E of aggressive drivers type ; After the vehicle is powered on, it continuously collects real-time driving characteristics and updates the driving style recognition results based on these characteristics. It also updates the stability energy threshold E corresponding to the current vehicle in real time based on these recognition results. type With differential braking initial weight k type .

6. The method for personalized stability control of a distributed drive-by-wire chassis according to claim 5, characterized in that, Before clustering the driving feature samples, factor dimensionality reduction is performed on the driving feature parameter matrix: the driving feature parameter matrix is ​​standardized to eliminate the influence of dimensions, resulting in a standardized matrix; the correlation coefficient matrix of the standardized matrix is ​​calculated and the eigenvalues ​​and corresponding eigenvectors are solved; principal factors are extracted according to the rule that the eigenvalue is greater than 1, and factor rotation is performed using the maximum variance method. The rotated factor loading matrix is ​​multiplied by the standardized matrix to obtain the score of each driving feature sample under each factor; the top three factors with a cumulative variance contribution rate greater than 85% are selected as the clustering input vector.

7. A distributed drive-by-wire chassis personalized stability control system, characterized in that, A distributed drive-by-wire chassis personalized stability control method according to any one of claims 1 to 6 includes a data acquisition module, a driving style identification module, a stability energy solution module, a model predictive control feedforward module, an LQR disturbance feedback correction module, and an actuator drive module. The data acquisition module is used to collect raw vehicle operation data in real time; the driving style identification module completes feature dimensionality reduction, clustering and classification, and outputs the matching stability energy threshold and differential braking initial weight; the stability energy solution module uses the two-stage dynamic roll axis method to calculate the normalized vehicle stability energy; the model predictive control feedforward module is used to solve the baseline feedforward control increment; the LQR disturbance feedback correction module is used to solve the disturbance feedback compensation increment; the actuator drive module receives the superimposed complete control increment and controls the four-wheel additional steering angle, wheel-side drive motor, and active suspension to complete the chassis stability coordinated control action.

Images