Traction force control method and system for cross-country road condition

Abstract

The invention relates to the technical field of vehicle traction control systems, in particular to a traction control method and system for cross-country road conditions, and the method comprises the steps: carrying out the preprocessing of a vehicle signal, predicting a wheel speed difference based on a hybrid model, generating a slip early-warning flag bit, a severity level and a control index through a naive Bayes classifier, and carrying out the calculation of the slip early-warning flag bit, the severity level and the control index. And the driving force or the braking force is adjusted in real time in combination with self-adaptive sliding mode control and neural network disturbance compensation. According to the invention, the wheel speed difference trend is predicted through the hybrid model to pre-judge the slip risk in advance, the naive Bayesian classifier is used to independently generate the early warning flag bit, the severity level and the control index to realize accurate decision, and the slip is inhibited in real time in combination with adaptive sliding mode control and neural network disturbance compensation. And based on the wheel speed deviation and the driver somatosensory dynamic self-learning optimization parameters, the manual calibration workload is remarkably reduced, and the control precision and response efficiency of the off-road working condition are comprehensively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vehicle traction control technology, in particular to a traction control method and system for off-road conditions. BACKGROUND

[0002] The existing traction control method for off-road conditions usually triggers control based on the real-time wheel speed difference threshold. When the deviation between the actual wheel speed of the driving wheel and the target wheel speed exceeds the preset threshold, intervention is made by adjusting the driving force or braking force. This method has significant shortcomings: first, there is a problem of control intervention lag, which cannot timely suppress the slipping phenomenon; second, different vehicle models need to be repeatedly calibrated, resulting in low efficiency; finally, the existing technology has poor adaptability to complex off-road conditions and insufficient control accuracy.

[0003] Therefore, there is an urgent need to provide a technical solution to solve the above problems. SUMMARY

[0004] To solve the above technical problems, the present application provides a traction control method and system for off-road conditions.

[0005] In a first aspect, the present application provides a traction control method for off-road conditions, and the technical solution of the method is as follows:

[0006] The vehicle input signal is subjected to validity verification and low-pass filtering to obtain a preprocessed signal;

[0007] Based on the preprocessed signal, the wheel speed difference of the driving wheel is predicted by a hybrid model combining extended Kalman filtering and convolutional neural network to obtain a wheel speed difference prediction value;

[0008] The wheel speed difference prediction value is input into a classification algorithm for processing to generate a slip warning flag, a slip severity level and a slip control index;

[0009] Based on the slip control index, the driving force or braking force is adjusted by a strategy combining adaptive sliding mode control and neural network disturbance compensation.

[0010] The traction control method for off-road conditions of the present application has the following beneficial effects:

[0011] The method of the present application predicts the wheel speed difference trend by a hybrid model to predict the risk of slipping in advance, uses a Naive Bayes classifier to independently generate a warning flag, a severity level and a control index to achieve accurate decision-making, combines adaptive sliding mode control and neural network disturbance compensation to suppress slipping in real time, and optimizes parameters based on wheel speed deviation and driver's dynamic sense to significantly reduce the workload of manual calibration and improve the control accuracy and response efficiency of off-road conditions.

[0012] On the basis of the above scheme, the off-road road condition traction control method of the present application can be further improved as follows.

[0013] In an optional manner, the step of verifying the validity of the vehicle input signal and performing low-pass filtering processing to obtain the preprocessed signal comprises:

[0014] Verify the validity flag of the vehicle input signal and the system function status to obtain a signal that passes verification;

[0015] Perform low-pass filtering processing on the verified signal to obtain a preprocessed signal.

[0016] In the above optional manner, the input signal reliability is ensured through a double verification mechanism, the validity flag is verified first to exclude invalid data interference, the system function is confirmed to be fault-free to avoid hardware error influence, and finally the high-frequency noise is eliminated through low-pass filtering, which significantly improves the signal quality, provides a stable input basis for the subsequent prediction module, and fundamentally reduces the risk of misjudgment.

[0017] In an optional manner, based on the preprocessed signal, the step of predicting the drive wheel speed difference through a hybrid model of fused extended Kalman filter and convolutional neural network comprises:

[0018] Process the preprocessed signal through extended Kalman filter to output a state estimation vector containing the complete state of the vehicle;

[0019] Extract the wheel speed difference component from the state estimation vector as a preliminary estimate, and process the preprocessed signal and the state estimation vector through a convolutional neural network to output a residual compensation value;

[0020] Add the preliminary wheel speed difference estimate and the residual compensation value to obtain the wheel speed difference prediction value.

[0021] In the above optional manner, the advantages of physical models and data-driven methods are innovatively fused, the extended Kalman filter is used to output a preliminary wheel speed difference estimate based on vehicle dynamics, and the convolutional neural network is used to learn the residual characteristics of sensor data and state estimation, and the two are superimposed to form a more accurate prediction result, effectively solving the problem of insufficient prediction accuracy of traditional single models in complex off-road conditions, and realizing slip prediction in a short time.

[0022] In an optional manner, the state estimation vector comprises:

[0023] Vehicle body longitudinal speed, transverse speed and yaw angular velocity;

[0024] Roll angle, pitch angle and their respective angular velocities;

[0025] Four-wheel angular velocity, suspension displacement and tire longitudinal force.

[0026] In the above optional manner, the key physical quantities such as vehicle body movement, posture change and tire force interaction are comprehensively covered, the longitudinal / lateral velocity, yaw rate and four-wheel dynamic parameters are included in the unified state space, a high-dimensional dynamic characterization system is established, a complete physical constraint basis is provided for the extended Kalman filter, the state estimation robustness of the model under extreme road conditions is significantly enhanced, and the prediction deviation caused by dimension loss is avoided.

[0027] In an optional manner, the step of outputting the residual compensation value by processing the preprocessed signal and the state estimation vector through the convolutional neural network comprises:

[0028] extracting the suspension height, acceleration and attitude angle from the preprocessed signal as original features;

[0029] extracting the longitudinal velocity, lateral velocity, roll angle and pitch angle from the state estimation vector as state features;

[0030] extracting the observation residual from the innovation vector of the extended Kalman filter as innovation features;

[0031] generating derived features based on the original features and the state features, the derived features including load transfer amount, average slip rate and road adhesion estimation;

[0032] outputting the residual compensation value based on the original features, the state features, the innovation features and the derived features.

[0033] In the above optional manner, a multi-source feature fusion strategy is adopted, the spatiotemporal features such as suspension height are extracted from the original signal, the estimated features such as velocity and posture are obtained from the state vector, and the innovation features (observation residual) of the extended Kalman filter are innovatively introduced, combined with the derived high-level features such as load transfer amount, to construct a multi-dimensional input tensor, so that the convolutional neural network can accurately capture the nonlinear residual of tire-road interaction and reduce the error of wheel speed difference prediction.

[0034] In an optional manner, the step of inputting the wheel speed difference prediction value into the classification algorithm to generate the slip warning flag, the slip severity level and the slip control index comprises:

[0035] inputting the wheel speed difference prediction value into a first naive Bayes classifier to generate a slip warning flag;

[0036] inputting the wheel speed difference prediction value into a second naive Bayes classifier to generate a slip severity level;

[0037] inputting the wheel speed difference prediction value into a third naive Bayes classifier to generate a slip control index.

[0038] In the above optional manner, the slip speed difference prediction value is processed in parallel by three independent naive Bayes classifiers, and the early warning flag, severity level and control index are output respectively, realizing multi-dimensional accurate quantification of the slip state, avoiding the problem of multi-task interference of a single classifier, and improving the decision reliability in rare off-road conditions by using the strong adaptability of the Bayesian probability model to few-sample abnormal working conditions.

[0039] In an optional manner, based on the slip control index, the steps of adjusting the driving force or braking force by fusing adaptive sliding mode control and neural network disturbance compensation strategy include:

[0040] Determining the target control quantity according to the slip control index, and calculating the equivalent control quantity based on the target control quantity;

[0041] Processing the wheel speed error by adaptive sliding mode control to generate an anti-disturbance quantity dynamically adjusted according to the error;

[0042] Estimating the unmodeled disturbance by a radial basis neural network to generate a disturbance compensation quantity;

[0043] Fusing the equivalent control quantity, the anti-disturbance quantity and the disturbance compensation quantity to generate a final control instruction to real-time inhibit the driving wheel slip and maintain the vehicle traction stability.

[0044] In the above optional manner, the target control quantity is dynamically generated based on the control index, the wheel speed error is real-time compensated by adaptive sliding mode control and the anti-disturbance quantity gain is automatically adjusted, and the radial basis neural network has online learning ability for unmodeled disturbances (such as road sudden changes), forming a double anti-disturbance mechanism, and finally fusing the output control instruction, so that the vehicle maintains the best traction stability on low adhesion road surfaces such as snow and mud, and the response speed of slip suppression is improved.

[0045] In an optional manner, it further includes:

[0046] The extended Kalman filter parameters in the hybrid model, the convolutional neural network weight parameters, the naive Bayes parameters in the classification algorithm, and the adaptive sliding mode gain parameters and neural network weight parameters in the control strategy are stored in the model database;

[0047] Real-time monitoring the deviation of the actual wheel speed from the target wheel speed generated based on the target control quantity;

[0048] If the deviation does not exceed the preset threshold and the driver operation data meets the somatosensory requirements, the parameters in the model database are maintained unchanged; otherwise, the parameters in the model database are adaptively updated according to the deviation data and the driver operation data.

[0049] In the optional mode, a double evaluation system of "objective deviation + subjective body sensation" is established, and the parameter updating is started only when the wheel speed deviation exceeds the threshold or the driver's operation feedback does not meet the body sensation requirement, so as to avoid system oscillation caused by excessive learning, and to be able to optimize according to the actual driving style and road condition characteristics, and reduce the workload of manual calibration.

[0050] In an optional mode, further comprising:

[0051] The updated parameters are fed back to the hybrid prediction model, the classification algorithm and the control strategy.

[0052] The wheel speed difference prediction, classification control and driving force adjustment are re-executed according to the updated parameters until the actual wheel speed difference is stabilized in the target range.

[0053] In the optional mode, a closed-loop feedback link of parameter updating-model re-execution is constructed, and the optimized parameters are injected into the prediction, classification and control modules in real time, and the iteration is executed until the wheel speed difference is stabilized in the target range, forming a self-evolving control closed loop, which is especially suitable for seasonal changes of off-road conditions (such as mud in rainy season / sand in dry season), and the control accuracy is continuously improved in long-term use.

[0054] In a second aspect, the present application provides a traction control system for off-road conditions, and the technical scheme of the system is as follows:

[0055] The traction control system for off-road conditions comprises a signal processing module, a vehicle state detection module, a main control module and a model database module.

[0056] The signal processing module is used for:

[0057] The validity flag of the vehicle input signal and the IBCU system function status are verified, and the signals that pass the verification are subjected to low-pass filtering processing, and the preprocessed signals are output to the vehicle state detection module.

[0058] The vehicle state detection module comprises a vehicle slip and loss of control prediction module and a data statistical classification module.

[0059] The vehicle slip and loss of control prediction module adopts an EKF-CNN hybrid model, which is used for:

[0060] The state estimation vector is output based on the vehicle dynamics model, and the sensor data and state characteristics are processed to output the residual compensation value.

[0061] The data statistical classification module adopts three independent naive Bayes classifiers, which are used for:

[0062] The wheel speed difference prediction value is input respectively, and the slip warning flag, the severity level and the control index are output.

[0063] The main control module is used for determining a target control amount according to a slip control index, and executing an ASMC-RBF fusion control law to output an adjustment instruction of driving force and braking force in real time.

[0064] The model database module is used for storing all key parameters, including: EKF covariance matrix parameters, CNN weights, Bayesian distribution parameters, ASMC gains and RBF network weights.

[0065] The wheel speed deviation is monitored and compared with a preset deviation threshold, and when the wheel speed deviation is greater than the preset threshold, parameter updating is triggered to realize parameter self-learning updating.

[0066] The updated parameters are fed back to the signal processing module, the vehicle state detection module and the main control module to form a closed loop.

[0067] The above description is only a summary of the technical solutions of the present application, in order to more clearly understand the technical means of the present application, the contents of the specification can be implemented, and in order to make the above and other purposes, characteristics and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0068] The accompanying drawings are only used to illustrate the embodiments and are not considered as limitations of the present application. Moreover, the same reference signs are used to represent the same parts throughout the drawings. In the drawings:

[0069] Figure 1 It is a flowchart of the off-road traction control method of the present application;

[0070] Figure 2 It is a schematic diagram of the input signal preprocessing process of the present application;

[0071] Figure 3 It is a schematic diagram of the convolutional neural network architecture of the present application;

[0072] Figure 4 It is a schematic diagram of the process from sensor data processing to final output;

[0073] Figure 5 It is a structural schematic diagram of the off-road traction control system of the present application;

[0074] Figure 6 It is a flowchart of the vehicle state detection module of the system of the present application;

[0075] Figure 7 It is a flowchart of the vehicle slip loss control prediction module of the system of the present application. DETAILED DESCRIPTION

[0076] Exemplary embodiments of the present application will be described in greater detail below with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it is understood that the present application can be embodied in various forms and should not be limited by the embodiments set forth herein.

[0077] Figure 1 A flowchart of an embodiment of a method for controlling traction on off-road terrain according to the present application is shown in Figure 1 as shown, comprising the following steps:

[0078] S1, validity check and low-pass filtering of vehicle input signals are performed to obtain preprocessed signals, in S1:

[0079] 1) The validity check refers to verifying the validity flag of the input signal and the system function state; wherein the validity flag is derived from the data validity identification provided by other message nodes of the vehicle or the internal modules of the integrated brake control unit (IBCU), and the function state refers to the absence of software or hardware failure of the IBCU; the role is to ensure signal reliability, prevent invalid data from entering subsequent calculation, and avoid misjudgment or control failure.

[0080] 2) The low-pass filtering process refers to applying a low-pass filter to smooth the input signal, and the filter coefficient is calibrated according to actual needs; the role is to eliminate high-frequency noise interference (such as sensor fluctuations), smooth the signal curve, and reduce model calculation error.

[0081] Specifically, first check whether the input signal is valid and whether the IBCU function is normal and fault-free. Only when both signal validity and function validity are checked, is the function module considered to have the prerequisite for running. Subsequently, low-pass filtering is performed on the verified signal. The main purpose of filtering is to eliminate high-frequency noise components in the signal, making the signal smooth, and providing more stable and reliable input data for the subsequent vehicle state prediction module. The filtering method can use a low-pass filter, and the filter coefficient can be calibrated and adjusted according to the vehicle type, sensor characteristics, and actual control effect.

[0082] S2, based on the preprocessed signal, the wheel speed difference is predicted by a hybrid model that fuses extended Kalman filter and convolutional neural network, to obtain a wheel speed difference prediction value, in S2:

[0083] 1) Extended Kalman filter (EKF) refers to a recursive estimation algorithm based on a nonlinear vehicle dynamics model, used to fuse multi-source sensor data, estimate vehicle state (such as speed, attitude) in real time, output preliminary estimate of wheel speed difference, and provide physical model-based prediction.

[0084] 2) Convolutional Neural Network (CNN) refers to a deep learning model used to process spatial relationships between features; the role is to learn and compensate for residual errors of EKF (such as unmodeled dynamics), output dynamic compensation values, and improve prediction accuracy.

[0085] 3) State estimation vector refers to the vector output by the extended Kalman filter, containing the key dynamic states of the vehicle. It at least contains the vehicle body longitudinal speed, transverse speed, yaw rate, roll angle, pitch angle and its angular velocity, four-wheel angular velocity, suspension displacement, and tire longitudinal force, etc. The role is to provide a preliminary estimation of the vehicle state based on physical model and sensor fusion.

[0086] 4) Residual compensation value refers to the error value between the preliminary estimation value of the wheel speed difference output by the extended Kalman filter and the true wheel speed difference value predicted by the convolutional neural network. The role is to compensate for the estimation deviation caused by unmodeled dynamics, parameter uncertainty, etc. in the EKF model, and improve the final prediction accuracy.

[0087] Specifically, the preprocessed signals (i.e. filtered suspension height and rate of change, lateral and longitudinal acceleration, pitch angle, roll angle, vehicle speed, wheel speed difference, etc.) after S1 processing are input into a hybrid prediction model that combines extended Kalman filter and convolutional neural network. First, the EKF module fuses multi-sensor data based on a simplified vehicle dynamics model (including vehicle body dynamics, suspension dynamics and tire dynamics), and outputs a state estimation vector containing the complete state of the vehicle, and extracts the preliminary estimation value of the drive wheel speed difference (such as front axle wheel speed difference, rear axle wheel speed difference) from it. At the same time, the CNN compensator module receives preprocessed signals, EKF state estimation vector, innovation vector (observation residual), and derived features (such as front axle load transfer, average slip rate, road adhesion estimation, etc.) generated based on these features as input. After training, the CNN outputs the residual prediction value between the EKF wheel speed difference estimation and the true value. Finally, the preliminary estimation value of the wheel speed difference output by the EKF is added to the residual compensation value predicted by the CNN to obtain the final, more accurate drive wheel speed difference prediction value.

[0088] S3, input the wheel speed difference prediction value into the classification algorithm for processing to generate slip warning flag, slip severity level and slip control index, in S3:

[0089] 1) Classification algorithm refers to a Naive Bayes classifier, which is used to classify the wheel speed difference prediction value based on a probability model and output discrete control indicators.

[0090] 2) Slip warning flag refers to a binary discrete output variable (for example: 0 means no slip, 1 means slip risk). The role is to serve as the basis for determining whether to trigger the warning system.

[0091] 3) Slip severity level: refers to a discrete multi-level output variable (e.g.: 1 = mild, 2 = moderate, 3 = severe). The role is to preliminarily determine the intervention intensity of the subsequent control strategy according to the predicted slip severity.

[0092] 4) Slip control index: refers to a discrete or continuous (processed by discretization) output variable, indicating the level of control intervention. The role is to directly map to specific control actions (such as driving force reduction percentage or brake force application amount) for precise vehicle control.

[0093] Specifically, the wheel speed difference prediction value obtained in the S2 step is input as the only input feature into the classification processing module based on the Naive Bayes algorithm, which includes three independent Naive Bayes classifiers, each responsible for outputting different control decision variables: the first classifier calculates the posterior probability based on the wheel speed difference prediction value, outputs a binary discrete variable, i.e. a slip warning flag (e.g. W = 0 no slip, W = 1 slip), for warning the driver; the second classifier calculates the posterior probability based on the wheel speed difference prediction value, outputs a discrete slip severity level (e.g. S = 1 mild, S = 2 moderate, S = 3 severe), for evaluating the severity of slip; the third classifier also calculates the posterior probability based on the wheel speed difference prediction value, outputs a discrete slip control index (e.g. C = 1, 2, 3..., corresponding to different driving force reduction ratios or brake force application levels), which directly quantifies the intensity of control intervention. The parameters of these classifiers (including: the prior probability of each category, the Gaussian distribution mean and variance of the wheel speed difference under each category) are pre-trained by the labeled data set and stored in the model database.

[0094] S4, based on the slip control index, adjust the driving force or brake force by fusing the adaptive sliding mode control and neural network disturbance compensation strategy in S4:

[0095] 1) Adaptive sliding mode control (ASMC) refers to a robust control algorithm that dynamically adjusts the gain to suppress disturbances, used to handle wheel speed error uncertainty and ensure control stability.

[0096] 2) Neural network disturbance compensation refers to a radial basis function neural network (RBF NN) used to estimate unmodeled disturbances, which compensates for unknown disturbances such as wind resistance or road unevenness, improving control accuracy.

[0097] 3) Equivalent control amount refers to the control amount calculated based on the nominal dynamics model of the vehicle, used to compensate for the known dynamic characteristics of the system.

[0098] 4) Anti-disturbance amount refers to the control component generated by adaptive sliding mode control, whose gain is dynamically self-adaptive to wheel speed error, used to resist system uncertainty, parameter variation and bounded disturbance.

[0099] 5) Disturbance compensation is the unmodeled dynamics and unknown disturbance (such as wind resistance, road unevenness) estimated by the radial basis neural network in real time, which is used to compensate for these difficult-to-accurately-model disturbances online.

[0100] Specifically, according to the slip control index (C) output by the S3 step, the corresponding target driving torque reduction or target braking torque (i.e., target control quantity) is queried or mapped from the model database. Then, a control strategy that combines adaptive sliding mode control and radial basis neural network disturbance compensation is used to accurately generate the final control instruction (u). First, based on the nominal vehicle driving wheel dynamics model and the current target wheel speed, the equivalent control quantity is calculated. Then, an integral sliding mode surface is designed, and an adaptive sliding mode control with a saturation function and an adaptive gain is applied to generate an anti-disturbance quantity, and a RBF neural network is constructed to process the input features (such as wheel speed error e, error derivative slip rate λ) in its hidden layer Gaussian radial basis function, online adjust the weights, and output the disturbance compensation quantity for unmodeled disturbances. Finally, the equivalent control quantity, the adaptive sliding mode anti-disturbance quantity, and the neural network disturbance compensation quantity are fused to generate the final control instruction for the vehicle driving system (adjust the driving torque T d ) or the braking system (adjust the braking torque T b ), so as to real-time inhibit the driving wheel slip and maintain the traction stability of the vehicle on the off-road road conditions.

[0101] The technical scheme of the embodiment predicts the wheel speed difference trend in advance to predict the slip risk, independently generates early warning flag, severity level and control index using the Naive Bayes classifier to realize accurate decision, real-time inhibits slip by combining adaptive sliding mode control and neural network disturbance compensation, and optimizes parameters based on wheel speed deviation and driver's dynamic feeling, which significantly reduces the manual calibration workload and improves the control accuracy and response efficiency on off-road conditions.

[0102] In an optional manner, the step of performing validity verification and low-pass filtering processing on the vehicle input signal to obtain a preprocessed signal comprises:

[0103] verifying the validity flag and the system function status of the vehicle input signal to obtain a verified signal;

[0104] performing low-pass filtering processing on the verified signal to obtain a preprocessed signal.

[0105] In the embodiment, first, signal and function validity check is performed, the purpose is to ensure that the data input to the traction control function module is reliable, and at the same time ensure that the system function is normal, to provide a reliable basis for subsequent calculation. Further:

[0106] Input signal acquisition, receiving input signals provided by other message nodes of the vehicle or other functional modules inside the integrated brake control unit. Vehicle input signals include: suspension height signal and rate of change, lateral and longitudinal acceleration, pitch angle, roll angle, vehicle speed, wheel speed difference.

[0107] Signal validity check, checking the validity flag bit attached to each of the above input signals. The flag bit is provided by the signal source (such as a sensor or an upstream calculation module) to indicate whether the signal is currently valid. Only signals with a "valid" validity flag bit are considered usable.

[0108] Functionality check, checking the functional status flag bit of the IBCU unit itself. The flag bit indicates whether the software and hardware of the IBCU are working properly and whether there are any faults. Only when the functional status flag bit indicates that the IBCU is functioning properly is the system considered to meet the operating conditions.

[0109] Processing of the check results, only when the validity flag bit of all input signals to be processed is "valid" and the functional status flag bit of the IBCU indicates "functioning properly", the signal and function validity check is considered "verified". Only after "verification", the subsequent signal filtering processing and control function calculation are allowed. Otherwise, if any input signal is invalid or the IBCU function is abnormal, the functional module will not be enabled to avoid making errors based on unreliable data.

[0110] As shown in Figure 2 After the signal and function validity checks, low-pass filtering is performed on the verified signals to eliminate high-frequency noise components in the input signals, smooth the signal curve, reduce signal fluctuations that interfere with subsequent model prediction and control algorithms, and improve system robustness. Further:

[0111] The filtering method uses a low-pass filter to process the verified original signals. The low-pass filter allows low-frequency signal components to pass through, while attenuating or suppressing high-frequency signal components. For each input signal that passes the validity check, a low-pass filtering algorithm is applied. The filter coefficient involved is a tunable parameter. Its specific value needs to be optimized and calibrated according to the actual control effect, sensor characteristics, signal noise characteristics, and specific vehicle or project requirements. For example, for signals with high noise (such as some suspension height sensor signals), a small alpha value (stronger filtering) is needed; for signals that require fast response (such as wheel speed), a large alpha value (weaker filtering) is needed.

[0112] The values of the input signals after low-pass filtering are the final preprocessed signals. These preprocessed signals will be output for further analysis and prediction.

[0113] In an alternative way, the step of predicting the wheel speed difference based on the preprocessed signals by a hybrid model of extended Kalman filter and convolutional neural network to obtain the wheel speed difference prediction value, comprises:

[0114] processing the preprocessed signals by the extended Kalman filter to output a state estimation vector containing the complete state of the vehicle;

[0115] extracting the wheel speed difference component from the state estimation vector as a preliminary estimation value, and processing the preprocessed signals and the state estimation vector by the convolutional neural network to output a residual compensation value;

[0116] adding the preliminary estimation value of the wheel speed difference and the residual compensation value to obtain the wheel speed difference prediction value.

[0117] In an alternative way, the state estimation vector comprises:

[0118] the longitudinal velocity, lateral velocity and yaw rate of the vehicle body;

[0119] the roll angle, pitch angle and the corresponding angular velocity of each;

[0120] the angular velocity of the four wheels, the suspension displacement and the longitudinal force of the tires.

[0121] In an alternative way, the step of processing the preprocessed signals and the state estimation vector by the convolutional neural network to output the residual compensation value, comprises:

[0122] extracting the suspension height, acceleration and attitude angle from the preprocessed signals as original features;

[0123] extracting the longitudinal velocity, lateral velocity, roll angle and pitch angle from the state estimation vector as state features;

[0124] extracting the observation residual from the innovation vector of the extended Kalman filter as innovation features;

[0125] generating derived features based on the original features and the state features, the derived features comprising: load transfer amount, average slip ratio and road adhesion estimation;

[0126] outputting the residual compensation value based on the original features, the state features, the innovation features and the derived features.

[0127] In this embodiment, first, the preprocessed signals are processed by the extended Kalman filter to output a state estimation vector containing the complete state of the vehicle.

[0128] The state vector x of the EKF contains key variables representing the complete dynamic state of the vehicle, and the state vector is defined as:

[0129]

[0130] where v x represents the vehicle body longitudinal velocity (m / s); v y represents the vehicle body lateral velocity (m / s); ω z represents the yaw rate (rad / s); φ represents the roll angle (rad); θ represents the pitch angle (rad); represents the roll rate (rad / s); represents the pitch rate (rad / s); ω fl represents the left front wheel angular velocity (rad / s); ω fr represents the right front wheel angular velocity (rad / s); ω rl represents the left rear wheel angular velocity (rad / s); ω rr represents the right rear wheel angular velocity (rad / s); z fl represents the left front suspension displacement (m); z fr represents the right front suspension displacement (m); z rl represents the left rear suspension displacement (m); z rr represents the right rear suspension displacement (m); F x,fl represents the left front tire longitudinal force (N); F x,fr represents the right front tire longitudinal force (N); F x,rl represents the left rear tire longitudinal force (N); F x,rr represents the right rear tire longitudinal force (N). The state vector integrates the vehicle body motion, wheel motion, suspension state, and tire force information.

[0131] The EKF performs state prediction based on a simplified nonlinear state transition model expressed as: x k+1 = f(x k , u k ) + w k ;

[0132] where x k+1 represents the state vector at time k+1; f(x) is the nonlinear state transition function; x k represents the state vector at time k; u k is the control input vector; w k is the process noise vector.

[0133] The key sub-models of the nonlinear state transition model include:

[0134] 1) Vehicle body dynamics model, expressed as:

[0135]

[0136] where m is the vehicle mass; I zIz for the moment of inertia around Z axis; F x,ij Ft for the tire longitudinal force; F y,ij Ft for the tire lateral force; l x,ij l for the longitudinal distance from the tire to the center of mass (positive for front and negative for rear); l y,ij l for the lateral distance from the tire to the center of mass (positive for left and negative for right).

[0137] 2) Suspension dynamics (linear) model, expressed as follows:

[0138]

[0139] where m s,ij is the unsprung mass; k s,ij is the suspension stiffness coefficient; c s,ij is the suspension damping coefficient; is the static tire vertical force; is the dynamic load transfer.

[0140] 3) Tire dynamics model, expressed as follows:

[0141] F x,ij = D x sin(C x arctan(B x kappa ij - E x (B x kappa ij - arctan(B x kappa ij )))) ;

[0142] where B x , C x , D x , E x are the Pacejka magic formula parameters; kappa ij is the tire slip ratio; slip ratio R is the tire effective radius; v x,ij is the tire center longitudinal velocity.

[0143] The EKF updates the state estimation using the vehicle sensor measurements. In the observation update stage of the extended Kalman filter, the core role of the observation model is to transform the original measurements of the vehicle sensors into the correction basis of the state estimation. The expression of the observation model is as follows: z k = h(x k ) + v k ;

[0144] where z k is the observation vector at time k; h(x) is the nonlinear observation function; v k is the observation noise vector.

[0145] The observation vector z includes:

[0146]

[0147] An example of the observation equation is:

[0148]

[0149] where a x is the longitudinal acceleration (m / s 2 ); a y is the lateral acceleration (m / s 2 ); and g is the gravitational acceleration (m / s 2 ).

[0150] The EKF recursive process includes a prediction step and an update step, further:

[0151] The prediction step is:

[0152]

[0153] where is the prior state estimate; P k|k-1 is the prior estimate covariance matrix; F k-1 is the state transition Jacobian matrix; P k-1|k1 is the posterior estimate covariance matrix; Q k is the process noise covariance matrix.

[0154] The update step is:

[0155]

[0156] P k|k = (I - K k H k ) P k|k-1

[0157] where v k is the innovation (observation residual); S k is the innovation covariance matrix; K k is the observation Jacobian matrix; R k is the observation noise covariance matrix; K k is the Kalman gain matrix; is the posterior state estimate; P k|k is the posterior estimate covariance matrix. Jacobian matrix:

[0158] Secondly, from the state estimate vector x output from the EKF, the left and right drive wheel angular velocity preliminary estimate values are extracted, expressed as:

[0159] Δω f,ekf =ω fl -ω fr ;

[0160] Δω r,ekf =ω rl -ω rr ;

[0161] Where, Δω f,ekf This represents the front axle wheel speed difference (left front - right front); Δω r,ekf This represents the rear axle wheel speed difference (left rear - right rear).

[0162] Target residual ∈ k This is the error between the initial estimate of the EKF wheel speed difference and the actual wheel speed difference. The residual is defined as: ∈ k =Δω true,k -Δω ekf,k ; where, ∈ k Estimate the residuals for EKF; Δω true,k Δω represents the actual wheel speed difference at time k. ekf,k The wheel speed difference estimated by EKF.

[0163] Feature engineering is performed, which involves constructing the CNN input feature tensor, including:

[0164] Raw features (sensor / preprocessed signal): Extract 8-dimensional features from the preprocessed signal: suspension height, lateral and longitudinal acceleration, pitch angle, and roll angle features, i.e., [a x ,a y ,φ,θ,z fl ,z fr ,z rl ,z rr ].

[0165] State features (EKF estimation): Five-dimensional features are extracted from the EKF state estimation vector: longitudinal velocity, lateral velocity, yaw rate, roll angle, and pitch angle.

[0166] Innovation feature (EKF observation residual): The innovation vector ν from the EKF update step k Extract the residual features related to the observations (with the same dimension as the observation vector, e.g., 14 dimensions), i.e.

[0167] Derived features (calculated based on the physical model): Additional features are calculated based on the original features and state features. The formula for derived features is as follows:

[0168] Among them, the average slip ratio κ ijrepresents the tire slip ratio estimated by the state; is the front axle load transfer amount (N); m is the mass of the vehicle (kg); h cg is the height of the center of mass; L is the wheelbase; t f is the front wheel track; K lat is the lateral load transfer coefficient.

[0169] All features (original 8 + state 5 + innovation 14 + derived 3 ~ total feature dimension D = 29) are organized in a time sequence into a fixed time window (e.g. the most recent T time steps) feature tensor X e R T×29 The generated T x 29-dimensional feature tensor will be used as input, delivered to the CNN architecture shown in FIG. 2 to predict the residual compensation value. A ring buffer is used for real-time updating. The update rule is: Figure 3

[0170] X k = [f k-T+1 , f k-T+2 ,..., f k ] T ;

[0171] where X k is the current time feature tensor; f k is the k-time feature vector; T is the time window length; f t = [sensors t , states t , v t , derived t ] represents the t-time feature vector.

[0172] For the architecture and processing of CNN, a standard CNN architecture containing convolutional layers, pooling layers and fully connected layers is adopted. The convolutional layer formula is:

[0173] where, is the output of the l-th layer at time step t; K is the convolution kernel size; D in is the number of input channels; is the convolution kernel weight; is the input feature map; b (l) is the bias term.

[0174] The output layer expression is:

[0175] where h is the fully connected layer output feature vector; W out is the output layer weight matrix; b out is the output layer bias vector.

[0176] The CNN is trained offline on the historical data set, and the loss function is:​

[0177]

[0178] wherein, is the loss function; N is the number of training samples; is the CNN predicted residual; λ is the L2 regularization coefficient; ||W||2 is the L2 norm of all weight parameters.

[0179] The trained weight parameters are stored in the model database for online use. After the EKF module design and the CNN residual compensator design, the online prediction process is entered, the final prediction value is calculated, the wheel speed difference preliminary estimation value output by the EKF is added to the residual compensation value predicted by the CNN, and the final, more accurate driving wheel speed difference prediction value is obtained, that is,

[0180]

[0181] wherein, is the hybrid predicted wheel speed difference value (rad / s); CNN(X k ) is the CNN predicted residual compensation value.

[0182] As shown in Figure 4 , the whole process from sensor data processing to final output is shown: the sensor collects the original measurement value to input the extended Kalman filter module to perform prediction and update, generate a state estimation vector and trigger the buffer to build a T x 29-dimensional feature tensor; the tensor is input into the convolutional neural network for forward propagation, and the residual compensation value is output to the output module, which is superimposed with the wheel speed difference preliminary estimation of the EKF to generate a high-precision prediction value, which is finally output to the decision system by the terminal module. The process is realized by parallel data flow design to achieve millisecond-level real-time prediction.

[0183] In an optional manner, the wheel speed difference prediction value is input into a classification algorithm to generate a slip warning flag, a slip severity level and a slip control index, comprising:

[0184] inputting the wheel speed difference prediction value into a first naive Bayes classifier to generate a slip warning flag;

[0185] inputting the wheel speed difference prediction value into a second naive Bayes classifier to generate a slip severity level;

[0186] inputting the wheel speed difference prediction value into a third naive Bayes classifier to generate a slip control index.

[0187] In this embodiment, the definitions of the main variables include:

[0188] 1) Wheel speed difference Δw (denoted as D), which represents the absolute value of the speed difference between the left and right drive wheels of the vehicle, in km / h or m / s. It reflects the degree of wheel slip, and the larger the difference, the higher the risk of slipping. It is a continuous variable and is used as an input feature.

[0189] 2) Slip warning flag (W), which is a binary flag indicating whether the vehicle is in a slipping state. It is used to trigger a warning system (such as a dashboard alert). For example, W = 0 indicates no slip (normal state); W = 1 indicates slip (warning needed). It is a discrete output (binary classification).

[0190] 3) Slip severity level (S), which represents a discrete level indicating the severity of the slip, affecting the degree of control strategy aggressiveness. For example, it is divided into three levels: S = 1 (mild slip), S = 2 (moderate slip), and S = 3 (severe slip). It is a discrete output (multi-classification).

[0191] 4) Slip control index (C), which is used to control the vehicle driving force (such as motor torque) and braking force (such as brake pressure). The larger the value, the stronger the control intervention (e.g., reducing driving force, increasing braking force). For example, it is discretized into three levels: C = 1 (mild intervention, reducing 10% torque), C = 2 (moderate intervention, reducing 30% torque), and C = 3 (severe intervention, reducing 50% torque), etc. In a continuous control system, discrete levels can be mapped to specific control values (such as C corresponding to torque reduction percentage). It is a discrete output (multi-classification), and Naive Bayes is a classification algorithm, so it needs to be discretized. If the original data is continuous, it can be processed by binning.

[0192] The principle of Naive Bayes algorithm is based on Bayes' theorem, and the posterior probability is calculated as:

[0193]

[0194] where Y is the output category (such as W, S, C); X is the input feature (here it is Δw, i.e. D); P(Y) is the prior probability, estimated from training data; P(X) is the evidence term, which is a normalization constant, calculated as P(X) = ∑ Y P(X|Y)P(Y); P(X|Y) is the likelihood probability, which assumes that the conditional distribution of D given Y. Assuming that the likelihood probability follows a Gaussian distribution (normal distribution):

[0195]

[0196] where Y is the output category (W, S, C); μ y and are the sample mean and variance of the category, respectively.

[0197] When making predictions, select the category with the highest posterior probability, i.e.:

[0198] Since the output has three variables (W, S, C), three independent Naive Bayes classifiers are trained (one for each output), that is, three independent Naive Bayes classifiers are used to output the following discrete decision variables:

[0199] First classifier: Generates a slip warning flag W∈{0,1};

[0200] Second classifier: Generates the severity level of slippage S∈{1,2,3} (1=mild, 2=moderate, 3=severe);

[0201] Third classifier: Generate slip control index C∈{1,2,3} (for example, C=1 corresponds to a 10% reduction in torque, C=2 corresponds to a 30% reduction, and C=3 corresponds to a 50% reduction).

[0202] The parameters of each classifier are derived from historical datasets. Obtained through offline training and stored in a model database, where d i Let w be the value of Δw. i ,s i ,c i The label is for annotation. Further:

[0203] 1) Regarding the estimation of parameters for the warning flag W:

[0204] Prior probability: P(W=0) = 1 - P(W=1);

[0205] Likelihood parameter (Gaussian distribution): Calculate the mean μ of Δw for class W=1. W＝1 and variance

[0206]

[0207] Where N1 is the number of samples where W = 1. Furthermore, μ for the class where W = 0 is calculated in the same way. W＝0 and

[0208] 2) Regarding the parameters for estimating the severity level S, assume three levels:

[0209] Prior probability:

[0210] Likelihood parameter (Gaussian distribution): Where, N k Let S be the number of samples for S = k.

[0211] 3) In estimating the parameters of the control index C, assume discretization into 3 levels:

[0212] Prior probability:

[0213] Likelihood parameter (Gaussian distribution, similar to S):

[0214]

[0215] where N c is the number of samples for C = c.

[0216] In the online prediction procedure, three outputs are calculated, namely given a new observation of the drive wheel speed difference value d (a concrete value), calculate W, S, C. Further:

[0217] 1) Calculate the slip warning flag W:

[0218] Calculate the likelihood:

[0219]

[0220] Calculate the posterior (ignore the denominator P(D = d), because it is unchanged when comparing):

[0221] P(W = 1 | D = d) ∝ P(D = d | W = 1)P(W = 1);

[0222] P(W = 0 | D = d) ∝ P(D = d | W = 0)P(W = 0);

[0223] Decision:

[0224]

[0225] Or use a threshold, such as P(W = 1 | D = d) > 0.5.

[0226] 2) Calculate the slip severity level S:

[0227] Calculate the posterior probability of each level k:

[0228] P(S = k | D = d) ∝ P(D = d | S = k)P(S = k), k ∈ {1, 2, 3};

[0229] where,

[0230] Decision:

[0231] 3) Calculate the slip control index C:

[0232] Calculate the posterior probability of each level c:

[0233] P(C = c | D = d) ∝ P(D = d | C = c)P(C = c), c e {1, 2,..., N};

[0234] wherein,

[0235] Decision:

[0236] Output can be directly used for control. For example, reduce the driving force by 10%; reduce the driving force by 50% and apply braking, and the like. In addition, the key coefficients involved in the detection algorithm are stored in the model database.

[0237] In an optional manner, based on the slip control index, the step of adjusting the driving force or the braking force by fusing the adaptive sliding mode control and the neural network disturbance compensation strategy comprises:

[0238] determining a target control quantity according to the slip control index, and calculating an equivalent control quantity based on the target control quantity;

[0239] processing the wheel speed error through adaptive sliding mode control to generate an anti-disturbance quantity dynamically adjusted according to the error;

[0240] estimating the unmodeled disturbance through a radial basis neural network to generate a disturbance compensation quantity;

[0241] fusing the equivalent control quantity, the anti-disturbance quantity and the disturbance compensation quantity to generate a final control instruction, so as to inhibit the driving wheel slip and maintain the vehicle traction stability in real time.

[0242] It should be noted that the vehicle slip warning sign, the slip severity level and the slip control index determine whether the driving force or the braking force of the vehicle needs to be controlled. The slip warning sign and the slip severity level determine whether the driving force or the braking force of the driving wheel of the vehicle needs to be controlled; the slip control index is an input condition for precise control of the driving force or the braking force. Each slip control index corresponds to a set of target driving force and target braking force of the driving wheel, and such data is stored in the model database. If the vehicle slip out-of-control condition under the current working condition is inhibited and the driving feeling requirement (corresponding to the driving wheel speed difference) is met, it is considered that the slip control index corresponds to the target driving force and the target braking force meeting the requirements; otherwise, self-learning adjustment is performed to meet the driving requirements (corresponding to the driving wheel speed difference Aw and the slip control index ). Therefore, the control scheme of fusing adaptive sliding mode control (ASMC) and neural network (NN) disturbance compensation can be realized.

[0243] In this embodiment, the dynamic equation of the vehicle drive wheels is:

[0244] Where J is the wheel's moment of inertia (kg·m) 2 ); ω is the actual wheel speed (rad / s); T d The driving torque (N·m, control input); T b R is the braking torque (N·m, control input); R is the wheel radius (m); F x (λ) represents the longitudinal force between the tire and the road surface (a nonlinear function, N); d(t) represents the unknown disturbance (wind resistance, road surface inequality); λ represents the tire slip ratio. (v is the vehicle speed).

[0245] The control objective is: wheel speed error e = ω ref -ω; e→0 and |e|<δ (preset error threshold). Where, ω ref ω represents the target wheel speed (rad / s); ω represents the actual wheel speed (rad / s).

[0246] For the design of Adaptive Sliding Membrane Control (ASMC):

[0247] 1) Define the sliding surface: Where k is the integral gain (>0), which determines the error convergence rate; s is the sliding mode variable, and the error exponential convergence occurs when s = 0.

[0248] 2) Control Law Design: Traditional Sliding Mode Control: u smc =u eq +u sw ,u sw =K·sgn(s); where, u eq The equivalent control quantity (compensating for known dynamics); u sw K is for switching control (suppressing disturbances); K is for fixed gain (causing chattering).

[0249] Adaptive Improvement (Eliminating Fixed Gain K): Disturbance Immunity in, φ is the adaptive gain (estimates the upper bound of the perturbation); sat(.) is the saturation function (replaces sgn(.) to reduce chattering); φ is the boundary layer thickness (>0).

[0250] Adaptive law: Where η is the adaptive rate, dynamically adjusted. To cover the disturbance.

[0251] For neural network (NN) perturbation compensation design:

[0252] 1) RBF neural network structure:

[0253] Input layer: (error, error derivative, slip ratio).

[0254] Hidden layer: Gaussian Radial Basis Function (RBF):

[0255] where c j is the center vector (preset); σ j is the width parameter (preset).

[0256] Output layer: disturbance compensation

[0257] where W = [w1, w2,..., w N ] T is the weight vector (online update).

[0258] 2) Weight adaptive law: where γ is the learning rate (> 0); κ is the attenuation coefficient (to prevent weight drift).

[0259] For the design of the fusion control law:

[0260] Final control command: where u eq is the equivalent control based on the nominal model; u smc is the adaptive sliding mode term, i.e., the disturbance compensation; is the N disturbance compensation term, i.e., the disturbance compensation.

[0261] Equivalent control: where, is the nominal tire force model (known part)

[0262] Final control law:

[0263] Based on Lyapunov function to prove the stability of the closed-loop system, ensuring that

[0264] where, represents the error between the ideal weight and the actual weight; represents the error between the ideal gain and the actual gain.

[0265] To further explain the specific process, after obtaining the slip control index ( ), the driving force or braking force is adjusted in real time through the following steps:

[0266] 1) Target control quantity determination and equivalent control calculation:

[0267] Firstly, according to the slip control index, the preset mapping relationship is queried from the model database to determine the specific target control amount (such as target driving torque reduction amount ΔT d or target braking torque T b ). Based on this target control amount, the expected target wheel speed (ω ref ) is set.

[0268] Subsequently, the equivalent control amount (u eq ) is calculated based on the dynamics equation. This calculation integrates the following elements:

[0269] the error between the target wheel speed and the actual wheel speed (wheel speed error); the system moment of inertia (wheel moment of inertia J); the tire-road interaction force nominal tire force model ( ). Among them, the role of the equivalent control amount is to compensate for the known vehicle dynamics characteristics, providing a reference for subsequent disturbance rejection control.

[0270] 2) Adaptive sliding mode disturbance rejection amount generation:

[0271] The wheel speed error (e) is processed through the adaptive sliding mode control mechanism to generate a dynamically adjusted disturbance rejection amount (u smc ), including:

[0272] Define the sliding mode surface as a measure of the system state deviation;

[0273] Use adaptive gain to estimate the disturbance upper bound in real time, and the update rate is dynamically adjusted by the adaptive law;

[0274] Use the saturation function (sat(s / φ)) to replace the traditional sign function to suppress control chattering.

[0275] The core function of the disturbance rejection amount is to resist parameter perturbation and time-varying disturbance, ensuring system robustness.

[0276] 3) Neural network disturbance compensation amount generation:

[0277] Estimate the unmodeled disturbance through the radial basis neural network to generate the disturbance compensation amount ( ), including:

[0278] The input layer receives the feature vector composed of the wheel speed error, error derivative, and tire slip rate.

[0279] The hidden layer performs nonlinear mapping through the Gaussian radial basis function;

[0280] The output layer weights are updated online according to the adaptive law.

[0281] The disturbance compensation amount is specifically used to offset the unmodeled dynamics (such as sudden changes in road adhesion)

[0282] 4) Fusion of control command generation and execution:

[0283] The outputs of the first three steps, i.e., the equivalent control amount (u eq ), the anti-disturbance amount (u smc ), and the disturbance compensation amount (u ), are fused to generate the final control command (u ). The command directly acts on the vehicle actuators, which are used to:

[0284] adjust the drive motor output torque (drive torque T d );

[0285] or, adjust the brake system pressure (brake torque T b ).

[0286] This fusion strategy can achieve real-time suppression of drive wheel slip and maintenance of traction stability under the guarantee of Lyapunov theory.

[0287] In an optional way, it also includes:

[0288] storing the extended Kalman filter parameters in the hybrid model, the convolutional neural network weight parameters, the naive Bayes parameters in the classification algorithm, and the adaptive sliding mode gain parameters and neural network weight parameters in the control strategy to the model database;

[0289] real-time monitoring of the deviation of the actual wheel speed from the target wheel speed generated based on the target control amount;

[0290] if the deviation does not exceed the preset threshold and the driver operation data meets the somatosensory requirements, the parameters in the model database are maintained unchanged; otherwise, the parameters in the model database are updated adaptively according to the deviation data and the driver operation data.

[0291] In an optional way, it also includes:

[0292] feedback the updated parameters to the hybrid prediction model, the classification algorithm, and the control strategy;

[0293] re-executing the wheel speed difference prediction, the classification control, and the drive force adjustment according to the updated parameters until the actual wheel speed difference is stabilized within the target range.

[0294] In this embodiment, first, centralized storage is performed, and the extended Kalman filter (EKF) parameters in the hybrid prediction model, the convolutional neural network weight parameters, the naive Bayes parameters in the classification algorithm, and the adaptive sliding mode gain and RBF neural network weight in the control strategy are stored to the model database.

[0295] The actual wheel speed is compared with the target wheel speed generated based on the target control amount in real time, and the wheel speed deviation is calculated: Δω = |ω - ω refIf Δω is less than or equal to the preset deviation threshold and the driver's operation data (such as the rate of change of accelerator pedal opening ≤ 5% / s) meets the sensory requirements, the parameters are kept unchanged; otherwise, the key parameters in the database are adaptively updated using the gradient descent method based on the deviation data Δω and the driver's operation data (such as rapid acceleration / braking signals).

[0296] To improve system adaptability, the updated parameters are synchronized in real time to the hybrid prediction model (updating the EKF covariance matrix and CNN weights), classification algorithm (correcting the Naive Bayes probability distribution), and control strategy (adjusting the sliding mode gain and RBF network weights). The entire process (wheel speed difference prediction, classification control, and driving force adjustment) is then re-executed based on the new parameters until the actual wheel speed difference stabilizes within the target range (e.g., ±0.5 rad / s).

[0297] Figure 5 A schematic diagram of an embodiment of a traction control system for off-road conditions provided by the present invention is shown. Figure 5 As shown, the system includes: a signal processing module, a vehicle state detection module, a main control module, and a model database module;

[0298] The signal processing module is used to: verify the validity flag of the vehicle input signal and the functional status of the IBCU system, and perform low-pass filtering on the verified signal to output the pre-processed signal to the vehicle status detection module.

[0299] like Figure 6 As shown, the vehicle status detection module includes: a vehicle skidding and loss of control prediction module and a data statistics and classification module;

[0300] The vehicle skidding and loss of control prediction module uses an EKF-CNN hybrid model, such as Figure 7 As shown, the prediction module sequentially executes EKF design (constructing a state estimation vector based on vehicle dynamics), CNN residual compensator design (processing sensor and state features to generate compensation values), and online prediction process (superimposing state estimates and compensation values ​​to output wheel speed difference prediction), forming a closed-loop prediction chain from dynamic modeling to data-driven correction, providing accurate input for the subsequent data statistics and classification module.

[0301] The vehicle skidding and loss of control prediction module is used to: output a state estimation vector based on the vehicle dynamics model, process sensor data and state characteristics, and output residual compensation values;

[0302] The data statistics and classification module employs three independent Naive Bayes classifiers for:

[0303] Input the predicted wheel speed difference value, and output the slippage warning flag, severity level, and control index respectively;

[0304] The main control module is configured to determine a target control amount according to a slip control index, and execute an ASMC-RBF fusion control law to output an adjustment instruction of driving force and braking force in real time.

[0305] The model database module is configured to store all key parameters, including an EKF covariance matrix parameter, a CNN weight, a Bayesian distribution parameter, an ASMC gain, and an RBF network weight.

[0306] The wheel speed deviation is monitored and compared with a preset deviation threshold, and when the wheel speed deviation is greater than the preset threshold, parameter updating is triggered to realize parameter self-learning updating.

[0307] The updated parameters are fed back to the signal processing module, the vehicle state detection module, and the main control module to form a closed loop.

[0308] The technical solution of the embodiment predicts the wheel speed difference trend in advance to predict the slip risk by using a hybrid model, independently generates a warning flag, a severity level, and a control index by using a Naive Bayes classifier to realize accurate decision-making, combines adaptive sliding mode control and neural network disturbance compensation to inhibit slip in real time, and optimizes parameters based on wheel speed deviation and driver's dynamic sense to significantly reduce the workload of manual calibration and comprehensively improve the control accuracy and response efficiency of off-road conditions.

[0309] The steps of the above-mentioned parameters and modules in the off-road traction control system of the embodiment for realizing corresponding functions can refer to the parameters and steps in the off-road traction control method of the embodiment described above, and will not be repeated here.

[0310] Although the embodiments of the present application have been shown and described above, it should be understood that the above-mentioned embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the present application.

Claims

1. A method of traction control for off-road conditions, characterized in that, The method comprises the following steps: validity check and low-pass filtering processing are performed on the vehicle input signal to obtain a pretreatment signal; based on the pretreatment signal, a hybrid model of extended Kalman filter and convolutional neural network is used to predict the drive wheel speed difference to obtain a wheel speed difference prediction value; the wheel speed difference prediction value is input into a classification algorithm for processing to generate a slip warning flag, a slip severity level and a slip control index; based on the slip control index, the driving force or braking force is adjusted through a strategy of fusing adaptive sliding mode control and neural network disturbance compensation.

2. The off-road traction control method according to claim 1, characterized in that, The step of performing validity check and low-pass filtering processing on the vehicle input signal to obtain a pretreatment signal comprises: validity flag and system function status of the vehicle input signal are verified to obtain a verified signal; low-pass filtering processing is performed on the verified signal to obtain a pretreatment signal.

3. The off-road traction control method according to claim 1, characterized in that, The step of predicting the drive wheel speed difference based on the pretreatment signal by fusing the hybrid model of extended Kalman filter and convolutional neural network to obtain a wheel speed difference prediction value comprises: the pretreatment signal is processed by the extended Kalman filter to output a state estimation vector containing the complete state of the vehicle; the wheel speed difference component is extracted from the state estimation vector as a preliminary estimation value, and the pretreatment signal and the state estimation vector are processed by the convolutional neural network to output a residual compensation value; the wheel speed difference preliminary estimation value and the residual compensation value are added to obtain a wheel speed difference prediction value.

4. The off-road traction control method according to claim 3, characterized in that, The state estimation vector comprises: vehicle body longitudinal speed, transverse speed and yaw angular speed; roll angle, pitch angle and their corresponding angular speeds; four-wheel angular speed, suspension displacement and tire longitudinal force.

5. The off-road traction control method according to claim 4, characterized in that, The step of processing the pretreatment signal and the state estimation vector by the convolutional neural network to output a residual compensation value comprises: suspension height, acceleration and attitude angle are extracted from the pretreatment signal as original features; longitudinal speed, transverse speed, roll angle and pitch angle are extracted from the state estimation vector as state features; observation residual is extracted from the innovation vector of the extended Kalman filter as innovation feature; derived features are generated based on the original features and the state features, including: load transfer amount, average slip rate and road adhesion estimation; the residual compensation value is output based on the original features, the state features, the innovation feature and the derived features.

6. The off-road traction control method according to claim 5, characterized in that, The step of inputting the wheel speed difference prediction value into a classification algorithm for processing to generate a slip warning flag, a slip severity level and a slip control index comprises: the wheel speed difference prediction value is input into a first naive Bayes classifier to generate a slip warning flag; the wheel speed difference prediction value is input into a second naive Bayes classifier to generate a slip severity level; the wheel speed difference prediction value is input into a third naive Bayes classifier to generate a slip control index.

7. The off-road traction control method according to claim 6, characterized in that, The step of adjusting the driving force or braking force based on the slip control index by fusing adaptive sliding mode control and neural network disturbance compensation strategy comprises: the target control amount is determined according to the slip control index, and the equivalent control amount is calculated based on the target control amount; the anti-disturbance amount dynamically adjusted according to the error is generated by processing the wheel speed error through adaptive sliding mode control; the disturbance compensation amount is generated by estimating the unmodeled disturbance through radial basis neural network. The equivalent control amount, the anti-disturbance amount and the disturbance compensation amount are fused to generate a final control instruction to suppress the drive wheel slip and maintain the vehicle traction stability in real time.

8. The off-road traction control method of claim 1, wherein, Further comprising: storing the extended Kalman filter parameters in the hybrid model, the convolutional neural network weight parameters, the naive Bayes parameters in the classification algorithm, and the adaptive sliding mode gain parameters and neural network weight parameters in the control strategy to the model database; monitoring the deviation between the actual wheel speed and the target wheel speed generated based on the target control amount in real time; if the deviation does not exceed the preset threshold and the driver operation data meets the somatosensory requirements, the parameters in the model database are maintained unchanged; otherwise, the parameters in the model database are adaptively updated according to the deviation data and the driver operation data.

9. The off-road traction control method of claim 8, wherein, Further comprising: feeding the updated parameters back to the hybrid prediction model, the classification algorithm and the control strategy; re-executing the wheel speed difference prediction, the classification control and the driving force adjustment according to the updated parameters until the actual wheel speed difference is stabilized within the target range.

10. An off-road traction control system characterized by, The traction control method for off-road conditions according to any one of claims 1-9.

Images