A method of estimating a main dynamic parameter of a vehicle

Abstract

The application is a kind of estimation method for main dynamic parameters of vehicle, comprising the following steps: step one: establishing vehicle longitudinal dynamics model and wheel dynamics model, jointly estimating vehicle state and parameters through unscented Kalman filtering algorithm, obtaining vehicle mass center height, mass center to vehicle front axle distance and vehicle mass; step two: establishing vehicle single-track yaw dynamics model, jointly estimating vehicle state and parameters through unscented Kalman filtering algorithm combined with the estimation results of vehicle mass center height, mass center to vehicle front axle distance and vehicle mass obtained by vehicle longitudinal dynamics model, obtaining vehicle yaw axis moment of inertia; step three: obtaining longitudinal velocity, longitudinal acceleration, lateral acceleration, yaw angular velocity and roll angular velocity, and training and testing the identification of roll moment of inertia through multi-layer feedforward neural network; step four: estimating vehicle roll moment of inertia through the trained neural network observer.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vehicle dynamics control, in particular to an estimation method of main vehicle dynamics parameters. BACKGROUND

[0002] With the development of automotive electronics technology, more and more vehicles are equipped with advanced auxiliary driving systems, such as anti-lock braking system, electronic stability system and active front steering, and the automation level of vehicles is increasingly developing towards unmanned driving. However, the reliability of unmanned driving system and stability control not only depends on the robustness and accuracy of control algorithm, but also depends on the accurate acquisition of vehicle state parameters. The changes of vehicle mass, center of mass height, distance from center of mass to front axle, and roll and yaw moment of inertia will directly affect the driving state of the vehicle. In the state equation of the vehicle dynamics model, the mass, center of mass height, distance from center of mass to front axle, and roll and yaw moment of inertia are important control parameters. Whether these parameters can be quickly and accurately obtained during vehicle driving and whether the control parameters can be corrected online have a significant impact on the effect of vehicle motion control.

[0003] Generally speaking, the traditional way to obtain vehicle dynamics parameters is to calibrate and measure through a specific test bench. Although this method can accurately and comprehensively obtain vehicle dynamics parameters, it also has obvious disadvantages. Firstly, this measurement method must be carried out on a specific test bench, which is not common and has high testing cost. Secondly, this method cannot update and correct vehicle parameters in real time, and when the vehicle parameters change significantly, the original parameters may no longer be applicable.

[0004] Currently, the most common method for obtaining vehicle parameters is online estimation using estimation algorithms. This involves establishing a vehicle dynamics model and collecting signals from existing onboard sensors, then using existing estimation algorithms such as least squares, Kalman filtering, and particle filtering to estimate the vehicle's dynamic parameters. However, existing methods for estimating vehicle dynamic parameters often decouple the coupling between vehicle motion and dynamic parameters to simplify the problem, focusing primarily on estimating and measuring only a subset of these parameters, such as vehicle mass and center of gravity position, or only vehicle mass and yaw moment of inertia. This approach struggles to effectively estimate systems with numerous parameters and strong nonlinearity, and it also considers a limited number of vehicle states and parameters. While various improved Kalman filtering algorithms are used to estimate multiple vehicle dynamic parameters, these methods are not comprehensive enough in terms of parameter and state estimation. Furthermore, estimating the vehicle's yaw moment of inertia using a fully model-based Kalman filtering algorithm requires obtaining the yaw angle state from sensors, but sensors capable of accurately measuring the yaw angle are prohibitively expensive. Summary of the Invention

[0005] This invention provides a method for estimating key dynamic parameters of a vehicle, the method comprising the following steps:

[0006] Step 1: Establish the vehicle's longitudinal dynamics model and wheel dynamics model, collect the vehicle's wheel speed, lateral and longitudinal acceleration and deceleration, yaw rate and longitudinal velocity signals, and use the unscented Kalman filter algorithm to jointly estimate the vehicle's state and parameters to obtain the vehicle's center of gravity height, distance from the center of gravity to the front axle and vehicle mass.

[0007] Step 2: Establish a vehicle monorail yaw dynamics model. During the vehicle's turning motion, collect the vehicle's wheel speed, lateral and longitudinal acceleration and deceleration, yaw rate, roll rate, and longitudinal velocity signals. Combine the vehicle's center of gravity height, distance from the center of gravity to the front axle, and vehicle mass estimation results obtained from the vehicle's longitudinal dynamics model using the unscented Kalman filter algorithm to jointly estimate the vehicle's state and parameters, and obtain the vehicle's yaw moment of inertia.

[0008] Step 3: Acquire data on longitudinal velocity, longitudinal acceleration, lateral acceleration, yaw rate, and roll rate, preprocess them, and train and test a multi-layer feedforward neural network to identify the roll inertia.

[0009] Step 4: Estimate the vehicle's roll inertia using the trained neural network observer.

[0010] Furthermore, in step 1, the vehicle longitudinal dynamics model and the wheel dynamics model are specifically as follows:

[0011]

[0012]

[0013]

[0014] where the total mass of the vehicle is denoted by ; the lumped rotational inertia of the front left and right wheels; the lumped rotational inertia of the rear left and right wheels; the effective radius of the tire is denoted by ; and are the front and rear driving torques of the lumped left and right wheels, respectively, and are the front and rear braking torques of the lumped left and right wheels, respectively; and denote the longitudinal forces of the front and rear tires, respectively, denotes the vehicle longitudinal velocity, and are the rotational angular velocities of the front and rear wheels, respectively, and the superscript · denotes differentiation with respect to the indicated quantity.

[0015] Further, in step 1, the Kalman filter estimates the state quantities , the parameter quantities , the input quantities , and the observation quantities ;

[0016] The state equation of the unscented Kalman filter process is :

[0017]

[0018]

[0019]

[0020] The parameter equation is :

[0021]

[0022]

[0023]

[0024] where denotes the vehicle longitudinal velocity, denotes the rotational angular velocity of the front wheels, denotes the rotational angular velocity of the rear wheels, denotes the vehicle mass, Indicates the height of the center of mass. This represents the distance from the center of gravity to the front axle. Indicates the driving torque of the front wheels. Indicates the driving torque of the rear wheels. Indicates the braking torque of the front wheels. Indicates the braking torque of the rear wheels. Indicates the angular velocity of the front wheel. The symbol indicates the angular velocity of the rear wheel, and the upper symbol · indicates the derivative of the quantity shown.

[0025] The final corrections for the state variables and parameters to be estimated are as follows:

[0026]

[0027]

[0028] Among them, Measured by the sensor, at this point, we can obtain the result. The state variables to be estimated for the vehicle at each time step , representing the vehicle's longitudinal velocity, the angular velocities of the front and rear wheels, and the parameters to be estimated, respectively. These are estimates of the vehicle's mass, center of gravity height, and distance from the center of gravity to the front axle, respectively.

[0029] Furthermore, in step 2, the yaw dynamics model of the vehicle is specifically as follows:

[0030]

[0031] Among them, subscript and These represent the front wheels and the rear wheels, respectively. and These represent the distances from the center of mass to the front and rear axes, respectively. The yaw rotational inertia of the vehicle; where Indicates the longitudinal force on the front wheel. and The lateral forces of the front and rear tires are represented respectively, ω represents the yaw rate of the vehicle, and the upper symbol · indicates the derivative of the indicated quantity.

[0032] Furthermore, in step 2, the state variable to be estimated by Kalman filtering... The number of parameters to be estimated Observation Select the vehicle yaw rate measured by the IMU;

[0033] The state equation for the unscented Kalman filtering process is:

[0034]

[0035] The parameter equation is:

[0036]

[0037] The final corrections of the state and parameter to be estimated are as follows:

[0038]

[0039]

[0040] wherein, is measured by the sensor, at this time, the state to be estimated of the vehicle at is obtained represents the lateral acceleration, represents the estimated value of the lateral angular velocity of the vehicle,

[0041] Further, in step 3, the longitudinal velocity, the lateral angular velocity, and the roll angular velocity are obtained by establishing a longitudinal-lateral-yaw-roll four-degree-of-freedom vehicle model similar to the actual vehicle, and the longitudinal-lateral-yaw-roll four-degree-of-freedom vehicle model is:

[0042]

[0043]

[0044]

[0045]

[0046] wherein, is the sprung mass, represents the lateral acceleration, represents the roll angle, is the height from the mass center to the roll rotation center, is the distance from the mass center to the front axle, is the lateral angular velocity, is the longitudinal force of the four wheels, is the lateral force of the four wheels, is the roll stiffness of the vehicle, is the roll damping of the vehicle, is the wheel track of the vehicle;

[0047] The longitudinal acceleration and the lateral acceleration are obtained by the following formula:

[0048]

[0049]

[0050] wherein, represents the longitudinal acceleration, represents the lateral acceleration.

[0051] Further, in step 3, the multi-layer feedforward neural network comprises an input layer, a hidden layer layer, a hidden layer layer, a hidden layer layer and an output layer;

[0052] the input vector of the input layer wherein, , , , , ; the output layer Y = [y], .

[0053] Further, in step 3, the multi-layer feedforward neural network weight adjustment process is:

[0054] the error of the output layer is:

[0055]

[0056] wherein is the output error of the i-th iteration, is the expected value of the i-th iteration, is the output value of the i-th iteration; In the weight adjustment phase, the adjustment is carried out layer by layer in reverse along the network;

[0057] First, the weights between the hidden layer and the output layer are adjusted

[0058] , the gradient of the error with respect to is calculated , and the adjustment is carried out in the reverse direction:

[0059]

[0060]

[0061] wherein, is the learning rate;

[0062] Then, the weights between the hidden layer and the hidden layer are adjusted , the gradient of the error with respect to is calculated , and the adjustment is carried out in the reverse direction:​​​

[0063]

[0064]

[0065] Continue to adjust the hidden layer and the hidden layer between the weights , the gradient of the error on , and adjust in the opposite direction along the direction:

[0066]

[0067]

[0068] Finally adjust the input layer and the hidden layer between the weights , the gradient of the error on , and adjust in the opposite direction along the direction:

[0069]

[0070]

[0071] Detect the neural network, define the detection error:

[0072]

[0073] If the detection error is less than or equal to 5%, it can be considered that the neural network observer training is completed, otherwise more data is collected to repeat the network training until the requirements are met.

[0074] The beneficial effects achieved by the present application are:

[0075] The present application first estimates the mass, the height of the center of mass, the distance from the center of mass to the front axle and the roll moment of inertia of the vehicle step by step by the method of joint state and parameter unscented Kalman filtering. The joint estimation of the state and the parameter can correct the estimation results with each other, and more accurate vehicle parameters can be obtained in real time. Then, the method of neural network is used to replace the dependence on high-precision sensors to estimate the roll moment of inertia. The roll moment of inertia corresponding to the vehicle state information obtained under different working conditions of the vehicle with known other vehicle dynamics parameters is learned by the neural network, and an online vehicle roll moment of inertia estimator is obtained, so that the roll moment of inertia of the vehicle is estimated in real time. Finally, comprehensive main vehicle parameters are obtained in real time, so that the cost of the vehicle dynamics parameter estimation method is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0076] Figure 1 ​​is a vehicle parameter estimation flowchart of an embodiment of the present application;

[0077] Figure 2 is a vehicle longitudinal dynamics model of an embodiment of the present application;

[0078] Figure 3 is a vehicle wheel dynamics model of an embodiment of the present application;

[0079] Figure 4 is a vehicle yaw dynamics model of an embodiment of the present application;

[0080] Figure 5 is a vehicle roll dynamics model of an embodiment of the present application;

[0081] Figure 6 is a filter process flowchart of an unscented Kalman filter algorithm of an embodiment of the present application;

[0082] Figure 7 is a neural network model of an embodiment of the present application. DETAILED DESCRIPTION

[0083] The technical solutions of the present application will be described in more detail below with reference to the accompanying drawings, which include but are not limited to the following embodiments.

[0084] As shown in the accompanying Figure 1 , the present application proposes a method for estimating main dynamics parameters of a vehicle, including the following steps:

[0085] Step one: establishing a vehicle longitudinal dynamics model and a vehicle wheel dynamics model, as shown in the accompanying Figure 2 and 3 , in the process of longitudinal motion of the vehicle, the vehicle is affected by the center of mass height, the distance from the center of mass to the front axle of the vehicle, and the mass of the vehicle, by making the vehicle accelerate and decelerate longitudinally, collecting the wheel speed, vehicle lateral and longitudinal acceleration, yaw angular velocity and longitudinal velocity signals of the vehicle, then through the unscented Kalman filter algorithm, jointly estimating the vehicle state and parameters, thereby obtaining the center of mass height, the distance from the center of mass to the front axle of the vehicle, and the mass of the vehicle.

[0086] In step one, the vehicle longitudinal dynamics model and the vehicle wheel dynamics model are specifically:

[0087]

[0088]

[0089]

[0090] The total mass of the vehicle is represented by ; The rotational inertia of the aggregated left and right front wheels; the rotational inertia of the left and right rear wheels, respectively, and ; and are the front and rear driving torques of the left and right wheels, respectively, and are the front and rear braking torques of the left and right wheels, respectively, which can be directly controlled by the in-wheel motors and used as inputs in this paper. Among them, and denote the longitudinal forces of the front and rear tires, respectively, denotes the longitudinal velocity of the vehicle, and are the rotational angular velocities of the front and rear wheels, respectively, and the superscript “•” denotes the differentiation of the indicated quantity.

[0091] The calculation of the longitudinal forces of the front and rear wheels is obtained by the following formula:

[0092]

[0093] The subscripts correspond to and are the front and rear, are the left and right lumped wheel longitudinal stiffnesses, are the slip ratios, are the front and rear wheel vertical loads.

[0094] The calculation of the slip ratios of the front and rear wheels is obtained by the following formula:

[0095]

[0096] correspond to the front and rear wheel angular velocities, is the wheel effective radius.

[0097] The calculation of the vertical loads of the front and rear wheels is obtained by the following formula:

[0098]

[0099]

[0100] wherein, is the height of the vehicle center of mass, is the distance from the center of mass to the front axle, is the mass of the vehicle, is the distance from the center of mass to the rear axle, is the acceleration of gravity, is the longitudinal acceleration, is the wheelbase of the vehicle.

[0101] The Kalman filter to be estimated state quantity , denotes the vehicle longitudinal speed, the rotational angular speed of the front and rear wheels, the parameter quantities to be estimated , are the vehicle mass, the mass center height and the distance from the mass center to the front axle, respectively, the input quantities , and are the collective front and rear tractive torques of the left and right wheels, respectively, and are the collective front and rear braking torques of the left and right wheels, respectively. The observation quantities The vehicle longitudinal speed and the rotational angular speed of the front and rear wheels measured by the GPS and the ABS are selected.

[0102] The unscented Kalman filter process is shown in the attached Figure 6 where the state equation is:

[0103]

[0104]

[0105]

[0106] The parameter equation is:

[0107]

[0108]

[0109]

[0110] The state evolution equation of the unscented Kalman filter process is:

[0111]

[0112]

[0113]

[0114] where, denotes the cyclic iteration to the th time, denotes the cyclic iteration to the th time, denotes the time step.

[0115] The parameter evolution equation is:

[0116]

[0117]

[0118]

[0119] wherein, denotes the parameter quantity.

[0120] The state observation equation of the unscented Kalman filter process is:

[0121]

[0122]

[0123]

[0124] The parameter observation equation is:

[0125]

[0126]

[0127]

[0128] in the formula, , and are the vehicle longitudinal speed and the front and rear wheel rotation angular velocity measured by the sensor.

[0129] The symmetric sampling method is used to select the sigma point set and its weight value:

[0130]

[0131]

[0132]

[0133]

[0134] in the formula, denotes the symmetrically sampled sigma point, and the superscript corresponds to and denote the state quantity and the parameter quantity, respectively. The superscripts denote that the loop is iterated to the th time; is the weight of each point when calculating the mean value, is the weight of each point when calculating the variance, is the adjustment parameter, wherein ; the parameter reflects the degree of deviation of the test point from the expected value, and controls the distribution range of the sigma point. The parameter reflects the high-order moment information of the state variable, and the optimal value is when the ; N denotes the number of sigma points.

[0135] The state time update for each sigma point is:

[0136]

[0137] where the state is the vehicle longitudinal speed, the front and rear wheel rotational angular speed, and the input is , which represent the front and rear wheel drive torques and the front and rear wheel braking torques, respectively. denotes the th sigma point of the state at the th prediction.

[0138] The parameter time update for each sigma point is:

[0139]

[0140] The parameter to be estimated is the vehicle mass, the mass center height, and the mass center to front axle distance, denotes the th sigma point of the parameter at the th prediction.

[0141] The prior estimate of the state and parameter to be estimated is obtained as:

[0142]

[0143]

[0144] where the superscript denotes the mean and denotes the value at time extrapolated from time .

[0145] The prior estimate of the variance of the state and parameter to be estimated is:

[0146]

[0147]

[0148] where is the process noise, and the superscript denotes the variance.

[0149] The observation update of the state and parameter is:

[0150]

[0151]

[0152] in, , To measure noise and thus obtain prior estimates of the corresponding observations:

[0153]

[0154]

[0155] in, This represents the prior estimate of the observed quantity. Observations representing state variables and parameter variables.

[0156] Prior estimate variances of observed values ​​for state and parameter quantities:

[0157]

[0158]

[0159] The covariances of state variables and observations, and the covariances of parameters and observations are respectively:

[0160]

[0161]

[0162] Correct the state variables and parameter variables:

[0163] The Kalman gains are as follows:

[0164]

[0165]

[0166] Perform variance updates for state variables and parameter variables:

[0167]

[0168]

[0169] The final corrections for the estimated state variables and parameters are as follows:

[0170]

[0171]

[0172] in, The state variables of the vehicle at time k can be obtained by measuring the data from the sensors. respectively, are the longitudinal velocity of the vehicle, the rotational angular velocity of the front and rear wheels, and the to-be-estimated parameter quantity respectively, are the estimated values of the vehicle mass, the mass center height, and the distance from the mass center to the front axle, so as to perform estimation of other parameters subsequently.

[0173] Step 2: Establish a single-track yaw dynamics model of the vehicle, as shown in FIG. 2. Figure 4 In the process of turning motion of the vehicle, wheel speed, vehicle lateral and longitudinal acceleration, yaw angular velocity, roll angular velocity, and longitudinal vehicle speed signals of the vehicle are collected, and then the estimation results of the mass center height, the distance from the mass center to the front axle, and the vehicle mass of the vehicle are obtained through the unscented Kalman filtering algorithm combined with the vehicle longitudinal dynamics model, so as to perform joint estimation on the vehicle state and parameters, thereby obtaining the vehicle yaw moment of inertia.

[0174] In step 2, the yaw dynamics model of the vehicle is specifically as follows:

[0175]

[0176] wherein subscripts and represent the front wheel and the rear wheel respectively; and represent the distance from the mass center to the front and rear axles respectively; is the vehicle yaw moment of inertia; wherein represents the front wheel longitudinal force, and represent the lateral force of the front tire and the rear tire respectively, ω represents the vehicle yaw angular velocity, represents the front wheel rotation angle, and the superscript “•” represents the differential of the indicated quantity.

[0177] The calculation of the front and rear wheel longitudinal forces can be obtained from the formula in step 1, and the lateral force is obtained from the following formula:

[0178]

[0179] The subscript corresponds to and are the front and rear, are the left and right lumped wheel cornering stiffnesses, is the cornering angle, are the vertical loads of the front and rear wheels.

[0180] The calculation of the front and rear wheel cornering angles is obtained from the following formula:

[0181]

[0182]

[0183] corresponding to the front wheel angle, is the lateral velocity, is the longitudinal velocity, and respectively represent the distance from the mass center to the front and rear axles, and ω represents the vehicle yaw rate.

[0184] The vertical loads of the front and rear wheels can be calculated by the formula in step one.

[0185] State variables to be estimated by Kalman filter ω represents the vehicle yaw rate, and parameter variables to be estimated is the yaw moment of inertia. The observation variable The vehicle yaw rate measured by the IMU is selected.

[0186] The unscented Kalman filter process is shown in FIG. 1, wherein the state equation is: Figure 6

[0187]

[0188] The parameter equation is:

[0189]

[0190] The state evolution equation of the unscented Kalman filter process is:

[0191]

[0192] The parameter evolution equation is:

[0193]

[0194] The state observation equation of the unscented Kalman filter process is:

[0195]

[0196] The parameter observation equation is:

[0197]

[0198] In the formula, ω represents the vehicle yaw rate measured by the sensor.

[0199] The symmetric sampling method is used to select the sigma point set and the weight value thereof:

[0200]

[0201]

[0202]

[0203] ​

[0204] where, denotes the symmetrically sampled sigma points, the superscript corresponds to and denote the state and parameter quantities, respectively. The subscripts denote the th iteration of the loop; is the weight of each point in calculating the mean, is the weight of each point in calculating the variance, is the scaling parameter, where ; the parameter reflects the degree of deviation of the test point from the expected value, controlling the distribution range of the sigma points. The parameter reflects the high-order moment information of the state variable, and the optimal value is when the state variable obeys the Gaussian distribution; ; denotes the number of sigma points.

[0205] The state quantity time update has:

[0206]

[0207] where the state quantity = [ω], denotes the vehicle yaw rate, denotes the th state quantity sigma point in the th prediction.

[0208] The parameter quantity time update has:

[0209]

[0210] The parameter quantity to be estimated is the vehicle yaw moment of inertia, denotes the th parameter quantity sigma point in the th prediction.

[0211] The prior estimate of the state quantity and parameter quantity to be estimated can be obtained:

[0212]

[0213]

[0214] The prior estimate of the variance of the state quantity and parameter quantity to be estimated is:

[0215]

[0216]

[0217] wherein is the process noise.

[0218] The observation of the state quantity and the parameter quantity is updated as:

[0219]

[0220]

[0221] wherein , is the measurement noise, and the prior estimation value of the observation corresponding thereto is obtained as:

[0222]

[0223]

[0224] The prior estimation variance of the observation of the state quantity and the parameter quantity is:

[0225]

[0226]

[0227] The covariance of the state quantity and the observation and the covariance of the parameter quantity and the observation are respectively:

[0228]

[0229]

[0230] The state quantity and the parameter quantity are corrected as:

[0231] The Kalman gain is respectively:

[0232]

[0233]

[0234] The variance of the state quantity and the parameter quantity is updated as:

[0235]

[0236]

[0237] Finally, the correction of the state quantity and the parameter quantity to be estimated is as follows:

[0238]

[0239]

[0240] wherein, are measured by sensors, at this time the state of the vehicle to be estimated at time t is the vehicle yaw rate, the parameter to be estimated is the estimated value of the vehicle yaw moment of inertia, in order to estimate other parameters subsequently.

[0241] Step three: on the basis of the vehicle parameters obtained above, collect the data required for neural network training, pre-process, and design a multi-layer feedforward neural network to train and test the identification of the roll moment of inertia.

[0242] In step three, on the basis of the vehicle parameters obtained above and other vehicle parameters that are not variable, an accurate four-degree-of-freedom vehicle model similar to the actual vehicle is established, as shown in FIGS. 1 and 2. Figure 4 and 5

[0243]

[0244]

[0245]

[0246]

[0247] wherein, is the sprung mass, is the lateral acceleration, is the roll angle, the total mass estimated in step one is subtracted from the unsprung mass to obtain is the height of the center of mass to the roll rotation center, the total center of mass height estimated in step one is subtracted from the height of the roll rotation center to the ground to obtain is the distance of the center of mass to the front axle, estimated in step one, is the yaw moment of inertia, estimated in step two; is the longitudinal force of the four wheels, is the lateral force of the four wheels, and the calculation formula is shown in steps one and two above; is the roll stiffness of the vehicle, is the roll damping of the vehicle, is the wheelbase of the vehicle, only the roll moment of inertia is changed, preferably is 300-1200 ​100 every interval Under various conditions, data collection is performed at different speeds for driving scenarios such as lane changes, double lane shifts, serpentine maneuvers, fishhook maneuvers, and curves, and the optimal data is collected. For 60-80 1 every interval In this case, the collected vehicle state data includes: longitudinal velocity, longitudinal acceleration, lateral acceleration, yaw rate, and roll rate. Among them, longitudinal velocity, yaw rate, and roll rate can be directly obtained from the four-degree-of-freedom vehicle state, while longitudinal acceleration and lateral acceleration are obtained from the following formulas.

[0248]

[0249]

[0250] For each operating condition, 10-20 sets of vehicle status data were collected. Data preprocessing was performed based on longitudinal speed, longitudinal acceleration, lateral acceleration, yaw rate, and roll rate. Data that did not clearly distinguish the roll moment of inertia was removed. The optimized data was divided into training set and test set, accounting for 75% and 25% of the total, respectively.

[0251] Step three employs a five-layer BP neural network, which consists of the following parts:

[0252] The network structure is as follows: Figure 7 As shown

[0253] The entire network consists of one input layer with I neurons and three hidden layers. layer, layer, The number of neurons in each layer are respectively , , The input layer consists of an output layer with only one neuron and the first neuron in the output layer. 1 neuron is denoted as ,hide The first layer 1 neuron is denoted as ,hide The first layer 1 neuron is denoted as ,hide The first layer 1 neuron is denoted as The neurons in the output layer are denoted as y, from arrive The connection weight is ,from arrive The connection weight is , arrive The connection weight is , The connection weight to y is ,use and These represent the input and output of each layer, respectively; the input vector of the input layer. ,in , , , , The first hidden layer output vector The output vector of the second hidden layer The output vector of the third hidden layer Output layer , .

[0254] Calculations of the network's weight matrix and the weight matrix from the input layer to the hidden H1 layer of the activation function:

[0255]

[0256] in Represents the flow from the input layer to the hidden layer. The weight matrix of the layer, This represents the output of the input layer neurons. Indicates hidden Layer neuron input;

[0257] hide Layer neuron input to neuron output calculation:

[0258]

[0259] Indicates hidden Layer neuron output, For activation functions;

[0260] hide Hidden layers Layer weight matrix calculation:

[0261]

[0262] in Indicates from hidden Hidden layers The weight matrix of the layer, Indicates hidden Neuron output, Indicates hidden Layer neuron input;

[0263] hide Layer neuron input to neuron output calculation:

[0264]

[0265] Hidden layer neuron output, f(x) = RELU as activation function

[0266] Hidden layer to hidden layer weight matrix calculation

[0267] Hidden layer neuron input, h = Wx

[0268]

[0269] Hidden layer neuron input to neuron output calculation

[0270]

[0271] Hidden layer neuron output, f(x) = RELU as activation function

[0272] Hidden layer to output layer weight matrix calculation

[0273]

[0274]

[0275] Output layer neuron input to neuron output calculation

[0276]

[0277] Output layer neuron output, f(x) = RELU as activation function

[0278] Neural network learning algorithm weight adjustment formula:​​​​​​​​​​​​​​​​​​​​​

[0279] The error of the output layer is:

[0280]

[0281] where is the output error of the th iteration, is the expected value of the th iteration, is the output value of the th iteration.

[0282] In the weight adjustment phase, the adjustment is made layer by layer in the reverse direction along the network. First, the weights between the hidden layer and the output layer are adjusted . According to the steepest descent method, the gradient of the error with respect to is calculated, and the adjustment is made in the reverse direction along this direction:

[0283]

[0284]

[0285] where is the learning rate.

[0286] Then, the weights between the hidden layer and the hidden layer are adjusted . According to the steepest descent method, the gradient of the error with respect to is calculated, and the adjustment is made in the reverse direction along this direction:

[0287]

[0288]

[0289] where is the learning rate.

[0290] Finally, the weights between the hidden layer and the hidden layer are adjusted . According to the steepest descent method, the gradient of the error with respect to is calculated, and the adjustment is made in the reverse direction along this direction:

[0291]

[0292]

[0293] where is the learning rate.

[0294] Finally, the weights between the input layer and the hidden layer are adjusted . According to the steepest descent method, the gradient of the error to is calculated, and the adjustment is carried out in the opposite direction:

[0295]

[0296]

[0297] wherein is the learning rate.

[0298] The neural network is detected, and the detection error

[0299]

[0300] If the detection error is less than or equal to 5%, it can be considered that the neural network observer training is completed, otherwise more data are collected to repeat the network training until the requirement is met.

[0301] Step four: the trained neural network observer is used to estimate the vehicle roll moment of inertia.

[0302] In step four, the vehicle is turned and the observation conditions of the neural network observer are met, the vehicle roll moment of inertia is estimated, if the estimation result fluctuates greatly, the least square method can be added after estimation for optimization, and finally a relatively accurate vehicle roll moment of inertia can be obtained.

[0303] The present application is not limited to the above specific embodiments, and those skilled in the art can implement the present application in other various specific embodiments according to the content disclosed in the embodiments and the drawings, therefore, any design using the design structure and idea of the present application, and making some simple changes or alterations, falls within the protection scope of the present application.

Claims

1. A method for estimating key dynamic parameters of a vehicle, characterized in that, The method for estimating the main dynamic parameters of a vehicle includes the following steps: Step 1: Establish the vehicle's longitudinal dynamics model and wheel dynamics model, collect the vehicle's wheel speed, lateral and longitudinal acceleration and deceleration, yaw rate and longitudinal velocity signals, and use the unscented Kalman filter algorithm to jointly estimate the vehicle's state and parameters to obtain the vehicle's center of gravity height, distance from the center of gravity to the front axle and vehicle mass. Step 2: Establish a vehicle monorail yaw dynamics model. During the turning motion of the vehicle, collect the vehicle's wheel speed, lateral and longitudinal acceleration and deceleration, yaw rate, roll rate and longitudinal velocity signals. Combine the vehicle's center of gravity height, distance from the center of gravity to the front axle and the vehicle's mass obtained from the vehicle's longitudinal dynamics model with the estimation results of the unscented Kalman filter algorithm to jointly estimate the vehicle's state and parameters, and obtain the vehicle's yaw moment of inertia. Step 3: Acquire data on longitudinal velocity, longitudinal acceleration, lateral acceleration, yaw rate, and roll rate, preprocess them, and train and test a multi-layer feedforward neural network to identify the roll inertia. Step 4: Estimate the vehicle's roll inertia using the trained neural network observer; In step 3, the longitudinal velocity, yaw rate, and roll rate are obtained by establishing a four-degree-of-freedom vehicle model that closely resembles the actual vehicle, which is: ； ； ； ； in, For the sprung mass, Indicates lateral acceleration. Indicates the roll angle. The height from the center of mass to the center of tilt rotation. This is the distance from the center of gravity to the front axle. For the moment of inertia of yaw rotation, The longitudinal force on the four wheels, The lateral forces on the four wheels, For vehicle roll stiffness, For vehicle roll damping, The wheelbase of the vehicle; The longitudinal acceleration and lateral acceleration are obtained by the following formula: ； ； in, Indicates longitudinal acceleration. Indicates lateral acceleration; The multilayer feedforward neural network includes an input layer and a hidden layer. Layers, hidden layers Layers, hidden layers Layers and output layers; Input vector of the input layer ,in , , , , Output layer Y=[y], .

2. The method for estimating the main dynamic parameters of a vehicle according to claim 1, characterized in that, In step 1, the vehicle longitudinal dynamics model and the wheel dynamics model are specifically as follows: ； ； ； Among them, the total mass of the vehicle is used express; The rotational inertia of the left and right front wheels; The rotational inertia of the left and right rear wheels is lumped together; the effective radius of the tire is expressed as... ; and These are the front and rear driving torques of the lumped left and right wheels, respectively. and These are the front and rear braking torques of the lumped left and right wheels, respectively; and These represent the longitudinal forces of the front and rear tires, respectively. Indicates the longitudinal speed of the vehicle. and The rotational angular velocities of the front and rear wheels are respectively indicated by the · symbol, which represents the derivative of the quantity shown.

3. The method for estimating the main dynamic parameters of a vehicle according to claim 2, characterized in that, In step 1, the state variables to be estimated by Kalman filtering are... The number of parameters to be estimated Input quantity Observation ; State equation of the unscented Kalman filter process for: ； ； ； The parametric equation is: ； ； ； in, Indicates the longitudinal speed of the vehicle. This indicates the angular velocity of the front wheels. This indicates the angular velocity of the rear wheel. Indicates vehicle mass. Indicates the height of the center of mass. This represents the distance from the center of gravity to the front axle. Indicates the driving torque of the front wheels. Indicates the driving torque of the rear wheels. Indicates the braking torque of the front wheels. Indicates the braking torque of the rear wheels. Indicates the angular velocity of the front wheel. The symbol indicates the angular velocity of the rear wheel, and the upper symbol · indicates the derivative of the quantity shown. The final corrections for the state variables and parameters to be estimated are as follows: ； ； in, Measured by the sensor, at this point, we can obtain the result. The state variables to be estimated for the vehicle at each time step , representing the vehicle's longitudinal velocity, the angular velocities of the front and rear wheels, and the parameters to be estimated, respectively. These are estimates of the vehicle's mass, center of gravity height, and distance from the center of gravity to the front axle, respectively.

4. The method for estimating the main dynamic parameters of a vehicle according to claim 1, characterized in that, In step 2, the yaw dynamics model of the vehicle is specifically as follows: ； Among them, subscript and These represent the front wheels and the rear wheels, respectively. and These represent the distances from the center of mass to the front and rear axes, respectively. The yaw rotational inertia of the vehicle; where This indicates the longitudinal force on the front wheel. and The lateral forces of the front and rear tires are represented respectively, ω represents the yaw rate of the vehicle, and the upper symbol · indicates the derivative of the indicated quantity.

5. The method for estimating the main dynamic parameters of a vehicle according to claim 4, characterized in that, In step 2, the state variables to be estimated by Kalman filtering are... The number of parameters to be estimated Observation Select the vehicle yaw rate measured by the IMU; The state equation for the unscented Kalman filtering process is: ； The parametric equation is: ； The final corrections for the state variables and parameters to be estimated are as follows: ； ； in, Measured by the sensor, at this point, we can obtain the result. The state variables to be estimated for the vehicle at each time step , representing the vehicle's yaw rate, is a parameter to be estimated. , which represents the estimated value of the vehicle's yaw moment of inertia.

6. The method for estimating the main dynamic parameters of a vehicle according to claim 1, characterized in that, In step 3, the weight adjustment process of the multilayer feedforward neural network is as follows: The error of the output layer is: ； In the formula For the first The output error of the next iteration For the first The expected value of the next iteration. For the first The output value of the next iteration; During the weight adjustment phase, adjustments are made layer by layer in reverse along the network. First adjust the hidden settings. Weights between the layer and the output layer Calculation error gradient Then adjust in the opposite direction: ； ； in, The learning rate; Adjust the hiding Layers and Hidden Weights between layers Calculation error gradient Then adjust in the opposite direction: ； ； Continue to adjust the hiding Layers and Hidden Weights between layers Calculation error gradient Then adjust in the opposite direction: ； ； Finally, adjust the input layer and hidden layer. Weights between layers Calculation error gradient Then adjust in the opposite direction: ； ； For neural networks, the detection error is defined as follows: ； If the detection error is less than or equal to 5%, the neural network observer training can be considered complete; otherwise, more data should be collected and the network training repeated until the requirements are met.

Images