A Vehicle Mass Estimation Method Based on Adaptive Extended Kalman Filter

Abstract

The present invention relates to a vehicle mass estimation method based on adaptive extended Kalman filtering, which includes the following steps: obtaining vehicle parameters; establishing a longitudinal dynamics model regarding vehicle mass and road slope; setting a vehicle mass estimator based on a neural network model, selecting input features of the neural network according to the model in step two, using vehicle mass as the output, and training the neural network model; setting a weight regulator to calculate a weight factor; embedding the mass pre-estimation of the neural network model into adaptive Kalman filtering, and correcting the vehicle mass pre-estimation value output by the neural network model based on the weight factor. By learning the non-linear relationship between vehicle states through the neural network, the influence caused by model mismatch is reduced. A weight factor is introduced to embed the result estimated by the neural network into adaptive Kalman filtering, adjust the weights of the Kalman filtering and the result estimated by the neural network, reduce the oscillation of the estimation result, and improve the accuracy and stability of the vehicle mass pre-estimation value.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving, and more particularly, to a vehicle mass estimation method based on adaptive extended Kalman filtering. Background Art

[0002] With the progress of industrial technology, autonomous vehicles are increasingly equipped with electronic control systems. To enable these control systems to work, accurate information about vehicle parameters must be collected. Vehicle mass is one of the most important characteristics, and is particularly important for systems that rely on vehicle dynamics, such as electronic stability controllers (ESC), electronic parking brakes (EPB), and power management strategies (PMS). However, due to different actual payloads, the mass varies greatly, especially for passenger cars and trucks, where the payload variation can range from 0% to 400%. If the accurate vehicle mass can be obtained in real time, the control effect of the vehicle will be significantly improved.

[0003] For existing vehicle mass estimation systems and methods, system models are established based on vehicle dynamics and kinematics and existing sensor data, and methods such as least squares and Kalman filtering are used to estimate the mass. However, due to changes in driving scenarios, the non-linearity of vehicle dynamics, and noise pollution in sensor data, the matching degree between the model-based estimator and the real vehicle decreases, and the accuracy drops. Summary of the Invention

[0004] To overcome the problem of low accuracy in vehicle mass estimation of the model-based method in the above-mentioned prior art, the present invention provides a vehicle mass estimation method based on adaptive extended Kalman filtering. By learning the non-linear relationship between vehicle states, a vehicle mass estimator based on a neural network is established to provide an accurate mass pre-estimation, and then the estimation result is stabilized by introducing adaptive Kalman filtering.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: a vehicle mass estimation method based on adaptive extended Kalman filtering, including the following steps:

[0006] Step 1: Obtain vehicle parameters;

[0007] Step 2: Establish a longitudinal dynamics model of vehicle mass and road slope according to the parameters in Step 1;

[0008] Step 3: Set a vehicle mass estimator based on a neural network model. According to the longitudinal dynamics model in Step 2, select parameters such as vehicle speed and acceleration as input features A t , and vehicle mass as the output, and train the neural network model;

[0009] Step 4: Set a weight regulator and calculate the weight factor;

[0010] Step 5: Embed the quality pre - estimation of the neural network model into the adaptive Kalman filter, and correct the pre - estimated vehicle mass value output by the neural network model based on the weight factor.

[0011] In the above - mentioned technical solution, through the neural network model in Step 3, the non - linear relationship between vehicle states is learned, a vehicle mass estimator based on the neural network is established, and a relatively accurate initial mass pre - estimation value is provided. After setting the weight of the output result of the neural network model, the initial mass pre - estimation is combined in the adaptive Kalman filter, and based on the weight of the neural network model, a vehicle mass estimation value with higher accuracy is output.

[0012] Preferably, in Step 2, the longitudinal dynamic model of mass and road slope is specifically:

[0013]

[0014] In the formula, T tq is the driving torque, i g is the transmission ratio, i 0 is the final drive ratio, η T is the mechanical efficiency of the transmission chain, r is the wheel rolling radius, C d is the air resistance coefficient, A f is the frontal area, ρ is the air density, v x is the vehicle speed, m is the vehicle mass, g 0 is the acceleration due to gravity, α is the slope angle, f is the rolling resistance coefficient, a x is the vehicle acceleration.

[0015] Preferably, in Step 3, the vehicle parameter data collected by the sensor is used as the training data set of the neural network model.

[0016] Preferably, after the neural network model training is completed in Step 3, the trained neural network model is deployed on the vehicle, and the real - time data of the vehicle - mounted sensor is used as the real - time input data of the neural network model, so that the neural network model outputs the vehicle mass pre - estimation value m nn .

[0017] Preferably, in Step 4, the working steps of the weight regulator are:

[0018] S4.1: Record the distribution characteristics of the training data set during the training process of the neural network model;

[0019] S4.2: Compare the distribution characteristics of the real-time input data with those of the training data set to determine the similarity between the real-time input data and the training data set. Output the confidence level τ of the estimation result obtained by the neural network based on the real-time data according to the similarity level. The confidence level τ is the weight factor. If the similarity level is high, the confidence level τ of the estimation result obtained by the neural network based on the real-time data is higher and the weight it occupies in the final estimation result is higher. If the similarity level is low, the confidence level of the estimation result obtained by the neural network based on the real-time data is lower and the weight it occupies in the final estimation result is lower.

[0020] Preferably, in the fifth step, the specific process is as follows:

[0021] S5.1: The Kalman filter selects the extended Kalman filter and establishes the extended Kalman filter expression equation according to the nonlinear system;

[0022] S5.2: Transform the Kalman filter expression equation based on the vehicle longitudinal dynamics model;

[0023] S5.3: Add the real-time output vehicle mass pre-estimation value m nn and the confidence level τ of the neural network model to the transformed equation, and output the estimated value of the vehicle current state quantity. The estimated value of the state quantity includes the estimated value of the vehicle mass.

[0024] Preferably, in step S5.1, for the nonlinear system x k = g(x k-1 , μ k-1 ) + w k-1 , y k = h(x k ) + v k , the expression equation of the extended Kalman filter estimator is as follows:

[0025] Calculate the prior estimate of the state quantity

[0026]

[0027] Calculate the prior estimate of the covariance matrix

[0028]

[0029] Calculate the Kalman gain K k

[0030]

[0031] Calculate the posterior estimate of the state quantity based on the observed quantity

[0032]

[0033] Update the posterior estimate P of the covariance matrix k

[0034]

[0035] where x k is the state quantity at the k-th step, y k is the observed quantity, μ is the control quantity, the process noise w follows the distribution N(0, Q w ), the observation noise v follows the distribution N(0, R v ), P is the covariance matrix, Q w is the process noise covariance matrix, R v is the measurement noise covariance matrix, is a matrix, g is the state transition function, h is the observation function.

[0036] In step S5.2: Select the variables of the extended Kalman filter equation

[0037] Select the state quantity as:

[0038]

[0039] Further expressed as:

[0040]

[0041] Select the observed quantity as:

[0042] y k =[v x

[0043] Establish the observation equation as follows:

[0044]

[0045] Calculate the Jacobian matrix A:

[0046]

[0047]

[0048]

[0049]

[0050] where v x is the vehicle speed, m is the vehicle mass, f k is the rolling resistance coefficient at time k, is the vehicle speed at time k, m k ​is the vehicle mass at time k, w k-1 is the process noise at time k-1, is the vehicle speed at time k-1, f k-1 is the rolling resistance coefficient at time k-1, m k-1 is the vehicle mass at time k-1, a x is the vehicle acceleration, ΔT is the acquisition time, H = [1 0 0].

[0051] Preferably, in step S5.3, the neural network model outputs the vehicle mass pre-estimation value m nn and the confidence level τ and adds them to the transformed equation, specifically:

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058] In the formula, is the expanded measurement noise covariance matrix; is the prior estimate of the covariance matrix; τ k is the confidence level at time k; H = [1 1 0]; is the noise covariance matrix of the neural network pre-estimation value; is the expanded state quantity; is the state quantity at time k.

[0059] is the state quantity vector, which includes the speed and mass of the vehicle at time k, that is, the vehicle mass estimation value is output.

[0060] Preferably, in step five, the adaptive Kalman filter is an adaptive unscented Kalman filter, and the specific process is as follows:

[0061] S5-1: Calculate 2n + 1 σ-point sets X

[0062]

[0063] In the formula: e represents the e-th column of the variance matrix P, and the square root can be calculated using Cholesky decomposition here; n is the dimension of the state vector; λ is the scaling ratio parameter, indicating the distance of the sampled σ-point from the mean value, λ = α 2(n + κ) - n; α, κ, and β are all parameters to be selected; P is the variance matrix.

[0064] S5-2: Calculate the weights corresponding to the σ-point set X

[0065]

[0066] In the formula: is the expected weight of the first σ-point; is the covariance weight of the first σ-point; is the expected weight of the remaining 2n σ-points; is the covariance weight of the remaining 2n σ-points; the subscript m represents expectation; c represents covariance;

[0067] S5-3: Use to obtain a set of σ-point sets and their corresponding weights

[0068]

[0069] In the formula, is the σ-point set obtained from ; is the posterior estimate of the state quantity; P k-1 is the posterior estimate of the covariance matrix;

[0070] S5-4: Substitute the above σ-point set into the state transition equation x k = g(x k-1 , μ k-1 ) + w k-1 to obtain the σ-point set Calculate the predicted value and the covariance matrix as follows:

[0071]

[0072] In the formula, is the prior estimate value of the state quantity; is the prior estimate of the covariance matrix;

[0073] S5-5: After obtaining the predicted value , use the unscented transform to obtain a new set of σ-points;

[0074] S5-6: Substitute the new σ-point set obtained in S5-5 into the observation equation y k = h(x k ) + v k to calculate the predicted observed quantity The above state transition equation and observation equation are both equations of a general nonlinear system;

[0075] S5-7: Sum the observed values obtained in S5-6 with weights to obtain its predicted mean and covariance, specifically:

[0076]

[0077] where is the estimated value of the observed value; is the covariance matrix of z k ; is the covariance matrix of x k and z k ;

[0078] S5-8: Calculate the Kalman gain matrix K k , specifically:

[0079]

[0080] where is the to the negative first power;

[0081] S5-9: Update the state quantity and the covariance matrix P k , specifically:

[0082]

[0083] where is the state quantity at time k, including the vehicle mass;

[0084] Preferably, the neural network model selects a multi-layer perceptron. Hyperparameters such as the number of network layers and the learning rate are selected according to data tuning. The deep learning framework on which the neural network model is based is not limited, including but not limited to Pytorch, Tensorflow, PaddlePaddle, MXNet; the neural network model can also select RNN, Transformer, etc.

[0085] Compared with the prior art, the beneficial effects of the present invention are: First, the neural network learns the non-linear relationship between vehicle states, reducing the impact caused by model mismatch. A weight factor is introduced to embed the result estimated by the neural network into the adaptive Kalman filter, adjusting the weights of the Kalman filter and the result estimated by the neural network, reducing the oscillation of the estimated result, and further improving the accuracy and stability of the vehicle mass prediction value. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 is a flowchart of a vehicle mass estimation method based on adaptive extended Kalman filter of the present invention;

[0087] Figure 2It is a schematic diagram of the adaptive extended Kalman filter of the present invention. Specific embodiments

[0088] The accompanying drawings are only for illustrative purposes and should not be construed as limitations on this patent; for better illustration of this embodiment, some components in the accompanying drawings will be omitted, enlarged or reduced, which do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the accompanying drawings may be omitted. The positional relationships described in the accompanying drawings are only for illustrative purposes and should not be construed as limitations on this patent.

[0089] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "long", "short", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the accompanying drawings are only for illustrative purposes and should not be construed as limitations on this patent. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0090] In the accompanying drawings of the embodiments of the present invention, for the convenience of reading and understanding, the front plate, rear plate and top plate in the casing structure are not shown.

[0091] The technical solutions of the present invention will be further specifically described below through specific embodiments and in conjunction with the accompanying drawings:

[0092] Embodiment 1

[0093] As Figure 1 shown in Embodiment 1 of a vehicle mass estimation method based on adaptive extended Kalman filtering, which includes the following steps:

[0094] Step 1: Obtain vehicle parameters;

[0095] Step 2: Establish a longitudinal dynamics model of vehicle mass and road slope according to the parameters in Step 1;

[0096] Step 3: Set a vehicle mass estimator based on a neural network model. According to the longitudinal dynamics model in Step 2, select parameters such as vehicle speed and acceleration as input features A t , and vehicle mass as the output, and train the neural network model;

[0097] Step 4: Set a weight regulator and calculate the weight factor;

[0098] Step 5: Embed the quality pre - estimation of the neural network model into the adaptive Kalman filter, and correct the vehicle mass pre - estimated value output by the neural network model based on the weight factor.

[0099] The working principle of this embodiment: Through the neural network model in Step 3, learn the non - linear relationship between vehicle states, establish a vehicle mass estimator based on the neural network, and provide a relatively accurate initial mass pre - estimated value. After setting the weights of the output results of the neural network model, combine the initial mass pre - estimation in the adaptive Kalman filter, and based on the weights of the neural network model, output a vehicle mass estimated value with higher accuracy.

[0100] The beneficial effects of this embodiment: First, learn the non - linear relationship between vehicle states through the neural network to reduce the influence caused by model mismatch. Introduce the weight factor, embed the result estimated by the neural network into the adaptive Kalman filter, adjust the weights of the Kalman filter and the neural network estimation result, reduce the oscillation of the estimation result, and further improve the accuracy and stability of the vehicle mass pre - estimated value.

[0101] Embodiment 2

[0102] As Figure 1-2 shown, an embodiment 2 of a vehicle mass estimation method based on adaptive extended Kalman filter includes the following steps:

[0103] Step 1: Obtain the vehicle parameters;

[0104] Step 2: Establish a longitudinal dynamics model of vehicle mass and road slope according to the parameters in Step 1. The longitudinal dynamics model of mass and road slope is specifically:

[0105]

[0106] In the formula, T tq is the driving torque, i g is the transmission ratio of the transmission, i 0 is the transmission ratio of the final drive, η T is the mechanical efficiency of the transmission chain, r is the wheel rolling radius, C d is the air resistance coefficient, A f is the frontal area, ρ is the air density, v x is the vehicle speed, m is the vehicle mass, g 0 is the acceleration due to gravity, α is the slope angle, f is the rolling resistance coefficient, a x is the vehicle acceleration.

[0107] Step 3: Set a vehicle mass estimator based on the neural network model. According to the longitudinal dynamics model in Step 2, select parameters such as vehicle speed and acceleration as the input features A of the neural network t, the vehicle mass is used as the output, and the neural network model is trained; the neural network model of this embodiment is based on a deep learning framework.

[0108] In this embodiment, the vehicle parameter data collected by the sensor is used as the training data set of the neural network model. After the neural network model is trained through the training data set, the trained neural network model is deployed on the vehicle, and the real-time data of the vehicle-mounted sensor is used as the real-time input data of the neural network model, so that the neural network model outputs the pre-estimated value m of the vehicle mass in real time. nn

[0109] Step four: Set a weight regulator to calculate the weight factor. The specific steps are as follows:

[0110] S4.1: Record the distribution characteristics of the training data set during the training process of the neural network model.

[0111] S4.2: Compare the distribution characteristics of the real-time input data with those of the training data set, determine the similarity between the real-time input data and the training data set, and output the confidence level τ of the estimation result obtained by the neural network based on the real-time data according to the similarity. The confidence level τ is the weight factor. If the similarity is high, the confidence level τ of the estimation result obtained by the neural network based on the real-time data is higher, and the weight in the final estimation result is higher. If the similarity is low, the confidence level of the estimation result obtained by the neural network based on the real-time data is lower, and the weight in the final estimation result is lower.

[0112] Step five: Embed the pre-estimation of the vehicle mass of the neural network model into the adaptive Kalman filter, and correct the pre-estimated value of the vehicle mass output by the neural network model based on the weight factor. The specific steps are as follows:

[0113] S5.1: The Kalman filter selects the extended Kalman filter, and establishes the extended Kalman filter expression equation according to the nonlinear system; for the nonlinear system x k = g(x k-1 , μ k-1 ) + w k-1 , y k = h(x k ) + v k , the expression equation of the extended Kalman filter estimator is as follows:

[0114] Calculate the prior estimate of the state quantity

[0115]

[0116] Calculate the prior estimate of the covariance matrix

[0117]

[0118] Calculate the Kalman gain K k

[0119]

[0120] Calculate the posterior estimate of the state quantity based on the observation quantity

[0121]

[0122] Update the posterior estimate P of the covariance matrix k

[0123]

[0124] where, x k is the state quantity at the k-th step, y k is the observation quantity, μ is the control quantity, the process noise w follows the distribution N(0, Q w ), the observation noise v follows the distribution N(0, R v ), P is the covariance matrix, Q w is the process noise covariance matrix, R v is the measurement noise covariance matrix, is a matrix, g is the state transition function, h is the observation function.

[0125] S5.2: Transform the Kalman filter expression equation based on the vehicle longitudinal dynamics model, specifically restrict the state quantity and the observation quantity, as follows:

[0126] Select the variables of the extended Kalman filter equation

[0127] Select the state quantity as:

[0128]

[0129] Further expressed as:

[0130]

[0131] Select the observation quantity as:

[0132] y k = [v x

[0133] Establish the observation equation as follows:

[0134]

[0135] Calculate the Jacobian matrix A:

[0136] ​

[0137]

[0138]

[0139]

[0140] wherein, v x is the vehicle speed, m is the vehicle mass, f k is the rolling resistance coefficient at time k, is the vehicle speed at time k, m k is the vehicle mass at time k, w k-1 is the process noise at time k-1, is the vehicle speed at time k-1, f k-1 is the rolling resistance coefficient at time k-1, m k-1 is the vehicle mass at time k-1, a x is the vehicle acceleration, ΔT is the acquisition time, H = [1 0 0].

[0141] S5.3: Add the vehicle mass pre-estimation value m nn and the confidence level τ output by the neural network model to the converted equation, and output the estimated value of the current state quantity of the vehicle. The estimated value of the state quantity includes the estimated value of the vehicle mass of the automobile, specifically:

[0142]

[0143]

[0144]

[0145]

[0146]

[0147]

[0148] wherein, is the expanded measurement noise covariance matrix; is the prior estimate of the covariance matrix τ k is the confidence level at time k; H = [1 1 0]; is the noise covariance matrix of the neural network pre-estimation value; is the expanded state quantity; is the state quantity at time k.

[0149] is the state quantity vector, which includes the speed and mass of the vehicle at time k of the automobile, that is, the estimated value of the vehicle mass is output.

[0150] Beneficial effects of this embodiment: First, the non-linear relationship between vehicle states is learned through a neural network to reduce the influence caused by model mismatch. A weight factor is introduced, and the result estimated by the neural network is embedded into the adaptive Kalman filter to adjust the weights of the Kalman filter and the result estimated by the neural network, reducing the oscillation of the estimation result and further improving the accuracy and stability of the vehicle mass prediction value.

[0151] Embodiment 3

[0152] Embodiment 3 of a vehicle mass estimation method based on adaptive extended Kalman filter. Based on the above 2, the difference from Embodiment 2 is that the adaptive extended Kalman filter in Step 5 is replaced by an adaptive unscented Kalman filter. The specific process is as follows:

[0153] S5-1: Calculate 2n + 1 σ-point sets X

[0154]

[0155] In the formula: e represents the e-th column of the variance matrix P. Here, the square root can be calculated using Cholesky decomposition; n is the dimension of the state vector; λ is the scaling ratio parameter, indicating the distance of the sampled σ-points from the mean, λ = α 2 (n + κ) - n; α, κ, and β are all parameters to be selected; P is the variance matrix.

[0156] S5-2: Calculate the weights corresponding to the σ-point set X

[0157]

[0158] In the formula: is the expected weight of the first σ-point; is the covariance weight of the first σ-point; is the expected weight of the remaining 2n σ-points; is the covariance weight of the remaining 2n σ-points; the subscript m represents the expectation; c represents the covariance;

[0159] S5-3: Use to obtain a set of σ-point sets and their corresponding weights

[0160]

[0161] In the formula, is obtained from to get the σ-point set; is the posterior estimate of the state quantity; P k-1 is the posterior estimate of the covariance matrix;

[0162] S5-4: Substitute the above σ-point set into the state transition equation to obtain the σ-point set Calculate the predicted value and covariance matrix as follows:

[0163]

[0164] In the formula, is the prior estimate of the state quantity; is the prior estimate of the covariance matrix

[0165] S5-5: Obtain the predicted value After that, use the unscented transform to obtain a new set of σ points;

[0166] S5-6: Substitute the new set of σ points obtained in S5-5 into the observation equation to calculate the predicted observed quantity

[0167] S5-7: Weighted sum the observed quantities obtained in S5-6 to obtain its predicted mean and covariance, specifically:

[0168]

[0169] In the formula, is the estimated value of the observed quantity; is z k 's covariance matrix; is x k and z k 's covariance matrix;

[0170] S5-8: Calculate the Kalman gain matrix K k , specifically:

[0171]

[0172] In the formula, is 's negative first power;

[0173] S5-9: Update the state quantity and covariance matrix P k , specifically:

[0174]

[0175] In the formula, is the state quantity at time k, including the vehicle mass.

[0176] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, rather than limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation manners here. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the claims of the present invention.

Claims

1. A vehicle mass estimation method based on adaptive extended Kalman filter, characterized in that, it includes the following steps: Step 1: Obtain vehicle parameters; Step 2: Establish a longitudinal dynamics model of vehicle mass and road slope according to the parameters in Step 1; Step 3: Set up a vehicle mass estimator based on a neural network model. According to the longitudinal dynamics model in Step 2, select vehicle speed and acceleration parameters as the input features of the neural network , with vehicle mass as the output, and train the neural network model; Step 4: Set a weight regulator and calculate the weight factor; Step 5: Embed the mass pre-estimation of the neural network model into the adaptive Kalman filter, and based on the weight factor, correct the pre-estimated value of the vehicle mass output by the neural network model. The specific process is as follows: S5.1: Select the extended Kalman filter for the Kalman filter, and establish the extended Kalman filter expression equation according to the nonlinear system; For a non-linear system , , the expression equation of the extended Kalman filter estimator is as follows: Prior estimate of the calculated state quantity Calculate the prior estimate of the covariance matrix Calculate the Kalman gain Posterior Estimation of State Quantities Calculated Based on Observed Quantities Update the posterior estimate of the covariance matrix Among them, is the state quantity of the th step, is the observed quantity, is the control quantity, and the process noise obeys the distribution , and the observation noise obeys the distribution . is the covariance matrix, is the process noise covariance matrix, is the measurement noise covariance matrix, , is a matrix, , is the state transition function, is the observation function; S5.2: Transform the Kalman filter expression equation based on the vehicle longitudinal dynamics model; Select variables of the extended Kalman filter equation The selected state variables are: Further expressed as: The selected observed variables are: The established observation equation is as follows: Calculate the Jacobian matrix : Wherein, is the vehicle speed, is the vehicle mass, is the rolling resistance coefficient at the k-th moment, is the vehicle speed at the k-th moment, is the vehicle mass at the k-th moment, is the process noise at the (k - 1)-th moment, is the vehicle speed at the (k - 1)-th moment, is the rolling resistance coefficient at the (k - 1)-th moment, is the vehicle mass at the (k - 1)-th moment, is the vehicle acceleration, is the acquisition time, H = ; S5.3: Output the real-time pre-estimated value of the vehicle mass by the neural network model and the confidence level into the converted equation, and output the estimated value of the current vehicle state quantity, where the estimated value of the state quantity includes the estimated value of the vehicle mass of the automobile.

2. According to the vehicle mass estimation method based on adaptive extended Kalman filter described in claim 1, characterized in that, in the said Step 2, the longitudinal dynamics model of mass and road slope is specifically: Wherein, is the driving torque, is the transmission ratio, is the final drive ratio, is the mechanical efficiency of the transmission chain, is the rolling radius of the wheel, is the air resistance coefficient, is the frontal area, is the air density, is the vehicle speed, is the vehicle mass, is the acceleration due to gravity, is the slope angle, is the rolling resistance coefficient, is the vehicle acceleration.

3. According to the vehicle mass estimation method based on adaptive extended Kalman filter described in claim 1, characterized in that, in the said Step 3, the vehicle parameter data collected by the sensor is used as the training data set of the neural network model.

4. According to the vehicle mass estimation method based on adaptive extended Kalman filter described in claim 3, characterized in that, After the neural network model is trained in Step 3, the trained neural network model is deployed to the vehicle, and the real-time data of in-vehicle sensors is used as the real-time input data of the neural network model, so that the neural network model outputs the pre-estimated value of vehicle mass in real time. .

5. According to the vehicle mass estimation method based on adaptive extended Kalman filter described in claim 4, characterized in that, in the said Step 4, the working steps of the weight regulator are: S4.1: Record the distribution characteristics of the training data set during the training process of the neural network model; S4.2: Compare the distribution characteristics of the real-time input data with those of the training data set to determine the similarity between the real-time input data and the training data set, and output the confidence level of the estimation result obtained by the neural network based on the real-time data according to the similarity level. , the confidence level is the weight factor.

6. According to the vehicle mass estimation method based on adaptive extended Kalman filter described in claim 1, characterized in that, In step S5.3, the neural network model outputs the pre-estimated value of the vehicle mass in real time and the confidence level are added to the converted equation, specifically: In the formula, is the expanded measurement noise covariance matrix; is the prior estimate of the covariance matrix; is the confidence level at time k; ; is the noise covariance matrix of the neural network pre-estimated value; is the expanded state quantity; is the state quantity at time k.

7. According to the vehicle mass estimation method based on adaptive extended Kalman filter described in claim 4, characterized in that, in Step 5, the adaptive Kalman filter is an adaptive unscented Kalman filter, and the specific process is as follows: S5-1: Calculate 2n + 1 point sets In the formula: represents the th column of the variance matrix P, where the square root can be calculated using Cholesky decomposition; n is the dimension of the state vector; is the scaling parameter, indicating the distance of the sampled points from the mean, ; , and are all parameters to be selected; P is the variance matrix; S5-2: Calculation Point set Corresponding weight Where: is the expected weight of the first point; is the covariance weight of the first point; are the expected weights of the remaining points; point; are the covariance weights of the remaining points; The subscript represents the expectation; represents the covariance; represents the covariance; S5-3: Using to obtain a set of point sets and their corresponding weights In the formula, is obtained from to get point set; is the posterior estimate of the state quantity; is the posterior estimate of the covariance matrix; S5-4: Substitute the above point set into the state transition equation to obtain point set , and calculate the predicted value and covariance matrix as follows: In the formula, is the prior estimated value of the state quantity; is the prior estimate of the covariance matrix; S5-5: Obtain the predicted value After that, use the unscented transform to obtain a new point set; S5-6: Substitute the new point set obtained in S5-5 into the observation equation to calculate the predicted observable quantity ; S5-7: The observed values obtained in S5-6 are weighted and summed to obtain the predicted mean and covariance, specifically: Wherein, is the estimated value of the observed quantity; is 's covariance matrix; is and 's covariance matrix; S5-8: Calculate the Kalman gain matrix , specifically as follows: In the formula, is to the power of negative one; S5-9: Update status quantity and covariance matrix , specifically as follows: In the formula, is the state quantity at time k, including vehicle mass.

8. According to the vehicle mass estimation method based on adaptive extended Kalman filter described in any one of claims 1-7, characterized in that, the deep learning framework on which the neural network model is based.

Images