A machine learning and dynamics based vehicle slope estimation method and system

Abstract

The application discloses a vehicle slope estimation method and system based on machine learning and dynamics, aiming at solving the problems of insufficient estimation accuracy, poor robustness and high sensor cost in the prior art. Necessary data information is acquired from a vehicle-mounted network bus and sensors, the data is preprocessed and divided into a training set and a verification set, the training set is used for training a feedforward neural network, an extended state observer based on vehicle longitudinal dynamics is constructed, and a variable weighting coefficient is used for solving. The application uses machine learning and vehicle dynamics to estimate the slope, has the advantages of high precision and strong robustness, and provides a reliable slope signal input for an automobile control system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vehicle parameter estimation, and more particularly, to a vehicle slope estimation method and system based on machine learning and dynamics. BACKGROUND

[0002] Road slope estimation, as an important part of vehicle parameter estimation, has a great influence on the handling performance, braking performance and fuel economy of the vehicle during driving. For example, the shift strategy of the automatic transmission needs to adjust the gear according to the slope information, which provides convenience for the driver; at the same time, the engine fuel consumption optimization, hybrid vehicle energy management, vehicle cruise control and other technologies also need slope information to cooperate to achieve the driving goal of the vehicle and the optimization of fuel economy.

[0003] The estimation method of automobile slope is mainly based on the slope estimation method of kinematics. This method uses the data collected by the accelerometer or other sensors directly, but in actual application, the sensor is greatly affected by environmental factors, has poor robustness, and has poor adaptability to different scenes; at the same time, many researchers use vehicle longitudinal dynamics method to estimate the slope, which has fast convergence speed and good stability, but this method is highly dependent on the accuracy of the vehicle longitudinal dynamics model, and the method is poor at low speed; in recent years, the vehicle slope estimation method based on machine learning has also made certain development, but the comprehensiveness of the training data will directly affect the neural network training result, and then directly affect the performance of the method.

[0004] Among the existing patents, such as Chinese patent authorization announcement number CN108960426B, authorization announcement date May 14, 2021, and invention name "Road slope comprehensive estimation system based on BP neural network", this invention uses the method of BP neural network, according to the data collected by the wheel vertical load acquisition unit, uses different road types to call different neural network modules for slope estimation, but the result of this method is affected by the training sample set, and its robustness needs to be improved; such as Chinese patent publication number CN106840097B, authorization announcement date May 25, 2021, and invention name "Road slope estimation method based on adaptive extended Kalman filter", this invention uses the extended Kalman filter estimation model, which can quickly converge and improve the real-time requirements of the automobile slope estimation algorithm, but this method has poor adaptability to some harsh working conditions and signal acquisition conditions.

[0005] In summary, among the current vehicle slope estimation methods, a high-precision and good-robustness automobile slope estimation method is needed, which is not affected by unknown parameters. Therefore, it is necessary to provide such a vehicle slope estimation method based on machine learning and dynamics. SUMMARY

[0006] The technical problem to be solved by the present application is to overcome the problems of low precision, poor robustness, high dependence on sensor results and related data in the existing vehicle slope estimation method, and to provide a vehicle slope estimation method and system based on machine learning and dynamics.

[0007] To solve the above technical problems, the present application is implemented by adopting the following technical solutions: comprising the following steps:

[0008] A vehicle slope estimation method based on machine learning and dynamics comprises the following steps:

[0009] Step one, constructing a vehicle longitudinal dynamics model and determining the dynamic parameters required by the vehicle longitudinal dynamics model;

[0010] Step two, collecting dynamic parameters and corresponding actual slope values of the vehicle under different working conditions, and then normalizing the dynamic parameters and corresponding actual slope values to construct a training data set and a verification data set;

[0011] Step three, constructing a neural network and training and verifying the neural network using the training data set and the verification data set to obtain a trained neural network;

[0012] Step four, designing a dynamics-based extended state observer according to the vehicle longitudinal dynamics model;

[0013] Step five, in the actual operation of the vehicle, inputting the real-time collected dynamic parameters into the trained neural network, and obtaining the vehicle slope estimation value based on the trained neural network and the slope estimation value based on the extended state observer according to the extended state observer and the slope estimation value based on the extended state observer to and According to the current vehicle driving condition, the k-time slope estimation value is obtained by fusion

[0014] In the formula, i=1, 2 or 3;

[0015] respectively represent the neural network estimation value at k time the dynamics-based extended state observer estimation value at k time and the estimation result at k-1 time

[0016] λ i (v x (k)) respectively represent the weight coefficient of the corresponding slope estimation value, and the weight coefficient is related to the current time speed.

[0017] In a further improvement, the longitudinal dynamics model of the vehicle in step one is as follows:

[0018]

[0019] In the formula, T e — Engine torque, in Nm

[0020] i g — Gearbox ratio, unitless

[0021] i0 — Transmission ratio of the main reducer, dimensionless

[0022] η t —Transmission efficiency, unitless

[0023] r – Wheel radius, in meters

[0024] C D —Air drag coefficient, unitless

[0025] A – Frontal area, in m² 2

[0026] ρ — air density, kg / m³ 3

[0027] v x —Longitudinal speed of the vehicle, in m / s

[0028] m — Total vehicle weight, in kg

[0029] g — acceleration due to gravity, unit: m / s² 2

[0030] f – Rolling resistance coefficient, dimensionless

[0031] θ – slope, in rad

[0032] δ — Rotational mass conversion factor, dimensionless

[0033] —Longitudinal acceleration of the vehicle, in m / s² 2 ;

[0034] The power parameters include the vehicle speed v at the previous moment and the moment before that. x (t), v x (t-1), the acceleration at the current time and the previous time. Engine torque T at the current moment and the previous moment e (t), T e (t-1) and the slope estimation results of the previous time step.

[0035] Further improvement, in the step two, the working conditions include highway working condition, mountain road working condition.

[0036] Further improvement, in the step three, the neural network is a feedforward neural network, the feedforward neural network is composed of an input layer and two hidden layers;

[0037] The current time and the vehicle speed v x (t) and the vehicle speed v x (t-1) of the previous time, the current time and the acceleration and the current time and the engine torque T e (t) and the engine torque T e (t-1) of the previous time, and the slope estimation result of the previous time As the neural network input, the output result is the slope estimation value of the current time

[0038] The ReLU function is selected as the activation function of the hidden layer, and the formula is as follows:

[0039]

[0040] In the formula, o i The weighted sum result of the input of the same unit of the previous layer; F() represents the ReLU function.

[0041] Further improvement, in the step four, the design method of the dynamics-based extended state observer is as follows: formula (1) is changed into the following form:

[0042]

[0043] Define the vehicle longitudinal speed as the state quantity and the engine torque as the control quantity, then formula (3) is changed into the standard form of the extended state observer:

[0044]

[0045] In the formula, x non The nonlinear part of the state observer; x is the longitudinal vehicle speed, is the longitudinal acceleration, and t is time;

[0046] In actual estimation, the result is calculated through iteration until convergence, as follows:

[0047]

[0048] In the formula, e(k) is the error of the state quantity, i.e. the vehicle speed at time k

[0049] h1, h2 - both represent iterative calculations, is the state estimation value, i.e., the longitudinal vehicle speed, x(k) is the actual longitudinal vehicle speed, is the nonlinear part of the state estimation value, u(k) is the control amount, i.e., the engine torque, b(k) is a constant, which is calculated from the vehicle inherent parameters, and e(k) is the estimation error of the longitudinal vehicle speed

[0050] slope estimation value is:

[0051]

[0052] wherein, is the nonlinear part of the state observer; x 2 (k) is the square of x(k).

[0053] Further improvement, lambda i (v x (k)) takes the value of the vehicle speed / 100.

[0054] A vehicle slope estimation system based on machine learning and dynamics, for running the vehicle slope estimation method based on machine learning and dynamics described above.

[0055] A vehicle slope estimation system based on machine learning and dynamics, for running the vehicle slope estimation method based on machine learning and dynamics described above.

[0056] The beneficial effects of the present application compared with the prior art are:

[0057] 1. The vehicle slope estimation method based on machine learning and dynamics described in the present application uses machine learning and neural network related knowledge, trains the algorithm using historical data, and increases the estimation accuracy of the vehicle slope estimation, especially the accuracy of some specific scenarios;

[0058] 2. The signal, i.e., data, used by the vehicle slope estimation method based on machine learning and dynamics described in the present application can be satisfied through the vehicle CAN bus and basic sensors, without using expensive on-board equipment, and has practical application value, reducing the cost of slope estimation during vehicle use;

[0059] 3. The vehicle slope estimation method based on machine learning and dynamics described in the present application adds an extended state observer based on dynamics on the basis of traditional machine learning, solves the problem of high dependence of machine learning on the training set, and improves the robustness of the overall algorithm;

[0060] 4. The vehicle slope estimation method based on machine learning and dynamics according to the present application can avoid the influence of abnormal acceleration signals or other signals that are greatly affected by the environment, and further improve the accuracy of the estimation results. BRIEF DESCRIPTION OF DRAWINGS

[0061] The present application will be further described below in conjunction with the accompanying drawings:

[0062] Figure 1 The design flowchart of the vehicle slope estimation method according to the present application is shown in the figure.

[0063] Figure 2 The schematic diagram of the feedforward neural network according to the present application is shown in the figure.

[0064] Figure 3 The data weighted fusion implementation method of the vehicle slope estimation method according to the present application is shown in the figure. DETAILED DESCRIPTION

[0065] The present application will be further described below in conjunction with the accompanying drawings:

[0066] The present application discloses a vehicle slope estimation method based on machine learning and dynamics. The present application uses the method of feedforward neural network and extended state observer to obtain the vehicle slope estimation value at the current time of each algorithm, and then fuses and weights the results according to the current vehicle speed to obtain reliable vehicle slope estimation data. The present application first analyzes the input signals of the neural network according to the longitudinal force state of the vehicle, mainly including the current time and the previous time vehicle speed v x (t), v x (t-1), the current time and the previous time acceleration and the current time and the previous time engine torque T e (t), T e (t-1), then the signals are preliminarily preprocessed into training set and verification set, the feedforward neural network is trained, and the final result is determined according to the feedback of the verification set, then the extended state observer based on the longitudinal dynamics of the vehicle is designed to ensure the robustness and accuracy of the vehicle slope estimation algorithm, and then the appropriate weighting factor is selected according to the real-time vehicle speed to further improve the reliability of the vehicle slope estimation result. The vehicle slope estimation method further improves the accuracy of slope estimation on the basis of the traditional method, and has stronger universality compared with other slope estimation methods based on machine learning. Next, the design idea of the vehicle slope estimation method based on machine learning and dynamics will be specifically described.

[0067] Reference is made to Figure 1The vehicle slope estimation method based on machine learning and dynamics includes vehicle longitudinal dynamics model construction and signal acquisition, signal preprocessing and dataset construction, neural network construction and training, design of an extended state observer based on dynamics, and weighted solving of the slope estimation value. The vehicle slope estimation method based on machine learning and dynamics is described in detail as follows.

[0068] The method comprises the following steps:

[0069] First, vehicle longitudinal dynamics model construction and signal acquisition

[0070] During driving, a vehicle is mainly subjected to five external forces, namely driving force, air resistance, rolling resistance, slope resistance, and acceleration resistance.

[0071] According to the force state of the current vehicle, the vehicle longitudinal dynamics balance equation can be listed as:

[0072]

[0073] In the formula, T e Engine torque, unit: Nm

[0074] i g Transmission ratio, unitless

[0075] i0Main reducer transmission ratio, unitless

[0076] η t Transmission efficiency, unitless

[0077] rWheel radius, unit: m

[0078] C D Air resistance coefficient, unitless

[0079] AWindward area, unit: m 2

[0080] ρAir density, kg / m 3

[0081] v x Vehicle longitudinal speed, unit: m / s

[0082] mVehicle longitudinal speed, unit: kg

[0083] gGravitational acceleration, unit: m / s 2

[0084] fRolling resistance coefficient, unitless

[0085] θSlope, unit: rad

[0086] δ — Rotational mass conversion factor, dimensionless

[0087] —Longitudinal acceleration of the vehicle, in m / s² 2

[0088] In the above information, the vehicle speed v at the current moment and the previous moment... x (t), v x (t-1), the acceleration at the current time and the previous time. And the engine torque T at the current moment and the previous moment. e (t), T e (t-1) serves as the input to the neural network. This information can be obtained directly or indirectly from wheel speed sensors, accelerometers, and the vehicle's CAN bus.

[0089] This data can be used to train neural networks and also as input signals for algorithms to calculate the slope at the current moment.

[0090] The second step is signal preprocessing and dataset construction.

[0091] After collecting the data mentioned in the first step under different scenarios (such as typical highway conditions, mountain road conditions, etc., which can be adjusted according to vehicle needs), it is necessary to perform necessary operations on this part of the signal and perform standardization processing such as unifying units to build training datasets and validation datasets.

[0092] The third step is the construction and training of the neural network.

[0093] See Figure 2 To obtain accurate road slope data, this method uses a feedforward neural network (FFNN), which consists of one input layer and two hidden layers. By combining current and historical information, the real-time slope of vehicles can be estimated more accurately.

[0094] The vehicle speed v at the current moment and the previous moment x (t), v x (t-1), the acceleration at the current time and the previous time. And the engine torque T at the current moment and the previous moment. e (t), T e (t-1), and the slope estimation results at the previous time step. As input to the neural network, the output is the slope estimate at the current moment.

[0095] The ReLU function is chosen as the activation function for the hidden layer, and its formula is as follows:

[0096]

[0097] In the formula, o i The weighted sum of the input of the same unit in the last layer

[0098] In the training process, select the appropriate learning rate, and input the preprocessed training data into the FFNN. At the same time, in order to avoid overfitting, when the prediction result of the validation set is deteriorated during the training process, the training is ended.

[0099] In actual work, input the real-time collected data into the trained neural network, and the vehicle slope estimation value based on machine learning can be obtained

[0100] Step 4: Design of the extended state observer based on dynamics

[0101] The purpose of the extended state observer (ESO) is to estimate the system disturbance according to the output of the system, which can meet the complex requirements in actual engineering.

[0102] As can be seen from formula (1), the longitudinal dynamics equation of the car is a nonlinear mathematical model, and the extended state observer can take the nonlinear part of the system as the extended state, which can improve the efficiency of observation while ensuring convergence, and is suitable for vehicle slope estimation. Formula (1) can be converted into the following form:

[0103]

[0104] Define the vehicle longitudinal speed as the state variable and the engine torque as the control variable, then formula (3) can be converted into the standard form of the extended state observer:

[0105]

[0106] In the formula, x non The nonlinear part of the state observer

[0107] In actual estimation, this method is calculated by iteration until the result converges, as follows:

[0108]

[0109] In the formula, e(k) is the error of the state variable, i.e. the vehicle speed at time k

[0110] The state variable at time k, the estimated value of the nonlinear part

[0111] h1, h2 are two iteration formulas

[0112] After sorting, the slope estimation value is:

[0113]

[0114] Fifth step, weighted solution of slope estimation value

[0115] Referring to Figure 3 In order to improve the real-time slope estimation value, the results of the feedforward neural network and the results of the extended state observer need to be fused according to the current vehicle driving conditions:

[0116]

[0117] In the formula, Slope estimation value at time k

[0118] Respectively represent the neural network estimation value at time k The extended state observer estimation value based on dynamics at time k And the estimation result at time k-1

[0119] λ i (v x (k)) Respectively represent the weight coefficient corresponding to the slope estimation value, and the weight coefficient is related to the current time speed

[0120] Considering that the vehicle longitudinal dynamics balance equation is accurate when the vehicle speed is high, the weight coefficient is related to the current time speed.

[0121] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not intended to limit the scope of protection of the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the essence and scope of the technical solutions of the present application.

Claims

1. A vehicle slope estimation method based on machine learning and dynamics, characterized in that: Includes the following steps: Step 1: Construct a longitudinal dynamics model of the vehicle and determine the dynamic parameters required for the longitudinal dynamics model. Step 2: Collect the vehicle's power parameters and corresponding actual slope values ​​under different working conditions, and then normalize the power parameters and corresponding actual slope values ​​to construct training datasets and validation datasets. Step 3: Construct a neural network, and train and validate the neural network using training and validation datasets to obtain a trained neural network; Step 4: Design a dynamic-based extended state observer based on the vehicle's longitudinal dynamics model; Step 5: During actual vehicle operation, the real-time collected dynamic parameters are input into the trained neural network, and the vehicle gradient estimate based on the trained neural network is obtained using the extended state observer. and slope estimates based on extended state observers right and The slope estimate at time k is obtained by fusing data based on the current vehicle driving conditions. In the formula, i = 1, 2 or 3; — These represent the neural network estimates at time k. Dynamically based extended state observer estimate at time k and the estimation results at time k-1 λ i (v x (k))——represents the weighting coefficients of the corresponding slope estimates, and the weighting coefficients are related to the vehicle speed at the current time; In step three, the neural network is a feedforward neural network, which consists of an input layer and two hidden layers. The vehicle speed v at the current moment and the previous moment x (t), v x (t-1), the acceleration at the current time and the previous time. And the engine torque T at the current moment and the previous moment. e (t), T e (t-1), and the slope estimation results at the previous time step. As input to the neural network, the output is the slope estimate at the current moment. The ReLU function is chosen as the activation function for the hidden layer, and its formula is as follows: In the formula, o i —The weighted summation of the inputs of the same unit from the previous layer; F() represents the ReLU function.

2. The vehicle slope estimation method based on machine learning and dynamics as described in claim 1, characterized in that, In step one, the longitudinal dynamics model of the vehicle is shown below: In the formula, T e — Engine torque, in Nm i g — Gearbox ratio, unitless i0 — Transmission ratio of the main reducer, dimensionless η t —Transmission efficiency, unitless r – Wheel radius, in meters C D —Air drag coefficient, unitless A – Frontal area, in m² 2 ρ — air density, kg / m³ 3 v x —Longitudinal speed of the vehicle, in m / s m — Total vehicle weight, in kg g — acceleration due to gravity, unit: m / s² 2 f – Rolling resistance coefficient, dimensionless θ – slope, in rad δ — Rotational mass conversion factor, dimensionless —Longitudinal acceleration of the vehicle, in m / s² 2 ; The power parameters include the vehicle speed v at the previous moment and the moment before that. x (t), v x (t-1), the acceleration at the current time and the previous time. Engine torque T at the current moment and the previous moment e (t), T e (t-1) and the slope estimation results of the previous time step.

3. The vehicle slope estimation method based on machine learning and dynamics as described in claim 1, characterized in that, In step two, the operating conditions include highway operating conditions and mountain road operating conditions.

4. The vehicle slope estimation method based on machine learning and dynamics as described in claim 2, characterized in that, In step four, the design method for the dynamic-based extended state observer is as follows: Equation (1) is transformed into the following form: Define the vehicle's longitudinal velocity as the state variable and the engine torque as the control variable, then equation (3) is transformed into the standard form of the extended state observer: In the formula, x non —The nonlinear part of the state observer; x is the longitudinal vehicle speed. The acceleration is longitudinal, and t is time. In practical estimation, iterative calculations are performed until the result converges, as shown below: In the formula, e(k) represents the state quantity at time k, i.e., the error in vehicle speed. h1, h2 — both represent iterative calculations. Here, x(k) is the estimated longitudinal speed, and x(k) is the actual longitudinal speed. The nonlinear part of the state estimate is represented by u(k), which is the control variable, i.e., the engine torque; b(k) is a constant calculated from the vehicle's inherent parameters; and e(k) is the estimation error of the longitudinal vehicle speed. Slope estimate for: in, For the nonlinear part of the state observer; x 2 (k) is the square of x(k).

5. The vehicle slope estimation method based on machine learning and dynamics as described in claim 1, characterized in that, λ i (v x The value of (k) is the vehicle speed / 100.

6. A vehicle slope estimation system based on machine learning and dynamics, characterized in that, The vehicle slope estimation system is used to run the vehicle slope estimation method based on machine learning and dynamics as described in any one of claims 1-5.

Images