State parameter estimation method of distributed driving special vehicle based on hybrid data driving

Abstract

The invention relates to a state parameter estimation method of a distributed driving special vehicle based on hybrid data driving, and belongs to the field of vehicle state parameter estimation and dynamics control, and the method comprises the steps: 1, constructing an eight-degree-of-freedom dynamics model and a tire model of the distributed driving special vehicle, the input and output vector design module is used for guiding input and output vector design of the LSTM neural network; 2, constructing a simulation environment of diversified working conditions and driving scenes by utilizing dynamics software, and collecting a multi-dimensional data set covering the vehicle speed, the steering wheel angle and the road adhesion coefficient; 3, designing an LSTM neural network architecture comprising an input layer, an LSTM layer, a full connection layer and a regression layer, and optimizing model parameters through offline training; and 4, deploying the trained model, and realizing real-time estimation of vehicle state parameters based on real-time sensor data. According to the method, state parameter estimation of the distributed driving special vehicle in a non-structural environment can be realized, and the accuracy and the real-time performance of parameter estimation are ensured at the same time.

Description

Technical Field

[0001] The present invention relates to the field of vehicle state parameter estimation and dynamic control, and particularly to a method for estimating state parameters of a distributed drive special vehicle based on hybrid data drive. Background Art

[0002] A special vehicle is a vehicle that can perform special operations in complex and uncertain environments such as jungles, mountains, grasslands, mud, deserts, and mining areas. For a distributed drive special vehicle, the motors are directly installed inside the hubs or at the wheel ends of each wheel, and the wheels are directly driven by the motors with the torque of each wheel independently controllable, enabling the power system control to respond precisely and quickly and providing more degrees of freedom for vehicle dynamic control.

[0003] Since a distributed drive special vehicle travels in complex and uncertain environments for a long time, higher requirements are put forward for the control of vehicle body stability during driving, and the real-time and accurate estimation of vehicle state parameters is the key basis and prerequisite for realizing stability control. There are mainly two methods for estimating vehicle motion state parameters: estimation based on a vehicle physical model and estimation based on data drive. Traditional methods mainly rely on a dynamic model and use Kalman filter (KF) and its variants (such as EKF, UKF) for estimation. These methods rely on an accurate dynamic model, but it is difficult to accurately determine model parameters under complex working conditions, resulting in a decrease in estimation accuracy. Observer methods (such as SMO, NLO) estimate the state by designing an observer, which can alleviate the uncertainty of model parameters, but have poor noise robustness and high computational complexity. In recent years, data-driven methods (such as RNN, CNN) have received attention due to their non-linear fitting ability. These methods learn the mapping relationship of state parameters through training data, without the need for an accurate dynamic model, but are highly dependent on high-quality training data, have limited generalization ability, and lack physical constraints, and the estimation results may violate the dynamic laws. The hybrid data-driven method combines the prior knowledge of the dynamic model and the fitting ability of the data-driven method and has been improved, but its robustness and real-time performance under complex working conditions still need to be enhanced. Summary of the Invention

[0004] Aiming at the above deficiencies of the prior art, the present invention proposes a method for estimating state parameters of a distributed drive special vehicle based on hybrid data drive, in order to enable the distributed drive special vehicle to accurately estimate vehicle state parameters in real time during driving in an unstructured environment, thereby ensuring the driving safety and handling stability of the distributed drive special vehicle in an unstructured environment.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] The characteristics of a method for estimating state parameters of a distributed drive special vehicle based on hybrid data drive according to the present invention are as follows: The method is carried out according to the following steps:

[0007] Step 1: Construct an eight-degree-of-freedom dynamics model and a tire model for a distributed-drive special vehicle, which are used to design the input and output vectors of the LSTM neural network;

[0008] Step 2: Use dynamics software to construct a simulation environment under different working conditions, so as to collect data for the input vector in the simulation environment, and obtain a vehicle state feature data set of the input vector under different working conditions;

[0009] Step 3: Construct an LSTM neural network, including: an input layer, two LSTM layers, a fully connected layer and a regression layer, and process the vehicle state feature data set of the input vector under different working conditions to obtain the predicted value of the output vector;

[0010] Step 4: Construct a multi-objective loss function for the LSTM neural network, including: estimation error, and dynamic constraints of longitudinal motion, lateral motion and roll motion;

[0011] Step 5: Offline train the LSTM neural network according to the vehicle state feature data set and the output vector until the multi-objective loss function converges, so as to obtain a vehicle state estimation model, which is used to estimate the vehicle state parameters of the real-time input sensor data.

[0012] The feature of a method for estimating state parameters of a distributed-drive special vehicle based on hybrid data drive according to the present invention also lies in that the step 1 includes:

[0013] Step 1.1: Establish an eight-degree-of-freedom dynamics model of longitudinal motion, lateral motion, yaw motion, roll motion and 4-wheel rotation of a distributed-drive special vehicle and a magic tire dynamics model;

[0014] Step 1.2: According to the eight-degree-of-freedom dynamics model and the magic tire dynamics model, infer the mapping relationship of each vehicle state parameter of the distributed-drive special vehicle, and determine the mapping function;

[0015] Step 1.3: According to the mapping function, use equation (1) to determine the input vector , output vector :

[0016]

[0017] In equation (1), , are longitudinal and lateral accelerations, is the yaw angular velocity, is the steering wheel angle, is the road surface adhesion coefficient, is the driving torque of the wheel on the j side of the i-th party, is the rotational speed of the wheel on the j side of the i-th party, where i ∈ {f, b}, f and b respectively represent the front and rear, and j ∈ {l, r}, l and r respectively represent the left and right, , are the longitudinal and lateral vehicle speeds, is the vehicle roll angle.

[0018] Further, the second step includes:

[0019] Step 2.1, in the steady-state, acceleration, and deceleration simulation environments of the vehicle, obtain the target speed curves under the three working conditions respectively;

[0020] Step 2.2, under various driving conditions in a complex unstructured environment, obtain the steering wheel angles in the low-frequency and high-frequency ranges respectively;

[0021] Step 2.3, on dry asphalt pavement, wet asphalt pavement, and snow-covered pavement, obtain the adhesion coefficients under the three pavement types respectively;

[0022] Step 2.4, after orderly combining the data sets in the three dimensions obtained in Steps 2.1 - 2.3, input them into a high-precision simulation dynamics model to obtain the input vector The vehicle state feature data set under different working conditions.

[0023] Further, the third step includes:

[0024] Step 3.1, the input layer composes the input vector The vehicle state feature data set under different working conditions into a vehicle state feature matrix with a dimension of , where is the length of each time step, h is the dimension number of the input vector , and h = 7;

[0025] Step 3.2, the first LSTM layer calculates the vehicle shallow state feature matrix through the long short-term memory mechanism according to , where represents the vehicle shallow state feature at the n-th time step, including the dynamic change information of all parameters in the input vector in the time dimension;

[0026] Step 4.3, the second LSTM layer calculates the vehicle deep state feature matrix through the long short-term memory mechanism according to , where represents the vehicle's deep - state features at the n - th time step, including the interaction information among all parameters in the input vector ;

[0027] Step 4.4: The fully - connected layer obtains the output state - parameter set , which respectively represent the predicted values of longitudinal and lateral vehicle speeds and vehicle roll angle:

[0028]

[0029] In formula (2), are two weight matrices of the fully - connected layer, are two bias vectors of the fully - connected layer, is the activation function.

[0030] Furthermore, in Step 4, the multi - objective loss function L is constructed using formula (3):

[0031]

[0032] In formula (3), is the estimation error, are the dynamic constraints of longitudinal motion, lateral motion, and roll motion, are three weights respectively, and there are:

[0033]

[0034] In formula (4), is the number of samples in the vehicle - state - feature dataset, is the longitudinal vehicle speed, lateral vehicle speed, and vehicle roll angle of the - th sample, is the predicted value of the longitudinal vehicle speed, lateral vehicle speed, and vehicle roll angle of the - th sample;

[0035]

[0036] In formula (5), is the time derivative of the predicted value of the longitudinal vehicle speed of the - th sample , is the longitudinal acceleration and yaw rate of the - th sample;

[0037]

[0038] In Equation (6), is the time derivative;

[0039] (7)

[0040] In Equation (7), is the moment of inertia of the unsprung mass about the x-axis, is the first and second time derivatives, is the unsprung mass, is the distance from the centroid of the unsprung mass to the roll center, is the equivalent roll damping of the vehicle suspension, is the equivalent roll stiffness of the vehicle suspension.

[0041] An electronic device according to the present invention includes a memory and a processor, characterized in that the memory is used to store a program for supporting the processor to execute the vehicle state parameter estimation method, and the processor is configured to execute the program stored in the memory.

[0042] A computer-readable storage medium according to the present invention, characterized in that when the computer program stored on the computer-readable storage medium is run by a processor, it executes the steps of the vehicle state parameter estimation method.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1. By combining the long short-term memory network with the dynamic constraints and the tire model, the present invention can accurately estimate the state parameters of the distributed drive special vehicle, including the longitudinal vehicle speed, the lateral vehicle speed and the roll angle, and ensure that the prediction results conform to the vehicle dynamics law, overcoming the problem of lack of physical consistency of the traditional data-driven model. The tire model introduces non-linear characteristics, enabling the model to meet the requirements of the distributed drive special vehicle for long-term driving in unstructured environments (such as muddy, sandy, steep slopes and other complex terrains), significantly improving the robustness of the estimation and the environmental adaptability.

[0045] 2. The long short-term memory network of the present invention adopts a lightweight design, with a simple model structure, and can operate efficiently on resource-limited embedded devices, meeting the requirements of real-time state estimation of distributed drive special vehicles in unstructured environments. The parameters of the dynamic model and the tire model can be flexibly adjusted to support different types of distributed drive special vehicles. The trained model is convenient for storage and migration, and can be quickly deployed to the control system of special vehicles, providing reliable support for vehicle control and safe operation in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is the eight-degree-of-freedom dynamic model diagram of the present invention;

[0047] Figure 2 This is the acquisition framework diagram of the vehicle state feature dataset of the present invention. Specific implementation mode

[0048] In this embodiment, a state parameter estimation method for a distributed drive special vehicle based on hybrid data drive is carried out according to the following steps:

[0049] Step 1: Construct an eight-degree-of-freedom dynamics model and a tire model of the distributed drive special vehicle for designing the input and output vectors of the LSTM neural network;

[0050] Step 1.1: Establish an eight-degree-of-freedom dynamics model of longitudinal motion, lateral motion, yaw motion, roll motion and four-wheel rotation of the distributed drive special vehicle and a magic tire dynamics model;

[0051] In specific implementation, the above-mentioned eight-degree-of-freedom dynamics model of the distributed drive special vehicle is as Figure 1 shown, and the differential equation expression is as formula (1):

[0052] (1)

[0053] In formula (1), is the vehicle mass, is the unsprung mass, is the longitudinal vehicle speed, is the lateral vehicle speed, is the front wheel steering angle, is the roll angle, is the yaw angular velocity, are the moments of inertia of the unsprung mass about the x and z axes respectively, is the distance from the center of mass to the front axle, is the distance from the center of mass to the rear axle, is the track width, are the longitudinal forces of the left front wheel, right front wheel, left rear wheel and right rear wheel respectively, are the lateral forces of the left front wheel, right front wheel, left rear wheel and right rear wheel respectively, is the wheel radius, is the rated moment of inertia of the wheel, represents the tire position number identifier, , , is the wheel rotational speed, is the wheel drive torque, is the wheel braking torque, is the lateral acceleration of the unsprung mass, is the equivalent roll stiffness of the vehicle suspension, is the equivalent roll damping of the vehicle suspension, is the distance from the centroid of the sprung mass to the roll center.

[0054] The differential equation expressions of the above magic tire dynamic model are as shown in Equations (2) and (3):

[0055] (2)

[0056] (3)

[0057] Equations (2) and (3) are the expressions of the longitudinal force and lateral force of the tire respectively, 、 are the peak factor and stiffness factor of the tire respectively, 、 are the curve shape factor and curvature factor respectively, is the tire slip ratio, is the tire sideslip angle, 、 are the horizontal drift and vertical drift of the curve respectively, is the road surface adhesion coefficient.

[0058] Step 1.2. According to the eight-degree-of-freedom dynamic model and the magic tire dynamic model, infer the mapping relationship of each vehicle state parameter of the distributed drive special vehicle to determine the mapping function;

[0059] In the specific implementation, the following five mapping functions are obtained according to Equations (1), (2), and (3): ; ; ; ; 。

[0060] Step 1.3. According to the mapping function, use Equation (4) to determine the input vector 、output vector of the LSTM neural network:

[0061] (4)

[0062] In Equation (4), 、 are the longitudinal and lateral accelerations, is the yaw angular velocity, is the steering wheel angle, is the road surface adhesion coefficient, is the driving torque of the wheel on the i-th side and j-th side, is the rotational speed of the wheel on the i-th side and j-th side, i ∈ {f, b}, f and b represent the front and rear respectively, j ∈ {l, r}, l and r represent the left and right respectively, , are the longitudinal and lateral vehicle speeds, and is the vehicle roll angle.

[0063] Step 2: Use dynamic software to construct a simulation environment under different working conditions, and then collect data on the input vector in the simulation environment to obtain a vehicle state feature dataset of the input vector under different working conditions;

[0064] Step 2.1, In the steady state, acceleration, and deceleration simulation environments of the vehicle, obtain the target speed curves under the three working conditions respectively;

[0065] In specific implementation, design 10 groups of different target speed curves to comprehensively cover the steady state, acceleration, and deceleration scenarios of the vehicle. The minimum set speed is 10 km / h, and the maximum speed is set to 100 km / h.

[0066] Step 2.2, Under various driving conditions in a complex unstructured environment, obtain the steering wheel angles in the low-frequency and high-frequency ranges respectively;

[0067] In specific implementation, first determine the frequency range of the steering wheel angle. In this embodiment, the steering wheel angle signal data of real vehicle driving in the DDD20 dataset (an open-source autonomous driving dataset from ETH Zurich) is selected. Perform Fourier transform on the steering wheel angle signal in the DDD20 dataset. The steering wheel rotation frequency is mainly concentrated in the range of 0.5 Hz. Considering the driving situation in a complex unstructured environment, the maximum rotation frequency of the steering wheel needs to reach 4 Hz. Therefore, the traditional low-frequency conditions are divided into three ranges (0.02 - 0.1 Hz, 0.1 - 0.2 Hz, 0.2 - 0.5 Hz), while the high-frequency conditions are concentrated in a relatively narrow frequency range (0.5 - 4 Hz).

[0068] Step 2.3, Under dry asphalt pavement, wet asphalt pavement, and snow-covered pavement, obtain the adhesion coefficients under the three pavement types respectively;

[0069] In specific implementation, the adhesion coefficient depends on various factors such as the road material, road surface condition, tire structure, tread pattern, and vehicle speed. Its specific value is difficult to directly measure. Therefore, three typical pavement situations of dry asphalt pavement, wet asphalt pavement, and snow-covered pavement are selected, corresponding to high-adhesion pavement, medium-adhesion pavement, and low-adhesion pavement respectively. In this embodiment, the adhesion coefficient of the high-adhesion pavement is 0.8, the adhesion coefficient of the medium-adhesion pavement is 0.5, and the adhesion coefficient of the low-adhesion pavement is 0.3.

[0070] Step 2.4, After orderly combining the datasets in the three dimensions obtained in Steps 2.1 - 2.3, input them into a high-precision simulation dynamics model to obtain the input vector Vehicle state feature datasets under different working conditions.

[0071] In specific implementation, the vehicle state feature dataset acquisition framework is as Figure 2 shown, and the relevant parameters of the simulation vehicle are set according to a certain distributed drive four-wheel special vehicle.

[0072] Step 3: Construct an LSTM neural network, including: an input layer, two LSTM layers, a fully connected layer, and a regression layer, and process the input vector for the vehicle state feature dataset under different working conditions to obtain the predicted value of the output vector;

[0073] Step 3.1: The input layer composes the input vector in the vehicle state feature dataset under different working conditions into a vehicle state feature matrix with a dimension of , where is the length of each time step, h is the dimension number of the input vector , and h = 7;

[0074] Step 3.2: The first LSTM layer calculates the vehicle shallow state feature matrix through the long short-term memory mechanism according to , where represents the vehicle shallow state feature at the nth time step, including the dynamic change information of all parameters in the input vector in the time dimension.

[0075] Step 3.3: The second LSTM layer calculates the vehicle deep state feature matrix through the long short-term memory mechanism according to , where represents the vehicle deep state feature at the nth time step, including the interaction information between all parameters in the input vector .

[0076] Step 3.4: The fully connected layer uses Equation (5) to obtain the output state parameter set , respectively represent the predicted values of the longitudinal and lateral vehicle speeds and the vehicle roll angle:

[0077] (5)

[0078] In Equation (5), are the two weight matrices of the fully connected layer, are the two bias vectors of the fully connected layer, and is the activation function.

[0079] Step 4: Use Equation (6) to construct the multi-objective loss function \(L\) of the LSTM neural network, including: estimation error, and dynamic constraints of longitudinal motion, lateral motion, and roll motion.

[0080] (6)

[0081] In Equation (6), is the estimation error, are the dynamic constraints of longitudinal motion, lateral motion, and roll motion, are three weights respectively, and there is:

[0082] (7)

[0083] In Equation (7), is the number of samples in the vehicle state feature dataset, is the longitudinal vehicle speed, lateral vehicle speed, and vehicle roll angle of the th sample, is the prediction value of the longitudinal vehicle speed, lateral vehicle speed, and vehicle roll angle of the th sample;

[0084] (8)

[0085] In Equation (8), is the time derivative of the predicted value of the longitudinal vehicle speed of the th sample , is the longitudinal acceleration and yaw rate of the th sample;

[0086] (9)

[0087] In Equation (9), is the time derivative of ;

[0088] (10)

[0089] In Equation (10), is the moment of inertia of the sprung mass about the x-axis, is 's first and second time derivatives, is the sprung mass, is the distance from the center of mass of the sprung mass to the roll center, is the equivalent roll damping of the vehicle suspension, is the equivalent roll stiffness of the vehicle suspension.

[0090] Step 5: Offline train the LSTM neural network according to the vehicle state feature dataset and the output vector until the multi-objective loss function converges, so as to obtain a vehicle state estimation model for estimating vehicle state parameters from the sensor data input in real time.

[0091] In this embodiment, an electronic device includes a memory and a processor. The memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0092] In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is run by a processor, it executes the steps of the above method.

Claims

1. A method for estimating state parameters of a distributed drive special vehicle based on hybrid data-driven, characterized in that The steps are as follows: Step 1: Construct an eight-degree-of-freedom dynamic model and a tire model of a distributed-drive special vehicle for designing the input and output vectors of the LSTM neural network; Step 2: Use dynamic software to construct a simulation environment under different working conditions, so as to collect data of the input vector in the simulation environment and obtain a vehicle state feature dataset of the input vector under different working conditions; Step 3: Construct an LSTM neural network, including: an input layer, two LSTM layers, a fully connected layer and a regression layer, and process the vehicle state feature dataset of the input vector under different working conditions to obtain the predicted value of the output vector; Step 4: Construct a multi-objective loss function of the LSTM neural network, including: estimation error, dynamic constraints of longitudinal motion, lateral motion and roll motion; Step 5: Offline train the LSTM neural network according to the vehicle state feature dataset and the output vector until the multi-objective loss function converges, so as to obtain a vehicle state estimation model for estimating vehicle state parameters from real-time input sensor data.

2. The state parameter estimation method of a distributed drive special vehicle based on hybrid data-driven according to claim 1, characterized in that The said Step 1 includes: Step 1.1: Establish an eight-degree-of-freedom dynamic model and a magic tire dynamic model for the longitudinal motion, lateral motion, yaw motion, roll motion and 4-wheel rotation of the distributed-drive special vehicle; Step 1.2: Infer the mapping relationship of each vehicle state parameter of the distributed-drive special vehicle based on the eight-degree-of-freedom dynamic model and the magic tire dynamic model, and determine the mapping function; Step 1.

3. Determine the input vector and output vector of the LSTM neural network using Equation (1) according to the mapping function , output vector : (1) In Equation (1), and are the longitudinal and lateral accelerations, is the yaw rate, is the steering wheel angle, is the road surface adhesion coefficient, is the driving torque of the wheel on the i-th side and j-th side, is the rotational speed of the wheel on the i-th side and j-th side, where i ∈ {f, b}, f and b represent the front and rear respectively, and j ∈ {l, r}, l and r represent the left and right respectively, and are the longitudinal and lateral vehicle speeds, is the vehicle roll angle.

3. A method for estimating state parameters of a distributed drive special vehicle based on a hybrid data-driven of LSTM and a dynamic model according to claim 2, characterized in that The said Step 2 includes: Step 2.1: Obtain the target speed curves under three working conditions respectively in the steady state, acceleration and deceleration simulation environments of the vehicle; Step 2.2: Obtain the steering wheel angles in the low-frequency and high-frequency ranges respectively under various driving conditions in a complex unstructured environment; Step 2.3: Obtain the adhesion coefficients under three road surface types respectively on dry asphalt road surface, wet asphalt road surface and snow-covered road surface; Step 2.4: After orderly combining the datasets in the three dimensions obtained in Steps 2.1 - 2.3, input them into the high-precision simulation dynamics model to obtain the input vector. The vehicle state feature datasets under different working conditions.

4. The state parameter estimation method for a distributed drive special vehicle based on hybrid data drive according to claim 3, wherein The said Step 3 includes: Step 3.

1. The input layer composes the input vector into a vehicle state feature matrix with a dimension of for the vehicle state feature data set under different working conditions, where is the length of each time step, h is the number of dimensions of the input vector , and h = 7; ​ Step 3.

2. The first LSTM layer calculates the vehicle shallow state feature matrix through the long short-term memory mechanism, where . Among them, represents the vehicle shallow state feature at the nth time step, including the dynamic change information of all parameters in the input vector in the time dimension; Step 4.

3. The second LSTM layer calculates the vehicle deep state feature matrix through the long short-term memory mechanism, where represents the vehicle deep state feature at the nth time step, including the interaction information between all parameters in the input vector ;​ Step 4.4: The fully connected layer obtains the output state parameter set by using Equation (2). , which respectively represent the predicted values of the longitudinal and lateral vehicle speeds and the vehicle roll angle. (2) In formula (2), are two weight matrices of the fully connected layer, are two bias vectors of the fully connected layer, is the activation function.

5. A method for estimating state parameters of a distributed drive special vehicle based on hybrid data-driven according to claim 4, characterized in that In the said Step 4, the multi-objective loss function L is constructed by using Equation (3); (3) In Equation (3), is the estimation error, are the dynamic constraints of longitudinal motion, lateral motion, and roll motion, are three weights respectively, and there is: (4) In formula (4), is the number of samples in the vehicle state feature dataset, is the longitudinal vehicle speed, lateral vehicle speed, and vehicle roll angle of the th sample, is the predicted value of the longitudinal vehicle speed, lateral vehicle speed, and vehicle roll angle of the th sample; (5) In formula (5), is the predicted value of the longitudinal vehicle speed of the th sample, and the time derivative of is the longitudinal acceleration and yaw rate of the th sample; (6) In formula (6), is the time derivative of; (7) In Equation (7), is the moment of inertia of the unsprung mass about the x-axis, is the first and second time derivatives of, is the unsprung mass, is the distance from the centroid of the unsprung mass to the roll center, is the equivalent roll damping of the vehicle suspension, is the equivalent roll stiffness of the vehicle suspension.

6. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program for supporting the processor to execute the vehicle state parameter estimation method according to any one of claims 1-5, and the processor is configured to execute the program stored in the memory.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program runs on the processor, it executes the steps of the vehicle state parameter estimation method according to any one of claims 1-5.

Images