A neural network-based data modeling method for multi-axle special vehicles

Abstract

The application discloses a kind of data modeling methods of multi-axis special vehicle based on neural network, including step 1.simplify multi-axis special vehicle dynamics model, establish single track model based on physical law;Step 2.based on the single track model based on physical law, refer to the recursive nature of physical model, establish the network model of closed loop structure capable of predicting the state variable of vehicle;Step 3.simulation training data set is generated using Trucksim simulation model;Step 4.closed loop network model is trained according to the data set obtained in step 3 in step 2, to obtain the closed loop network model after training;The method uses the network model of closed loop structure to carry out data modeling, reduces the demand of network to data, simultaneously constructs a network training method under closed loop condition, improves the generalization ability of closed loop network model, with the characteristics of high fidelity and generalization ability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle modeling, and particularly relates to a data modeling method for a multi-axle special vehicle based on a neural network. BACKGROUND

[0002] Multi-axle special vehicles usually run in harsh working conditions, and the structure of the vehicle itself is complex, so the model of the multi-axle special vehicle has strong nonlinear characteristics and is difficult to model accurately. However, accurately modeling vehicle dynamics is of great importance in vehicle control, state estimation and decision planning. Modeling for vehicle dynamics behavior is a process of abstracting the dynamic behavior of the vehicle into mathematical equations. According to different perspectives of cognitive vehicle dynamics, models can be divided into physical models and data models.

[0003] Most vehicle modeling is based on the physical modeling method of the first principle. The characteristic behavior of vehicle dynamics is established in the Newtonian mechanics system, and the high fidelity of the model depends on the accuracy of the dynamic equation and the physical parameters. However, a complete vehicle dynamics system is a complex strong nonlinear system. For example, when modeling the vehicle tire, the accuracy of the model depends more on the experimental parameters in the empirical model. This process requires a special experimental platform to test under specific experimental conditions. Therefore, when physically modeling the vehicle, it is very difficult and costly for researchers to obtain complete parameters about the vehicle. Vehicle dynamics modeling based on physical models has to be simplified according to the application requirements and the limitations of known model parameters, and different degrees of linearization assumptions are made, which makes it difficult for physical models to accurately reflect the real vehicle dynamics.

[0004] In recent years, deep learning has developed rapidly. Through offline training of big data, network models can fit any function to discover potential laws from data, and are widely used in intelligent decision-making, state prediction and trajectory planning in automatic driving. Data-driven network models rely on the big data generated by the modeling object, and the model optimizes the network parameters from the data. In the process of vehicle driving, the vehicle-mounted sensor can provide rich vehicle state information, providing a rich data set for establishing a vehicle dynamics neural network model, so some scholars have begun to use data-driven methods to model vehicle dynamics, such as:

[0005] In the literature "Pan Y, Nie X, Li Z, et al. Data-driven vehicle modeling of longitudinal dynamics based on a multibody model and deep neural networks [J]. Measurement, 2021, 180: 109541- ", Pan et al. proposed a neural network to establish a vehicle longitudinal dynamics model, the network input is the vehicle longitudinal speed and front wheel twist, the output is the vehicle braking distance; compared with the results of Carsim simulation test, the simulation results show that the network model can accurately predict the vehicle longitudinal speed and travel distance, the model can be applied to vehicle real-time simulation and control;

[0006] In the literature "RUTHERFORD S J, COLE D J 2010. Modelling nonlinear vehicle dynamics with neural networks. International Journal of Vehicle Design [J], 53: 260-287", Mauro et al. proposed a structured network model to model the vehicle longitudinal dynamics, the network input is the longitudinal speed, the change of the center of mass height, the throttle pedal and the gear change relationship, according to the different sources of vehicle longitudinal force components, the input information is classified, and a structured network is constructed, the simulation results show that the structured network model has better generalization, and improves the fidelity of data modeling;

[0007] With the deepening of vehicle modeling, in the literature "LIO M, BORTOLUZZI D, ROSATI G 2019. Modelling longitudinal vehicle dynamics with neural networks. Vehicle System Dynamics [J], 58: 1-19", Simon et al. used multilayer perceptron to establish a vehicle longitudinal dynamics model, and analyzed in detail the influence of network initialization weight and network size on network model generalization ability, the research found that the network weight initialization affects the final result of network training, with the increase of network size, it is helpful to improve the network convergence level;

[0008] In the literature "Cao X, Li H, Liu C, et al. Vehicle Longitudinal and Lateral Dynamics Modeling by Deep Neural Network [C] / / 2021 IEEE International Conference on Real-time Computing and Robotics (RCAR), 2021", Cao et al. constructed a hierarchical network for constructing a vehicle lateral dynamics model, the network model is divided into two networks, the first layer is used to predict the vehicle yaw rate, and the second layer combines the results of the first layer to predict the speed slip. However, due to the overall feedforward structure of the network, when designing the input of the network, the model needs to input the steering command, longitudinal acceleration, speed and historical yaw rate, which requires the network model to predict the state information, and also requires the corresponding sensors to collect the corresponding data during the actual measurement process. Excessive data requirements limit the practical application of data modeling;

[0009] In order to realize the recursive update of the network model, in the literature "Hermansdorfer L, Trauth R, Betz J, et al. End-to-End Neural Network for Vehicle Dynamics Modeling [C] / / 20206th IEEE Congress on Information Science and Technology (CiSt), 2020", Leonhard et al. use the trained network to continuously update the input, so that the network model can be driven like a physical model. Although the network realizes a closed-loop structure during testing, Leonhard et al. trained the network based on an open-loop network structure, and all network inputs used collected data. Although the network parameters obtained based on the open-loop structure can be deployed in the same network module, when the network model is tested, the network structure has been changed to a closed-loop, so when the closed-loop network module loads the model parameters under the open-loop structure, the model has low fidelity during testing.

[0010] Therefore, in order to improve the fidelity of the vehicle dynamics closed-loop network modeling, it is necessary to design a data modeling method for multi-axle special vehicles based on neural network to solve the problems existing in the prior art and improve the generalization ability of the closed-loop network model. SUMMARY

[0011] In view of the above problems, the present application aims to provide a neural network-based data modeling method for multi-axle special vehicles, which uses a closed-loop network model for data modeling, reduces the network's demand for data, and at the same time, constructs a network training method under closed-loop conditions, improves the generalization ability of the closed-loop network model, and has the characteristics of high fidelity and generalization ability.

[0012] To achieve the above-mentioned purpose, the technical solution adopted by the present application is as follows:

[0013] A neural network-based data modeling method for multi-axle special vehicles, comprising

[0014] Step 1. Simplify the multi-axle special vehicle dynamics model and establish a single-track model based on physical laws;

[0015] Step 2. Based on the recursive properties of the physical model, a network model with closed-loop structure capable of predicting the state variables of the vehicle is established on the basis of the single-track model;

[0016] Step 3. Use the Trucksim simulation model to generate a simulated training data set;

[0017] Step 4. Perform closed-loop training on the closed-loop network model established in Step 2 according to the data set obtained in Step 3 to obtain the trained closed-loop network model.

[0018] Preferably, the establishment process of the single-track model in Step 1 comprises

[0019] Step 1.1. Perform force analysis on the multi-axle special vehicle and establish a lateral dynamics model of the vehicle:

[0020]

[0021]

[0022] Wherein in formula (1) and formula (2), m is the mass of the whole vehicle; F i,lat is the tire lateral force of the i-th axle; I z is the moment of inertia; V x is the longitudinal speed of the vehicle; γ is the yaw rate of the vehicle; L i is the longitudinal distance from the axle center to the center of mass; V y is the lateral speed of the vehicle;

[0023] Step 1.2. Linearize the tire model shown in formula (1) and formula (2), and when the side slip angle is small, the tire model has an approximately linear relationship with the side slip angle as follows:

[0024] F i,lat (α)=C i α i (3)

[0025]

[0026]

[0027] wherein in formula (3), C i is the tire lateral stiffness under a specific vertical load, δ i is the tire steering angle;

[0028] Step 1.3. Assuming that the tire and the ground maintain a pure rolling contact relationship during driving, each wheel has only one steering center, and the modeling vehicle is except that the third axle is locked by a machine, the steering angle of the three-axle wheel is always 0, and the remaining tires can be freely steered, the steering angle of each wheel of the vehicle during driving is:

[0029]

[0030] δ3= 0 (7)

[0031]

[0032]

[0033] In the single-track model, each steering angle of the wheel is calculated by the steering wheel angle δ;

[0034] Step 1.4. Substituting formula (3) to formula (9) into formula (1) and formula (2), the linearized tire model-based lateral dynamics model is Euler forward expanded to obtain a discretized lateral dynamics model:

[0035] X(k+1) = AX(k) + BU(k) (10)

[0036] wherein in formula (10),

[0037] X(k+1) = [V y (k+1), γ(k+1)] T ,

[0038] X(k) = [V y (k), γ(k)] T ,

[0039]

[0040] δ(k) is the steering wheel angle of the control signal at time k, V y (k) and γ(k) are the lateral velocity and yaw rate predicted by the dynamic model at time k, and V x (k) is the longitudinal velocity of the vehicle collected at time k.

[0041] Preferably, the network model of the closed-loop structure in step 2 comprises a recurrent neural network module GRU and a fully connected layer FNN, and the recurrent neural network module GRU comprises a reset gate and an update gate, wherein

[0042] The initial state of the recurrent neural network module GRU is h0, and the input sequence of the recurrent neural network module GRU is xdrive network, wherein the xdrive network comprises a vehicle steering wheel angle δ, a longitudinal speed V x and a lateral speed V y and a yaw rate γ output by the fully connected layer FNN.

[0043] The input of the fully connected layer FNN is the output of the recurrent neural network module GRU, and the output of the fully connected layer FNN is a lateral speed V y and a yaw rate γ of the vehicle.

[0044] Preferably, the calculation process of the network model in step 2 comprises

[0045] Step 2.1. First, calculate the update gate z t , and the input vector x t is subjected to a linear transformation and multiplied by the weight matrix W z ; h0 is usually initialized as a zero vector, which is also subjected to a linear transformation and multiplied by the weight matrix U z ; add the two parts and input them into the activation function Sigmoid to compress the result to between 0 and 1, and the calculation equation is as shown in equation (11):

[0046] z t = σ(W z x t + U z h0) (11)

[0047] Wherein, x t is the input vector at the t-th time step, and h0 is initialized as a zero vector;

[0048] Step 2.2. Calculate the reset gate r t in the same way as the update gate, and the input data and the initialized zero vector input vector h0 are subjected to a linear transformation and then added, and the added result is also input into the activation function Sigmoid to compress the result to between 0 and 1, and the calculation equation is as shown in equation (12):

[0049] r t = σ(W r x t + U r h0) (12)

[0050] Step 2.3. New memory content h' in GRU unit i The output result contains the reset gate, and the specific calculation is shown in equation (13):

[0051] h' t = tanh(Wx t + r t ⊙ Uh t-1 ) (13)

[0052] The output h of GRU t The calculation method is shown in equation (14), which retains the information of the current unit and passes it to the next unit:

[0053] h t = z t ⊙ h0+ (1-z t )⊙ h t ’ (14)

[0054] Step 2.4. For the fully connected layer FNN, receive the feature vector h t output by GRU, and finally output the lateral kinematic state information, the calculation process is shown in equation (15):

[0055]

[0056] Preferably, the process of training the closed-loop network model in step 4 comprises

[0057] Step 4.1. Data normalization processing is performed on the data set obtained in step 3

[0058] For each type of input data of the network, the sample mean μ and sample standard deviation σ in the training set are calculated respectively, then converted according to equation (11), and the inverse normalization of the network output data is performed using equation (12) to map the actual state information, where

[0059]

[0060] x = σx * + μ (12)

[0061] Step 4.2. Set the loss function

[0062] The loss function is the mean square error, which measures the difference between the network output and the actual label, and adds the L2 regularization function, which changes the influence of the regularization term on the network by adjusting the size of the hyperparameter λ, and the loss function of the entire network is shown in equation (13):

[0063]

[0064] wherein in formula (13), J mse_loss represents a network loss function, m is the number of sample data in a single input network, Y neural networks is the output of the network model, Y label is the label value of the network output, λ is the penalty coefficient of the regularization term, W i represents the i-th weight, and N represents the total number of network weights.

[0065] Step 4.3. Set the dynamic learning rate update strategy based on RMSprop.

[0066] Step 4.4. Closed-loop training of the closed-loop network model.

[0067] Preferably, the process of the dynamic learning rate update strategy based on RMSprop in step 4.3 comprises

[0068] (1) Initialize the global learning rate ε, the decay rate ρ, the small constant δ, and the cumulative squared gradient γ.

[0069] wherein ε n+1 = κ·ε n , and κ is the initialization scaling coefficient.

[0070] (2) Obtain m training samples from the training set.

[0071] (3) Calculate the gradient g

[0072]

[0073] (4) Calculate the cumulative squared gradient γ

[0074] γ n+1 = ργ n + (1-ρ)g⊙g.

[0075] (5) Calculate the network parameter update

[0076]

[0077] (6) Network parameter update

[0078] θ n+1 = θ n -Δθ n .

[0079] The beneficial effects of the present application are: the present application discloses a data modeling method for multi-axis special vehicle based on neural network, compared with the prior art, the improvement of the present application is that:

[0080] Aiming at the lateral dynamic behavior of multi-axle special vehicle, the application proposes a data modeling method based on neural network; the network framework main body with closed loop structure is adopted, the state information of network output is used as input for predicting the next time state, realizing the recursive update of data modeling; for the closed loop network model, the training strategy of closed loop structure is designed, the intermediate variable is introduced in the network model, so that the network still maintains the closed loop structure in the training stage; the network module adopts the combination mode of loop gating unit and full connection network; the data set is generated by the Trucksim simulation model verified by real vehicle, and the analysis result shows that: the closed loop training method can make the network with closed loop structure have better fidelity, effectively improve the generalization ability of closed loop network model, and has the advantages of high fidelity and generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 The flow chart of the data modeling method of the multi-axle special vehicle based on neural network of the application.

[0082] Figure 2 The single rail model diagram of the application.

[0083] Figure 3 The physical model updating process diagram of the application.

[0084] Figure 4 The closed loop network model diagram of the application.

[0085] Figure 5 The experimental platform diagram of the application.

[0086] Figure 6 The experimental path curve diagram of the application.

[0087] Figure 7 The longitudinal velocity curve diagram of the application.

[0088] Figure 8 The steering wheel angle curve diagram of the application.

[0089] Figure 9 The lateral acceleration verification curve diagram of the application.

[0090] Figure 10 The yaw angular velocity verification curve diagram of the application.

[0091] Figure 11 The data set driving route diagram of the application.

[0092] Figure 12 The data set driving speed diagram of the application.

[0093] Figure 13 The training flow chart of the application.

[0094] Figure 14 Loss function variation comparison chart of the present application.

[0095] Figure 15 Lateral velocity prediction result comparison chart of the present application.

[0096] Figure 16 Yaw rate prediction result comparison chart of the present application.

[0097] Wherein: in Figure 13 In the figure (a) is the training flow chart of the closed-loop training method of the present application, and the figure (b) is the training flow chart of the open-loop training method of the prior art. DETAILED DESCRIPTION

[0098] In order to enable those skilled in the art to better understand the technical solutions of the present application, the technical solutions of the present application will be further described below in combination with the drawings and examples.

[0099] Example 1: refer to the attached Figures 1-16 A neural network-based data modeling method for multi-axle special vehicles is provided, taking a certain type of special five-axle vehicle as an example, and the specific process includes

[0100] Step 1. Simplify the multi-axle special vehicle dynamics model, and establish a single-track model based on physical laws, which includes

[0101] Step 1.1. The single-track model is a commonly used simplified vehicle dynamics model, as shown in Figure 2 After force analysis of the vehicle, the lateral dynamics model of the vehicle is established as shown in formula (1) and formula (2):

[0102]

[0103]

[0104] Wherein in formula (1) and formula (2), m is the mass of the vehicle; F i,lat is the tire lateral force of the i-th axle; I z is the moment of inertia; V x is the longitudinal speed of the vehicle; γ is the yaw rate of the vehicle; L i is the longitudinal distance from the axle center to the center of mass; V y is the lateral speed of the vehicle;

[0105] Step 1.2. Linearize the tire model shown in formula (1) and formula (2). When the side slip angle is small, the tire model has an approximately linear relationship with the side slip angle as shown in formula (3)-(5):

[0106] F i,lat (α)=C i α i(3)

[0107]

[0108]

[0109] wherein C i is the tire lateral stiffness under specific vertical load, and δ i is the tire steering angle in formula (4) and formula (5) ;

[0110] Step 1.3. Assuming that the tire and the ground maintain a pure rolling contact relationship during driving, at this time each wheel has only one steering center, and the modeling vehicle is except that the third axle is locked by a mechanical lock, the steering angle of the three-axle wheel is always 0, and the remaining tires can be freely steered, so the steering angle of each wheel during driving of the vehicle is:

[0111]

[0112] δ3=0 (7)

[0113]

[0114]

[0115] Therefore, in the single-track model, each steering angle of the wheel can be calculated using the steering wheel angle δ;

[0116] Step 1.4. Substituting formula (3) to formula (9) into formula (1) and formula (2), based on the linearized tire model, the lateral dynamics model is Euler forward expanded to obtain the discretized lateral dynamics model, the structure is shown in formula (10) :

[0117] X(k+1)=AX(k)+BU(k) (10)

[0118] wherein in formula (10),

[0119] X(k+1)=[V y (k+1),γ(k+1)] T ,

[0120] X(k)=[V y (k),γ(k)] T ,

[0121]

[0122] wherein δ(k) is the steering wheel angle of the control signal at time k, V y (k) and γ(k) are the lateral velocity and yaw rate predicted by the dynamic model at time k, V x(k) is the longitudinal velocity of the vehicle collected at time k;

[0123] Step 2. Based on the physical law-based monorail model shown in Equation (10), establish a closed-loop network model that can continuously update and predict the state variables of the vehicle. The specific process includes...

[0124] Step 2.1. Equation (10) shows that the physical model is a recursive model, and the recursive process is as follows: Figure 3 As shown, by initializing the model's predictive variables and then inputting control signals and necessary system variables into the model, the model can continuously update and predict the vehicle's state variables; in the lateral dynamics model, the steering wheel angle δ and the longitudinal velocity V x These are the control signals and system state variables input to the model, including the lateral velocity V. y The yaw rate γ is a predictable state variable in the model;

[0125] Step 2.2. Therefore, to implement the recursive update function in the physical model, a system is built as follows: Figure 4 The closed-loop network model shown maps the lateral dynamics of the vehicle. In the closed-loop network model, the model structure is connected end-to-end, and the network outputs the lateral velocity V. y And the yaw rate γ, which is also used as input information to the network along with the vehicle steering wheel angle δ and longitudinal velocity V. x The information together forms a new input sequence x to drive the network, which predicts the vehicle's lateral dynamics state information at the next time step. Figure 4 In this context, h0 represents the initial state of the GRU module, which is usually set to a sequence of 0s;

[0126] pass Figure 4 We can see that the closed-loop network model consists of a recursive network module GRU and a fully connected layer FNN; the main parameters of the recursive network module GRU and the fully connected layer FNN are shown in Table 1 and Table 2.

[0127] Table 1 GRU Module Parameters

[0128]

[0129] Table 2 FNN Module Parameters

[0130]

[0131] like Figure 4 As shown, the recursive network module GRU described in this embodiment has two gates: a reset gate and an update gate; x t Given the input vector at time step t, the specific process of computation for the network model includes:

[0132] Step 2.1. First, calculate the update gate z t , the input vector x t goes through a linear transformation and multiplies with the weight matrix W z ; h0is usually initialized as a zero vector, which also goes through a linear transformation and multiplies with the weight matrix U z ; add the two parts and input into the activation function Sigmoid to compress the result to 0 to 1, the calculation equation is shown in (11):

[0133] z t = σ(W z x t + U z h0) (11)

[0134] where x t is the input vector of the t-th time step, and h0is initialized as a zero vector;

[0135] Step 2.2. Reset gate r t is calculated in the same way as the update gate, the input data and the initialized zero vector input vector h0go through a linear transformation and then add, and the result of the addition is also input into the activation function Sigmoid to compress the result to 0 to 1, the calculation equation is shown in (12):

[0136] r t = σ(W r x t + U r h0) (12)

[0137] Step 2.3. New memory content h' t in the GRU unit contains the output result of the reset gate, and the specific calculation is shown in equation (13):

[0138] h' t = tanh(Wx t + r t ⊙ Uh t-1 ) (13)

[0139] The output h t of the GRU is calculated as shown in equation (14), which retains the information of the current unit and passes it to the next unit:

[0140] h t = z t ⊙ h0+ (1-z t ) ⊙ h t ’ (14)

[0141] Step 2.4. For the fully connected layer FNN, the feature vector h output by the recurrent network module GRU is received t The final output lateral kinematics state information is calculated as shown in equation (15):

[0142]

[0143] Step 3. A simulation training data set is generated using the Trucksim simulation model, and the specific process includes

[0144] Step 3.1. The simulation model of the target vehicle is built using Trucksim, and the simulation model is experimentally verified: data modeling relies on large data sets, but the cost of obtaining data through experiments is relatively high; therefore, a simulation model of the target vehicle is built using Trucksim to generate a data set, and a turning condition under the condition of real vehicle experiment is used to verify the accuracy of the Trucksim simulation model, as shown in the experimental platform Figure 5 The steering wheel angle sensor uses a UX-2D type steering wheel rotation torque detector; the Speedbox-mini can be used to collect vehicle driving process longitudinal speed, longitudinal acceleration, lateral acceleration and other information; the GPS and INS antennas are placed on the roof to measure vehicle attitude angle and angular velocity information; the sensor signals are output to the data acquisition system through the CAN bus, and Race Technology and DEWEsoft-X3 are used for post-processing of experimental data;

[0145] The experimental path is shown in Figure 6 The vehicle speed varies between 20 km / h and 40 km / h, and the generated data set is used to model the lateral dynamics, so the lateral acceleration and yaw rate are selected as the indicators for verifying the Trucksim model; the vehicle longitudinal speed information and steering wheel angle information shown in Figure 7 and Figure 8 are input into Trucksim; the lateral acceleration and yaw rate simulation results are compared with the experimental data, and the comparison results are shown in Figure 9 Figure 10 The experimental data and simulation data are basically consistent;

[0146] Figure 9 Figure 10 ​Further analysis of the results, the experimental test data and TruckSim model output data error comparison results shown in Table 3, the lateral acceleration root mean square error is 0.0023, the maximum error is 0.0735g; The root mean square error of yaw rate is 0.0359, and the maximum error is 0.3006 deg / s; Therefore, the simulation data and experimental data are basically the same, which verifies that the simulation model built in Trucksim has good fidelity in reflecting the lateral dynamic characteristics of the vehicle, and the main physical parameters of the simulation model are shown in Table 4;

[0147] Table 3: Trucksim model error

[0148]

[0149] Table 4: Trucksim simulation vehicle parameters

[0150]

[0151]

[0152] Step 3.2. Training conditions

[0153] The training set is used to train the network model offline, and the generalization ability and model accuracy of the network model are closely related to the data set; When making the data set, a random driving route is set as shown in Figure 4 , the initial position of the vehicle is Figure 11 , the origin position, and the vehicle advances along the curve stretching direction, the vehicle longitudinal speed during driving changes between 30km / h and 40km / h, the change curve is shown in Figure 12 ; The data set is divided into two parts according to the driving process, of which the solid line path is used to make the training set, and the dotted line part is used to make the test set; The data information includes vehicle longitudinal speed, steering wheel angle, yaw rate and vehicle lateral speed, which is obtained by Simulink and Trucksim joint simulation, and the sampling frequency is set to 100Hz;

[0154] Step 4. Train the closed-loop network model established in step 2 according to the data set obtained in step 3, the specific process includes

[0155] Step 4.1. Data normalization processing is performed on the data set obtained in step 3

[0156] In deep learning, the normalization of training data (training set) can improve the convergence speed and model accuracy of training, so that the network avoids gradient explosion in the training process; at the same time, it is beneficial to ensure that the influence of each input state on the final output result of the network is equal; the commonly used normalization method is Max-Min normalization; in order to reduce the influence of the difference between the maximum and minimum values in the training set and the test set, Standard normalization is used when preprocessing data; for each type of input data of the network, the sample mean μ and sample standard deviation σ in the training set are calculated respectively, and then converted according to formula (11); since the data set used by the network in the training process is normalized, the network output data needs to be inverse normalized as shown in formula (12) when compared with the simulation data, so as to map the actual state information:

[0157]

[0158] x = σx * + μ (12)

[0159] Step 4.2. Set the loss function

[0160] The loss function is a standard for measuring the generalization ability of the network model. In order to avoid the phenomenon of overfitting of the network model in the training process, which leads to the reduction of the generalization ability of the model, adding a network weight regularization function to the loss function is a common overfitting avoidance technique; when training the network, the mean square error is used as the loss function to measure the difference between the network output and the actual label, and the L2 regularization function is added to limit the weight to remain in a small value range. By adjusting the size of the hyperparameter λ, the influence of the regularization term on the network can be changed, so that the network has better generalization ability. The loss function of the whole network is shown in formula (13):

[0161]

[0162] In formula (13), J mse_loss represents the network loss function, m is the number of sample data in a single input network, Y neural networks is the output of the network model, Y label is the label value of the network output, λ is the penalty coefficient of the regularization term, which is 0.001, W i represents the i-th weight, and N represents the total number of network weights.

[0163] Step 4.3. Set the dynamic learning rate update strategy based on RMSprop

[0164] In the network backpropagation process, the RMSprop adaptive learning rate optimizer is used to calculate the gradient of updating the model parameters, and the parameter update process is shown in Table 5;

[0165] First, the basic parameters learning rate ε, decay rate ρ and small constant δ are initialized, where the small constant δ is used to avoid the divergence of parameter update process caused by the accumulated square gradient being 0 when calculating the parameter update Δθ n ; As an adaptive optimizer, RMSprop uses the hyperparameter decay rate ρ to weaken the influence of the accumulated gradient when calculating the gradient, so that the parameter update process is not reduced due to the increase of the number of iterations, and by setting a reasonable hyperparameter ρ, the gradient disappearance during network training can be avoided.

[0166] Table 5: RMSprop optimizer update strategy

[0167]

[0168] In the training of the network, with the increase of the number of training cycles, the constantly reduced learning rate can make the network achieve better training effect; therefore, on the basis of RMSprop, a dynamic adjustment strategy is adopted for the global learning rate, and the global learning rate is reduced to 0.6 of the original after every 100 cycles of network training, and the pseudo code for dynamic updating of the global learning rate is shown in Table 6:

[0169] Table 6: Dynamic updating strategy of global learning rate

[0170]

[0171] Step 4.4. Closed-loop training of closed-loop network model

[0172] Generally, the network training process and the structure of the network model are in an open-loop structure, the network input and output are not connected in a closed loop, and the input data are all from the collected sample data; the training process is as shown in the solid line part, at this time the network model is in an open-loop structure. Figure 13

[0173] However, for horizontal dynamics modeling, a network model with a closed-loop structure needs to be established; if the model parameters are still trained in an open-loop manner, the network structure must be destroyed so that the network is trained under open-loop conditions; in order to construct a training method suitable for closed-loop networks, the data information of the state variables lateral velocity and yaw rate are not provided in the input process of the network, but intermediate states are introduced to replace them in the training of the network model; before the network iteration, the intermediate state variables lateral velocity and yaw rate are initialized to 0, and the collected longitudinal velocity and steering wheel angle information are input into the network together, and the data labels are the collected yaw rate and lateral velocity; when starting training, the forward propagation is performed in turn, the loss function is calculated, the parameters are updated in reverse, and the state information predicted by the network forward propagation is used to update the intermediate state variables, and the longitudinal velocity and steering angle information in the data set are used as new input sequences for the next batch of training.​

[0174] Step 4.5. Training convergence comparison

[0175] With Figure 13 (b) shows the open-loop training method, the loss function value trend of the two groups of training methods is shown in Figure 14 ; it can be seen that after 500 cycles of training, the network of the two groups of methods can converge to a stable level; under the condition of open-loop training, the loss function finally converges to 0.00053; compared with the closed-loop condition, the loss function converges to a lower level; Figure 14

[0176] Step 5. Model generalization analysis process, which includes

[0177] In the test set test phase, compared with the open-loop training method of Figure 13 (b), the results are shown in Figure 15 and Figure 16 , the comparison standard takes Trucksim simulation data as the reference value; the prediction of lateral velocity is shown in Figure 15 , the trend of the prediction results of the two groups of models is consistent with the simulation data, but at the vehicle turning point, the network model obtained by closed-loop training has higher fidelity; error analysis of the prediction results of the two groups can be obtained, and the error comparison results are shown in Table 7, the maximum error of the model under closed-loop training is 0.0785km / h, and the root mean square error is 0.0083, which is reduced by 24.01% and 59.71% compared with the model error of open-loop training;

[0178] Table 7: Error comparison of lateral velocity prediction results

[0179]

[0180] The comparison results of yaw rate are shown in Figure 15 , the model obtained by closed-loop training is closer to the simulation data of Trucksim, and the error comparison analysis of the prediction results of the two groups of models is shown in Table 8; the maximum error of the prediction results of the model obtained by closed-loop training is 0.3418deg / s, and the root mean square error is 0.0359, which is reduced by 49.49% and 77.74% compared with the model error of open-loop training;

[0181] Table 8: Error comparison of yaw rate prediction results

[0182]

[0183] ​By comparing the prediction results of lateral velocity and yaw rate in the test set, the closed-loop training can make the closed-loop network model have higher fidelity; although in the network training stage, the open-loop training can make the loss function reduce to a lower level; but the network is in an open-loop structure during the training process, and the input of the network is all from the collected data; however, in the test stage, the output data of the network is used for the input of the next moment to realize the recursive updating capability of the network; but when the predicted result is inconsistent with the actual state, the input of the collected data into the network is equivalent to inputting the wrong data into the network, which reduces the fidelity of the model.

[0184] For the network model of the closed-loop structure, the embodiment proposes a closed-loop training method, which introduces an intermediate state variable, so that the network is still in a closed-loop structure during training, and the network can be trained under the condition of only inputting longitudinal velocity and lateral velocity information, thereby ensuring the consistency of the model structure and input data of the closed-loop network model in the training stage and the test stage. Therefore, the closed-loop training described in the embodiment can make the closed-loop network have better fidelity.

[0185] Therefore, for the problem that the dynamics model of the special vehicle is difficult to accurately model, the embodiment builds a network model of a closed-loop structure based on a data modeling method to represent the lateral dynamics of the vehicle, designs a training method suitable for the closed-loop network as described above, compares the influence of different network modules on the generalization ability of the closed-loop network model, and draws the following conclusions:

[0186] 1. The closed-loop network model built in the embodiment reduces the demand of the network for input data, and the network model only needs longitudinal velocity and lateral velocity information to predict the lateral dynamics motion state of the vehicle during the test process, and the network model can be driven like a physical model;

[0187] 2. When training the network model of the closed-loop structure, by introducing the method of the intermediate state variable, the network iteratively updates the information of the intermediate state variable during the training process, so that the model can be trained under the condition of only inputting longitudinal velocity and steering wheel angle information, and the model always maintains a closed-loop structure. Compared with the open-loop training in the prior art, the closed-loop training method designed in the embodiment effectively improves the model precision.

[0188] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for data modeling of multi-axle specialty vehicles based on neural networks, characterized by: Comprising Step 1. Simplify the multi-axle special vehicle dynamics model, and establish a single-track model based on physical laws; Step 2. On the basis of the single-track model, refer to the recursive nature of the physical model, and establish a network model capable of predicting the closed-loop structure of the state variables of the vehicle; Step 3. Use the Trucksim simulation model to generate a simulation training dataset; Step 4. According to the dataset obtained in step 3, the closed-loop network model established in step 2 is closed-loop trained to obtain a trained closed-loop network model; The network model with a closed-loop structure in step 2 comprises a GRU recursive network module and a fully connected layer FNN, and the GRU recursive network module comprises a reset gate and an update gate, wherein: The initial state of the recursive network module GRU is as follows: And the input sequence of the recursive network module GRU is The driving network, the Drive network includes vehicle steering wheel angle Longitudinal velocity Lateral velocity output by the fully connected FNN layer and yaw rate ; The input of the full connection layer FNN is the output of the recurrent network module GRU, and the output of the full connection layer FNN is the lateral speed of the vehicle and the yaw angular velocity ; The calculation process of the network model in step 2 comprises Step 2.

1. First, calculate the update gate , the input vector goes through a linear transformation and is multiplied by the weight matrix ; The hidden state, usually initialized as a zero vector, also goes through a linear transformation and is multiplied by the weight matrix ; the two parts are added together and input into the activation function Sigmoid, which compresses the result to between 0 and 1, and the calculation equation is shown in equation (11): （11） wherein, is the input vector for the t-th time step, is initialized to a zero vector; Step 2.

2. Reset the door The calculation method is the same as the update door, and the input data is input to the vector After linear transformation and addition, the same result of addition is input to the activation function Sigmoid, and the result is compressed to between 0 and 1. The calculation equation is shown in equation (12): （12） Step 2.

3. New memory content in the GRU cell of the recurrent network module The output result contains the reset gate, which is calculated as shown in equation (13): （13） The output of the recurrent network module GRU The computational method is shown in equation (14), which preserves the information of the current cell and passes it to the next cell: （14） Step 2.

4. For the fully connected layer FNN, receive the feature vector output by the recurrent network module GRU The final output lateral kinematic state information is calculated as shown in equation (15): （15）。 2. The data modeling method for multi-axle special vehicle based on neural network according to claim 1, characterized in that: The establishment process of the single-track model in step 1 comprises Step 1.

1. Force analysis is performed on the multi-axle special vehicle to establish a lateral dynamics model of the vehicle: （1） （2） wherein in formula (1) and formula (2), is the total vehicle mass; is the first is the lateral force of the tire of the axle; is the moment of inertia; is the vehicle longitudinal speed; is the vehicle yaw rate; is the longitudinal distance from the axle center to the center of mass; is the lateral speed of the vehicle; Step 1.

2. Linearize the tire model shown in formula (1) and formula (2), and when the side slip angle is small, the tire model has an approximately linear relationship with the side slip angle as shown in formula (3): （3） （4） （5） wherein in formula (3), is the tire lateral stiffness under a specific vertical load, formula (4) and formula (5) in is the tire cornering angle; Step 1.

3. Assume that the tire and the ground maintain a pure rolling contact relationship during driving, and each wheel has only one steering center. The modeling vehicle has a third axle that is mechanically locked, and the steering angle of the three-axle wheel is always 0. The remaining tires can freely steer, and the steering angle of each wheel during vehicle driving is as shown in formula (4): （6） （7） （8） （9） In the single-track model, each wheel steering angle is given by the steering wheel angle calculation; Step 1.

4. Substitute formula (3) to formula (9) into formula (1) and formula (2), and perform Euler forward expansion on the lateral dynamics model based on the linearized tire model to obtain a discretized lateral dynamics model as shown in formula (10): （10） In formula (10), ， ， ， , is steering wheel angle at time t, and is lateral velocity and yaw rate predicted by the power model at time t, is vehicle longitudinal velocity collected at time t.

3. The data modeling method for multi-axle special vehicle based on neural network according to claim 1, characterized in that: The process of training the closed-loop network model in step 4 comprises Step 4.

1. Data normalization is performed on the dataset obtained in step 3 For each type of input data of the network, the sample mean in its training set is calculated and the sample standard deviation Then, according to equation (11), the data output by the network is inversely normalized to map the actual state information, where （11） （12） Step 4.

2. Set the loss function The loss function is measured by the mean square error between the network output and the actual label, and an L2 regularization function is added, and the size of the regularization term changes the influence of the network by adjusting the hyperparameter , the loss function of the whole network is shown in equation (13): （13） wherein in formula (13), represents a network loss function, m is the number of sample data in a single input network, is the output of the network model, is the label value of the network output, is a penalty coefficient of a regularization term, and is 0.001, represents the first weight, and N represents the total number of weights of the network. Step 4.

3. Set the dynamic learning rate update strategy based on RMSprop; Step 4.

4. Closed-loop train the closed-loop network model.

4. The data modeling method for multi-axle special vehicle based on neural network according to claim 3, characterized in that: The process of the dynamic learning rate update strategy based on RMSprop in step 4.3 comprises (1) initialize global learning rate , decay rate , small constant , and accumulated squared gradient ; wherein , is an initialization scaling factor; (2) Obtain m training samples from the training set; (3) calculating the gradient ; ； (4) Calculate cumulative squared gradient ; ； (5) Calculate the network parameter update ； (6) Network parameter update 。

Images