1. A vehicle dynamics prediction model based on a time-lag feedback neural network is characterized in that a long-time memory neural network is used for designing a neural network vehicle dynamics prediction model with time-delay output feedback, all unknown or unmodeled vehicle dynamics change effects are arranged in a data set, and the model can learn the tire nonlinear effect of a vehicle and potential unknown dynamics state changes including load transfer.

2. The vehicle dynamics prediction model based on the time-lag feedback neural network as claimed in claim 1, wherein the concrete structure of the neural network vehicle dynamics prediction model is as follows:

the first layer is an input layer which has 7 characteristic inputs, namely a yaw angular velocity r and a transverse velocity U_yLongitudinal speed U_xFront wheel steering angle delta, longitudinal resultant force F of vehicle front axle tire_xfThe output of the sum network model becomes the derivative output value of the yaw rate and the transverse speed of the input information through time delay feedbackThe data of each input feature comprises vehicle dynamics information of 4 time steps; the second layer is an LSTM1 network layer, and the hidden layer is designed to have 128 hidden units; the third layer is an activation layer, and the activation function is selected as a Relu function; the fourth layer is an LSTM2 network layer, the hidden layer is designed to have 128 hidden units and only outputs the information after the last LSTM-CELL calculation; the fifth layer is a regression output layer, and the hidden layer is designed to have 2 hidden units.

3. The vehicle dynamics prediction model based on the time-lag feedback neural network is characterized in that the vehicle dynamics prediction model adopts a two-layer long-and-short time memory neural network structure, and can predict the derivative of the current vehicle yaw velocity and the current lateral velocity by including vehicle control and state information with 4 time step states in input data, particularly state output parameters including time-lag feedback in the state input data; the forward calculation method of the neural network vehicle dynamics prediction model is as follows:

h＝{x_t,…,x_t-T}

a_l＝max(0,z₁)

wherein x is_tRepresenting delayed feedback status output, control and status input information in a single time step, h representing x comprising a plurality of historical time step information_tData, W_lstm{1,2}∈(w_i,w_f,w_g,w_o),b_lstm{1，2}∈(b_i,b_f,b_g,b_o) Represents the learned network weights and bias parameters, w, of the 2 LSTM network layers_i,w_f,w_g,wo,b_i,b_f,b_g,b_oRespectively representing the weight matrixes of an input gate layer, a forgetting gate layer, a Tanh layer and an output gate layer in the LSTM network and the bias matrixes of the input gate layer, the forgetting gate layer, the Tanh layer and the output gate layer;to output the regression layer weight transpose matrix, b_FCIs the bias matrix of the output regression layer; in the equation of the network model, F_lstmIs an abbreviation of LSTM network model. a _ l represents an active layer, z_iI is 1,2 represents the weighted output of different network layers; predicted output of networkAndis defined as:andwhere Δ t is 50ms, which is the sampling frequency of the signal.

4. A training data acquisition method of a vehicle dynamics prediction model based on a time-lag feedback neural network is characterized by comprising a virtual data set and a real data set; the virtual data set includes: the low fidelity can explain the multi-time-step virtual data acquisition of the nonlinear dynamics of the vehicle and the multi-time-step virtual data acquisition of high fidelity vehicle dynamics software CarSim.

5. The method for acquiring the training data of the vehicle dynamics prediction model based on the time-lag feedback neural network as claimed in claim 4, wherein the method for acquiring the low-fidelity interpretable vehicle nonlinear dynamics multi-time-step virtual data comprises the following steps:

analyzing the complex unmanned vehicle according to a Newton second law to obtain a stress balance equation along an x axis, a y axis and a z axis, and designing a vehicle nonlinear dynamic model, wherein the unmanned vehicle (front wheel drive and steering) has translation in longitudinal, transverse and vertical directions and rotation in 3 directions of rolling, pitching and yawing; wherein the lateral and yaw motions are substantially produced by the steering maneuver; the nonlinear vehicle dynamics model is therefore represented by a differential equation as:

wherein U is the speed at the vehicle's center of mass; u shape_x、U_yThe speeds of the vehicle mass center along the x and y directions of the vehicle body coordinate system are respectively; alpha is alpha_f,α_rRespectively are front and rear wheel side deflection angles; beta is the vehicle mass center slip angle;r is the vehicle yaw rate; a and b are the distances from the mass center of the vehicle to the front and rear axes; m is the vehicle mass, I_zThe moment of inertia of the vehicle around the z-axis of the center of mass; f_yf,F_yrThe lateral resultant force of the front axle tire and the rear axle tire is respectively; f_xfThe longitudinal resultant force received by the front axle tire; delta is a front wheel corner;

aiming at the nonlinear characteristics generated in the driving process of the vehicle under different road conditions, which are caused by the tire when the tire turns, the Fiala model of the tire is introduced for expanding the application range of the vehicle model, and the lateral force F of the tire is introduced_yThe calculation formula of (2) is as follows:

wherein C is_αAnd μ is the tire cornering stiffness and road adhesion coefficient; f_zIs the tire vertical load; α is the tire slip angle; alpha is alpha_satIs the tire saturated slip angle; the formula for calculating the slip angle of the front tire and the rear tire is as follows:

when the vehicle runs at a high speed, the weight transmission can increase or reduce the magnitude of the vertical force borne by the tire, so that the magnitude of the lateral force of the tire is influenced; l is the vehicle wheel base, wherein the vertical force calculation formula of the front axle and the rear axle is as follows:

the delay of the lateral force experienced by each tire when the vehicle is traveling at low speeds is modeled by the tire slack length, and the tire delay is characterized by a front and rear tire slip angle first derivativeThe calculation formula is as follows:

wherein σ_f,σ_rThe front and rear tire slack lengths.

6. The method for obtaining training data of a time-lag feedback neural network-based vehicle dynamics prediction model according to claim 5, further comprising performing theoretical numerical limits on low-fidelity interpretable vehicle nonlinear dynamics model inputs, the theoretical numerical limits being as follows:

U_x∈(1m/s，33m/s)

δ∈(-25°，25°)

F_x∈(-μmg,power_lim/U_x)

r∈(-μg/U_x,μg/U_x)

U_y∈(3aμmgU_x/LC_α+bμg/U_x,-3aμmgU_x/LC_α-bμg/U_x)

wherein mu is a road adhesion coefficient, and power _ lim is the maximum power of the vehicle;

randomly sampling data of a low-fidelity interpretable vehicle nonlinear dynamic model to obtainThe output tag signal is subjected to Euler integration processing on the output signal and fed back to a random sampling signal input end to be used as a part of input signals at the next moment;

and finally, performing multi-time-step processing on the input and output signals to obtain a low-fidelity interpretable vehicle nonlinear dynamics multi-time-step virtual data set.

7. The method for acquiring the training data of the vehicle dynamics prediction model based on the time-lag feedback neural network as claimed in claim 4, wherein the high fidelity vehicle dynamics software CarSim multi-time-step virtual data acquisition comprises the following steps:

s1, setting a road transverse and longitudinal section, an adhesion coefficient and vehicle parameters in CarSim high-fidelity vehicle dynamics simulation software, specifically selecting the vehicle type grade, the length, the width, the height, the wheelbase, the minimum ground clearance, the trim mass and the finished vehicle rotational inertia;

s2, analyzing the limit relation between the longitudinal vehicle speed and the front wheel steering angle: analyzing the limit relation between the longitudinal vehicle speed and the front wheel steering angle by using a minimum total variance method, and further determining a reasonable threshold value of the vehicle control quantity;

s3 selecting random inputs with uniform distribution to ensure successful training of neural network vehicle dynamics models, sine waves with different frequencies and amplitudes are added as input signals u (t) in order to create continuous random input signals, the formula being:

where d is the total number of sinusoids, a_mIs the maximum amplitude of the input, ω_mIs the maximum frequency of the input,. phi.is the initial phase angle of the sine wave, j represents the jth sine wave；

S4 selects the appropriate phase of the signal to ensure that the power of the signal is evenly distributed over these frequencies according to the schroeder phase equation:

wherein phi₁Is an arbitrary value; the total number of sinusoids should be as large as possible, the maximum amplitude of the signal being the sum of the individual amplitudes, the maximum frequency being determined by the maximum frequency of the individual sinusoids;

s5 selects d as 100, max frequency ω_mMaximum amplitude a of the sum of the combined sinusoidal steering inputs, 2Hz_mDetermining according to a limit relation between the longitudinal speed of the vehicle and the front wheel steering angle;

s6 performs joint simulation on the combined sinusoidal input of the vehicle model in the high-fidelity vehicle dynamics software CarSim under the limit relation between the longitudinal vehicle speed and the front wheel steering angle through the Matlab/Simulink to obtain continuous output response signals, performs multi-time-step processing on the output signals, and finally obtains a multi-time-step virtual data set of the high-fidelity vehicle dynamics software CarSim.

8. The method for acquiring the training data of the time-lag feedback neural network-based vehicle dynamics prediction model according to claim 7, wherein the specific process of S2 includes:

assuming that the front wheel steering input control quantity is x (t), the lateral dynamic response is y (t), and the total variance of the vehicle dynamic response is shown in formula (7):

in the formula, x₀，y₀Steady state values for input x (t) and response y (t),x₀is a constant; setting a threshold E of the total variance of the dynamic response of the vehicle to ensure the safety and the operation stability of the vehicle₀E is less than or equal to E₀；

E is set according to the motion condition of the vehicle and by combining the operation experience that a driver can ensure safe and stable running during emergency steering of the vehicle₀Threshold values, whereby the front wheel steering angle input is calculated:

wherein when the longitudinal speed is 20m/s, E is taken₀0.15, the maximum safe front wheel angle was calculated to be 3.

9. The method for acquiring the training data of the time-lag feedback neural network-based vehicle dynamics prediction model according to claim 4, wherein the method for acquiring the real data set comprises the following steps:

the unmanned vehicle is provided with a wheel force sensor, an S-Motion DTI sensor, an MSW DTI sensor and an integrated navigation system (GPS global positioning system + inertial navigation), and receives data from a vehicle CAN bus; operating the unmanned vehicle to perform various vehicle tests such as a drift test, a snake test, a single-shift test, a sine frequency sweep steering test, a steady-state rotation test, a quasi-static linear acceleration and deceleration test and an ISO double-shift running test under a dry asphalt pavement, a wet-skid silt pavement and an ice-snow pavement, and acquiring actual unmanned vehicle dynamic data under high, medium and low pavement adhesion coefficients; smoothing the collected data by using a second-order Butterworth low-pass filter with the cutoff frequency of 3Hz to filter the influence of high-frequency behaviors on vehicle dynamics; and finally, carrying out data synchronization and multi-time-step processing to obtain a real dynamic data set of the actual unmanned vehicle.

10. A training method of a vehicle dynamics prediction model based on a time-lag feedback neural network is characterized by comprising two stages of training, specifically comprising the following steps:

dividing the virtual data set and the real data set into a training set of 70%, a verification set of 15% and a test set of 15%; randomizing the data to break time correlation of the data set, ensuring that each data sample consists of a time-related state trajectory, and no correlation exists between any given two data samples; the Loss function is selected as a mean square error, the optimizer is selected as Adam, the batch size is set to be 1000, the learning rate is set to be 0.0001, the learning framework based on Pythroch performs learning optimization training on the network model, and the optimization training algorithm is as follows:

wherein N is the total number of training samples,is the derivative truth-label data of the vehicle yaw rate and lateral velocity at the next time,predicting vehicle yaw rate and derivative output values of lateral velocity for a network

In the first stage, the network learns by using a vehicle dynamics virtual data set to obtain a pre-training model, and carries out generalization test on unknown data; inputting 15% of test set data into network for calculation, and obtaining predicted output value of networkTagging with test set dataCalculating a Mean Square Error (MSE);

the generalized test judgment conditions are as follows: if it is notThe generalization test passes;

after the test is passed, performing model weight optimization of a second stage by using a vehicle dynamics real data set, namely loading a pre-training model obtained by using a virtual data set in the first stage, a training set of the real data set and data in a verification set, and performing learning training on the network model by using the optimization training algorithm again; finally, a vehicle dynamics prediction model based on a time-lag feedback neural network and attached to an actual unmanned vehicle is obtained;

the virtual dataset, the real dataset are obtained by the method of claims 4-9.

Background

With the continuous improvement of the requirements of drivers on the safety, the maneuverability and the riding comfort of vehicles and the increasing maturity of control theories, the research on vehicle intelligent technology is concerned widely. The development of a dynamic model-based control technology for unmanned vehicles can achieve better road utilization rate and higher safety, but also needs to adapt to various complex driving environments, such as driving on roads with different road adhesion coefficients and curvature changes or achieving safe and stable emergency obstacle avoidance operation under emergency conditions.

The vehicle dynamic mathematical model based on physical derivation usually makes certain idealized assumptions to simplify the vehicle model during modeling, which results in that the real dynamic response of the vehicle in the driving process, such as load transfer of the front and rear axles of the vehicle, high-order dynamic response of tires, and the like, cannot be accurately calculated through experimental data. Although the data-driven model can be changed constantly according to the change of the external environment of the vehicle, the solution is difficult during the estimation of model parameters due to the fact that the nonlinear algorithm is used for carrying out optimization solution, and the problem of instantaneity exists. Based on historical empirical analysis of the manned vehicle, accurate predictions of future states are needed based on the current operating state of the vehicle. The unmanned vehicle can tightly combine planning and control by establishing an accurate prediction model, so that the overall control performance of the unmanned vehicle is further improved. However, the unmanned vehicle is a complex dynamic system, particularly under extreme conditions, the vehicle system and related subsystems can show highly nonlinear and strong coupling characteristics, and expanding the model dimension can improve the model accuracy, but can increase the modeling difficulty and bring challenges to the rapid solution of the algorithm. Therefore, how to establish the unmanned vehicle dynamics prediction model under the condition of simultaneously considering the complexity and the fidelity of the model becomes an important problem which needs to be solved urgently at present.

Disclosure of Invention

In order to solve the technical problems, the invention provides a vehicle dynamics prediction model modeling and training data acquisition method and a training method based on a time-lag feedback neural network. The training data acquisition specifically comprises the following steps:

in vehicle dynamics virtual data acquisition: the method comprises the steps of low-fidelity interpretable vehicle nonlinear dynamics multi-time-step virtual data acquisition and high-fidelity vehicle dynamics software CarSim multi-time-step virtual data acquisition.

In the low-fidelity interpretable vehicle nonlinear dynamics multi-time-step virtual data acquisition, different fidelity models such as Fiala tire nonlinear models, longitudinal load transfer of front and rear shafts, tire relaxation effects, variable road adhesion coefficients and the like are selectively added to the vehicle nonlinear dynamics based on the Newton's second law, so that low-fidelity interpretable vehicle nonlinear dynamics models with different complexity degrees are obtained. The input of the model is limited by theoretical value, and the input of the low-fidelity interpretable vehicle nonlinear dynamics model is reasonably and randomly sampled to obtainThe output tag signal is processed by Euler integration and fed back to the input end of the random sampling signal as a part of input signal of the next moment. Finally, the input and output signals are processed in multiple time steps (the term "multiple time step processing" means that a complete vehicle dynamics data time series is processed into a plurality of vehicle dynamics data series consisting of 4 time steps, for example, the data 12345678 is processed into: 1234/2345/3456 … …), obtaining a low fidelity interpretable vehicle nonlinear dynamics multi-time-step virtual dataset.

In a high-fidelity vehicle dynamics software CarSim multi-time-step virtual data acquisition module, aiming at the whole vehicle parameter configuration of an actual unmanned vehicle object, the cross-section, the longitudinal section, the attachment coefficient and the vehicle parameters of a road in the CarSim high-fidelity vehicle dynamics simulation software are modified, for example, the vehicle type grade is selected, the length, the width, the height, the axle distance, the minimum ground clearance, the setup quality, the whole vehicle rotational inertia and the like are modified, and a data acquisition joint simulation model is established in Matlab/Simulink. Analyzing the limit relation between the longitudinal speed and the front wheel steering angle, considering the limitation of vehicle control input based on the driving experience of people, carrying out combined simulation on the combined sinusoidal input under the limit relation between the longitudinal speed and the front wheel steering angle of a vehicle model in high-fidelity vehicle dynamics software CarSim by Matlab/Simulink to obtain continuous output response signals, carrying out multi-time-step processing on input and output data, and finally obtaining a multi-time-step virtual data set of the high-fidelity vehicle dynamics software CarSim.

In the actual unmanned vehicle dynamics real data acquisition module: the unmanned vehicle is provided with a wheel force sensor, an S-Motion DTI sensor, an MSW DTI sensor and an integrated navigation system (GPS global positioning system + inertial navigation) and receives data from a vehicle CAN bus. The method is characterized in that an unmanned vehicle is operated to perform various vehicle tests such as a drift test, a snake test, a single-shift test, a sine frequency sweep steering test, a steady-state rotation test, a quasi-static linear acceleration and deceleration test, an ISO double-shift running test and the like on a dry asphalt pavement, a wet and slippery sediment pavement and an ice and snow pavement, and actual unmanned vehicle dynamic data under high, medium and low pavement adhesion coefficients are obtained. And smoothing the collected data by using a second-order Butterworth low-pass filter with the cut-off frequency of 3Hz to filter the influence of high-frequency behaviors such as suspension vibration on vehicle dynamics. And finally, carrying out data synchronization and multi-time-step processing to obtain a real data set of the dynamics of the actual unmanned vehicle.

In the design of a vehicle dynamics prediction model based on a time-lag feedback neural network: the method is mainly characterized in that a neural network vehicle dynamics model (TDFB-NNVM) with time delay output feedback is designed by utilizing a long-time and short-time memory neural network, and is different from the learning of parameters in a single-track model based on physics. The neural network vehicle model adopts a double-layer long-time memory neural network structure, and the activation layer is selected as a Relu function. The model predicts the current vehicle yaw rate and lateral velocity derivatives by including vehicle control and state information with 4 time-step states in the input data, particularly state output parameters including time-lapse feedback in the state input data.

In the two-stage training of the vehicle dynamics prediction model based on the time-lag feedback neural network, the obtained virtual data set and the real data set are divided into a 70% training set, a 15% verification set and a 15% testing set. The data is randomized to break the time dependence of the data set, ensuring that each data sample consists of a time-dependent state trace, but that there is no correlation between any given two data samples. The Loss function is selected as a mean square error, the optimizer is selected as Adam, the batch size is set to be 1000, the learning rate is set to be 0.0001, the network model is learned and trained based on a learning framework of Pythroch, in the first stage, the network learns by using a vehicle dynamics virtual data set to obtain a pre-training model, and the generalization test of unknown data is carried out. And after the test is passed, performing model weight optimization of a second stage by using the vehicle dynamics real data set to finally obtain a time-lag feedback neural network-based vehicle dynamics prediction model attached to the actual unmanned vehicle.

The invention has the beneficial effects that:

1. the invention provides a vehicle dynamics virtual and actual data collection method under the condition of multiple roads, and lays a data foundation for vehicle dynamics modeling. Firstly, selectively adding different fidelity models based on the nonlinear dynamics of the vehicle, thereby obtaining low-fidelity interpretable multi-time-step virtual data sets with different complexity degrees of the nonlinear dynamics model of the vehicle; secondly, acquiring multi-time-step virtual data of a high-fidelity dynamic model through high-fidelity vehicle dynamics software CarSim; and finally, arranging an actual unmanned vehicle dynamics data acquisition device to acquire a vehicle dynamics real data set. The vehicle dynamics virtual data set is wide in freedom degree selection range and low in acquisition cost, meanwhile, the demand of real vehicle data can be reduced, the vehicle dynamics real data set can optimize weight parameters for the model again, and the accuracy of actual vehicle dynamics prediction response is improved.

2. The neural network vehicle dynamics prediction model designed based on the time-lag feedback idea can identify various complex dynamics behaviors including limit cycles, chaos and bifurcation in the vehicle running process, can learn all unknown or unmodeled vehicle dynamics change effects and highly nonlinear and strong coupling characteristics of the vehicle, and can make appropriate prediction on the road surface on which the vehicle runs without clear road surface friction estimation.

3. The invention lays a good foundation for developing a high-performance motion controller for the unmanned vehicle by accurately predicting the dynamic state of the vehicle at the next moment.

Drawings

FIG. 1 is a flow chart of vehicle dynamics prediction model training and data acquisition based on a time-lag feedback neural network

FIG. 2 is a vehicle dynamics virtual data acquisition module;

FIG. 3 is a flow chart of a low fidelity interpretable vehicle nonlinear dynamics multi-time step virtual data acquisition;

FIG. 4 is a non-linear kinetic model of a vehicle;

FIG. 5 is a flow chart of high fidelity vehicle dynamics software CarSim multi-time step virtual data acquisition;

FIG. 6 is a graph of vehicle front wheel angle versus longitudinal speed characteristics;

FIG. 7 is a real data acquisition module of actual unmanned vehicle dynamics;

FIG. 8 is a vehicle dynamics prediction model based on a time-lag feedback neural network;

FIG. 9 is a two-stage training structure diagram of a vehicle dynamics prediction model based on a time-lag feedback neural network;

Detailed Description

The invention will be further explained with reference to the drawings.

FIG. 1 is a flow chart of vehicle dynamics prediction model training and data acquisition based on a time-lag feedback neural network, including two parts of training data acquisition and model design training.

Fig. 2 is a vehicle dynamics virtual data acquisition module, which includes a low fidelity interpretable vehicle nonlinear dynamics multi-time-step virtual data acquisition module and a high fidelity vehicle dynamics software CarSim multi-time-step virtual data acquisition module, as follows:

FIG. 3 is a flow chart of a low fidelity interpretable vehicle non-linear dynamics multi-time step virtual data acquisition. The complex unmanned vehicle is analyzed according to Newton's second law to obtain stress balance equations along the x-axis, the y-axis and around the z-axis, and a nonlinear dynamical model of the vehicle is designed, as shown in FIG. 4. The unmanned vehicle (front wheel drive and steering) has the translation in the longitudinal direction, the transverse direction and the vertical direction and the rotation in 3 directions of rolling, pitching and yawing. Wherein the lateral and yaw motions are substantially generated by the steering maneuver. The nonlinear vehicle dynamics model may be expressed in terms of differential equations as:

wherein U is the speed at the vehicle's center of mass; u shape_x，U_yThe speeds of the vehicle mass center along the x and y directions of the vehicle body coordinate system are respectively; alpha is alpha_f，α_rRespectively are front and rear wheel side deflection angles; beta is the vehicle mass center slip angle; r is a carYaw angular velocity of the vehicle; a and b are the distances from the mass center of the vehicle to the front and rear axes; m is the vehicle mass, I_zThe moment of inertia of the vehicle around the z-axis of the center of mass; f_yf.F_yrThe lateral resultant force of the front axle tire and the rear axle tire is respectively; f_xfThe longitudinal resultant force received by the front axle tire; delta is the front wheel angle.

The nonlinear characteristics generated in the running process of the vehicle under different road conditions are caused by the fact that the tire turns, so that in order to expand the application range of a vehicle model, a Fiala model of the tire is introduced, and the lateral force F of the tire_yThe calculation formula of (2) is as follows:

wherein C is_αAnd μ is the tire cornering stiffness and road adhesion coefficient; f_zIs the tire vertical load; α is the tire slip angle; alpha is alpha_satIs the tire saturated slip angle. The formula for calculating the slip angle of the front tire and the rear tire is as follows:

longitudinal weight transfer also affects vehicle dynamics when the vehicle is in high performance. This effect is due to the vehicle accelerating or braking which increases or decreases the vertical forces experienced on each tire, thereby affecting the dynamics of the vehicle. Where h is the height to the center of gravity of the vehicle, g is the acceleration due to gravity, a_xIs the vehicle longitudinal acceleration, and L is the vehicle wheelbase. When used in conjunction with a non-linear tire model, weight transfer may increase or decrease the tire loadThe magnitude of the vertical force further influences the magnitude of the lateral force of the tire. Front and rear axle vertical force F_zf，F_zrThe calculation formula of (2) is as follows:

another major contributing factor to the model during low speed driving is the tire slack length, from which the delay in the lateral force experienced by each tire can be modeled. The retardation of the tire is characterized by the first derivative of the slip angle between the front and rear tiresThe calculation formula is as follows:

wherein σ_f，σ_rThe front and rear tire slack lengths.

Different fidelity models such as Fiala tire nonlinear model, longitudinal load transfer of front and rear shafts, tire relaxation effect, variable road adhesion coefficient and the like are selectively added to the vehicle nonlinear dynamics based on Newton's second law, so that low-fidelity interpretable vehicle nonlinear dynamics models with different complexity degrees are obtained. The input to the model is subject to theoretical numerical limits as follows:

U_x∈(1m/s，33m/s)

δ∈(-25°，25°)

F_xf∈(-μmg，power_lim/U_x)

r∈(-μg/U_x，μg/U_x)

U_y∈(3aμmgU_x/LC_α+bμg/U_x，-3aμmgU_x/LC_α-bμg/U_x)

where μ is the road adhesion coefficient and power _ lim is the maximum vehicle power.

Obtaining after reasonable random sampling of inputs to a low-fidelity interpretable vehicle nonlinear dynamical modelThe output tag signal is processed by Euler integration and fed back to the input end of the random sampling signal as a part of input signal of the next moment. And finally, performing multi-time-step processing on the input and output signals to obtain a low-fidelity interpretable vehicle nonlinear dynamics multi-time-step virtual data set.

FIG. 5 is a flow chart of high fidelity vehicle dynamics software CarSim multi-time step virtual data acquisition. According to the whole vehicle parameters and road parameters of an actual unmanned vehicle object, the cross-section, the attachment coefficient and the vehicle parameters of the road in the CarSim high-fidelity vehicle dynamics simulation software are modified, for example, the vehicle type grade is selected, the length, the width, the height, the wheel base, the minimum ground clearance, the setup mass, the whole vehicle rotational inertia and the like are modified, and a data acquisition joint simulation model is established in Matlab/Simulink.

And analyzing the limit relation between the longitudinal vehicle speed and the front wheel turning angle. The longitudinal speed and the front wheel steering angle of the vehicle do not have a mutual relation when the vehicle normally runs, generally when the vehicle runs at a low speed, in order to pursue maneuvering flexibility, operations such as turning around, large turning and the like can be carried out, so the input front wheel steering angle is relatively large; when the vehicle runs at a high speed, the input front wheel steering angle is relatively small to ensure safety and stability. However, when the vehicle is in an emergency obstacle avoidance working condition, a large front wheel steering angle is input at a medium-high speed in order to ensure safe and stable running, so that a limit relationship of mutual influence exists between the longitudinal vehicle speed and the front wheel steering angle. The total variance method can comprehensively explain the 'compliance degree' of the vehicle under certain operation, and is a comprehensive index for evaluating the operation stability of the vehicle, so that the minimum total variance method is used for carrying out theoretical analysis on the limit relation between the longitudinal vehicle speed and the front wheel rotation angle, and further determining the reasonable threshold value of the vehicle control quantity.

Assuming that the front wheel steering input control quantity is x (t), the lateral dynamic response is y (t), and the total variance of the vehicle dynamic response is shown in formula (7):

in the formula, x₀，y₀Steady state values for input x (t) and response y (t),to reflect the error. Since the front wheel steering speed of the vehicle is fast in an emergency, let x be assumed₀As a constant, the total variance calculation can be simplified as:

in order to ensure the safety and the operation stability of the vehicle, a threshold value of the total variance of the dynamic response of the vehicle is required to be set, namely E is less than or equal to E₀. E is set according to the motion condition of the vehicle and by combining the operation experience that a driver can ensure safe and stable running during emergency steering of the vehicle₀And (3) a threshold value, so that the reasonable input of the front wheel steering angle is calculated:

according to the experience of the driver and relevant theoretical data, when the longitudinal speed is 20m/s, taking E₀0.15, the maximum safe front wheel angle was calculated to be 3. The curve of the front wheel angle versus longitudinal speed characteristic obtained using the total variance method is shown in fig. 6. When the speed of the vehicle is very low, the front wheel turning angle can be very large, but the front wheel turning angle can be rapidly increased along with the increase of the vehicle speedThe speed decreases, and when the speed increases to a certain degree, the front wheel turning angle decreases slowly. The corresponding front wheel steering angles at medium and low vehicle speeds are corrected to a certain extent in consideration of the actual steering condition of the vehicle in the running process.

After the driver inputs a control command, the vehicle obtains a certain vehicle dynamic response. Thus, the training input may be similar to the driver's control input, and the training input signal may be selected to take into account the limitations of the vehicle control input based on human driving experience. When danger occurs during driving, a driver can suddenly change the front wheel steering angle of the vehicle to avoid obstacles, but generally, continuous steering input with low frequency is applied to the vehicle during normal driving. Therefore, the maximum control input frequency is selected to be 2Hz based on human intuition when driving.

A single sinusoid may identify some dynamic systems, while others require random inputs with uniform distribution. In practice, the steering inputs applied are relatively random over a relatively long period of time during which the vehicle is operating. Therefore, selecting random inputs with a uniform distribution may ensure successful training of the neural network vehicle dynamics model and compliance with the actual vehicle behavior.

To create a continuous random input signal, sinusoids with different frequencies and amplitudes are added as input signal u (t), the formula:

where d is the total number of sinusoids, a_mIs the maximum amplitude of the input, ω_mIs the maximum frequency of the input, #_jIs the initial phase angle of the sine wave, j represents the jth sine wave.

The combined sine wave method for system identification requires the power of the signal to be placed precisely at each frequency. The phase of the signal may be selected to ensure that the power of the signal is evenly distributed over these frequencies according to the schroeder phase equation:

wherein phi₁Any value may be selected. The total number of sinusoids should be as large as possible to increase the amount of information displayed to the system. The maximum amplitude of the signal is the sum of the individual amplitudes, and the maximum frequency is determined by the maximum frequency of the individual sine waves.

The maximum frequency omega is selected to use d as 100_m2Hz combined sinusoids. Maximum amplitude a of the combined sinusoidal steering input sum_mDetermined from the limit relationship between the vehicle longitudinal speed and the front wheel steering angle in fig. 6: a is_mEqual to the maximum front wheel rotation angle that can be reached at a certain vehicle speed. For example, when the vehicle speed is 20m/s, a_mEqual to a maximum front wheel angle of 3.

The method comprises the steps of carrying out combined simulation on a vehicle model in high-fidelity vehicle dynamics software CarSim through combined sinusoidal input under the limit relation of longitudinal vehicle speed and front wheel steering angle on Matlab/Simulink to obtain continuous output response signals, carrying out multi-time-step processing on the data, and finally obtaining a multi-time-step virtual data set of the high-fidelity vehicle dynamics software CarSim.

Fig. 7 is a real data acquisition module of actual unmanned vehicle dynamics. The unmanned vehicle is provided with a wheel force sensor, an S-Motion DTI sensor, an MSW DTI sensor and an integrated navigation system (GPS global positioning system + inertial navigation) and receives data from a vehicle CAN bus. The method is characterized in that an unmanned vehicle is operated to perform various vehicle tests such as a drift test, a snake test, a single-shift test, a sine frequency sweep steering test, a steady-state rotation test, a quasi-static linear acceleration and deceleration test, an ISO double-shift running test and the like on a dry asphalt pavement, a wet and slippery sediment pavement and an ice and snow pavement, and actual unmanned vehicle dynamic data under high, medium and low pavement adhesion coefficients are obtained. And smoothing the collected data by using a second-order Butterworth low-pass filter with the cut-off frequency of 3Hz to filter the influence of high-frequency behaviors such as suspension vibration on vehicle dynamics. And finally, carrying out data synchronization and multi-time-step processing to obtain a real data set of the dynamics of the actual unmanned vehicle.

FIG. 8 is a vehicle dynamics prediction model based on a time-lag feedback neural network. The method is mainly characterized in that a neural network vehicle dynamics model (TDFB-NNVM) with time delay output feedback is designed by utilizing a long-time and short-time memory neural network, and is different from the learning of parameters in a single-track model based on physics.

The neural network vehicle dynamics model specifically adopts the following structure: the first layer is an input layer which has 7 characteristic inputs, namely a yaw angular velocity r and a transverse velocity U_yLongitudinal speed U_xFront wheel steering angle delta, longitudinal resultant force F of vehicle front axle tire_xfThe output of the sum network model becomes the derivative output value of the yaw rate and the transverse speed of the input information through time delay feedbackThe data for each input feature contains vehicle dynamics information for a total of 4 time steps. The second layer is the LSTM1 network layer, and the hidden layer design has 128 hidden units. The third layer is an activation layer, and the activation function is selected to be a Relu function. The fourth layer is the LSTM2 network layer, the hidden layer is designed to have 128 hidden units, and only outputs the information after the last LSTM-CELL calculation. The fifth layer is a regression output layer, and the hidden layer is designed to have 2 hidden units.

The model predicts the current vehicle yaw rate and lateral velocity derivatives by including vehicle control and state information with 4 time-step states in the input data, particularly state output parameters including time-lapse feedback in the state input data.

The forward calculation method of the learned neural network vehicle dynamics model is as follows:

h＝{x_t，…，x_t-T}

a_l＝max(0，z₁)

wherein x is_tRepresenting delayed feedback status output, control and status input information in a single time step, h representing x comprising a plurality of historical time step information_tData, W_lstm{1，2}∈(w_i，w_f，w_g，w_o)，b_lstm{1，2}∈(b_i，b_f，b_g，b_o) Represents the learned network weights and bias parameters, w, of the 2 LSTM network layers_i，w_f，w_g，w_o，b_i，b_f，b_g，b_oRespectively expressed as weight matrix of an input gate layer, a forgetting gate layer, a Tanh layer and an output gate layer in the LSTM network and bias matrix of the input gate layer, the forgetting gate layer, the Tanh layer and the output gate layer.To output the regression layer weight transpose matrix, b_FCIs the bias matrix of the output regression layer. Wherein in the equation of the network model, F_lstmIs an abbreviation of LSTM network model. a _ l represents an active layer, z_iAnd i is 1, and 2 represents weighted output of different network layers. Predicted output of networkAndis defined as:andwhere Δ t is 50ms, which is the sampling frequency of the signal.

FIG. 9 is a diagram of a two-stage training architecture of a vehicle dynamics prediction model based on a time-lag feedback neural network. The obtained virtual data set and the real data set are divided into a training set of 70%, a verification set of 15% and a test set of 15%. The data is randomized to break the time dependence of the data set, ensuring that each data sample consists of a time-dependent state trace, but that there is no correlation between any given two data samples. The Loss function is selected as mean square error MSE, the optimizer is selected as Adam, the batch size is set to be 1000, the learning rate is set to be 0.0001, the network model is subjected to learning training based on a learning framework of Pythroch, and an optimization training algorithm is as follows:

wherein N is the total number of training samples,is the derivative truth-label data of the vehicle yaw rate and lateral velocity at the next time,derivative output values of the vehicle yaw rate and lateral speed are predicted for the network.

In the first stage, the network learns by using a vehicle dynamics virtual data set to obtain a pre-training model, and performs a generalization test on unknown data. Inputting data of 15% test set into network for calculationBy obtaining predicted output values of the networkTagging with test set dataThe mean square error MSE is calculated.

The generalized test judgment conditions are as follows: if it is notThe generalization test passes.

After the test is passed, model weight optimization of the second stage is carried out by utilizing the vehicle dynamics real data set, namely, a pre-training model obtained by utilizing the virtual data set in the first stage, a training set of the real data set and data in a verification set are loaded, and the optimization training algorithm is utilized again to carry out learning training on the network model. And finally obtaining a vehicle dynamics prediction model based on a time-lag feedback neural network, which is attached to the actual unmanned vehicle.

The above-listed series of detailed descriptions are merely specific illustrations of possible embodiments of the present invention, and they are not intended to limit the scope of the present invention, and all equivalent means or modifications that do not depart from the technical spirit of the present invention are intended to be included within the scope of the present invention.