A mechanism analysis-data driven vehicle dynamics series hybrid model, intelligent vehicle trajectory tracking control method and controller

Abstract

The application discloses a mechanism analysis-data driven vehicle dynamics series hybrid model, an intelligent automobile track tracking control method and a controller, proposes a mechanism analysis-data driven vehicle dynamics series hybrid model, the model not only has the advantages of good system mechanics background and physical meaning of the mechanism model, but also has the advantages of not needing any prior knowledge and only relying on sample data of the data driven model, and the advantages of the two are complementary. The model can supplement part of the unmodeled dynamics in the computer mechanism model and improve the global calculation precision of the model, has the ability of implicitly understanding different road adhesion conditions, and lays a good model foundation for intelligent automobile motion control algorithm design. In the control method, an accurate vehicle dynamics model is introduced as a prediction model, the prediction ability of the future state output of the vehicle is effectively improved, nonlinear constraint conditions are increased, and the control precision, stability and reliability are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent vehicle dynamics control, and particularly relates to a mechanism analysis-data driven vehicle dynamics series hybrid model, an intelligent vehicle trajectory tracking control method and a controller. BACKGROUND

[0002] With the increasing requirements of drivers on the safety, maneuverability and ride comfort of vehicles and the increasing maturity of control theory, intelligent vehicle technology research has attracted widespread attention. The control technology of autonomous vehicles based on vehicle dynamics model can achieve better road utilization 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 implementing safe and stable emergency obstacle avoidance operations in emergency working conditions.

[0003] Trajectory tracking control, as an important supporting technology for realizing automatic driving of vehicles, mainly uses control algorithms to calculate the front wheel steering angle, so that the lateral position error and heading deviation of the vehicle in the process of driving along the reference trajectory are as small as possible under the action of the underlying execution device, while ensuring that the vehicle has certain stability and driving safety. At present, most control methods are to calculate physical quantities such as vehicle yaw rate to describe vehicle motion based on the vehicle dynamic mathematical model derived based on physics, and then design a feedback control system for tracking. However, the vehicle mechanism analysis model based on theoretical derivation usually makes idealized assumptions to simplify during modeling, which makes it difficult to establish an accurate mathematical model close to the actual research object. In addition, the trajectory tracking control method based on the model is usually affected by parameter perturbation, uncertainty disturbance, time delay and actuator saturation constraint, etc., so that some mathematical assumptions no longer hold, resulting in that the intelligent vehicle will lose the ability of trajectory tracking and stability. In recent years, with the progress of big data and artificial intelligence technology, data-driven complex system analysis methods have developed rapidly, providing a new solution to the modeling problem of vehicle nonlinear dynamic systems. However, the related basic theory of applying data-driven technology to establish vehicle dynamics model still needs to be improved, and there are bottlenecks in the explainability and stability of the model. The mechanism model has practical physical meaning, and the data-driven model only relies on sample data modeling. The fusion of the two can realize the complementary advantages between models, which has great potential. Therefore, how to establish an accurate vehicle dynamics model and use the high-precision model obtained in the development and design of automatic driving motion control algorithm based on modern control theory has become an important problem to be solved. SUMMARY

[0004] In view of the above problems, the application provides a mechanism analysis-data driven vehicle dynamics series hybrid model, and based on the model, an intelligent automobile trajectory tracking control method and controller based on the vehicle dynamics hybrid model are provided, including model design and model-based control algorithm.

[0005] The above model and model control algorithm of the application can be integrated in a vehicle controller.

[0006] The application has the following beneficial effects:

[0007] 1. The application provides a mechanism analysis-data driven vehicle dynamics series hybrid model, which has the advantages of good system mechanics background and physical significance of the mechanism model, and the advantage of not requiring any prior knowledge and only relying on sample data of the data driven model, and the advantages of the two are complementary. In addition, the model can supplement part of the unmodeled dynamics in the computer mechanism model and improve the global calculation accuracy of the model, and has the ability to implicitly understand different road adhesion conditions, thereby laying a good model foundation for intelligent automobile motion control algorithm design.

[0008] 2. In the control algorithm, an accurate vehicle dynamics model is introduced as a prediction model, which effectively improves the prediction ability of the control algorithm for the future state output of the vehicle in the actual driving environment, increases the nonlinear constraint condition, further explores the potential of the vehicle from the perspective of dynamics control, improves the control precision, stability and reliability, and ensures the safe and stable operation of the intelligent automobile. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 The intelligent automobile trajectory tracking control flowchart based on the vehicle dynamics hybrid model;

[0010] Figure 2 The vehicle mechanism analysis model diagram;

[0011] Figure 3 The mechanism analysis-data driven vehicle dynamics series hybrid model structure diagram;

[0012] Figure 4 Embedding schematic diagram for predictive model control algorithm; DETAILED DESCRIPTION

[0013] The application will be further described below with reference to the accompanying drawings.

[0014] Figure 1 The intelligent vehicle trajectory tracking control flowchart based on a vehicle dynamics hybrid model is as follows:

[0015] A mechanism analysis-data driven vehicle dynamics series hybrid model is constructed as a prediction model. Vehicle state and control data are processed by the mechanism model for front-end calculation, and the level is combined and used as the input of the data driven module. The long short-term memory network is used as the backbone network to realize the nonlinear correlation feature extraction of time series data and the final model output calculation.

[0016] The weight parameters of the neural network are extracted for model forward calculation. The Euler integral is used to complete the discretization of the prediction model, and the model predictive control trajectory tracking algorithm is designed. The feedforward feedback control algorithm is designed to realize the longitudinal control of the vehicle while providing the external input required by the prediction model in the lateral control, and to realize the precise trajectory tracking control effect.

[0017] Figure 2 The vehicle mechanism analysis model diagram is as follows. U is the speed at the vehicle mass center; U x ,U y are the speeds of the vehicle mass center along the x and y directions of the vehicle body coordinate system, respectively; α f ,α r are the front and rear wheel side slip angles, respectively; β is the vehicle mass center side slip angle; r is the vehicle yaw rate; a and b are the distances from the vehicle mass center to the front and rear axles; m is the vehicle mass, I z is the moment of inertia of the vehicle around the z-axis of the mass center; F yf ,F yr are the lateral forces received by the front and rear tires, respectively; F xf is the longitudinal force received by the front tire; δ is the front wheel steering angle; e is the lateral deviation; x LA is the longitudinal preview distance; e LA is the lateral deviation at the preview point; and Δψ is the heading deviation.

[0018] Based on reasonable assumptions to simplify the vehicle model:

[0019] (1) Ignoring the influence of vehicle aerodynamics;

[0020] (2) Assuming that it is running on a horizontal road, ignoring the influence of vehicle vertical motion;

[0021] (3) The vehicle is assumed to be a single-track model, and the wheel camber angle and lateral load transfer are ignored;

[0022] (4) The vehicle steering system is ignored, and the input of the vehicle dynamics model is the front wheel steering angle;

[0023] (5) The vehicle and suspension are assumed to be rigid systems, and the motion coupling effect is ignored;

[0024] (6) The longitudinal-lateral coupling effect of the tire force is ignored, and only the tire cornering characteristics are considered.

[0025] The vehicle is set to be front-wheel drive, and a single-track model of the lateral dynamics based on the mechanism analysis theory is established, and the differential equation is expressed as:

[0026]

[0027] In order to expand the application range of the vehicle model, the Fiala nonlinear tire model is introduced, that is:

[0028]

[0029]

[0030] In the formula, C α and μ are the tire cornering stiffness and road adhesion coefficient; F z is the tire vertical load; α is the tire cornering angle; and α sat is the tire saturation cornering angle. The front and rear tire cornering angle calculation formulas are:

[0031]

[0032] When the vehicle is in high-performance driving, the vehicle will increase or decrease the vertical force borne on each tire due to acceleration or braking, and the use of the nonlinear tire model will affect the calculation of the tire lateral force, and then affect the vehicle lateral dynamics. The longitudinal load transfer effect is introduced, and the calculation formulas of the front and rear axle vertical forces F zf and F zr are:

[0033]

[0034] In the formula, h CG is the height to the center of gravity of the vehicle, g is the acceleration of gravity, and L is the wheelbase of the vehicle.

[0035] Figure 3 The structure diagram of the mechanism analysis-data driven vehicle dynamics series hybrid model is constructed. The input characteristics of the MD-SHM model are the dynamic state and control information of the vehicle, the state information includes the yaw rate r, the lateral velocity U y and the longitudinal velocity U x, control information contains front wheel angle δ and vehicle longitudinal force F xf Each input feature vector contains vehicle control and state data combined with 3 time steps of history information. These input feature vectors are input into the mechanism analysis model established in the foregoing, and the yaw rate differential value and lateral velocity differential value of 4 time steps are calculated under the premise of considering the tire nonlinearity and load transfer effect, so as to realize data preprocessing based on the mechanism model. The data calculated by the mechanism model are used as the initial vehicle history dynamic benchmark input with physical significance of the neural network model, are combined with the input feature vectors, and new vehicle state and control data of 4 time steps are obtained and input into the data-driven neural network model connected in series after the mechanism model.

[0036] The vehicle state and control data of 4 time steps obtained by the mechanism model calculation and combination are typical time series data, and there are certain time sequence correlation characteristics in these data. The MD-SHM model selects the long short-term memory network (LSTM) which has great advantages in time series data feature extraction as the backbone network of the data-driven module, and combines the fully connected neural network. The vehicle state and control data with 7 dynamic characteristics and 4 time step sequences are input into the data-driven module connected in series after the mechanism model. The module encodes the data time sequence characteristics through the first layer of LSTM, maps to the high-dimensional feature space through the fully connected layer, and decodes the characteristics through the second layer of LSTM, and then transmits to the regression layer to calculate the final required yaw rate differential value and lateral velocity differential value. In summary, the data is forward propagated through the MD-SHM, and the continuous time state data sequence generated during the application vehicle operation is used to calculate and map the current state change amount of the vehicle.

[0037] Since the single-track model has interpretability and determinacy, and the neural network model has strong nonlinearity modeling capability, by merging the two cooperative models, the calculation defects of the single-track model can be supplemented, the initial calculation value of the single-track model can be optimized, the lost information of the single-track model can be completed, the unmodeled effects in the single-track model can be estimated, the global calculation performance of the model can be improved, and the system background and physical meaning are contained to some extent, and better interpretability is obtained. In summary, the MD-SHM model improves the modeling accuracy and calculation efficiency while improving the extrapolation capability of the neural network model.

[0038] The forward calculation process of the MD-SHM model is as follows:

[0039]

[0040] In the formula, x t is the vehicle state data of a single time step, u t is the vehicle control data of a single time step, X t h t X t W lstm{1,2} ∈(w i ,w f ,w g ,w o )W FC{1,2} ,b lstm{1,2} ∈(b i ,b f ,b g ,b o ),b FC{1,2} F lstm z i f STM

[0041] Virtual running data of vehicle mechanism analysis model and real vehicle state data of intelligent car in actual driving process are collected for training. First, the pre-training model is obtained by training using virtual vehicle dynamics data, and then the weight parameters of the neural network are updated using real vehicle dynamics data set.

[0042] The key to successful network training is to design a correct loss function to adjust the internal parameters of the network so that the loss value decreases continuously. In the training process, the internal parameters of the network are adjusted through error back propagation to make the loss value decrease continuously. The smaller the loss value is, the higher the fitting accuracy of the network to the data is. The training loss function is defined by applying Euler integral method as follows:

[0043]

[0044] In the formula, r, U y are the true values of the vehicle yaw rate and lateral velocity state measurement, are the predicted values of the vehicle yaw rate and lateral velocity at the next time of the network calculated by Euler integral, N is the training number. W, b are the weight and bias matrix of the neural network to be trained in the model, f MD-SHM is the forward calculation formula of MD-SHM model shown in formula (5).

[0045] The details of network training are as follows: the loss function is selected as mean square error MSE, the optimizer is selected as Adam, the batchsize is set to 1000, the learning rate is set to 0.001, and the network model is learned and trained based on Pytorch learning framework.​

[0046] After the training, the neural network weight parameters are extracted, which are used for forward calculation to obtain the complete mechanism analysis-data driven vehicle dynamics series hybrid model of the autonomous vehicle.

[0047] Figure 4 The schematic diagram is embedded for the predictive model control algorithm. In Figure 1 The vehicle trajectory tracking error model is established based on the single-track model shown in the figure:

[0048]

[0049] wherein, is the speed of the vehicle along the reference trajectory at the arrival driving distance s.

[0050] Considering the established mechanism-data series hybrid vehicle dynamics model, the historical and current state and control quantity are taken as the input of the MD-SHM model, and are expressed as follows:

[0051]

[0052] In the formula, the state quantity x(t) is [r, U y , U x ] t , the control quantity u(t) is [δ, F xf ] t The front wheel steering angle of the vehicle in the trajectory tracking process is obtained by the model predictive control algorithm optimization, and the longitudinal control is realized separately, so that the longitudinal speed and the front wheel longitudinal force of the vehicle can be regarded as the external input of the lateral dynamics, which will be calculated outside the model predictive control optimization problem.

[0053] In order to realize the longitudinal speed control of the vehicle, a feedforward control and feedback control algorithm are designed to track the pre-calculated expected longitudinal acceleration and longitudinal speed curve. In addition, in the process of Figure 4 In the model prediction and solving of the lateral control quantity, the longitudinal speed and the front wheel longitudinal force of the vehicle are important external input quantities of the lateral dynamics.

[0054] Based on the point mass assumption and the friction circle theory, the speed planning of the reference trajectory is carried out, wherein the acceleration limited friction circle is defined as follows:

[0055]

[0056] In the formula, a y is the lateral acceleration of the vehicle mass center, a x is the longitudinal acceleration of the vehicle mass center, μ is the road adhesion coefficient, and the vehicle acceleration limit coefficient ξ is 0.533.

[0057] Based on the assumption of steady turning of the vehicle, the lateral acceleration of the vehicle when it is turning with a constant turning radius R can be determined as follows:

[0058]

[0059] When the vehicle is tracking a variable-curvature trajectory, its longitudinal acceleration will vary with the trajectory curvature K. Substituting equation (9) into equation (10) and limiting the friction circle, we have:

[0060]

[0061] Differentiating equation (11) with respect to time t, we have: x

[0062]

[0063] Substituting equation (12) into equation (11), we have:

[0064]

[0065] Solving the differential equation by reverse integration of equation (13), we have the expected longitudinal velocity of each trajectory point in the variable-curvature trajectory as:

[0066]

[0067]

[0068] wherein is the distance between the two trajectory points in front of and behind the trajectory point, and L is the total length of the trajectory, so is the distance at the end of the variable-curvature trajectory where the reverse integration starts. The end of the integration is located at the starting point of the variable-curvature trajectory.

[0069] The speed planning is performed on the variable-curvature reference trajectory to obtain the expected longitudinal acceleration a x,des and the longitudinal velocity U x,des curve, and substituted into the solution of the longitudinal control input. Based on the point mass assumption, the current speed of the vehicle, the driving position along the reference trajectory, and the front wheel longitudinal force are measured, and the front wheel longitudinal force is solved by the feedforward term that compensates for the expected acceleration and the resistance and the feedback term that tracks the expected speed, as shown in the following equation:

[0070]

[0071]

[0072] wherein k p is the proportional gain of the difference between the planned expected speed and the measured actual speed, and s​​t Distance reached at the reference trajectory point for travel time t. Assuming small lateral error and heading deviation of the point mass model relative to the reference trajectory, the longitudinal velocity external input required in the MPC prediction horizon is substituted into the calculation

[0073]

[0074]

[0075] The position of the next trajectory point in the prediction horizon is obtained by Euler integration:

[0076]

[0077] Δt represents the discrete time interval, s t+1 , s t represents the trajectory distance s at time t+1 and time t.

[0078] In order to apply the MD-SHM model to the design of the model predictive control algorithm, realize the continuous time prediction of the model in the model predictive control optimization problem, and discretize it, specifically use the Euler integration method to convert the continuous dynamics into discrete dynamics, and ensure that the system control signal in the optimization problem remains constant between stages. In each stage of optimization, each discrete state and control variable is considered as an optimization variable, as shown in the following formula:

[0079] x(t+Δt)=x(t)+Δt·f MD-SHM (h t )

[0080] Since the input quantity of the MD-SHM model has historical state and control quantity, it is difficult to calculate the high-order integral term when implemented in model predictive control (MPC). By selecting the Euler method, the MD-SHM model can be calculated the minimum number of times in the optimization problem, and the discrete dynamic state information at the next stage time can be calculated within the optimization range, thereby improving the calculation efficiency and meeting the real-time requirements of the control algorithm.

[0081] After the discretization of the MD-SHM model is completed, it needs to be embedded in the model predictive control algorithm to realize the prediction of the future trajectory error state. The MD-SHM model will be calculated in each optimization problem along the prediction horizon. In the first stage of the optimization problem, the input of the MD-SHM model only includes the measured historical and current state information and the current optimized control information; in the subsequent three stages, the model input is composed of the measured and optimized vehicle state and control information. From the fifth stage, the model input only includes the optimized vehicle state and control information. For example Figure 4 ​As shown, the MPC prediction problem with 3 delay states shows how the input of each MD-SHM model in each stage varies and eventually consists of only the optimization state and control information.

[0082] In each optimization problem, the MD-SHM model calculates the predicted value of the vehicle state at the next time, and these state values are substituted into the calculation to obtain the trajectory tracking error output of the vehicle at the future time within the prediction time domain. In addition, it should be noted that in the actual control process, the control time domain is generally required to be less than the prediction time domain, that is, N c <N p When the prediction time domain of the optimization solution is greater than the control time domain, the control increment is set to 0, that is:

[0083] u(t+i)=u(t),i=N c ,N c +1,……,N p -1

[0084] In the design of the model predictive control algorithm based on dynamics, due to the complexity of the control unit inside the vehicle, the mechanical saturation and physical constraints of the actuator, it is necessary to add control quantity constraints, control increment constraints, vehicle dynamics constraints, etc. to ensure that the control algorithm can be accurately and quickly solved. The specific constraint conditions are as follows:

[0085] (1) Restrict the front wheel steering angle control quantity, and the control quantity obtained by solving is less than the steering limit of the vehicle bottom actuator, and the limit is ±32°, which is expressed as follows:

[0086] δ min (t+j)≤δ(t+j)≤δ max (t+j),j=0,1,…,N c -1

[0087] δ max 、δ min represent the maximum and minimum values of the constrained front wheel steering angle control quantity.

[0088] (2) Restrict the front wheel steering angle control increment, and limit the control increment in each sampling period within a reasonable range to avoid sudden changes in the control quantity and ensure the continuity of the control quantity, and the limit is ±2.25°, which is expressed as follows:

[0089] Δδ min (t+j)≤Δδ(t+j)≤Δδ max (t+j),j=0,1,…,N c -1

[0090] Δδ min 、Δδ max represent the minimum and maximum values of the constrained front wheel steering angle control increment.

[0091] (3) Restrict the side slip angle of the mass center, which is limited within a reasonable range to improve the stability of the vehicle. The vehicle runs stably on a dry asphalt road with good adhesion, and the limit value of the side slip angle of the mass center is ±12°; the vehicle runs on an icy road with low adhesion, and the limit value is ±2°. In this paper, the adhesion coefficient of the road is uncertain, so the side slip angle of the mass center is restricted to the larger limit value of ±12°, which is expressed as follows:

[0092] x β，min (t+j)≤x β (t+j)≤x β，max (t+j),j=0,1,…,N p -1

[0093] x β x β，min x β，max x β x

[0094] Reasonably designing the objective function in the model predictive control algorithm is an important basis for ensuring that the intelligent vehicle can quickly and stably track the reference trajectory. In the design of the model control algorithm based on MD-SHM, in order to clearly compare with the model trajectory tracking control algorithm based on NVM, the trajectory tracking objective function commonly used is selected, the deviation of the system state quantity and the optimization of the control quantity are added, and the objective function in the following form is adopted:

[0095]

[0096] In the formula, the front wheel steering angle control increment Δδ is directly taken as the optimization variable, Q e Q Δψ Q Δδ are the weight matrices of the lateral error e, the heading deviation Δψ and the control increment Δδ, respectively, N p N c are the prediction time domain and the control time domain, respectively. The first term indicates the ability of the system to follow the reference trajectory, and requires the lateral error and the heading deviation of the vehicle following the reference trajectory to be as small as possible, thereby improving the trajectory tracking effect. The second term indicates the requirement for the constraint of the control increment, and ensures that the control variable changes as quickly and smoothly as possible. The overall goal of the objective function is to enable the controlled object to quickly, accurately and stably track the trajectory.

[0097] Based on the above objective function and constraint conditions, the trajectory tracking control algorithm based on the MD-SHM prediction model solves the following nonlinear optimization problem with constraints in each control cycle, as shown below:

[0098] minimize

[0099] subject to h1=h measure

[0100]

[0101] x(t+1)=x(t)+Δt·f MD-SHM (h t ), t=1,…N p -1

[0102] δ min (t)≤δ(t)≤δ max (t), t=1,,N c

[0103] Δδ min (t)≤Δδ(t)≤Δδ max (t), t=1,…,N c

[0104] x β，min (t)≤x β (k)≤x β，max (t), t=1,…,N p

[0105] In the formula, h measure The multi-time-step vehicle state data measured within each control cycle is set as the initial vehicle state quantity h1 input into the control algorithm. By solving the nonlinear optimization problem shown in the above equation within a certain control cycle, a series of control increments in the control time domain within that cycle are obtained:

[0106]

[0107] The first control increment in the obtained control increment sequence is used to calculate the actual control quantity acting on the controlled system, i.e.:

[0108] δ(t) = δ(t-1) + Δδ(t)

[0109] Under the influence of the control variables, the vehicle generates new dynamic state variables, which are then passed to the next control cycle for further optimization. By continuously solving and iterating through the above problems, the trajectory tracking control of intelligent vehicles can be achieved.

[0110] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent methods or modifications that do not depart from the technology of the present invention should be included within the scope of protection of the present invention.

Claims

1. A mechanism analysis-data driven hybrid model of vehicle dynamics, characterized in that, The input features of the mechanism analysis-data driven vehicle dynamics series hybrid model are the dynamic state information and control information of the vehicle, the dynamic state information includes yaw rate r, lateral velocity U y and longitudinal velocity U x , and the control information includes front wheel steering angle δ and vehicle longitudinal force F xf ; Each input feature vector contains vehicle control and state data currently combined with 3 time steps of historical information, which is input into the mechanism model to calculate the yaw rate differential value and lateral velocity differential value for a total of 4 time steps, realizing data preprocessing based on the mechanism model; The data calculated by the mechanism model is used as the initial vehicle historical dynamic benchmark input with physical meaning of the data-driven model, which is combined with the input feature vector to obtain a new 4-time-step vehicle state and control data combination, which is input into the data-driven model connected in series after the mechanism model, and the data-driven model is used to realize nonlinear correlation feature extraction of time series data and final model output calculation; The mechanism model in the vehicle dynamics series hybrid model is designed as follows: The vehicle is set to be front-wheel drive, and a single-track model of lateral dynamics based on mechanism analysis theory is established, and the differential equation is expressed as: where U is the velocity at the vehicle center of mass; U x , U y are the velocities along the x, y directions of the vehicle body coordinate system at the vehicle center of mass; a f , a r are the front and rear wheel side slip angles; β is the vehicle center of mass side slip angle; r is the vehicle yaw rate; a, b are the distances from the vehicle center of mass to the front and rear axles; m is the vehicle mass, I z is the vehicle moment of inertia about the z axis at the center of mass; F yf, F yr are the lateral forces on the front and rear axles; F xf is the longitudinal force on the front axle; δ is the front wheel steering angle, and e is the lateral deviation; x LA is the longitudinal preview distance, and e LA is the lateral deviation at the preview point. The Fiala nonlinear tire model is introduced, that is: where C α and μ are the tire cornering stiffness and the road adhesion coefficient; F z is the tire vertical load; α is the tire cornering angle; α sat is the tire saturation cornering angle, and the front and rear tire cornering angle calculation formulas are: The longitudinal load transfer effect is introduced, and the vertical forces F zf ,F zr The calculation formula is as follows: In the formula, h CG It is the height of the vehicle's center of gravity, g is the acceleration due to gravity, and L is the vehicle's wheelbase.

2. The mechanism analysis-data driven, concatenated hybrid model of vehicle dynamics according to claim 1, characterized in that The data-driven model adopts a long short-term memory neural network model LSTM as the backbone network of the data-driven model, and combines a fully connected neural network to input the vehicle state and control data with 7 dynamic characteristics and 4 time step sequences into the data-driven model. The first layer LSTM encodes the data time series features, and the high-dimensional feature space is mapped through the fully connected layer. The second layer LSTM realizes feature decoding and is transmitted to the regression layer to calculate the final required yaw rate differential value and lateral velocity differential value.

3. The mechanism analysis-data driven, concatenated hybrid model of vehicle dynamics according to claim 1, characterized in that, It also includes training a neural network and updating weight parameters: Virtual running data of the vehicle mechanism analysis model and real vehicle state data of the intelligent vehicle in the actual driving process are collected for MD-SHM training; first, the virtual vehicle dynamics data is used for training to obtain a pre-trained model, and then the real vehicle dynamics data set is used to update the weight parameters of the neural network; A loss function is designed to adjust the internal parameters of the neural network so that the loss value decreases continuously. In the training process, the internal parameters of the network are adjusted through error back propagation to make the loss value decrease continuously. The smaller the loss value, the higher the fitting accuracy of the network to the data. The training loss function is defined using the Euler integral method as follows: In the formula, wherein, r, U y is the state measurement true value of vehicle yaw rate and lateral velocity, is the predicted value of vehicle yaw rate and lateral velocity at the next time of the network calculated by Euler integral of the model output, and N is the training number. The specific settings of network training: the loss function is selected as mean square error MSE, the optimizer is selected as Adam, the batch size is set to 1000, the learning rate is set to 0.001, and the network model is learned and trained based on the Pytorch learning framework.

4. An intelligent vehicle trajectory tracking control method based on a vehicle dynamics hybrid model, characterized in that, It includes: S1: Construct a mechanism analysis-data driven vehicle dynamics series hybrid model MD-SHM as claimed in any one of claims 1-3, the vehicle state and control data are calculated and processed by the mechanism model at the front end, combined and used as the input of the data-driven model, and the long short-term memory neural network is used to realize nonlinear correlation feature extraction of time series data and final model output calculation. S2: extract the weight parameters of the neural network to perform the forward calculation of the MD-SHM model, design a model predictive control trajectory tracking algorithm, design a feedforward feedback control algorithm to provide the external input required by the MD-SHM model in the lateral control while realizing the longitudinal control of the vehicle, and achieve precise trajectory tracking control effect. 5.The intelligent vehicle trajectory tracking control method based on a hybrid vehicle dynamics model according to claim 4, characterized in that, In the S2, the forward calculation process of the MD-SHM is as follows: [x t ,u t ]=[r,U y ,U x ,δ,F xf ] t h t = {X t ,…,X t-3} where x t is the vehicle state data for a single time step, u t is the vehicle control data for a single time step, is the single-track model computation data reference input, X t is the combined vehicle state and control data set for a level, h t is X t for multiple time steps W lstm{1,2} ∈(w i ,w f ,w g ,w o ),W lstm{1,2} ∈(w i ,w f ,w g ,w o ),W FC{1,2} ,b lstm{1,2} ∈(b i ,b f ,b g ,b o ),b FC{1,2} are the weight parameters of different network layers obtained by training, F lstm is the operation function abbreviation of the LSTM network, and z is the calculation output of different network layers. 6.The intelligent vehicle trajectory tracking control method based on a hybrid vehicle dynamics model according to claim 4, characterized in that, In the S2, the model predictive control trajectory tracking algorithm is designed as follows: A vehicle trajectory tracking error model is established: Considering the established mechanism-data serial hybrid vehicle dynamics model, the historical and current state and control variables are taken as the input of the MD-SHM model, and are represented as follows: h t = {X t ,…,X t-3} In the formula, the state quantity x(t) is [r, U y , x ] t , the control quantity u(t) is [δ, F xf ] t ; The front wheel steering angle of the vehicle in the trajectory tracking process is obtained by optimization of the model predictive control trajectory tracking algorithm, and the longitudinal control is realized separately.

7. The intelligent vehicle trajectory tracking control method based on a hybrid model of vehicle dynamics according to claim 4, characterized in that, In the S2, the feedforward feedback control algorithm is designed as follows: Based on the point mass assumption and the friction circle theory, the speed planning of the reference trajectory is performed, wherein the acceleration-limited friction circle is defined as follows: In the formula, a y is the lateral acceleration of the vehicle mass center, a x is the longitudinal acceleration of the vehicle mass center, μ is the road adhesion coefficient, and the vehicle acceleration limiting coefficient ξ is 0.533; The lateral acceleration of the vehicle when it is in a constant turning radius R is determined, and is shown in the following formula: When the vehicle tracks a variable-curvature trajectory, its longitudinal acceleration will change with the trajectory curvature κ, and by substituting formula (9) into formula (10) in the limited friction circle, we have: U x Taking the differential, we get: Substitute formula (12) into formula (11) to obtain: By performing reverse integration on formula (13) to solve the differential equation, the expected longitudinal speed of each trajectory point in the variable-curvature trajectory is obtained, which is: In the formula, Let L be the distance between two consecutive trajectory points. The total trajectory distance L is divided into n smaller trajectories formed by every two trajectory points. Therefore... At the end of the variable curvature trajectory The distance from where the inverse integration begins is the point where the integration ends at the starting point of the variable curvature trajectory; The variable-curvature reference trajectory is speed planned to obtain a desired longitudinal acceleration a x,des and longitudinal velocity U x,des curve, and substituted into the solution of the longitudinal control input. Based on the point mass assumption, the current speed of the vehicle, the driving position along the reference trajectory, the front wheel longitudinal force is solved, which consists of a feedforward term that compensates for the desired acceleration and resistance, and a feedback term that tracks the desired speed, as shown in the formula: where k p is the proportional gain of the difference between the planned desired speed and the measured actual speed, s t is the distance to the reference trajectory point reached at the travel time t, which is substituted into the calculation of the longitudinal speed external input required in the MPC prediction horizon is approximated as: The position of the next trajectory point in the prediction horizon is obtained by Euler integration: Δt denotes a discrete time interval, s t+1 , s t denotes the trajectory distance s at time t+1 and time t. 8.The intelligent vehicle trajectory tracking control method based on a hybrid vehicle dynamics model according to claim 4, characterized in that, The S2 also includes the discretization of the MD-SHM model: Using the Euler integration method, the continuous dynamics is converted into discrete dynamics, and in each stage of optimization, each discrete state and control variable is regarded as an optimization variable, as shown in the following formula: x(t + At) = x(t) + At - f MD-SHM (h t ) Since the input of the MD-SHM model includes historical state and control variables, when implemented in MPC, the Euler method is selected to realize the minimum number of calculations of the MD-SHM model in the optimization problem, and to calculate the discrete dynamics state information at the next stage within the optimization range; The discretized MD-SHM model is embedded in the model predictive control trajectory tracking algorithm to realize the prediction of the future trajectory error state: The MD-SHM model will be calculated in each optimization problem along the prediction horizon, and in the first stage of the optimization problem, the input of the MD-SHM model only includes the measured historical and current state information and the currently optimized control information; in the subsequent three stages, the model input is composed of the measured and optimized vehicle state and control information; from the fifth stage, the model input only includes the optimized vehicle state and control information, and the MPC prediction problem with 3 delay states shows how the input of each MD-SHM model changes in each stage and is finally composed of only the optimized state and control information; In each optimization problem, the MD-SHM model calculates the predicted value of the vehicle state at the next time, and these state values are substituted into the calculation to obtain the trajectory tracking error output of the vehicle at future time within the prediction horizon; wherein the prediction horizon is smaller than the control horizon, i.e. N c N p When the prediction horizon is larger than the control horizon at the optimization solution, the control increment is set to 0, i.e. u(t+i) = u(t), i = N c ,N c +1,..., N p -1; And the specific constraint conditions are designed as follows: (1) Restrict the front wheel steering angle control amount, the control amount obtained is less than the steering limit of the vehicle bottom layer actuator, the limit is ±32°, which is expressed as follows: δ min (t + j) < δ(t + j) < δ max (t + j), j = 0, 1,..., N c -1 (2) Restrict the front wheel steering angle control increment, the control increment in each sampling period is limited within a reasonable range to avoid sudden changes in the control amount and ensure the continuity of the control amount, the limit is ±2.25°, which is expressed as follows: Δδ min (t + j) ≤ Δδ(t + j) ≤ Δδ max (t + j), j = 0, 1,..., N c -1 (3) Restrict the center of mass side slip angle, the center of mass side slip angle is limited within a reasonable range to improve the stability of the vehicle, the vehicle is stably driven on a well-adhered dry asphalt road, the limit value of the center of mass side slip angle is ±12°; the vehicle is driven on an ice and snow road with a lower adhesion coefficient, the limit value is ±2°, which is expressed as follows: x β，min (t+j)≤x β (t+j)≤x β，max (t+j),j=0,1,…,N p -1 And design the target function in the following form: where the front wheel steering angle control increment Δδ is directly taken as the optimization variable, Q e Δψ Δδ are weight matrices of the lateral error e, the heading deviation Δψ and the control increment Δδ, respectively, N p c are the prediction horizon and the control horizon, respectively.​​​ Based on the above target function and constraint conditions, the trajectory tracking algorithm based on the MD-SHM model solves the following nonlinear optimization problem with constraints in each control period, as shown below: subject to h1 = h measure x(t + 1) = x(t) + At- f MD-SHM (h t ), t = 1,... N p -1 δ min (t)≤δ(t)≤δ max (t),t=1,,N c Δδ min (t)≤Δδ(t)≤Δδ max (t),t=1,…,N c x β，min (t)≤x β (k)≤x β，max (t),t=1,…,N p By solving the nonlinear optimization problem shown in the above formula in a certain control period, a series of control increments in the control time domain in the period are obtained: The first control increment in the sequence of control increments obtained is used to calculate the control amount actually acting on the controlled system, that is: δ(t) = δ(t-1) + Δδ(t) The vehicle will produce new dynamic state quantities under the action of the control amount and pass them to the next control period for the next optimization solution, so as to realize the trajectory tracking control of the intelligent automobile through the continuous solving and iteration of the above problem.

9. An intelligent vehicle trajectory tracking controller based on a hybrid model of vehicle dynamics, characterized in that, The controller is built-in with the control method of any one of claims 4-8.

Images