Predictive control method for vehicle automatic driving based on dynamic neural network Hammerstein model

Abstract

The present invention belongs to the technical field of vehicle automatic driving prediction control, and discloses a vehicle automatic driving prediction control method based on a dynamic neural network Hammerstein model, including: S1. Constructing a Hammerstein model based on a dynamic neural network, the Hammerstein model includes a static nonlinear module and a dynamic linear module connected in series, and the static nonlinear module includes a multi-layer feedforward neural network and a map table for reflecting the relationship curve between the tire side slip angle and the tire lateral force; S2. Designing a prediction controller based on the Hammerstein model, and the prediction controller calculates the control quantity of the vehicle according to the expected state quantity and the actual state quantity output by the vehicle system in real-time feedback, and inputs the control quantity into the vehicle control system. In summary, the Hammerstein model of the present invention has the advantages of high modeling accuracy and rapidity, thereby effectively ensuring the accuracy of vehicle nonlinear modeling and reducing modeling complexity.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle automatic driving prediction control, and specifically relates to a vehicle automatic driving prediction control method based on a dynamic neural network Hammerstein model. Background Art

[0002] The vehicle dynamics model is of great significance for the realization of vehicle motion planning based on model predictive control. Tires and steering systems, as important subsystems of the vehicle system, have highly nonlinear characteristics. In order to improve the vehicle control accuracy, the nonlinear dynamic characteristics of the vehicle need to be fully considered.

[0003] Currently, in the field of vehicle autonomous driving controller design, vehicle dynamics modeling includes two types of models: linear and nonlinear. For specific working conditions, most studies use classic linear models such as 2DOF, 3DOF, and 5DOF to design controllers. Among them, the two-degree-of-freedom vehicle model takes into account the yaw and lateral degrees of freedom, and only considers the linear characteristics of the tires, simplifying the entire vehicle into a single vehicle model. The three-degree-of-freedom vehicle model adds the longitudinal degree of freedom, and the center of mass sideslip angle and yaw angular velocity will change with the longitudinal speed, which is more in line with the actual vehicle system. In addition, the vehicle model further considers the suspension and steering systems, and the analysis of the tire model is extended from the linear area to the nonlinear area, such as 12DOFs, 17DOFs, 19DOFs, etc. Compared with the linear model, the nonlinear model can analyze the motion state of the vehicle under combined working conditions.

[0004] In summary, for vehicle autonomous driving control, the requirements for vehicle modeling accuracy are increasing. Linear models are difficult to adapt to the requirements of controller design under complex working conditions. At the same time, the complexity of nonlinear vehicle coupling models poses higher challenges to controller design, and it is difficult to fully and accurately analyze the vehicle's lateral and longitudinal coupling characteristics. Therefore, in the design of vehicle autonomous driving controllers, facing the problem of difficult modeling of its complex nonlinear lateral and longitudinal coupling systems, the applicant proposed in the present invention a vehicle autonomous driving control method based on a dynamic neural network, constructing a Hammerstein lateral and longitudinal coupling model, and predictive control based on the lateral and longitudinal coupling model. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a vehicle automatic driving predictive control method based on a dynamic neural network Hammerstein model. Specifically, the present invention describes the nonlinear characteristics of tire side slip through a Map table, thereby reducing the difficulty of modeling the nonlinear side slip characteristics of the tire and effectively reducing the computational complexity.

[0006] To achieve the above purpose, the present invention provides the following technical solutions

[0007] A vehicle automatic driving predictive control method based on a dynamic neural network Hammerstein model, comprising:

[0008] S1. Constructing a Hammerstein model based on a dynamic neural network, wherein the Hammerstein model includes a static nonlinear module and a dynamic linear module connected in series, and the static nonlinear module includes a multi-layer feedforward neural network and a map table for reflecting a curve of a relationship between a tire slip angle and a tire lateral force;

[0009] S2. Design a predictive controller based on the Hammerstein model, and the predictive controller calculates the control quantity of the vehicle according to the expected state quantity and the actual state quantity output by the vehicle system in real-time feedback, and inputs the control quantity into the vehicle control system.

[0010] Preferably, the dynamic linear module is a linear state space equation, and the equation expression is:

[0011] In the formula, is the intermediate variable output by the static nonlinear module, A and B are constant matrices to be identified, and v x (k), v y (k) is the state quantity at time k, v x (k+1),v y (k+1), is the state quantity at time k+1.

[0012] Preferably, the intermediate variable output by the static nonlinear module hour:

[0013] Obtain the transition variable corresponding to the initial variable u2(k) through the Map table

[0014] The initial variable u1(k) and the transition variable are obtained by the multi-layer feedforward neural network The corresponding intermediate variable

[0015] Among them, the initial variables u1(k) and u2(k) are the resultant longitudinal force F of the vehicle on the X-axis at time k. x (k) and the front wheel steering angle δ f (k), transition variable Is the tire slip angle α in the Map table f Corresponding tire lateral force F y .

[0016] Preferably, the transition variable corresponding to the initial variable u2(k) is obtained through the Map table hour:

[0017] The conversion between front wheel steering angle and tire slip angle is as follows: l f is the distance from the center of mass of the vehicle to the front axle, and the tire slip angle is obtained through the current vehicle status information and the front wheel steering angle;

[0018] Set the tire slip angle range [-a,a];

[0019] The range of the tire slip angle is divided into N intervals, and each interval is numbered in turn. When the tire slip angle α is within the range of the slip angle [-a, a], the interval number is Store the corresponding tire lateral force F y .

[0020] Preferably, the Hammerstein model is combined with a soft constraint objective function in the predictive controller, the soft constraint objective function is used to correspond to the constraints on the desired state quantity, the actual state quantity and the control quantity increment, and the expression of the soft constraint objective function is:

[0021]

[0022] In the formula, k represents the current moment of the model, k+i represents the i-th moment predicted by the model, and x j (·) is the j-th predicted state quantity predicted by the Hammerstein model, r j (·) is the jth reference state quantity, Δu j (·) is the jth control variable change, Δu(n+i)=u(n+i)-u(n+i-1), and the control variables u1 and u2 are the resultant longitudinal forces F on the X-axis respectively. x and the front wheel steering angle δ f , N p is the prediction time domain, N c is the control time domain, Q is the predictive control output weight coefficient matrix, and S is the control increment weight coefficient matrix.

[0023] Preferably, the state quantities in the desired state quantity and the actual state quantity both include the longitudinal speed v x , lateral velocity v y and yaw rate

[0024] Preferably, the control variable includes the front wheel steering angle δ f and longitudinal force F x .

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] (1) High precision: The neural network has nonlinear mapping capabilities and a rich and variable network structure. The multi-layer feedforward neural network and the map table are combined into a static nonlinear module in the Hammerstein model, which can fully reflect the nonlinear coupling characteristics of the vehicle system.

[0027] (2) Reducing modeling complexity: In the present invention, the static nonlinear module and the dynamic linear module of the Hammerstein model are separately constructed, thereby effectively reducing the modeling complexity of the model of the present invention compared with the traditional nonlinear dynamic neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a schematic diagram of the predictive control principle of the present invention, in which MPC is a predictive controller;

[0029] Figure 2 It is a structural diagram of the Hammerstein model based on the dynamic neural network in the present invention;

[0030] Figure 3 A map table reflecting the relationship curve between tire slip angle and tire lateral force in the present invention;

[0031] Figure 4 The topological structure of the Hammerstein model of the dynamic neural network in the present invention;

[0032] Figure 5 This is a flow chart of the genetic algorithm in the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] See also Figure 1-5 As shown, the present invention provides a vehicle automatic driving predictive control method based on a dynamic neural network Hammerstein model. In the method, a Hammerstein model based on a dynamic neural network is first constructed, and then a predictive controller based on the Hammerstein model is designed. Finally, the predictive controller is used to realize the predictive control of the vehicle automatic driving.

[0035] S1. Hammerstein model based on dynamic neural network:

[0036] Specifically, according to Figure 2The designed Hammerstein model includes a static nonlinear module and a dynamic linear module connected in series.

[0037] a) The static nonlinear module includes a multi-layer feedforward neural network and a map table for reflecting the relationship curve between tire slip angle and tire lateral force;

[0038] In the Map table, data is read by hash table lookup method, including the following steps:

[0039] The conversion between front wheel steering angle and tire slip angle is as follows: l f is the distance from the center of mass of the vehicle to the front axle, and the tire slip angle is obtained through the current vehicle status information and the front wheel steering angle;

[0040] Set the tire slip angle range [-a,a];

[0041] The range of the tire slip angle is divided into N intervals, and each interval is numbered in turn. When the tire slip angle α is within the range of the slip angle [-a, a], the interval number corresponds to The data stored in the interval is the tire lateral force F y .

[0042] b) The dynamic linear module is a linear state space equation:

[0043] In the formula, is the intermediate variable output by the static nonlinear module, A and B are constant matrices to be identified, and v x (k), v y (k) is the state quantity at time k, v x (k+1),v y (k+1), is the state quantity at time k+1.

[0044] For the Hammerstein model designed above, its parameters should be trained before application. Specifically during training:

[0045] Provide the initial variables u1(k), u2(k) and the state v at time k x (k), v y (k) And about the initial variables u1(k), u2(k) and the state quantity v at time k x (k), v y (k) All can be obtained from actual vehicle systems or vehicle dynamics simulation software;

[0046] Obtain the transition variable corresponding to the initial variable u2(k) through the Map table

[0047] The initial variable u1(k) and the transition variable are obtained by the multi-layer feedforward neural network The corresponding intermediate variable

[0048] Generate the state v at time k+1 through the dynamic linear module x (k+1),v y (k+1),

[0049] Obtain the state quantity actually output by the vehicle at time k+1, and verify whether the error between the state quantity generated above and the state quantity actually output is within an acceptable error range. If yes, the training of the Hammerstein model is completed.

[0050] In the above, the initial variables u1(k) and u2(k) are the resultant longitudinal force δ of the vehicle on the X-axis at time k. f (k) and front wheel angle F x (k), transition variable Is the tire slip angle α in the Map table f Corresponding tire lateral force F y In addition, in the above training, the updating of weights and thresholds in the multi-layer feedforward neural network and the linear model is realized by a real-time loop learning algorithm.

[0051] exist Figure 4 In the figure, the topology of the dynamic neural network Hammerstein model consists of a Map and five neural network layers.

[0052] Input layer neurons

[0053] Hidden layer 1 neuron output Z (2) =f(ω 1 Z (1) +b 1 ), where ω 1 is the weight, b 1 is the threshold, and the activation function is the hyperbolic sine function

[0054] The output of the middle layer neuron is Z (3) =ω 2 Z (2) +b 2 ,ω 2 is the weight, b 2 is the threshold value;

[0055] The output of the neuron in hidden layer 2 is Z(4) =ω 3 Z (3) +νx(k),ω 3 ,ν are weights;

[0056] The output layer neuron outputs Z (5) =ω 4 Z (4) ,ω 4 is the weight.

[0057] Real-time recurrent learning algorithm: The layers in the recurrent neural network are connected in time, so the gradient updates the parameters in the network along the time channel. Based on the following optimization objectives, Where x(k) is the state quantity of the vehicle system's desired output and the state quantity output by the neural network Hammerstein model to update the i-th row and j-th column element ν in the weight ν ij For example, the update rules are as follows:

[0058]

[0059] η is the learning rate, is the derivative with respect to the time term.

[0060] in

[0061]

[0062]

[0063] In addition to the real-time loop learning algorithm, parameters can also be updated through time back-propagation algorithm and particle swarm algorithm.

[0064] S2. Design a predictive controller based on the Hammerstein model:

[0065] The above trained Hammerstein model is used as the design basis, combined with the soft constraint objective function to correspond to the constraint of the expected state quantity, the actual state quantity and the control quantity increment. The expression of the soft constraint objective function is:

[0066] In the formula, k represents the current moment of the model, k+i represents the i-th moment predicted by the model, and x j (·) is the j-th predicted state quantity predicted by the Hammerstein model, r j (·) is the jth reference state quantity, Δu j (·) is the jth control variable change, Δu(n+i)=u(n+i)-u(n+i-1), and the control variables u1 and u2 are the resultant longitudinal forces F on the X-axis respectively.x and the front wheel steering angle δ f , N p is the prediction time domain, N c is the control time domain, Q is the predictive control output weight coefficient matrix, and S is the control increment weight coefficient matrix.

[0067] S3. Use the above prediction controller to realize the prediction control of vehicle automatic driving:

[0068] The state variables include the longitudinal velocity v x , lateral velocity v y and yaw rate The control variables include the front wheel steering angle δ f and the front wheel longitudinal force F x .

[0069] In summary, combined Figure 1 The optimization calculation process of the predictive controller based on the Hammerstein model designed by the present invention is as follows:

[0070] The actual status information of the vehicle system is obtained by sensors;

[0071] Solve the optimal control quantity of the vehicle based on genetic algorithm;

[0072] Determine whether the optimal solution is feasible; if feasible, determine it as the control variable and act on the vehicle system; if not, solve it again.

[0073] Specific, combined Figure 5 Genetic algorithm flow chart, the solution process is as follows:

[0074] The population is initialized, the population size is N;

[0075] Evaluate the population and use the constrained objective function as the fitness of the individual;

[0076] Evolutionary computation includes selection, crossover, and mutation. For the selection operation, taking the roulette method as an example, the probability of individual i being selected is P i :

[0077]

[0078] Among them, F i is the fitness of individual i in the population, k is the coefficient;

[0079] Perform crossover operation, taking real number crossover method as an example, chromosome c m With c n The crossover operation at the jth position is as follows:

[0080] c mj =c mj (1-d)+cnj d

[0081] c nj =c nj (1-d)+c mj d

[0082] d is a random number in [0,1];

[0083] Mutation operation, gene c ij For example:

[0084]

[0085] Among them, c max and c min are the upper and lower bounds of the gene, G is the number of iterations, GEN is the maximum number of iterations, r is a random number in [0,1], and r1 is a random number in [0,1];

[0086] Determine whether the maximum number of iterations is met. If so, output the best solution; if not, return to the iteration.

[0087] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A vehicle automatic driving predictive control method based on a dynamic neural network Hammerstein model, characterized in that: include: S1. Constructing a Hammerstein model based on a dynamic neural network, wherein the Hammerstein model includes a static nonlinear module and a dynamic linear module connected in series, and the static nonlinear module includes a multi-layer feedforward neural network and a map table for reflecting a curve of a relationship between a tire slip angle and a tire lateral force; S2. Design a predictive controller based on the Hammerstein model, and the predictive controller calculates the control amount of the vehicle according to the expected state amount and the actual state amount output by the vehicle system in real-time feedback, and inputs the control amount into the vehicle control system; The dynamic linear module is a linear state space equation, and the equation expression is: In the formula, is the intermediate variable output by the static nonlinear module, A and B are constant matrices to be identified, and v x (k), v y (k) is the state quantity at time k, v x (k+1), v y (k+1), is the state quantity at time k+1; The intermediate variables output by the static nonlinear module hour: Obtain the transition variable corresponding to the initial variable u2(k) through the Map table The initial variable u1(k) and the transition variable are obtained by the multi-layer feedforward neural network The corresponding intermediate variable Among them, the initial variables u1(k) and u2(k) are the resultant longitudinal force F of the vehicle on the X-axis at time k. x (k) and the front wheel steering angle δ f (k), transition variable Is the tire slip angle α in the Map table f Corresponding tire lateral force F y ; Obtain the transition variable corresponding to the initial variable u2(k) through the Map table When , the hash search method is used, which includes the following steps: The conversion between front wheel steering angle and tire slip angle is as follows: l f is the distance from the center of mass of the vehicle to the front axle, and the tire slip angle is obtained through the current vehicle status information and the front wheel steering angle; Set the tire slip angle range [-a, a]; The range of the tire slip angle is divided into N intervals, and each interval is numbered in turn. When the tire slip angle α is within the range of the slip angle [-a, a], the interval number is After being processed by the hash function, the tire longitudinal force F stored in the output interval number is y ; The state quantities in the desired state quantity and the actual state quantity both include a longitudinal velocity, a lateral velocity, and a yaw angular velocity.

2. The vehicle automatic driving prediction control method based on the dynamic neural network Hammerstein model according to claim 1 is characterized in that: In the predictive controller, the Hammerstein model is combined with a soft constraint objective function, and the soft constraint objective function is used to correspond to the constraint of the desired state quantity, the actual state quantity and the control quantity increment, and the expression of the soft constraint objective function is: In the formula, k represents the current moment of the model, k+i represents the i-th moment predicted by the model, and x j (·) is the j-th predicted state quantity predicted by the Hammerstein model, r j (·) is the jth reference state quantity, Δu j (·) is the jth control variable change, Δu(n+i)=u(n+i)-u(n+i-1), and the control variables u1 and u2 are the resultant longitudinal forces F on the X-axis respectively. x and the front wheel steering angle δ f , N p is the prediction time domain, N c is the control time domain, Q is the predictive control output weight coefficient matrix, and S is the control increment weight coefficient matrix.

3. The vehicle automatic driving prediction control method based on the dynamic neural network Hammerstein model according to claim 1 is characterized in that: The control quantity includes the longitudinal force F x and the front wheel steering angle δ f .

Images