A lateral and longitudinal decoupling control method for lane-changing scenarios based on parameter adaptation and its application

Abstract

This invention belongs to the technical field of chassis control for autonomous vehicles and provides a method and application for lateral and longitudinal decoupling control of lane-changing scenarios based on parameter adaptation. First, based on a hierarchical planning lane-changing trajectory generation process, a reasonable lane-changing trajectory is generated in conjunction with a path planning structure. Then, based on a genetic optimization and neural network adaptive trajectory tracking process, optimal lane-changing control parameters are generated through test data optimization. Finally, a neural network algorithm is used to train a control parameter model to achieve an optimal control process for specific lane-changing conditions. This invention achieves high-real-time, high-comfort, and high-stability driving tasks by adaptively identifying lane-changing scenarios. It also adjusts its control behavior based on real-time identification of lane-changing conditions, thereby achieving high-performance control output.

Description

Technical Field

[0001] The present invention belongs to the technical field of chassis control of autonomous vehicles, and in particular relates to a lateral and longitudinal decoupling control method and application for lane change scenarios based on parameter adaptation. Background Art

[0002] Autonomous vehicles navigating complex traffic environments need to be able to flexibly change lanes based on real-time road conditions and the dynamics of surrounding vehicles. This process must not only ensure efficiency—switching from the current lane to the target lane as quickly as possible—but also ensure comfort, avoiding excessive acceleration, angular velocity, and lateral excursion. Therefore, lane changes require comprehensive lateral and longitudinal control processes to ensure the vehicle smoothly completes lane changes along a reasonable trajectory. Designing control methods tailored to lane-changing scenarios is a key research topic in the field of vehicle chassis control.

[0003] Due to the complexity and variability of lane change scenarios and the real-time requirements of the control process, existing lane change control methods are difficult to apply to all lane change scenarios. First, basic lane change control methods such as PID and LQR are mostly based on fixed control parameters. However, during actual testing, vehicle dynamics vary with factors such as load, speed, and road surface. These fixed-parameter algorithms cannot recognize the current state of the scene and are therefore incapable of addressing all lane change scenarios, resulting in reduced control effectiveness. Second, existing intelligent or nonlinear algorithms, such as reinforcement learning and novel MPC, can achieve control effectiveness through long-term computation. However, these algorithms incur significant computational overhead, impacting their real-time performance and hindering their practical application in the field of real-world vehicle control. Therefore, a lane change control method that balances algorithm efficiency and practical effectiveness is urgently needed. This method, which adaptively identifies lane change scenarios to achieve high-real-time, high-comfort, and high-stability driving tasks, and combines real-time identification of lane change conditions to adjust its control behavior, ultimately achieving high-performance control output, is needed. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention provides a lateral and longitudinal decoupling control method for lane changing scenarios based on parameter adaptation. First, based on the hierarchical planning lane change trajectory generation process, a reasonable lane change trajectory is generated in combination with the path planning structure. Then, based on the adaptive trajectory tracking process of genetic optimization and neural network, the optimal lane change control parameters are generated through test data optimization. Finally, the control parameter model is trained in combination with the neural network algorithm to achieve the optimal control process for specific lane changing conditions.

[0005] The present invention provides a method for lateral and longitudinal decoupling control in a lane change scenario based on parameter adaptation, which specifically includes the following steps:

[0006] Step 1: Lane change scenario initialization and trajectory cluster generation: Initialize the lane change scenario, abstract the lane change process of the autonomous vehicle into a vector transition process, convert the lane change trajectory into a combination of driving scenario elements, use a quintic polynomial for pre-planning, calculate multiple sets of applicable results, generate lane change trajectory clusters, and retain the lane change scenario elements for future use.

[0007] Preferably, the specific steps include:

[0008] The lane-changing process of an autonomous vehicle is abstracted as a vector transition process, that is, a trajectory of the motion vector from the starting point to the end point, which needs to meet the constraints of speed direction and position orientation. The speed direction of the host vehicle at the initial moment of lane change is defined as the x-axis, and the y-axis is obtained by rotating 90° counterclockwise. The lane-changing process is converted into a combination of the following scene elements: the initial speed of the host vehicle V1, the horizontal axis speed of the end point V x2 , terminal longitudinal axis speed V y2 , track longitudinal spacing D long , track lateral spacing D later ;

[0009] Based on the above scenario elements, use the fifth-order polynomial to calculate the results applicable to the array;

[0010] The quintic polynomial planning process is decomposed into three calculation formulas: position, velocity, and acceleration:

[0011] The position calculation formula is:

[0012] x(t)=a0+a1t+a2t 2 +a3t 3 +a4t 4 +a5t 5

[0013] y(t)=b0+b1t+b2t 2 +b3t 3 +b4t 4 +b5t 5

[0014] The speed calculation formula is:

[0015] x′(t)=a1+2a2t+3a3t 2 +4a4t 3 +5a5t 4

[0016] y′(t)=b1+2b2t+3b3t 2 +4b4t 3 +5b5t 4

[0017] The acceleration calculation formula is:

[0018] x″(t)=2a2+6a3t+12a4t 2 +20a5t 3

[0019] y″(t)=2b2+6b3t+12b4t 2 +20b5t 3

[0020] Where a0, a1, a2, a3, a4, a5, b0, b1, b2, b3, b4, and b5 are the coefficients of the polynomial; x(t) is the position of the object on the x-axis at time t; y(t) is the position of the object on the y-axis at time t; x′(t) is the first-order derivative of x(t), which is the instantaneous velocity of the object on the x-axis at time t; y′(t) is the first-order derivative of y(t), which is the instantaneous velocity of the object on the y-axis at time t; x″(t) is the second-order derivative of x(t), which is the instantaneous acceleration of the object on the x-axis at time t; y″(t) is the second-order derivative of y(t), which is the instantaneous acceleration of the object on the y-axis at time t.

[0021] The trajectory is expressed as the horizontal and vertical position, speed, and acceleration information after time t from the lane change starting point.

[0022] Step 2: Design trajectory evaluation indicators and select the optimal trajectory: Design comfort and traffic efficiency indicators to reasonably evaluate the lane change trajectory. The comfort indicator is expressed as the extreme value of the acceleration change rate during the lane change process, and the traffic efficiency indicator is expressed as the length of time to complete the lane change task. Design normalized weights, perform weighted calculations on each sub-indicator, and select the planned trajectory with the highest evaluation score as the optimal lane change trajectory.

[0023] Preferably, the specific steps include:

[0024] 21) The comfort index is expressed as the extreme value of the acceleration change rate during the lane change process. The real-time acceleration change rate in the lane change trajectory data is calculated using the following formula:

[0025]

[0026] Where, Indicates the extreme value of the acceleration rate of change during the lane change in this scenario, a i+Δt is the acceleration value at time i+Δt, a i is the acceleration value at time i, Δt is the time interval;

[0027] The traffic efficiency index is expressed as the length of time it takes to complete a lane change task. The lane change time is calculated by combining the relationship between distance and speed during the process. The calculation formula is as follows:

[0028]

[0029] Where T represents the lane changing time, s represents the total length of the lane changing trajectory, Indicates the average vehicle speed during the process.

[0030] 22) Assuming the weights of both indicators are 0.5, the calculation formula is as follows:

[0031]

[0032] S=0.5×S1+0.5×S2

[0033] Where S1 represents the normalized value of the acceleration change rate, Indicates the acceleration change rate during the current lane change process, represents the maximum permissible acceleration rate of change, S2 represents the standardized value of lane change time, Indicates the countdown of the current lane change process. represents the inverse of the maximum allowed time, S represents the final lane change trajectory score, and the best lane change trajectory with the highest evaluation score is selected as the pre-lane change trajectory.

[0034] Step 3: Design the trajectory tracking algorithm and obtain the parameters to be adjusted: Use the lateral and longitudinal decoupling control algorithm to separately control the lateral and longitudinal trajectory tracking processes. A PID algorithm that considers both speed and position loops is used in the longitudinal direction, while an LQR algorithm is used in the lateral direction. Build a lane change trajectory tracking control algorithm model to obtain the control parameters required for adjustment.

[0035] Preferably, the specific steps include:

[0036] 31) The inner loop of the longitudinal dual-loop PID algorithm is the speed loop, and the outer loop is the position loop. The parameters that need to be adjusted are: the proportional, differential, and integral coefficients of the inner and outer loops, a total of 6 parameters. The control parameters in the steps need to be adjusted for any scenario.

[0037] The output of the outer loop PID controller is:

[0038]

[0039] Where u1 is the intermediate variable, e1 is the outer loop error, K p1 , K i1 , K d1 are the proportional, integral and differential coefficients of the outer loop PID controller respectively.

[0040] The output of the inner loop PID controller is:

[0041]

[0042] Where u2 is the control signal, e2 is the inner loop error, K p2 , K i2 , K d2 are the proportional, integral and differential coefficients of the inner loop PID controller respectively.

[0043] The calculation method of inner and outer ring errors is as follows:

[0044] e1=yd-y

[0045] e2=u1-u

[0046] Where y d is the expected output of the system, y is the actual output of the system, and u is the actual input of the system.

[0047] 32) The lateral LQR algorithm obtains optimal control through linear quadratic adjustment. During control, the weight coefficient matrices Q and R are set in advance. Q and R are diagonal matrices, and the elements on the diagonal are selected for adjustment.

[0048] For the LQR algorithm, the basic steps are as follows:

[0049] First, the dynamic equations of the system are expressed as a state space model:

[0050] x(t+1)=Ax(t)+Bu(t)

[0051] Where x(t) is the state vector of the system, u(t) is the control input vector of the system, A and B are the matrices of the system, A is the state transfer matrix, and B is the control input matrix;

[0052] Secondly, the cost function J(u) is designed in quadratic form to measure the performance of the system:

[0053]

[0054] Among them, Q is the penalty weight matrix of the state error, R is the penalty weight matrix of the control input, L is the cross-term weight matrix between the state and control input, Q and R are both diagonal matrices and positive definite matrices;

[0055] Finally, the algebraic Riccati equation is used to solve the optimal feedback gain matrix to minimize the cost function, that is:

[0056] P=A T PA-A T PB(R+B T PB)- 1 B T PA+Q

[0057] Where P is a symmetric positive definite matrix;

[0058] K=(R+B T PB)- 1 B T PA

[0059] Where K is the optimal feedback gain matrix, which is used to calculate the optimal control input;

[0060] u(n)=-Kx(n)

[0061] Among them, x(n) is the state vector at the current moment, and u(n) is the optimal control input.

[0062] Step 4: Design control parameter ranges and trajectory tracking indicators: Design the value ranges of each control parameter and design trajectory tracking effect evaluation indicators from two aspects: trajectory tracking effect and lane change comfort.

[0063] Preferably, the specific steps include:

[0064] 41) Initialize the value range of each control parameter determined in step 3 and set a range limit for each control parameter: K p1 and K p2 Between 0.01 and 100, K i1 and K i2 Between 0.001 and 10, K d1 and K d2 Between 0.0001 and 1, due to the different dimensions, they are first logarithmized and then the subsequent genetic optimization operation is performed; the Q matrix and R matrix in the LQR algorithm are both symmetric positive definite matrices, and only the diagonal elements need to be adjusted. The dimension of the Q matrix is ​​consistent with the dimension of the system state vector, that is, if the state vector is n-dimensional, then the Q matrix is ​​an n×n matrix, and the dimension of the R matrix is ​​the same as the dimension of the system control vector, that is, if the control vector is m-dimensional, then the R matrix is ​​an m×m matrix; the values ​​of the diagonal elements of the Q and R matrices are both between 0 and 10.

[0065] 42) Design the evaluation indicators of the actual tracking effect of the trajectory, as follows:

[0066] Track tracking effect: It is characterized by the maximum value of the track deviation during the actual tracking control process:

[0067] Tra=max(e ij )

[0068]

[0069] Where Tra is the calculated maximum trajectory offset, e ij is the Euclidean distance between the theoretical trajectory and the actual trajectory calculated at time t.

[0070] Lane change comfort: characterized by the acceleration change during the actual lane change process:

[0071]

[0072] Step 5: Genetic optimization of the optimal control parameters in the lane change scenario: Through the three genetic operators of selection, crossover, and mutation, encoding and decoding operations, and individual fitness calculation, the optimal control parameters in this scenario are optimized.

[0073] Preferably, the specific steps include:

[0074] 51) Genetic operator initialization parameter design: Design encoding, decoding, selection, crossover, and mutation methods, and give the values ​​of genetic algorithm hyperparameters.

[0075] The designed encoding method is 0-1 binary encoding, which can restore the overall optimization effect of the genetic algorithm to the greatest extent; the decoding method is: converting from 0-1 encoding to the original real number value; the selection method is: roulette selection; the crossover probability is: single-point crossover, the crossover probability is 0.3; the mutation method is: single-point mutation, the mutation probability is 0.05.

[0076] 52) Based on the parameter value range in step 4, a genetic algorithm is used to quickly search for the genetic code with the highest fitness. For each lane change scenario combination A, the optimal control parameter combination B for this scenario is obtained through genetic algorithm optimization. The fitness function is the actual tracking effect evaluation indicator C designed above. The maximum number of iterations and population size are designed to find the algorithm parameters with the best control effect in this lane change scenario.

[0077] Step 6: Automatic cycle test of batch lane change scenario: the initial speed V1 of the main vehicle obtained in step 1, the end horizontal axis speed V x2 , terminal longitudinal axis speed V y2 , track longitudinal spacing D long , track lateral spacing D later A total of five lane-changing scenario elements are defined. The value range of each scenario element type is defined and discretely processed to obtain all lane-changing scenarios requiring calculation of optimal control parameters. A simulation test platform is selected, and an automated cyclic simulation test process is designed to perform automated trajectory generation and evaluation, trajectory tracking simulation testing, and control parameter genetic optimization. Combined with the automated simulation test platform, steps one, two, three, four, five, and six are repeated, and the test scenario and control parameter data are saved.

[0078] Preferably, the specific steps include:

[0079] 61) The value ranges of each scene element type are as follows: V1 and V x2 The value range is 0-30m / s, V y2The value range is 0-10m / s, D long The value range is 10-100m, D later The value range is 5-20m; the design speed discrete interval is 0.5m / s, and the distance discrete interval is 0.5m. By combining the above-mentioned multiple scenario elements, all lane change scenarios that require calculation of optimal control parameters are obtained.

[0080] 62) The simulation test platform is fundamental to the successful operation of the entire system. It requires the ability to generate and evaluate trajectories, simulate and test automated trajectory tracking algorithms, implement genetic algorithms, and subsequently train and apply neural networks. Considering these multiple requirements, the MATLAB computing platform combined with PreScan simulation software was selected as the preferred implementation.

[0081] 63) Combined with the automated simulation test platform, repeat steps 1, 2, 3, 4, 5, and 6 until the optimal control parameter calculation for the batch lane change scenario is completed. The scenario state parameters and the corresponding optimal control parameters are saved in pairs for the subsequent neural network training process.

[0082] Step 7: Preprocessing of lane change scenarios and control parameter data: Preprocess the data saved in step 6 and divide them into training and test sets.

[0083] Preferably, the specific steps include:

[0084] 71) The data saved in step 6 contains the following: the input parameters are the values ​​of the scene elements, and the output parameters are the optimal control parameters of the lane change trajectory tracking algorithm. The data is processed and all input and output parameters are normalized using the maximum and minimum normalization principle. The formula is as follows:

[0085]

[0086] Where x is a value of the original data, x' is the normalized value, max(X) and min(X) are the maximum and minimum values ​​of the original data, respectively.

[0087] 72) The above data is divided into training set and test set in proportion for subsequent training and testing of neural network algorithm.

[0088] Step 8. Initialize and train the BP neural network: Establish the network's input, output, and internal hierarchy. After training, save the network and use it for adaptive control in different lane-changing scenarios.

[0089] Preferably, the specific steps include:

[0090] 81) Select a BP neural network as the network structure for the control parameter prediction model. Ensure that the input of the BP neural network is the values ​​of the elements of the lane change scenario, with the number of input layer nodes matching the number of scene elements. The output is the values ​​of the control parameters for that scenario, with the number of output layer nodes matching the number of control parameters. Set the number of layers and nodes in the intermediate layers.

[0091] 82) Set the activation function, initial network weights, set network training parameters, start the network training process, wait for the training to complete and save the network model for adaptive control in different lane change scenarios.

[0092] Furthermore, the activation functions of the hidden layer and the output layer are both set to tansig functions; random numbers are used to initialize the network weights; and the network training parameters including the learning rate and the maximum number of iterations are set.

[0093] The neural network training process is carried out by combining the above network architecture and training data. After the training is completed, the network is saved and can be used for adaptive control in different lane change scenarios. Subsequently, the lane change scene elements can be directly input to output the control parameters.

[0094] The present invention also provides an application of a method for horizontal and vertical decoupling control of a lane-changing scenario based on parameter adaptation. The above-mentioned method for horizontal and vertical decoupling control of a lane-changing scenario based on parameter adaptation is applied in a specific lane-changing scenario. First, a specific lane-changing task is designed. An arbitrary lane-changing task is selected. The lane-changing trajectory is abstracted as a combination of the scene elements in step 1, that is, the initial vehicle speed V1 of the main vehicle, the terminal horizontal axis speed V x2 , terminal longitudinal axis speed V y2 , track longitudinal spacing D long , track lateral spacing D later The five scene elements are assigned values, and several trajectory paths are planned using step one. The optimal lane change trajectory is selected using the evaluation indicators in step two. Secondly, the lane change scenario is input into the trained neural network model, and the optimal control parameters under the current lane change state are output. The optimal control parameters are applied to the optimal lane change trajectory generated in the above steps to execute the tracking control process.

[0095] Beneficial effects of the present invention:

[0096] The present invention provides a method for decoupling lane-changing scenarios based on parameter adaptation. This method is a lane-changing control method that balances algorithm efficiency and actual effect. On the one hand, it generates a reasonable lane-changing trajectory based on a hierarchical planning process combined with a path planning structure. On the other hand, it uses a genetic optimization and neural network adaptive trajectory tracking process to generate optimal lane-changing control parameters through test data optimization, and trains the control parameter model in combination with a neural network algorithm to achieve an optimal control process for specific lane-changing conditions. The present invention conducts a comprehensive evaluation from the perspectives of comfort and traffic efficiency, and designs lane-changing trajectory evaluation indicators. To achieve better control effects, the present invention controls the horizontal and vertical trajectory tracking processes separately, using a PID algorithm that considers both speed and position in the longitudinal direction and an LQR algorithm in the lateral direction. High real-time, high comfort, and high stability driving tasks are achieved through adaptive identification of lane-changing scenarios, and the vehicle adjusts its own control behavior in combination with real-time identification of lane-changing conditions, thereby achieving high-performance control effect output. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 This is a schematic diagram of the overall flow of a lateral and longitudinal decoupling control method for lane changing scenarios based on parameter adaptation according to the present invention.

[0098] Figure 2 Schematic diagram of the lane change scenario deconstruction of the present invention.

[0099] Figure 3 Schematic diagram of the horizontal and vertical trajectory tracking process of the present invention.

[0100] Figure 4 Schematic diagram of the genetic optimization process of the present invention.

[0101] Figure 5 This is a schematic diagram of the deployment effect of the simulation test platform of the present invention.

[0102] Figure 6 This is a schematic diagram of the application effect of the BP neural network of the present invention. DETAILED DESCRIPTION

[0103] like Figure 1 As shown, the present invention provides a lateral and longitudinal decoupling control method for lane change scenarios based on parameter adaptation, which specifically includes the following steps:

[0104] Step 1: Lane change scenario initialization and trajectory cluster generation: Initialize the lane change scenario, abstract the lane change process of the autonomous vehicle into a vector transition process, convert the lane change trajectory into a combination of driving scenario elements, use a quintic polynomial for pre-planning, calculate multiple sets of applicable results, generate lane change trajectory clusters, and retain the lane change scenario elements for future use.

[0105] Traditional lane change control only considers the generation of a single lane change trajectory. This single lane change trajectory, calculated by fitting mathematical formulas based on the initial and target points, cannot fully meet the multiple requirements of comfort, safety, and traffic efficiency during the lane change process. Therefore, the present invention considers a multi-trajectory generation method that converts the trajectory fitting process into a polynomial equation solution process. The specific steps include:

[0106] like Figure 2 As shown in the figure, the lane change process of the autonomous vehicle is abstracted as a vector transition process, that is, a trajectory of the motion vector from the starting point to the end point, which needs to meet the constraints of speed direction and position orientation. The speed direction of the main vehicle at the initial moment of lane change is set as the x-axis, and the y-axis is obtained by rotating 90° counterclockwise. The lane change process is converted into a combination of the following scene elements: the initial speed of the main vehicle V1, the horizontal axis speed of the end point V x2 , terminal longitudinal axis speed V y2 , track longitudinal spacing D long , track lateral spacing D later ;

[0107] Based on the above scenario elements, 10 sets of applicable results were calculated using a quintic polynomial. The quintic polynomial has the advantages of third-order continuous differentiability and continuous curvature without sudden changes, which can better simulate actual lane change curves and decouple the horizontal and vertical directions.

[0108] The quintic polynomial planning process is decomposed into three calculation formulas: position, velocity, and acceleration:

[0109] The position calculation formula is:

[0110] x(t)=a0+a1t+a2t 2 +a3t 3 +a4t 4 +a5t 5

[0111] y(t)=b0+b1t+b2t 2 +b3t 3 +b4t 4 +b5t 5

[0112] The speed calculation formula is:

[0113] x′(t)=a1+2a2t+3a3t 2 +4a4t 3 +5a5t 4

[0114] y′(t)=b1+2b2t+3b3t 2 +4b4t 3 +5b5t 4

[0115] The acceleration calculation formula is:

[0116] x″(t)=2a2+6a3t+12a4t 2 +20a5t 3

[0117] y″(t)=2b2+6b3t+12b4t 2 +20b5t 3

[0118] Where a0, a1, a2, a3, a4, a5, b0, b1, b2, b3, b4, and b5 are the coefficients of the polynomial; x(t) is the position of the object on the x-axis at time t; y(t) is the position of the object on the y-axis at time t; x′(t) is the first-order derivative of x(t), which is the instantaneous velocity of the object on the x-axis at time t; y′(t) is the first-order derivative of y(t), which is the instantaneous velocity of the object on the y-axis at time t; x″(t) is the second-order derivative of x(t), which is the instantaneous acceleration of the object on the x-axis at time t; y″(t) is the second-order derivative of y(t), which is the instantaneous acceleration of the object on the y-axis at time t.

[0119] The trajectory is expressed as the horizontal and vertical position, speed, and acceleration information after time t from the lane change starting point.

[0120] Step 2: Design trajectory evaluation indicators and select the optimal trajectory: Design lane change trajectory evaluation indicators, conduct a comprehensive evaluation from the perspectives of comfort and traffic efficiency, and thus determine the optimal lane change trajectory. The specific steps include:

[0121] 21) Design comfort index and traffic efficiency index to reasonably evaluate the lane change trajectory. The comfort index is expressed as the extreme value of the acceleration change rate during the lane change process, and the traffic efficiency index is expressed as the length of time to complete the lane change task.

[0122] The comfort index is expressed as the extreme value of the acceleration change rate during the lane change process. The real-time acceleration change rate in the lane change trajectory data is calculated using the following formula:

[0123]

[0124] Where, Indicates the extreme value of the acceleration rate of change during the lane change in this scenario, a i+Δt is the acceleration value at time i+Δt, a i is the acceleration value at time i, Δt is the time interval;

[0125] The traffic efficiency index is expressed as the time it takes to complete a lane change task. The lane change time is calculated by combining the relationship between distance and speed during the process. The calculation formula is as follows:

[0126]

[0127] Where T represents the lane changing time, s represents the total length of the lane changing trajectory, Indicates the average vehicle speed during the process.

[0128] 22) Design normalized weights, perform weighted calculations on each sub-indicator, and select the planned trajectory with the highest evaluation score as the optimal lane change trajectory.

[0129] The weights of both indicators are set to 0.5, and the calculation formula is as follows:

[0130]

[0131] S=0.5×S1+0.5×S2

[0132] Where S1 represents the normalized value of the acceleration change rate, Indicates the acceleration change rate during the current lane change process, represents the maximum permissible acceleration rate of change, S2 represents the standardized value of lane change time, Indicates the countdown of the current lane change process. represents the inverse of the maximum allowed time, S represents the final lane change trajectory score, and the best lane change trajectory with the highest evaluation score is selected as the pre-lane change trajectory.

[0133] Step 3: Design the trajectory tracking algorithm and obtain the parameters to be adjusted: Use the lateral and longitudinal decoupling control algorithm to build a lane change trajectory tracking control algorithm model and obtain the control parameters required for adjustment.

[0134] The lateral and longitudinal motions are decoupled by a fifth-order polynomial. To achieve better control effects, the present invention controls the lateral and longitudinal trajectory tracking processes separately. The PID algorithm considering both speed and position loops is used for the longitudinal direction, while the LQR algorithm is used for the lateral direction. Figure 3 As shown, PID control is simple to operate and can quickly implement the control process, while the LQR algorithm has a fast nonlinear solution capability and can search for the optimal result under the current state. Therefore, the present invention designs the longitudinal control as a dual-loop PID algorithm and the lateral algorithm as an LQR algorithm. However, the overall control algorithm requires adjusting many parameters. The specific steps include:

[0135] 31) The inner loop of the longitudinal dual-loop PID algorithm is the speed loop, and the outer loop is the position loop. The parameters that need to be adjusted are: the proportional, differential, and integral coefficients of the inner and outer loops, a total of 6 parameters.

[0136] The control parameters in this step need to be adjusted for any scenario.

[0137] The output of the outer loop PID controller is:

[0138]

[0139] Where u1 is the intermediate variable, e1 is the outer loop error, K p1 , K i1 , K d1 are the proportional, integral and differential coefficients of the outer loop PID controller respectively.

[0140] The output of the inner loop PID controller is:

[0141]

[0142] Where u2 is the control signal, e2 is the inner loop error, K p2 , K i2 , K d2 are the proportional, integral and differential coefficients of the inner loop PID controller respectively.

[0143] The calculation method of inner and outer ring errors is as follows:

[0144] e1=yd-y

[0145] e2=u1-u

[0146] Where y d is the expected output of the system, y is the actual output of the system, and u is the actual input of the system.

[0147] 32) The lateral LQR algorithm obtains optimal control through linear quadratic adjustment, but the weight coefficient matrices Q and R need to be set in advance during control. Q and R are diagonal matrices, and the elements on the diagonal are selected for adjustment.

[0148] For the LQR algorithm, the basic steps are as follows:

[0149] First, the dynamic equations of the system are expressed as a state space model:

[0150] x(t+1)=Ax(t)+Bu(t)

[0151] Where x(t) is the state vector of the system, u(t) is the control input vector of the system, and A and B are the matrices of the system.

[0152] Secondly, the cost function J(u) is designed in quadratic form to measure the performance of the system:

[0153]

[0154] Among them, Q is the penalty weight matrix of the state error, R is the penalty weight matrix of the control input, L is the cross-term weight matrix between the state and control input, Q and R are both diagonal matrices and positive definite matrices;

[0155] Finally, the algebraic Riccati equation is used to solve the optimal feedback gain matrix to minimize the cost function, that is:

[0156] P=A T PA-A T PB(R+B T PB)- 1 B T PA+Q

[0157] Where P is a symmetric positive definite matrix;

[0158] K=(R+B T PB)- 1 B T PA

[0159] Where K is the optimal feedback gain matrix, which is used to calculate the optimal control input;

[0160] u(n)=-Kx(n)

[0161] Among them, x(n) is the state vector at the current moment, and u(n) is the optimal control input.

[0162] Step 4: Design of control parameter range and trajectory tracking index: Design the value range of each control parameter and trajectory tracking effect evaluation index. Specific steps include:

[0163] 41) Initialize the value range of each control parameter determined in step 3 and set a range limit for each control parameter: K p1 and K p2 Between 0.01 and 100, K i1 and K i2 Between 0.001 and 10, K d1 and K d2 Between 0.0001 and 1, due to the different dimensions, they are first logarithmized and then the subsequent genetic optimization operation is performed; the Q matrix and R matrix in the LQR algorithm are both symmetric positive definite matrices, and only the diagonal elements need to be adjusted. The dimension of the Q matrix is ​​consistent with the dimension of the system state vector, that is, if the state vector is n-dimensional, then the Q matrix is ​​an n×n matrix, and the dimension of the R matrix is ​​the same as the dimension of the system control vector, that is, if the control vector is m-dimensional, then the R matrix is ​​an m×m matrix; the values ​​of the diagonal elements of the Q and R matrices are both between 0 and 10.

[0164] 42) Design trajectory tracking effect evaluation indicators from two aspects: trajectory tracking effect and lane change comfort:

[0165] The actual trajectory tracking process involves the autonomous vehicle tracking the optimal lane change trajectory described above based on the control algorithm output. However, due to issues such as algorithm parameters, control effectiveness, and simulation accuracy, the trajectory tracking control results may be subject to certain errors and instabilities. This paper considers the potential instability of the trajectory tracking process and designs evaluation indicators for the actual trajectory tracking effect, specifically as follows:

[0166] Track tracking effect: It is characterized by the maximum value of the track offset during the actual tracking control process:

[0167] Tra=max(e ij )

[0168]

[0169] Where Tra is the calculated maximum trajectory offset, e ij is the Euclidean distance between the theoretical trajectory and the actual trajectory calculated at time t.

[0170] Lane change comfort: characterized by the acceleration change during the actual lane change process:

[0171]

[0172] The rate of change calculation method is the same as that in step 2, but here it is the rate of change of the vehicle itself during the actual trajectory tracking process.

[0173] Step 5: Genetic optimization of optimal control parameters in lane change scenarios: Through the three genetic operators of selection, crossover, and mutation, encoding and decoding operations, and individual fitness calculation, the optimal control parameters in this scenario are optimized. The specific steps include:

[0174] 51) Genetic operator initialization parameter design: Design encoding, decoding, selection, crossover, and mutation methods, and give the values ​​of genetic algorithm hyperparameters.

[0175] The designed encoding method is 0-1 binary encoding, which can restore the overall optimization effect of the genetic algorithm to the greatest extent; the decoding method is: converting from 0-1 encoding to the original real number value; the selection method is: roulette selection; the crossover probability is: single-point crossover, the crossover probability is 0.3; the mutation method is: single-point mutation, the mutation probability is 0.05.

[0176] 52) Based on the parameter value range in step 4, a genetic algorithm is used to quickly search for the genetic code with the highest fitness, thereby finding the algorithm parameters with the best control effect in the lane change scenario.

[0177] The optimal control parameter calculation process based on genetic algorithm is as follows Figure 4As shown in the figure, for each lane change scenario combination A, the optimal control parameter combination B in this scenario is obtained through genetic algorithm optimization. The fitness function is the actual tracking effect evaluation index C designed above. The maximum number of iterations is designed to be 100, and the population size is 50.

[0178] Step 6: Automatic cycle test of batch lane change scenario: Design a batch lane change scenario, repeat steps 1, 2, 3, 4, 5, and 6, and save the test scenario and control parameter data. The specific steps include:

[0179] 61) Define the value range of each scene element type abstracted in step 1 and perform discretization to obtain all lane change scenarios that require calculation of optimal control parameters. The steps are as follows: Step 1 obtains the initial speed V1 of the main vehicle, the end horizontal axis speed V x2 , terminal longitudinal axis speed V y2 , track longitudinal spacing D long , track lateral spacing D later There are a total of 5 lane change scene elements.

[0180] The design value range is as follows: V1 and V x2 The value range is 0-30m / s, V y2 The value range is 0-10m / s, D long The value range is 10-100m, D later The value range is 5-20m; the design speed discrete interval is 0.5m / s, and the distance discrete interval is 0.5m. By combining the above-mentioned multiple scenario elements, all lane change scenarios that require calculation of optimal control parameters are obtained.

[0181] 62) Select a suitable simulation test platform, design an automated cyclic simulation test process, implement automated trajectory generation and evaluation, trajectory tracking simulation test, and control parameter genetic optimization process.

[0182] The simulation test platform is the foundation for the successful operation of the entire system. It needs to have the ability to generate and evaluate trajectories, simulate and test automated trajectory tracking algorithms, implement genetic algorithms, and train and apply neural networks in subsequent steps. Considering the various task requirements, this embodiment uses the MATLAB computing platform combined with PreScan simulation software. The corresponding relationship between the overall algorithm and the platform is as follows: Figure 5 shown.

[0183] 63) Combined with the automated simulation test platform, repeat steps 1, 2, 3, 4, 5, and 6 until the optimal control parameter calculation for the batch lane change scenario is completed. The scenario state parameters and the corresponding optimal control parameters are saved in pairs for the subsequent neural network training process.

[0184] Step 7: Preprocessing of lane change scenarios and control parameter data: Preprocess the data saved in step 6 and divide it into training set and test set. The specific steps include:

[0185] 71) The data content saved in step 6 is: the input parameters are the values ​​of the scene elements, and the output parameters are the optimal control parameters of the lane change trajectory tracking algorithm.

[0186] When analyzing the data, different types of data have their own value ranges and discrete step lengths, which leads to data inconsistency. It is difficult to directly input the data into the neural network for training, so the data needs to be processed. All input and output parameters are normalized using the maximum and minimum normalization principle. The formula is as follows:

[0187]

[0188] Where x is a value of the original data, x' is the normalized value, max(X) and min(X) are the maximum and minimum values ​​of the original data, respectively.

[0189] 72) Divide the above data into training set and test set in proportion.

[0190] In this embodiment, all lane change scenarios and control parameter corresponding data are divided into training sets and test sets in a ratio of 8:2 for subsequent training and testing of the neural network algorithm.

[0191] Step 8: Initialize and train the BP neural network: Establish the network's input, output, and internal hierarchy. After training, save the network and use it for adaptive control in different lane-changing scenarios. Specific steps include:

[0192] 81) Figure 6 As shown in the figure, the BP neural network has excellent nonlinear prediction capabilities and good generalization ability. Therefore, the BP neural network is selected as the network structure of the control parameter prediction model. The input of the BP neural network is determined to be the values ​​of the elements of the lane change scenario, with the number of input layer nodes consistent with the number of scene elements. The output is the control parameter values ​​for the scenario, with the number of output layer nodes consistent with the number of control parameters. The number of layers and nodes in the intermediate layer is set.

[0193] In this embodiment, the number of input layer nodes is consistent with the number of scene elements, which is 5; the number of output layer nodes is consistent with the number of control parameters, which is 6+m+n, where m and n represent the dimensions of the LQR control vector and the state vector, respectively; the number of intermediate layers is set to 1, and the number of nodes is consistent with the number of output layer nodes.

[0194] 82) Set the activation function, initial network weights, learning rate, maximum number of iterations and other network training parameters, start the network training process, and save the network model after the training is completed for adaptive control in different lane change scenarios.

[0195] Since the present invention is a regression prediction problem, the activation functions of the hidden layer and the output layer are set to tansig functions. This function can provide nonlinear mapping capabilities and the gradient is relatively easy to calculate. Random numbers are used to initialize the network weights. The learning rate is set to 0.05 and the maximum number of iterations is 20,000.

[0196] The neural network training process is carried out by combining the above network architecture and training data. After the training is completed, the network is saved and can be used for adaptive control in different lane change scenarios. Subsequently, the lane change scene elements can be directly input to output the control parameters.

[0197] Step 9: Adaptive parameter control for lane-changing tasks: Design a specific lane-changing task and perform a complete trajectory generation-adaptive tracking parameter control process. Specific steps include:

[0198] The above steps provide an overall algorithm and model for trajectory generation and tracking control. This step explains how to apply this algorithm in a specific lane change scenario.

[0199] 91) Design a specific lane-changing task, abstract the lane-changing trajectory into a combination of the scene elements in step 1, use spline curves to plan multiple trajectories, and select the optimal lane-changing trajectory based on the indicators.

[0200] Choose any lane change task, that is, the initial speed of the main vehicle is V1, the end horizontal axis speed is V x2 , terminal longitudinal axis speed V y2 , track longitudinal spacing D long , track lateral spacing D later The five scene elements are assigned values, 10 paths are planned using step one, and the evaluation index of step two is used to select the best lane change trajectory for tracking.

[0201] 92) The lane change scenario is input into the trained neural network model, the optimal control parameters under the current lane change state are output, and the trajectory tracking control algorithm is applied to the specific lane change scenario.

[0202] After the trajectory generation is completed, the trajectory tracking control work can be performed. The values ​​of the above scene elements are input into the neural network model, and the corresponding control parameter values ​​can be output. After adjusting the parameters of each control algorithm, the optimal trajectory generated by the above steps can be used to execute the tracking control process.

Claims

1. A method for lateral and longitudinal decoupling control in lane change scenarios based on parameter adaptation, characterized by: The following steps are involved: Step 1: Lane change scenario initialization and trajectory cluster generation: Initialize the lane change scenario, abstract the lane change process of the autonomous vehicle into a vector transition process, and convert the lane change trajectory into a combination of lane change scenario elements, including the initial speed of the main vehicle. , terminal horizontal axis speed , terminal longitudinal axis speed , longitudinal spacing of tracks , track lateral spacing ; Use quintic polynomials for pre-planning, calculate array applicable results, generate lane change trajectory clusters, and retain lane change scenario elements for backup; Step 2: Design trajectory evaluation indicators and select the optimal trajectory: Design comfort and traffic efficiency indicators to reasonably evaluate the lane change trajectory. The comfort indicator is expressed as the extreme value of the acceleration change rate during the lane change process, and the traffic efficiency indicator is expressed as the length of time to complete the lane change task. Design normalized weights, perform weighted calculations on each sub-indicator, and select the planned trajectory with the highest evaluation score as the optimal lane change trajectory. Step 3: Design a trajectory tracking algorithm and obtain the parameters to be adjusted: A lateral and longitudinal decoupling control algorithm is used to separately control the lateral and longitudinal trajectory tracking processes. A PID algorithm that considers both speed and position is used in the longitudinal direction, while an LQR algorithm is used in the lateral direction. A lane change trajectory tracking control algorithm model is constructed to obtain the required control parameters. Step 4: Design control parameter ranges and trajectory tracking indicators: Design the value ranges of each control parameter and design trajectory tracking effect evaluation indicators based on trajectory tracking effect and lane change comfort. The steps are as follows: 41) Initialize the value range of each control parameter determined in step 3 and set a range limit for each control parameter: and Between 0.01 and 100, and Between 0.001 and 10, and Between 0.0001 and 1, first adjust the control parameters Calculate the logarithm and then perform subsequent genetic optimization operations; the Q matrix and R matrix in the LQR algorithm are both symmetric positive definite matrices, and only the diagonal elements are adjusted. The dimension of the Q matrix is ​​consistent with the dimension of the system state vector, that is, if the state vector is n-dimensional, then the Q matrix is ​​an n×n matrix, and the dimension of the R matrix is ​​the same as the dimension of the system control vector, that is, if the control vector is m-dimensional, then the R matrix is ​​an m×m matrix; the values ​​of the diagonal elements of the Q and R matrices are both between 0 and 10; 42) Design the trajectory tracking effect function as follows: Trajectory tracking effect function: It is characterized by the maximum value of the trajectory offset during the actual tracking control process: ， ， Where, is the calculated maximum value of the trajectory offset, is the Euclidean distance between the theoretical trajectory and the actual trajectory calculated at time t; Lane change comfort: characterized by the acceleration change during the actual lane change process: ， Step 5: Genetic optimization of optimal control parameters in the lane change scenario: Through the three genetic operators of selection, crossover, and mutation, encoding and decoding operations, and individual fitness calculation, the optimal control parameters for this scenario are optimized. The steps are as follows: 51) Genetic operator initialization parameter design: design encoding, decoding, selection, crossover, and mutation methods, and provide genetic algorithm hyperparameter values; The design encoding method is 0-1 binary encoding; the decoding method is: converting from 0-1 encoding to the original real number value; the selection method is: roulette selection; the crossover probability is: the crossover method is: single-point crossover; the mutation method is: single-point mutation; 52) Based on the range of control parameter values, a genetic algorithm is used to quickly search for the genetic code with the highest fitness. For each lane change scenario element combination A, the genetic algorithm is used to optimize the optimal control parameter combination B for this scenario. The fitness function is the trajectory tracking effect function C. The maximum number of iterations and population size are designed to find the algorithm parameters with the best control effect for this lane change scenario. Step 6: Automatic cycle test of batch lane change scenario: the initial speed of the main vehicle obtained in step 1 , terminal horizontal axis speed , terminal longitudinal axis speed , longitudinal spacing of tracks , track lateral spacing There are a total of five lane change scenario elements. The value range of each scenario element type is defined and discretized to obtain all lane change scenarios that require calculation of optimal control parameters. Select a simulation test platform, design an automated cyclic simulation test process, perform automated trajectory generation and evaluation, trajectory tracking simulation testing, and control parameter genetic optimization process; Combined with the automated simulation test platform, repeat steps 1, 2, 3, 4, 5, and 6, and save the test scenario and control parameter data; Step 7: Preprocessing lane change scenarios and control parameter data: Normalize the data saved in step 6 and divide it into training and test sets; Step 8. Initialize and train the BP neural network: establish the network's input, output, and internal layers; After training is complete, save the network and use it for adaptive control in different lane-changing scenarios.

2. The method for lateral and longitudinal decoupling control in lane change scenarios based on parameter adaptation according to claim 1, characterized in that: Step one includes: The lane-changing process of an autonomous vehicle is abstracted as a vector transition process, that is, a trajectory of the motion vector from the starting point to the end point, which needs to meet the constraints of speed direction and position orientation. The speed direction of the host vehicle at the initial moment of lane change is defined as the x-axis, and the y-axis is obtained by rotating 90° counterclockwise. The lane-changing process is converted into a combination of the following scene elements: the initial speed of the host vehicle , terminal horizontal axis speed , terminal longitudinal axis speed , longitudinal spacing of tracks , track lateral spacing ; Based on the above scenario elements, use the fifth-order polynomial to calculate the results applicable to the array; The quintic polynomial planning process is decomposed into three calculation formulas: position, velocity, and acceleration: The position calculation formula is: ， ， The speed calculation formula is: ， ， The acceleration calculation formula is: ， ， Where, are the coefficients of the polynomial; It's in time When the object is Position on the axis; It's in time When the object is Position on the axis; yes The first derivative of is at time When the object is Instantaneous speed on the axis; yes The first derivative of is at time When the object is Instantaneous speed on the axis; yes The second derivative of When the object is Instantaneous acceleration on the axis; yes The second derivative of When the object is Instantaneous acceleration on the axis; The trajectory is expressed as the horizontal and vertical position, speed, and acceleration information after time t from the lane change starting point.

3. The method for lateral and longitudinal decoupling control in lane change scenarios based on parameter adaptation according to claim 1, characterized in that: Step 2 includes: 21) The comfort index is expressed as the extreme value of the acceleration change rate during the lane change process. The real-time acceleration change rate in the lane change trajectory data is calculated using the following formula: ， Where, Indicates the extreme value of the acceleration change rate during the lane change in this scenario, for The acceleration value at the moment, for The acceleration value at the moment, is the time interval; The traffic efficiency index is expressed as the time it takes to complete a lane change task. The lane change time is calculated by combining the relationship between distance and speed during the process. The calculation formula is as follows: ， Where, represents the lane changing time, s represents the total length of the lane changing trajectory, Indicates the average vehicle speed during the process; 22) Assume that the weights of both indicators are 0.5, and the calculation formula is as follows: ， ， ， in, represents the normalized value of the rate of change of acceleration, Indicates the acceleration change rate during the current lane change process, Indicates the maximum permissible rate of acceleration change, represents the normalized value of lane change time, Indicates the countdown of the current lane change process. represents the inverse of the maximum allowed time, S represents the final lane change trajectory score, and the best lane change trajectory with the highest evaluation score is selected as the pre-lane change trajectory.

4. The method for lateral and longitudinal decoupling control in lane change scenarios based on parameter adaptation according to claim 1, characterized in that: Step three includes: 31) The inner loop of the longitudinal dual-loop PID algorithm is the speed loop, and the outer loop is the position loop. The parameters that need to be adjusted are: the proportional, differential, and integral coefficients of the inner and outer loops, a total of 6 parameters; The output of the outer loop PID controller is: ， Where, is an intermediate variable, is the outer loop error, are the proportional, integral and differential coefficients of the outer loop PID controller respectively; The output of the inner loop PID controller is: ， Where, is the control signal, is the inner loop error, are the proportional, integral and differential coefficients of the inner loop PID controller respectively; The calculation method of inner and outer ring errors is as follows: ， ， Where, is the expected output of the system, is the actual output of the system, is the actual input to the system; 32) The horizontal LQR algorithm obtains optimal control through linear quadratic adjustment. The weight coefficient matrices Q and R are set in advance during control. Q and R are diagonal matrices, and the elements on the diagonal are selected for adjustment. For the LQR algorithm, the basic steps are as follows: First, the dynamic equations of the system are expressed as a state space model: ， Where, is the state vector of the system, is the control input vector of the system, A and B are the matrices of the system, A is the state transfer matrix, and B is the control input matrix; Secondly, design the cost function in quadratic form , used to measure the performance of the system: ， Among them, Q is the penalty weight matrix of the state error, R is the penalty weight matrix of the control input, L is the cross-term weight matrix between the state and control input, Q and R are both diagonal matrices and positive definite matrices; Finally, the algebraic Riccati equation is used to solve the optimal feedback gain matrix to minimize the cost function, that is: ， Where P is a symmetric positive definite matrix; ， Where K is the optimal feedback gain matrix, which is used to calculate the optimal control input; ， in, is the state vector at the current moment, is the optimal control input.

5. The method for lateral and longitudinal decoupling control in lane change scenarios based on parameter adaptation according to claim 1, characterized in that: Step six includes: 61) The value ranges for each scene element type are as follows: and The value range is 0-30m / s, The value range is 0-10m / s, The value range is 10-100m. The value range is 5-20m; the design speed discrete interval is 0.5m / s, and the distance discrete interval is 0.5m. By combining the above multiple scenario elements, all lane change scenarios requiring calculation of optimal control parameters are obtained; 62) The simulation test platform is implemented by combining the MATLAB computing platform with the PreScan simulation software; 63) Combined with the automated simulation test platform, repeat steps 1, 2, 3, 4, 5, and 6 until the optimal control parameter calculation for the batch lane change scenario is completed. The scenario state parameters and the corresponding optimal control parameters are saved in pairs for subsequent neural network training.

6. The method for lateral and longitudinal decoupling control in lane change scenarios based on parameter adaptation according to claim 1, characterized in that: Step seven includes: 71) The data saved in step 6 contains the following: the input parameters are the values ​​of the scene elements, and the output parameters are the optimal control parameters of the lane change trajectory tracking algorithm. The data is processed and all input and output parameters are normalized using the maximum and minimum normalization principle. The formula is as follows: ， Where, is a value of the original data, is the normalized value, and are the maximum and minimum values ​​of the original data respectively; 72) Divide the above data into training set and test set in proportion for subsequent training and testing of neural network algorithm.

7. The method for lateral and longitudinal decoupling control in lane change scenarios based on parameter adaptation according to claim 1, characterized in that: Step eight includes: 81) Select a BP neural network as the network structure of the control parameter prediction model. The input of the BP neural network is determined to be the values ​​of the elements of the lane change scenario, with the number of input layer nodes consistent with the number of scene elements. The output is the control parameter values ​​for the scenario, with the number of output layer nodes consistent with the number of control parameters. Set the number of layers and nodes in the middle layer. 82) Set the activation function, initial network weights, set network training parameters, start the network training process, wait for the training to complete and save the network model for adaptive control in different lane change scenarios.

8. An application of a lateral and longitudinal decoupling control method for lane changing scenarios based on parameter adaptation, wherein the lateral and longitudinal decoupling control method for lane changing scenarios based on parameter adaptation according to any one of claims 1 to 7 is applied in a specific lane changing scenario, firstly designing a specific lane changing task, arbitrarily selecting a lane changing task, and abstracting the lane changing trajectory into a combination of the scenario elements in step 1, i.e., the initial speed of the host vehicle , terminal horizontal axis speed , terminal longitudinal axis speed , longitudinal spacing of tracks , track horizontal spacing The five scene elements are assigned values, and several trajectory paths are planned using step one. The optimal lane change trajectory is selected using the evaluation indicators in step two. Secondly, the lane change scenario is input into the trained neural network model, and the optimal control parameters under the current lane change state are output. The optimal control parameters are applied to the optimal lane change trajectory generated in the above steps to execute the tracking control process.

Images