Event-triggered MPC collision avoidance method for autonomous vehicles based on neural networks

Abstract

This invention relates to a neural network-based event-triggered MPC collision avoidance method for autonomous vehicles, belonging to the field of autonomous vehicle collision avoidance technology. It includes: obtaining the presence of obstacles based on sensor modules; determining path planning trigger conditions when obstacles are present; training a neural network based on vehicle state data to obtain vehicle disturbance compensation when the path planning trigger conditions are met, and adding the vehicle disturbance compensation to the vehicle dynamics model; designing an MPC obstacle avoidance trajectory planner based on a cost function and constraints for the vehicle dynamics model, and outputting a planned reference trajectory. This invention adds disturbance compensation to the vehicle's nominal model, improving modeling accuracy, reducing trigger frequency, and decreasing computational load. Compared with traditional event-triggered MPC methods, neural network-based event-triggered MPC can reduce the number of triggers and decrease tracking errors.

Description

Technical Field

[0001] This invention relates to the field of collision avoidance technology for autonomous vehicles, and specifically to a neural network-based event-triggered MPC collision avoidance method for autonomous vehicles. Background Technology

[0002] In the field of autonomous driving, path planning and tracking have become research hotspots both domestically and internationally. Faced with complex traffic environments and roads, efficient collision avoidance strategies are crucial for autonomous vehicles to complete their designated tasks. Intelligent vehicles consider road conditions and vehicle dynamics to search for the optimal path, thereby adapting to different traffic scenarios and vehicle states. The motion planner, based on the vehicle's current position, target position, and surrounding environment information, uses algorithms and calculations to find a feasible collision-free path, guiding the object from its starting position to its destination. Simultaneously, the motion controller provides control force, ensuring the vehicle travels along a predetermined trajectory to achieve collision avoidance driving tasks. By proactively planning trajectories adapted to different traffic environments based on various driving scenarios and needs, the system addresses obstacle avoidance issues, avoids potential collision risks, ensures collision avoidance stability, and meets the real-time requirements of autonomous vehicles.

[0003] Traditional MPC algorithms plan and track collision-free trajectories at a fixed frequency, which cannot adapt to rapidly changing conditions in real-world driving environments. In particular, rapid changes may lead to discrepancies between the planned path and the actual required path. While event-triggered MPC algorithms execute vehicle motion and planning at different frequencies, their nominal models do not account for disturbance errors, resulting in control strategies that cannot adequately adapt to changes in the real-world environment. Therefore, it is necessary to incorporate real-time feedback to compensate for model uncertainties and dynamically adjust the planned path to make the generated control strategy more robust. Summary of the Invention

[0004] The technical problem to be solved by this invention is:

[0005] Traditional MPC algorithms, while capable of planning and tracking at a fixed frequency, cannot adapt to real-time environmental changes. Event-triggered MPC algorithms, which execute planning and control only when triggering conditions are met, offer greater flexibility in responding to changes, but they do not account for interference errors, resulting in poor robustness of the generated control strategy and an inability to effectively address uncertainties in real-world environments. To overcome these shortcomings, this invention provides a neural network-based event-triggered MPC collision avoidance method for autonomous vehicles.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A collision avoidance method for event-triggered MPC autonomous vehicles based on neural networks, characterized by comprising:

[0008] The presence of obstacles is determined based on the sensing module;

[0009] When obstacles are present, the path planning trigger conditions are determined;

[0010] When the path planning triggering condition is met, a neural network is trained based on vehicle state data to obtain vehicle disturbance compensation, and the vehicle disturbance compensation is added to the vehicle dynamics model.

[0011] Based on the cost function and constraints, an MPC obstacle avoidance trajectory planner is designed for the vehicle dynamics model, and the planned reference trajectory is output.

[0012] A further technical solution of the present invention: the determination of the path planning triggering condition specifically includes:

[0013] The system acquires the vehicle tracking error and compares it with a set threshold. When the vehicle tracking error exceeds the set threshold, path planning is triggered.

[0014] A further technical solution of the present invention: the vehicle tracking error is the error between the vehicle's planned reference path and the vehicle's actual path, specifically:

[0015]

[0016] Where Δ is the trigger level, O is the path tracking error weight matrix, and t is the weight matrix for adjusting the contribution of errors in different directions to the triggering conditions. k+1 ' represents the trigger time, indicating that the trajectory tracking error has reached the trigger level; ρ is the upper bound of the perturbation; β is the trigger parameter; T is the length of the prediction time; and I represents the neural network learning vector. It is the largest eigenvalue of the matrix. A is the state vector of the reference trajectory. k For the system matrix, B k For the control matrix, D k Represents the error matrix. These are the control input vectors, S and t. k These are the predicted end time and start time, respectively. The start time is also the event trigger time.

[0017] A further technical solution of the present invention: the neural network includes an input layer, a hidden layer, and an output layer;

[0018] The input vector i(t) of the input layer is [v x (t) v y (t) w(t) δ(t)F x [(t)], where v x (t), v y (t), w(t), δ(t), F x(t) represent the vehicle's longitudinal velocity, lateral velocity, angular velocity, front wheel steering angle, and tire longitudinal force, respectively;

[0019] The output data of the output layer is the difference between the updated predicted state and the actual measured state, which is used to train the weight parameters of the model for the next time step.

[0020]

[0021] in, These represent the longitudinal velocity, lateral velocity, and angular velocity corresponding to the reference path, respectively.

[0022] A further technical solution of the present invention: A Gaussian function is selected as the activation function for training the neural network, that is:

[0023]

[0024] Among them, c j Let σ be the center vector of the Gaussian function, representing the center position of the function in the feature space; j The width of the Gaussian function is denoted by ; the third output layer is the overall output of the neural network, calculated through a linear combination of the hidden layers to determine the network's predicted output.

[0025]

[0026] Where m is the number of hidden layer nodes, the training error can be expressed as:

[0027]

[0028] During training, the total error is minimized by adjusting the weight parameters, and the updated weight matrix is ​​substituted into the disturbance compensation term of the vehicle motion model to improve the accuracy of trajectory prediction.

[0029] A further technical solution of the present invention: The vehicle disturbance compensation is added to the vehicle dynamics model to obtain a vehicle dynamics model with the added vehicle disturbance compensation term.

[0030]

[0031] Where nn(·) is an additional neural network function used to estimate the modeling error of the nominal model. It is the derivative of the state vector of the reference trajectory. It is the state vector of the reference trajectory. It controls the input vector.

[0032] A further technical solution of the present invention: the MPC obstacle avoidance trajectory planner is specifically as follows:

[0033]

[0034]

[0035]

[0036]

[0037]

[0038] s∈[t k , t k +T]

[0039] in, The cost function J is the control input sequence that minimizes, driving the vehicle to follow the expected path and avoid obstacles. J is the total cost function.

[0040] A computer system is characterized by comprising: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method described above.

[0041] A computer-readable storage medium is characterized by storing computer-executable instructions, which, when executed, are used to implement the above-described method.

[0042] A computer program product is characterized by including computer-executable instructions, which, when executed, are used to implement the above-described method.

[0043] The beneficial effects of this invention are as follows:

[0044] This invention provides a neural network-based event-triggered MPC collision avoidance method for autonomous vehicles. The event-triggered MPC method ensures that the reference path is replanned only when the tracking error reaches a trigger level, thus performing path planning and motion control at different time intervals and improving the adaptability of trajectory tracking. It utilizes neural networks to approximate the uncertainties in vehicle modeling. Adding disturbance compensation to the vehicle's nominal model improves modeling accuracy, reduces the trigger frequency, and decreases computational load. Compared to traditional event-triggered MPC methods, the neural network-based event-triggered MPC can reduce the number of triggers and decrease tracking error.

[0045] 1. This invention utilizes the learning ability of neural networks to compensate for the uncertainty of modeling, and introduces the training results of neural networks into the cost function calculation of the model predictive control algorithm, thereby handling unpredictable environmental disturbances and improving the accuracy of the model.

[0046] 2. Vehicle kinematics and dynamics models were introduced into the model predictive control algorithm, and the objective function of the collision avoidance strategy was designed. Constraints on vehicle state and control inputs, considering actuator saturation, were formulated under the collision avoidance strategy to ensure the feasibility of the planned trajectory.

[0047] 3. By formulating triggering rules, the reference path is replanned only at the event trigger time. Vehicle motion planning and control are executed at different frequencies, which improves the vehicle's adaptability to trajectory tracking errors and achieves precise vehicle control. Attached Figure Description

[0048] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0049] Figure 1 Flowchart of the method of this invention.

[0050] Figure 2 Vehicle planar motion diagram.

[0051] Figure 3 Diagram of a single hidden layer neural network.

[0052] Figure 4 Vehicle lateral collision avoidance constraints.

[0053] Figure 5 Event-triggered MPC based on neural networks.

[0054] Figure 6 Tracking error graph of event-triggered MPC based on neural networks.

[0055] Figure 7 Trajectory tracking based on traditional event-triggered MPC.

[0056] Figure 8 Tracking error diagram for trajectory tracking based on traditional event-triggered MPC. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0058] This invention provides a neural network-based event-triggered MPC collision avoidance method for autonomous vehicles. It establishes a kinematic and dynamic model of the vehicle, considering driving safety and stability, and formulates vehicle state constraints and control input constraints to ensure the feasibility of trajectory planning. A vehicle motion model based on a neural network is established to estimate the modeling error of the nominal model, handle unpredictable environmental disturbances, and improve the model's accuracy. Triggering rules are formulated using path tracking errors, replanning the reference path only when triggering conditions are met, and performing vehicle motion and planning at different frequencies, thus improving the adaptability of trajectory tracking.

[0059] Includes the following:

[0060] 1. Vehicle kinematic and dynamic models

[0061] First, a vehicle motion model was designed, solving the path planning and tracking problems, enabling the designed motion controller to execute and track a reference collision-free trajectory. The vehicle's planar motion is as follows: Figure 2 As shown, G represents the vehicle's center of gravity, l f and l r The distance from the vehicle's center of gravity to the front and rear wheels is represented by δ, where δ is the front wheel steering angle and ψ is the vehicle yaw angle. r Let be the desired yaw angle. Then the vehicle kinematic equation can be expressed as:

[0062]

[0063] The vehicle dynamics equations are:

[0064]

[0065] Where m is the total mass of the vehicle, I z Let F be the vehicle's vertical moment of inertia. xf (t) and F xr (t) represents the longitudinal tire force of the front and rear wheels, respectively, F yf (t) and F yr (t) represent the lateral tire forces of the front and rear wheels, respectively. This dynamic model can more accurately describe the basic characteristics of a vehicle traveling at high speed. Since vehicle motion control includes longitudinal adjustment and lateral tracking, to simplify the controller design, the vehicle's front wheel steering angle and tire longitudinal force are used as control inputs, resulting in the state-space expression of the vehicle system:

[0066]

[0067] The above expression is a practical model of vehicle motion, including indescribable external disturbances. D k Let sinψ represent the error matrix. Under normal driving conditions, since the heading angle error is small, we assume sinψ e ≈ψe ,cosψ e If ≈1, then the system matrix is ​​represented as:

[0068]

[0069] The control matrix is ​​represented as:

[0070]

[0071] Among them, C f and C r These represent the lateral stiffness of the front and rear wheels, respectively.

[0072] 2. Interference Error Compensation Based on Neural Networks

[0073] Since actual vehicle motion is affected by uncertain environmental disturbances, adding disturbance compensation to the vehicle dynamics model can reduce prediction errors and improve model accuracy. For a nominal vehicle system without modeling errors and external disturbances, the state-space expression is:

[0074]

[0075] Design a vehicle motion model with disturbance compensation, whose state-space expression is as follows:

[0076]

[0077] Where nn(·) is an additional neural network function used to estimate the nominal modeling error. Assuming that the vehicle model error is mainly caused by the dynamics of the system equations, the input vector for training the neural network is independently extracted from the vehicle model, i.e.:

[0078] i(t) = [v x (t) v y (t) w(t) δ(t) F x (t)] (8)

[0079] Among them, v x (t), v y (t), w(t), δ(t), F x (t) represent the vehicle's longitudinal velocity, lateral velocity, angular velocity, front wheel steering angle, and tire longitudinal force, respectively;

[0080] The difference between the updated predicted state and the actual measured state is used as the output layer data for online learning of the neural network, and is used to train the weight parameters of the model for the next time step.

[0081]

[0082] To simplify computation, a single-hidden-layer neural network approximator is chosen to handle uncertainties in structure and parameters. The network structure is as follows: Figure 3 As shown.

[0083] This invention selects a Gaussian function as the activation function for the neural network, that is:

[0084]

[0085] Among them, c j Let σ be the center vector of the Gaussian function, representing the center position of the function in the feature space. j This represents the width of the Gaussian function. The third output layer is the overall output of the neural network, calculated through a linear combination of the hidden layers to determine the network's predicted output.

[0086]

[0087] Where m is the number of hidden layer nodes, the training error can be expressed as:

[0088]

[0089] During training, the total error is minimized by adjusting the weight parameters, and the updated weight matrix is ​​substituted into the disturbance compensation term of the vehicle motion model to improve the accuracy of trajectory prediction.

[0090] 3. Path planning trigger rule formulation

[0091] By designing trigger rules, it's possible to determine when to activate a new collision avoidance strategy, thus allowing for flexible adjustment of the frequency of path planning and control. The basic process is as follows:

[0092] 1. Define the triggering conditions for activating a new collision avoidance strategy based on vehicle motion status information;

[0093] 2. When the triggering condition is met, the path planning algorithm is started to find a safe and feasible path, avoid obstacles, and reach the target point within the given time range;

[0094] 3. Considering the vehicle's dynamic constraints and environmental changes, the model predictive control algorithm is used to calculate the control force at each time step on the planned path and generate an adaptive control strategy.

[0095] 4. At each time step, control the vehicle according to the calculated control force to make it move along the planned trajectory;

[0096] 5. Set the next event trigger point. When the vehicle reaches this point, activate the new model predictive control algorithm for path planning and control, and repeat the above process until the vehicle reaches its destination.

[0097] Through the above steps, asynchronous execution of path planning is achieved. This invention considers external interference and modeling errors, and utilizes path tracking errors to formulate triggering rules. The triggering time is defined as:

[0098]

[0099] Where Δ represents the trigger level, and O is the path tracking error weight matrix, adjusting the contribution of errors in different directions to the triggering conditions. k+1 ' is the trigger time, indicating that the trajectory tracking error has reached the trigger level. The trigger level Δ can be expressed as:

[0100]

[0101] Where ρ is the upper bound of the perturbation, and β is the triggering parameter. The triggering time is t when the triggering condition is met within the prediction time range. k+1 This means the trajectory tracking error reaches the trigger level; furthermore, a reference trajectory needs to be generated at the initial position, and vehicle motion planning needs to be performed when the vehicle reaches the end of the planned trajectory. The prediction time is defined as T, and if the triggering rule is not met within the prediction range, the trigger time is t. k +T. The total trigger condition expression is as follows:

[0102] t k+1 =min(t) k+1 ',t k +T) (15)

[0103] 4. Collision Avoidance Controller Design

[0104] To ensure driving safety, the vehicle's status needs to be limited to a permissible range, allowing the vehicle to change paths along a planned trajectory.

[0105] The vehicle restrictions are as follows:

[0106]

[0107] The above constraints can be used to avoid potential collisions with obstacles. These constraints can be summarized as follows:

[0108]

[0109] The goal of a collision avoidance strategy is to find a collision-free path to the destination and calculate the control parameters within the predicted horizon. The cost function is defined by considering both control cost and drive stability; its expression is:

[0110]

[0111] In the formula, R is the control force weight matrix; M is the control input change weight matrix; Q is the state deviation weight matrix; T is the prediction time domain length; and J is the total cost function.

[0112] Based on the vehicle motion model, combined with the cost function and constraints, the expression for the collision avoidance strategy is obtained as follows:

[0113]

[0114] By solving the above optimization problem, the control variables for planning the path within the prediction range were obtained, thus achieving the goal of driving the vehicle to follow the expected path and avoid obstacles.

[0115] Based on the above theoretical description and formula derivation, an event-triggered MPC collision avoidance method for autonomous vehicles based on neural networks is obtained according to an embodiment of the present invention, specifically including the following steps:

[0116] Step 1: Determine the presence of obstacles based on the sensing module;

[0117] Step 2: When obstacles exist, determine the path planning trigger conditions;

[0118] Step 3: When the path planning triggering condition is met, train a neural network based on vehicle state data to obtain vehicle disturbance compensation, and add the vehicle disturbance compensation to the vehicle dynamics model.

[0119] Step 4: Based on the cost function and constraints, design the MPC obstacle avoidance trajectory planner for the vehicle dynamics model and output the planned reference trajectory.

[0120] 5. Simulation Experiment

[0121] In the simulation environment, considering lane-changing driving scenarios with road obstacles, a neural network-based MPC algorithm is used to generate a reference trajectory and the actual tracking path. The reference trajectory is generated when the tracking error reaches a threshold. Alternatively, vehicle motion planning is also required when the vehicle reaches the end of the planned trajectory. In the simulation, the prediction step size is set to 30, and the target lane centerline is set to y = 4. The simulation results are as follows: Figure 5 As shown:

[0122] As shown in the diagram above, the algorithm plans a total of three reference paths. The first reference trajectory is generated at the initial position, and the planned path can avoid obstacles; however, tracking errors are inevitable during the tracking process. Once a trigger condition is met, motion planning is re-performed at a longitudinal position of 50m, generating corresponding control forces to drive the vehicle to follow the new reference path with minimal error. The third trigger point is activated at a longitudinal position of 108m, and the generated reference path guides the vehicle to travel smoothly along the center line of the target lane. The tracking error diagram is shown below:

[0123] exist Figure 6 During the process, the path tracking error remained within 0.15m. The lower bound of the tracking error was reached at the second trigger point (longitudinal position 50m), and the upper bound was reached at the third trigger point (longitudinal position 108m). Because the reference path was replanned after the trigger conditions were met, the error remained within the allowable range. Compared to the MPC collision avoidance method proposed in this invention, traditional event-triggered MPC is used for planning and tracking within the same driving scenario. The tracking trajectory is shown in the figure below:

[0124] Depend on Figure 7 As can be seen, under event-triggered MPC, the planner triggered and generated a total of 5 reference paths. The final path after vehicle stabilization deviated somewhat from the target lane centerline. This is because, without neural network simulation of uncertainty, the motion planner cannot make corresponding compensation adjustments based on feedback from actual tracking errors.

[0125] Tracking error such as Figure 8 As shown, from Figure 8 It can be seen that the method is poorly adaptable to tracking errors and the tracking error is large. The event-triggered MPC obstacle avoidance strategy based on neural networks makes up for the shortcomings of the method.

[0126] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. A collision avoidance method for event-triggered MPC autonomous vehicles based on neural networks, characterized in that, include: The presence of obstacles is determined based on the sensing module; When obstacles are present, the path planning trigger conditions are determined; The determination of the path planning trigger conditions is specifically as follows: The vehicle tracking error is obtained, and a judgment is made between the vehicle tracking error and a set threshold. When the vehicle tracking error exceeds the set threshold, path planning is triggered. The vehicle tracking error is the error between the planned reference path and the actual path taken by the vehicle, specifically: Where Δ represents the trigger level, and O is the path tracking error weight matrix, which adjusts the contribution of errors in different directions to the triggering conditions. The trigger time indicates that the trajectory tracking error has reached the trigger level. ρ The upper bound of the perturbation. β The trigger parameter is T, where T is the length of the prediction time. Represents the learning vector of the neural network. It is the largest eigenvalue of the matrix. It is the state vector of the reference trajectory. For the system matrix, For the control matrix, Represents the error matrix. It is the control input vector. S , t k These are the predicted end time and start time, respectively, with the start time also being the event trigger time; When the path planning triggering condition is met, a neural network is trained based on vehicle state data to obtain vehicle disturbance compensation, and the vehicle disturbance compensation is added to the vehicle dynamics model. An MPC obstacle avoidance trajectory planner is designed based on the cost function and constraints of the vehicle dynamics model, and outputs a planned reference trajectory; the MPC obstacle avoidance trajectory planner is specifically as follows: in, The cost function J is minimized to correspond to the control input sequence, which drives the vehicle to follow the expected path and avoid obstacles. J Let be the total cost function.

2. The event-triggered MPC collision avoidance method for autonomous vehicles based on neural networks according to claim 1, characterized in that, The neural network includes an input layer, a hidden layer, and an output layer; The input vector of the input layer ,in, , , , , These represent the vehicle's longitudinal velocity, lateral velocity, angular velocity, front wheel steering angle, and tire longitudinal force, respectively. The output data of the output layer is the difference between the updated predicted state and the actual measured state, which is used to train the weight parameters of the model for the next time step. in, , , These represent the longitudinal velocity, lateral velocity, and angular velocity corresponding to the reference path, respectively.

3. The event-triggered MPC collision avoidance method for autonomous vehicles based on neural networks according to claim 2, characterized in that, When training a neural network, a Gaussian function is chosen as the activation function, that is: in, Let be the center vector of the Gaussian function, representing the center position of the function in the feature space; The width of the Gaussian function is denoted by ; the third output layer is the overall output of the neural network, calculated through a linear combination of the hidden layers to determine the network's predicted output. in, m Let be the number of hidden layer nodes. The training error can be expressed as: During training, the total error is minimized by adjusting the weight parameters, and the updated weight matrix is ​​substituted into the disturbance compensation term of the vehicle motion model to improve the accuracy of trajectory prediction.

4. The event-triggered MPC collision avoidance method for autonomous vehicles based on neural networks according to claim 1, characterized in that, Adding the vehicle disturbance compensation to the vehicle dynamics model yields a vehicle dynamics model with the added vehicle disturbance compensation term: in, It is an additional neural network function used to estimate the modeling error of the nominal model. It is the derivative of the state vector of the reference trajectory. It is the state vector of the reference trajectory. It controls the input vector.

5. A computer system, characterized in that... include: One or more processors, a computer-readable storage medium for storing one or more programs, wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of claim 1.

6. A computer-readable storage medium, characterized in that... The device stores computer-executable instructions, which, when executed, are used to implement the method of claim 1.

7. A computer program product, characterized in that... It includes computer-executable instructions, which, when executed, are used to implement the method of claim 1.

Images