Event-triggered trajectory tracking method, system, and medium for a distributed drive vehicle

Abstract

This invention provides an event-triggered trajectory tracking method, system, and medium for a distributed drive vehicle. The trajectory tracking method includes: using an MPC controller to calculate a predicted state sequence x of the vehicle based on a reference trajectory. ref and control command sequence u. When the MPC controller triggers, it controls x. ref The system updates the parameter u and maps u to the torque distribution of the four wheels to achieve reference trajectory tracking. When the MPC controller is in sleep mode, it uses the previously updated u to achieve reference trajectory tracking. In this invention, when the MPC is in sleep mode, it directly extracts the corresponding control commands sequentially from the cached control command sequence based on the time step and issues them to the four independent wheels. Furthermore, it fuses the data-driven global optimization RL triggering mechanism with the rule-based triggering mechanism of multi-dimensional physical rules through a logical OR gate, significantly reducing the triggering frequency of the MPC controller while ensuring vehicle tracking performance.

Description

Technical Field

[0001] This invention relates to the field of vehicle control technology, and in particular to an event-triggered trajectory tracking method, system, and medium for distributed drive vehicles. Background Technology

[0002] Distributed drive vehicles offer a high degree of freedom in vehicle trajectory tracking control due to the independent controllability of torque at each wheel. However, under complex driving conditions, existing model predictive control (MPC) architectures and event-triggered mechanisms face the following pressing technical challenges: 1. Open-loop integral offset problem in differential steering systems. To reduce the computational load on the onboard controller, event-triggered mechanisms are often introduced into the MPC framework. However, unlike traditional position-based direct mechanical steering, differential steering is essentially a "force control" system. During the "dormant (open-loop hold)" phase of event triggering, the rate of change of the front wheel steering angle is determined by both the issued differential torque and the actual dynamic return torque. Since traditional predictive models cannot perfectly cover the dynamic delay in the establishment of wheel rotational inertia and tire slip ratio, this leads to a slight mismatch in reality. Under the action of the integral element inherent in differential steering, this torque deviation accumulates rapidly over time, causing the actual front wheel steering angle to deviate significantly from the expected state. If only the traditional external lateral tracking error is used as the trigger condition, when the system is awakened, the lateral error is already very large, causing the controller to frequently output excessive extreme torque for violent correction, which can easily lead to severe overshoot and oscillation.

[0003] 2. Global optimization bottleneck between computing power allocation and tracking accuracy. Existing event triggering strategies mostly adopt fixed rules or fixed thresholds based on physical states. Setting thresholds too strictly will lead to a persistently high triggering frequency, failing to release computing power; setting thresholds too loosely will sacrifice tracking accuracy when facing complex curvatures or road disturbances. Single rule constraints are difficult to adapt to the computing power requirements under all operating conditions, and lack the ability to autonomously explore and dynamically seek the optimal balance between computing frequency and tracking performance while ensuring the bottom line of system dynamic safety. Summary of the Invention

[0004] To address the technical problems of high computational load in traditional MPC and the inability of differential steering systems to guarantee trajectory tracking accuracy at low computational frequencies, this invention provides an event-triggered trajectory tracking method, system, and medium for distributed drive vehicles.

[0005] This invention is achieved through the following technical solution: Firstly, this invention proposes an event-triggered trajectory tracking method for distributed drive vehicles, which uses an MPC controller based on vehicle differential steering dynamics to calculate the vehicle's predicted state sequence x according to the vehicle's reference trajectory. refand control command sequence u. When the MPC controller triggers, it controls x. ref The MPC controller updates u and maps u to the torque distribution of the four wheels to achieve reference trajectory tracking. When the MPC controller is in sleep mode, it uses the previously updated u to achieve reference trajectory tracking. The methods for triggering the MPC controller include: triggering the MPC controller to update when a rule-triggered mechanism or RL-triggered mechanism outputs a trigger command.

[0006] The rule-triggered mechanism includes: calculating the vehicle's current position based on the allocated torque, and comparing it with the reference trajectory to calculate the lateral error e. y (t). x ref Predicted front wheel steering angle δ pred (t) and the current true front wheel steering angle δ real (t) Compare and calculate the angle prediction error e δ (t). The current heading angle is compared with the forward aiming time T of the reference trajectory. preview The heading angle is compared to calculate the abrupt change in heading angle ΔΨ. When e y (t), e δ When (t) and ΔΨ are greater than the corresponding threshold, a trigger command is output.

[0007] The RL triggering mechanism includes: designing a multi-objective reward function and using it to train the network parameters of the RL agent. The trained RL agent infers whether to trigger based on the vehicle state and the predicted state. The multi-objective reward function includes: a rule overreach penalty R when the rule-based triggering mechanism is triggered. rule , when e y (t) Training round termination penalty R when the corresponding threshold is exceeded fail .

[0008] Secondly, this invention proposes an event-triggered trajectory tracking system for a distributed drive vehicle, which uses the event-triggered trajectory tracking method for a distributed drive vehicle described in the first aspect. The event-triggered trajectory tracking system includes: a trajectory tracking controller and an adaptive dual event trigger. The trajectory tracking controller uses an MPC controller based on vehicle differential steering dynamics to calculate the vehicle's predicted state sequence x according to the vehicle's reference trajectory. ref and control command sequence u. When the MPC controller triggers, it controls x. ref The system updates `u` and maps it to the torque distribution of the four wheels to achieve reference trajectory tracking. When the MPC controller is in sleep mode, it uses the previously updated `u` to achieve reference trajectory tracking. An adaptive dual-event trigger is used to activate the MPC controller when a trigger command is output by either the rule-triggered mechanism or the RL-triggered mechanism. The rule-triggered mechanism includes: calculating the vehicle's current position based on the allocated torque and comparing it with the reference trajectory to calculate the lateral error `e`. y (t). xref Predicted front wheel steering angle δ pred (t) and the current true front wheel steering angle δ real (t) Compare and calculate the angle prediction error e δ (t). The current heading angle is compared with the forward aiming time T of the reference trajectory. preview The heading angle is compared to calculate the abrupt change in heading angle ΔΨ(t). When e y (t), e δ When ΔΨ(t) and ΔΨ(t) are greater than the corresponding thresholds, a trigger command is output. The RL triggering mechanism includes: designing a multi-objective reward function and using it to train the network parameters of the RL agent. The trained RL agent infers and outputs the trigger command based on the vehicle state and the predicted state. The multi-objective reward function includes: the rule overreach penalty R when the rule triggering mechanism is triggered. rule , when e y (t) Training round termination penalty R when the corresponding threshold is exceeded fail .

[0009] Thirdly, the present invention provides a computer-readable storage medium storing a computer program / instructions. When the computer program / instructions are executed by a processor, they implement the steps of the event-triggered trajectory tracking method for a distributed driven vehicle in the first aspect.

[0010] The beneficial effects of this invention are as follows: This invention employs a trajectory tracking strategy based on differential steering and an adaptive dual-event triggering strategy. When the MPC (Master Control Controller) is updated, it not only issues the current control command sequence but also saves the entire predicted state sequence for subsequent cycle angle mismatch assessment. When the MPC is in a dormant state, it directly extracts the corresponding control commands from the cached control command sequence based on the time step and issues them to the four independent wheels. Above the physical hard constraint rule triggering mechanism, an RL (Research and Development) agent is introduced as a high-level decision-making center to autonomously explore cost-effective dormant opportunities. This data-driven global optimization RL triggering mechanism is fused with the multi-dimensional physical rule extreme condition rule triggering mechanism through a logical OR gate. This significantly reduces the triggering frequency of the MPC controller while ensuring vehicle tracking performance, thereby reducing computational power consumption. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1This is a flowchart of an event-triggered trajectory tracking method for distributed-drive vehicles; Figure 2 This is a schematic diagram of the differential steering principle; Figure 3 This is a graph showing the results of comparing lateral errors in the simulation; Figure 4 This is a comparison of the centroid sideslip angles in the simulation. Figure 5 This is a comparison chart of the cumulative number of MPC triggers in the simulation. Detailed Implementation

[0013] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0015] To address the problems of traditional MPC in trajectory tracking control of distributed drive vehicles, this invention proposes an event-triggered trajectory tracking method for distributed drive vehicles. This method integrates a comprehensive architecture of an "adaptive dual event trigger" and a "trajectory tracking controller." Please refer to... Figure 1 First, the system calculates the vehicle's spatial tracking error and forward curvature abrupt changes in real time, and introduces the "corner prediction mismatch" index to detect chassis open-loop integral drift early. Then, the scheduling of computing resources is taken over by a dual event trigger system that integrates data-driven and physical rule-based approaches: the RL (Reinforcement Learning) agent autonomously explores cost-effective sleep opportunities to reduce computing power consumption under multi-objective reward-driven conditions. The rule-based triggering mechanism acts as a safety shield, forcibly triggering a fallback for extreme and dangerous conditions; the two mechanisms achieve a comprehensive decision through a logical "OR". When the MPC controller is triggered and awakened, control commands are generated by the trajectory tracking controller: the MPC controller combines vehicle differential steering dynamics to output a control command sequence u. Then, fully considering dynamic axle load transfer and tire adhesion limits, the control command sequence u is precisely and optimally mapped to the four independent wheels. This closed-loop architecture effectively achieves global dynamic optimization of algorithm computing power and tracking accuracy.

[0016] This embodiment proposes an event-triggered trajectory tracking method for distributed drive vehicles. It incorporates differential steering dynamics into the prediction time domain, employs an event-triggered architecture, and after the MPC controller completes its solution, extracts and caches the complete predicted state sequence x. ref and control command sequence u. When the MPC controller is triggered, the predicted state sequence x is... ref The control command sequence u is updated, and not only is the current control command sequence u sent to the torque distribution layer, but the entire predicted state sequence x is also updated. ref This information is saved for corner mismatch assessment in subsequent cycles. When the MPC controller is in a dormant state without being triggered, the corresponding control quantities are directly extracted from the cached control command sequence u (i.e., the last updated control command sequence u) according to the time step and issued accordingly. Then, while adhering to the tire dynamic adhesion ellipse constraint, and with minimizing the overall load rate as the basic objective, the control command sequence u is accurately and safely mapped to the four independent wheels.

[0017] Specifically, traditional trajectory tracking MPC typically uses the front wheel steering angle as the control input. However, in the event-triggered trajectory tracking method of this invention, the front wheel steering angle is generated by the differential torque of the left and right wheels of the front axle. To accurately describe this dynamic process in the prediction time domain, an 8-DOF extended vehicle prediction model incorporating steering system dynamics is established. Predicted state sequence x ref Set as the vehicle's state variable x: The control command sequence u, as the vehicle's control variable, is: v x v y , r, Φ, Y, X, δ, dT actual These are, respectively, longitudinal vehicle speed, lateral vehicle speed, yaw rate, heading angle, abscissa and ordinate in the global coordinate system, front wheel steering angle, and actual front axle differential torque. F x,total dT cmd M z,add These represent the desired total longitudinal force, the desired front axle differential torque, and the additional yaw moment, respectively. The nonlinear continuous-time state equations of the vehicle prediction model. It consists of the following parts: (1) Vehicle planar dynamics and kinematics: ; In the formula, m is the total vehicle mass, and F yf F is the lateral force on the front wheel. yr F is the lateral force of the rear wheel, δ is the front wheel steering angle, and F is the lateral force of the rear wheel. drag For air resistance, I z Let L be the moment of inertia of yaw rotation. rLet be the distance from the rear axle to the center of mass. The kinematic equations follow conventional coordinate transformations and are approximated using a linear tire model: This is to ensure real-time solution efficiency. Where F y{f,r} C represents the lateral force of the front and rear wheels. {f,r} For the lateral stiffness of the front and rear wheels, α {f,r} The front and rear wheel slip angles.

[0018] (2) Differential steering system dynamics. The basic principle of differential steering is as follows: Figure 2 As shown, due to the kingpin offset distance r σ The presence of the kingpin allows the longitudinal driving force of the left and right front wheels to generate torque around the kingpin. By controlling the driving torque of the left and right wheels, a differential torque can be generated, thereby overcoming the return torque and producing front wheel steering. Based on this, a dynamic model of the steering actuator can be established as follows: ; In the formula, J eff b eff These represent the equivalent moment of inertia and equivalent damping of the steering actuator, respectively. τ a The frictional resistance torque τ of the steering actuator. f M' is the sum of the restoring torques of the left and right front wheels around their respective kingpins. f The differential torque can be expressed as: ; In the formula, τ dsl τ dsr These are the torques exerted by the left and right wheels around the kingpin, respectively. F xfr F xfl These represent the longitudinal forces on the left and right wheels of the front axle, respectively. γ is the caster angle. σ is the kingpin inclination angle. The second derivative of the front wheel steering angle and the frictional resistance torque of the steering actuator are generally small and can be considered as bounded disturbances that can be neglected. Therefore, the differential steering dynamics can be simplified as follows: .

[0019] After obtaining the vehicle prediction model, in order to achieve online real-time optimization of the MPC controller, the nonlinear continuous-time state equation of the vehicle prediction model is converted using the fourth-order Runge-Kutta method. Discretization. Within each control step, the trajectory tracking control is transformed into a standard nonlinear programming (NLP) problem, and solved efficiently using an interior-point method solver (IPOPT). Specifically, to achieve high-precision trajectory tracking while ensuring control smoothness, a cost function J containing tracking error, control variables, and control increments is designed. The output of the MPC controller is solved by minimizing the cost function J to obtain the predicted state sequence x. ref And the control instruction sequence u. The cost function J is expressed as: ; In the formula, N represents the prediction time domain, x(k|t) is the state variable at the k-th step at time t, and x... ref (k | t) is the reference state variable at time t and step k, u(k | t) is the control variable at time t and step k, Δu(k | t) is the increment of the control variable at time t and step k, Q, R, R d These correspond to the positive definite weight matrices. In each control cycle, the MPC controller searches for a set of future control sequences that minimize the cost function J, while satisfying various physical constraints (such as maximum motor torque and maximum tire slip angle). This future control sequence that minimizes the cost function J is the predicted state sequence x. ref And the control command sequence u. To ensure the vehicle's dynamic stability under extreme conditions and prevent sideslip or fishtailing, a hard constraint boundary based on the tire slip angle is introduced in the prediction time domain N: ; Based on physical limits, this embodiment sets a maximum sideslip angle threshold. α i (k | t) represents the front and rear wheel slip angles at time t and step k. f and r represent the front and rear wheels, respectively. Simultaneously, physical actuator boundary constraints are applied to the input control variables and the front wheel steering angle. Unlike traditional MPC which only extracts the first control input, this embodiment uses an event-triggered architecture. After solving the problem, the MPC controller must extract and cache the complete predicted state sequence x. ref And the control command sequence u. When the MPC controller triggers an update, it not only sends the current control command sequence u to the torque distribution layer, but also sends the entire predicted state sequence x. ref It is passed to the event triggering layer for corner mismatch assessment in subsequent cycles. When the MPC controller is in a non-triggered sleep state, the corresponding control quantity is directly extracted from the cached control instruction sequence u according to the time step and issued.

[0020] After obtaining the required control command sequence u, torque distribution is performed based on multi-objective optimization. Specifically, based on the tire friction ellipse theory, an increase in the force on the tire in one direction will inevitably lead to a decrease in the available adhesion in another direction. To strictly control the lateral tracking error within the ideal range, sufficient lateral force margin must be reserved for the tire to maintain the vehicle's lateral stability. Therefore, minimizing the tire's overall load factor (i.e., adhesion utilization) is chosen as the basic optimization objective for torque distribution. To reduce the computational dimensionality, considering that vehicle stability mainly depends on the lateral force margin, the lateral safety boundary is indirectly improved by minimizing the longitudinal tire force. The optimization objective function L can be defined as: ; In the formula, fl, fr, rl, and rr represent the front left, front right, rear left, and rear right wheels, respectively. i To adjust the penalty coefficient for each round of control weights. F xi F zi These represent the longitudinal driving force and vertical load of each wheel, respectively. μ i Let be the road adhesion coefficient at the corresponding wheel. To ensure that the combined longitudinal forces of the four wheels can accurately track the control command sequence u, the control input vector U is defined as follows: The control command sequence u is set as the control command vector V: .in, M steer For steering torque, R eff Let F be the effective rolling radius of the wheel. xfl F xfr F xrl F xrr These represent the longitudinal forces on the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively. The control input vector U and control command vector V satisfy the following relationship: BU = V. B represents the command mapping matrix, which can be determined by the vehicle's geometric kinematic parameters. ; In the formula, B f L is the track width of the vehicle's front wheels. f B is the distance from the front axle to the center of gravity. r The rear wheel track is represented by this matrix. This matrix reflects the mechanical coupling relationship between longitudinal force, yaw moment, and differential steering moment. Furthermore, control variables are inevitably subject to physical truncation limitations in practical engineering. The torque output by each hub motor cannot exceed the motor's maximum external characteristic peak value, nor can it exceed the frictional physical limits under current road conditions. Therefore, the following inequality constraints are established: ; Boundary conditions are defined as: maximum longitudinal driving force Among them, T max Let L be the maximum allowable output torque of the motor. Considering that in real-world driving scenarios, vehicles may face unsolvable situations where control variable requirements exceed road surface limits, hard constraints often lead to solver crashes. Therefore, the constraints are softened, transforming the original optimization objective into an L2-norm quadratic cost function L containing slack variables: ; In the formula, W u It is a diagonal weighted matrix, and its , used to normalize the longitudinal adhesion utilization weights of different wheels. λ is a scalar coefficient balancing the main objective of optimization and the equality constraints. W v It is a proportionality coefficient, and its Used to determine the priority of target control, by increasing the additional yaw moment and differential steering moment in W. v The proportion of W in the equation can be used to prioritize sacrificing some longitudinal speed following performance to ensure tracking accuracy when the dynamic margin is insufficient. Here, diag(·) represents the diag function. vFx,total W vMz,add W vMsteer They are F x,total M z,add M steer The coefficients. Since the cost function L is quadratic (with a squared term), and the constraints (maximum motor torque, road friction ring) are linear or convex, it constitutes a "convex optimization problem". In each control cycle (e.g., every 50 milliseconds), the cost function L is solved efficiently online in milliseconds using numerical optimization engines such as CaSADi, OSQP, or the internal effective set method to obtain the required control input vector U. Then, based on tire parameters and other parameters, it is converted into the torque command of the corresponding wheel motor: left front wheel torque T. fl Right front wheel torque T fr Left rear wheel torque T rl Right rear wheel torque T rr This not only ensures the global optimality of the control allocation results, but also meets the stringent requirements for high-frequency response of the vehicle under extreme operating conditions.

[0021] In this embodiment, the methods for triggering the MPC controller include: a rule-based triggering mechanism based on multidimensional physical rules and an RL-based triggering mechanism based on adaptive reinforcement learning. When the rule-based triggering mechanism or the RL-based triggering mechanism outputs a trigger command, the MPC controller is triggered to update. The following section uses the predicted state sequence x as an example. ref The control command sequence u provides a detailed explanation of the two triggering mechanisms.

[0022] I. Rule Triggering Mechanism.

[0023] Traditional fixed-time triggered MPCs typically require extremely high computation frequencies to ensure tracking accuracy under highly dynamic conditions, placing a significant burden on the computing power of the vehicle controller. Conventional event-triggered mechanisms often use the vehicle's pose tracking error as the sole trigger condition. However, the design of event-triggered mechanisms for differential steering systems must fully consider their unique underlying physical execution logic. MPCs usually assume that the issued front axle differential torque can be instantly and losslessly converted into a longitudinal force difference on the ground and generate steering torque, thus ignoring the dynamic delay in establishing wheel inertia and tire slip ratio. Under traditional high-frequency time-triggered closed-loop control, this nonlinear mismatch in the physical model is often masked by high-frequency state feedback. However, this mismatch is amplified during the "open-loop hold" phase of the event-triggered mechanism. Unlike traditional direct mechanical steering (angle control), differential steering is essentially a "force control" system. Position control can usually maintain state consistency under short-term open-loop hold. In differential steering, however, the rate of change of the front wheel steering angle is determined by both the differential torque issued by the actuator and the dynamic self-centering torque of the tires on the ground. During the open-loop phase of the MPC's pause calculation, the establishment of the actual vehicle's wheel acceleration and slip ratio requires dynamic response time. The resulting self-aligning torque and differential response will inevitably deviate slightly from the MPC's internal predictions. This slight torque-level deviation directly leads to errors in the front wheel steering angular velocity. The integral element within the control system causes this angular velocity error to accumulate over time, ultimately resulting in a significant deviation of the actual wheel steering angle from the MPC's expected angle. To address these issues, this embodiment provides a rule-triggered mechanism that includes state feedback, internal predictive monitoring, and look-ahead aiming, aiming to maintain control performance equivalent to high-frequency time-triggered mechanisms at a lower calculation frequency. Specifically, the rule-triggered mechanism includes the following three core conditions: (1) Trajectory tracking error boundary trigger.

[0024] To constrain the vehicle's spatial position deviation, the vehicle's current position is calculated based on the distributed torque, and then compared with the reference trajectory in the Frenet coordinate system to calculate the lateral error e. y (t): ; In the formula, x(t) is the ordinate of the vehicle at time t, x ref (t) represents the ordinate of the reference trajectory at time t, Ψ ref (t) represents the heading angle of the reference trajectory at time t, and y(t) represents the x-coordinate of the vehicle at time t. ref (t) represents the x-coordinate of the reference trajectory at time t. When the lateral error e y (t) Output a trigger command when the set safety envelope is exceeded: e y，th This is the tolerance threshold for lateral errors.

[0025] (2) Differential steering prediction mismatch trigger.

[0026] This is the core mechanism for suppressing open-loop integral drift in differential steering. At each step, the predicted state sequence x generated by the previous MPC controller solution is extracted. ref The predicted front wheel steering angle δ corresponding to the current moment pred (t), and compared with the actual front wheel steering angle δ fed back by the current physical chassis. real (t) Compare and calculate the angle prediction error e δ (t): ; δ th This is the corner prediction error threshold. This rule can trigger the MPC controller to correct the state in advance when the internal state of the chassis shows a divergent trend but has not yet had a significant impact on the macro trajectory.

[0027] (3) Forward curvature mutation triggering.

[0028] The core advantage of the MPC controller lies in its predictive capability. To prevent the vehicle from missing the optimal turning opportunity due to being in a dormant state before entering a curve, a forward-looking aiming mechanism is introduced. This involves comparing the current heading angle with the forward aiming time T of the reference trajectory. preview The heading angle is compared and the sudden change in heading angle ΔΨ is calculated to determine the road conditions ahead: ; In the formula, Ψ ref (t) is the heading angle of the reference trajectory at time t, T preview The timeout is set to 0.2–1 seconds. In this embodiment, after 0.5 seconds of pre-aiming, when the heading abruptly changes beyond the threshold ΔΨ... th At that time, the MPC controller is forcibly awakened to deal with the upcoming high curvature road section.

[0029] (4) Deadlock state triggered.

[0030] The number of steps t that the MPC controller has taken in the last calculation of the control command sequence u. idle Greater than or equal to the maximum keep-alive interval N max At this time, the rule triggering mechanism outputs a triggering command to prevent the MPC controller from entering a deadlock state if the control commands are exhausted.

[0031] In summary, the event Rule is triggered when any one of the above conditions is met. trigger That is, set it to 1 (output trigger command): .

[0032] II. RL Triggering Mechanism.

[0033] While rule-based triggering mechanisms can effectively suppress integral drift in differential steering systems and ensure driving safety, their fixed thresholds are difficult to adapt to optimal computational power allocation across all operating conditions. Setting the threshold too strictly leads to persistently high trigger frequencies, negating the purpose of event triggering. Setting the threshold too leniently sacrifices tracking accuracy. Therefore, this embodiment proposes a dual-trigger architecture that integrates data-driven and rule-based fallback mechanisms. This architecture prioritizes RL triggering mechanisms while using rule-based triggering mechanisms as a safety baseline, achieving a globally optimal adaptive balance between computational power consumption and tracking performance. Specifically, to transform the event-triggered scheduling problem into a task solvable by an RL agent, this embodiment constructs the following Markov decision process: 1) State Space: To enable the RL agent to fully perceive the vehicle's real-time state, internal actuator hysteresis, and future road conditions, a 7-dimensional continuous state space is defined. : ; Among them, e y For lateral error, e Ψ β is the heading angle error, used to assess the current space tracking error. β is the center-of-gravity sideslip angle, and r is the yaw rate, used to assess the chassis's dynamic stability. δ The angle prediction error is calculated from the predicted state sequence x. ref One dimension of data is extracted and used to perceive the open-loop mismatch state unique to differential steering. Steering angle prediction error e δ Together with other six-dimensional data, it defines the state space s t .

[0034] 2) Action Space: The action space is defined as a discrete binary action set A: A∈{0,1}. The output a... t =1 indicates that the MPC controller is awakened at the current moment to predict the state sequence x. ref Update the control instruction sequence u. Output a t =0 indicates that the system remains in a dormant state, using the predicted state sequence x from the previous cycle. ref and control instruction sequence u.

[0035] After the state space is constructed, a multi-objective function is designed based on the training samples in the state space s. tThe network parameters of the RL agent are trained by maximizing the cumulative reward. This RL agent can employ algorithms such as Q-learning, Deep Q-Network (DQN), Policy Gradient, and PPO. The trained RL agent infers and outputs trigger commands based on the vehicle state and predicted state. The design of this multi-objective function is crucial in guiding the RL agent to balance "control performance" and "computational cost," influencing the RL agent's preference for action outputs. Specifically, this embodiment designs a multi-objective reward function that includes dead zone tolerance, stability constraints, and trigger rule penalties: R t =R track +R stab +R comp +R rule +R fail ; Among them, R t Represents the reward value. R track R stab R comp R rule R fail These represent the penalties for tracking error, dynamic stability, computational cost, rule overreach, and training round termination, respectively. The meaning of each sub-item is explained in detail below: 1) Tracking error penalty R with dead zone track .

[0036] To prevent the RL agent from becoming overly sensitive to minor perturbations, a non-perceptible dead zone mechanism is introduced. When the lateral error and heading angle error are at extremely small thresholds (such as on a straight road segment), no penalty is imposed, encouraging the agent to remain dormant within this range. ; In the formula, w1 and w2 represent the lateral error penalty coefficient and the heading angle error penalty coefficient, respectively. y,db e Ψ,db These represent the dead zone thresholds for lateral error and heading angle error, respectively, which are set to 0.05m and 0.01rad in this embodiment.

[0037] 2) Dynamic stability penalty R stab It is used to suppress dangerous slippage and punish excessive sideslip angle β and yaw rate r: ; In the formula, w3 and w4 represent the penalty coefficients for the center of mass sideslip angle and the yaw rate, respectively.

[0038] 3) Computing power consumption penalty R compThis is the driving force that causes the RL agent to reduce its computation frequency. Whenever the RL agent actively chooses to trigger computation (a t When =1), a constant single-step action penalty value is applied, i.e., the computing power consumption penalty R. comp Its size directly determines how "stingy" the RL agent is with computing power.

[0039] 4) Penalty for exceeding authority in rule rules (R) rule When an RL agent issues a hibernation command due to insufficient exploration (a t (If the value is 0), but the rule-triggered mechanism assesses that the current state is approaching danger and forcibly triggers it, it imposes a high penalty for overstepping authority on the RL agent. This can forcibly correct the dangerous "lazy" behavior of the RL agent.

[0040] 5) Training round termination penalty R fail When the lateral error e y When (t) exceeds a certain error range (e.g., 1m), a very large termination penalty is imposed.

[0041] Through this designed multi-objective reward function, when the RL agent is in a certain state space s, t Select trigger (a) t =1) or hibernation (a t After `=0`, multiple penalty terms in the multi-objective reward function will jointly provide an immediate reward. Actions with higher rewards have a higher probability of being selected in subsequent training. Finally, after training convergence, the RL agent will output the probability distribution of two actions based on its current 7-dimensional state (including lateral error, centroid sideslip angle, yaw rate, rotation prediction error, sleep steps, and look-ahead curvature change). During the execution phase, the action with the higher probability is typically chosen as the RL agent's output. Simultaneously, this output will be logically ORed with the underlying multi-dimensional physical rules (lateral error exceeding limits, rotation mismatch exceeding thresholds, curvature mutation, and maximum keep-alive interval) to obtain the final trigger command S. trigger : ; Through this fusion mechanism, reinforcement learning is responsible for finding the most cost-effective hibernation opportunities under most safe operating conditions, while physical rules act as a shield, forcibly taking over control when encountering extreme disturbances or insufficient generalization ability of RL. This dual architecture not only overcomes the performance limitations of fixed rules but also compensates for the fatal flaw of pure data-driven algorithms lacking hard safety boundaries. Under the constraint of the safety shield, the RL agent autonomously decides whether to wake up MPC, achieving a balance between computing power and tracking performance.

[0042] To verify the effectiveness of the event-triggered trajectory tracking method for distributed drive vehicles proposed in this invention, a comprehensive performance evaluation of the adaptive dual event-triggered mechanism was conducted. The experiment employed a dual lane-change scenario with a total length of approximately 200m and a maximum lateral offset of 3.5m. The cruising speed was set to 20m / s, the road surface adhesion coefficient to be 0.85, and the simulation sampling period to be 0.05s. The main parameters of the test vehicle are shown in Table 1.

[0043] Table 1: Main parameters of the speed measuring vehicle.

[0044]

[0045] The experiment included three control groups: 1. Fixed frequency triggering: The traditional time-triggered mechanism performs MPC solution at each step (consuming 100% of computing resources), serving as an ideal benchmark for tracking performance.

[0046] 2. Pure rule-based event triggering: a triggering strategy that relies solely on fixed threshold physical rules (lateral error, aiming curvature, and corner mismatch).

[0047] 3. Adaptive dual event triggering (i.e., this invention): a fusion triggering strategy dominated by reinforcement learning and backed by rules.

[0048] Simulation results show that all three strategies can safely complete the double lane change task and perform well. Figure 3 As stated above, the maximum lateral error was always strictly controlled within 0.25 m. Then, the sideslip angles of the vehicle's center of gravity were compared for the three strategies, as follows: Figure 4 As shown, the adaptive dual-trigger mechanism achieves performance comparable to the traditional time-triggered mechanism. Finally, the cumulative number of triggers is compared, such as... Figure 5As shown, within a 10-second simulation period: the traditional fixed-frequency full-load operation performed approximately 200 calculations (defined as a 100% computing power baseline). Pure rule-based event triggering reduced the number of calculations to 118 (triggering rate 58.1%). The dual-triggering mechanism (RL-based ET-MPC) proposed in this invention only triggered 80 times, with a triggering rate as low as 39.9%. It is noteworthy that the red line exhibits a stepped distribution characteristic: in the intervals of 0~2s, 6.5~7.5s, and 8.5~10s, where the curvature is relatively gentle or the vehicle is in a steady state, the red line shows an extremely long horizontal segment, indicating that the RL agent fully utilizes the "unnoticeable dead zone" to achieve long-term deep sleep. In the interval of 2~4s, where the curvature changes abruptly, the triggering density is adaptively increased. This fully demonstrates that the architecture integrating data-driven and physical rules successfully achieves optimal computing power scheduling. The simulation results show that, under the dual lane change condition, the event-triggered trajectory tracking method for distributed drive vehicles proposed in this invention not only ensures high tracking accuracy and vehicle stability, but also significantly reduces the average computational load of the on-board controller to 39.9% of that of the traditional time-triggered mechanism.

[0049] In another embodiment, an event-triggered trajectory tracking system for a distributed drive vehicle is also proposed, which uses the event-triggered trajectory tracking method for a distributed drive vehicle described in the above embodiments. The event-triggered trajectory tracking system includes: a trajectory tracking controller and an adaptive dual event trigger. The trajectory tracking controller uses an MPC controller based on vehicle differential steering dynamics to calculate the vehicle's predicted state sequence x according to the vehicle's reference trajectory. ref and control command sequence u. When the MPC controller triggers, it controls x. ref The system updates `u` and maps it to the torque distribution of the four wheels to achieve reference trajectory tracking. When the MPC controller is in sleep mode, it uses the previously updated `u` to achieve reference trajectory tracking. An adaptive dual-event trigger is used to activate the MPC controller when a trigger command is output by either the rule-triggered mechanism or the RL-triggered mechanism. The rule-triggered mechanism includes: calculating the vehicle's current position based on the allocated torque and comparing it with the reference trajectory to calculate the lateral error `e`. y (t). x ref Predicted front wheel steering angle δ pred (t) and the current true front wheel steering angle δ real (t) Compare and calculate the angle prediction error e δ (t). The current heading angle is compared with the forward aiming time T of the reference trajectory. preview The heading angle is compared to calculate the abrupt change in heading angle ΔΨ(t). When e y (t), e δWhen ΔΨ(t) and ΔΨ(t) are greater than the corresponding thresholds, a trigger command is output. The RL triggering mechanism includes: designing a multi-objective reward function and using it to train the network parameters of the RL agent. The trained RL agent infers and outputs the trigger command based on the vehicle state and the predicted state. The multi-objective reward function includes: tracking error penalty R... track Dynamic stability penalty R stab Computing power consumption penalty R comp Rule triggering mechanism triggers penalty for exceeding authority (R) rule , when e y (t) Training round termination penalty R when the corresponding threshold is exceeded fail .

[0050] In another embodiment, a computer-readable storage medium is also provided, which stores a computer program / instructions. When the computer program / instructions are executed by a processor, the steps of the event-triggered trajectory tracking method for a distributed driven vehicle in the above embodiments are implemented. The computer-readable storage medium may include, but is not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0051] In another embodiment, a computer program product is also provided, comprising a computer program / instructions. The computer program / instructions are used to cause a computer to perform the steps of the event-triggered trajectory tracking method for a distributed driven vehicle as described in the above embodiments. The computer program / instructions may exist in a computer-readable medium in forms including, but not limited to, source files, executable files, and installation package files. Accordingly, the computer program / instructions may be executed by a computer in ways including, but not limited to: the computer directly executing the instructions; the computer compiling the instructions and then executing the corresponding compiled program; the computer reading and executing the instructions; or the computer reading and installing the instructions and then executing the corresponding installed program.

[0052] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0053] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for event-triggered trajectory tracking of a distributed driven vehicle, characterized in that, The MPC controller is used to calculate the vehicle's predicted state sequence x based on the vehicle's differential steering dynamics and the vehicle's reference trajectory. ref and control command sequence u; when the MPC controller is triggered, x ref The system updates u and maps u to the torque distribution of the four wheels to achieve reference trajectory tracking; when the MPC controller is in sleep mode, it uses the previously updated u to achieve reference trajectory tracking. Methods for triggering the MPC controller include: triggering the MPC controller to update when the rule triggering mechanism or RL triggering mechanism outputs a triggering command; The rule triggering mechanism includes: calculating the lateral error e by comparing the vehicle position with the reference trajectory. y (t); x ref Predicting the front wheel steering angle δ pred (t) and the actual front wheel steering angle δ real (t) Compare and calculate the angle prediction error e δ (t); Current heading angle and forward aiming time T of the reference trajectory preview The heading angle comparison is used to calculate the heading angle abrupt change ΔΨ; when e y (t), e δ When (t) or ΔΨ is greater than the corresponding threshold, output a trigger command; The RL triggering mechanism includes: designing a multi-objective reward function to train the network parameters of the RL agent; the trained RL agent is based on the vehicle state and x. ref The instruction to trigger the inference output; the multi-objective reward function includes: tracking error penalty R. track Dynamic stability penalty R stab Computing power consumption penalty R comp Rule triggering mechanism triggers penalty for exceeding authority (R) rule , when e y (t) Training round termination penalty R when the corresponding threshold is exceeded fail .

2. The event-triggered trajectory tracking method for distributed drive vehicles according to claim 1, characterized in that, Predicted state sequence x ref Set as the vehicle's state variable x: The control command sequence u serves as the vehicle's control variable. ; Among them, v x v y , r, Φ, Y, X, δ, dT actual These are, respectively, longitudinal vehicle speed, lateral vehicle speed, yaw rate, heading angle, abscissa and ordinate in the global coordinate system, front wheel steering angle, and actual front axle differential torque; F x,total dT cmd M z,add These are the expected total longitudinal force, the expected front axle differential torque, and the additional yaw moment, respectively. Design a cost function J that includes tracking error, control variables, and control increments, and solve for the output of the MPC controller to obtain the predicted state sequence x by minimizing the cost function J. ref The control instruction sequence u; the cost function J is expressed as: ； In the formula, x(k|t) is the state variable at time t, and x is the state variable at step k. ref (k | t) represents the reference state variable at step k at time t, N represents the prediction time domain, u(k | t) represents the control variable at step k at time t, and Δu(k | t) represents the increment of the control variable at step k at time t; Q, R, R d The corresponding positive definite weight matrices.

3. The event-triggered trajectory tracking method for distributed drive vehicles according to claim 1, characterized in that, Methods for mapping u to torque distribution across four wheels include: Define the control input vector U: The control command sequence u is set as the control command vector V: Among them, F xfl F xfr F xrl F xrr These represent the longitudinal forces on the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively; F x,total M z,add M steer These are the desired total longitudinal force, the additional yaw moment, and the steering moment, respectively; the control input vector U and the control command vector V satisfy: BU=V; B represents the command mapping matrix; After solving the L2 norm quadratic cost function L containing slack variables, the longitudinal forces of each of the four wheels are obtained, and then converted into the corresponding left front wheel torque T based on the tire parameters. fl Right front wheel torque T fr Left rear wheel torque T rl Right rear wheel torque T rr The expression for the cost function L is: ； In the formula, W u W is a diagonal weighted matrix, where λ is the scalar coefficient and W is the weighted coefficient. v This is the proportionality coefficient.

4. The event-triggered trajectory tracking method for distributed drive vehicles according to claim 1, characterized in that, To prevent vehicles from missing the optimal turning opportunity due to being in a dormant state before entering a curve, a change in heading angle ΔΨ is introduced, and its calculation formula is as follows: ； In the formula, Ψ ref (t) represents the heading angle of the reference trajectory at time t; T preview Set to 0.2 to 1 second; when ΔΨ is greater than the threshold, the MPC controller is forcibly triggered.

5. The event-triggered trajectory tracking method for distributed drive vehicles according to claim 1, characterized in that, The rule-triggered mechanism also includes: when the number of steps that the MPC controller has taken in the last calculated control instruction sequence u is greater than the maximum keep-alive interval, the rule-triggered mechanism outputs a trigger instruction.

6. The event-triggered trajectory tracking method for distributed drive vehicles according to claim 1, characterized in that, Define the state space s t for: ； Define the action space as a discrete binary action set A: A∈{0,1}; Among them, e y For lateral error, e Ψ Here, β is the heading angle error, β is the sideslip angle of the center of mass, r is the yaw rate, and e is the yaw rate. δ For the angle prediction error, t idle ΔΨ is the number of steps taken in the last calculated control command sequence u by the MPC controller, and ΔΨ is the change in heading angle. Based on the training samples, the multi-objective function in the state space s t The RL agent is updated by maximizing cumulative rewards to influence its preference for action outputs.

7. The event-triggered trajectory tracking method for distributed drive vehicles according to claim 1, characterized in that, The reward value R of a multi-objective reward function t The sum of its subterms: R t =R track +R stab +R comp +R rule +R fail ; In the formula, the tracking error penalty R track Used for lateral error e y and heading angle error e Ψ Punishment will be imposed; Dynamic stability penalty R stab Used to penalize excessive centroid sideslip angle β and yaw rate r; Computing power consumption penalty R comp This is used to encourage RL agents to reduce their computation frequency.

8. The event-triggered trajectory tracking method for distributed drive vehicles according to claim 1, characterized in that, The expression for vehicle differential steering dynamics is: ； In the formula, δ is the front wheel steering angle, and b eff For the equivalent damping of the steering actuator, M' f For differential torque, τ f It is the sum of the restoring torques of the left and right front wheels around their respective kingpins.

9. An event-triggered trajectory tracking system for a distributed drive vehicle, characterized in that, It uses the event-triggered trajectory tracking method for distributed drive vehicles as described in any one of claims 1 to 8; The event-triggered trajectory tracking system includes: The trajectory tracking controller uses the MPC controller based on vehicle differential steering dynamics to calculate the vehicle's predicted state sequence x according to the vehicle's reference trajectory. ref and control command sequence u; when the MPC controller is triggered, x ref The system updates u and maps u to the torque distribution of the four wheels to achieve reference trajectory tracking; when the MPC controller is in sleep mode, it uses the previously updated u to achieve reference trajectory tracking. An adaptive dual-event trigger is used to trigger the MPC controller to update when a trigger command is output by either the rule-triggered mechanism or the RL-triggered mechanism. The rule-triggered mechanism includes: calculating the vehicle's current position based on the allocated torque and comparing it with the reference trajectory to calculate the lateral error e. y (t); x ref Predicted front wheel steering angle δ pred (t) and the current true front wheel steering angle δ real (t) Compare and calculate the angle prediction error e δ (t); The current heading angle is compared with the forward aiming time T of the reference trajectory. preview The heading angle is compared to calculate the abrupt change in heading angle ΔΨ(t); when e y (t), e δ When ΔΨ(t) and ΔΨ(t) are greater than the corresponding thresholds, a trigger command is output; the RL triggering mechanism includes: designing a multi-objective reward function and using it to train the network parameters of the RL agent; the trained RL agent infers and outputs the trigger command based on the vehicle state and the predicted state; the multi-objective reward function includes: tracking error penalty R track Dynamic stability penalty R stab Computing power consumption penalty R comp Rule triggering mechanism triggers penalty for exceeding authority (R) rule , when e y (t) Training round termination penalty R when the corresponding threshold is exceeded fail .

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the event-triggered trajectory tracking method for a distributed drive vehicle as described in any one of claims 1 to 8.

Images