Motion control method for realizing modularized bus stop parking

Abstract

The invention discloses a motion control method for realizing the parking of a modularized bus station, which is characterized by comprising the following steps of: constructing a motion control model and determining a target function; acquiring module bus data and environmental sensing data, and inputting the data into the trained first neural network model to obtain a weight coefficient; weighting the objective function through the weight coefficient, and solving under the motion control model to obtain an optimal control quantity; according to the invention, the first neural network model is utilized to obtain the weight coefficients of multiple targets based on the input features such as the bus state, and the problem of multi-target conflict is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving vehicle planning and control, and more particularly to a motion control method for realizing the docking of a modular bus at a station. Background Art

[0002] Currently, traditional public transportation systems use fixed-capacity buses. This fixed capacity can lead to higher passenger wait times and higher operating costs, as passenger demand and bus supply may not always match. Emerging modular buses offer greater flexibility by varying bus capacity by adjusting the number of modules in the fleet, thus offering a potential solution to this problem.

[0003] Current research in the field of modular buses is mainly focused on scheduling, operation, and timetable optimization, and has made significant progress. In order to further promote the implementation of modular buses, it is also crucial to study related control issues. However, current research related to modular bus control is relatively scarce. According to real-world scenarios, bus control can be divided into on-road and on-platform control. Current control-related research is all focused on the docking or separation of modular buses on-road, and there is a lack of control research related to platform scenarios. How modular buses can smoothly and energy-efficiently drive to the platform and accurately stop at the designated window is related to both the riding experience and the operator's image, so it is very important. At present, there is no research on modular bus motion control for this process.

[0004] Parallel parking scenarios for private cars are similar to those for modular buses, offering the potential for leveraging these approaches. In this field, NMPC (Nonlinear Model Predictive Control) has been highly regarded for its ease of handling high-precision problems and constructing constrained multi-objective optimization models. Therefore, NMPC has become a potential solution for modular bus parking control. The vehicle dynamics model and the weights of the objective function are key factors influencing NMPC performance. In the parallel parking field, a common approach is to manually derive a dynamics model based on physical laws. Weights are then manually tuned offline through trial and error, with fixed weights used online. The current research validates the feasibility of these strategies for achieving high-performance parallel parking. However, due to the numerous differences between modular buses and private cars, these strategies struggle to meet the performance requirements of parking scenarios.

[0005] On the one hand, modular buses can flexibly dock and detach, so their dynamic systems may undergo significant changes. This flexibility requires a more adaptable vehicle dynamics model. Enumerating all conditions and manually deriving a physical model is extremely labor-intensive, and the necessary physical parameters may not be available. On the other hand, modular buses have high requirements for smoothness and low energy consumption, and they must dock precisely within designated windows. However, the objective functions corresponding to these requirements may conflict, making it difficult to achieve optimal overall performance.

[0006] Therefore, how to cope with the variability of modular buses, balance the conflicts among multiple objectives, and realize safe, efficient and comfortable entry and stop control of modular buses in platform scenarios is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0007] In view of this, the present invention provides a motion control method for realizing modular bus stop parking, which uses a first neural network model to obtain weight coefficients of multiple targets based on input features such as bus status to solve the problem of multi-target conflict.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A motion control method for achieving modular bus stop parking includes the following steps:

[0010] Build a motion control model and define the objective function.

[0011] Obtain module bus data and environmental perception data, and input them into the trained first neural network model to obtain weight coefficients.

[0012] The objective function is weighted by the weight coefficient, and the optimal control quantity is obtained by solving it under the motion control model.

[0013] Preferably, the first neural network model fuses the input features based on the encoder module in the Transformer, then passes through a multi-layer perceptron with a residual structure, and finally passes through the output layer to obtain the weight value.

[0014] Preferably, the input features include the quadratic error between the current posture of the bus and the desired posture, the current speed, the current control amount, the mass and the moment of inertia of the bus.

[0015] Preferably, the first neural network model is trained by reinforcement learning, and the reward function used during training is:

[0016] R=(J energy +J smooth +J precise )

[0017] Among them, R is the reward function, J energy is the actual energy consumption of public transportation, J smooth is the total acceleration, J precise Represents the deviation between the final docked position and the expected position.

[0018] Preferably, the step of constructing a motion control model and clarifying an objective function specifically includes:

[0019] Construct the objective function:

[0020]

[0021] Where J is the objective function; ζ r =ζ-ζ d , is the longitudinal and transverse position deviation of the bus, ζ=[s,e,θ] T is the actual position, and its components are the longitudinal position, lateral position and heading angle in the Frenet coordinate system, ζ d is the corresponding expected value; u=[f t ,δ] T is the control quantity for connecting to the bus, and its components are tire driving force and steering angle, Q p ,Q a ,Q j ,Q t is the weight coefficient matrix;

[0022] Construct vehicle dynamics models;

[0023] Constructing constraints, wherein the constraints include dynamic constraints based on the vehicle dynamics model:

[0024]

[0025] Where x=[v x ,v y ,ω] T Represents the state quantity, Represents the inverse of the state vector, v x ,v y ,ω represent the longitudinal and lateral velocities and angular velocity respectively, and F is the vehicle dynamics model.

[0026] Preferably, the constraint conditions also include lane boundary restriction constraints based on the shape of the modular bus and physical constraints required for the state and control quantities of the docking bus.

[0027] Preferably, the lane boundary restriction constraints based on the shape of the modular bus include:

[0028] e+l f sin(θ)+w lcos(θ)≤G l (s)

[0029] el r sin(θ)+w l cos(θ)≤G l (s)

[0030] e+l f sin(θ)-w r cos(θ)≥G r (s)

[0031] el r sin(θ)-w r cos(θ)≥G r (s)

[0032] Among them, w l and w r Respectively represent the distance from the center of mass to the left and right ends of the bus, l f ,l r Respectively represent the distance from the center of mass to the front and rear ends of the bus, G l ,G r Represent the functions of the upper and lower lane boundaries respectively;

[0033] The physical constraints required for the state and control quantities of the bus docking include:

[0034] u min ≤u≤u max

[0035] Among them, u min ,u max The physical limit of the controlled quantity.

[0036] Preferably, the constructing of the vehicle dynamics model specifically includes:

[0037] Get the status of each axle and use the obtained status sequence as input information.

[0038] The second neural network model is used to extract the feature vector and output the resultant force and resultant moment.

[0039] The resultant acceleration is calculated based on the resultant force and moment, and the longitudinal and transverse coupling terms are added to obtain the model output.

[0040] Preferably, the constructing of the vehicle dynamics model further includes training parameters in the second neural network model based on a labeled data set.

[0041] Through the above technical solutions, it can be seen that compared with the prior art, the present invention discloses a motion control method for realizing modular bus stop docking. By constructing a motion control model and designing a specific objective function, a first neural network model is used to output weight coefficients based on input features such as bus status through a Transformer and a multi-layer perceptron fusion, and the model is trained by reinforcement learning. The weight coefficients are weighted to the objective function to solve the optimal control quantity and solve the multi-objective conflict problem; at the same time, a second neural network model is used to extract the features and output the resultant force and torque according to the axle state sequence, and a vehicle dynamics model is constructed in combination with the resultant acceleration and coupling terms. The parameters are trained based on the data set to realize adaptive modeling of the modular bus's variable dynamics system, thereby realizing safe, efficient and comfortable entry and docking control of the modular bus in the platform scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0043] Figure 1 This is a schematic diagram of the framework of a motion control method for realizing modular bus stop parking proposed by the present invention.

[0044] Figure 2 This is a schematic diagram of the module bus entering the station and stopping scene.

[0045] Figure 3 Diagram of different vehicle parameters used for experimental validation.

[0046] Figure 4 is the fitting error diagram of the kinetic model.

[0047] Figure 5 This is the high-speed extrapolation effect diagram of the dynamic model.

[0048] Figure 6 A graph of the overall performance results obtained for different objective function weight settings. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] like Figure 1 and Figure 2 The embodiment of the present invention discloses a motion control method for realizing modular bus stop parking, comprising the following steps:

[0051] S1: Build a motion control model and define the objective function.

[0052] S2: Obtain module bus data and environmental perception data, and input them into the trained first neural network model to obtain weight coefficients.

[0053] S3: Weight the objective function by the weight coefficient and solve it under the motion control model to obtain the optimal control quantity.

[0054] Among them, the motion control model is based on NMPC theory and is constructed offline.

[0055] As a feasible implementation method, the constructed motion control model includes the following specific features:

[0056] Objective function,The objective function is designed by comprehensively considering minimizing the docking error, minimizing the energy consumption and maximizing the comfort, specifically:

[0057]

[0058] Among them, J is the objective function, t f is the final stop time; r =ζ-ζ d ,ζ d Contains the corresponding expected value, ζ=[s,e,θ] T is the longitudinal and transverse position and heading angle of the bus, and its components are speed, longitudinal position in Frenet coordinate system, transverse position and heading angle; u=[f t ,δ] T is the control quantity for connecting to the bus, and its components are tire driving force and steering angle; Q p ,Q a ,Q j ,Q t is the weight coefficient matrix, which is used to represent the weight coefficients of each state variable or control quantity for docking error, energy consumption, comfort and final state.

[0059] in, is the error term, representing the minimum docking error, u T Q a u represents energy consumption, To reflect the impact degree, to reflect the comfort.

[0060] Constraints:

[0061] (1) The bus state and control variables of the module must satisfy the dynamic constraints:

[0062]

[0063] Where x=[v x ,v y ,ω] T Indicates the state quantity. x ,v y ,ω represent the longitudinal and lateral velocities and angular velocity respectively. F defines the vehicle dynamics model.

[0064] Select a road model based on a Cartesian coordinate system to describe the relationship between speed and position:

[0065]

[0066] in, and are the rates of change of longitudinal position, lateral position and heading angle respectively; θ is the heading angle; v x and v y are the components of the velocity on the x-axis and y-axis respectively.

[0067] (2) The shape of the modular bus cannot violate lane boundary restrictions:

[0068] e+l f sin(θ)+w l cos(θ)≤G l (s)

[0069] el r sin(θ)+w l cos(θ)≤G l (s)

[0070] e+l f sin(θ)-w r cos(θ)≥G r (s)

[0071] el r sin(θ)-w r cos(θ)≥G r (s)

[0072] Among them, l f ,l r Represents the distance from the center of mass to the front and rear ends of the bus. l ,G r are functions representing the upper and lower lane boundaries respectively.

[0073] (3) The state and control quantities of the bus connection must meet physical constraints:

[0074] u min ≤u≤u max

[0075] Among them, u min ,u max The physical limit of the controlled quantity.

[0076] (4) The initial state of the model must be equal to the current state of the bus: x0 = x(0).

[0077] In order to further implement the above technical solution, a vehicle dynamics model is constructed based on PINN. The specific method is as follows:

[0078] Get the status of each axle and use the obtained state sequence as input information; for each axle, take the following input information:

[0079] x in =[f t ,f y ,l,δ,f t δ] T

[0080] Among them, f t ,f y are the longitudinal force and lateral force of the tire on the axle; δ is the steering angle of the tire; l is the distance from the axle to the center of mass; f y ≈c y α, c y ,α are the tire's cornering stiffness and slip angle respectively. For the front axle, α≈δ-(v y +l f ω) / v x For the rear axle, α≈δ-(v y -l r ω) / v x .

[0081] A second neural network model (such as a recurrent neural network model) is used to extract feature vectors and output the resultant force and torque;

[0082]

[0083] in, is the feature vector containing force and torque information (can be five-dimensional). Gelu(*) is the activation function. hi For hidden features (forces and moments calculated from the previous axis). in ,w hi ,b is a learnable parameter.

[0084]

[0085] Among them, f ais the resultant force and moment, w a ,b a is a learnable parameter, w a The dimension can be selected as 3×3.

[0086] The resultant acceleration is calculated based on the resultant force and moment, and the longitudinal and transverse coupling terms are added to obtain the model output.

[0087]

[0088] y=a+[v y ω,-v x ω,0] T

[0089] Among them, y is the final output of the neural dynamics model, that is, the second neural network model, a is the combined acceleration, represents element-wise division; w1, w2, w3 are learnable parameters.

[0090] Furthermore, the parameters in the second neural network model are trained based on the labeled data set. During the training process, the training set is subjected to the above-mentioned forward propagation process to obtain a predicted value, and the model loss is calculated according to the predicted value and the parameters are updated.

[0091] As a feasible implementation method, the present invention realizes dynamic adjustment of multiple target weights in the objective function through a first neural network model, and trains the network offline based on reinforcement learning.

[0092] Specifically, the input features of the network include: the quadratic error between the current position of the bus and the desired position, the current speed, the current control amount, the mass and moment of inertia of the bus. The output of the network is Q p ,Q a ,Q j The values ​​in represent the weight coefficients of each state variable or control quantity for docking error, energy consumption, and comfort.

[0093] The network structure of the first neural network model: First, a linear layer is used to increase the dimensionality of the input features to 14. The input features are fused using the encoder module in the Transformer. The encoder module is set to two layers and two heads, with a model dimension of 14. The input features are then passed through a multilayer perceptron with a residual structure, with a dimension of 14×14×14. Finally, the output layer has a dimension of 14×7.

[0094] The reward function required to train the network based on reinforcement learning is set up as follows:

[0095] R=(J energy +J smooth +J precise )

[0096] Where R is the reward function. energy is the actual energy consumption of public transportation. smooth is the total acceleration. J precise Represents the deviation between the final docking position and the expected position, which can be defined as follows:

[0097]

[0098] Among them, s d ,e d are the desired vertical and horizontal positions respectively. m is the threshold of the docking error.

[0099] Based on the first network model and the reward function, a deep reinforcement learning algorithm, such as the SAC algorithm, is used to train the network. The Critic's network architecture can be the same as the Actor's. The default loss function is used for both.

[0100] In this embodiment, at the beginning of each generation of training, the current state information of the modular bus is obtained, including: the quadratic error between the current position of the bus and the desired position, the current speed, the current control amount, the mass and moment of inertia of the bus, and this data is organized into a format suitable for input into the first neural network model. Then, the model fuses the input features based on the encoder module of the Transformer to capture the correlation between the modular bus data and the environmental perception data. The fused features are further processed by a multi-layer perceptron with a residual structure, and finally the weight values ​​are obtained through the output layer. These weight coefficients are used to subsequently weight the objective function. Because the objective function is constantly changing, the newly generated weight coefficients will affect the importance of each sub-objective in the objective function (such as docking accuracy, comfort, efficiency, etc.). For example, if the current road is congested, the weight of avoiding sudden braking (to ensure comfort) in the objective function may be increased; if the bus stop is approaching, the weight of docking accuracy in the objective function may be increased. Under the updated objective function and motion control model, an appropriate optimization algorithm (such as gradient descent or Newton's method) is used to solve for the optimal control variable. The resulting optimal control variable is then applied to the bus, causing it to perform corresponding actions such as acceleration, deceleration, and steering. This optimal control variable is designed to ensure that the bus operates optimally under the current objective function and bus state, such as arriving at a bus stop at an appropriate speed and acceleration.

[0101] After the bus performs an action, it interacts with the environment. The environment updates the bus's state (such as location and speed changes) based on the bus's action and generates corresponding rewards (or penalties). For example, if the bus stops accurately at the bus stop, it will receive a higher reward; if the bus collides or deviates from the route, it will receive a penalty.

[0102] Finally, data such as the new state of the bus after executing the action, the reward obtained, the control amount executed, etc. are collected and used to train the first neural network model. The parameters of the model are adjusted through reinforcement learning algorithms (such as deep Q network, policy gradient algorithm, etc.) so that the weight coefficients generated by the model in the future can better adapt to the changes in the objective function.

[0103] Below, the present invention further verifies the above method with specific examples.

[0104] Based on Carla, a docking scene is created. Based on Carsim / Trucksim, three sizes of vehicles are provided: small, medium and large. For specific parameters, see Figure 3 . Write the algorithm in python. The initial condition of the reinforcement learning scenario is set to v x ∈[4,6]m / s,v y ∈[-0.05,0.05]m / s,ω∈[-0.05,0.05]rad / s,s∈[-40,-20]m,e∈[4,6]m,θ∈[-1,1]°. The desired docking position is [0,1,0]. The vehicle is considered docked if the distance from the desired position is less than 0.05m, the heading angle deviation is less than 1°, and the speed is less than 0.01m / s. The entire training process consists of 600 rounds.

[0105] The original NMPC model serves as the baseline, while the vehicle dynamics model uses a common physics-based single-scale model. This dynamics model is then linearized to form the LMPC control model, which serves as another baseline. The method using a neural dynamics model and adaptive weights is called enhanced model predictive control (RMPC). The overall performance statistics of the three methods on different vehicle models are shown in the table below.

[0106] Table 1 Overall performance statistics of different methods

[0107]

[0108] As can be seen from the table above, the proposed RMPC outperforms the baseline methods in all performance indicators, especially the docking error. Experiments show that LMPC and NMPC have large errors in their dynamic models, resulting in large deviations from the true value model and often leading to unsolvable problems.

[0109] Under the scenario settings corresponding to Table 1, the fitting error of the dynamic model is as follows Figure 4 As shown in the figure, it can be seen that the neural dynamics model has higher accuracy. Let the initial lateral velocity and angular velocity be 0, and apply sinusoidal signals of driving force and steering angle with amplitudes of f respectively. t =1KN,δ f= 0.1rad, period is 2 seconds. Gradually increase the longitudinal velocity, the extrapolated result is as follows Figure 5 We found that although the neural dynamics model was trained only on data with speeds less than 6 m / s, it also showed strong extrapolation to high-speed conditions.

[0110] The Bayesian optimization, expert-tuned objective function weights, and randomly generated objective function weights are used as benchmarks, and the reward value calculated by formula (31) is used to evaluate the overall performance. In the randomly generated scenario, the proposed neural network-based parameter adjustment method is compared with the results of each benchmark. Figure 6 As shown in Figure 2, it can be seen that the overall performance of the proposed method is significantly better than that of other methods.

[0111] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0112] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A motion control method for realizing modular bus stop parking, characterized in that: The following steps are included: constructing a motion control model and clarifying the objective function; Obtaining module bus data and environmental perception data, and inputting them into the trained first neural network model to obtain weight coefficients; The objective function is weighted by the weight coefficient, and the optimal control quantity is obtained by solving it under the motion control model.

2. The motion control method for realizing modular bus stop parking according to claim 1 is characterized in that: The first neural network model fuses the input features based on the encoder module in the Transformer, then passes through a multi-layer perceptron with a residual structure, and finally passes through the output layer to obtain the weight value.

3. The motion control method for realizing modular bus stop parking according to claim 2 is characterized in that: The input features include the quadratic error between the current posture of the bus and the desired posture, the current speed, the current control amount, the mass and the moment of inertia of the bus.

4. A motion control method for realizing modular bus stop parking according to claim 1, 2 or 3, characterized in that: The first neural network model is trained by reinforcement learning, and the reward function used in training is: R\(J energy +J smooth +J precise ) Among them, R is the reward function, J energy is the actual energy consumption of public transportation, J smooth is the total acceleration, J precise Represents the deviation between the final docking position and the expected position.

5. The motion control method for realizing modular bus stop parking according to claim 1 is characterized in that: The construction of the motion control model and the clarification of the objective function specifically include: Construct the objective function: Where J is the objective function; ζ r =ζ-ζ d , is the longitudinal and transverse position deviation of the bus, ζ=[s,e,θ] T is the actual position, and its components are the longitudinal position, lateral position and heading angle in the Frenet coordinate system, ζ d is the corresponding expected value; u=[f t ,δ] T is the control quantity for connecting to the bus, and its components are tire driving force and steering angle, Q p ,Q a ,Q j ,Q t is the weight coefficient matrix; Construct vehicle dynamics models; Constructing constraints, wherein the constraints include dynamic constraints based on the vehicle dynamics model: Where x=[v x ,v y ,ω] T Represents the state quantity, represents the inverse of the state vector, v x ,v y ,ω represent the longitudinal and lateral velocities and angular velocity respectively, and F is the vehicle dynamics model.

6. The motion control method for realizing modular bus stop parking according to claim 5, characterized in that: The constraints also include lane boundary restriction constraints based on the shape of the modular bus and physical constraints required for the state and control quantities of the docking bus.

7. The motion control method for realizing modular bus stop parking according to claim 6, characterized in that: The lane boundary restriction constraints based on the shape of the modular bus include: e+l f sin(θ)+w l cos(θ)≤G l (s) en r sin(θ)+w l cos(θ)≤G l (s) e+l f sin(θ)-w r cos(θ)≥G r (s) en r sin(θ)-w r cos(θ)≥G r (s) Among them, w l and w r Respectively represent the distance from the center of mass to the left and right ends of the bus, l f ,l r Respectively represent the distance from the center of mass to the front and rear ends of the bus, G l ,G r Represent the functions of the upper and lower lane boundaries respectively; The physical constraints required for the state and control quantities of the bus docking include: in min Oh, oh. max Among them, u min ,u max The physical limit of the controlled quantity.

8. The motion control method for realizing modular bus stop parking according to claim 5, characterized in that: The constructing of the vehicle dynamics model specifically includes: Get the status of each axle and use the obtained status sequence as input information; The second neural network model is used to extract the feature vector and output the resultant force and moment; The resultant acceleration is calculated based on the resultant force and moment, and the longitudinal and transverse coupling terms are added to obtain the model output.

9. The motion control method for realizing modular bus stop parking according to claim 8, characterized in that: The constructing of the vehicle dynamics model also includes training parameters in the second neural network model based on a labeled data set.

Images