A dynamic graph attention vehicle platoon cooperative control method for cut-in disturbance

Abstract

The application discloses a dynamic graph attention vehicle platoon cooperative control method for cutting-in disturbance. The method belongs to the technical field of intelligent networked vehicle cooperative control and artificial intelligence control. The method specifically comprises the following steps: establishing a vehicle queue model; performing leading vehicle control based on rolling time domain acceleration optimization; calculating the reference acceleration of the longitudinal control of the leading vehicle, and obtaining the optimal control acceleration at any time through the construction of a target function and minimization operation; performing following vehicle control based on reinforcement learning; defining the reinforcement learning state, action, reward and constraint condition of the following vehicle, constructing the reinforcement learning strategy input of the following vehicle based on a multi-head attention mechanism, and outputting the control action of the following vehicle through the reinforcement learning strategy; and performing cutting-in vehicle control based on reinforcement learning; defining the reinforcement learning state, action, reward and constraint condition of the cutting-in vehicle, and outputting the control action of the cutting-in vehicle through the reinforcement learning strategy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent connected vehicle cooperative control and artificial intelligence control technology, specifically relating to a dynamic graph attention fleet cooperative control method for intrusive disturbances. Background Technology

[0002] With the development of intelligent driving and vehicle-road cooperative technologies, vehicle platoon cooperative control has become an important means to improve road traffic efficiency and enhance driving safety. Existing platoon cooperative control methods are usually based on fixed formation structures and preset communication topologies, achieving longitudinal following control by acquiring the motion states of the preceding vehicle, the lead vehicle, or adjacent vehicles. However, in real traffic environments, situations such as external vehicles cutting in, changes in platoon members, and adjustments to inter-vehicle interaction relationships frequently occur, making it difficult for control methods under fixed topologies to adapt to dynamically changing multi-vehicle interaction scenarios. Existing vehicle platoon cooperative control methods mainly include rule-based control methods, optimization-based control methods, and learning-based control methods. Rule-based methods have simple structures and low computational overhead, but rely on fixed control parameters and preset following relationships, making it difficult to dynamically adjust control strategies according to changes in the intensity of inter-vehicle interaction. Optimization-based methods can handle constraints such as safe distance, acceleration, and comfort, but usually require pre-determining the interaction objects and topology structure, and are prone to problems such as discontinuous control response and increased computational burden in topology-changing scenarios such as vehicle cut-in.

[0003] Compared to the methods mentioned above, learning-based cooperative control methods can learn the potential relationships between vehicles from traffic data, exhibiting better scenario adaptability than traditional methods. Graph neural networks, in particular, can represent vehicles as nodes and vehicle relationships as edges, making them suitable for describing multi-vehicle coupling relationships in vehicle platoons. However, existing methods still have significant shortcomings: First, most methods still use fixed communication topologies in vehicle platoon control, making it difficult to further distinguish the importance of different effective communication edges in cooperative control; second, while some methods can calculate inter-vehicle relationship weights, these weights are mostly used as feature extraction results and fail to directly apply to the positions of effective communication edges in the vehicle platoon communication topology matrix, making it difficult to transform vehicle relationship weights into communication edge weights usable for control decisions; third, existing platoon cooperative control methods are insufficient in describing topology changes during entry scenarios, and traditional fixed topologies or hard-switching topologies struggle to achieve continuous and smooth adjustments to communication relationships and control reference relationships. Against this backdrop, how to construct a platoon cooperative control method that can accurately characterize inter-vehicle communication relationships, dynamically distinguish the importance of effective communication edges, and adapt to topology changes during external vehicle entry has become a key technical problem urgently needing to be solved in the field of intelligent connected vehicle cooperative control. Summary of the Invention

[0004] To address the problems of fixed communication topology, difficulty in distinguishing the importance of effective communication edges, difficulty in directly incorporating vehicle relationship weights into control decisions, and the potential for abrupt changes in control response caused by hard topology switching in existing vehicle platooning cooperative control methods in scenarios involving external vehicle intrusion, this invention provides a dynamic graph attention-based platooning cooperative control method oriented towards intrusion disturbances.

[0005] The method specifically involves a vehicle queue consisting of three queue vehicles and one cutting-in vehicle. The lead vehicle is vehicle 1, followed by vehicles 2 and 3, with vehicle 2 preceding vehicle 3. The cutting-in vehicle is vehicle 4, which, after entering the queue, becomes the preceding vehicle of vehicle 2. The method includes the following steps: S1. Establish a vehicle queuing model: including a longitudinal following model of the vehicle queuing, a cutting-in behavior model of the cutting-in vehicle, a longitudinal dynamics model of the vehicle, a lateral dynamics model of the vehicle, and a communication topology model based on a multi-head attention mechanism; S2. Perform lead vehicle control based on rolling time-domain acceleration optimization: calculate the reference acceleration for longitudinal control of the lead vehicle, and obtain the optimal control acceleration at any time by constructing an objective function and minimizing the operation; S3. Perform vehicle following control based on reinforcement learning: Define the vehicle following reinforcement learning state, action, reward and constraint conditions, construct the vehicle following reinforcement learning policy input based on multi-head attention mechanism, and output the vehicle following control action through the reinforcement learning policy; S4. Perform reinforcement learning-based vehicle control: Define the reinforcement learning state, action, reward and constraint of the vehicle, and output the vehicle control action through the reinforcement learning strategy.

[0006] Furthermore, the construction of the vehicle platoon longitudinal following model is specifically as follows: for any following vehicle... Define vehicle With vehicles The actual longitudinal distance between them is used to construct the expected following distance based on the current driving speed. The distance error between the actual longitudinal distance and the expected following distance is calculated, and the vehicle distance is then calculated. With vehicles Speed ​​error between; The construction of the cutting-in behavior model of the cutting-in vehicle is as follows: a cutting-in process factor is introduced. The cutting-in process factor is characterized by the continuous change of the longitudinal control influence of vehicle 2 from the initial stage of cutting-in to the completion of cutting-in through the lateral position of vehicle 4 during the cutting-in process. The cutting-in process factor is smoothed and mapped to obtain a smooth cutting-in factor.

[0007] Furthermore, the construction of the vehicle's longitudinal dynamics model is as follows: Calculate the total driving resistance experienced by the vehicle, which includes rolling resistance, air resistance, and gradient resistance; construct the vehicle's longitudinal acceleration based on the driving force and the total driving resistance experienced by the vehicle. The construction of the vehicle lateral dynamics model is specifically as follows: establish the relationship between the front and rear wheel slip angles and lateral forces, obtain the lateral forces of the front and rear wheels, and obtain the rate of change of the vehicle's center of gravity slip angle and the rate of change of the vehicle's yaw rate through the lateral forces of the front and rear wheels.

[0008] Furthermore, the construction of the communication topology model based on the multi-head attention mechanism is specifically as follows: Establish a soft handover communication topology matrix to describe the basic communication relationships between vehicles during the handover process. Further construct an effective communication side mask based on the soft handover communication topology matrix to ensure that the multi-head attention mechanism only applies to the currently allowed information interaction relationships. Construct node features and edge features. Node features uniformly represent the vehicle node state, while edge features describe the relative interaction relationship between the controlled vehicle and the information sending vehicle. Feature mapping is performed on a single attention head. For any attention head, query vector, key vector, value vector, and edge vector are constructed respectively to extract the control requirement features of the controlled vehicle, the state features of the information sending vehicle, and the interaction features of the communication edges. The original attention score on the communication edge is calculated, which simultaneously considers the control requirements of the controlled vehicle, the state features of the information sending vehicle, and the interaction features between vehicles. The original attention scores on all communication edges are arranged in matrix form to obtain the original score matrix of a single attention head. Normalized attention weight calculation involves normalizing the elements in the original score matrix of a single attention head to obtain the vehicle weights for that single attention head. For vehicles The normalized attention weights of the information are arranged in matrix form to obtain the attention weight matrix of a single attention head. The attention weight matrix is ​​then corrected to obtain the corrected attention weight matrix for a single attention head. Multi-head communication topology matrix fusion involves fusing the modified attention weight matrices of multiple attention heads to obtain a multi-head fused communication topology matrix. The multi-head attention fused communication information received by the vehicle is constructed by multiplying the relative state value vectors between vehicles in the multiple attention heads with the multi-head fused communication topology matrix.

[0009] Furthermore, the reference acceleration for longitudinal control of the leading vehicle is calculated by discretizing the standard driving cycle into a reference speed sequence and obtaining the reference acceleration based on the changes in the reference speed at adjacent times.

[0010] Furthermore, the objective function comprises three terms: a velocity tracking error term, an acceleration tracking error term, and an acceleration change term; the velocity tracking error term is the value within the prediction time domain. The speed of time and The square of the reference velocity difference at time points, and the acceleration tracking error term within the prediction time domain. acceleration at time and The square of the reference acceleration difference at time points, with the acceleration change term being the value within the prediction time domain. Acceleration at time and prediction in the time domain The square of the acceleration difference at each moment.

[0011] Furthermore, when performing reinforcement learning-based vehicle following control, the reward function is composed of a weighted sum of a relative error term and a control smoothness term, specifically: ; in, Is the vehicle in Rewards earned at any time; It is the set of information interaction vehicles corresponding to the vehicle. ； Is the vehicle in The distance error between the time and the information interaction vehicle; Is the vehicle in The speed error relative to the information interaction vehicle at any given moment; Is the vehicle in The acceleration error relative to the information interaction vehicle at any given moment; Is the vehicle in The acceleration correction at any given time; Is the vehicle in The acceleration of time; , , , and This is the weighting coefficient for the corresponding reward item; It is the reward scaling factor; It is the longitudinal following vehicle number; It is the vehicle number for information exchange.

[0012] Furthermore, the reinforcement learning policy input based on the multi-head attention mechanism is constructed as follows: For any following vehicle, the two vehicles that interact with it are identified. First, the relative state vector of the following vehicle relative to the number of vehicles interacting with it is constructed. Based on the relative state vector, the value matrix in the attention mechanism is obtained. Then, the relative state vector of the following vehicle is weighted using multi-head fusion attention weights to obtain the attention-enhanced relative state vector of the following vehicle. The attention-enhanced relative state vector of the following vehicle is vertically concatenated to obtain the policy input vector of the following vehicle.

[0013] Furthermore, when performing reinforcement learning-based control of the intervening vehicle, the reward function of the intervening vehicle is defined as follows:

[0014] in, Is the car cutting in Rewards earned at any time; It is the reward scaling factor; Is Lateral deviation of the target lane at any given time; It is the weighting coefficient for lateral deviation penalty; It is the weighting coefficient for the smoothness of movement; Is the car cutting in Reinforcement of learning actions at all times; It is the square norm of the change in action at adjacent time points; It is a safety risk penalty coefficient; Is The safety risk indicator value at any given time is set to 1 when the cutting vehicle collides, crosses the boundary, or otherwise fails to meet the safety evaluation conditions; otherwise, it is set to 0.

[0015] The beneficial effects of the method are as follows: 1. This invention models vehicle platoons and externally intervening vehicles as a unified dynamic graph structure and dynamically assigns weights to effective communication edges between vehicles based on a multi-head attention mechanism, enabling the interaction relationships between vehicles to change in real time according to operating status and control requirements. Each attention head addresses control objectives such as spacing maintenance, speed consistency, comfort constraints, and collision risk suppression, modifying the effective communication edges in the communication topology matrix, thereby allowing vehicle relationship weights to directly influence the collaborative control decision-making process. Thus, this invention overcomes the limitation of fixed communication topologies in distinguishing the importance of information sources, improving the interactive perception capability, collaborative control accuracy, and safety control capability of vehicle platoons in complex dynamic scenarios such as vehicle incursions.

[0016] 2. This invention introduces a smoothing entry factor to continuously characterize the process of an external vehicle gradually merging into the vehicle queue from an adjacent lane, and constructs a soft-switching communication topology matrix accordingly. This allows the communication reference relationship and control reference relationship to change smoothly during the entry process. Compared to traditional fixed topologies or hard-switching topologies, this invention avoids problems such as abrupt changes in control reference, acceleration fluctuations, and discontinuous queue response caused by the position change of the entering vehicle, thereby improving control smoothness, vehicle following stability, and overall queue robustness during the entry process.

[0017] 3. This invention constructs a collaborative control framework that combines lead vehicle rolling time-domain acceleration optimization control with follower vehicle reinforcement learning control. This framework enables the lead vehicle to smoothly track standard driving conditions, while the follower vehicles adaptively generate control actions based on weighted communication information, vehicle state errors, and dynamic constraints. By comprehensively considering constraints such as spacing error, speed error, acceleration, jerk, and minimum safe distance during the control process, it can reduce speed fluctuations and acceleration abrupt changes while ensuring safe following. Therefore, this invention improves the longitudinal consistency, driving comfort, training adaptability, and collaborative control reliability of vehicle platoons in dynamic traffic environments. Attached Figure Description

[0018] Figure 1 This is an architecture diagram of a dynamic graph attention-based vehicle cooperative control method for intrusive disturbances as described in an embodiment of the present invention; Figure 2 This is a schematic diagram of the vehicle queue as described in an embodiment of the present invention; Figure 3 This is a schematic diagram of the multi-head attention mechanism described in an embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0020] Example 1 This embodiment proposes a dynamic graph attention-based vehicle platoon cooperative control method for intrusion disturbances, aiming to address the problems of existing vehicle platoon cooperative control methods in scenarios involving external vehicle intrusion, such as fixed communication topology, difficulty in distinguishing the importance of effective communication edges, difficulty in directly incorporating vehicle relationship weights into control decisions, and the potential for abrupt changes in control response caused by hard topology switching. When an external vehicle enters the vehicle platoon, the relative positional relationships, communication reference relationships, and control reference relationships between vehicles in the platoon change. Existing fixed topologies typically only indicate whether communication relationships exist between vehicles, making it difficult to describe the process of an intruding vehicle gradually integrating into the platoon, and also difficult to distinguish the importance of different effective communication edges to the current control decision, easily leading to abrupt changes in control reference, vehicle acceleration fluctuations, and decreased platoon stability. To address these issues, this invention models the vehicle platoon and externally intruding vehicles as a unified dynamic graph structure, introduces a smoothing intrusion factor to achieve continuous adjustment of communication topology relationships, and employs a multi-head attention mechanism to dynamically assign weights to effective communication edges, enabling the communication edge weights to adaptively change according to objectives such as spacing control, speed control, comfort control, and collision risk control. By constructing reinforcement learning control inputs based on dynamic weighted topology, active collaborative control is achieved during the entry process, thereby improving the stability, safety, comfort, and robustness of the fleet.

[0021] The specific technical solution architecture is as follows: Figure 1 As shown, the specific technical solution is as follows: S1: Establish a vehicle queuing model The vehicle platoon consists of three platoon vehicles and one cutting-in vehicle, such as Figure 2 As shown, the leading vehicle is vehicle 1, the following vehicles are vehicles 2 and 3, and the cutting-in vehicle is vehicle 4. Cutting-in vehicle 4 cuts into the lane containing the vehicle queuing from the adjacent lane and is positioned between vehicle 1 and vehicle 2 after the cut-in. The vehicle queuing model includes a longitudinal following model, a cutting-in vehicle behavior model, a longitudinal dynamics model, a lateral dynamics model, and a communication topology model based on a multi-head attention mechanism.

[0022] S1-1: Establish a longitudinal following model for vehicle platoons For any following vehicle, the actual vehicle spacing is defined as follows to describe its longitudinal relative position to the reference vehicle:

[0023] in, Indicates vehicle With vehicles The actual longitudinal spacing between them Indicates vehicle The vertical position, Indicates vehicle The vertical position, Indicates vehicle The length of the vehicle body.

[0024] Reference vehicles refer to neighboring vehicles that have a valid information exchange relationship with the currently following vehicle. They are not limited to the vehicle directly in front of them. The specific reference objects are determined by the current communication topology, including the leading vehicle, the vehicle in front, and the cutting-in vehicle that establishes a valid communication relationship with it during the cutting-in process.

[0025] To ensure vehicle safety during longitudinal following, a desired following distance needs to be constructed based on the current driving speed. The desired following distance is defined as: in, Indicates vehicle The expected distance between the two sides, Indicates the minimum safe distance. Indicates the expected headway. Indicates vehicle The current speed.

[0026] To measure the deviation between the current following state and the desired following state of the vehicle being followed, it is necessary to calculate the spacing error and speed error, which are defined as follows:

[0027] in, Indicates vehicle and vehicles Spacing error, Indicates vehicle and vehicles Speed ​​error, Indicates reference vehicle speed, Indicates vehicle The speed.

[0028] S1-2: Establish a model of the vehicle's cutting-in behavior. Vehicle 4 cuts into the lane where the queuing is located from the adjacent lane. Initially, vehicle 4 does not yet exert a stable direct leading influence on vehicle 2. As the cut-in process progresses, vehicle 4's longitudinal control influence on vehicle 2 gradually increases. After the cut-in is complete, vehicle 4 becomes the direct leading vehicle of vehicle 2. To describe this continuous process, a cut-in progress factor is introduced, defined as:

[0029] in, Indicates the cut-in process factor. Indicates that vehicle 4 is at time The horizontal position, This indicates the lateral position of vehicle 4 when it begins to cut in. This indicates the lateral position of vehicle 4 in the target lane when it completes its cut-in. This represents the amplitude limiting function.

[0030] To avoid abrupt changes in the entry process at the beginning and end, a smoothing mapping can be applied to the original entry process factor to obtain a smooth entry factor, which is defined as:

[0031] in, This represents the smoothing cut-in factor for vehicle 4. This represents the initial entry process factor for vehicle 4.

[0032] S1-3: Establish the vehicle longitudinal dynamics model During longitudinal movement, vehicles are mainly affected by rolling resistance, air resistance, and gradient resistance.

[0033] in, Indicates vehicle The total driving resistance experienced, Indicates vehicle The rolling resistance it experiences Indicates vehicle air resistance, Indicates vehicle The slope resistance it experiences Indicates vehicle The overall vehicle quality Represents gravitational acceleration. Indicates the rolling resistance coefficient. Indicates the road slope angle. Indicates air density, Indicates the air drag coefficient. This indicates the vehicle's frontal area.

[0034] To describe the effect of longitudinal forces on a vehicle's acceleration or deceleration, the longitudinal acceleration is derived from the resultant force relationship in the longitudinal direction and is defined as follows: in, Indicates vehicle longitudinal acceleration, Indicates vehicle The driving force.

[0035] S1-4: Establish the vehicle's lateral dynamics model To describe the lateral forces generated by the front and rear wheels of a vehicle during lateral movement, we first establish the relationship between the front and rear wheel slip angles and the lateral forces, which is defined as follows:

[0036] in, Indicates vehicle Front wheel slip angle, Indicates vehicle Rear wheel slip angle, Indicates vehicle Front wheel angle, Indicates vehicle Side slip angle of the center of mass, Indicates vehicle The distance from the center of gravity to the front axle. Indicates vehicle The distance from the center of gravity to the rear axle. Indicates vehicle yaw rate, Indicates vehicle Front wheel lateral force, Indicates vehicle Rear wheel lateral force, Indicates vehicle Front wheel lateral stiffness, Indicates vehicle Rear wheel lateral stiffness.

[0037] To describe the effect of tire lateral force on vehicle lateral attitude, it is necessary to establish the dynamic relationship between the center of gravity sideslip angle and yaw rate, which is defined as:

[0038] in, Indicates vehicle Rate of change of centroid sideslip angle Indicates vehicle Rate of change of yaw rate, Indicates vehicle Front wheel lateral force, Indicates vehicle Rear wheel lateral force, Indicates vehicle Moment of inertia about the vertical axis.

[0039] The relationship between lateral position and heading angle update is defined as follows:

[0040] in, Indicates vehicle The lateral position at the next moment. Indicates vehicle The horizontal position at the current moment, Indicates vehicle lateral position change rate Indicates vehicle The heading angle at the next moment, Indicates vehicle The heading angle at the current moment, Indicates vehicle Rate of change of heading angle Indicates the sampling period.

[0041] S1-5: Establish a communication topology model based on multi-head attention mechanism The input to the multi-head attention mechanism consists of the controlled vehicle's own state, the information-sending vehicle's state, and the relative states between vehicles. The controlled vehicle's own state represents the current control requirements, the information-sending vehicle's state represents the external information source, and the relative states between vehicles represent the interaction relationships corresponding to the communication edges.

[0042] S1-5-1: Basic Communication Topology Matrix To describe the basic communication relationships between vehicles during the handover process, a soft handover communication topology matrix is ​​first presented. The communication topology matrix is ​​defined as follows:

[0043] in, This represents the soft handover communication topology matrix. Represents vehicles in the basic communication topology Receiving vehicles The communication edge weights of information This represents the smoothing cut-in factor.

[0044] To ensure that the multi-head attention mechanism only applies to currently allowed information interaction relationships, it is necessary to further construct an effective communication side mask based on the basic communication topology matrix, which is defined as:

[0045] in, Indicates vehicle Receiving vehicles Effective communication side mask during information transmission. This indicates an indicator function.

[0046] S1-5-2: Construction of Node and Edge Features To uniformly represent the state of vehicle nodes, node features are defined as follows:

[0047] in, Indicates vehicle Node embedding features, This represents the node feature mapping function.

[0048] To describe the relative interaction between the controlled vehicle and the information sending vehicle, the edge feature is defined as:

[0049] in, Indicates vehicle Receiving vehicles The information is the edge embedding feature corresponding to the information. This represents the edge feature mapping function.

[0050] S1-5-3: Feature Mapping of a Single Attention Head The multi-head attention mechanism described above includes multiple attention heads, each corresponding to a control focus direction, such as... Figure 3 As shown. Different attention heads can focus on different task requirements such as spacing maintenance, speed consistency, comfort constraints, and collision risk mitigation. For any attention head, to extract the control requirement features of the controlled vehicle, the state features of the information-sending vehicle, and the interaction features of the communication edges, query vectors, key vectors, value vectors, and edge vectors are constructed respectively, defined as follows:

[0051] in, Indicates the first A person's attention is focused on the vehicle The query vector, Indicates the first A person's attention is focused on the vehicle The key vector, Indicates the first A person's attention is focused on the vehicle With vehicles The relative state value vectors between them Indicates the first A person's attention is focused on the vehicle With vehicles The edge vectors of the relative states between them. Indicates the first A query mapping matrix for each attention head. Indicates the first The key mapping matrix of each attention head. Indicates the first The value mapping matrix of each attention head, Indicates the first The edge feature mapping matrix of each attention head.

[0052] To measure the importance of a given information-sending vehicle to the current control decision of a controlled vehicle, a raw attention score needs to be calculated on the communication edge. This score considers the control requirements of the controlled vehicle, the state characteristics of the information-sending vehicle, and the interaction characteristics between the vehicles. Based on these three types of characteristics, the raw attention score is defined as follows:

[0053] in, Indicates the first A person's attention is focused on the vehicle For vehicles The original attention score, Indicates the first Each attention head contains feature dimensions.

[0054] Arranging the original attention scores on all communication edges in matrix form yields the [number of points]. The original score matrix of each attention head is defined as follows:

[0055] in, Indicates the first The original score matrix of each attention head. This indicates the total number of vehicles.

[0056] S1-5-4: Normalized Attention Weight Calculation To obtain the importance weights of different communication edges, the normalized attention weight is defined as:

[0057] in, Indicates the first A person's attention is focused on the vehicle For vehicles Normalized attention weights of information This represents a very small positive number used to avoid a denominator of zero.

[0058] Arranging all normalized attention weights in matrix form yields the... The attention weight matrix for each attention head is defined as follows:

[0059] in, Indicates the first Attention weight matrix for each attention head.

[0060] Normalized attention weights are used to correct effective communication edges in the soft handover communication topology matrix. The corrected communication edge weights have two meanings: first, the fundamental changes in communication edges caused by the handover process; and second, the changes in information importance under different control objectives. To obtain the corrected communication edge weights for a single attention head, they are defined as follows:

[0061] in, Indicates the first The communication edge weights after attention head correction To represent the corrected result of all communication edges under a single attention head, the corrected communication topology matrix is ​​defined as:

[0062] in, Indicates the first The communication topology matrix after attention head correction This indicates element-wise multiplication of matrices.

[0063] S1-5-5: Multi-head Communication Topology Matrix Fusion Since different attention heads correspond to different control objectives, the same communication edge may have different weights under different attention heads. To obtain the communication topology matrix for integrated control decision-making, it is necessary to fuse the correction results of multiple attention heads. The multi-head fused communication topology matrix is ​​defined as follows:

[0064] in, This represents the communication topology matrix after multi-head fusion. Indicates the number of attention heads. Indicates the first The fusion coefficient of each attention head.

[0065] To ensure that the fusion coefficients of different attention heads have a consistent scale, the following constraints can be set:

[0066] Ultimately, the weighted aggregation result of the communication information received by the controlled vehicle is defined as:

[0067] in, Indicates vehicle Received multi-head attention fusion communication information, Represents vehicles in the multi-head fusion communication topology matrix Receiving vehicles Communication edge weights of information.

[0068] S2: Leading vehicle control based on rolling time-domain acceleration optimization The lead vehicle object is identified as Vehicle 1 and is used to generate the baseline motion state for convoy operation. First, the standard driving cycle is discretized into a reference speed sequence, and then the reference acceleration is obtained based on the changes in the reference speed at adjacent time points, serving as the target input for the longitudinal control of the lead vehicle.

[0069] in, Indicates that the lead car is in Reference speed at each control moment, Indicates that the lead car is in Reference acceleration at each control moment, Indicates that the lead car is in Reference speed at each control moment, This indicates the sampling period of the control system.

[0070] To ensure the lead vehicle smoothly tracks the standard driving cycle, an objective function is constructed considering speed tracking error, acceleration tracking error, and acceleration change. The optimal control acceleration at any given time is then solved by minimizing this objective function. The objective function is expressed as: in, Indicates that the lead car is in The optimization objective function at each control time point. Indicates the length of the prediction time domain. Indicates the step number in the prediction time domain. Indicates the prediction of the first time domain The acceleration at that moment. Indicates the prediction of the first time domain The speed at that moment Indicates the first Reference speed at any given moment. Indicates the first Reference acceleration at a given moment. This represents the speed tracking error weighting coefficient. This represents the acceleration tracking error weighting coefficient. This represents the weighting coefficient for the change in acceleration.

[0071] To determine the control acceleration of the lead vehicle at the current moment, the acceleration sequence in the predicted time domain is used as the variable to be optimized:

[0072] in, Indicates that the lead car is in Predicted acceleration control sequence at each control time. Indicates the prediction of the first time domain The acceleration to be optimized at each moment.

[0073] The predicted acceleration control sequence is solved by minimizing the objective function:

[0074] in, Indicates the first The optimal acceleration control sequence is obtained by solving for each control time step. This represents the value of the variable that minimizes the objective function.

[0075] To meet real-time control requirements, only the first acceleration in the sequence is taken as the control acceleration at the current moment, and the solution is recalculated at the next control moment.

[0076] in, This represents the first acceleration in the optimal acceleration control sequence.

[0077] S3: Reinforcement Learning-Based Longitudinal Following Vehicle Control The longitudinal following vehicle includes vehicle 2 and vehicle 3. The longitudinal following vehicle control based on reinforcement learning includes the following steps: S3-1: Define reinforcement learning states, actions, rewards, and constraints. For vehicle 2, its input state matrix is ​​defined as:

[0078] in, It is the input state matrix of vehicle 2 at time t; , and These are the longitudinal displacement, longitudinal velocity, and longitudinal acceleration of vehicle 2 at time t. , and These are the longitudinal displacement, longitudinal velocity, and longitudinal acceleration of vehicle 4 at time t. , and These are the longitudinal displacement, longitudinal velocity, and longitudinal acceleration of vehicle 1 at time t. , and These are the distance error, speed error, and acceleration error of vehicle 2 relative to vehicle 4; , and These are the distance error, speed error, and acceleration error of vehicle 2 relative to vehicle 1; It is the smooth cut-in factor for vehicle 4.

[0079] For vehicle 3, its input state matrix consists of five column vectors. The first column represents the vehicle's own motion state, the second and third columns represent the motion states of vehicle 2 and vehicle 1, respectively, and the fourth and fifth columns represent the relative states of vehicle 3 with respect to vehicle 2 and vehicle 3 with respect to vehicle 1, respectively. An identifier is introduced in the last row, defined as follows:

[0080] in, It is the input state matrix of vehicle 3 at time t; , and These are the longitudinal displacement, longitudinal velocity, and longitudinal acceleration of vehicle 3 at time t. , and These are the longitudinal displacement, longitudinal velocity, and longitudinal acceleration of vehicle 2 at time t. , and These are the longitudinal displacement, longitudinal velocity, and longitudinal acceleration of vehicle 1 at time t. , and These are the distance error, speed error, and acceleration error of vehicle 3 relative to vehicle 2; , and These are the distance error, speed error, and acceleration error of vehicle 3 relative to vehicle 1.

[0081] To describe the action output of the reinforcement learning controller, the action is set as a correction relative to the acceleration of vehicle 1, defined as follows:

[0082] in, Is the vehicle in The reinforcement learning action at any moment is defined as the vehicle's... The amount of correction relative to the vehicle's acceleration at any given moment; Is the vehicle in The target control acceleration at any given moment; It is vehicle 1 in Longitudinal acceleration at time t; It is the longitudinal following vehicle number.

[0083] To guide vehicles in maintaining reasonable spacing, consistent speed, and smooth longitudinal control, the reward function is a weighted average of a relative error term and a control smoothness term, defined as follows:

[0084] in, Is the vehicle in Rewards earned at any time; It is the set of information interaction vehicles corresponding to the vehicle. ； Is the vehicle in The distance error between the time and the information interaction vehicle; Is the vehicle in The speed error relative to the information interaction vehicle at any given moment; Is the vehicle in The acceleration error relative to the information interaction vehicle at any given moment; Is the vehicle in The acceleration correction at any given time; Is the vehicle in The acceleration of time; , , , and This is the weighting coefficient for the corresponding reward item; It is the reward scaling factor; It is the longitudinal following vehicle number; It is the vehicle number for information exchange.

[0085] To ensure that the motion output, final control acceleration, and vehicle operating state of vehicles 2 and 3 meet the dynamic and safety requirements, the longitudinal following vehicle control constraints are defined as follows:

[0086] in, This indicates the lower limit of reinforcement learning actions; This indicates the upper limit of reinforcement learning actions. Indicates the lower limit of the vehicle's longitudinal acceleration; This indicates the upper limit of longitudinal acceleration. Indicates the lower limit of longitudinal velocity; This indicates the upper limit of longitudinal speed. Indicates the lower limit of jerk; This indicates the upper limit of jerk.

[0087] S3-2: Constructing reinforcement learning policy input based on multi-head attention mechanism For vehicle 2, the information interaction vehicles corresponding to vehicle 2 include vehicle 4 and vehicle 1. First, the relative state vector of vehicle 2 with respect to vehicle 4 and vehicle 1 is constructed, defined as follows:

[0088] in, It is the relative state vector of vehicle 2 with respect to vehicle 4; It is the relative state vector of vehicle 2 with respect to vehicle 1.

[0089] Based on the relative state vector, the value matrix in the attention mechanism is obtained:

[0090] in, Indicates the first The value vector of vehicle 2 relative to vehicle 4 in the attention head; Indicates the first The value vector of vehicle 2 relative to vehicle 1 in each attention head; Indicates the first A value vector mapping matrix for each attention head is used to map the relative state vector to the value feature space of the attention mechanism; This indicates the attention header number.

[0091] Then, the relative state vector of vehicle 2 is weighted using multi-head fusion attention weights to obtain the attention-enhanced relative state vector of vehicle 2, defined as follows:

[0092] in, It is the attention enhancement relative state vector of vehicle 2 relative to vehicle 4; It is the attention enhancement relative state vector of vehicle 2 relative to vehicle 1; It is the multi-head fusion attention weight corresponding to the information received by vehicle 2 from vehicle 4; It is the multi-head fusion attention weight corresponding to the information received by vehicle 2 from vehicle 1.

[0093] The attention enhancement relative state vectors of vehicle 2 are concatenated vertically to obtain the policy input vector of vehicle 2, as defined below:

[0094] in, It is the attention enhancement policy input vector of the vehicle 2 input policy network.

[0095] Inputting the policy input vector of vehicle 2 into the policy network yields the Gaussian distribution of the actions, defined as follows:

[0096] in, It is the average action output by the vehicle 2 strategy network; It is the standard deviation of the action output by the vehicle 2-strategy network; It is the policy network mapping function for vehicle 2; These are the policy network parameters for vehicle 2.

[0097] Vehicle 2 samples the acceleration correction from the Gaussian distribution of the motion and generates the final control acceleration, defined as follows:

[0098] in, It is the zero-mean, unit variance Gaussian noise corresponding to vehicle 2.

[0099] For vehicle 3, construct its relative state vector with respect to vehicles 2 and 1, construct the policy input vector in the same way, and output the final action.

[0100] S4: Reinforcement Learning-Based Vehicle Control The vehicle to be switched in is vehicle 4. During the switching process, vehicle 4 does not receive information from the vehicle queue communication topology.

[0101] S4-1: Define reinforcement learning states, actions, rewards, and constraints. The status of the vehicle entering the lane consists only of its own longitudinal motion status, its own lateral motion status, target lane deviation information, entry process information, and historical control information, as defined below:

[0102] in, It is the input state matrix of vehicle 4 at time t. It is the acceleration of vehicle 4 at time 4. It is the lateral position of vehicle 4 at time 4. It is the lateral velocity of vehicle 4 at time t. It is the heading angle of vehicle 4 at time 4; It is the heading angular velocity of vehicle 4 at time t. It is the lateral deviation of vehicle 4 relative to the center line of the target lane; It is the longitudinal acceleration correction of vehicle 4 at the previous moment; It is the front wheel steering angle correction amount of vehicle 4 at the previous moment.

[0103] The lateral deviation of the target lane is defined as follows:

[0104] in, It is the lateral position of the target lane centerline after vehicle 4 has entered the lane.

[0105] To input the policy network, the vehicle's 4-state matrix is ​​expanded into a one-dimensional state vector, defined as follows:

[0106] in, It is a one-dimensional state vector of the vehicle's 4-input policy network; It is a matrix vectorization operation.

[0107] The motion of vehicle 4 is set as longitudinal acceleration correction and front wheel steering angle correction, defined as follows:

[0108] in, It is the vehicle's reinforcement learning action at any moment; Vehicle 4 at time The longitudinal acceleration correction amount; Vehicle 4 at time The front wheel steering angle correction amount.

[0109] The final control input is defined as follows:

[0110] in, It is the target longitudinal acceleration of vehicle 4 at time t. This represents the target longitudinal acceleration of vehicle 4 at the previous moment; It is the target front wheel steering angle of vehicle 4 at time 4; It is the target front wheel steering angle of vehicle 4 at time 4.

[0111] The reward function for vehicle 4 is defined as follows:

[0112] in, It is the reward that vehicle 4 receives at any given moment; It is the reward scaling factor; It is the weighting coefficient for lateral deviation penalty; It is the weighting coefficient for the smoothness of movement; It is the reinforcement learning action of vehicle 4 in the previous moment; It is the square norm of the change in action at adjacent time points; It is a safety risk penalty coefficient; It is a safety risk indicator. It takes a value of 1 when vehicle 4 collides, crosses the boundary, or otherwise fails to meet the safety evaluation conditions; otherwise, it takes a value of 0.

[0113] To ensure that the longitudinal and lateral movements of vehicle 4, as well as the cutting-in process, meet the requirements of vehicle dynamics, lane boundaries, and cutting-in process, the control constraints for the cutting-in vehicle are defined as follows:

[0114] in, This indicates the lower limit of the longitudinal acceleration correction for the vehicle. This indicates the upper limit of the longitudinal acceleration correction for the vehicle. This indicates the lower limit of the steering angle correction for the four front wheels of the vehicle; This indicates the upper limit of the steering angle correction for the four front wheels of the vehicle. This indicates the lower limit of the longitudinal acceleration of the vehicle's four targets; This indicates the upper limit of the longitudinal acceleration of the vehicle's four targets. This indicates the lower limit of the front wheel steering angle of the vehicle's four targets; This indicates the upper limit of the front wheel steering angle of the vehicle's four targets. This indicates the lower limit of the vehicle's longitudinal speed; This indicates the maximum longitudinal speed of the vehicle. This indicates the lower limit of the vehicle's 4-speed jerk. This indicates the maximum acceleration limit of the vehicle's 4-speed acceleration.

Claims

1. A dynamic graph attention vehicle platoon cooperative control method for cut-in disturbance, the vehicle platoon is composed of three platoon vehicles and one cut-in vehicle, wherein the leading vehicle is vehicle 1, the following vehicles are vehicle 2 and vehicle 3, vehicle 2 is the front vehicle of vehicle 3, and the cut-in vehicle is vehicle 4, which becomes the front vehicle of vehicle 2 after cutting in, characterized in that, The method comprises the following steps: S1, establishing a vehicle platoon model: including a vehicle platoon longitudinal following model, a cut-in vehicle cut-in behavior model, a vehicle longitudinal dynamics model, a vehicle lateral dynamics model, and a communication topology model based on a multi-head attention mechanism; S2, performing leading vehicle control based on rolling horizon acceleration optimization: calculating the reference acceleration of the longitudinal control of the leading vehicle, and obtaining the optimal control acceleration at any time through the construction of an objective function and minimization operation; S3, performing following vehicle control based on reinforcement learning: defining the reinforcement learning state, action, reward, and constraint condition of the following vehicle, constructing the reinforcement learning strategy input of the following vehicle based on the multi-head attention mechanism, and outputting the control action of the following vehicle through the reinforcement learning strategy; S4, performing cut-in vehicle control based on reinforcement learning: defining the reinforcement learning state, action, reward, and constraint condition of the cut-in vehicle, and outputting the control action of the cut-in vehicle through the reinforcement learning strategy.

2. The dynamic graph attention platoon cooperative control method for cut-in perturbation according to claim 1, characterized in that, The construction of the vehicle platoon longitudinal following model is specifically: for any following vehicle , defining the actual longitudinal distance between the vehicle and the vehicle , constructing the expected following distance according to the current driving speed, calculating the distance error of the actual longitudinal distance and the expected following distance, calculating the speed error between the vehicle and the vehicle ; The construction of the cut-in vehicle cut-in behavior model is as follows: a cut-in process factor is introduced, which is represented by the lateral position of the vehicle 4 in the cut-in process, and is used to represent the continuous change of the influence of the vehicle 2 from the initial cut-in to the completion of the longitudinal control; the cut-in process factor is smoothed to obtain a smooth cut-in factor.

3. The dynamic graph attention platoon cooperative control method for cut-in perturbation according to claim 2, characterized in that, The construction of the vehicle longitudinal dynamics model is as follows: The total driving resistance of the vehicle is calculated, including rolling resistance, air resistance, and slope resistance; the vehicle longitudinal acceleration is constructed according to the driving force of the vehicle and the total driving resistance of the vehicle; The construction of the vehicle lateral dynamics model is as follows: the relationship between the side slip angle of the vehicle front and rear wheels and the lateral force is established, the lateral force of the vehicle front and rear wheels is obtained, and the vehicle mass center side slip angle change rate and vehicle yaw rate change rate are obtained through the lateral force of the vehicle front and rear wheels.

4. The dynamic graph attention car platoon cooperative control method for cut-in perturbation according to claim 3, characterized in that, The construction of the communication topology model based on the multi-head attention mechanism is as follows: A soft switching communication topology matrix is established to describe the basic communication relationship between vehicles in the cut-in process, and an effective communication edge mask is further constructed based on the soft switching communication topology matrix to ensure that the multi-head attention mechanism only acts on the information interaction relationship that is currently allowed to exist; Node features and edge features are constructed, the node features uniformly represent the vehicle node state, and the edge features describe the relative interaction relationship between the controlled vehicle and the information sending vehicle; Feature mapping of a single attention head is performed, for any attention head, query vectors, key vectors, value vectors, and edge vectors are constructed respectively, and the control demand features of the controlled vehicle, the state features of the information sending vehicle, and the interaction features of the communication edge are extracted; The original attention score on the communication edge is calculated, which considers the control demand of the controlled vehicle, the state features of the information sending vehicle, and the interaction features between vehicles, and all the original attention scores on the communication edges are arranged in matrix form to obtain the original score matrix of a single attention head; The normalized attention weight calculation is to normalize the elements in the original score matrix of a single attention head to obtain the normalized attention weight of the vehicle The normalized attention weight calculation is to normalize the elements in the original score matrix of a single attention head to obtain the normalized attention weight of the vehicle The normalized attention weight calculation is to normalize the elements in the original score matrix of a single attention head to obtain the normalized attention weight of the vehicle The normalized attention weight calculation is to normalize the elements in the original score matrix of a single attention head to obtain the normalized attention weight of the vehicle The multi-head communication topology matrix fusion fuses the modified attention weight matrices of multiple attention heads, obtains a multi-head fused communication topology matrix, and constructs multi-head attention fused communication information received by the vehicle according to a product of a relative state value vector between the vehicles in the multiple attention heads and the multi-head fused communication topology matrix.

5. The dynamic graph attention platoon cooperative control method for cut-in perturbation according to claim 4, characterized in that, The reference acceleration of the leading vehicle longitudinal control is specifically: discretizing the standard driving cycle into a reference speed sequence, and obtaining the reference acceleration according to the change of the reference speed at adjacent time points.

6. The dynamic graph attention platoon cooperative control method for cut-in perturbation according to claim 5, characterized in that, The target function includes three terms, respectively, a velocity tracking error term, an acceleration tracking error term and an acceleration variation term; the velocity tracking error term is the square of the difference between the velocity at the time point and the reference velocity at the time point in the prediction time domain, the acceleration tracking error term is the square of the difference between the acceleration at the time point and the reference acceleration at the time point in the prediction time domain, and the acceleration variation term is the square of the difference between the acceleration at the time point and the acceleration at the time point in the prediction time domain.

7. The dynamic graph attention platoon cooperative control method for cut-in perturbation according to claim 6, characterized in that, When the following vehicle control based on reinforcement learning is performed, the reward function is composed of a relative error term and a control smoothness term, and is specifically: ； wherein, is the reward obtained by the vehicle at the time instant ; is the set of information interaction vehicles corresponding to the vehicle ； is the distance error of the vehicle relative to the information interaction vehicles at the time instant ; is the speed error of the vehicle relative to the information interaction vehicles at the time instant ; is the acceleration error of the vehicle relative to the information interaction vehicles at the time instant ; is the acceleration correction amount of the vehicle at the time instant ; is the jerk of the vehicle at the time instant ; , , , and are the weight coefficients of the corresponding reward items; is the reward scaling coefficient; is the longitudinal following vehicle number; is the information interaction vehicle number.

8. The dynamic graph attention platoon cooperative control method for cut-in perturbation according to claim 7, characterized in that, The input of the reinforcement learning strategy based on the multi-head attention mechanism is specifically: for any following vehicle, two vehicles that interact with the following vehicle are determined, a relative state vector of the following vehicle relative to the information interaction vehicle quantity is first constructed, a value matrix in the attention mechanism is obtained according to the relative state vector, then the relative state vector of the following vehicle is weighted by using the multi-head fusion attention weight, and an attention enhanced relative state vector of the following vehicle is obtained, and the attention enhanced relative state vector of the following vehicle is longitudinally spliced to obtain a strategy input vector of the following vehicle.

9. The dynamic graph attention platoon cooperative control method for cut-in perturbation according to claim 8, wherein, When the cut-in vehicle control based on reinforcement learning is performed, the reward function of the cut-in vehicle is defined as follows: wherein, is the reward obtained by the cut-in vehicle at ; is the reward scaling coefficient; is the target lane lateral deviation at ; is the lateral deviation penalty weight coefficient; is the action smoothness term weight coefficient; is the reinforcement learning action of the cut-in vehicle at ; is the square norm of the action change amount at the adjacent time; is the safety risk penalty coefficient; is the safety risk indicator at , which takes the value of 1 when the cut-in vehicle has a collision, crosses the boundary, or other conditions that do not meet the safety evaluation, and otherwise takes the value of 0.

Images