Intelligent connected vehicle platoon resilience control method under mixed traffic flow

Abstract

This invention discloses a resilient control method for intelligent connected vehicle platooning under mixed traffic flow, belonging to the technical field of vehicle control methods. The method includes the following steps: evaluating V2V communication quality and characterizing random delays; establishing a prediction model of HDV clustering behavior and a CAV error dynamics equation; dynamically fusing onboard sensor information and V2V shared information based on communication quality; establishing a delay-augmented state-space model, incorporating communication delays into a fixed-dimensional state-space representation; designing and solving a model predictive control optimization problem to obtain nominal control commands that satisfy constraints; designing an online residual learning compensation module to compensate for model mismatch and unmodeled dynamics in real time; and fusing and executing control commands through a resilience gating mechanism to achieve resilient platooning control. This method does not require a large amount of offline training data, improving practicality and deployability.

Description

Technical Field

[0001] This invention relates to the field of vehicle control methods, and in particular to a method for intelligent connected vehicle platooning resilience control under mixed traffic flow. Background Technology

[0002] With the rapid development of intelligent transportation systems, connected and automated vehicles (CAVs) have become a significant driving force for the transformation of modern transportation systems. Through the deep integration of vehicle-to-everything (V2X) communication and automatic control technologies, CAV platooning control can significantly improve road capacity, reduce fuel consumption, decrease traffic accident rates, and effectively suppress traffic flow oscillations. However, the large-scale adoption of CAV technology is a gradual process, and future road traffic will long present a complex situation of mixed traffic flow between CAVs and human-driven vehicles (HDVs). This mixed traffic flow environment brings unprecedented challenges: HDV driving behavior is significantly random and unpredictable; its rapid acceleration and deceleration can easily trigger traffic fluctuations that propagate upstream within the platoon, creating "phantom traffic jams." Simultaneously, vehicle-to-vehicle (V2V) communication, on which CAVs rely, faces reliability issues such as random latency, packet loss, and intermittent interruptions in real-world environments. This dual superposition of physical layer disturbances and information layer uncertainties makes it difficult for traditional platooning control methods to guarantee the safety and stability of the system.

[0003] Currently, research on hybrid traffic formation control mainly falls into three categories. Model-based methods, such as Model Predictive Control (MPC), offer theoretical guarantees and constraint handling capabilities, but their performance is limited under model mismatch and strong disturbances. Data-driven deep reinforcement learning methods can adapt to complex scenarios, but suffer from low sample efficiency and difficulty in ensuring security. In recent years, physical-information fusion learning methods, by combining mechanistic models with data-driven compensation, have shown promising application prospects. However, existing methods generally focus on nominal performance optimization, paying insufficient attention to the system's resilience under extreme disturbances and lacking adaptive mechanisms for communication quality degradation, making it difficult to achieve smooth performance degradation and rapid recovery. Facing the practical deployment needs of future intelligent transportation systems, hybrid traffic formation control technology is developing towards resilience and adaptability. By establishing a communication quality perception mechanism, designing multi-source information adaptive fusion strategies, and constructing a learning framework with physical constraint guarantees, it is hoped that the formation system can achieve stable operation, smooth degradation, and rapid recovery capabilities under dual disturbances, providing reliable technical support for the large-scale application of intelligent connected vehicle technology.

[0004] Currently, research on intelligent connected vehicle (HV) platooning control in mixed traffic flows has become a core direction in the field of intelligent transportation, encompassing several key technological paths. First, model-based predictive control methods provide theoretical guarantees and optimization frameworks for control decisions by establishing vehicle dynamics models and platooning cooperative constraints. Second, data-driven learning control methods utilize the nonlinear fitting capabilities of intelligent algorithms such as neural networks to achieve adaptive control strategies in complex scenarios. Furthermore, physical-information fusion learning methods combine mechanistic models with data compensation to improve system robustness while maintaining physical interpretability. All of these studies focus on cooperative platooning control, leveraging onboard communication and intelligent decision-making technologies to lay the foundation for improving the efficiency, safety, and stability of mixed traffic flows. However, existing methods still have significant shortcomings in their resilience to the dual uncertainties of HDV behavioral disturbances and V2V communication quality degradation.

[0005] The existing research directions mainly include three types: model-based predictive control methods, data-based neural network control methods, and physical information fusion observation control methods.

[0006] The first type: Model-based methods typically construct longitudinal dynamics and communication topology models of vehicle platoons, design distributed or centralized model predictive controllers, and solve for the optimal control input through rolling optimization. This method can achieve cooperative adaptive cruise control under nominal operating conditions, ensuring platoon stability and following performance. A typical example is the connected vehicle platoon control method based on distributed MPC, which improves inter-vehicle cooperation by designing multi-component cost functions and iteratively optimizing control strategies. However, this type of method heavily relies on accurate vehicle dynamics models and HDV behavior prediction models, facing significant model mismatch problems in mixed traffic environments: on the one hand, HDV driving behavior has strong randomness and driving style differences, making it difficult to accurately characterize with deterministic models; on the other hand, random delays and packet loss in V2V communication cause temporal inconsistencies in platoon state information, and existing MPC methods lack adaptive adjustment mechanisms for communication quality degradation, resulting in insufficient real-time response and poor robustness when the channel deteriorates.

[0007] The second approach, data-driven methods, primarily utilizes the nonlinear expressive capabilities of deep learning to implement end-to-end formation control strategies by constructing neural network controllers. This method adapts to complex traffic scenarios through offline training or online learning, overcoming the difficulties of accurate modeling to some extent. A typical example is the formation control method based on a composite neural network controller, which combines a neural network with a linear quadratic regulator (LQR), maintaining basic dynamic performance even when the neural network training effect is poor. However, purely data-driven methods suffer from large training sample requirements and low learning efficiency, especially in diverse scenarios with mixed traffic flows, requiring a large amount of simulation or real vehicle data to cover various combinations of operating conditions. More critically, this method lacks explicit guarantees of physical constraints, making it difficult to provide theoretical proof of safety and stability. Under dynamic changes in communication quality or sudden HDV disturbances, the generalization ability and safety margin of the control strategy are difficult to verify, posing a risk of control failure or drastic performance degradation.

[0008] The third approach focuses on physical-information fusion, combining model and observer techniques to achieve state estimation and collaborative control. This method utilizes vehicle dynamics models and real-time observation information to design state observers, compensating for the influence of unmeasurable states or measurement noise. This allows for the design of intermittent control or event-triggered control protocols to reduce communication overhead. A typical example is the intelligent connected vehicle platooning control method in a zero-trust environment, which uses augmented systems to design observers and defines event-triggered strategies to achieve intermittent control. This method integrates physical models and data feedback to some extent, improving the robustness of the control system. However, existing fusion methods primarily focus on nominal performance optimization or communication resource conservation, neglecting the system's resilience under extreme disturbances: they lack adaptive compensation mechanisms for communication quality awareness, failing to dynamically adjust control strategy weights based on real-time channel conditions; and they lack robust designs for smooth performance degradation and rapid recovery, making it difficult to guarantee the safe and stable operation and continuous service quality of the platooning system under superimposed disturbances. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide an intelligent connected vehicle platooning resilience control method for mixed traffic flow that maintains the interpretability and constraint satisfaction guarantee of model-based methods, overcomes the problems of low sample efficiency and difficult security verification of pure data-driven methods, can continuously adapt to environmental changes, does not require a large amount of offline training data, and improves practicality and deployability.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for intelligent connected vehicle platooning resilience control under mixed traffic flow, comprising the following steps: S1: Establish communication quality assessment and Markov channel model to evaluate V2V communication quality and characterize random delay; S2: Construct a hybrid platoon vehicle dynamics model, establish an HDV clustering behavior prediction model and CAV error dynamics equations; S3: Design a heterogeneous information fusion strategy for communication quality perception, dynamically fusing on-board sensor information and V2V shared information based on communication quality; S4: Establish a delay augmented state space model to incorporate communication delay into a fixed-dimensional state space representation; S5: Design and solve the model predictive control optimization problem to obtain the nominal control command that satisfies the constraints; S6: Design an online residual learning compensation module to compensate for model mismatch and unmodeled dynamics in real time; S7: Achieve formation resilience control by fusing and executing control commands through a resilience gating mechanism.

[0011] The beneficial effects of adopting the above technical solution are as follows: First, the residual learning framework based on physical information fusion maintains both the interpretability and constraint satisfaction guarantee of model-based methods, while overcoming the problems of low sample efficiency and difficulty in verifying security of pure data-driven methods; Second, through resilient gating mechanisms and heterogeneous information fusion strategies, resilient control capabilities under dual perturbations are achieved, including maintaining core functions, smooth performance degradation, and rapid recovery; Third, through online learning and adaptive fusion, the system can continuously adapt to environmental changes without requiring a large amount of offline training data, thus improving practicality and deployability. Attached Figure Description

[0012] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0013] Figure 1 This is the main flowchart of the method described in the embodiments of the present invention; Figure 2 This is a flowchart of the communication quality assessment and Markov channel modeling process in the method described in this invention; Figure 3 This is a flowchart of the mixed formation vehicle dynamics modeling process in the method described in this invention; Figure 4 This is a flowchart of the heterogeneous information fusion strategy in the method described in this invention; Figure 5 This is a flowchart of the delay augmented state space modeling process in the method described in this invention; Figure 6 This is a flowchart of the model predictive control optimization process in the method described in this invention; Figure 7 This is a flowchart of the online residual learning compensation process in the method described in this invention; Figure 8 This is a flowchart of the control instruction fusion process for the resilience gating mechanism in the method described in this invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0015] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0016] Overall, such as Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for intelligent connected vehicle platooning resilience control under mixed traffic flow, the method comprising the following steps: Step S1: Establish a communication quality assessment and Markov channel model to evaluate V2V communication quality and characterize random delay; Step S2: Construct a hybrid platoon vehicle dynamics model, establish an HDV clustering behavior prediction model and a CAV error dynamics equation; Step S3: Design a heterogeneous information fusion strategy for communication quality perception, and dynamically fuse vehicle sensor information and V2V shared information based on communication quality. Step S4: Establish a delay augmented state space model, incorporating communication delay into a fixed-dimensional state space representation; Step S5: Design and solve the model predictive control optimization problem to obtain the nominal control command that satisfies the constraints; Step S6: Design an online residual learning compensation module to compensate for model mismatch and unmodeled dynamics in real time; Step S7: The formation's resilience is achieved by fusing and executing control commands through a resilience gating mechanism.

[0017] The above steps will be explained in detail below with specific methods: like Figure 2 As shown, the method for establishing the communication quality assessment and Markov channel model in step S1 includes the following steps: S1-1: Establish a delay evolution model based on Age of Information (AoI). To quantify the degree of information obsolescence caused by random end-to-end delays, packet loss, and communication interruptions, Age of Information (AoI) is used as a metric, denoted as... The discrete-time recursive evolution model of AoI is as follows: in, I k Package arrival indicator, indicating time k Whether the intelligent connected vehicle successfully received new information; when I k When =1, the intelligent connected vehicle successfully received new information. I k =0 The intelligent connected vehicle did not receive any information; D k The end-to-end delay for receiving new information; To control the sampling period; The maximum communication latency that the system can tolerate S1-2: Establishing a two-state Markov chain model to characterize channel state switching. To characterize the stochastic nature of the communication link, a two-state Markov chain model is used to describe the channel state. ,in This indicates a good channel condition. This represents a degraded (Bad) channel state. The channel state transition probability is defined as: in, To be in good condition Transfer to deteriorated state The transition probability, with a value range of . ; To recover from the deteriorated state Transfer to good condition The transition probability, with a value range of . .

[0018] S1-3: Construct a data packet reception model based on channel state.

[0019] Given channel state (in ), data packet reception event Follows Bernoulli distribution: in, Channel state The probability of successfully receiving the next data packet. Specifically, This represents the probability of successful reception under good channel conditions. Represents the probability of successful reception under degraded channel conditions, satisfying .

[0020] S1-4: Establish an end-to-end communication delay model based on channel state.

[0021] When data packets are successfully received ( End-to-end communication delay The model is as follows: in, As a reference delay; The unit delay increment; Let be a random variable that follows a geometric distribution, with parameter . ( (This is the current channel state), and it is decremented by 1 to allow for zero additional delay. The parameter is The geometric distribution. Parameters are used under good channel conditions. Use smaller parameters in degraded channel conditions (satisfy This is used to generate a heavy-tailed delay distribution, which reproduces the observed delay spike behavior.

[0022] S1-5: Construct online communication quality metrics.

[0023] To explicitly quantify the information uncertainty caused by communication quality degradation, an online communication quality metric is introduced. Based on the above AoI evolution process and data packet reception process, The discrete-time implementation is as follows: in, As a sensitivity factor for information freshness, it characterizes the degree of impact of AoI on communication quality; The length of the sliding time window used to evaluate communication reliability; In the time window The number of nominal data packets expected to be received (rounded down to the nearest integer). The range of values ​​is When communication quality is good Approaching 1, when communication quality degrades Decrease.

[0024] like Figure 3 As shown, the specific method for constructing the platooning vehicle dynamics model under mixed traffic flow in step S2 includes the following steps: S2-1: Define the topology and communication relationships of intelligent connected vehicle platoons. Define the local platoon unit containing... The vehicle, led by the navigation CAV (numbered as...) ), and follow-up HDV and the accused CAV (number) The vehicle index set is composed of ( ). The assembly of all vehicles in the formation is denoted as The CAV set is HDV collection is CAV Real-time acquisition of information from vehicles ahead using onboard sensors The status information is received from the pilot CAV via V2V communication. Status information.

[0025] S2-2: Establish the longitudinal dynamics model of the vehicles in the formation. The longitudinal dynamics of each vehicle in the formation satisfy the following equations: in, For vehicles At any moment The longitudinal position; For vehicles longitudinal velocity; For vehicles longitudinal acceleration; For vehicles With the vehicle immediately in front The distance between vehicles; For vehicles With the vehicle immediately in front The relative speed between them.

[0026] S2-3: Establish a general form of error dynamics based on the desired spacing strategy. (Vehicle) With the car in front The general dynamic form of the spacing error and relative velocity error between them is: in, The expected headway; The desired stationary distance; Actual vehicle spacing Spacing from the expected distance The difference is defined as .

[0027] S2-4: Establish a dynamic model for HDV car-following behavior. The car-following behavior of HDVs in formation is described by the following dynamic model: in, For HDV index, ; For the headway gain; This is the relative velocity gain; This is the desired speed function based on the vehicle spacing; For vehicle spacing variables; This represents the minimum spacing in a stationary state. The maximum distance threshold for maintaining maximum vehicle speed; This represents the maximum expected speed.

[0028] S2-5: Establish the dynamics of HDV spacing error and relative velocity error. Based on the general form in step S2-3, HDV With the car in front The dynamics of the spacing error and relative velocity error between them are as follows: in, , HDV Compared to the car in front The expected headway and expected stationary distance; , These are the spacing error and the relative speed error, respectively.

[0029] S2-6: Establish a delay dynamics model for the CAV actuator. Under adverse conditions such as communication delay, packet loss, or intermittent interruptions, the CAV... The dynamics can be described as follows: in, The desired acceleration command; The actuator time constant characterizes the dynamics of the transmission system and the execution delay; The communication delay (in continuous-time representation) defined in step S1. For CAV With the car in front The distance between vehicles. For CAV i The acceleration.

[0030] S2-7: Estimation of macroscopic behavioral parameters of clustered HDVs based on traffic wave theory. This is achieved through macroscopic traffic wave speeds. Estimated HDV With CAV Time-varying expected time intervals between and desired stationary distance : in, For CAV Compared to its immediate predecessor HDV The expected time between the locomotive headway; For CAV Compared to its immediate predecessor HDV The expected static distance between them; For HDV With CAV The actual distance between vehicles is defined as ; For HDV speed; This indicates that CAV takes into account the impact of traffic waves. Compared to CAV The time-varying time interval.

[0031] S2-8: Establish CAV With Leading CAV The error dynamics equation between CAV. With CAV The error dynamics between them can be written as: in, For CAV Compared to CAV The spacing error is defined as ; For CAV With CAV The actual distance between vehicles; For CAV Compared to CAV The relative velocity is defined as ; The term represents the clustered disturbance, indicating the impact of the preceding vehicle's HDV. And Leading CAV The unpredictability and clustering disturbances caused by macroscopic estimation errors; To lead CAV The acceleration.

[0032] S2-9: Establish CAV With the HDV in front The error dynamics equation between CAV. With HDV The error dynamics between them can be written as: in, For CAV Compared to the HDV in front Spacing error; For the HDV in front The acceleration.

[0033] like Figure 4 As shown, the specific method for designing a heterogeneous information fusion strategy for communication quality awareness in S3 includes the following steps: S3-1: Define the blending spacing error and the blending speed error. Define the blending spacing error. and fusion speed error as follows: in, To meet the HDV in front Error fusion weights associated with information sources; To cooperate with CAV Error fusion weights associated with information sources.

[0034] S3-2: Establish normalization constraints for the fusion weights. The fusion weights satisfy the normalization constraints: S3-3: Define the fusion ratio function for adaptive communication quality. Define the fusion ratio function. Taking into account both the timeliness and reliability of information: in, This is the fusion ratio function, with a range of values ​​of [value range missing]. ; To assign to CAV under fully trusted conditions The nominal weight upper bound satisfies ; The communication quality confidence level defined in step 1.6 has a value range of [value range missing]. .

[0035] S3-4: Calculate the fusion weights of the preceding vehicle information source. Based on the normalization constraints of step S3-2 and the fusion ratio function of step S3-3, the preceding vehicle HDV... Information source fusion weight The calculation is as follows: in, The range of values ​​is When communication quality is good ( )hour, As communication quality increases, the system relies more on V2V shared navigation CAV information; when communication quality degrades ( )hour, The system automatically steers by relying more on information about the vehicle ahead obtained from onboard sensors.

[0036] S3-5: Update the fusion weights and calculate the fusion error state based on real-time communication quality. At each control time... Perform the following calculation process: S3-5-1: Obtain the real-time communication quality confidence score output from step S1-5 ; S3-5-2: Calculate the fusion ratio according to the formula in step S3-3. ; S3-5-3: Calculate the weight of the vehicle in front according to the formula in step S3-4. ; S3-5-4: Obtain the original error status output from steps S2-8 and S2-9: , , , ; S3-5-5: Calculate the fusion spacing error according to the formula in step S3-1: S3-5-6: Calculate the fusion speed error according to the formula in step S3-1: like Figure 5 As shown, the method for establishing the delayed augmented state space model in step S4 specifically includes the following steps: S4-1: Define the low-dimensional state vector and the perturbation vector. Define the low-dimensional state vector. and perturbation vector as follows: in This is a disturbance to the lumped system.

[0037] S4-2: Construct the continuous-time fusion error state dynamics equation. The continuous-time dynamics equation for the fusion error state is: in, For communication delay The control input; the model matrix is ​​represented as: , in, For time-varying system matrices; To control the input matrix; The input matrix is ​​a time-varying perturbation matrix.

[0038] S4-3: The model parameters are discretized using the zeroth-order preservation assumption. Within each sampling interval, the parameters are... , and Under the zeroth-order preservation assumption, these are considered piecewise constants, denoted as follows: , and .in, For the first Each sampling time, This is the index for discrete time steps.

[0039] S4-4: Discretization of continuous-time dynamics using the forward Euler method. The discretization of continuous-time dynamics using the forward Euler method yields the discrete-time state-space equations: in, For the first Discrete state vectors at each time step; For delay Control input after the step; The number of discrete delay steps is determined by the communication delay. Divide by sampling period Rounded down to the nearest integer; The upper bound of the discrete delay steps is given by the maximum tolerable delay. Divide by Rounded up, the discrete-time model matrix is: , , in, For discrete system matrices; For discrete control input matrix; The input matrix is ​​a discrete perturbation matrix.

[0040] S4-5: Introduce delayed augmented state vectors to construct fixed-dimensional models.

[0041] S4-5-1: Introducing a Delay-Augmented State Vector : in They are respectively k -1, k -2, k Time-based control input.

[0042] S4-5-2: Define the delayed selection operator Define the input history vector Then the delayed input satisfies .

[0043] , in For the first One element is 1, and the rest are 0. dimensional standard basis vectors S4-6: Construct a delay-free linear system representation. Define the control increment. The delayed-free linear system representation can be obtained as follows: The system matrix is ​​defined as follows: , , , in, To augment the system matrix; To augment the control increment input matrix; For augmented perturbation input matrix; Input the historical update matrix; express A zero matrix with 3 rows and 3 columns; This represents the zero vector, which is arranged in 3 rows and 1 column.

[0044] S4-7: Construction A stacking state prediction model with a step prediction range.

[0045] S1: By predicting the range Freeze within step and Construct the stacked prediction formula. Define the stacked prediction state vector. Stacked control incremental vector and stacked perturbation vector : , , in, Indicates at time Based on current information, time Augmented state prediction; Indicates at time For time Control increment; Indicates at time For time The disturbance prediction; This represents the number of steps for predicting the range.

[0046] S2: Obtain the standard linear state prediction model: The model matrix is ​​defined as follows: , , in, This is the state transition matrix; To control the incremental influence matrix; This is the disturbance effect matrix.

[0047] like Figure 6 As shown, the design and solution of the model predictive control optimization problem in S5 specifically includes the following steps: S5-1: Construct the MPC multi-objective cost function. The prediction range is... The step-by-step MPC optimization problem is expressed as: in, For a moment The total cost function; For at any time Based on current information, time state vector The prediction; For the prediction step index; Number of steps for the prediction range; This is the state weight matrix; To control the incremental weight scalar; This is the terminal state weight matrix.

[0048] weight matrix , and Defined as: in These represent the weighting coefficients related to spacing error, velocity error, and acceleration, respectively. This indicates an increase in the amount of punishment controlled; These represent the terminal weighting coefficients for spacing error, velocity error, and acceleration, respectively.

[0049] S5-2: Construct constraints for the MPC optimization problem.

[0050] S5-2-1: Define control increment constraints. Control increment Upper and lower bound constraints must be satisfied: in, To control the lower bound of the increment; To control the upper bound of the increment.

[0051] S5-2-2: Define control input constraints. Control input The control increment is obtained by accumulating control increments, and must satisfy upper and lower bound constraints: in, To control the lower bound of the input; To control the upper limit of the input; This is the control input from the previous moment.

[0052] S5-2-3: Define state vector constraints. Predict the state vector. It is necessary to start from the augmented state vector Extract from, and satisfy upper and lower bound constraints: in, This is the lower bound of the state vector; The upper bound of the state vector; the state selection matrix. ,in It is a 3-order identity matrix. 3 lines The zero matrix of columns.

[0053] S5-3: Reconstruct the MPC optimization problem into a standard quadratic programming form. Utilize the stacking state prediction model from step S4-7. The MPC optimization problem is refactored into a standard QP form: st in This is the optimal control input correction value; This is the coefficient matrix of the quadratic terms; This is the vector of linear term coefficients; , These are the state weight diagonal matrix and the penalty control diagonal matrix, respectively. Choose a diagonal matrix for the state; This is the inequality constraint coefficient matrix; The corresponding inequality boundary vectors; for N 3D identity matrix; It is a length of N A column vector of all 1s; It is a standard lower triangular cumulative matrix; To control the selection matrix; These are the upper and lower bounds of the superposition of state vectors, respectively.

[0054] S5-4: Extracting the optimal control increment sequence The first element is used as the nominal control increment at the current moment: Here, [·]1 represents the extraction of the first element of the vector.

[0055] like Figure 7 As shown, the method for designing the online residual learning compensation module in S6 includes the following steps: S6-1: Construct the input set for the physical controller. At time step... Controller input set based on physical model Defined as: in for The predicted system disturbance; for CAV With CAV The expected static distance between them.

[0056] S6-2: Characterizes the input-output mapping of the physical controller. The controller based on the physical model calculates the nominal control command online through QP-MPC in step 5. The mapping relationship is expressed as follows: in, Represents a physical model of a mixed formation; For parameters of the hybrid formation physical model; This represents the MPC optimization solution process defined in step S5; This represents a linear control law mapping.

[0057] S6-3: Constructing the input set for the residual learning module. This input set is used to compensate online for reference mismatch caused by mixed traffic disturbances, communication quality degradation, and equivalent errors caused by time-varying estimation. Defined as: in, Nominal control command. Raw perceptual error vector. Defined as: S6-4: Constructing a single-step state prediction model. The single-step prediction error of the augmented state is used as a monitoring signal. The single-step prediction is constructed using actual execution control increments that include residual compensation, based on parameters. Single-step prediction of the augmented state from time k to time k+1 Defined as: S6-5 establishes the loss function for the neural network. The single-step prediction error loss function is constructed using the squared L2 norm: in, For a moment Based on parameters The loss function; For a moment The observed actual augmented state; Let represent the squared 2-norm of a vector.

[0058] S6-6: An approximate gradient strategy is used for online parameter updates. When a constraint is activated, a zero-gradient approximation is applied to maintain update stability. The gradient update formula for the network parameters is: in, For the updated network parameters; The learning rate; Represents the loss function For parameters The gradient.

[0059] S6-7: Calculate the residual control correction term using a neural network. The residual learning module outputs the online correction term for the control command through a neural network. : in, Represents the residual learning model; These are the parameters of the residual learning model; To parameterize Neural network functions; This is a residual control correction term.

[0060] like Figure 8 As shown, the specific method for fusing and executing control commands through a resilient gating mechanism in step S7 includes the following steps: S7-1: Calculate the nominal control increment. Transfer the nominal control command... Convert to control increment form: in, To control the increment in name; Nominal control instructions; This is the actual control input from the previous moment.

[0061] S7-2: Incremental control is integrated through a resilient gating mechanism that is aware of communication quality.

[0062] in, This is the final control increment after fusion; The communication quality confidence level output in steps S1-6; This refers to the residual control correction term output in step S6-4; This is the projection operator.

[0063] S7-3: Calculate the final control command and ensure constraints are satisfied. The final control command for the current time step is obtained by superimposing the control input from the previous time step and fusing the control increment. in, This is the final control command; S7-4: Update the control history buffer.

[0064] Execute control commands Then, update the input history vector. For use in the next moment: in, The updated input history vector contains the most recent Step control input; The upper bound of the discrete delay steps defined in step S4-4. This history vector will be used in the next time step to construct the augmented state vector in step S4-5. .

[0065] The proposed method combines physical model-based model predictive control with communication-aware online residual learning. The physical model module provides an interpretable, constraint-aware nominal control skeleton, preserving the constraint handling capabilities and theoretical guarantees of MPC. The residual learning module compensates for model mismatch and unmodeled dynamics in real time through an online neural network. A resilient gating mechanism based on communication quality awareness is introduced. The V2V communication status is evaluated in real time using a communication quality metric based on information age. The fusion weights of onboard sensor information and V2V shared information are dynamically adjusted according to the communication quality, so that the contribution of the residual compensation term is explicitly adjusted by the communication quality confidence. When the communication quality is good, high-performance control is achieved by fully utilizing cooperative information. When the communication quality degrades, the control mode automatically and smoothly switches to a control mode that relies on local sensors and physical models, avoiding instability introduced by stale information and learning compensation.

[0066] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for intelligent connected vehicle platooning resilience control under mixed traffic flow, characterized in that... Includes the following steps: S1: Establish communication quality assessment and Markov channel model to evaluate V2V communication quality and characterize random delay; S2: Construct a hybrid platoon vehicle dynamics model, establish an HDV clustering behavior prediction model and CAV error dynamics equations; S3: Design a heterogeneous information fusion strategy for communication quality perception, dynamically fusing on-board sensor information and V2V shared information based on communication quality; S4: Establish a delay augmented state space model to incorporate communication delay into a fixed-dimensional state space representation; S5: Design and solve the model predictive control optimization problem to obtain the nominal control command that satisfies the constraints; S6: Design an online residual learning compensation module to compensate for model mismatch and unmodeled dynamics in real time; S7: Achieve formation resilience control by fusing and executing control commands through a resilience gating mechanism.

2. The intelligent connected vehicle platooning resilience control method under mixed traffic flow as described in claim 1, characterized in that, The method for establishing communication quality assessment and Markov channel model in S1 includes the following steps: S1-1, Establish a delayed evolution model based on information age, using information age (AoI) as the metric, denoted as... The discrete-time recursive evolution model of AoI is as follows: in, I k Indicates time k Whether the intelligent connected vehicle successfully received new information; when I k When =1, the intelligent connected vehicle successfully received new information. I k =0 The intelligent connected vehicle did not receive any information; D k The end-to-end delay for receiving new information; To control the sampling period; This represents the maximum communication latency that the system can tolerate. S1-2, Establishing a two-state Markov chain model to characterize channel state switching: A two-state Markov chain model is used to describe the channel state. ,in Indicates good channel condition. The channel state transition probability is defined as follows: ； ； in, To be in good condition Transfer to deteriorated state The transition probability, with a value range of . ; To recover from the deteriorated state Transfer to good condition The transition probability, with a value range of . ; S1-3, Construct a data packet reception model based on channel state: Given channel state (in ), data packet reception event Follows Bernoulli distribution: in, Channel state The probability of successfully receiving the next data packet; This represents the probability of successful reception under good channel conditions. Represents the probability of successful reception under degraded channel conditions, satisfying ; S1-4: Establish an end-to-end communication delay model based on channel state; When data packets are successfully received ( End-to-end communication delay The model is as follows: in, As a reference delay; The unit delay increment; Let be a random variable that follows a geometric distribution, with parameter . , The current channel state is decremented by 1 to allow for zero additional delay; The parameter is Geometric distribution; using parameters under good channel conditions Use smaller parameters in degraded channel conditions This is to generate a heavy-tailed delay distribution and reproduce the observed delay spike behavior; S1-5: Construct online communication quality metrics; Introducing online communication quality metrics Based on the above AoI evolution process and data packet reception process, The discrete-time implementation is as follows: in, As a sensitivity factor for information freshness, it characterizes the degree of impact of AoI on communication quality; The length of the sliding time window used to evaluate communication reliability; In the time window The expected number of nominal data packets to be received. The range of values ​​is When communication quality is good Approaching 1, when communication quality degrades Decrease.

3. The intelligent connected vehicle platooning resilience control method under mixed traffic flow as described in claim 1, characterized in that, The method for constructing a platooning vehicle dynamics model under mixed traffic flow in S2 includes the following steps: S2-1 defines the topology and communication relationships of intelligent connected vehicle platoons: Considering a mixed convoy of multiple vehicles, define a local formation unit containing Vehicles, including the Navigator CAV, numbered ; Follow-up The HDV and the allegedly involved CAV, numbered The vehicle index set is defined as follows: The assembly of all vehicles in the formation is denoted as The CAV set is HDV collection is CAV The vehicle acquires real-time status information of the vehicle ahead through onboard sensors and receives navigation CAV data via V2V communication. Status information; S2-2, Establish the longitudinal dynamics model of the platooned vehicles: The longitudinal dynamics of each vehicle in the formation satisfy the following equations: in, For vehicles At any moment The longitudinal position; For vehicles longitudinal velocity; For vehicles longitudinal acceleration; For vehicles With the vehicle immediately in front The distance between vehicles; For vehicles With the vehicle immediately in front The relative speed between them; S2-3, Establish the general form of error dynamics based on the desired spacing strategy: Based on the above definition, a vehicle With the car in front The general dynamic form of the spacing error and relative velocity error between them is: in, The expected headway; The desired stationary distance; Actual vehicle spacing Spacing from the expected distance The difference is defined as ; S2-4, Establish the dynamic model of HDV car-following behavior: The following dynamic model describes the car-following behavior of HDVs in formation: in, For HDV index, ; For the headway gain; This is the relative velocity gain; This is the desired speed function based on the vehicle spacing; For vehicle spacing variables; This represents the minimum spacing in a stationary state. The maximum distance threshold for maintaining maximum vehicle speed; The maximum expected speed; S2-5, Establish the dynamics of HDV spacing error and relative velocity error: According to the general form in steps S2-3, HDV With the car in front The dynamics of the spacing error and relative velocity error between them are as follows: in, , HDV Compared to the car in front The expected headway and expected stationary distance; , These are the spacing error and the relative speed error, respectively. S2-6, Establish the delay dynamics model of the CAV actuator; CAV The dynamics can be described as follows: in, The desired acceleration command; The actuator time constant characterizes the dynamics of the transmission system and the execution delay; The communication delay defined in step 1; For CAV With the car in front The distance between vehicles; For CAV i The acceleration; S2-7, Estimating macroscopic behavioral parameters of clustered HDVs based on traffic wave theory: Combining traffic wave theory with the Newell car-following model, macroscopic traffic wave speeds are used to... Estimated HDV With CAV Time-varying expected time intervals between and desired stationary distance : in, For CAV Compared to its immediate predecessor HDV The expected time between the locomotive headway; For CAV Compared to its immediate predecessor HDV The expected static distance between them; For HDV With CAV The actual distance between vehicles is defined as ; For HDV speed; This indicates that CAV takes into account the impact of traffic waves. Compared to CAV The time-varying time interval; S2-8, Establish CAV With Leading CAV Error dynamics equation between: CAV With CAV The error dynamics between them can be written as: in, For CAV Compared to CAV The spacing error is defined as ; For CAV With CAV The actual distance between vehicles; For CAV Compared to CAV The relative velocity is defined as ; The term represents the clustered disturbance, indicating the impact of the preceding vehicle's HDV. And Leading CAV The unpredictability and clustering disturbances caused by macroscopic estimation errors; To lead CAV The acceleration; S2-9, Establish CAV With the HDV in front Error dynamics equation between: CAV With HDV The error dynamics between them can be written as: in, For CAV Compared to the HDV in front Spacing error; For the HDV in front The acceleration.

4. The intelligent connected vehicle platooning resilience control method under mixed traffic flow as described in claim 1, characterized in that, The specific method for designing a heterogeneous information fusion strategy for communication quality awareness in S3 includes the following steps: S3-1, define the fusion spacing error and fusion speed error: Define the fusion spacing error and fusion speed error as follows: in, To meet the HDV in front Error fusion weights associated with information sources; To cooperate with CAV Error fusion weights associated with information sources; S3-2, Establish normalization constraints for the fusion weights: The fusion weights satisfy the normalization constraint: S3-3, Define the fusion ratio function for adaptive communication quality: Define the fusion ratio function : in, This is the fusion ratio function, with a range of values ​​of [value range missing]. ; To assign to CAV under fully trusted conditions The nominal weight upper bound satisfies ; The communication quality confidence level is given by the value range of . ; S3-4, Calculate the fusion weights of the preceding vehicle information source: Based on the normalized constraint sum and the fusion ratio function, the preceding vehicle HDV Information source fusion weight The calculation is as follows: in, The range of values ​​is When communication quality is good, As communication quality increases, the system relies more on V2V shared navigation CAV information; when communication quality degrades, The system automatically steers by relying more on information about the vehicle ahead obtained from onboard sensors.

5. The intelligent connected vehicle platooning resilience control method under mixed traffic flow as described in claim 1, characterized in that, The method for establishing the delayed augmented state-space model in S4 includes: S4-1, Define the low-dimensional state vector and perturbation vector: Define a low-dimensional state vector and perturbation vector as follows: in For lumped system disturbances; S4-2, Construct the continuous-time fusion error state dynamics equation: The continuous-time dynamic equation for the fusion error state is: in, For communication delay The control input; the model matrix is ​​represented as: , in, For time-varying system matrices; To control the input matrix; The input matrix is ​​a time-varying perturbation matrix; S4-3 uses the zeroth-order preservation assumption to discretize the model parameters: Within each sampling interval, , and Under the zeroth-order preservation assumption, these are considered piecewise constants, denoted as follows: , and ;in, For the first Each sampling time, For discrete time step index; S4-4, Discretization of continuous-time dynamics using the forward Euler method: The continuous-time dynamics are discretized using the forward Euler method, resulting in the discrete-time state-space equations: in, For the first Discrete state vectors at each time step; For delay Control input after the step; The number of discrete delay steps is determined by the communication delay. Divide by sampling period Rounded down to the nearest integer; The upper bound of the discrete delay steps is given by the maximum tolerable delay. Divide by Rounding up, the discrete-time model matrix is: , , in, For discrete system matrices; For discrete control input matrix; The input matrix is ​​a discrete perturbation matrix; S4-5, introduce delayed augmented state vectors to construct fixed-dimensional models: Introducing Delayed Augmented State Vector : in They are respectively k -1, k -2, k Time-based control input; Define delayed selection operator Define the input history vector Then the delayed input satisfies ; , in For the first One element is 1, and the rest are 0. dimensional standard basis vectors S4-6, Constructing a delay-free linear system representation: Define control increment We obtain the linear system representation with delayed freedom: The system matrix is ​​defined as follows: , , , in, To augment the system matrix; To augment the control increment input matrix; For augmented perturbation input matrix; Input the historical update matrix; express A zero matrix with 3 rows and 3 columns; Represents a 3x1 zero vector; S4-7, Construction Stack state prediction model with step prediction range: By predicting the range Freeze within step and Construct the stacked prediction formula; define the stacked prediction state vector. Stacked control incremental vector and stacked perturbation vector : , , in, Indicates at time Based on current information, time Augmented state prediction; Indicates at time For time Control increment; Indicates at time For time The disturbance prediction; Number of steps for the prediction range; The standard linear state prediction model is obtained: The model matrix is ​​defined as follows: , ， in, This is the state transition matrix; To control the incremental influence matrix; This is the disturbance effect matrix.

6. The intelligent connected vehicle platooning resilience control method under mixed traffic flow as described in claim 1, characterized in that, The method for designing and solving the model predictive control optimization problem in S5 includes the following steps: S5-1, Construct the MPC multi-objective cost function: MPC optimizes the control law by predicting system behavior and imposing constraints within a finite predictability range. Only the first control action is applied at a time, and this process is repeated at each time step; the predictability range is... The step-by-step MPC optimization problem is expressed as: in, For a moment The total cost function; For at any time Based on current information, time state vector The prediction; For the prediction step index; Number of steps for the prediction range; This is the state weight matrix; To control the incremental weight scalar; This is the terminal state weight matrix; weight matrix , and Defined as: in These represent the weighting coefficients related to spacing error, velocity error, and acceleration, respectively. This indicates an increase in the amount of punishment controlled; These represent the terminal weighting coefficients for spacing error, velocity error, and acceleration, respectively. S5-2, Constructing constraints for the MPC optimization problem: Define control increment constraints, control increment Upper and lower bound constraints must be satisfied: in, To control the lower bound of the increment; To control the upper bound of the increment; Define control input constraints, control input The control increment is obtained by accumulating control increments, and must satisfy upper and lower bound constraints: in, To control the lower bound of the input; To control the upper limit of the input; This is the control input from the previous moment; Define state vector constraints and predict state vectors. It is necessary to start from the augmented state vector Extract from, and satisfy upper and lower bound constraints: in, This is the lower bound of the state vector; The upper bound of the state vector; the state selection matrix. ,in It is a 3-order identity matrix. 3 lines The zero matrix of columns; S5-3 refactors the MPC optimization problem into a standard quadratic programming form: Using stacked state prediction models The MPC optimization problem is refactored into a standard QP form: st in This is the optimal control input correction value; This is the coefficient matrix of the quadratic terms; This is the vector of linear term coefficients; , These are the state weight diagonal matrix and the penalty control diagonal matrix, respectively. Choose a diagonal matrix for the state; This is the inequality constraint coefficient matrix; The corresponding inequality boundary vectors; for N 3D identity matrix; It is a length of N A column vector of all 1s; It is a standard lower triangular cumulative matrix; To control the selection matrix; These are the upper and lower bounds of the superposition of state vectors, respectively; S5-4: Extracting the optimal control increment sequence: The first element is used as the nominal control increment at the current moment: Here, [·]1 represents the extraction of the first element of the vector.

7. The intelligent connected vehicle platooning resilience control method under mixed traffic flow as described in claim 1, characterized in that, The specific method for designing the online residual learning compensation module in S6 includes: S6-1, Construct the input set for the physical controller: At time step Controller input set based on physical model Defined as: in for The predicted system disturbance; for CAV With CAV The expected static distance between them; S6-2, characterizing the input-output mapping of the physical controller: The physical model-based controller calculates nominal control commands online using QP-MPC. The mapping relationship is expressed as follows: in, Represents a physical model of a mixed formation; For parameters of the hybrid formation physical model; This represents the defined MPC optimization solution process; Represents a linear control law mapping; S6-3, Construct the input set for the residual learning module: Input set of the residual learning module Defined as: in, Nominal control instructions; raw sensing error vector Defined as: S6-4, Construct a single-step state prediction model; The single-step prediction error of the augmented state is used as the monitoring signal; the single-step prediction is constructed using the actual execution control increment including residual compensation, based on parameters. Single-step prediction of the augmented state from time k to time k+1 Defined as: S6-5, Establish the loss function for the neural network: The single-step prediction error loss function is constructed using the squared norm of the second law: in, For a moment Based on parameters The loss function; For a moment The observed actual augmented state; The squared norm of a vector is denoted by 2. S6-6: Online parameter update using an approximate gradient strategy; When constraints are activated, a zero gradient approximation is applied to maintain update stability; the gradient update formula for network parameters is: in, For the updated network parameters; The learning rate; Represents the loss function For parameters The gradient; S6-7, Calculate the residual control correction term using a neural network: The residual learning module outputs online correction terms for control commands through a neural network. : in, Represents the residual learning model; These are the parameters of the residual learning model; To parameterize Neural network functions; This is a residual control correction term.

8. The intelligent connected vehicle platooning resilience control method under mixed traffic flow as described in claim 1, characterized in that, The method for fusing and executing control commands through a resilient gating mechanism in S7 includes the following steps: S7-1, Calculate the nominal control increment: nominal control command Convert to control increment form: in, To control the increment in name; Nominal control instructions; This is the actual control input from the previous moment; S7-2 integrates incremental control through a resilient gating mechanism that is aware of communication quality. in, This is the final control increment after fusion; For communication quality confidence; This is a residual control correction term; For projection operators; S7-3, Calculate the final control command and ensure that constraints are satisfied; The final control command for the current moment is obtained by superimposing the control input from the previous moment and fusing the control increment: in, This is the final control command; S7-4, Update control history buffer: Execute control commands Then, update the input history vector. For use in the next moment: in, The updated input history vector contains the most recent Step control input; This is an upper bound on the discrete delay steps; this history vector will be used to construct the augmented state vector at the next time step. .

Images