Vehicle-road cooperative control architecture system based on data-mechanism coupled modeling, and construction method

Abstract

Disclosed are a vehicle-road cooperative control architecture system based on data-mechanism coupled modeling, and a construction method. To address the challenges of traditional mechanism-based modeling for autonomous driving, a multi-agent system modeling method driven by the fusion of data and mechanisms, and a federated reinforcement learning-based vehicle-road cooperative swarm optimization method are provided, establishing a vehicle decision-making model parameter updating technique based on multi-dimensional experience sharing, and thereby resolving interpretability and generalization challenges of purely data-driven models. A rule-based driving safety field is built to achieve rule-guided data-driven training; a secondary planning-control framework based on an intelligent chassis is constructed, providing state quantity input based on chassis feedback, and resolving problems related to purely data-driven models such as questionable reliability, reliance on large-scale data, and a lack of transparency and interpretability in the decision-making process; quantitative comfort metrics are constructed to filter for an optimal strategy for the current environment, and a balance between sample efficiency and model robustness is achieved by synthesizing shared models benefiting from different environments.

Description

Vehicle-Road Cooperative Control Architecture System and Construction Method Based on Data Mechanism Coupling Modeling Technical Field

[0001] This invention belongs to the field of transportation and relates to a vehicle-road cooperative control architecture system and its construction method based on data mechanism coupling modeling. Background Technology

[0002] In the field of autonomous driving, existing systems are massive cyber-physical systems. Achieving high-level autonomous driving through traditional mechanistic modeling methods is extremely difficult. The shortcomings are mainly reflected in limited environmental perception, difficulties in interaction modeling and scene understanding, inefficient decision-making and planning, and difficulty in ensuring driving safety, which directly limits the application of autonomous driving technology in complex traffic situations.

[0003] In application, the main challenges of autonomous driving algorithms can be summarized as the trust problem and the interpretability problem. The trust problem manifests primarily in the fact that rule-based methods heavily rely on manually configured rules, making it difficult to handle complex driving environments. Conversely, end-to-end data-driven methods heavily depend on training samples; insufficient training with long-tail cases can lead to serious decision-making errors, raising concerns about the algorithm's reliability in application. The interpretability problem stems mainly from the opaque structure of black-box algorithms, resulting in insufficient interpretability of data-driven methods, making it impossible to explain the control process and decision-making basis. Furthermore, the lack of explanation for control algorithms directly impacts their evaluation and improvement, further reducing their reliability.

[0004] Introducing federated learning to address the trustworthiness problem of control algorithms is a novel idea. By transferring and aggregating model parameters, it improves the transferability and robustness of algorithms, thereby alleviating the trust crisis in control algorithms. However, the parameter aggregation process in federated learning is also difficult to interpret, which negates the benefits it brings. In summary, existing autonomous driving algorithms face challenges in balancing algorithm trustworthiness and interpretability.

[0005] Summary of the Invention

[0006] To address the aforementioned technical challenges, this invention provides a vehicle-road cooperative control architecture based on data-mechanism coupling modeling. It proposes a multi-agent system modeling method driven by data and mechanism fusion, a vehicle-road cooperative group optimization method based on federated reinforcement learning, and establishes a vehicle decision-making model parameter update technology based on multi-dimensional experience sharing. This solves the interpretability and generalization problems of purely data-driven models. Leveraging roadside advantages, a rule-based driving safety field is built, enabling rule-guided data-driven training. Addressing the difficulty of traditional mechanism modeling in autonomous driving, a data-mechanism coupling driven model is proposed, constructing a quadratic programming control framework based on an intelligent chassis. It innovatively proposes state input based on chassis feedback, resolving issues such as questionable reliability of purely data-driven models, reliance on large-scale data, and opaque and uninterpretable decision-making processes. Based on the sensitive range of human perception, weighted lateral acceleration, yaw rate acceleration, and longitudinal acceleration are introduced to construct a comfort quantification index to screen locally optimal strategies for the current environment. By synthesizing a globally shared model that benefits from different environments, a balance between sample efficiency and model robustness is achieved.

[0007] The present invention provides a technical solution for a vehicle-road cooperative control architecture based on data mechanism coupling modeling, comprising three parts: rule-guided coupling modeling, data mechanism coupling planning and control, and data mechanism coupling evaluation.

[0008] The rule-guided coupled modeling part mainly involves processing bird's-eye view information at the roadside to build a driving safety field, and establishing a fusion reward function to guide the training process of the data-driven control algorithm. The coupled modeling part includes three components: a roadside information processing module, a safety field modeling module, and a reward function modeling module.

[0009] The roadside information processing module primarily leverages the roadside's superior field of view to transform the bird's-eye view image into a semantic bird's-eye view. It then uses the dynamic information within this semantic bird's-eye view to model the interactions between intelligent connected vehicles on the roadside. Finally, it transmits the safety field information and the semantic bird's-eye view to the vehicle via V2I communication. The semantic bird's-eye view is a four-channel matrix, containing static road and lane information, as well as dynamic desired path and vehicle information.

[0010] The safety field modeling module is established using the following equation: φ=a x / b y =l v / w v

[0011] Among them, S sta C represents the static safety field strength. aThe static safety field strength coefficient is represented by x0 and y0, which represent the coordinates of the static risk center O(x0, y0), ε represents the safety field shape coefficient, and a represents the static safety field strength coefficient. x and b y This represents the appearance coefficient of the intelligent connected vehicle, where φ is the aspect ratio of the intelligent connected vehicle, and l v Indicates the length of the vehicle, w v Indicates the width of the vehicle.

[0012] When the intelligent connected vehicle moves, the risk center O(x0,y0) of the Gaussian safety field will shift to a new risk center O′(x′0,y′0) as the vehicle moves:

[0013] Where, k v Indicates the moving adjustment factor, and The sign is related to the direction of motion; β represents the transfer vector of the connected vehicle. The angle between the coordinate axes and the coordinate system in the Cartesian coordinate system. This represents the velocity vector of the connected vehicle. A virtual vehicle of length l′ is formed in the dynamic safety field under the influence of risk center transfer. v Width is w′ v S dyn The dynamic safety field strength is represented by the new aspect ratio φ′=a′. x / b′ y =l′ v / w′ v .

[0014] The reward function modeling module considers both driving expectation and driving safety: r = rexpectation + r safety

[0015] Where r represents the reward function, rexpectation represents the reward function related to driving expectations, and r safety This represents a reward function related to driving safety.

[0016] The reward function related to driving expectations includes both lateral and longitudinal aspects. Firstly, regarding lateral driving expectations: d0 = min(||a 28,28 -bdesired route|| 2 r1 lateral =-log 1.1 (|d0|+1) r2 lateral =-10*|sin(radians(θ))| r lateral =r1 lateral +r2 lateral

[0017] Where d0 represents the distance from the center line of the lane, a 28,28 R1 represents the vehicle center, bdesired route represents the desired route, and r1 represents the desired route. lateral R2 represents the lateral distance reward function. lateral Let r represent the heading angle reward function, θ represent the heading angle deviation at the vehicle end, and r represent the heading angle reward function. lateral This represents the expected reward function for lateral driving.

[0018] Secondly, there are longitudinal driving expectations: d min =min(||a 28,28 -b x,y || 2 ) r2longitudinal=-|v ego –9| rlongitudinal=r1longitudinal+r2longitudinal

[0019] Where, d min b represents the minimum distance between autonomous vehicles. x,y The center of the other vehicle is represented by x, the time of collision is represented by v. ego Let r1 represent the vehicle speed, r2 represent the distance reward function, r2 represent the speed reward function, and rlongitudinal represent the longitudinal driving expected reward function.

[0020] Finally, the expected reward function for driving is established using the following formula: rexpectation = r lateral +rlongitudinal

[0021] The reward function related to driving safety is calculated by the roadside driving safety field from two aspects: driving safety and driving aggression. First, driving safety:

[0022] Among them, R i,j (t) represents the driving risk posed by intelligent connected vehicle j to intelligent connected vehicle i. Let k represent the field strength of intelligent connected vehicle j with respect to intelligent connected vehicle i. c This represents the risk perception coefficient. Let θ represent the speed of the intelligent connected vehicle j at time t. i,j (t) represents the angle between intelligent connected vehicle i and intelligent connected vehicle j at time t, r Risk Let f represent the reward function related to driving risk. Risk (ξ) represents the driving risk score, R thrτ represents the risk threshold. rc This indicates the duration exceeding the risk threshold.

[0023] Among them, R j,i (t′) represents the driving risk posed by intelligent connected vehicle i to intelligent connected vehicle j. Let i represent the field strength of intelligent connected vehicle i for intelligent connected vehicle j. Let θ represent the speed of the intelligent connected vehicle i at time t′. j,i (t′) represents the angle between intelligent connected vehicle j and intelligent connected vehicle i at time t′, r Agg Let f represent the reward function related to driving aggression. Agg (ξ) represents the integral of driving aggression.

[0024] The driving safety-related reward function is established using the following formula: r Safety =r Risk +r Agg

[0025] The data mechanism coupled planning and control part mainly involves constructing a reinforcement learning and intelligent chassis secondary planning and control architecture based on chassis feedback. This is achieved through real-time dynamic coupling of the vehicle chassis mechanism model and the reinforcement learning algorithm. The coupled planning and control part comprises three components: a vehicle-side information processing module, a reinforcement learning primary planning and control module, and an intelligent chassis secondary planning and control module.

[0026] The vehicle-side information processing module primarily involves each intelligent connected vehicle acquiring semantic bird's-eye view information from the roadside via V2I communication, and cropping the global semantic bird's-eye view based on its own vehicle position sensor information. The cropped semantic bird's-eye view, combined with sensor information and the results of the intelligent chassis secondary planning control, consists of three parts of state variables that are stacked over two consecutive frames before being used as input to the reinforcement learning primary planning control module.

[0027] The reinforcement learning-based primary planning control module primarily uses the state variables output by the vehicle-side information processing modules of each intelligent connected vehicle as input to the reinforcement learning neural network, and outputs the corresponding primary planning control result. The primary planning control process includes steering wheel control, throttle control, and brake control, where: steering∈[-1,1], throttle∈[0,1], brake∈[0,1].

[0028] Where steering represents steering wheel control, throttle represents accelerator control, and brake represents brake control, the action space of this invention is ∈ [-1,1]. 2Two types of motion control are used: steering wheel control and accelerator-brake control. For accelerator-brake control, [-1,0] represents brake control, and [0,1] represents accelerator control. This invention uses a Beta distribution as the output of the reinforcement learning linear programming control: Beta=B(α,β)

[0029] Here, α and β represent two parameters of the Beta distribution. This invention further samples the corresponding action control variables from the Beta distribution. The Beta distribution is used in this invention because it has several advantages over the Gaussian distribution commonly used in model-free reinforcement learning. One advantage is its ability to dynamically simulate sample distributions of various shapes. Furthermore, unlike the Gaussian distribution, which extends to infinity in both positive and negative directions, the Beta distribution has bounded support from 0 to 1 and does not require mandatory constraints. In summary, the Beta distribution provides a flexible, variable, and bounded method for simulating various sample distributions and handling bounded variables, making it a superior choice to the Gaussian distribution.

[0030] The intelligent chassis secondary planning control module mainly performs secondary planning control based on the output of the reinforcement learning primary planning control module and the desired path, using the results of the secondary planning control. The output of the secondary planning control directly controls the torque and steering angle of the vehicle's four wheels, and the control variables are then transmitted as part of the state variables to the reinforcement learning primary planning control processes of each intelligent connected vehicle. The secondary planning control process establishes the coordination and cooperation problem between intelligent chassis systems as an optimization problem considering global performance indicators, described by the following equation:

[0031] Among them, U i (t) represents the control quantity of intelligent subsystem i at time t, J i J i This represents the total cost function of intelligent subsystems i and j. This represents the predicted state of intelligent subsystem i at a future time. This represents the control quantity of intelligent subsystem i at a future time. This represents the assumed state of the neighbor intelligent subsystem j. Let λ represent the hypothetical control input of the neighbor intelligent subsystem j. i This represents the coupling coefficient of the cost function between intelligent subsystems. X i Let W represent the predicted state of intelligent subsystem i, m represent the number of intelligent subsystems, and W represent the predicted state of intelligent subsystem i. i U represents the reference state sequence of intelligent subsystem i. i This represents the sequence of control inputs for intelligent subsystem i. Represents the sequence of state weight coefficients. This represents the sequence of control weight coefficients for intelligent subsystem i.

[0032] Where, x i (k) represents the state of the intelligent subsystem i at time k in the future, X k U represents the state sequence of the intelligent subsystem at time k in the future. k This represents the sequence of control variables for intelligent subsystem j at time k in the future. F represents the transpose matrix of the control input sequence of intelligent subsystem i, where const represents a constant. ii G ii F ij G ij F jj G jj F jk G jk F represents the computation matrix. ij G ij The structure is as follows:

[0033] in, Represents the state parameters of sub-agent i. The coupling state parameters represent those between sub-agents i and j. N represents the coupling state parameter. p N represents the prediction time domain. c This represents the control time domain; other computational matrices are similar. This can be represented by the following three formulas:

[0034] Among them, Q i R represents the state weight coefficient of intelligent subsystem i. i This represents the control weight coefficient of intelligent subsystem i. The transpose of the state sequence of intelligent subsystem j at future time k is given by the following matrix. Q represents the transpose matrix of the reference state sequence of the intelligent subsystem j. j This represents the state weight coefficient of the intelligent subsystem j. This represents the transpose of the control sequence of intelligent subsystem j at future time k. G represents the transpose of the future k-time state sequence of the intelligent subsystem j. ji This represents a computation matrix, with the structure described above.

[0035] The data mechanism coupling assessment section mainly involves building a quantitative index for comfort, and achieving coupled assessment and group optimization through a neural network screening mechanism based on the vehicle chassis mechanism model. The coupled assessment section comprises three parts: a comfort index modeling module, a neural network screening module, and a neural network parameter aggregation module.

[0036] The comfort index modeling module mainly uses the three state quantities that are most noticeable to the human body—lateral acceleration, yaw rate acceleration, and longitudinal acceleration—as measurement indicators based on the degree of state change when the vehicle is in motion. It then plots the sensitive intervals of human perception and uses the weighted sum of squares of the three state quantities as the quantitative indicator.

[0037] in, S represents a quantitative indicator of human comfort based on a smart chassis. i (t) represents the lateral acceleration, yaw acceleration, and longitudinal acceleration at time t, ω i The weighted parameters i∈[0,m] represent the lateral acceleration, yaw acceleration, and longitudinal acceleration, respectively.

[0038] The neural network screening module primarily uses the constructed human comfort quantification index based on the intelligent chassis as the screening criterion to select the reinforcement learning neural network parameters corresponding to the intelligent connected vehicle that provides the best perceived human comfort. The screening process is as follows:

[0039] in, This indicates that at time t, the parameters φ of the intelligent connected vehicle itself... t,i And another intelligent connected vehicle parameter φ t,i′ New network parameters obtained through aggregation.

[0040] The neural network parameter aggregation module primarily acquires the neural network parameters of the intelligent connected vehicle through roadside V2I communication, calculates shared neural network parameters through parameter averaging, and finally sends these parameters to the vehicle via V2I communication to achieve experience sharing until the network converges. The parameter averaging is performed according to the following equation:

[0041] in, φ′ represents the shared neural network parameters at time m, N represents the number of intelligent connected vehicles, and φ′ represents the number of intelligent connected vehicles. m,i Let m represent the neural network parameters of the i-th intelligent connected vehicle at time m.

[0042] The technical solution for building a vehicle-road cooperative control architecture based on data mechanism coupling modeling in this invention includes the following steps:

[0043] Step 1: Rule-guided coupled modeling. First, information processing is performed at the roadside, leveraging the roadside's field of view advantage to transform the bird's-eye view image into a semantic bird's-eye view. Dynamic information from this semantic bird's-eye view is then used to model the interactions between intelligent connected vehicles at the roadside. Finally, safety field information and the semantic bird's-eye view are transmitted to the vehicle via V2I communication, where a safety field is established to model the intelligent connected vehicle interaction process from the roadside perspective. Finally, reward function modeling is performed considering both driving expectations and driving safety.

[0044] Step 2: Perform data mechanism coupled planning and control. A reinforcement learning and intelligent chassis secondary planning and control architecture based on chassis feedback is constructed. Data mechanism coupled planning and control is achieved through real-time dynamic coupling between the intelligent chassis system and the reinforcement learning algorithm. The coupled planning and control process includes three parts: vehicle-side information processing, reinforcement learning primary planning and control, and intelligent chassis secondary planning and control. First, vehicle-side information processing is performed. Each intelligent connected vehicle obtains semantic bird's-eye view information from the roadside via V2I communication and crops the global semantic bird's-eye view based on its own vehicle position sensor information. The cropped semantic bird's-eye view, combined with sensor information and the intelligent chassis secondary planning and control results, is stacked over two consecutive frames and used as the input for reinforcement learning primary planning and control. Then, primary planning and control is performed through reinforcement learning. The state variables output from the information processing of each intelligent connected vehicle are used as the input to the reinforcement learning neural network, outputting the corresponding primary planning and control result. Finally, secondary planning control of the intelligent chassis is performed. Based on the output of the primary planning control of reinforcement learning and the desired path, secondary planning control based on the intelligent chassis multi-subsystem is performed. The output of the secondary planning control will directly control the torque and steering angle of the four wheels of the vehicle, and the control quantity will be transmitted again as part of the state quantity to the primary planning control process of each intelligent connected vehicle.

[0045] Step 3: Conduct data mechanism coupling evaluation. Establish a quantitative comfort index and achieve coupling evaluation through a neural network screening mechanism based on the vehicle chassis mechanism model. Use the quantitative comfort index as the basis to quantitatively evaluate the control effect of intelligent connected vehicles.

[0046] Step 4: Neural Network Screening. Using the established quantitative indicators of human comfort based on the intelligent chassis as the screening criteria, the reinforcement learning neural network parameters corresponding to the intelligent connected vehicle that best achieves human-perceived comfort are selected.

[0047] Step 5: Perform neural network parameter aggregation. This mainly involves acquiring the neural network parameters of the intelligent connected vehicle through roadside V2I communication, calculating shared neural network parameters through parameter averaging, and finally distributing them to the vehicle via V2I communication to achieve experience sharing until the network converges.

[0048] Preferably, in step 1, the reward function modeling considers both driving expectation and driving safety: r = rexpectation + r safety

[0049] Where r represents the reward function, rexpectation represents the reward function related to driving expectations, and r safety This represents a reward function related to driving safety.

[0050] The reward function related to driving expectations includes both lateral and longitudinal aspects. Firstly, regarding lateral driving expectations: d0 = min(||a 28,28 -bdesired route|| 2 r1 lateral =-log 1.1 (|d0|+1) r2 lateral =-10*|sin(radians(θ))| r lateral =r1 lateral +r2 lateral

[0051] Where d0 represents the distance from the center line of the lane, a 28,28 R1 represents the vehicle center, bdesired route represents the desired route, and r1 represents the desired route. lateral R2 represents the lateral distance reward function. lateral Let r represent the heading angle reward function, θ represent the heading angle deviation at the vehicle end, and r represent the heading angle reward function. lateral This represents the expected reward function for lateral driving.

[0052] Secondly, there are longitudinal driving expectations: d min =min(||a 28,28 -b x,y || 2 ) r2longitudinal=-|v ego –9| rlongitudinal=r1longitudinal+r2longitudinal

[0053] Where, d min b represents the minimum distance between autonomous vehicles. x,y The center of the other vehicle is represented by x, the time of collision is represented by v. egoLet r represent the vehicle speed, r1longitudinal represent the distance reward function, r2longitudinal represent the speed reward function, and rlongitudinal represent the longitudinal driving expectation reward function. Finally, the driving expectation-related reward function is established using the following formula: rexpectation = r lateral +rlongitudinal

[0054] The reward function related to driving safety is calculated by the roadside driving safety field from two aspects: driving safety and driving aggression. First, driving safety:

[0055] Among them, R i,j (t) represents driving risk, k represents the electric field strength between intelligent connected vehicle i and intelligent connected vehicle j. c This represents the risk perception coefficient. Let θ represent the speed of the intelligent connected vehicle j at time t. i,j (t) represents the angle between intelligent connected vehicle i and intelligent connected vehicle j at time t, r Risk Let f represent the reward function related to driving risk. Risk (ξ) represents the driving risk score, R thr τ represents the risk threshold. rc This indicates the duration exceeding the risk threshold.

[0056] Among them, R j,i (t′) represents the driving risk. This represents the field strength between intelligent connected vehicle j and intelligent connected vehicle i. Let θ represent the speed of the intelligent connected vehicle i at time t′. j,i (t′) represents the angle between intelligent connected vehicle j and intelligent connected vehicle i at time t′, r Agg Let f represent the reward function related to driving aggression. Agg (ξ) represents the driving aggression score. The driving safety-related reward function is established using the following formula: r Safety =r Risk +r Agg

[0057] Preferably, in step 2, the single-stage planning and control process includes steering wheel control, throttle control, and brake control, wherein: steering∈[-1,1] throttle∈[0,1] brake∈[0,1]

[0058] Where steering represents steering wheel control, throttle represents accelerator control, and brake represents brake control, the action space of this invention is ∈ [-1,1]. 2 Two types of motion control are used: steering wheel control and accelerator-brake control. For accelerator-brake control, [-1,0] represents brake control, and [0,1] represents accelerator control. This invention uses a Beta distribution as the output of the reinforcement learning linear programming control: Beta=B(α,β)

[0059] Here, α and β represent two parameters of the Beta distribution. This invention further samples the corresponding action control variables from the Beta distribution. The Beta distribution is used in this invention because it has several advantages over the Gaussian distribution commonly used in model-free reinforcement learning. One advantage is its ability to dynamically simulate sample distributions of various shapes. Furthermore, unlike the Gaussian distribution, which extends to infinity in both positive and negative directions, the Beta distribution has bounded support from 0 to 1 and does not require mandatory constraints. In summary, the Beta distribution provides a flexible, variable, and bounded method for simulating various sample distributions and handling bounded variables, making it a superior choice to the Gaussian distribution.

[0060] Preferably, in step 2, the intelligent chassis secondary planning control process establishes the coordination and cooperation problem between intelligent chassis systems as an optimization problem considering global performance indicators, described by the following equation:

[0061] Among them, U i (t) represents the control quantity of intelligent subsystem i at time t, J i The overall cost function of the intelligent subsystem is represented by the following: This represents the predicted state of intelligent subsystem i at a future time. This represents the control quantity of intelligent subsystem i at a future time. This represents the assumed state of the neighbor intelligent subsystem j. Let λ represent the hypothetical control input of the neighbor intelligent subsystem j. i This represents the coupling coefficient of the cost function between intelligent subsystems. X i W represents the predicted state of intelligent subsystem i. i U represents the reference state sequence of intelligent subsystem i. i This represents the sequence of control inputs for intelligent subsystem i. Represents the sequence of state weight coefficients. This represents the sequence of control weight coefficients for intelligent subsystem i.

[0062] Among them, Fii G ii F ij G ij F jj G jj F jk G jk They represent the computational matrices, x and x respectively. i (k) represents the state of the intelligent subsystem i at time k in the future, X k U represents the state sequence of the intelligent subsystem at time k in the future. k This represents the sequence of control variables for intelligent subsystem j at time k in the future. represents the transpose matrix of the control input sequence of intelligent subsystem i, where const represents a constant. This can be represented by the following three formulas:

[0063] Among them, Q i R represents the state weight coefficient of intelligent subsystem i. i This represents the control weight coefficient of intelligent subsystem i. The transpose of the state sequence of intelligent subsystem j at future time k is given by the following matrix. Q represents the transpose matrix of the reference state sequence of the intelligent subsystem j. j This represents the state weight coefficient of the intelligent subsystem j. This represents the transpose of the control sequence of intelligent subsystem j at future time k. Let be the transpose of the state sequence of intelligent subsystem j at future time k.

[0064] Preferably, in step 3, the comfort index modeling mainly involves using the three state variables most noticeable to the human body—lateral acceleration, yaw acceleration, and longitudinal acceleration—based on the degree of state change during vehicle operation, and plotting the sensitive intervals of human perception accordingly. The weighted sum of squares of the three state variables is then used as the quantitative indicator.

[0065] in, S represents a quantitative indicator of human comfort based on a smart chassis. i (t) represents the lateral acceleration, yaw acceleration, and longitudinal acceleration at time t, ω i The weighted parameters i∈[0,m] represent the lateral acceleration, yaw acceleration, and longitudinal acceleration, respectively.

[0066] Preferably, in step 4, the screening process includes:

[0067] in, This indicates that at time t, the parameters φ of the intelligent connected vehicle itself... t,i And another intelligent connected vehicle parameter φ t,i′ New network parameters obtained through aggregation.

[0068] Preferably, in step 5, the parameters are averaged according to the following equation:

[0069] in, φ′ represents the shared neural network parameters at time m, N represents the number of intelligent connected vehicles, and φ′ represents the number of intelligent connected vehicles. m,i Let m represent the neural network parameters of the i-th intelligent connected vehicle at time m.

[0070] The beneficial effects of this invention are:

[0071] (1) A multi-agent system modeling method driven by data and mechanism fusion was proposed, a vehicle-road cooperative group optimization method based on federated reinforcement learning was proposed, and a vehicle decision model parameter update technology based on multi-dimensional experience sharing was established, which solved the interpretability and generalization problems of pure data-driven models.

[0072] (2) By leveraging the advantages of the roadside, a rule-based driving safety field was built, realizing data-driven training guided by rules. To address the difficulty of modeling traditional autonomous driving mechanisms, a data mechanism coupled driving model was proposed, and a secondary planning control framework based on the intelligent chassis was constructed. A state input based on chassis feedback was innovatively proposed, solving problems such as doubts about the credibility of pure data-driven systems, reliance on large-scale data, and opaque and unexplainable decision-making processes. Based on the sensitive range of human perception, lateral acceleration, yaw angle acceleration, and longitudinal acceleration were weighted to construct a comfort quantification index to screen locally optimal strategies for the current environment. By synthesizing a globally shared model that benefits from different environments, a balance between sample efficiency and model robustness was achieved. Attached Figure Description

[0073] Figure 1. Schematic diagram of the rule-guided coupling modeling proposed in this invention;

[0074] Figure 2. Schematic diagram of data mechanism coupled planning and control proposed in this invention;

[0075] Figure 3. Schematic diagram of the data mechanism coupling evaluation proposed in this invention; Detailed Implementation

[0076] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings, but the content of the present invention is not limited thereto.

[0077] This invention provides a vehicle-road cooperative control architecture based on data mechanism coupling modeling, specifically including the following steps:

[0078] (1) Rule-guided coupling modeling is performed, as shown in Figure 1. First, information processing is performed at the roadside, utilizing the roadside's field of view advantage to transform the bird's-eye view image into a semantic bird's-eye view. Then, dynamic information from the semantic bird's-eye view is used to model the interaction between intelligent connected vehicles at the roadside. Finally, safety field information and the semantic bird's-eye view are transmitted to the vehicle via V2I communication. Then, a safety field is established at the vehicle, and the interaction process of intelligent connected vehicles is modeled from the roadside perspective, established through the following equations: φ=a x / b y =l v / w v

[0079] Among them, S sta C represents the static safety field strength. a The static safety field strength coefficient is represented by x0 and y0, which represent the coordinates of the static risk center O(x0, y0). x and b y This represents the appearance coefficient of the intelligent connected vehicle, where φ is the aspect ratio of the intelligent connected vehicle, and l v Indicates the length of the vehicle, w v Indicates the width of the vehicle.

[0080] When the intelligent connected vehicle moves, the risk center O(x0,y0) of the Gaussian safety field will shift to a new risk center O′(x′0,y′0) as the vehicle moves:

[0081] Where, k v Indicates the moving adjustment factor, and The sign is related to the direction of motion; β represents the transfer vector of the connected vehicle. The angle between the vehicle and the coordinate axes in the Cartesian coordinate system. A virtual vehicle of length l′ is formed in the dynamic safety field under the influence of risk center transfer. v Width is w′ v S dyn The dynamic safety field strength is represented by the new aspect ratio φ′=a′. x / b′ y =l′ v / w′ v .

[0082] Finally, the reward function is modeled from the perspectives of driving expectation and driving safety: r = rexpectation + r safety

[0083] Where r represents the reward function, rexpectation represents the reward function related to driving expectations, and r safety This represents the reward function related to driving safety. The reward function related to driving expectation includes both lateral and longitudinal aspects. First, the lateral driving expectation: d0 = min(||a 28,28 -bdesired route|| 2 r1 lateral =-log 1.1 (|d0|+1) r2 lateral =-10*|sin(radians(θ))| r lateral =r1 lateral +r2 lateral

[0084] Where d0 represents the distance from the center line of the lane, a 28,28 R1 represents the vehicle center, bdesired route represents the desired route, and r1 represents the desired route. lateral R2 represents the lateral distance reward function. lateral Let r represent the heading angle reward function, θ represent the heading angle deviation at the vehicle end, and r represent the heading angle reward function. lateral This represents the expected reward function for lateral driving.

[0085] Secondly, there are longitudinal driving expectations: d min =min(||a 28,28 -b x,y || 2 ) r2longitudinal=-|v ego –9| rlongitudinal=r1longitudinal+r2longitudinal

[0086] Where, d min b represents the minimum distance between autonomous vehicles. x,y The center of the other vehicle is represented by x, the time of collision is represented by v. ego Let r represent the vehicle speed, r1longitudinal represent the distance reward function, r2longitudinal represent the speed reward function, and rlongitudinal represent the longitudinal driving expectation reward function. Finally, the driving expectation-related reward function is established using the following formula: rexpectation = r lateral +rlongitudinal

[0087] The reward function related to driving safety is calculated by the roadside driving safety field from two aspects: driving safety and driving aggression. First, driving safety:

[0088] Among them, R i,j (t) represents the driving risk posed by intelligent connected vehicle j to intelligent connected vehicle i. Let k represent the field strength of intelligent connected vehicle j with respect to intelligent connected vehicle i. c This represents the risk perception coefficient. Let θ represent the speed of the intelligent connected vehicle j at time t. i,j (t) represents the angle between intelligent connected vehicle i and intelligent connected vehicle j at time t, r Risk Let f represent the reward function related to driving risk. Risk (ξ) represents the driving risk score, R thr τ represents the risk threshold. rc This indicates the duration exceeding the risk threshold.

[0089] Among them, R j,i (t′) represents the driving risk posed by intelligent connected vehicle i to intelligent connected vehicle j. Let i represent the field strength of intelligent connected vehicle i for intelligent connected vehicle j. Let θ represent the speed of the intelligent connected vehicle i at time t′. j,i (t′) represents the angle between intelligent connected vehicle j and intelligent connected vehicle i at time t′, r Agg Let f represent the reward function related to driving aggression. Agg (ξ) represents the driving aggression score. The driving safety-related reward function is established using the following formula: r Safety =r Risk +r Agg

[0090] (2) Data mechanism coupled planning control is performed, as shown in Figure 2. A reinforcement learning and intelligent chassis secondary planning control architecture based on chassis feedback is constructed. Data mechanism coupled planning control is achieved through real-time dynamic coupling of the vehicle chassis mechanism model and reinforcement learning algorithm. The coupled planning control process includes three parts: vehicle-side information processing, reinforcement learning primary planning control, and intelligent chassis secondary planning control. First, vehicle-side information processing is performed. Each intelligent connected vehicle obtains semantic bird's-eye view information from the roadside through V2I communication and clips the global semantic bird's-eye view based on the vehicle's position sensor information. The clipped semantic bird's-eye view, combined with sensor information and intelligent chassis secondary planning control results, will be stacked after two consecutive frames and used as the input for reinforcement learning primary planning control. Then, primary planning control is performed through reinforcement learning. The state variables output from the information processing of each intelligent connected vehicle are used as the input to the reinforcement learning neural network, and the corresponding primary planning control result is output. The single-stage planning and control process includes steering wheel control, throttle control, and brake control, where: steering ∈ [-1,1], throttle ∈ [0,1], and brake ∈ [0,1].

[0091] Where steering represents steering wheel control, throttle represents accelerator control, and brake represents brake control, the action space of this invention is ∈ [-1,1]. 2 Two types of motion control are used: steering wheel control and accelerator-brake control. For accelerator-brake control, [-1,0] represents brake control, and [0,1] represents accelerator control. This invention uses a Beta distribution as the output of the reinforcement learning linear programming control: Beta=B(α,β)

[0092] Here, α and β represent two parameters of the Beta distribution. This invention further samples the corresponding action control variables from the Beta distribution. The Beta distribution is used in this invention because it has several advantages over the Gaussian distribution commonly used in model-free reinforcement learning. One advantage is its ability to dynamically simulate sample distributions of various shapes. Furthermore, unlike the Gaussian distribution, which extends to infinity in both positive and negative directions, the Beta distribution has bounded support from 0 to 1 and does not require mandatory constraints. In summary, the Beta distribution provides a flexible, variable, and bounded method for simulating various sample distributions and handling bounded variables, making it a superior choice to the Gaussian distribution.

[0093] Finally, secondary planning control of the intelligent chassis is performed. Based on the output of the reinforcement learning primary planning control and the desired path, secondary planning control based on the intelligent chassis multi-subsystem is conducted. The output of the secondary planning control directly controls the torque and steering angle of the vehicle's four wheels, and the control variables are again transmitted as part of the state variables to the reinforcement learning primary planning control processes of each intelligent connected vehicle. The secondary planning control process establishes the coordination and cooperation problem between intelligent chassis systems as an optimization problem considering global performance indicators, described by the following equation:

[0094] Among them, U i (t) represents the control quantity of intelligent subsystem i at time t, J i The overall cost function of the intelligent subsystem is represented by the following: This represents the predicted state of intelligent subsystem i at a future time. This represents the control quantity of intelligent subsystem i at a future time. This represents the assumed state of the neighbor intelligent subsystem j. Let λ represent the hypothetical control input of the neighbor intelligent subsystem j. i This represents the coupling coefficient of the cost function between intelligent subsystems. X i W represents the predicted state of intelligent subsystem i. i U represents the reference state sequence of intelligent subsystem i. i This represents the sequence of control inputs for intelligent subsystem i. Represents the sequence of state weight coefficients. This represents the sequence of control weight coefficients for intelligent subsystem i.

[0095] Among them, F ii G ii F ij G ij F jj G jj F jk G jk They represent the computational matrices, x and x respectively. i (k) represents the state of the intelligent subsystem i at time k in the future, X k U represents the state sequence of the intelligent subsystem at time k in the future. k This represents the sequence of control variables for intelligent subsystem j at time k in the future. represents the transpose matrix of the control input sequence of intelligent subsystem i, where const represents a constant. This can be represented by the following three formulas:

[0096] Among them, Q i R represents the state weight coefficient of intelligent subsystem i. i This represents the control weight coefficient of intelligent subsystem i. This represents the transpose of the state sequence of intelligent subsystem i at future time k. Q represents the transpose matrix of the reference state sequence of the intelligent subsystem j. j This represents the state weight coefficient of the intelligent subsystem j. This represents the transpose of the control sequence of intelligent subsystem j at future time k. Let be the transpose of the state sequence of intelligent subsystem j at future time k.

[0097] (3) Conduct data mechanism coupling evaluation, as shown in Figure 3. A comfort quantification index is established, and coupling evaluation is achieved through a neural network screening mechanism based on the vehicle chassis mechanism model. The comfort quantification index serves as the basis for quantitative evaluation of the control effect of intelligent connected vehicles. The comfort index modeling mainly involves using the degree of state change during vehicle operation, taking lateral acceleration, yaw acceleration, and longitudinal acceleration—the three most noticeable state quantities for human perception—as measurement indicators. Sensitive intervals for human perception are plotted sequentially, and the weighted sum of squares of the three state quantities is used as the quantification index.

[0098] in, S represents a quantitative indicator of human comfort based on a smart chassis. i (t) represents the lateral acceleration, yaw acceleration, and longitudinal acceleration at time t, ω i The weighted parameters i∈[0,m] represent the lateral acceleration, yaw acceleration, and longitudinal acceleration, respectively.

[0099] (4) Neural network selection. Using the established quantitative indicators of human comfort based on the intelligent chassis as the selection criteria, the reinforcement learning neural network parameters corresponding to the intelligent connected vehicle that best achieves human-perceived comfort are selected. The selection process is as follows:

[0100] in, This indicates that at time t, the parameters φ of the intelligent connected vehicle itself... t,i And another intelligent connected vehicle parameter φ t,i′ New network parameters obtained through aggregation.

[0101] (5) Neural network parameter aggregation. This mainly involves acquiring the neural network parameters of the intelligent connected vehicle through roadside V2I communication, calculating shared neural network parameters through parameter averaging, and finally distributing these parameters to the vehicle via V2I communication to achieve experience sharing until the network converges. The parameter averaging is performed according to the following equation:

[0102] in, φ′ represents the shared neural network parameters at time m, N represents the number of intelligent connected vehicles, and φ′ represents the number of intelligent connected vehicles. m,i Let m represent the neural network parameters of the i-th intelligent connected vehicle at time m.

[0103] In summary, this invention provides a vehicle-road cooperative control architecture based on data-mechanism coupling modeling, proposes a multi-agent system modeling method driven by data and mechanism fusion, a vehicle-road cooperative group optimization method based on federated reinforcement learning, and establishes a vehicle decision model parameter update technology based on multi-dimensional experience sharing, solving the interpretability and generalization problems of purely data-driven models. It leverages roadside advantages to build a rule-based driving safety field, realizing rule-guided data-driven training. Addressing the difficulty of traditional mechanism modeling in autonomous driving, it proposes a data-mechanism coupling driven model, constructs a quadratic programming control framework based on an intelligent chassis, and innovatively proposes state variable input based on chassis feedback, solving problems such as questionable credibility of purely data-driven models, reliance on large-scale data, and opaque and uninterpretable decision-making processes. Based on the sensitive range of human perception, it introduces weighted lateral acceleration, yaw angle acceleration, and longitudinal acceleration to construct a comfort quantification index to screen locally optimal strategies for the current environment. By synthesizing a globally shared model that benefits from different environments, it achieves a balance between sample efficiency and model robustness.

[0104] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent methods or modifications that do not depart from the technology of the present invention should be included within the scope of protection of the present invention.

Claims

1. A vehicle-road cooperative control architecture system based on data mechanism coupling modeling, characterized in that, include: The system comprises three parts: rule-guided coupled modeling, data mechanism coupled planning and control, and data mechanism coupled evaluation. The rule-guided coupled modeling part mainly builds a driving safety field by processing bird's-eye view information at the road end, and guides the data-driven control training process by establishing a fusion reward function. The data mechanism coupled planning and control part mainly constructs a reinforcement learning and intelligent chassis secondary planning and control algorithm based on chassis feedback. Data mechanism coupled planning and control is realized through real-time dynamic coupling of the vehicle chassis mechanism model and the reinforcement learning algorithm. The data mechanism coupling evaluation section is used to build a comfort quantification index, and achieves coupling evaluation and group optimization through neural network screening based on the vehicle chassis mechanism model.

2. The vehicle-road cooperative control architecture system based on data mechanism coupling modeling according to claim 1, characterized in that, The rule-guided coupled modeling component includes: a roadside information processing module, a safety field modeling module, and a reward function modeling module; The roadside information processing module mainly utilizes the roadside's field of vision advantage to transform images from a bird's-eye view into semantic bird's-eye views. It then uses the dynamic information in the semantic bird's-eye views to model the interactions between intelligent connected vehicles on the roadside. Finally, it transmits the safety field information and semantic bird's-eye views to the vehicle via V2I communication. The safety field modeling module establishes the safety field using the following equation: φ=a x / b y =l v / w v Among them, S sta C represents the static safety field strength. a The static safety field strength coefficient is represented by x0 and y0, which represent the coordinates of the static risk center O(x0, y0), ε represents the safety field shape coefficient, and a represents the static safety field strength coefficient. x and b y This represents the appearance coefficient of the intelligent connected vehicle, where φ is the aspect ratio of the intelligent connected vehicle, and l v Indicates the length of the vehicle, w v Indicates vehicle width; The reward function modeling module considers both driving expectation and driving safety: r=rexpectation+r safety Where r represents the reward function, rexpectation represents the reward function related to driving expectations, and r safety Indicates driving safety All-related reward function.

3. The vehicle-road cooperative control architecture system based on data mechanism coupling modeling according to claim 2, characterized in that, The semantic bird's-eye view is a 4-channel matrix that includes static road and lane information, as well as dynamic expected path and vehicle information.

4. The vehicle-road cooperative control architecture system based on data mechanism coupling modeling according to claim 2, characterized in that, When the intelligent connected vehicle moves, the risk center O(x0,y0) of the Gaussian safety field will shift to a new risk center O′(x′0,y′0) as the vehicle moves: Where, k v Indicates the moving adjustment factor, and The sign is related to the direction of motion; β represents the transfer vector of the connected vehicle. The angle between the vehicle and the coordinate axes in the Cartesian coordinate system forms a virtual vehicle of length l′ under the influence of the risk center transfer in the dynamic safety field. v Width is w′ v S dyn The dynamic safety field strength is represented by the new aspect ratio φ′=a′. x / b′ y =l′ v / w′ v .

5. The vehicle-road cooperative control architecture system based on data mechanism coupling modeling according to claim 2, characterized in that, The driving expectation-related reward function includes both horizontal and vertical aspects. Lateral driving expected reward function: d0=min(||a 28,28 -bdesiredroute|| 2 ) r1 lateral =-log 1.1 (|d0|+1) r2 lateral =-10*|sin(radians(θ))| r lateral ＝r1 lateral +r2 lateral Where d0 represents the distance from the center line of the lane, a 28,28 R1 represents the vehicle center, bdesiredroute represents the desired path, and r1 represents the desired path. lateral R2 represents the lateral distance reward function. lateral Let r represent the heading angle reward function, θ represent the heading angle deviation at the vehicle end, and r represent the heading angle reward function. lateral This represents the expected reward function for lateral driving. Longitudinal driving expected reward function: d min =min(||a 28,28 -b x,y || 2 ) r2longitudinal=-|v ego –9| rlongitudinal=r1longitudial+r2longitudial Where, d min b represents the minimum distance between autonomous vehicles. x,y The center of the other vehicle is represented by x, the time of collision is represented by v. ego Let r1 represent the vehicle speed, r2 represent the distance reward function, rlongitudinal represent the speed reward function, and rlongitudinal represent the expected longitudinal driving reward function. The driving expectation-related reward function is established using the following formula: rexpectation=r lateral +rlongitudinal。 6. The vehicle-road cooperative control architecture system based on data mechanism coupling modeling according to claim 2, characterized in that, The driving safety-related reward function is calculated by the roadside driving safety field from two aspects: driving safety and driving aggression. The first aspect is driving safety: Among them, R i,j (t) represents the driving risk posed by intelligent connected vehicle j to intelligent connected vehicle i. Let represent the field strength of intelligent connected vehicle j for intelligent connected vehicle i. Let θ represent the speed of the intelligent connected vehicle j at time t. i,j (t) represents the angle between intelligent connected vehicle i and intelligent connected vehicle j at time t, r Risk Let f represent the reward function related to driving risk. Risk (ξ) represents the driving risk score, R thr τ represents the risk threshold. rc Indicates the duration exceeding the risk threshold; Among them, R j,i (t′) represents the driving risk posed by intelligent connected vehicle i to intelligent connected vehicle j. Let i represent the field strength of intelligent connected vehicle i for intelligent connected vehicle j. Let θ represent the speed of the intelligent connected vehicle i at time t′. j,i (t′) represents the angle between intelligent connected vehicle j and intelligent connected vehicle i at time t′, r Agg Let f represent the reward function related to driving aggression. Agg (ξ) represents the integral of driving aggression; The driving safety-related reward function is established using the following formula: r Safety ＝r Risk +r Agg 。 7. The vehicle-road cooperative control architecture system based on data mechanism coupling modeling according to claim 1, characterized in that, The data mechanism coupled planning and control part includes: a vehicle-side information processing module, a reinforcement learning primary planning and control module, and an intelligent chassis secondary planning and control module; The vehicle-side information processing module of the intelligent connected vehicle uses the semantic bird's-eye view information provided by the roadside and the vehicle's position sensor information to crop the global semantic bird's-eye view. The cropped semantic bird's-eye view, combined with the sensor information and the intelligent chassis secondary planning control results, is stacked into two consecutive frames and used as the input of the reinforcement learning primary planning control module. The reinforcement learning primary planning control module takes the state variables output by the vehicle-side information processing module as input to the reinforcement learning neural network and outputs the corresponding primary planning control results, including steering wheel control, throttle control, and brake control, wherein: syeering∈[-1,1] throttle∈[0,1] brake∈[0,1] Where syeering represents the steering wheel control input, throttle represents the accelerator control input, and brake represents the brake control input; the Beta distribution is used as the output of the reinforcement learning primary programming control. Beta = B(α,β) Where α and β represent two parameters of the Beta distribution; The intelligent chassis secondary planning control module implements intelligent chassis secondary planning control based on the output of the reinforcement learning primary planning control module and the desired path. The output control quantity will directly control the torque and steering angle of the vehicle's four wheels, and the control quantity will be transmitted again as part of the state quantity to the reinforcement learning primary planning control module of each intelligent connected vehicle.

8. The vehicle-road cooperative control architecture system based on data mechanism coupling modeling according to claim 7, characterized in that the intelligent chassis secondary planning control module establishes the coordination and cooperation problem between intelligent chassis systems as an optimization problem considering global performance indicators, described by the following equations: in, U i (t) represents the control quantity of intelligent subsystem i at time t, J i J j This represents the total cost function of intelligent subsystems i and j. This represents the predicted state of intelligent subsystem i at a future time. This represents the control quantity of intelligent subsystem i at a future time. This represents the assumed state of the neighbor intelligent subsystem j. Let λ represent the hypothetical control input of the neighbor intelligent subsystem i. i This represents the coupling coefficient of the cost function between intelligent subsystems. X i W represents the predicted state of intelligent subsystem i. i U represents the reference state sequence of intelligent subsystem i. i This represents the sequence of control inputs for intelligent subsystem i. Represents the sequence of state weight coefficients. This represents the sequence of control weight coefficients for intelligent subsystem i; Among them, F ii G ii F ij G ij F jj G jj F jk G jk They represent the computational matrices, x and x respectively. i (k) represents the state of the intelligent subsystem i at time k in the future, X k U represents the state sequence of the intelligent subsystem at time k in the future. k This represents the sequence of control variables for intelligent subsystem j at time k in the future. represents the transpose matrix of the control input sequence of intelligent subsystem i, where const represents a constant; This can be represented by the following three formulas: Among them, Q i R represents the state weight coefficient of intelligent subsystem i. i This represents the control weight coefficient of intelligent subsystem i. The transpose of the state sequence of intelligent subsystem j at future time k is given by the following matrix. Q represents the transpose matrix of the reference state sequence of the intelligent subsystem j. j This represents the state weight coefficient of the intelligent subsystem j. This represents the transpose of the control sequence of intelligent subsystem j at future time k. Let be the transpose of the state sequence of intelligent subsystem j at future time k.

9. The vehicle-road cooperative control architecture system based on data mechanism coupling modeling according to claim 1, characterized in that, The data mechanism coupling evaluation part includes a comfort index modeling module, a neural network screening module, and a neural network parameter aggregation module; The comfort index modeling module, based on the degree of change in the vehicle's state during driving, uses lateral acceleration, yaw acceleration, and longitudinal acceleration—the three state quantities most noticeable to the human body—as measurement indicators, and sequentially plots the sensitive intervals of human perception, using the weighted sum of squares of the three state quantities as the quantitative indicator. in, S represents a quantitative indicator of human comfort based on a smart chassis. i (t) represents the lateral acceleration, yaw acceleration, and longitudinal acceleration at time t, ω i The weighted parameters represent the lateral acceleration, yaw acceleration, and longitudinal acceleration, i∈[0,m], where m=3 represent the lateral, yaw, and longitudinal accelerations, respectively. The neural network screening module uses the constructed human comfort quantification index based on the intelligent chassis as the screening criterion to screen the reinforcement learning neural network parameters corresponding to the intelligent connected vehicle that achieves the best perceived human comfort. The screening process is as follows: in, This indicates that at time t, the parameters φ of the intelligent connected vehicle itself... t,i And another intelligent connected vehicle parameter φ t,i′ New neural network parameters obtained through aggregation; The neural network parameter aggregation module acquires the neural network parameters of the intelligent connected vehicle through roadside V2I communication. The shared neural network parameters are calculated by averaging the parameters, and then sent to the vehicle via V2I communication to achieve experience sharing until the network converges. The parameter averaging is performed according to the following equation: in, φ′ represents the shared neural network parameters at time m, N represents the number of intelligent connected vehicles, and φ′ represents the number of intelligent connected vehicles. m,i Let m represent the neural network parameters of the i-th intelligent connected vehicle at time m.

10. A method for constructing a vehicle-road cooperative control architecture system based on data mechanism coupling modeling, characterized in that, Including the following: Step 1: Perform rule-guided coupled modeling. First, information processing is performed at the roadside, leveraging the roadside's field of view advantage to transform the bird's-eye view image into a semantic bird's-eye view. Dynamic information from the semantic bird's-eye view is then used to model the interaction between intelligent connected vehicles at the roadside. Finally, safety field information and the semantic bird's-eye view are transmitted to the vehicle via V2I communication, where a safety field is established to model the intelligent connected vehicle interaction process from the roadside perspective. Finally, reward function modeling is performed considering both driving expectations and driving safety. Step 2: Construct a data mechanism coupled planning and control algorithm; Construct an architecture based on chassis feedback reinforcement learning and intelligent chassis secondary planning and control, and realize data mechanism coupled planning and control through real-time dynamic coupling of intelligent chassis system and reinforcement learning algorithm; The data mechanism coupled planning and control algorithm includes vehicle-side information processing, reinforcement learning-based primary planning and control, and intelligent chassis-based secondary planning and control. First, vehicle-side information processing is performed. Each intelligent connected vehicle obtains semantic bird's-eye view information from the roadside via V2I communication. Based on the vehicle's position sensor information, the global semantic bird's-eye view is cropped. The cropped semantic bird's-eye view, combined with sensor information and the intelligent chassis-based secondary planning and control results, is stacked over two consecutive frames and used as input for reinforcement learning-based primary planning and control. Then, primary planning and control is performed through reinforcement learning. The state variables output from the information processing of each intelligent connected vehicle are used as input to the reinforcement learning neural network, outputting the corresponding primary planning and control results. Finally, intelligent chassis-based secondary planning and control is performed. Based on the reinforcement learning-based primary planning and control output and the desired path, secondary planning and control based on the intelligent chassis is performed. The secondary planning and control output directly controls the torque and steering angle of the vehicle's four wheels, and the output control quantity is again transmitted as part of the state variables to the reinforcement learning-based primary planning and control process of each intelligent connected vehicle. Step 3: Conduct data mechanism coupling evaluation; build a comfort quantification index, and achieve coupling evaluation through a neural network screening mechanism based on the vehicle chassis mechanism model. Based on the human comfort quantification index, realize the quantitative evaluation of the control effect of intelligent connected vehicles. Step 4: Perform neural network screening; use quantitative indicators of human comfort as the screening criteria to screen the reinforcement learning neural network parameters corresponding to the intelligent connected vehicle that provides the best perceived comfort to the human body. Step 5: Perform neural network parameter aggregation; obtain the neural network parameters of the intelligent connected vehicle through roadside V2I communication, calculate the shared neural network parameters by averaging the parameters, and finally send them to the vehicle through V2I communication to achieve experience sharing until the network converges; In step 1, the reward function modeling considers both driving expectation and driving safety. r＝r expection +r safety Where: r represents the reward function, r expection Let r represent the expected reward function related to driving. safety This represents a reward function related to driving safety. The reward function related to driving expectations includes both lateral and longitudinal aspects. First, let's consider lateral driving expectations: d0=min(||a 28,28 -bdesiredroute|| 2 ) r1 lateral =-log 1.1 (|d0|+1) r2 lateral =-10*|sin(radians(θ))| r lateral ＝r1 lateral +r2 lateral Where d0 represents the distance from the center line of the lane, a 28,28 R1 represents the vehicle center, bdesiredroute represents the desired path, and r1 represents the desired path. lateral R2 represents the lateral distance reward function. lateral Let r represent the heading angle reward function, θ represent the heading angle deviation at the vehicle end, and r represent the heading angle reward function. lateral This represents the expected reward function for lateral driving. Secondly, there are longitudinal driving expectations: d min =min(||a 28,28 -b x,y || 2 ) r2longitudinal=-|v ego –9| rlongitudinal=r1longitudinal+r2longitudinal Where, d min b represents the minimum distance between autonomous vehicles. x,y The center of the other vehicle is represented by x, the time of collision is represented by v. ego Let r1 represent the vehicle speed, r2 represent the distance reward function, rlongitudinal represent the speed reward function, and rlongitudinal represent the expected longitudinal driving reward function. Finally, the expected reward function for driving is established using the following formula: rexpectation=r lateral +rlongitudinal The reward function related to driving safety is calculated by the roadside driving safety field from two aspects: driving safety and driving aggression. First, driving safety: Among them, R i,j (t) represents driving risk, k represents the electric field strength between intelligent connected vehicle i and intelligent connected vehicle j. c This represents the risk perception coefficient. Let θ represent the speed of the intelligent connected vehicle j at time t. i,j (t) represents the angle between intelligent connected vehicle i and intelligent connected vehicle j at time t, r Risk Let f represent the reward function related to driving risk. Risk (ξ) represents the driving risk score, R thr τ represents the risk threshold. rc Indicates the duration exceeding the risk threshold; Among them, R j,i (t′) represents the driving risk. This represents the field strength between intelligent connected vehicle j and intelligent connected vehicle i. Let θ represent the speed of the intelligent connected vehicle i at time t′. j,i (t′) represents the angle between intelligent connected vehicle j and intelligent connected vehicle i at time t′, r Agg Let f represent the reward function related to driving aggression. Agg (ξ) represents the integral of driving aggression; The driving safety-related reward function is established using the following formula: r Safety ＝r Risk +r Agg ； In step 2, the single-stage planning and control process includes three parts: steering wheel control, throttle control, and brake control. steering∈[-1,1] throttle∈[0,1] brake∈[0,1] Where, steering represents steering wheel control, throttle represents accelerator control, and brake represents brake control. quantity; Using the Beta distribution as the output of primary planning control in reinforcement learning: Beta = B(α,β) Where α and β represent two parameters of the Beta distribution; In step 2, the intelligent chassis secondary planning control process establishes the coordination and cooperation problem between intelligent chassis systems as an optimization problem considering global performance indicators, described by the following equations: Among them, U i (t) represents the control quantity of intelligent subsystem i at time t, J i The total cost function of intelligent subsystem i is represented by the function. This represents the predicted state of intelligent subsystem i at a future time. This represents the control quantity of intelligent subsystem i at a future time. This represents the assumed state of the neighbor intelligent subsystem j. Let λ represent the hypothetical control input of the neighbor intelligent subsystem i. i This represents the coupling coefficient of the cost function between intelligent subsystems. X i W represents the predicted state of intelligent subsystem i. i U represents the reference state sequence of intelligent subsystem i. i This represents the sequence of control variables for intelligent subsystem i. Represents the sequence of state weight coefficients. This represents the sequence of control weight coefficients for intelligent subsystem i; Among them, F ii G ii F ij G ij F jj G jj F jk G jk They represent the computational matrices, x and x respectively. i (k) represents the state of the intelligent subsystem i at time k in the future, X k U represents the state sequence of the intelligent subsystem at time k in the future. k This represents the sequence of control variables for intelligent subsystem j at time k in the future. This matrix represents the transpose of the control input sequence of intelligent subsystem i, where const represents a constant. This can be represented by the following three formulas: Among them, Q i R represents the state weight coefficient of intelligent subsystem i. i This represents the control weight coefficient of intelligent subsystem i. The transpose of the state sequence of intelligent subsystem j at future time k is given by the following matrix. Q represents the transpose matrix of the reference state sequence of the intelligent subsystem j. j This represents the state weight coefficient of the intelligent subsystem j. This represents the transpose of the control sequence of intelligent subsystem j at future time k. The transpose of the state sequence of intelligent subsystem j at future time k is represented by . In step 3, the comfort index modeling is based on the degree of state change of the vehicle during driving. The three most noticeable state variables for human perception—lateral acceleration, yaw rate acceleration, and longitudinal acceleration—are used as measurement indicators. Sensitive intervals for human perception are plotted sequentially, and the weighted sum of squares of the three state variables is used as the quantitative indicator. in, S represents a quantitative indicator of human comfort based on a smart chassis. i (t) represents the lateral acceleration, yaw acceleration, and longitudinal acceleration at time t, ω i The weighted parameters represent the lateral acceleration, yaw acceleration, and longitudinal acceleration, i∈[0,m], where m=3 represent the lateral, yaw, and longitudinal accelerations, respectively. In step 4, the screening process includes: in, This indicates that at time t, the parameters φ of the intelligent connected vehicle itself... t,i And another intelligent connected vehicle parameter φ t,i′ New network parameters obtained through aggregation; In step 5, the parameters are averaged according to the following equation: in, φ′ represents the shared neural network parameters at time m, N represents the number of intelligent connected vehicles, and φ′ represents the number of intelligent connected vehicles. m,i Let m represent the neural network parameters of the i-th intelligent connected vehicle at time m.

Images