Intelligent networked automobile dynamics simulation method and system based on physical engine

Abstract

The invention provides an intelligent networked automobile dynamics simulation method and system based on a physical engine, and relates to the technical field of intelligent networked automobiles, and the method comprises the steps: obtaining vehicle parameters and road condition information of an intelligent networked automobile, optimizing a multi-agent cooperative control model through a self-adaptive optimization algorithm based on the vehicle parameters and the road condition information, and carrying out the optimization of the multi-agent cooperative control model. Constructing a digital twinborn model, and mapping the optimized multi-agent cooperative control model to an actual intelligent networked automobile control system; the method comprises the following steps: establishing a distributed edge computing network of an intelligent networked automobile group, constructing a decentralized knowledge base by utilizing a block chain technology, forming a decision-making experience data set, designing a multi-level attention mechanism to extract target information based on the decision-making experience data set and multi-modal data of a current traffic scene, and generating a final vehicle control instruction. And transmitting the final vehicle control instruction to a vehicle execution system to realize dynamic adaptive group cooperative control of the intelligent networked vehicle.

Description

Technical Field

[0001] The present invention relates to intelligent networked vehicle technology, and in particular to an intelligent networked vehicle dynamics simulation method and system based on a physical engine. Background Art

[0002] Against the backdrop of the rapid development of intelligent connected vehicles, vehicle dynamics simulation technology has gradually become an important means of improving traffic safety and efficiency. Traditional dynamics simulation methods mostly rely on simple physical models and cannot fully consider complex road conditions and interactions between vehicles. With the rise of deep learning technologies such as graph neural networks and recurrent neural networks, combined with a dynamic simulation environment of a multi-level physical engine, the behavior of vehicles in multiple scenarios can be simulated more accurately. This method can not only improve simulation accuracy, but also provide data support for the collaborative control of intelligent connected vehicles. Through the combination of adaptive optimization algorithms and digital twin models, real-time decision support can be achieved, thereby promoting the development of intelligent transportation systems.

[0003] Therefore, the patent of this invention proposes a method and system for dynamic simulation of intelligent connected vehicles based on a physical engine. By constructing a multi-level physical engine environment, introducing a multi-agent collaborative control model, and building a digital twin closed loop with the actual vehicle control system, it achieves a high degree of simulation of the group behavior of intelligent connected vehicles and direct verification and optimization of control strategies, providing key technical support for the rapid iterative development of intelligent connected vehicle control systems. Summary of the invention

[0004] The embodiments of the present invention provide a method and system for dynamics simulation of an intelligent connected vehicle based on a physical engine, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] A method for dynamics simulation of an intelligent connected vehicle based on a physical engine is provided, comprising:

[0007] Obtain vehicle parameters and road condition information of the intelligent connected vehicle, and based on the vehicle parameters and road condition information, construct a multi-level physical engine simulation environment including macro traffic flow, meso vehicle dynamics, and micro components. In the multi-level physical engine simulation environment, use graph neural networks and recurrent neural networks to construct a multi-agent collaborative control model of the intelligent connected vehicle and deploy it to the multi-level physical engine simulation environment, and simulate the group behavior of intelligent connected vehicles in multiple scenarios. Based on the simulation results, use an adaptive optimization algorithm to optimize the multi-agent collaborative control model, and construct a digital twin model to map the optimized multi-agent collaborative control model to the actual intelligent connected vehicle control system;

[0008] Establish a distributed edge computing network for the intelligent connected vehicle group, receive real-time data from on-board sensors and workshop communication data, perform local data processing and decision calculations through edge nodes, output initial vehicle control instructions, deploy self-organizing learning algorithms based on initial vehicle control instructions and environmental data, generate individual control strategies through dynamic weight adjustment and local rule updates, use individual control strategies and corresponding scene features as input, use blockchain technology to build a decentralized knowledge base, form a decision experience data set, design a multi-level attention mechanism based on the decision experience data set and the multimodal data of the current traffic scene to extract target information, and output scene understanding results and decision features;

[0009] A dynamic scene graph is constructed by utilizing the scene understanding results and decision-making features. Based on the dynamic scene graph and historical trajectory data, multi-step prediction results and corresponding decision suggestions are output. The multi-step prediction results, decision suggestions and local state information of each vehicle are used as input to design a swarm intelligence emergence mechanism to generate coordinated behavior instructions that can optimize the overall traffic flow. The coordinated behavior instructions, individual control strategies and predictive decision results are integrated to generate final vehicle control instructions through a multi-objective decision optimization algorithm. The final vehicle control instructions are transmitted to the vehicle execution system to achieve dynamic adaptive swarm collaborative control of intelligent connected vehicles.

[0010] In an optional embodiment,

[0011] Based on the vehicle parameters and road condition information, a multi-level physical engine simulation environment including macro traffic flow, meso vehicle dynamics and micro components is constructed. In the multi-level physical engine simulation environment, a multi-agent collaborative control model of intelligent connected vehicles is constructed using graph neural networks and recurrent neural networks and deployed to the multi-level physical engine simulation environment. The group behaviors of intelligent connected vehicles in multiple scenarios are simulated, including:

[0012] At the macroscopic traffic flow level, vehicles are abstracted as particles and roads are abstracted as channels. The interaction rules between particles are defined, and the differential equations of the traffic flow model are solved using numerical integration methods to obtain the macroscopic motion trajectory of each vehicle. At the mesoscopic vehicle dynamics level, vehicles are simplified as multi-body systems, and the vehicle dynamics differential equations are established to solve the vehicle's position, velocity, acceleration, and angular velocity, and simulate the interaction force between the vehicle and the road. At the microscopic component level, geometric modeling and physical modeling are performed on the key components of the vehicle, and the internal state changes and output responses of the simulated components are simulated through simulation, ultimately forming a multi-level physical engine simulation environment.

[0013] In a multi-level physical engine simulation environment, graph neural networks are used to model the interaction between agents. Vehicles are abstracted as nodes in the graph, and wireless communication connections between vehicles are abstracted as edges in the graph. Node features are iteratively updated through a message passing mechanism. A recurrent neural network is used to model the sequential decision-making process of a single agent, unifying the state changes and decision iterations of the vehicle into the cyclic transmission of hidden states.

[0014] The output of the graph neural network is used as the input of the recurrent neural network to form an end-to-end multi-agent collaborative control model. The multi-agent collaborative control model is deployed in a multi-level physical engine simulation environment. The vehicle status and road condition information in the simulation environment are used as the input of the model. The multi-agent collaborative control model is used to process the input vehicle status and road condition information to generate a distributed control strategy for intelligent connected vehicles, including the expected speed, expected acceleration, and expected lane for each vehicle.

[0015] The distributed control strategy is used as the control input of a multi-level physical engine simulation environment to simulate the group behavior of intelligent connected vehicles in multiple scenarios, including platoon formation, lane change and collision avoidance.

[0016] In an optional embodiment,

[0017] The multi-agent collaborative control model is used to process the input vehicle status and road condition information to generate the distributed control strategy of the intelligent connected vehicle, including:

[0018] Constructing an environmental dynamics model, which receives historical states and action sequences as input, assigns weights to states and actions at different moments through an attention mechanism, uses an external memory mechanism to maintain memory units to store and retrieve long-term information, and performs convolution operations in the time and space dimensions through a spatiotemporal convolutional network to output predicted future states;

[0019] The constructed environmental dynamics model is trained using historical trajectory data. Based on the trained environmental dynamics model, a nonlinear model predictive controller is designed. The input is the current state and the predicted future state. The random MPC idea is introduced to simulate uncertainty and output a preliminary control sequence.

[0020] In the multi-agent collaborative control scenario, the nonlinear model predictive control problem is decomposed into multiple sub-problems, each sub-problem corresponds to an agent, each agent receives local state information, and performs collaborative optimization through a distributed MPC architecture to generate a preliminary distributed control strategy. By combining numerical optimization methods with heuristic optimization techniques, the optimization objectives and constraints are received, and the preliminary distributed control strategy is optimized through iterative search to obtain the optimized distributed control strategy, including the expected speed, expected acceleration and expected lane for each vehicle.

[0021] In an optional embodiment,

[0022] Based on the decision-making experience dataset and the multimodal data of the current traffic scene, a multi-level attention mechanism is designed to extract target information, and the output scene understanding results and decision features include:

[0023] Build a multi-task learning framework, which includes a shared feature extractor based on convolutional neural networks and multi-layer attention mechanism, a scene understanding subtask module, and a decision feature extraction subtask module;

[0024] Receive a traffic scene data set including scene input data, scene understanding labels and decision labels; use the shared feature extractor to extract a common scene representation from the scene input data to obtain a shared feature, input the shared feature into a scene understanding subtask module, perform semantic segmentation, target detection and relationship recognition tasks, and obtain a scene understanding result; at the same time, input the shared feature into a decision feature extraction subtask module, perform behavior prediction and risk assessment tasks, and obtain a decision feature extraction result;

[0025] Based on the scene understanding results and scene understanding labels, the subtask loss functions of semantic segmentation, object detection and relationship recognition are calculated respectively. Based on the decision feature extraction results and decision labels, the subtask loss functions of behavior prediction and risk assessment are calculated respectively. A dynamic weight adjustment strategy is adopted, and the learning difficulty and convergence speed information of each subtask is input. The weight coefficients of the subtask loss functions are adaptively calculated, and the weighted subtask loss functions are summed to obtain the joint loss function.

[0026] The joint loss function is input into the optimization algorithm, and the adaptive moment estimation method with learning rate decay and gradient clipping is used to calculate the parameter gradient and update the parameters of the shared feature extractor, scene understanding subtask module and decision feature extraction subtask module. The iteration is repeated until the preset termination condition is met, and the trained multi-task learning model is output.

[0027] Using the trained multi-task learning model, new traffic scene data is input and scene understanding results and decision feature extraction results are output simultaneously.

[0028] In an optional embodiment,

[0029] Using the scene understanding results and decision features, a dynamic scene graph is constructed. Based on the dynamic scene graph and historical trajectory data, multi-step prediction results and corresponding decision suggestions are output, including:

[0030] Construct a multi-level traffic rules knowledge graph, organize traffic concepts, relationships and constraints in the form of ontology, use description logic to realize the formal representation and reasoning of traffic rules, and obtain a traffic rules knowledge base;

[0031] Construct a multimodal encoding-decoding network structure. The encoder uses a graph convolutional neural network to extract features and aggregate information from dynamic scene graphs. It learns the hidden representation of nodes through multi-layer graph convolution operations and captures the interactive relationship between nodes. The decoder uses a long short-term memory network or a gated recurrent unit regression to generate vehicle trajectories. At each time step, the vehicle position at the next moment is predicted based on the trajectory points generated in the previous time step and the scene features output by the encoder.

[0032] In the constructed encoder, a multi-head self-attention mechanism is used to calculate the mutual influence between vehicles. Each attention head focuses on different interaction patterns to obtain multi-scale dependencies. In the decoder, a masked self-attention mechanism is used to generate future trajectories. The mask operation ensures that only information from previous moments can be accessed when predicting the trajectory points at the current moment, thus obtaining a preliminary trajectory prediction result.

[0033] The trajectory prediction is modeled as a learning problem of conditional probability distribution. The initial trajectory prediction results are used to train the model parameters by maximizing the conditional log-likelihood function. The stochastic gradient descent algorithm with adaptive learning rate is used to minimize the difference between the predicted trajectory and the true trajectory. In the inference stage, the Monte Carlo sampling method is used to generate multiple predicted trajectories according to the learned conditional probability distribution to obtain the trajectory prediction probability distribution.

[0034] Based on the obtained trajectory prediction probability distribution, the Monte Carlo simulation method is used to calculate the collision probability and minimum safe distance between vehicles, and the smoothness of vehicle driving is evaluated through the geometric characteristics and time characteristics of the trajectory. The trajectory prediction results and the road network topology are used to construct a time-varying traffic flow network, calculate the expected time for vehicles to arrive at their destinations and the impact of path selection on overall traffic efficiency, and obtain scenario evaluation results;

[0035] The obtained trajectory prediction probability distribution and scenario evaluation results are combined with the constructed traffic rules knowledge base, and decision recommendations are generated through probabilistic symbolic reasoning. Combined with the current state of the vehicle, trajectory prediction probability distribution and rule constraints in the knowledge graph, multi-step prediction results and corresponding decision recommendations are finally output.

[0036] In an optional embodiment,

[0037] Using trajectory prediction results and road network topology, we construct a time-varying traffic flow network, calculate the expected time for vehicles to arrive at their destinations and the impact of path selection on overall traffic efficiency, and obtain scenario evaluation results including:

[0038] The road network is divided into several traffic zones, each of which corresponds to a node in the network. According to the topological structure and geometric characteristics of the road network, the spectral clustering algorithm is used to divide the road network, the similarity matrix between nodes is defined, the matrix is ​​spectrally decomposed, the eigenvectors corresponding to the first k largest eigenvalues ​​are taken to form a matrix, and each row of the matrix is ​​clustered to obtain the traffic zone division result of the road network.

[0039] The trajectory prediction model is used to predict the vehicle trajectory in each time slice. For each cell, the number of vehicles falling into the current traffic cell is counted to obtain the traffic flow of each traffic cell in each time slice.

[0040] According to the road network topology, the number of roads connecting two adjacent traffic zones is counted. In each time slice, the traffic flow transferred from one traffic zone to another adjacent traffic zone is calculated. If the transfer probability of vehicles on each connecting road section is equal, the transfer probability between adjacent traffic zones is calculated based on the traffic flow and the number of connecting road sections.

[0041] According to the traffic zone division results, traffic flow in traffic zones and the transfer probability between traffic zones, a time-varying traffic flow network is constructed, in which nodes are traffic zones and node attributes are dynamic traffic flow in traffic zones. For each vehicle, the shortest time path is calculated on the time-varying traffic flow network using a dynamic programming algorithm based on the traffic zone where its starting position is located and the traffic zone where its destination is located. For each vehicle's actual driving path, the actual time it takes to reach the destination is calculated. The difference between the actual arrival time and the shortest expected time is defined as the time difference, and the overlap between the actual driving path and the shortest time path is defined as the path deviation.

[0042] The time differences and path deviations of all vehicles are weighted averaged respectively, and the weighted average of the time differences and the weighted average of the path deviations are weighted summed according to the preset trade-off parameters to obtain the efficiency evaluation value of the traffic scenario and the scenario evaluation result.

[0043] In an optional embodiment,

[0044] The final vehicle control command is generated through a multi-objective decision optimization algorithm, and the final vehicle control command is transmitted to the vehicle execution system to realize the dynamic adaptive group cooperative control of intelligent networked vehicles, including:

[0045] Establish a vehicle-cloud communication architecture and build a multi-level environment model in the cloud, including a static map layer, a dynamic prediction layer, and a semantic understanding layer. The static map layer models unchanging elements such as roads and lane lines, the dynamic prediction layer predicts the future movement of traffic participants based on physical motion models and trajectory history, and the semantic understanding layer analyzes the overall characteristics of traffic scenes.

[0046] Based on the environmental model, the optimal control strategy is solved in the cloud using a pre-built decision model to obtain a desired state sequence of the vehicle;

[0047] The desired state sequence is sent to the vehicle side, and fine-tuned by executing the optimization module on the vehicle side to generate coordinated behavior instructions and individual control strategies, integrate the coordinated behavior instructions, individual control strategies and desired state sequence, and generate final vehicle control instructions through a multi-objective decision optimization algorithm, and transmit the final vehicle control instructions to the vehicle execution system;

[0048] The actual execution results and environmental feedback of the vehicle execution system are recorded and used as local training data. The decision model is parameterized into a neural network. The vehicle uses the local training data to update the model parameters and upload the parameter gradients to the cloud. The cloud aggregates the average parameter gradients to update the global decision model and distributes it to all vehicles to form an optimized decision model. The optimized decision model is used to repeatedly execute the vehicle control instruction generation step to realize dynamic adaptive group collaborative control of intelligent connected vehicles.

[0049] According to a second aspect of the embodiments of the present invention,

[0050] Provided is a dynamics simulation system for intelligent connected vehicles based on a physical engine, including:

[0051] The first unit is used to obtain vehicle parameters and road condition information of intelligent connected vehicles, and based on the vehicle parameters and road condition information, construct a multi-level physical engine simulation environment including macro traffic flow, meso vehicle dynamics and micro components. In the multi-level physical engine simulation environment, a multi-agent collaborative control model of intelligent connected vehicles is constructed by using graph neural networks and recurrent neural networks and deployed to the multi-level physical engine simulation environment, and the group behavior of intelligent connected vehicles in multiple scenarios is simulated. Based on the simulation results, the multi-agent collaborative control model is optimized by using an adaptive optimization algorithm, and a digital twin model is constructed to map the optimized multi-agent collaborative control model to the actual intelligent connected vehicle control system;

[0052] The second unit is used to establish a distributed edge computing network for a group of intelligent connected vehicles, receive real-time data from on-board sensors and vehicle-to-vehicle communication data, perform local data processing and decision-making calculations through edge nodes, output initial vehicle control instructions, deploy self-organizing learning algorithms based on initial vehicle control instructions and environmental data, generate individual control strategies through dynamic weight adjustment and local rule updates, use individual control strategies and corresponding scene features as input, use blockchain technology to build a decentralized knowledge base, form a decision experience data set, design a multi-level attention mechanism based on the decision experience data set and multimodal data of the current traffic scene to extract target information, and output scene understanding results and decision features;

[0053] The third unit is used to use the scene understanding results and decision-making features to construct a dynamic scene graph, output multi-step prediction results and corresponding decision suggestions based on the dynamic scene graph and historical trajectory data, take the multi-step prediction results, decision suggestions and local state information of each vehicle as input, design a group intelligence emergence mechanism, generate coordinated behavior instructions that can optimize the overall traffic flow, integrate the coordinated behavior instructions, individual control strategies and predictive decision results, generate final vehicle control instructions through a multi-objective decision optimization algorithm, transmit the final vehicle control instructions to the vehicle execution system, and realize dynamic adaptive group collaborative control of intelligent connected vehicles.

[0054] According to a third aspect of the embodiments of the present invention,

[0055] An electronic device is provided, comprising:

[0056] processor;

[0057] a memory for storing processor-executable instructions;

[0058] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0059] According to a fourth aspect of the embodiments of the present invention,

[0060] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0061] In this embodiment,

[0062] By constructing a multi-level physical engine simulation environment, a highly realistic simulation of the group behavior of intelligent connected vehicles can be achieved, providing a real and reliable digital twin platform for the development and testing of collaborative control algorithms.

[0063] Using graph neural networks and recurrent neural networks to build a multi-agent collaborative control model can effectively handle the complex interactions and timing dependencies between vehicles, and improve the intelligence level of vehicle collaborative control.

[0064] The collaborative control model is optimized through an adaptive optimization algorithm so that it can dynamically adapt to different traffic scenarios and improve the robustness and generalization ability of the control strategy. A distributed edge computing network is established to realize real-time processing of vehicle perception data and low-latency execution of decision-making calculations, thereby improving the system response speed. A decentralized knowledge base is constructed using blockchain technology to realize the safe and reliable sharing of vehicle decision-making experience and promote the continuous evolution of vehicle intelligence. Multi-step prediction of vehicle future behavior is achieved based on dynamic scene graphs and trajectory prediction, making decisions more forward-looking and predictive. A group intelligence emergence mechanism is designed to optimize the overall traffic flow, while ensuring the safety of individual vehicles and taking into account the overall efficiency of the traffic system. A multi-objective decision optimization algorithm is used to generate control instructions that take into account multiple performance indicators, achieving a balance between multiple objectives such as safety, comfort, and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 A schematic diagram of a flow chart of a method for dynamics simulation of an intelligent connected vehicle based on a physical engine according to an embodiment of the present invention;

[0066] Figure 2 It is a schematic diagram of the structure of a smart connected vehicle dynamics simulation system based on a physical engine according to an embodiment of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0069] Figure 1 FIG. 1 is a flow chart of a method for dynamics simulation of an intelligent connected vehicle based on a physical engine according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0070] S101. Obtain vehicle parameters and road condition information of the intelligent connected vehicle, and based on the vehicle parameters and road condition information, construct a multi-level physical engine simulation environment including macro traffic flow, meso vehicle dynamics and micro components. In the multi-level physical engine simulation environment, use graph neural networks and recurrent neural networks to construct a multi-agent collaborative control model of the intelligent connected vehicle and deploy it to the multi-level physical engine simulation environment, and simulate the group behavior of intelligent connected vehicles in multiple scenarios. Based on the simulation results, use an adaptive optimization algorithm to optimize the multi-agent collaborative control model, and construct a digital twin model to map the optimized multi-agent collaborative control model to the actual intelligent connected vehicle control system.

[0071] Among them, vehicle parameters include physical properties of the vehicle, such as mass, size, tire characteristics, etc., as well as dynamic characteristics of the vehicle, such as maximum acceleration, maximum speed, steering performance, etc. Road condition information includes road geometry, road friction coefficient, traffic flow, etc.

[0072] Based on the acquired vehicle parameters and road condition information, a multi-level physical engine simulation environment is constructed. A physical engine is a computer software that can simulate the motion laws of the physical world, such as rigid body dynamics, collision detection, etc. The multi-level physical engine simulation environment includes a macro traffic flow layer, a meso vehicle dynamics layer, and a micro component layer. The macro traffic flow layer simulates the traffic flow changes of the entire road network; the meso vehicle dynamics layer simulates the motion state of a single vehicle; and the micro component layer simulates the working state of the internal components of the vehicle. By constructing a multi-level physical engine simulation environment, the performance of intelligent connected vehicles can be comprehensively evaluated.

[0073] In a multi-level physical engine simulation environment, a multi-agent collaborative control model of intelligent connected vehicles is constructed using graph neural networks and recurrent neural networks. Graph neural networks can model the interaction between vehicles, and recurrent neural networks can model the sequential decision-making process of vehicles. Multi-agents refer to multiple individuals with autonomous decision-making capabilities, and collaborative control refers to the control strategy that multiple agents reach a consensus on through interaction. Intelligent connected vehicles can be regarded as a multi-agent system, where each vehicle is an agent, and they achieve collaborative control through vehicle-to-vehicle communication and vehicle-road collaboration. The constructed multi-agent collaborative control model needs to be deployed in a physical engine simulation environment for training and testing.

[0074] In the physical engine simulation environment, the group behavior of intelligent connected vehicles in multiple scenarios is simulated, including highways, urban roads, intersections, etc., and the interactive behaviors of vehicles in different scenarios are simulated, such as lane changing, following, and avoiding. Based on the simulation results, the multi-agent collaborative control model is optimized using an adaptive optimization algorithm to improve the generalization and robustness of the model. The adaptive optimization algorithm can automatically adjust the optimization direction and step size according to the loss function of the model. Common algorithms include Adam and Adagrad. The optimized multi-agent collaborative control model can better adapt to different traffic scenarios.

[0075] In order to apply the optimized multi-agent collaborative control model to the actual intelligent connected vehicle system, a digital twin model needs to be built. A digital twin is a virtual representation of a physical entity that can synchronize the state of the physical entity in real time. By building a digital twin model of an intelligent connected vehicle, the optimized control strategy can be mapped to the control system of the actual vehicle to achieve virtual-real synchronization. The digital twin model needs to interact with the physical engine simulation environment, receive simulation data, and feedback control instructions.

[0076] In an optional embodiment,

[0077] Based on the vehicle parameters and road condition information, a multi-level physical engine simulation environment including macro traffic flow, meso vehicle dynamics and micro components is constructed. In the multi-level physical engine simulation environment, a multi-agent collaborative control model of intelligent connected vehicles is constructed using graph neural networks and recurrent neural networks and deployed to the multi-level physical engine simulation environment. The group behaviors of intelligent connected vehicles in multiple scenarios are simulated, including:

[0078] At the macroscopic traffic flow level, vehicles are abstracted as particles and roads are abstracted as channels. The interaction rules between particles are defined, and the differential equations of the traffic flow model are solved using numerical integration methods to obtain the macroscopic motion trajectory of each vehicle. At the mesoscopic vehicle dynamics level, vehicles are simplified as multi-body systems, and the vehicle dynamics differential equations are established to solve the vehicle's position, velocity, acceleration, and angular velocity, and simulate the interaction force between the vehicle and the road. At the microscopic component level, geometric modeling and physical modeling are performed on the key components of the vehicle, and the internal state changes and output responses of the simulated components are simulated through simulation, ultimately forming a multi-level physical engine simulation environment.

[0079] In a multi-level physical engine simulation environment, graph neural networks are used to model the interaction between agents. Vehicles are abstracted as nodes in the graph, and wireless communication connections between vehicles are abstracted as edges in the graph. Node features are iteratively updated through a message passing mechanism. A recurrent neural network is used to model the sequential decision-making process of a single agent, unifying the state changes and decision iterations of the vehicle into the cyclic transmission of hidden states.

[0080] The output of the graph neural network is used as the input of the recurrent neural network to form an end-to-end multi-agent collaborative control model. The multi-agent collaborative control model is deployed in a multi-level physical engine simulation environment. The vehicle status and road condition information in the simulation environment are used as the input of the model. The multi-agent collaborative control model is used to process the input vehicle status and road condition information to generate a distributed control strategy for intelligent connected vehicles, including the expected speed, expected acceleration, and expected lane for each vehicle.

[0081] The distributed control strategy is used as the control input of a multi-level physical engine simulation environment to simulate the group behavior of intelligent connected vehicles in multiple scenarios, including platoon formation, lane change and collision avoidance.

[0082] For example, first, at the macroscopic traffic flow level, vehicles are abstracted as particles and roads are abstracted as channels. Particles refer to individual vehicles in the traffic flow, and the relationship between vehicles is described by defining the interaction rules between particles, such as the headway time and speed difference between vehicles. Channels refer to the roads on which vehicles travel, which can be modeled according to geometric features such as the width and curvature of the road. By using numerical integration methods to solve the differential equations of the traffic flow model, the motion trajectory of each vehicle at a macroscopic scale can be obtained, reflecting the spatiotemporal evolution of the overall traffic flow. For example, in a highway scenario, a traffic flow model can be established by defining the following behavior and lane changing behavior between vehicles to simulate the motion trajectory of traffic under different density and speed conditions.

[0083] Secondly, at the meso-vehicle dynamics level, the vehicle is simplified into a multi-body system. A multi-body system refers to a system composed of multiple rigid bodies connected by mechanical structures such as hinges and springs. By establishing the vehicle's dynamic differential equations, the vehicle's position, velocity, acceleration, angular velocity and other state variables can be solved. At the same time, it is also necessary to simulate the interaction between the vehicle and the road, such as the friction between the tires and the road surface, the support force of the suspension, etc. For example, for a four-wheeled car, it can be simplified into a multi-body system consisting of a body and four wheels, and then the vehicle's dynamic equations are established based on the vehicle's geometric parameters and physical parameters, such as mass, moment of inertia, suspension stiffness, etc., and the vehicle's motion response under acceleration, braking, steering and other working conditions is solved.

[0084] Thirdly, at the micro-component level, geometric modeling and physical modeling of key vehicle components are required. Geometric modeling refers to the establishment of a three-dimensional solid model based on the geometric features of the component, such as its shape and size. Physical modeling refers to the establishment of a physical field model, such as stress field, temperature field, etc., based on the material properties and mechanical properties of the component. Through simulation solution, the internal state changes and output responses of components under actual working conditions, such as stress distribution and thermal deformation, can be simulated. For example, for an engine cylinder block, a three-dimensional solid model can be established based on its structural diagram, and then the elastic-plastic constitutive relationship of the cylinder block material, the thermodynamic state equation of the gas in the cylinder, etc., are considered to establish a thermal-structural coupling physical field model of the cylinder block, and the stress distribution and deformation of the cylinder block under the action of combustion pressure and thermal load are analyzed through finite element simulation.

[0085] After building a multi-level physical engine simulation environment, it is necessary to introduce intelligent algorithms to model the behavior decisions of vehicle agents. Graph neural networks are used to model the interaction between agents. Vehicles are abstracted as nodes in the graph, and wireless communication connections between vehicles are abstracted as edges in the graph. The characteristics of the nodes can represent the state information of the vehicle, such as position, speed, etc., and the characteristics of the edges can represent the relative relationship between vehicles, such as distance, speed difference, etc. By defining the message passing function on the nodes and edges, the iterative update of node features is realized, so that the nodes can fuse their own features and the features of neighboring nodes. For example, each node can pass its own state information to its neighboring nodes. After receiving the message, the neighboring node fuses it with its own state and updates its own features. After multiple rounds of iterations, the characteristics of the nodes will contain global information, which is conducive to decision-making.

[0086] For the sequential decision-making process of a single agent, a recurrent neural network is used for modeling. A recurrent neural network is a type of neural network suitable for processing sequence data. By introducing a loop structure, the output at the current moment depends not only on the input at the current moment, but also on the hidden state at the previous moment. The vehicle's state changes and decision iterations are unified into the loop transfer of the hidden state, so that the current decision not only considers the current environmental information, but also the previous historical information. For example, the vehicle's speed, acceleration, and the state of surrounding vehicles over a period of time can be input into the recurrent neural network to generate the hidden state at the current moment, which is then mapped to the decision action at the current moment through the fully connected layer.

[0087] In order to achieve end-to-end multi-agent collaborative control, the output of the graph neural network is used as the input of the recurrent neural network to form a complete model architecture. The graph neural network is responsible for modeling the interaction relationship between the agents, while the recurrent neural network is responsible for making sequential decisions. The model is deployed in a multi-level physical engine simulation environment, and the vehicle status and road condition information in the simulation environment are used as the input of the model. The model processes the input information and generates a distributed control strategy for the intelligent connected vehicle, including the expected speed, expected acceleration and expected lane of each vehicle. These control instructions will serve as the input of the simulation environment to drive the movement of the virtual vehicle. In the simulation environment, various test scenarios can be set, such as convoy formation, lane change, collision avoidance, etc., to evaluate the performance of the model.

[0088] In this embodiment, a multi-level, high-fidelity virtual simulation environment for intelligent connected vehicles is constructed, covering modeling and simulation functions from macro traffic flow to micro components, providing a platform support for the development and testing of vehicle intelligent algorithms; information interaction and collaborative control between vehicles and vehicles and between vehicles and infrastructure are realized, breaking through the "single-vehicle intelligence" bottleneck of traditional autonomous driving, and realizing the coordination and optimization of vehicle groups; an end-to-end deep learning method is used to realize vehicle intelligent decision-making for complex traffic scenarios, which has stronger environmental adaptability and robustness than traditional rules or optimization methods.

[0089] In an optional embodiment,

[0090] The multi-agent collaborative control model is used to process the input vehicle status and road condition information to generate the distributed control strategy of the intelligent connected vehicle, including:

[0091] Constructing an environmental dynamics model, which receives historical states and action sequences as input, assigns weights to states and actions at different moments through an attention mechanism, uses an external memory mechanism to maintain memory units to store and retrieve long-term information, and performs convolution operations in the time and space dimensions through a spatiotemporal convolutional network to output predicted future states;

[0092] The constructed environmental dynamics model is trained using historical trajectory data. Based on the trained environmental dynamics model, a nonlinear model predictive controller is designed. The input is the current state and the predicted future state. The random MPC idea is introduced to simulate uncertainty and output a preliminary control sequence.

[0093] In the multi-agent collaborative control scenario, the nonlinear model predictive control problem is decomposed into multiple sub-problems, each sub-problem corresponds to an agent, each agent receives local state information, and performs collaborative optimization through a distributed MPC architecture to generate a preliminary distributed control strategy. By combining numerical optimization methods with heuristic optimization techniques, the optimization objectives and constraints are received, and the preliminary distributed control strategy is optimized through iterative search to obtain the optimized distributed control strategy, including the expected speed, expected acceleration and expected lane for each vehicle.

[0094] For example, an environmental dynamics model is first constructed. The model receives historical states and action sequences, analyzes and predicts environmental changes. In order to address the complexity of temporal and spatial information processing, the model uses an attention mechanism to assign different weights to states and actions at different times. This mechanism can help the model pay more attention to key actions at key time points and improve the accuracy of predictions.

[0095] At the same time, an external memory mechanism is introduced to maintain a long-term information storage and retrieval system. This mechanism helps the model access previous experience and data when processing current input, thereby more fully understanding the dynamic changes of the environment.

[0096] In order to process spatiotemporal information, the environmental dynamics model is also modeled through a spatiotemporal convolutional network. This network architecture allows the model to perform convolution operations in both the temporal and spatial dimensions, thereby more efficiently extracting information and outputting future state predictions. This prediction will provide an important reference for the subsequent control strategy design.

[0097] After the environmental dynamics model is built, it is then trained to achieve good predictive capabilities using historical trajectory data. Using the trained model, a nonlinear model predictive controller is designed to handle the current state and predicted future state. At this point, the controller needs to introduce a stochastic model predictive control (MPC) approach to deal with the uncertainty in the environment, which can generate a preliminary control sequence to guide the vehicle's behavior.

[0098] In the scenario of multi-agent collaborative control, the nonlinear model predictive control problem is decomposed into multiple sub-problems, so that each sub-problem corresponds to an independent agent. Each agent receives its local state information, and based on this information, collaborative optimization is performed through a distributed MPC architecture. This process enables the agents to communicate and coordinate with each other, thereby improving the overall control effect.

[0099] Finally, the initially generated distributed control strategy is optimized by combining numerical optimization methods with heuristic optimization techniques. At this stage, the system receives the optimization objectives and constraints, and gradually adjusts the strategy through iterative search to obtain the final optimized control strategy. This optimized strategy will provide each vehicle with guidance on the desired speed, acceleration, and lane, ensuring that the vehicle responds efficiently and safely in a dynamic traffic environment.

[0100] In this embodiment, the environmental dynamics model makes full use of the attention mechanism, external memory and spatiotemporal convolution to accurately capture environmental features and improve prediction accuracy. The nonlinear model predictive controller combines the idea of ​​stochastic MPC to improve the robustness and adaptability of control. The distributed MPC architecture realizes the collaborative optimization of multiple agents, reduces computational complexity and improves scalability. The combination of numerical optimization and heuristic search obtains high-quality control strategies while satisfying constraints. The overall solution can realize collaborative interaction between agents, improve the overall performance of the system, and has broad application prospects.

[0101] S102. Establish a distributed edge computing network for a group of intelligent connected vehicles, receive real-time data from on-board sensors and workshop communication data, perform local data processing and decision calculations through edge nodes, output initial vehicle control instructions, and deploy a self-organizing learning algorithm based on the initial vehicle control instructions and environmental data. Generate individual control strategies through dynamic weight adjustment and local rule updates, use individual control strategies and corresponding scene features as input, use blockchain technology to build a decentralized knowledge base, form a decision-making experience data set, and design a multi-level attention mechanism based on the decision-making experience data set and multimodal data of the current traffic scene to extract target information, and output scene understanding results and decision features.

[0102] The distributed edge computing network consists of multiple edge nodes, each of which can process local vehicle data and communicate with other nodes to exchange information. On-board sensors collect vehicle status data in real time, such as position, speed, acceleration, etc., and transmit the data to edge nodes through vehicle-to-vehicle communication. Edge nodes process and fuse local data and output initial vehicle control instructions, such as acceleration, deceleration, steering, etc.

[0103] Based on the initial vehicle control instructions and environmental data, the edge node deploys a self-organizing learning algorithm to generate a personalized vehicle control strategy through dynamic weight adjustment and local rule update. Self-organizing learning is a distributed machine learning paradigm in which each node learns autonomously based on local observations and interactions and continuously updates its own knowledge and strategies. Dynamic weight adjustment refers to dynamically adjusting the weights of different features to highlight key information based on the vehicle's state and environmental changes. Local rule update refers to updating the node's decision rules based on vehicle feedback and interaction to improve the adaptability of the strategy.

[0104] In order to make full use of the historical experience of the group, personalized control strategies and corresponding scene features are used as inputs, and blockchain technology is used to build a decentralized knowledge base. Blockchain is a distributed ledger technology that can record and share data securely and reliably. By putting personalized strategies and scene features on the chain, a globally shared decision-making experience data set is formed. The decentralized knowledge base can prevent single point failures and data tampering, and ensure the reliability and consistency of knowledge.

[0105] Based on the decision experience data set and the multimodal data of the current traffic scene, a multi-level attention mechanism is designed to extract key scene understanding and decision features. Multimodal data includes sensor data such as video, radar, and lidar, as well as environmental data such as high-precision maps and weather. The multi-level attention mechanism extracts target information from local to global through hierarchical attention calculations, and outputs scene understanding results and decision features. The scene understanding results include information such as the location, speed, and intention of traffic participants such as vehicles, pedestrians, and obstacles. Decision features include the topological structure of the road, traffic signs and markings, and road priority.

[0106] In an optional embodiment,

[0107] Based on the decision-making experience dataset and the multimodal data of the current traffic scene, a multi-level attention mechanism is designed to extract target information, and the output scene understanding results and decision features include:

[0108] Build a multi-task learning framework, which includes a shared feature extractor based on convolutional neural networks and multi-layer attention mechanism, a scene understanding subtask module, and a decision feature extraction subtask module;

[0109] Receive a traffic scene data set including scene input data, scene understanding labels and decision labels; use the shared feature extractor to extract a common scene representation from the scene input data to obtain a shared feature, input the shared feature into a scene understanding subtask module, perform semantic segmentation, target detection and relationship recognition tasks, and obtain a scene understanding result; at the same time, input the shared feature into a decision feature extraction subtask module, perform behavior prediction and risk assessment tasks, and obtain a decision feature extraction result;

[0110] Based on the scene understanding results and scene understanding labels, the subtask loss functions of semantic segmentation, object detection and relationship recognition are calculated respectively. Based on the decision feature extraction results and decision labels, the subtask loss functions of behavior prediction and risk assessment are calculated respectively. A dynamic weight adjustment strategy is adopted, and the learning difficulty and convergence speed information of each subtask is input. The weight coefficients of the subtask loss functions are adaptively calculated, and the weighted subtask loss functions are summed to obtain the joint loss function.

[0111] The joint loss function is input into the optimization algorithm, and the adaptive moment estimation method with learning rate decay and gradient clipping is used to calculate the parameter gradient and update the parameters of the shared feature extractor, scene understanding subtask module and decision feature extraction subtask module. The iteration is repeated until the preset termination condition is met, and the trained multi-task learning model is output.

[0112] Using the trained multi-task learning model, new traffic scene data is input and scene understanding results and decision feature extraction results are output simultaneously.

[0113] Exemplarily, an end-to-end network structure consisting of a shared feature extractor, a scene understanding subtask module, and a decision feature extraction subtask module is first designed.

[0114] The shared feature extractor uses a combination of convolutional neural networks and multi-layer attention mechanisms to learn high-level semantic feature representations from raw traffic scene input data. Convolutional neural networks have the characteristics of local receptive fields and parameter sharing, and can efficiently capture low-level features such as textures and edges in images; while the attention mechanism can adaptively focus on key information in the scene by weighted aggregation of feature responses from different regions or channels. By stacking multiple convolutional layers and attention layers, the shared feature extractor gradually maps the scene data to a compact feature space to obtain shared features rich in semantic information.

[0115] Based on the shared features, the scene understanding subtask module further performs semantic segmentation, object detection, and relationship recognition tasks. Semantic segmentation aims to label each pixel of the image and generate a segmentation mask consistent with the input resolution. A common approach is to use an encoder-decoder structure, first extract multi-scale features through the encoder, and then use the decoder to gradually restore the spatial resolution and predict the category probability of each pixel. Object detection detects the target object of interest in the image and gives its category and location information. Most mainstream object detection methods are based on region candidates and feature pyramids. First, a series of dense candidate boxes are generated, and then the feature areas corresponding to the candidate boxes are classified and regressed. Relationship recognition is to infer the interactive relationship between objects, such as front and back positions, identity, etc., based on the detection of the target object. Graph neural network is a commonly used modeling tool that can effectively mine the association pattern between objects by aggregating neighborhood information at nodes.

[0116] For decision-making planning tasks, the decision feature extraction subtask module adopts a similar feature learning mechanism, but focuses more on the prediction of future trajectories and behaviors. Among them, the behavior prediction task is to predict the movement trajectory of traffic participants in the future based on the historical observation sequence. An effective idea is to model it as a sequence generation problem, input past observations, and output future coordinate sequences. Recurrent neural networks and their variants such as LSTM, GRU, etc. are widely used in trajectory prediction tasks due to their long and short-term memory capabilities. Risk assessment quantifies the safety and comfort of each predicted trajectory to assist the decision module in selecting the optimal driving plan. A scoring function can be designed to comprehensively consider factors such as the collision risk and acceleration changes of the trajectory to give a scalar risk value.

[0117] During the model training phase, a traffic scene dataset containing scene input data, scene understanding labels, and decision labels is constructed. The scene input data can be information collected by sensors such as vehicle-mounted cameras and lidars, or it can be a preprocessed raster representation. The scene understanding labels include pixel-level semantic segmentation truth values, object-level detection box coordinates and categories, and triple representations of object relationships. The decision labels are the position coordinate sequences and risk values ​​of traffic participants in the next few time steps. Using these annotated data, corresponding loss functions can be designed for different subtasks, such as cross entropy loss commonly used for semantic segmentation, focal loss commonly used for target detection, and mean square error for trajectory prediction.

[0118] In order to balance the learning progress of different tasks, a dynamic weight adjustment strategy is used to adaptively adjust the weight coefficient of each subtask loss function. Intuitively, if a task is more difficult to learn or converges slowly, we want to give it more attention and increase the weight coefficient of its loss function accordingly. One feasible approach is to design an indicator to measure the difficulty of the task, such as performance on the validation set, or gradient statistics of parameter updates. The learning status of each subtask can be evaluated regularly, and its weight coefficient can be dynamically calculated so that different tasks can achieve synchronous convergence during training.

[0119] After obtaining the weighted loss functions of each subtask, they are summed to form a unified joint loss function. This loss function is used as an indicator to measure the performance of the entire multi-task learning framework, and the optimization goal is to minimize the joint loss function. Considering the volatility of the loss function and parameter gradients during training, the adaptive moment estimation method Adam is used for optimization, and the learning rate decay and gradient clipping mechanisms are introduced to stabilize the training process and avoid the gradient explosion problem. By setting a reasonable batch size and number of iterations, forward propagation and back propagation can be repeatedly performed on the training set to update the parameters in the shared feature extractor, scene understanding subtask module, and decision feature extraction subtask module until the preset convergence conditions are met.

[0120] After training, a multi-task learning model integrating scene understanding and decision planning functions is obtained. When faced with new traffic scene inputs, a common scene representation is first obtained through a shared feature extractor, and then input into the scene understanding subtask module and the decision feature extraction subtask module respectively, and a series of prediction results including semantic segmentation masks, target detection boxes, object relationships, future trajectories, and risk scores are output at the same time. These results can provide comprehensive environmental cognition and risk analysis for downstream decision control modules, and assist unmanned vehicles in making safe and intelligent behavioral decisions.

[0121] In this embodiment, through a shared feature learning mechanism, the two major tasks can promote each other and optimize collaboratively. At the same time, the end-to-end learning paradigm also greatly simplifies the design and development process of the unmanned driving system, reducing the burden of manual parameter adjustment and feature engineering. In addition, the framework uses a dynamic weight adjustment strategy to balance the learning progress of different tasks, which can alleviate the optimization problems in multi-task learning and improve the model convergence speed and generalization performance. Finally, by introducing advanced technologies such as attention mechanisms and graph neural networks, the framework can adaptively focus on key information in the scene and mine semantic associations between objects, thereby obtaining a more accurate and comprehensive understanding of the environment.

[0122] S103. Utilize the scene understanding results and decision-making features to construct a dynamic scene graph, output multi-step prediction results and corresponding decision suggestions based on the dynamic scene graph and historical trajectory data, take the multi-step prediction results, decision suggestions and local state information of each vehicle as input, design a swarm intelligence emergence mechanism, generate coordinated behavior instructions that can optimize the overall traffic flow, integrate the coordinated behavior instructions, individual control strategies and predictive decision results, generate final vehicle control instructions through a multi-objective decision optimization algorithm, transmit the final vehicle control instructions to the vehicle execution system, and realize dynamic adaptive swarm collaborative control of intelligent connected vehicles.

[0123] The dynamic scene graph is constructed by using the scene understanding results and decision features combined with static high-precision map information. Static high-precision maps contain accurate information about fixed facilities such as roads, lanes, and traffic signs. Based on the static map, the dynamic scene graph adds the real-time status of dynamic traffic participants to generate a structured scene representation containing spatiotemporal relationships. The dynamic scene graph can provide global environmental perception and semantic understanding, providing information support for intelligent decision-making.

[0124] Based on the dynamic scene graph and historical trajectory data, the scene prediction results for multiple time steps in the future and the corresponding decision suggestions are output. Scene prediction predicts the future movement trajectory of traffic participants by learning their movement patterns and interaction rules. Decision suggestions propose behavioral strategies for multiple time steps in the future, such as acceleration, deceleration, lane change, etc., based on the prediction results and traffic rules. Multi-step prediction and decision-making can improve the predictability and initiative of vehicles and reduce the impact of emergencies.

[0125] Taking multi-step prediction results, decision suggestions, and local state information of each vehicle as input, a swarm intelligence emergence mechanism is designed to generate swarm coordination behavior instructions for overall traffic flow optimization. The swarm intelligence emergence mechanism stimulates the emergence characteristics of the entire system by designing local interaction rules for individual vehicles, resulting in a global optimization effect. Swarm coordination behavior instructions include the relative position, relative speed, and formation form between vehicles, with the goal of improving road traffic efficiency and safety.

[0126] Finally, the group coordination behavior instructions, individual control strategies and predictive decision results are integrated, and real-time vehicle control instructions, such as steering angle and acceleration, are generated through a multi-objective decision optimization algorithm. Multi-objective decision optimization requires balancing different performance requirements, such as safety, efficiency, and comfort, to select the optimal control instructions. The final control instructions control the chassis motor, steering system, etc. through the on-board execution system to achieve vehicle motion control. Throughout the process, the vehicle continuously iterates and optimizes the control strategy based on real-time feedback to achieve active adaptation to the dynamic traffic environment.

[0127] In an optional embodiment,

[0128] Using the scene understanding results and decision features, a dynamic scene graph is constructed. Based on the dynamic scene graph and historical trajectory data, multi-step prediction results and corresponding decision suggestions are output, including:

[0129] Construct a multi-level traffic rules knowledge graph, organize traffic concepts, relationships and constraints in the form of ontology, use description logic to realize the formal representation and reasoning of traffic rules, and obtain a traffic rules knowledge base;

[0130] Construct a multimodal encoding-decoding network structure. The encoder uses a graph convolutional neural network to extract features and aggregate information from dynamic scene graphs. It learns the hidden representation of nodes through multi-layer graph convolution operations and captures the interactive relationship between nodes. The decoder uses a long short-term memory network or a gated recurrent unit regression to generate vehicle trajectories. At each time step, the vehicle position at the next moment is predicted based on the trajectory points generated in the previous time step and the scene features output by the encoder.

[0131] In the constructed encoder, a multi-head self-attention mechanism is used to calculate the mutual influence between vehicles. Each attention head focuses on different interaction patterns to obtain multi-scale dependencies. In the decoder, a masked self-attention mechanism is used to generate future trajectories. The mask operation ensures that only information from previous moments can be accessed when predicting the trajectory points at the current moment, thus obtaining a preliminary trajectory prediction result.

[0132] The trajectory prediction is modeled as a learning problem of conditional probability distribution. The initial trajectory prediction results are used to train the model parameters by maximizing the conditional log-likelihood function. The stochastic gradient descent algorithm with adaptive learning rate is used to minimize the difference between the predicted trajectory and the true trajectory. In the inference stage, the Monte Carlo sampling method is used to generate multiple predicted trajectories according to the learned conditional probability distribution to obtain the trajectory prediction probability distribution.

[0133] Based on the obtained trajectory prediction probability distribution, the Monte Carlo simulation method is used to calculate the collision probability and minimum safe distance between vehicles, and the smoothness of vehicle driving is evaluated through the geometric characteristics and time characteristics of the trajectory. The trajectory prediction results and the road network topology are used to construct a time-varying traffic flow network, calculate the expected time for vehicles to arrive at their destinations and the impact of path selection on overall traffic efficiency, and obtain scenario evaluation results;

[0134] The obtained trajectory prediction probability distribution and scenario evaluation results are combined with the constructed traffic rules knowledge base, and decision recommendations are generated through probabilistic symbolic reasoning. Combined with the current state of the vehicle, trajectory prediction probability distribution and rule constraints in the knowledge graph, multi-step prediction results and corresponding decision recommendations are finally output.

[0135] For example, first, a multi-level traffic rules knowledge graph is constructed, and traffic concepts, relationships and constraints are organized in the form of ontology. The description logic is used to realize the formal representation and reasoning of traffic rules, and the traffic rules knowledge base is obtained. Among them, ontology is a formal method for knowledge representation, which can express the hierarchical relationship and logical relationship between concepts. Description logic is a logical language used for ontology representation, which can define concepts, roles and individuals, and describe the relationship and constraints between them through logical axioms. Through ontology and description logic, the concepts, relationships and constraints in traffic rules can be structured and a multi-level traffic rules knowledge graph can be constructed. For example, concepts such as "vehicle" and "road" and relationships such as "driving on" and "turning" can be defined, and the constraints between them can be described through axioms, such as "vehicles driving on roads" and "vehicles cannot pass when the light is red".

[0136] Then, a multimodal encoder-decoder network structure is designed for vehicle trajectory prediction. The encoder uses a graph convolutional neural network to extract features and aggregate information from the dynamic scene graph. The graph convolutional neural network is a deep learning model that processes graph-structured data. It propagates and aggregates features on the graph by defining convolution operations. The dynamic scene graph represents vehicles, roads, and the relationships between them in the form of a graph. Nodes represent vehicles and roads, and edges represent the interactions between them. Through multi-layer graph convolution operations, the hidden representation of nodes can be learned to capture the interaction patterns between vehicles and between vehicles and roads. The decoder uses a long short-term memory network (LSTM) or a gated recurrent unit (GRU) to regress and generate vehicle trajectories. LSTM and GRU are two commonly used recurrent neural network structures that can model long-term dependencies of time series data. At each time step, the position of the vehicle at the next moment is predicted based on the trajectory points generated at the previous time step and the scene features output by the encoder. Through this encoder-decoder architecture, the sequence generation of vehicle trajectories can be achieved.

[0137] In the constructed encoder, a multi-head self-attention mechanism is used to calculate the mutual influence between vehicles. The self-attention mechanism calculates the similarity between elements in the sequence and obtains the dependency relationship between each element and other elements, which can capture global context information. The multi-head self-attention mechanism uses multiple attention heads for parallel calculation, and each head focuses on different feature subspaces, which can model multi-scale interaction patterns. Through the multi-head self-attention mechanism, the dependency relationship between vehicles at different scales can be obtained, and the mutual influence between vehicles can be more comprehensively characterized. In the decoder, the masked self-attention mechanism is used to generate future trajectories. When calculating the attention score at the current moment, the masked self-attention shields the information of the future moment through the mask operation, ensuring that only the trajectory information of the previous moment can be accessed when predicting the current position, avoiding information leakage. Through the masked self-attention mechanism, trajectory prediction results that conform to the temporal causal relationship can be obtained.

[0138] Next, the trajectory prediction problem is modeled as a learning problem of conditional probability distribution. Given the historical trajectory and scene information, the conditional probability distribution of the future trajectory is learned. The model parameters are trained by maximizing the conditional log-likelihood function using the preliminary trajectory prediction results obtained by the encoder-decoder network. The conditional log-likelihood function describes the logarithmic probability of the occurrence of the true trajectory sequence under given conditions. By maximizing this likelihood function, the predicted trajectory distribution can be made as close to the true trajectory distribution as possible. At the same time, the model is optimized using the stochastic gradient descent algorithm with adaptive learning rate. By dynamically adjusting the learning rate, convergence is accelerated and local optimality is avoided. In the inference stage, the Monte Carlo sampling method is used to generate multiple possible future trajectories based on the learned conditional probability distribution to obtain the probability distribution of trajectory prediction. Monte Carlo sampling samples multiple trajectories from the learned distribution by random sampling, which can approximately represent the uncertainty and diversity of the trajectory.

[0139] Based on the predicted probability distribution of trajectories, the Monte Carlo simulation method is used to evaluate the safety and traffic efficiency of the scenario. By sampling and simulating multiple predicted trajectories, the collision probability and minimum safe distance between vehicles can be calculated. The collision probability reflects the possibility of a vehicle collision, and the minimum safe distance indicates the minimum distance that needs to be maintained between vehicles to avoid a collision. By analyzing the geometric characteristics (such as curvature, acceleration) and time characteristics (such as speed, time interval) of the trajectory, the smoothness and comfort of vehicle driving can be evaluated. Using the predicted vehicle trajectories and road topology, a time-varying traffic flow network is constructed, and the expected time for the vehicle to arrive at the destination and the impact of different path choices on the overall traffic efficiency are calculated through graph theory algorithms. By comprehensively considering factors such as safety, smoothness and traffic efficiency, a comprehensive evaluation of traffic scenarios can be performed.

[0140] Finally, the trajectory prediction probability distribution, scenario evaluation results and traffic rules knowledge base are combined to generate decision recommendations through probabilistic symbolic reasoning. Probabilistic symbolic reasoning is a reasoning method that combines probabilistic graphical models with symbolic logic, which can perform logical reasoning under uncertain conditions. By representing information such as vehicle status, trajectory prediction results and traffic rules as variables and conditional probabilities in probabilistic graphical models, and using symbolic logic to describe the constraint relationship between them, decision recommendations that comply with traffic rules and take uncertainty into account can be obtained. The decision recommendation includes the vehicle's driving strategy and the corresponding confidence level, which indicates the credibility of the decision. Through probabilistic symbolic reasoning, a vehicle decision plan that takes into account traffic rules, safety and efficiency can be obtained.

[0141] In this embodiment, the knowledge graph is used to perform a structured representation of traffic rules so that the decision-making process complies with traffic regulations and constraints. The vehicle trajectory is predicted by a deep learning method, and the interaction between vehicles and scene information are considered to improve the accuracy of trajectory prediction. The multi-head self-attention mechanism and the masked self-attention mechanism are used to model the complex dependencies between vehicles and generate trajectory prediction results that conform to time causality. The trajectory prediction problem is modeled as a learning problem of conditional probability distribution. The model is optimized by maximizing the conditional log-likelihood and stochastic gradient descent to obtain the trajectory prediction distribution in a probabilistic sense. Based on the trajectory prediction distribution, the traffic scene is comprehensively evaluated by comprehensively considering multiple factors such as safety, smoothness and traffic efficiency. Probabilistic symbolic reasoning is used to combine knowledge graphs, trajectory prediction and scene evaluation to generate vehicle decision recommendations that take into account traffic rules and uncertainty.

[0142] In an optional embodiment,

[0143] Using trajectory prediction results and road network topology, we construct a time-varying traffic flow network, calculate the expected time for vehicles to arrive at their destinations and the impact of path selection on overall traffic efficiency, and obtain scenario evaluation results including:

[0144] The road network is divided into several traffic zones, each of which corresponds to a node in the network. According to the topological structure and geometric characteristics of the road network, the spectral clustering algorithm is used to divide the road network, the similarity matrix between nodes is defined, the matrix is ​​spectrally decomposed, the eigenvectors corresponding to the first k largest eigenvalues ​​are taken to form a matrix, and each row of the matrix is ​​clustered to obtain the traffic zone division result of the road network.

[0145] The trajectory prediction model is used to predict the vehicle trajectory in each time slice. For each cell, the number of vehicles falling into the current traffic cell is counted to obtain the traffic flow of each traffic cell in each time slice.

[0146] According to the road network topology, the number of roads connecting two adjacent traffic zones is counted. In each time slice, the traffic flow transferred from one traffic zone to another adjacent traffic zone is calculated. If the transfer probability of vehicles on each connecting road section is equal, the transfer probability between adjacent traffic zones is calculated based on the traffic flow and the number of connecting road sections.

[0147] According to the traffic zone division results, traffic flow in traffic zones and the transfer probability between traffic zones, a time-varying traffic flow network is constructed, in which nodes are traffic zones and node attributes are dynamic traffic flow in traffic zones. For each vehicle, the shortest time path is calculated on the time-varying traffic flow network using a dynamic programming algorithm based on the traffic zone where its starting position is located and the traffic zone where its destination is located. For each vehicle's actual driving path, the actual time it takes to reach the destination is calculated. The difference between the actual arrival time and the shortest expected time is defined as the time difference, and the overlap between the actual driving path and the shortest time path is defined as the path deviation.

[0148] The time differences and path deviations of all vehicles are weighted averaged respectively, and the weighted average of the time differences and the weighted average of the path deviations are weighted summed according to the preset trade-off parameters to obtain the efficiency evaluation value of the traffic scenario and the scenario evaluation result.

[0149] Exemplarily, the entire road network is first divided into several traffic zones, each of which corresponds to a node in the time-varying traffic flow network. When dividing the traffic zones, the topological structure and geometric characteristics of the road network are taken into consideration, and the spectral clustering algorithm is used to divide the road network. Specifically, the method defines a similarity matrix between road nodes, performs spectral decomposition on the similarity matrix, takes the eigenvectors corresponding to the first k largest eigenvalues ​​to form a feature matrix, and then performs K-means clustering on each row of the feature matrix, and finally obtains the traffic zone division result of the road network. This division method makes full use of the topological and geometric information of the road network and can obtain a reasonable traffic zone division.

[0150] After obtaining the traffic zone division, this method uses the trajectory prediction model to predict the vehicle trajectory in each time slice and obtain the predicted position of each vehicle in each time slice. Then, based on the predicted position, for each traffic zone, the number of vehicles falling into the traffic zone is counted to obtain the traffic flow of each traffic zone in each time slice. This step realizes the association of vehicle trajectory data with traffic zones, and obtains time series data reflecting the dynamic changes of traffic zone flow.

[0151] Next, the method counts the number of roads connecting two adjacent traffic cells according to the road network topology. In each time slice, the traffic flow from one traffic cell to another adjacent traffic cell is calculated. Assuming that the transfer probability of vehicles on each connecting road segment is equal, the transfer probability between adjacent traffic cells can be calculated based on the traffic flow and the number of connecting road segments. This step obtains the dynamic transfer relationship between traffic cells and describes the transfer law of traffic flow between traffic cells.

[0152] Based on the above traffic zone division results, traffic zone traffic flow and traffic zone transfer probability, this method constructs a time-varying traffic flow network. In this network, nodes represent traffic zones, node attributes are dynamic traffic flows within traffic zones, edges represent the connection relationship between traffic zones, and edge weights are the transfer probabilities between traffic zones. The time-varying traffic flow network integrates the road network structure, traffic flow dynamics and traffic flow transfer rules, providing a basis for further analyzing the efficiency of the transportation system.

[0153] After constructing the time-varying traffic flow network, the method further proposes an evaluation scheme for the efficiency of traffic scenarios. For each vehicle, the shortest time path is calculated on the time-varying traffic flow network using a dynamic programming algorithm based on the traffic zone where its starting position is located and the traffic zone where its destination is located. By defining the state variable as the shortest expected time for the vehicle to reach the destination traffic zone from the current traffic zone, establishing the state transfer equation and solving it, the shortest time path of the vehicle can be obtained. This step obtains the optimal path of the vehicle in the traffic network, laying the foundation for evaluating the efficiency of the actual path.

[0154] Next, for each vehicle's actual driving path, calculate the actual time it takes to reach the destination. The difference between the actual arrival time and the shortest expected time is defined as the time difference, and the overlap between the actual driving path and the shortest time path is defined as the path deviation. The time difference reflects the degree of deviation between the vehicle's actual driving time and the ideal shortest time, and the path deviation reflects the degree of difference between the vehicle's actual path selection and the optimal path. These two indicators characterize the efficiency of vehicle driving from the perspective of time and path.

[0155] Finally, the method performs weighted average of the time differences and path deviations of all vehicles, and weights and sums the weighted average of the time differences and the weighted average of the path deviations according to the preset trade-off parameters to obtain the efficiency evaluation value of the traffic scene. The efficiency evaluation value comprehensively considers the time differences and path deviations of all vehicles, and obtains the overall efficiency level of the entire traffic scene by weighted average. The smaller the efficiency evaluation value, the higher the overall efficiency of the traffic scene, and the vehicle can reach the destination faster and more accurately.

[0156] Assume that a city's road network consists of 1,000 road segment nodes, which are divided into 20 traffic zones by spectral clustering algorithm. Within one hour during the morning rush hour, the trajectory prediction model is used to predict the trajectory of 10,000 vehicles, and the predicted position of each vehicle in each time slice is obtained. Then, the number of vehicles in each traffic zone is counted to obtain the traffic flow of the 20 traffic zones in each time slice. According to the road network topology, the number of roads connecting adjacent traffic zones is counted, and the traffic flow and transition probability between traffic zones are calculated. A time-varying traffic flow network is constructed, which contains 20 nodes and several edges with strip weights. For a vehicle starting from traffic zone A and going to traffic zone B, a dynamic programming algorithm is applied on the time-varying traffic flow network to obtain its shortest time path, which is compared with the actual driving path, and the time difference and path deviation are calculated. The time difference and path deviation of all vehicles are weighted averaged to obtain the traffic efficiency evaluation value of the entire morning rush hour, so as to judge the traffic conditions of this period.

[0157] In this embodiment, the road network structure and vehicle trajectory data are fully utilized to construct a dynamically evolving traffic flow network model to characterize the spatiotemporal distribution and transfer rules of traffic flow. A quantitative evaluation index for traffic scene efficiency is proposed to judge the efficiency of vehicle travel from the two perspectives of time difference and path deviation, providing a basis for traffic management and optimization. The overall traffic efficiency evaluation value is obtained by weighted average, which comprehensively considers the travel efficiency of all vehicles and can objectively reflect the operating status of the entire traffic scene. The evaluation results can provide decision support for traffic planning, traffic control, and travel guidance, and improve the overall efficiency of the traffic system by optimizing the traffic network and guiding vehicle path selection.

[0158] In an optional embodiment,

[0159] The final vehicle control command is generated through a multi-objective decision optimization algorithm, and the final vehicle control command is transmitted to the vehicle execution system to realize the dynamic adaptive group cooperative control of intelligent networked vehicles, including:

[0160] Establish a vehicle-cloud communication architecture and build a multi-level environment model in the cloud, including a static map layer, a dynamic prediction layer, and a semantic understanding layer. The static map layer models unchanging elements such as roads and lane lines, the dynamic prediction layer predicts the future movement of traffic participants based on physical motion models and trajectory history, and the semantic understanding layer analyzes the overall characteristics of traffic scenes.

[0161] Based on the environmental model, the optimal control strategy is solved in the cloud using a pre-built decision model to obtain a desired state sequence of the vehicle;

[0162] The desired state sequence is sent to the vehicle side, and fine-tuned by executing the optimization module on the vehicle side to generate coordinated behavior instructions and individual control strategies, integrate the coordinated behavior instructions, individual control strategies and desired state sequence, and generate final vehicle control instructions through a multi-objective decision optimization algorithm, and transmit the final vehicle control instructions to the vehicle execution system;

[0163] The actual execution results and environmental feedback of the vehicle execution system are recorded and used as local training data. The decision model is parameterized into a neural network. The vehicle uses the local training data to update the model parameters and upload the parameter gradients to the cloud. The cloud aggregates the average parameter gradients to update the global decision model and distributes it to all vehicles to form an optimized decision model. The optimized decision model is used to repeatedly execute the vehicle control instruction generation step to realize dynamic adaptive group collaborative control of intelligent connected vehicles.

[0164] For example, we first use the powerful computing and storage capabilities of the cloud to build a multi-level environmental model, including a static map layer, a dynamic prediction layer, and a semantic understanding layer. The static map layer models unchanging elements such as roads and lane lines to provide basic navigation information for vehicles. The dynamic prediction layer predicts the future movement of surrounding traffic participants based on physical motion models and historical trajectory data, allowing vehicles to respond in advance. The semantic understanding layer analyzes the overall characteristics of the traffic scene, understands the current traffic conditions, and provides higher-level information for decision-making.

[0165] After building the environment model in the cloud, the pre-trained decision model is used to solve the optimal control strategy. The decision model can be an intelligent model built based on reinforcement learning, deep learning and other technologies. It is trained with a large amount of simulation and real vehicle data and can make the best decision in a complex and changing traffic environment. The cloud decision model solves a series of expected vehicle states based on the current environmental state, indicating the speed, acceleration, steering and other control quantities that the vehicle should achieve in the future.

[0166] The desired state sequence generated in the cloud is sent to the vehicle through vehicle-to-cloud communication. After the vehicle receives the desired state sequence, it is first fine-tuned through the execution optimization module. The execution optimization module takes into account the vehicle's own dynamic characteristics and constraints, refines the desired state sent from the cloud, and generates practical behavior instructions and individual control strategies. Behavioral instructions can be high-level instructions such as lane changing, acceleration, and deceleration, while individual control strategies are direct controls on actuators such as throttles, brakes, and steering. The execution optimization module integrates behavioral instructions, individual control strategies, and desired state sequences, and uses a multi-objective decision optimization algorithm to generate the final control instructions while meeting multiple goals such as safety, comfort, and efficiency.

[0167] The vehicle execution system drives the throttle, brake, steering and other actuators according to the final control instructions received, so that the vehicle can make corresponding motion control. At the same time, the on-board sensors record the vehicle's driving status and environmental feedback in real time, such as speed, acceleration, surrounding obstacles and other information, to form local training data. These local training data are used for real-time updates of the decision model on the vehicle side. The decision model is parameterized in the form of a neural network. The vehicle side uses local training data to update the parameters of the neural network using incremental learning, and uploads the parameter gradients to the cloud. The cloud collects the parameter gradients uploaded by all vehicles, aggregates and averages them, updates the global decision model, and obtains a decision model that adapts to the latest traffic environment. The updated global decision model is then distributed to each vehicle, so that the vehicle side obtains the optimized decision model. This process is iterated continuously, so that the decision model can continue to learn and evolve, and continuously improve the vehicle's intelligent driving capabilities.

[0168] In this embodiment, through the multi-level environmental model constructed by vehicle-cloud collaboration, the vehicle can obtain a more comprehensive and long-term environmental perception, predict future changes in traffic conditions, and provide a reliable basis for decision-making. The decision model trained in the cloud can handle complex traffic scenarios and make the best decision. Through continuous online learning, the decision model can adapt to the ever-changing traffic environment, making the vehicle adaptive. Through vehicle-cloud-vehicle information interaction, vehicles can share perception information and decision-making experience, realize collaborative control of vehicle groups, and improve overall traffic efficiency and safety. Offloading computationally intensive tasks such as decision model training and environmental modeling to the cloud reduces the computing pressure of the vehicle system and reduces the cost of vehicle-mounted equipment. The cloud can generate personalized decision models based on the characteristics and driving preferences of different vehicles to provide users with a differentiated driving experience.

[0169] Figure 2 FIG. 1 is a schematic diagram of a structure of a dynamics simulation system for an intelligent connected vehicle based on a physical engine according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0170] The first unit is used to obtain vehicle parameters and road condition information of intelligent connected vehicles, and based on the vehicle parameters and road condition information, construct a multi-level physical engine simulation environment including macro traffic flow, meso vehicle dynamics and micro components. In the multi-level physical engine simulation environment, a multi-agent collaborative control model of intelligent connected vehicles is constructed by using graph neural networks and recurrent neural networks and deployed to the multi-level physical engine simulation environment, and the group behavior of intelligent connected vehicles in multiple scenarios is simulated. Based on the simulation results, the multi-agent collaborative control model is optimized by using an adaptive optimization algorithm, and a digital twin model is constructed to map the optimized multi-agent collaborative control model to the actual intelligent connected vehicle control system;

[0171] The second unit is used to establish a distributed edge computing network for a group of intelligent connected vehicles, receive real-time data from on-board sensors and vehicle-to-vehicle communication data, perform local data processing and decision-making calculations through edge nodes, output initial vehicle control instructions, deploy self-organizing learning algorithms based on initial vehicle control instructions and environmental data, generate individual control strategies through dynamic weight adjustment and local rule updates, use individual control strategies and corresponding scene features as input, use blockchain technology to build a decentralized knowledge base, form a decision experience data set, design a multi-level attention mechanism based on the decision experience data set and multimodal data of the current traffic scene to extract target information, and output scene understanding results and decision features;

[0172] The third unit is used to use the scene understanding results and decision-making features to construct a dynamic scene graph, output multi-step prediction results and corresponding decision suggestions based on the dynamic scene graph and historical trajectory data, take the multi-step prediction results, decision suggestions and local state information of each vehicle as input, design a group intelligence emergence mechanism, generate coordinated behavior instructions that can optimize the overall traffic flow, integrate the coordinated behavior instructions, individual control strategies and predictive decision results, generate final vehicle control instructions through a multi-objective decision optimization algorithm, transmit the final vehicle control instructions to the vehicle execution system, and realize dynamic adaptive group collaborative control of intelligent connected vehicles.

[0173] According to a third aspect of the embodiments of the present invention,

[0174] An electronic device is provided, comprising:

[0175] processor;

[0176] a memory for storing processor-executable instructions;

[0177] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0178] A fourth aspect of the embodiments of the present invention is:

[0179] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0180] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamics simulation of intelligent connected vehicles based on a physical engine, characterized in that: include: Obtain vehicle parameters and road condition information of the intelligent connected vehicle, and based on the vehicle parameters and road condition information, construct a multi-level physical engine simulation environment including macro traffic flow, meso vehicle dynamics, and micro components. In the multi-level physical engine simulation environment, use graph neural networks and recurrent neural networks to construct a multi-agent collaborative control model of the intelligent connected vehicle and deploy it to the multi-level physical engine simulation environment, and simulate the group behavior of intelligent connected vehicles in multiple scenarios. Based on the simulation results, use an adaptive optimization algorithm to optimize the multi-agent collaborative control model, and construct a digital twin model to map the optimized multi-agent collaborative control model to the actual intelligent connected vehicle control system; Establish a distributed edge computing network for the intelligent connected vehicle group, receive real-time data from on-board sensors and workshop communication data, perform local data processing and decision calculations through edge nodes, output initial vehicle control instructions, deploy self-organizing learning algorithms based on initial vehicle control instructions and environmental data, generate individual control strategies through dynamic weight adjustment and local rule updates, use individual control strategies and corresponding scene features as input, use blockchain technology to build a decentralized knowledge base, form a decision experience data set, design a multi-level attention mechanism based on the decision experience data set and the multimodal data of the current traffic scene to extract target information, and output scene understanding results and decision features; A dynamic scene graph is constructed by utilizing the scene understanding results and decision-making features. Based on the dynamic scene graph and historical trajectory data, multi-step prediction results and corresponding decision suggestions are output. The multi-step prediction results, decision suggestions and local state information of each vehicle are used as input to design a swarm intelligence emergence mechanism to generate coordinated behavior instructions that can optimize the overall traffic flow. The coordinated behavior instructions, individual control strategies and predictive decision results are integrated to generate final vehicle control instructions through a multi-objective decision optimization algorithm. The final vehicle control instructions are transmitted to the vehicle execution system to achieve dynamic adaptive swarm collaborative control of intelligent connected vehicles.

2. The method according to claim 1, characterized in that Based on the vehicle parameters and road condition information, a multi-level physical engine simulation environment including macro traffic flow, meso vehicle dynamics and micro components is constructed. In the multi-level physical engine simulation environment, a multi-agent collaborative control model of intelligent connected vehicles is constructed using graph neural networks and recurrent neural networks and deployed to the multi-level physical engine simulation environment. The group behaviors of intelligent connected vehicles in multiple scenarios are simulated, including: At the macroscopic traffic flow level, vehicles are abstracted as particles and roads are abstracted as channels. The interaction rules between particles are defined, and the differential equations of the traffic flow model are solved using numerical integration methods to obtain the macroscopic motion trajectory of each vehicle. At the mesoscopic vehicle dynamics level, vehicles are simplified as multi-body systems, and the vehicle dynamics differential equations are established to solve the vehicle's position, velocity, acceleration, and angular velocity, and simulate the interaction force between the vehicle and the road. At the microscopic component level, geometric modeling and physical modeling are performed on the key components of the vehicle, and the internal state changes and output responses of the simulated components are simulated through simulation, ultimately forming a multi-level physical engine simulation environment. In a multi-level physical engine simulation environment, graph neural networks are used to model the interaction between agents. Vehicles are abstracted as nodes in the graph, and wireless communication connections between vehicles are abstracted as edges in the graph. Node features are iteratively updated through a message passing mechanism. A recurrent neural network is used to model the sequential decision-making process of a single agent, unifying the state changes and decision iterations of the vehicle into the cyclic transmission of hidden states. The output of the graph neural network is used as the input of the recurrent neural network to form an end-to-end multi-agent collaborative control model. The multi-agent collaborative control model is deployed in a multi-level physical engine simulation environment. The vehicle status and road condition information in the simulation environment are used as the input of the model. The multi-agent collaborative control model is used to process the input vehicle status and road condition information to generate a distributed control strategy for intelligent connected vehicles, including the expected speed, expected acceleration, and expected lane for each vehicle. The distributed control strategy is used as the control input of a multi-level physical engine simulation environment to simulate the group behavior of intelligent connected vehicles in multiple scenarios, including platoon formation, lane change and collision avoidance.

3. The method according to claim 2, characterized in that The multi-agent collaborative control model is used to process the input vehicle status and road condition information to generate the distributed control strategy of the intelligent connected vehicle, including: Constructing an environmental dynamics model, which receives historical states and action sequences as input, assigns weights to states and actions at different moments through an attention mechanism, uses an external memory mechanism to maintain memory units to store and retrieve long-term information, and performs convolution operations in the time and space dimensions through a spatiotemporal convolutional network to output predicted future states; The constructed environmental dynamics model is trained using historical trajectory data. Based on the trained environmental dynamics model, a nonlinear model predictive controller is designed. The input is the current state and the predicted future state. The random MPC idea is introduced to simulate uncertainty and output a preliminary control sequence. In the multi-agent collaborative control scenario, the nonlinear model predictive control problem is decomposed into multiple sub-problems, each sub-problem corresponds to an agent, each agent receives local state information, and performs collaborative optimization through a distributed MPC architecture to generate a preliminary distributed control strategy. By combining numerical optimization methods with heuristic optimization techniques, the optimization objectives and constraints are received, and the preliminary distributed control strategy is optimized through iterative search to obtain the optimized distributed control strategy, including the expected speed, expected acceleration and expected lane for each vehicle.

4. The method according to claim 1, characterized in that Based on the decision-making experience dataset and the multimodal data of the current traffic scene, a multi-level attention mechanism is designed to extract target information, and the output scene understanding results and decision features include: Build a multi-task learning framework, which includes a shared feature extractor based on convolutional neural networks and multi-layer attention mechanism, a scene understanding subtask module, and a decision feature extraction subtask module; Receive a traffic scene data set including scene input data, scene understanding labels and decision labels; use the shared feature extractor to extract a common scene representation from the scene input data to obtain a shared feature, input the shared feature into a scene understanding subtask module, perform semantic segmentation, target detection and relationship recognition tasks, and obtain a scene understanding result; at the same time, input the shared feature into a decision feature extraction subtask module, perform behavior prediction and risk assessment tasks, and obtain a decision feature extraction result; Based on the scene understanding results and scene understanding labels, the subtask loss functions of semantic segmentation, object detection and relationship recognition are calculated respectively. Based on the decision feature extraction results and decision labels, the subtask loss functions of behavior prediction and risk assessment are calculated respectively. A dynamic weight adjustment strategy is adopted, and the learning difficulty and convergence speed information of each subtask is input. The weight coefficients of the subtask loss functions are adaptively calculated, and the weighted subtask loss functions are summed to obtain the joint loss function. The joint loss function is input into the optimization algorithm, and the adaptive moment estimation method with learning rate decay and gradient clipping is used to calculate the parameter gradient and update the parameters of the shared feature extractor, scene understanding subtask module and decision feature extraction subtask module. The iteration is repeated until the preset termination condition is met, and the trained multi-task learning model is output. Using the trained multi-task learning model, new traffic scene data is input and scene understanding results and decision feature extraction results are output simultaneously.

5. The method according to claim 1, characterized in that Using the scene understanding results and decision features, a dynamic scene graph is constructed. Based on the dynamic scene graph and historical trajectory data, multi-step prediction results and corresponding decision suggestions are output, including: Construct a multi-level traffic rules knowledge graph, organize traffic concepts, relationships and constraints in the form of ontology, use description logic to realize the formal representation and reasoning of traffic rules, and obtain a traffic rules knowledge base; Construct a multimodal encoding-decoding network structure. The encoder uses a graph convolutional neural network to extract features and aggregate information from dynamic scene graphs. It learns the hidden representation of nodes through multi-layer graph convolution operations and captures the interactive relationship between nodes. The decoder uses a long short-term memory network or a gated recurrent unit regression to generate vehicle trajectories. At each time step, the vehicle position at the next moment is predicted based on the trajectory points generated in the previous time step and the scene features output by the encoder. In the constructed encoder, a multi-head self-attention mechanism is used to calculate the mutual influence between vehicles. Each attention head focuses on different interaction patterns to obtain multi-scale dependencies. In the decoder, a masked self-attention mechanism is used to generate future trajectories. The mask operation ensures that only information from previous moments can be accessed when predicting the trajectory points at the current moment, thus obtaining a preliminary trajectory prediction result. The trajectory prediction is modeled as a learning problem of conditional probability distribution. The initial trajectory prediction results are used to train the model parameters by maximizing the conditional log-likelihood function. The stochastic gradient descent algorithm with adaptive learning rate is used to minimize the difference between the predicted trajectory and the true trajectory. In the inference stage, the Monte Carlo sampling method is used to generate multiple predicted trajectories according to the learned conditional probability distribution to obtain the trajectory prediction probability distribution. Based on the obtained trajectory prediction probability distribution, the Monte Carlo simulation method is used to calculate the collision probability and minimum safe distance between vehicles, and the smoothness of vehicle driving is evaluated through the geometric characteristics and time characteristics of the trajectory. The trajectory prediction results and the road network topology are used to construct a time-varying traffic flow network, calculate the expected time for vehicles to arrive at their destinations and the impact of path selection on overall traffic efficiency, and obtain scenario evaluation results; The obtained trajectory prediction probability distribution and scenario evaluation results are combined with the constructed traffic rules knowledge base, and decision recommendations are generated through probabilistic symbolic reasoning. Combined with the current state of the vehicle, trajectory prediction probability distribution and rule constraints in the knowledge graph, multi-step prediction results and corresponding decision recommendations are finally output.

6. The method according to claim 5, characterized in that Using trajectory prediction results and road network topology, we construct a time-varying traffic flow network, calculate the expected time for vehicles to arrive at their destinations and the impact of path selection on overall traffic efficiency, and obtain scenario evaluation results including: The road network is divided into several traffic zones, each of which corresponds to a node in the network. According to the topological structure and geometric characteristics of the road network, the spectral clustering algorithm is used to divide the road network, the similarity matrix between nodes is defined, the matrix is ​​spectrally decomposed, the eigenvectors corresponding to the first k largest eigenvalues ​​are taken to form a matrix, and each row of the matrix is ​​clustered to obtain the traffic zone division result of the road network. The trajectory prediction model is used to predict the vehicle trajectory in each time slice. For each cell, the number of vehicles falling into the current traffic cell is counted to obtain the traffic flow of each traffic cell in each time slice. According to the road network topology, the number of roads connecting two adjacent traffic zones is counted. In each time slice, the traffic flow transferred from one traffic zone to another adjacent traffic zone is calculated. If the transfer probability of vehicles on each connecting road section is equal, the transfer probability between adjacent traffic zones is calculated based on the traffic flow and the number of connecting road sections. According to the traffic zone division results, traffic flow in traffic zones and the transfer probability between traffic zones, a time-varying traffic flow network is constructed, in which nodes are traffic zones and node attributes are dynamic traffic flow in traffic zones. For each vehicle, the shortest time path is calculated on the time-varying traffic flow network using a dynamic programming algorithm based on the traffic zone where its starting position is located and the traffic zone where its destination is located. For each vehicle's actual driving path, the actual time it takes to reach the destination is calculated. The difference between the actual arrival time and the shortest expected time is defined as the time difference, and the overlap between the actual driving path and the shortest time path is defined as the path deviation. The time differences and path deviations of all vehicles are weighted averaged respectively, and the weighted average of the time differences and the weighted average of the path deviations are weighted summed according to the preset trade-off parameters to obtain the efficiency evaluation value of the traffic scenario and the scenario evaluation result.

7. The method according to claim 1, characterized in that The final vehicle control command is generated through a multi-objective decision optimization algorithm, and the final vehicle control command is transmitted to the vehicle execution system to realize the dynamic adaptive group cooperative control of intelligent networked vehicles, including: Establish a vehicle-cloud communication architecture and build a multi-level environment model in the cloud, including a static map layer, a dynamic prediction layer, and a semantic understanding layer. The static map layer models unchanging elements such as roads and lane lines, the dynamic prediction layer predicts the future movement of traffic participants based on physical motion models and trajectory history, and the semantic understanding layer analyzes the overall characteristics of traffic scenes. Based on the environmental model, the optimal control strategy is solved in the cloud using a pre-built decision model to obtain a desired state sequence of the vehicle; The desired state sequence is sent to the vehicle side, and fine-tuned by executing the optimization module on the vehicle side to generate coordinated behavior instructions and individual control strategies, integrate the coordinated behavior instructions, individual control strategies and desired state sequence, and generate final vehicle control instructions through a multi-objective decision optimization algorithm, and transmit the final vehicle control instructions to the vehicle execution system; The actual execution results and environmental feedback of the vehicle execution system are recorded and used as local training data. The decision model is parameterized into a neural network. The vehicle uses the local training data to update the model parameters and upload the parameter gradients to the cloud. The cloud aggregates the average parameter gradients to update the global decision model and distributes it to all vehicles to form an optimized decision model. The optimized decision model is used to repeatedly execute the vehicle control instruction generation step to realize dynamic adaptive group collaborative control of intelligent connected vehicles.

8. A physical engine-based intelligent connected vehicle dynamics simulation system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain vehicle parameters and road condition information of intelligent connected vehicles, and based on the vehicle parameters and road condition information, construct a multi-level physical engine simulation environment including macro traffic flow, meso vehicle dynamics and micro components. In the multi-level physical engine simulation environment, a multi-agent collaborative control model of intelligent connected vehicles is constructed by using graph neural networks and recurrent neural networks and deployed to the multi-level physical engine simulation environment, and the group behavior of intelligent connected vehicles in multiple scenarios is simulated. Based on the simulation results, the multi-agent collaborative control model is optimized by using an adaptive optimization algorithm, and a digital twin model is constructed to map the optimized multi-agent collaborative control model to the actual intelligent connected vehicle control system; The second unit is used to establish a distributed edge computing network for a group of intelligent connected vehicles, receive real-time data from on-board sensors and vehicle-to-vehicle communication data, perform local data processing and decision-making calculations through edge nodes, output initial vehicle control instructions, deploy self-organizing learning algorithms based on initial vehicle control instructions and environmental data, generate individual control strategies through dynamic weight adjustment and local rule updates, use individual control strategies and corresponding scene features as input, use blockchain technology to build a decentralized knowledge base, form a decision experience data set, design a multi-level attention mechanism based on the decision experience data set and multimodal data of the current traffic scene to extract target information, and output scene understanding results and decision features; The third unit is used to use the scene understanding results and decision-making features to construct a dynamic scene graph, output multi-step prediction results and corresponding decision suggestions based on the dynamic scene graph and historical trajectory data, take the multi-step prediction results, decision suggestions and local state information of each vehicle as input, design a group intelligence emergence mechanism, generate coordinated behavior instructions that can optimize the overall traffic flow, integrate the coordinated behavior instructions, individual control strategies and predictive decision results, generate final vehicle control instructions through a multi-objective decision optimization algorithm, transmit the final vehicle control instructions to the vehicle execution system, and realize dynamic adaptive group collaborative control of intelligent connected vehicles.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images