Vehicle formation dynamic modeling and simulation method based on physical coding deep learning

Abstract

The invention discloses a vehicle formation dynamic modeling and simulation method based on physical coding deep learning. The method comprises the steps that driving data of a vehicle formation are collected and preprocessed; based on the preprocessed data, constructing a parameter coding calculation graph of physical constraints; constructing a multi-scale trajectory feature learning network to learn behavior features of the vehicle formation from the preprocessed data; constructing a non-autoregression parameter decoder to map behavior characteristics of the vehicle formation into time-varying parameters required by a physically constrained parameter coding calculation graph, and iteratively calculating a future trajectory of the vehicle formation in combination with a pilot vehicle state; and in combination with the future trajectory of the vehicle formation, model training is performed through preprocessed data, and a formation dynamic model with both precision and physical interpretability is generated. According to the method, the defect that a traditional data driving model lacks physical interpretability and analyzability is overcome, meanwhile, the conservative limitation of a pure physical model is avoided, the real formation speed and spacing can be reproduced with high precision, and closed-loop analysis of formation stability is supported.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation technology, and specifically relates to a vehicle formation dynamic modeling and simulation method based on physical coding deep learning. Background Art

[0002] Vehicle platooning control is a core research area in intelligent transportation systems. Its goal is to improve traffic efficiency, reduce energy consumption, and enhance safety through cooperative control between vehicles. Traditional platooning dynamics modeling methods can be divided into two main categories: purely physical model-driven and data-driven. Physical model-driven methods construct vehicle dynamics equations based on classical car-following theory (e.g., the Pipes model and the optimal speed model) and achieve platooning stability through linear or nonlinear feedback control. These models offer clear physical interpretability but rely on strong assumptions (e.g., idealized driver behavior and homogeneous vehicle dynamics). This makes it difficult to replicate the complex interactions and heterogeneous driving behaviors found in real-world traffic scenarios, resulting in conservative simulation results and limited generalization. Data-driven methods (e.g., LSTM and Transformer) use deep learning to directly learn platooning dynamics from trajectory data. While these methods can capture complex nonlinear relationships, their lack of physical constraints can lead to outputs that violate fundamental physical laws (e.g., vehicle stability conditions), making them difficult to support closed-loop analysis of platooning safety. Furthermore, platooning dynamics involve multi-vehicle spatiotemporal coupling and disturbance propagation. Traditional methods that directly extend the single-vehicle model can amplify modeling errors. In terms of the spatial dimension, vehicle-to-vehicle interactions exhibit forward causality (perturbations from the lead vehicle propagate along the platoon), necessitating modeling of formation-level topological features. In terms of the temporal dimension, driving decisions exhibit historical dependencies (following behavior is influenced by the multi-step state of the preceding vehicle), necessitating the capture of multi-scale temporal characteristics. In recent years, physical-informed deep learning has attempted to integrate physical priors with data-driven models. However, existing PIDL frameworks often employ weak constraints on the residuals of physical equations through loss functions, failing to encode physical laws within the model architecture. This results in limited interpretability and difficulty ensuring dynamic stability. Furthermore, traditional platooning modeling methods lack theoretical gaps in disturbance propagation analysis and multi-vehicle coordinated response, failing to meet the requirements for high-precision, interpretable, and scalable modeling of vehicle platoons. Summary of the Invention

[0003] To address the deficiencies in the prior art, the present invention provides a vehicle formation dynamic modeling and simulation method based on physical coding deep learning, which overcomes the defects of traditional data-driven models that lack physical interpretability and analyzability, while avoiding the conservative limitations of pure physical models. It can reproduce the actual formation speed and spacing with high precision and support closed-loop analysis of formation stability.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] In a first aspect, a method for dynamic modeling and simulation of vehicle formations based on physical coding deep learning is provided, comprising: collecting driving data of the vehicle formation and preprocessing it to obtain preprocessed data; constructing a parameter coding calculation graph of physical constraints based on the preprocessed data; constructing a multi-scale trajectory feature learning network to learn the behavioral characteristics of the vehicle formation from the preprocessed data; constructing a non-autoregressive parameter decoder to map the behavioral characteristics of the vehicle formation into time-varying parameters required by the parameter coding calculation graph of physical constraints, and iteratively calculating the future trajectory of the vehicle formation in combination with the state of the leading vehicle; and performing model training on the preprocessed data in combination with the future trajectory of the vehicle formation to generate a formation dynamic model with both accuracy and physical interpretability.

[0006] Furthermore, the driving data of the vehicle formation meets the following conditions: (a) the distance between adjacent vehicles in the vehicle formation is no more than 120m, and the speed difference is no more than 3m / s; (b) the formation consists of at least 4 vehicles; (c) the formation operation state is maintained for no less than 10s.

[0007] Furthermore, a parameter encoding calculation graph of physical constraints is constructed, including: constructing the linear time-varying dynamic equations of the vehicle formation,

[0008]

[0009] in, Indicates the first n The car at time step t The acceleration of Is the first in the team n The car at time step t The distance between the front of the vehicle and the front of the vehicle, and are the mean values of speed and spacing in the historical time domain respectively; Indicates the speed of the nth vehicle and the speed of the n-1th vehicle The difference, =0 is the expected value of the speed difference, the speed feedback term <0, spacing feedback item >0, speed difference feedback item >0 is a necessary and sufficient condition for the equation to have physical meaning, and it is also a sufficient condition for the vehicle to meet the local stability condition. The parameters of the physical constraints are encoded in the calculation graph to be determined, and the physical constraints are realized through the Softplus activation function. represents the three-dimensional real space, represents the parameter vector of the dynamics model of the nth vehicle at time step t.

[0010] Furthermore, the parameter encoding computation graph satisfies the following physical constraints: (aa) parameter vector Each component of must satisfy: the speed feedback coefficient is negative, the spacing feedback coefficient is positive, and the relative speed feedback coefficient is positive; (bb) describes the relationship between vehicle acceleration and state deviation through a linear time-varying equation in the form of:

[0011]

[0012] in, is the state vector of the nth vehicle at time step t, is the expected state vector of the nth vehicle, defined by the historical state mean.

[0013] Furthermore, the multi-scale trajectory feature learning network includes: a vehicle-level feature extraction module, which is used to extract multi-scale time series features of vehicle historical trajectories from the pre-processed data using a Mamba state-space model, and generate vehicle-level historical features with probabilistic driving feature distribution through a variational encoder; a formation-level feature fusion module, which is used to capture the dynamic impact of the forward vehicle state on subsequent vehicles in the formation from the pre-processed data based on the fleet's unidirectional topological causal attention mechanism, and generate formation-level interaction features.

[0014] Furthermore, the non-autoregressive parameter decoder fuses formation-level interaction features with vehicle-level historical features through a temporal causal cross-attention mechanism, generates the parameters required for the physically constrained parameter encoding calculation graph, and adopts a multi-step parameter sharing strategy to reduce decoding complexity.

[0015] Furthermore, the future trajectory of the vehicle formation is calculated, including: (a1) using the Softplus activation function to activate the output parameters of the non-autoregressive parameter decoder to implement physical constraints on the predicted parameters; (a2) calculating the vehicle speed, spacing, and speed difference based on the vehicle formation state at the current time step; (a3) calculating the acceleration of the vehicles in the formation based on the predicted parameters and state corresponding to the current time step; (a4) updating the next step of the formation operation state based on the vehicle kinematics; (a5) repeating (a2) to (a4) to realize the future trajectory of the vehicle formation in the future time domain.

[0016] Furthermore, the model training includes: integrating the multi-scale trajectory feature learning network with the parameter encoding computation graph of physical constraints, and using trajectory slices for end-to-end training. Its loss function consists of the following parts: (b1) prediction loss: the mean square error of formation speed and spacing; (b2) KL divergence loss: the consistency of the output distribution of the constrained variational encoder and the prior distribution. The loss function expression is as follows:

[0017]

[0018] in, Is to control the KL divergence loss The hyperparameters of is the prediction loss and L is the total loss.

[0019] In a second aspect, a vehicle formation dynamic modeling and simulation device based on physical coding deep learning is provided, comprising a storage medium and a processor; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the vehicle formation dynamic modeling and simulation method based on physical coding deep learning described in the first aspect.

[0020] Compared with the existing technology, the beneficial effects achieved by the present invention are as follows: the present invention constructs a parameter encoding calculation graph of physical constraints; constructs a multi-scale trajectory feature learning network to learn the behavioral characteristics of vehicle formations from historical driving data; constructs a non-autoregressive parameter decoder to map the behavioral characteristics of the vehicle formation into the time-varying parameters required by the parameter encoding calculation graph of physical constraints, and iteratively calculates the future trajectory of the vehicle formation in combination with the state of the pilot vehicle; combines the future trajectory of the vehicle formation to perform model training and generate a formation dynamic model with both accuracy and physical interpretability; overcomes the defects of traditional data-driven models that lack physical interpretability and analyzability, while avoiding the conservative limitations of pure physical models, can reproduce the real formation speed and spacing with high precision, and supports closed-loop analysis of formation stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a schematic diagram of the main process of a vehicle platoon dynamic modeling and simulation method based on physical coding deep learning provided by an embodiment of the present invention;

[0022] Figure 2 It is a schematic diagram of the network architecture of an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] Example 1 A vehicle formation dynamic modeling and simulation method based on physical coding deep learning includes: collecting and preprocessing driving data of the vehicle formation to obtain preprocessed data; constructing a parameter coding calculation graph of physical constraints based on the preprocessed data; constructing a multi-scale trajectory feature learning network to learn the behavioral characteristics of the vehicle formation from the preprocessed data; constructing a non-autoregressive parameter decoder to map the behavioral characteristics of the vehicle formation into time-varying parameters required by the parameter coding calculation graph of physical constraints, and iteratively calculating the future trajectory of the vehicle formation in combination with the state of the pilot vehicle; and performing model training on the preprocessed data in combination with the future trajectory of the vehicle formation to generate a formation dynamic model with both accuracy and physical interpretability.

[0025] like Figure 1 、 Figure 2 As shown, the present invention embeds domain knowledge into the model architecture to construct a formation dynamic model with high precision and physical interpretability, providing a reliable theoretical tool for the modeling and simulation of vehicle formations.

[0026] Step 1: Collect the driving data of the vehicle formation and preprocess it to obtain preprocessed data.

[0027] Based on the natural driving data set, the continuous following relationship recognition algorithm is used to extract the trajectory data of real vehicle formations. The formation must meet the following conditions:

[0028] (1) The distance between adjacent vehicles in a formation shall not exceed 120 m, and the speed difference shall not exceed 3 m / s;

[0029] (2) The formation consists of at least 4 vehicles;

[0030] (3) Maintain the formation operation status for at least 10 seconds.

[0031] This paper uses the HIGH-SIM natural driving dataset to extract trajectory segments of platoons containing 4-10 vehicles. The required data features include: timestamp, position, velocity, acceleration, speed difference and spacing between adjacent vehicles in the platoon.

[0032] Data screening rules include:

[0033] Geometric constraints: Distance between adjacent vehicles , speed difference ;

[0034] Time constraints: formation duration ≥ 10 seconds, sampling frequency 10 Hz (total data points ≥ 100);

[0035] Outlier removal: Remove aggressive driving clips with an absolute acceleration value greater than 3m / s².

[0036] Data augmentation, including:

[0037] Time series shifting: intercept the trajectory with a sliding window (window length 40 steps, step length 1 step) to generate trajectory slice samples;

[0038] Speed perturbation: Randomly scale the overall speed of the formation by ±10% to simulate different traffic density scenarios and improve the model's feature expression ability and robustness.

[0039] Step 2: Based on the preprocessed data, construct a parameter-encoded computational graph (PeCG) of physical constraints.

[0040] Based on the generalized car-following model, the linear time-varying dynamic equations of vehicle platooning are constructed, the parameters required for vehicle platooning modeling are determined, and the model parameter signs are constrained by the Softplus activation function to meet the local stability conditions of the vehicles.

[0041] The modeling and simulation problem of vehicle platooning can be formulated as follows: construct the vehicle platooning dynamics equation, determine the required parameters based on the data obtained in step 1, and further predict the future trajectory.

[0042] The parameter encoding computational graph is designed based on professional knowledge in the transportation field. It constructs a mapping from parameters to future trajectories and realizes the gradient transfer between predicted trajectories and parameters through a differentiable design.

[0043] The vehicle platoon modeling problem is formulated as:

[0044] ,

[0045] Among them, P and F are recorded as the modeled historical trajectory length and the generated future trajectory length; N represents the number of following vehicles in the convoy, represents the state vector of the nth vehicle at time step t, represents the parameters of the vehicle platoon dynamics model at time step t, Represents the sequence from vehicle formation state to The mapping, Indicates passing Parameterized vehicle platooning dynamics model. represents the speed of the nth vehicle in the convoy at time step t, represents the headway between the nth vehicle in the convoy and its preceding vehicle at time step t, represents the speed difference between the nth vehicle in the convoy and its preceding vehicle at time step t, represents the state vector of vehicles 1 to N from time step tP to time step t, represents the state vector of vehicles 1 to N from time step t+1 to time step t+F, represents the state vector of vehicles 1 to N at time step t, It represents the speed of the leading vehicle (numbered 0) from time step t+1 to time step t+F.

[0046] The vehicle platoon modeling task can be divided into: 1) Design based on domain knowledge 2) Based on real platoon driving data Considering the scalability and flexibility of the platoon modeling framework, the linear time-varying dynamic equations of the vehicles are constructed with the help of the vehicle following domain knowledge:

[0047]

[0048] in, Indicates the first n The car at time step t The acceleration of Is the first in the team n The car at time step t The distance between the front of the vehicle and the front of the vehicle, and are the mean values of speed and spacing in the historical time domain, respectively, which represent the expected state, meaning that the vehicle hopes that the speed and spacing will remain consistent with the historical state as much as possible; Indicates the speed of the nth vehicle and the speed of the n-1th vehicle The difference, =0 is the expected value of the speed difference, which means that the vehicle hopes to keep the speed of the preceding vehicle as consistent as possible. <0, spacing feedback item >0, speed difference feedback item >0 is a necessary and sufficient condition for the equation to have physical meaning, and it is also a sufficient condition for the vehicle to meet the local stability condition. The physical constraints are implemented by the Softplus activation function for the undetermined parameters in the parameter encoding computation graph (PeCG) of the physical constraints. Represents a three-dimensional real space. The impact of the vehicle formation operation state on individual vehicles is reflected in To support the determination of the actual formation data , a physical constraint parameter encoding computation graph (PeCG) is designed for gradient transfer, and the physical constraints are implemented through the Softplus activation function to meet the local stability conditions of the vehicle.

[0049] The parameter-encoded computation graph satisfies the following physical constraints:

[0050] (aa) parameter vector Each component of must satisfy: the speed feedback coefficient is negative, the spacing feedback coefficient is positive, and the relative speed feedback coefficient is positive;

[0051] (bb) The relationship between vehicle acceleration and state deviation is described by a linear time-varying equation in the form of:

[0052]

[0053] in, represents the parameter vector of the nth vehicle dynamics model at time step t, is the expected state vector of the nth vehicle, defined by the historical state mean, is the expected state vector of the nth vehicle, defined by the historical state mean.

[0054] Step 3: Construct a multi-scale trajectory feature learning network (MTFLN) to learn the behavioral characteristics of the vehicle formation from the preprocessed data.

[0055] The multi-scale trajectory feature learning network is used to learn the vehicle platoon operation characteristics from the trajectories extracted in step 1, and extract important factors affecting the platoon operation from both the vehicle and platoon levels to provide valuable information for subsequent parameter prediction.

[0056] Multi-scale trajectory feature learning network, including:

[0057] The vehicle-level feature extraction module (VFL) uses the Mamba state-space model (SSM) to extract multi-scale temporal features of the vehicle's historical trajectory from preprocessed data and generates vehicle-level historical features with probabilistic driving feature distributions through a variational encoder (VE). The variational encoder (VE) is used to generate probabilistic driving feature distributions to model the randomness of driver behavior.

[0058] The vehicle-level feature extraction module is divided into an individual vehicle trajectory feature learning module and a feature uncertainty learning module. The former uses the Mamba state space model (SSM) to simulate the temporal changes of vehicle operation status and extract hierarchical features from the trajectory at different time scales. The calculation formula for this part is expressed as:

[0059]

[0060] in, is the output of the vehicle-level feature extraction module, representing the trajectory feature encoding, Represents the dimension of the model, and R represents the real number space; and Represents the learnable parameters and operators in the trajectory feature learning submodule; Indicates vehicle N From the time step tP To time step t The state vector of .

[0061] The feature uncertainty learning module introduces a variational encoder (VE) to generate a probabilistic driving feature distribution and model the randomness of driver behavior. The calculation formula for this part is expressed as:

[0062] , ,

[0063] in, is the predicted mean of the variational encoder output, is the prediction variance of the variational encoder output; driving feature distribution Depend on and Defined for the multivariate Gaussian distribution with mean and variance. Indicates that from the distribution Driving feature samples sampled in . , , , They represent the operators in the feature uncertainty learning submodule, the generated feature samples of vehicle n, the learnable weights and Construct the diagonal matrix.

[0064] The vehicle-level feature extraction module shares parameters among all vehicles in the formation and is applied independently.

[0065] The formation-level feature fusion module (PFL) is used to capture the dynamic impact of the forward vehicle state on subsequent vehicles in the formation from preprocessed data based on the unidirectional topological causal attention mechanism of the fleet, and generate formation-level interaction features.

[0066] The calculation formula of the PFL module is:

[0067]

[0068] in, represents the convoy trajectory features learned by the formation-level feature fusion module, represents the operator of the formation-level feature fusion module, represents the feature samples of vehicles 1 to N generated by the feature uncertainty learning submodule, Represents the learnable weights of the formation-level feature fusion module.

[0069] While it's difficult for a vehicle to observe the status of multiple vehicles ahead, in a tightly packed platoon, even small changes in the status of the preceding vehicle can be gradually amplified, significantly impacting vehicles further down the platoon. The PFL module utilizes a causal attention mechanism to learn formation-level interaction features based on the causal relationships of the platoon topology. The robust feature representation and selection capabilities of the attention mechanism enable the model to effectively quantify the intensity of vehicle interactions within the platoon and generate formation-level spatiotemporal interaction features. The introduction of the causal mask based on a unidirectional topological structure is based on the reasonable assumption that during platooning, the vast majority of vehicles make driving decisions based on the traffic conditions ahead. This assumption further enhances the model's applicability to platoons of varying sizes.

[0070] Step 4: Construct a non-autoregressive parameter decoder (NARP Decoder) to map the behavioral characteristics of the vehicle formation into the time-varying parameters required by the parameter encoding calculation graph of physical constraints, and iteratively calculate the future trajectory of the vehicle formation based on the state of the pilot vehicle.

[0071] The non-autoregressive parameter decoder is used to map the vehicle platoon operation features extracted in step 3 to the parameters required by PeCG in step 2, building a bridge from historical trajectories to parameters.

[0072] The formation-level features are fused with the vehicle-level historical features, and the non-autoregressive generation of the parameters required in the prediction time domain is achieved based on the temporal causal attention mechanism.

[0073] A multi-time-step parameter sharing strategy (parameters are updated every 5 steps) is adopted to enhance the model's ability to model the fleet's operating characteristics, avoid simple parameter fitting of the model, and reduce computational complexity.

[0074] The task of the NARP decoder is to decode the parameters required to predict PeCG in the time domain. For any vehicle n in the formation, The system incorporates multi-level spatiotemporal features of the vehicle-platoon, replicates them along the time dimension, and achieves time awareness through relative time encoding. Using these as queries, with vehicle trajectory feature encodings as keys and values, it applies temporal causal cross-attention to further extract trajectory features relevant to future driving decisions and parameters. To prevent the model from unrealistically leveraging distant future information for prediction, the decoder employs a non-autoregressive prediction method to accelerate model inference, improving online trajectory prediction and generation capabilities.

[0075]

[0076] in, 、 、 denote the output features, operators and learnable weights of the parameter decoder, respectively, represents the features of vehicle n corresponding to the formation-level feature fusion module, represents the trajectory features of vehicle n learned by the vehicle trajectory feature learning module.

[0077] Step 5: Combine the future trajectory of the vehicle formation and use the preprocessed data to perform model training to generate a formation dynamic model that is both accurate and physically interpretable.

[0078] Calculate the future trajectory of the vehicle formation, including:

[0079] (a1) The Softplus activation function is used to activate the output parameters of the non-autoregressive parameter decoder to implement physical constraints on the prediction parameters. The calculation formula is as follows:

[0080]

[0081] in, Represents the dynamic characteristics of the vehicle from 1 to N output by the parameter decoder.

[0082] (a2) Calculate the vehicle speed, spacing, and speed difference based on the vehicle formation state at the current time step. The speed difference calculation formula is as follows:

[0083]

[0084] in, represents the speed of the next k+1 steps predicted based on the speed state at time step t, represents the speed of the convoy’s leading vehicle at time step t+k+1, Indicates the first to N -1 car based on time step t The speed of the running state prediction in the next k+1 steps, It represents the speed of all vehicles in the fleet predicted in the next k+1 steps based on the operating status at time step t, and T represents the transpose symbol of a vector or matrix.

[0085] (a3) Calculate the acceleration of the vehicles in the formation based on the predicted parameters and states corresponding to the current time step. The calculation formula is as follows:

[0086]

[0087] in, represents the acceleration of all vehicles in the fleet predicted for the next k steps based on their operating status at time step t, represents the dynamical parameters of the prediction network at time step t for the future k-th step, represents the state vector of all vehicles in the fleet predicted for the next k steps based on their running status at time step t, It represents the expected state vector of all vehicles in the fleet for the next k steps based on the operating status at time step t.

[0088] (a4) Update the next platooning state based on vehicle kinematics. The calculation formula is as follows:

[0089]

[0090] in, It represents the headway of all vehicles in the fleet predicted based on the running status of time step t in the future k+1 steps. represents the speed of all vehicles in the fleet predicted for the next k steps based on their running status at time step t, It represents the headway of all vehicles in the fleet predicted based on the running status of time step t in the future k+1 steps. Indicates the time interval between each time step.

[0091] (a5) Repeat (a2) to (a4) to realize the future trajectory of the vehicle formation in the future time domain.

[0092] The output parameters (a6) can be used to analyze the stability of vehicle platoon operation.

[0093] Step 6: Model training and loss function.

[0094] The model integrates MTFLN and PeCG and adopts trajectory slicing end-to-end training. Its loss function consists of the following parts:

[0095] (b1) Prediction loss: mean square error of formation speed and spacing;

[0096] (b2) KL divergence loss: constrains the consistency of the variational encoder output distribution with the prior distribution;

[0097] The loss function expression is as follows:

[0098]

[0099] in, Is to control the KL divergence loss The hyperparameter of this case is 0.01. is the prediction loss and L is the total loss.

[0100] Step 7: Formation stability analysis

[0101] Stop-and-go oscillations in real traffic are a manifestation of traffic flow instability. A good vehicle platooning dynamics model should be able to describe the chordal stability associated with traffic oscillations. The parameters in PeCG can be directly used to calculate and analyze the chordal stability of vehicle platoons. The calculation formula is as follows:

[0102]

[0103] By calculating the amplitude of the disturbance transfer function, the chord stability of the formation dynamics is judged and compared with the stability evolution of real traffic scenarios.

[0104] This invention overcomes the lack of physical interpretability and analyzability of traditional data-driven models by encoding domain knowledge into a deep learning framework, while avoiding the conservative limitations of pure physical models. It can reproduce the actual formation speed and spacing with high precision, and supports closed-loop analysis of formation stability, providing theoretical support for simulation testing and control optimization of intelligent connected vehicle formations.

[0105] Example 2 Based on the vehicle formation dynamic modeling and simulation method based on physical coding deep learning described in Example 1, this embodiment provides a vehicle formation dynamic modeling and simulation device based on physical coding deep learning, including a storage medium and a processor; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the vehicle formation dynamic modeling and simulation method based on physical coding deep learning described in Example 1.

[0106] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A vehicle platoon dynamic modeling and simulation method based on physical coding deep learning, characterized by: include: Collecting and preprocessing the driving data of the vehicle formation to obtain preprocessed data; constructing a parameter encoding computation graph of physical constraints based on the preprocessed data; Constructing a multi-scale trajectory feature learning network to learn the behavioral characteristics of the vehicle formation from the pre-processed data; A non-autoregressive parameter decoder is constructed to map the behavioral characteristics of the vehicle formation into the time-varying parameters required by the parameter encoding computation graph of physical constraints. The future trajectory of the vehicle formation is iteratively calculated based on the state of the pilot vehicle. Combined with the future trajectory of the vehicle formation, the model is trained using the preprocessed data to generate a formation dynamic model that is both accurate and physically interpretable.

2. The vehicle formation dynamic modeling and simulation method based on physical coding deep learning according to claim 1 is characterized in that: The driving data of the vehicle platoon meets the following conditions: (a) The distance between adjacent vehicles in a platoon shall not exceed 120 m, and the speed difference shall not exceed 3 m / s; (b) The formation consists of at least four vehicles; (c) Maintain the formation operation status for not less than 10 seconds.

3. The vehicle formation dynamic modeling and simulation method based on physical coding deep learning according to claim 1 is characterized in that: Construct a parameter-encoded computational graph of physical constraints, including: Construct the linear time-varying dynamic equations of vehicle formations, ， in, Indicates the first n The car at time step t The acceleration of Is the first in the team n The car at time step t The distance between the front of the vehicle and the front of the vehicle, and are the mean values of speed and spacing in the historical time domain respectively; Indicates the speed of the nth vehicle and the speed of the n-1th vehicle The difference, =0 is the expected value of the speed difference, the speed feedback term <0, spacing feedback item >0, speed difference feedback item >0 is a necessary and sufficient condition for the equation to have physical meaning, and it is also a sufficient condition for the vehicle to meet the local stability condition. The parameters of the physical constraints are encoded in the calculation graph to be determined, and the physical constraints are realized through the Softplus activation function. represents the three-dimensional real space, represents the parameter vector of the dynamics model of the nth vehicle at time step t.

4. The vehicle formation dynamic modeling and simulation method based on physical coding deep learning according to claim 3 is characterized in that: The parameter-encoded computation graph satisfies the following physical constraints: (aa) parameter vector Each component of must satisfy: the speed feedback coefficient is negative, the spacing feedback coefficient is positive, and the relative speed feedback coefficient is positive; (bb) The relationship between vehicle acceleration and state deviation is described by a linear time-varying equation in the form of: ， in, is the state vector of the nth vehicle at time step t, is the expected state vector of the nth vehicle, defined by the historical state mean.

5. The vehicle formation dynamic modeling and simulation method based on physical coding deep learning according to claim 1 is characterized in that: Multi-scale trajectory feature learning network, including: A vehicle-level feature extraction module is used to extract multi-scale temporal features of vehicle historical trajectories from the preprocessed data using a Mamba state-space model, and generate vehicle-level historical features with probabilistic driving feature distribution through a variational encoder; The formation-level feature fusion module is used to capture the dynamic impact of the state of the forward vehicle in the formation on the subsequent vehicles from the preprocessed data based on the unidirectional topological causal attention mechanism of the fleet, and generate formation-level interaction features.

6. The vehicle formation dynamic modeling and simulation method based on physical coding deep learning according to claim 5 is characterized in that: The non-autoregressive parameter decoder fuses formation-level interaction features with vehicle-level historical features through a temporal causal cross-attention mechanism, generates the parameters required for the physically constrained parameter encoding computation graph, and adopts a multi-step parameter sharing strategy to reduce decoding complexity.

7. The vehicle formation dynamic modeling and simulation method based on physical coding deep learning according to claim 1 is characterized in that: Calculate the future trajectory of the vehicle formation, including: (a1) Softplus activation function is used to activate the output parameters of the non-autoregressive parameter decoder to implement physical constraints on the prediction parameters; (a2) Calculate vehicle speed, spacing, and speed difference based on the vehicle formation state at the current time step; (a3) Calculate the acceleration of the vehicles in the formation based on the predicted parameters and states corresponding to the current time step; (a4) Update the next formation operation state based on vehicle kinematics; (a5) Repeat (a2) to (a4) to realize the future trajectory of the vehicle formation in the future time domain.

8. The vehicle formation dynamic modeling and simulation method based on physical coding deep learning according to claim 1 is characterized in that: The model training includes: By integrating the multi-scale trajectory feature learning network with the parameter encoding computation graph of physical constraints, trajectory slices are used for end-to-end training. The loss function consists of the following parts: (b1) Prediction loss: mean square error of formation speed and spacing; (b2) KL divergence loss: constrains the consistency of the variational encoder output distribution with the prior distribution; The loss function expression is as follows: ， in, Is to control the KL divergence loss The hyperparameters of is the prediction loss and L is the total loss.

9. A vehicle formation dynamic modeling and simulation device based on physical coding deep learning, characterized in that: including storage media and processors; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the vehicle formation dynamic modeling and simulation method based on physical coding deep learning according to any one of claims 1 to 8.

Images