Vehicle dynamics modeling method based on vehicle and road cloud fusion mechanism and data driving

Abstract

The invention belongs to the technical field of intelligent traffic and vehicle and road cloud integration, and discloses a vehicle dynamics modeling method based on a vehicle and road cloud fusion mechanism and data driving, comprising the following steps: S1, constructing a three-degree-of-freedom mechanism model, and determining a continuous physical evolution reference; s2, constructing a dual-input stream state feature processing mechanism for realizing alignment of cloud discrete features and vehicle end continuous features; s3, constructing a deep learning network composed of a gating circulation unit GRU and a feed-forward neural network FNN, wherein the deep learning network is used for capturing a nonlinear residual error which cannot be described by a mechanism model; s4, establishing a cloud serial fusion prediction architecture, and generating a final fusion prediction state; and S5, carrying out loss function design and model training based on physical constraints. According to the method, the mapping problem of discrete control and continuous state in vehicle-road cloud integration is solved, and high-precision prediction of the dynamic state of the vehicle at the future moment is realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation and vehicle-road-cloud integration technology, specifically involving a vehicle dynamics modeling method based on vehicle-road-cloud fusion mechanism and data-driven approach. Background Technology

[0002] In a vehicle-road-cloud integrated system, the cloud control platform undertakes the core tasks of global coordination and decision-making control. Unlike single-vehicle intelligence, cloud control faces significant spatial-temporal heterogeneity challenges: control commands generated in the cloud are typically discrete control sequences based on sampling periods, while the motion of a vehicle on a real road is a continuous state evolution governed by physical laws. Existing vehicle dynamics modeling methods are mainly divided into two categories: one is a physical mechanism model based on Newton-Euler's laws, which has good physical interpretability but lacks accuracy when dealing with tire nonlinear saturation and complex external disturbances; the other is a data-driven deep learning model, which has strong fitting ability but lacks physical constraints and has weak generalization ability.

[0003] Existing patent CN111898199A (A Data-Driven Modeling Method for Vehicle Dynamics) proposes using a Deep Extended Dynamic Mode Decomposition (Deep EDMD) network to approximate the Koopman operator, attempting to map the nonlinear dynamic system into a linear system for processing. However, this method is essentially still a global linear approximation of the nonlinear system. When the vehicle enters the highly nonlinear tire force saturation region, the expressive power of the linear operator is limited, making it difficult to capture complex transient changes. Furthermore, this method mainly focuses on the approximation of mathematical operators, lacking direct physical constraints on the vehicle's physical structure (such as tire lateral stiffness changes), resulting in insufficient robustness of the model under extreme conditions. Existing patent CN113657036A (A Vehicle Dynamics Simulation Implementation Method Based on Neural Networks and Physical Models) proposes a parameter-identification fusion method, using neural networks to estimate key parameters in the physical model in real time (such as front and rear wheel lateral stiffness). The method then substitutes the data into a bicycle model for calculation. However, this method is limited by the structural assumptions of the physical model itself. For example, the three-degree-of-freedom model ignores the influence of the vehicle's vertical load transfer on the lateral force. Even if the neural network can accurately estimate the lateral stiffness, it cannot compensate for the systematic errors caused by "model structural deficiencies" (such as the lack of roll degrees of freedom). In other words, this method can only optimize parameters and cannot correct model structural defects. Existing patent CN116992573A (A Data-Driven Modeling Method for Vehicle Dynamics Applicable to IVCPS) proposes to use a GRU-FNN network and a multi-step state feature (MTSSF) sliding window to construct a surrogate model. However, its feature engineering part (MTSSF) mainly focuses on the stacking of historical states and fails to distinguish the essential differences between "internal state evolution" (inertial domain) and "external control excitation" (forced domain) from a physical property perspective, which makes the network prone to information aliasing when extracting features.

[0004] More importantly, the aforementioned existing technologies are primarily geared towards local vehicle-side computing, neglecting the unique "cloud-edge spatiotemporal mismatch" problem in the vehicle-road-cloud integrated scenario: control commands issued from the cloud are typically based on long-period discrete sequences (such as jump commands every 100ms), while vehicles evolve continuously in the physical world. Existing models directly input discrete control sequences into the network, ignoring the nonlinear excitation effect of control commands on continuous physical states during sampling intervals (such as transient response lag caused by command step), leading to a "virtual-real asynchrony" phenomenon in open-loop prediction by the cloud model.

[0005] Therefore, there is an urgent need for a modeling method that can combine the determinism of physical mechanisms with the nonlinear compensation capability driven by data for vehicle-road-cloud integrated scenarios, and solve the problem of aligning discrete control in the cloud with continuous state in the vehicle. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a vehicle dynamics modeling method based on the vehicle-road-cloud fusion mechanism and data-driven approach, which aims to solve the problems of low accuracy, difficulty in alignment, and poor physical consistency of a single model in the vehicle-road-cloud integrated system when processing discrete control sequences in the cloud and continuous physical states at the vehicle end.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A vehicle dynamics modeling method based on vehicle-road-cloud fusion mechanism and data-driven approach includes the following steps:

[0009] S1. Construct a three-degree-of-freedom mechanism model to establish a benchmark for continuous physical evolution;

[0010] S2. Construct a dual-input stream state feature processing mechanism to align discrete features in the cloud with continuous features on the vehicle.

[0011] S3. Construct a deep learning network consisting of gated recurrent units (GRU) and feedforward neural networks (FNN) to capture nonlinear residuals that cannot be described by mechanistic models;

[0012] S4. Establish a cloud-based serial fusion prediction architecture to generate the final fusion prediction status;

[0013] S5. Design of loss function and model training based on physical constraints.

[0014] Furthermore, the specific content of step S1 is as follows: Establish a nominal model including longitudinal, lateral, and yaw motions, i.e., a three-degree-of-freedom mechanism model. This model serves as the physical kernel describing the continuous motion laws of the vehicle. This model is used to receive the initial state value of the vehicle and perform numerical integration derivation over continuous time steps to calculate the nominal physical prediction value that conforms to the physical laws. , This represents the predicted state vector of the vehicle at the next moment.

[0015] Furthermore, step S2 includes the following sub-steps:

[0016] S2.1 The sliding window mechanism is designed to decouple the vehicle-road-cloud system data into the following two feature flows:

[0017] I. Continuous state flow: contains historical vehicle motion states, characterizing the inertial evolution of the system;

[0018]

[0019]

[0020] In the formula, It is a continuous state flow, consisting of a sequence of vehicle historical motion states of length L, which characterizes the internal state evolution trend of the vehicle under inertia and damping. This represents the state feature vector at time t; These represent the longitudinal velocity, lateral velocity, yaw rate, and lateral acceleration at each moment.

[0021] II. Discrete Control Flow: Contains a sequence of control commands, representing external stimuli;

[0022]

[0023]

[0024] In the formula, It is a discrete control flow, consisting of a sequence of historical operation instructions of length L, which characterizes the forced perturbation of the dynamic system by external excitation; This represents the control feature vector at time t; These represent the front wheel steering angle, drive torque request, and braking pressure, respectively.

[0025] S2.2 Feature embedding and alignment fusion;

[0026] Will and Two independent feature embedding layers are input separately and mapped to the same high-dimensional latent feature space. Then, the two features are fused in the channel dimension to form an aligned spatiotemporal feature tensor.

[0027] Furthermore, the deep learning network constructed in step S3 includes a temporal feature extraction layer and a nonlinear mapping and decoding layer;

[0028] The temporal feature extraction layer is the GRU layer. The GRU layer takes the fused features output from step S2 as input and outputs a hidden state vector. ;

[0029] The nonlinear mapping and decoding layer is the FNN layer, which includes hidden layers and regression output layers.

[0030] Hidden layer: The number of neurons is set to 64, and the ReLU activation function is used;

[0031] Regression output layer: The number of neurons is the same as the number of state variables to be predicted, and the Linear activation function is used;

[0032] The deep learning network ultimately outputs the predicted dynamic residual values ​​under the current operating conditions. .

[0033] Furthermore, the specific content of step S4 is as follows:

[0034] The nominal physical prediction value calculated in step S1 is compared with the dynamic residual prediction value output in step S3. For linear superposition, the calculation formula is:

[0035]

[0036] In the formula, This indicates the unnormalization operation; This represents the high-fidelity prediction state generated after serial fusion in the cloud, and is the final prediction value.

[0037] Furthermore, step S5 includes the following sub-steps:

[0038] S5.1 Design the loss function;

[0039] The total loss function is a weighted mean square error combined with an L2 regularization term. :

[0040]

[0041] In the formula, Batch size; The number of state variables; For the first The physical weight coefficients of each state variable; This represents the true value of the j-th state variable in the i-th sample; This represents the model prediction value of the j-th state variable in the i-th sample; The regularization coefficient is used. The parameters representing the model;

[0042] S5.2 Model training based on feedback from the loss function;

[0043] The model is trained using historical driving data stored in the cloud, enabling it to learn the continuous response patterns of the vehicle under discrete cloud commands.

[0044] The Adam optimizer was used for parameter updates, with an initial learning rate of 0.001.

[0045] An early stopping mechanism is introduced: if the validation set loss does not decrease within 15 consecutive epochs during training, training is automatically terminated and the current optimal weights are saved.

[0046] Beneficial effects:

[0047] This invention provides a vehicle dynamics modeling method based on vehicle-road-cloud fusion mechanism and data-driven approach. Through deep integration of mechanism and data-driven approaches, it solves the mapping problem between discrete control and continuous state in vehicle-road-cloud integration. Deployed on a cloud control platform, this method utilizes sampled states uploaded from the vehicle and control sequences sent from the cloud to achieve high-precision prediction of the vehicle's future dynamic state.

[0048] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0049] Figure 1 This is a flowchart of a vehicle dynamics modeling method based on vehicle-road-cloud fusion mechanism and data-driven approach according to the present invention;

[0050] Figure 2 This is a schematic diagram of dual-input stream state feature (DISF) extraction.

[0051] Figure 3 This is a diagram of the architecture of a fusion dynamics model deployed in the cloud. Detailed Implementation

[0052] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0053] like Figures 1-3 As shown, this invention provides a vehicle dynamics modeling method based on vehicle-road-cloud fusion mechanism and data-driven approach, specifically including the following steps:

[0054] S1. Construct a three-degree-of-freedom mechanism model to establish a benchmark for continuous physical evolution.

[0055] A three-degree-of-freedom dynamic model of the vehicle is established, neglecting the high-frequency motion of the suspension. This model serves as the physical kernel describing the continuous motion of the vehicle. For any given moment... Model receiving Vehicle state vector at time 1 and control input vector .in For longitudinal velocity, For lateral velocity, This is the yaw rate. The mechanistic model assumes the vehicle is... to Within tiny time steps, it follows the laws of rigid body motion. Even control commands issued from the cloud... Even though it is discrete (e.g., updated every 100ms), the mechanistic model is still calculated using high-precision numerical integration. Nominal physical prediction value at time , This represents the predicted state vector of the vehicle at the next moment. This step provides a baseline prediction that conforms to the laws of physics, explains the low-frequency linear characteristics that dominate vehicle dynamics, and ensures that the model can still output a reasonable physical state when data is missing.

[0056] S2. Design a dual-input stream state feature (DISF) mechanism to align discrete features in the cloud with continuous features on the vehicle.

[0057] To address the characteristics of the data stream in the vehicle-road-cloud system, a length of [length missing] is introduced. Time-sliding window mechanism (preferred in this embodiment) This window decouples the input data into two feature streams with different physical properties at the feature level.

[0058] Step 2.1 Extract the "continuous state stream" (State Stream, This data stream contains a sequence of the vehicle's historical motion states within a time window, characterizing the internal state evolution trend of the vehicle system under inertia and damping. Mathematically, it is expressed as:

[0059]

[0060]

[0061] In the formula, It is a continuous state flow, consisting of a sequence of vehicle historical motion states of length L, which characterizes the internal state evolution trend of the vehicle under inertia and damping. This represents the state feature vector at time t; These represent the longitudinal velocity, lateral velocity, yaw rate, and lateral acceleration at each moment.

[0062] Step 2.2 Extract the "Discrete Control Flow" (Control Stream) This data stream contains a sequence of historical operating commands from the cloud or the driver, characterizing the forced perturbation of the dynamic system by external stimuli. Mathematically, it is expressed as:

[0063]

[0064]

[0065] In the formula, It is a discrete control flow, consisting of a sequence of historical operation instructions of length L, which characterizes the forced perturbation of the dynamic system by external excitation; This represents the control feature vector at time t; These represent the front wheel steering angle, drive torque request, and braking pressure, respectively.

[0066] Step 2.3 Feature Embedding and Alignment Fusion: To address the differences in physical dimensions and distribution between discrete control commands and continuous state responses, the original data is not directly concatenated. Instead, the features are... and Two independent feature embedding layers are input separately and mapped to the same high-dimensional latent feature space. Subsequently, the two features are fused along the channel dimension to form an aligned spatiotemporal feature tensor, which serves as the input to the subsequent neural network.

[0067] S3. Construct a deep learning network consisting of gated recurrent units (GRU) and feedforward neural networks (FNN) specifically designed to capture nonlinear residuals (such as tire lateral force saturation) that cannot be described by mechanistic models.

[0068] Step 3.1 Temporal Feature Extraction (GRU Layer): GRU is chosen as the core temporal unit because it reduces parameters by approximately 30% compared to LSTM, making it more suitable for control scenarios with high real-time requirements. GRU dynamically adjusts the information flow through reset and update gates. The fused features output from Step 2 are input into the GRU layer, and GRU outputs a hidden state vector. This vector encodes the cumulative hysteresis effect of historical discrete control sequences on the current continuous state of the vehicle (i.e., the system's "memory").

[0069] Step 3.2 Nonlinear Mapping and Decoding (FNN Layer):

[0070] Design a feedforward neural network as a decoder, comprising the following two layers:

[0071] 1. Hidden Layer: The number of neurons is set to 64, and the ReLU activation function is used. ReLU can alleviate the gradient vanishing problem and enhance the network's ability to express complex dynamic characteristics. Dropout regularization (dropout rate set to 0.2) is introduced in this layer to prevent the model from overfitting under specific conditions.

[0072] 2. Regression Output Layer: Set the number of neurons to match the number of state variables to be predicted (6 in this example). Since the dynamic residuals can be positive or negative and are unbounded, this layer must use a linear activation function; saturation functions such as Sigmoid or Tanh are strictly prohibited. Finally, the FNN outputs the predicted dynamic residual values ​​under the current operating conditions. .

[0073] S4: Implement cloud-based serial fusion to generate high-fidelity predicted states. A serial hybrid modeling strategy, primarily based on physical models and supplemented by data-driven approaches, is adopted.

[0074] The nominal physical prediction value calculated in step S1 The predicted dynamic residuals output from step S3 Perform linear superposition. The calculation formula is:

[0075]

[0076] in, This indicates the inverse normalization operation. Since the input and output of a neural network are typically normalized within the range [0, 1] or [-1, 1], they must be restored to their true physical dimensions (e.g., mean and variance) using the statistical parameters of the training set before fusion. , This serial architecture features an embedded fault-oriented safety mechanism. When cloud data transmission is obstructed or the neural network output is abnormal, the system can automatically ignore residual terms and smoothly degenerate into classic mechanistic model predictions, ensuring the most basic control safety.

[0077] S5: Loss Function Design and Model Training Based on Physical Constraints. To enable the cloud-based model to accurately learn the pattern of "discrete instructions driving continuous response," a composite loss function incorporating physical weights is constructed for offline training or online fine-tuning.

[0078] Step 5.1 Construct a loss function that includes weighted mean square error and regularization term, focusing on constraining key stability indicators such as yaw rate;

[0079] Loss function definition: The total loss function is defined as the weighted mean squared error combined with the L2 regularization term. :

[0080]

[0081] in, Set the batch size (to 64). The number of state variables. For the first The physical weighting coefficients of each state variable. Considering the critical role of lateral stability in autonomous driving control, this embodiment specifically sets the weighting coefficient of yaw rate to... This is to penalize the prediction error of the variable in particular. Let be the regularization coefficient, set to . This is used to suppress model overfitting.

[0082] Step 5.2 Train the model based on the feedback of the loss function;

[0083] The model is trained using historical driving data stored in the cloud, enabling it to learn the continuous response patterns of the vehicle under discrete cloud commands.

[0084] Optimizer and Training Strategy: The Adam optimizer is used for parameter updates, with an initial learning rate of 0.001. An early stopping mechanism is introduced: During training, if the validation set loss does not decrease within 15 consecutive epochs, training is automatically terminated and the current optimal weights are saved to ensure the model's generalization ability under unknown conditions.

[0085] It is hereby declared that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A vehicle dynamics modeling method based on vehicle-road-cloud fusion mechanism and data-driven approach, characterized in that, Includes the following steps: S1. Construct a three-degree-of-freedom mechanism model to establish a benchmark for continuous physical evolution; S2. Construct a dual-input stream state feature processing mechanism to align discrete features in the cloud with continuous features on the vehicle. S3. Construct a deep learning network consisting of gated recurrent units (GRU) and feedforward neural networks (FNN) to capture nonlinear residuals that cannot be described by mechanistic models; S4. Establish a cloud-based serial fusion prediction architecture to generate the final fusion prediction status; S5. Design of loss function and model training based on physical constraints.

2. The vehicle dynamics modeling method based on vehicle-road-cloud fusion mechanism and data-driven approach according to claim 1, characterized in that, The specific content of step S1 is as follows: Establish a nominal model including longitudinal, lateral, and yaw motions, i.e., a three-degree-of-freedom mechanism model. This model serves as the physical kernel describing the continuous motion laws of the vehicle. This model is used to receive the initial state values ​​of the vehicle and perform numerical integration derivation over continuous time steps to calculate the nominal physical prediction value that conforms to the physical laws. , This represents the predicted state vector of the vehicle at the next moment.

3. The vehicle dynamics modeling method based on vehicle-road-cloud fusion mechanism and data-driven approach according to claim 2, characterized in that, Step S2 includes the following sub-steps: S2.1 The sliding window mechanism is designed to decouple the vehicle-road-cloud system data into the following two feature flows: I. Continuous state flow: contains historical vehicle motion states, characterizing the inertial evolution of the system; In the formula, It is a continuous state flow, consisting of a sequence of vehicle historical motion states of length L, which characterizes the internal state evolution trend of the vehicle under inertia and damping. This represents the state feature vector at time t; These represent the longitudinal velocity, lateral velocity, yaw rate, and lateral acceleration at each moment. II. Discrete Control Flow: Contains a sequence of control commands, representing external stimuli; In the formula, It is a discrete control flow, consisting of a sequence of historical operation instructions of length L, which characterizes the forced perturbation of the dynamic system by external excitation; This represents the control feature vector at time t; These represent the front wheel steering angle, drive torque request, and braking pressure, respectively. S2.2 Feature embedding and alignment fusion; Will and Two independent feature embedding layers are input separately and mapped to the same high-dimensional latent feature space. Then, the two features are fused in the channel dimension to form an aligned spatiotemporal feature tensor.

4. The vehicle dynamics modeling method based on vehicle-road-cloud fusion mechanism and data-driven approach according to claim 3, characterized in that: The deep learning network constructed in step S3 includes a temporal feature extraction layer and a nonlinear mapping and decoding layer; The temporal feature extraction layer is the GRU layer. The GRU layer takes the fused features output from step S2 as input and outputs a hidden state vector. ; The nonlinear mapping and decoding layer is the FNN layer, which includes hidden layers and regression output layers. Hidden layer: The number of neurons is set to 64, and the ReLU activation function is used; Regression output layer: The number of neurons is the same as the number of state variables to be predicted, and the Linear activation function is used; The deep learning network ultimately outputs the predicted dynamic residual values ​​under the current operating conditions. .

5. The vehicle dynamics modeling method based on vehicle-road-cloud fusion mechanism and data-driven approach according to claim 4, characterized in that, The specific content of step S4 is as follows: The nominal physical prediction value calculated in step S1 is compared with the dynamic residual prediction value output in step S3. For linear superposition, the calculation formula is: In the formula, This indicates the unnormalization operation; This represents the high-fidelity prediction state generated after serial fusion in the cloud, and is the final prediction value.

6. The vehicle dynamics modeling method based on vehicle-road-cloud fusion mechanism and data-driven approach according to claim 5, characterized in that, Step S5 includes the following sub-steps: S5.1 Design the loss function; The total loss function is a weighted mean square error combined with an L2 regularization term. : In the formula, Batch size; The number of state variables; For the first The physical weight coefficients of each state variable; This represents the true value of the j-th state variable in the i-th sample; This represents the model prediction value of the j-th state variable in the i-th sample; The regularization coefficient is used. The parameters representing the model; S5.2 Model training based on feedback from the loss function; The model is trained using historical driving data stored in the cloud, enabling it to learn the continuous response patterns of the vehicle under discrete cloud commands. The Adam optimizer was used for parameter updates, with an initial learning rate of 0.

001. An early stopping mechanism is introduced: if the validation set loss does not decrease within 15 consecutive epochs during training, training is automatically terminated and the current optimal weights are saved.

Images