CN-122021340-A - Vehicle dynamics modeling method based on vehicle-road cloud fusion mechanism and data driving

Abstract

The invention belongs to the technical field of intelligent transportation and vehicle road cloud integration, and discloses a vehicle dynamics modeling method based on a vehicle road cloud fusion mechanism and data driving, which comprises the following steps of S1, constructing a three-degree-of-freedom mechanism model, and establishing a continuous physical evolution reference; S2, constructing a dual input stream state characteristic processing mechanism for realizing alignment of cloud discrete and vehicle-end continuous characteristics, S3, constructing a deep learning network composed of a gate control circulation unit GRU and a feedforward neural network FNN for capturing non-linear residual errors which cannot be described by a mechanism model, S4, building a cloud serial fusion prediction framework to generate a final fusion prediction state, and S5, designing a loss function based on physical constraint and training the model. The method solves the mapping problem of discrete control and continuous state in the vehicle-road cloud integration, and realizes high-precision prediction of the dynamic state of the vehicle at the future moment.

Inventors

• ZHAO MIN
• TANG ZHONGYUAN
• SUN DAIHUA

Assignees

• 重庆大学

Dates

Publication Date

20260512

Application Date

20260227

Claims (6)

1. 1. A vehicle dynamics modeling method based on a vehicle road cloud fusion mechanism and data driving is characterized by comprising the following steps: s1, constructing a three-degree-of-freedom mechanism model, and establishing a continuous physical evolution reference; S2, constructing a dual input stream state feature processing mechanism for realizing alignment of cloud discrete and vehicle-end continuous features; S3, constructing a deep learning network consisting of a gate control circulation unit GRU and a feedforward neural network FNN, and capturing nonlinear residual errors which cannot be described by a mechanism model; s4, establishing a cloud serial fusion prediction framework, and generating a final fusion prediction state; s5, loss function design and model training based on physical constraint.
2. 2. The vehicle dynamics modeling method based on vehicle road cloud fusion mechanism and data driving as set forth in claim 1, wherein the specific contents of the step S1 are that a nominal model including longitudinal, lateral and yaw movements, namely a three-degree-of-freedom mechanism model is established as a physical kernel describing a continuous motion rule of a vehicle, the model is used for receiving a vehicle state initial value and carrying out numerical integral deduction of continuous time steps to calculate a nominal physical predicted value conforming to a physical law , Representing the predicted state vector of the vehicle at the next time.
3. 3. The vehicle dynamics modeling method based on the vehicle road cloud fusion mechanism and the data driving according to claim 2, wherein the step S2 includes the following sub-steps: s2.1, designing a sliding window mechanism to decouple the vehicle road cloud system data into the following two paths of characteristic flows: I. a continuous state flow, which comprises a historical vehicle motion state and characterizes the inertial evolution of the system; In the formula, The vehicle is a continuous state flow, is composed of a vehicle history motion state sequence with the length of L, and represents the evolution trend of the internal state of the vehicle under the inertia and damping actions; a state feature vector representing time t; respectively representing the longitudinal speed, lateral speed, yaw rate and lateral acceleration at the moment; II. A discrete control flow comprising a sequence of control instructions characterizing an external stimulus; In the formula, The system is a discrete control flow, and is composed of a historical operation instruction sequence with the length of L, and the forced disturbance of external excitation to a dynamics system is represented; a control feature vector representing time t; Respectively representing front wheel rotation angle, driving torque request and braking pressure; s2.2, feature embedding and alignment fusion; Will be And Two independent feature embedding layers are respectively input and mapped to the same high-dimensional potential feature space, and then two paths of features are fused in the channel dimension to form an aligned space-time feature tensor.
4. 4. The vehicle dynamics modeling method based on the vehicle road cloud fusion mechanism and data driving as claimed in claim 3, wherein the deep learning network constructed in the step S3 comprises a time sequence feature extraction layer and a nonlinear mapping and decoding layer; the time sequence feature extraction layer is a GRU layer, and the GRU layer takes the fusion features output in the step S2 as input to output a hidden state vector ; The nonlinear mapping and decoding layer is a FNN layer, and the FNN layer comprises a hidden layer and a regression output layer; the hidden layer is formed by setting the number of the neuron nodes to 64 and adopting a ReLU activation function; The regression output layer is used for enabling the number of neurons to be consistent with the number of state variables to be predicted and adopting a Linear activation function; The deep learning network finally outputs the predicted value of the dynamic residual error under the current working condition 。
5. 5. The vehicle dynamics modeling method based on the vehicle road cloud fusion mechanism and the data driving according to claim 4, wherein the specific contents of the step S4 are as follows: The nominal physical predicted value calculated in the step S1 and the dynamic residual predicted value output in the step S3 are combined And performing linear superposition, wherein the calculation formula is as follows: In the formula, Representing an inverse normalization operation; the high-fidelity prediction state generated after cloud serial fusion is the final prediction value.
6. 6. The vehicle dynamics modeling method based on the vehicle road cloud fusion mechanism and the data driving according to claim 5, wherein the step S5 includes the following sub-steps: s5.1, designing a loss function; using weighted mean square error combined with L2 regularization term as total loss function : In the formula, Is of batch size; The number of the state variables; is the first Physical weight coefficients for the individual state variables; Representing the true value of the jth state variable in the ith sample; a model predictor representing a jth state variable in an ith sample; Is a regularization coefficient; Parameters representing the model; S5.2, model training is carried out based on feedback of the loss function; Training the model by utilizing historical driving data stored in the cloud, so that the model can learn a continuous response rule of the vehicle end under the drive of cloud discrete instructions; Adopting an Adam optimizer to update parameters, wherein the initial learning rate is 0.001; And introducing an early-stopping mechanism, namely automatically stopping training and storing the current optimal weight if the verification set Loss does not drop in 15 continuous epochs in the training process.

Description

Vehicle dynamics modeling method based on vehicle-road cloud fusion mechanism and data driving Technical Field The invention belongs to the technical field of intelligent transportation and vehicle road cloud integration, and particularly relates to a vehicle dynamics modeling method based on a vehicle road cloud fusion mechanism and data driving. Background In the vehicle-road cloud integrated system, a cloud control platform bears the core tasks of global coordination and decision control. Unlike bicycle intelligence, cloud control faces significant "space-time" heterogeneous challenges in that cloud-generated control instructions are typically discrete control sequences based on sampling periods, while movement of a vehicle on a real road is a continuous state evolution governed by laws of physics. The existing vehicle dynamics modeling method is mainly divided into two types, namely a physical mechanism model based on Newton-Euler's law, which has good physical interpretability but insufficient precision when dealing with nonlinear saturation of tires and complex external disturbance, and a deep learning model based on data driving, which has strong fitting capacity but weak generalization capacity and lacks physical constraint. The prior patent CN111898199a, a vehicle dynamics data driven modeling method, proposes to approximate Koopman operators using a network of depth-extended dynamic modal decomposition (Deep EDMD), in an attempt to map nonlinear dynamics systems into linear systems for processing. However, this approach is still essentially a global linear approximation to a nonlinear system, where the expression capability of the linear operator is limited when the vehicle enters a highly nonlinear tire force saturation region, and complex transient mutations are difficult to capture. In addition, the method mainly focuses on approximation of mathematical operators, and lacks direct physical constraint on a physical structure of a vehicle (such as tire cornering stiffness change), so that robustness of the model under extreme working conditions is insufficient. The prior patent CN113657036A (vehicle dynamics simulation implementation method based on a neural network and a physical model) provides a fusion method of parameter identification type, and key parameters (such as front and rear wheel cornering stiffness) in the physical model are estimated in real time by using the neural network) And substituting the model into a bicycle model for calculation. However, this approach is limited by the structural-Constrained (structural-Constrained) assumption of the physical model itself. For example, the three degree of freedom model itself ignores the effect of vehicle vertical load transfer on lateral force, and even if the neural network can accurately estimate cornering stiffness, it is unable to compensate for systematic errors due to "model structural loss" (e.g., lack of roll degrees of freedom). In other words, the method can only optimize parameters and cannot correct the defects of the model structure. The prior patent CN116992573a, a vehicle dynamics data driven modeling method suitable for IVCPS, although proposes the use of a GRU-FNN network and a multi-time step state feature (MTSSF) sliding window to build a proxy model. However, its feature engineering section (MTSSF) focuses mainly on stacking of historical states, failing to distinguish from physical properties the essential differences of "internal state evolution" (inertial domain) and "external control stimulus" (forcing domain), resulting in a network that is prone to information aliasing when extracting features. More importantly, the prior art is mainly oriented to vehicle-end local calculation, and the problem of special 'cloud-end space-time mismatch' in a vehicle-road cloud integrated scene is ignored, wherein a control instruction issued by a cloud end is usually based on a long-period discrete sequence (such as a jump instruction once every 100 ms), and a vehicle continuously evolves in the physical world. The existing model directly inputs a discrete control sequence into a network, and ignores nonlinear excitation action of a control instruction on a continuous physical state in a sampling gap (such as transient response lag caused by instruction step), so that the cloud model is easy to generate 'virtual-real asynchronous' phenomenon in open-loop prediction. Therefore, a modeling method for integrating a vehicle-road cloud-oriented scene, combining certainty of a physical mechanism and data-driven nonlinear compensation capability, and solving the problem of cloud discrete control and vehicle-end continuous state alignment is needed. Disclosure of Invention In view of the above, the invention aims to provide a vehicle dynamics modeling method based on a vehicle-road cloud fusion mechanism and data driving, which aims to solve the problems of low precision, difficult alignment and poor physical consistency of a single mod