A physical information neural network vehicle system identification method

Abstract

The application discloses a physical information neural network vehicle system identification method, comprising the following steps: S1, collecting original data, preprocessing, and constructing time series input based on the preprocessed original data; S2, performing feature embedding processing on the time series input, and outputting time sequence feature representation; S3, constructing a vehicle dynamics system according to the time sequence feature representation; S4, constructing a physical residual based on the vehicle dynamics system, and determining physical consistency; S5, using the trained Transform model to complete the identification of the vehicle state and the dynamics parameters. The application can maintain high state estimation accuracy under complex working conditions such as sudden acceleration, sudden turning and low adhesion, has good robustness to sensor noise and partial state unobservable conditions, and is suitable for application scenarios such as vehicle stability control, automatic driving and vehicle dynamics modeling.

Description

Technical Field

[0001] This invention belongs to the field of vehicle system identification technology, specifically relating to a physical information neural network method for vehicle system identification. Background Technology

[0002] With the development of intelligent driving and advanced driver assistance systems, vehicle dynamics modeling and vehicle system identification technologies have been widely researched and applied in areas such as vehicle stability control, path tracking control, autonomous driving decision-making and control, and vehicle state estimation. Vehicle system identification typically estimates vehicle dynamic state variables and model parameters using sensor data collected during vehicle operation, provided that the vehicle dynamics model structure is known or partially known. It is an important application direction of control theory and system identification in vehicle engineering.

[0003] In existing research and engineering practice, vehicle system identification methods are mainly based on analytical vehicle dynamics models, and their theoretical foundation comes from classical system identification theory and state estimation methods. Commonly used techniques include least squares, recursive least squares, and Kalman filtering based on state-space models. These methods are usually based on a well-defined vehicle dynamics model and estimate key vehicle dynamic parameters and states through parameterized models, forming a relatively systematic research and application system in the field of vehicle engineering.

[0004] With the improvement of computing power and the enrichment of data acquisition methods, data-driven modeling methods have been gradually introduced into vehicle dynamics research. Some studies employ machine learning methods such as artificial neural networks and support vector machines to model the relationship between vehicle inputs, states, and outputs, thereby assisting in vehicle dynamics analysis and state estimation. These methods differ from traditional system identification methods based on analytical models in their research paradigm, focusing instead on data-driven function mapping modeling. Physical-Informed Neural Networks (PINNs) are modeling methods that embed physical control equations, conservation relationships, or constraints into the neural network training process. Their research originated in the fields of numerical solutions to partial differential equations and inverse problems of physical systems. By introducing physical constraint terms into the loss function, the PINN method enables the neural network to satisfy the requirements of physical equations while fitting data, and has been widely applied to modeling problems of continuous physical systems such as fluid mechanics, solid mechanics, electromagnetic fields, and heat conduction.

[0005] In existing technologies, vehicle dynamics systems are typically described using a nonlinear dynamic model with one degree of freedom. Researchers have proposed various methods for vehicle system identification. One traditional approach relies on analytical vehicle dynamics models, combined with parameter estimation or state estimation algorithms to achieve system identification, such as methods based on least squares or filtering theory. These methods usually require a relatively accurate vehicle dynamics model structure, and their identification accuracy and adaptability are limited when the model makes many simplification assumptions or has significant parameter uncertainties.

[0006] With the development of machine learning technology, data-driven vehicle system identification methods have gradually emerged. These methods typically utilize neural networks to establish a mapping relationship between vehicle state, control input, and time. While these methods reduce the reliance on analytical vehicle dynamics models to some extent, due to the lack of explicit physical constraints, their prediction results are prone to deviating from the laws of vehicle dynamics when faced with complex operating conditions or changes in data distribution.

[0007] In other engineering and physical modeling fields, to improve the physical consistency of data-driven models, the Physics-Informed Neural Network (PINN) method has been proposed. This type of method introduces system dynamics equations or physical constraints during neural network training, enabling the network to satisfy physical requirements while fitting observed data. However, existing PINN methods are mainly geared towards modeling continuum mechanics, fluid mechanics, or general dynamic systems. Their direct application in vehicle system identification remains limited, and they struggle to effectively handle characteristics such as strong coupling of control inputs and non-stationary state evolution in vehicle systems.

[0008] On the other hand, in the field of time series modeling, models such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs) are widely used to characterize the temporal evolution of system states. These methods typically employ recursive structures to model time series, which can capture local temporal dependencies to some extent. However, these models are primarily geared towards general time series prediction tasks and struggle to effectively characterize the global temporal dependencies between states and control inputs in vehicle dynamics systems. Furthermore, their physical constraints are usually introduced empirically or locally, making it difficult to guarantee the physical consistency and stability of system identification results under complex driving conditions.

[0009] Despite the progress made in vehicle system identification, existing technologies still have the following shortcomings in practical applications:

[0010] First, background techniques have limited ability to model complex multi-degree-of-freedom systems. In vehicle dynamics systems, significant nonlinear coupling relationships exist between degrees of freedom. Existing PINN methods based on fully connected neural networks primarily focus on single-moment state mapping, making it difficult to effectively characterize the coupling evolution relationships between different degrees of freedom over time, thus affecting system identification accuracy. Second, background techniques lack the ability to model time dependencies. Existing PINN methods typically do not explicitly model global dependencies over long time series, resulting in insufficient ability to describe the dynamic response of vehicles under complex conditions (such as sharp turns, acceleration / deceleration transitions, etc.), and easily overlooking the influence of historical conditions on the current state. Furthermore, background techniques lack robustness to noise and sparse data. In real-world vehicle systems, sensor measurement data often suffers from noise interference and uneven sampling. Existing point-level supervision-based system identification methods are sensitive to single-moment measurement errors, easily leading to unstable identification results when data quality is low.

[0011] In summary, existing vehicle system identification methods still have shortcomings in terms of model dependency, physical consistency, and time series modeling capabilities. There is an urgent need for a new vehicle system identification method that can integrate vehicle dynamics constraints, global time dependencies, and the advantages of data-driven approaches. Summary of the Invention

[0012] To address the above problems, this invention proposes a physical information neural network-based vehicle system identification method.

[0013] The technical solution of the present invention is: a physical information neural network vehicle system identification method comprising the following steps:

[0014] S1. Collect raw data, preprocess it, and construct a time series input based on the preprocessed raw data;

[0015] S2. Perform feature embedding processing on the time series input, and feed the time series input after feature embedding processing into the Transformer model to output the time series feature representation.

[0016] S3. Construct the vehicle dynamics system based on the temporal characteristics;

[0017] S4. Based on the vehicle dynamics system, construct physical residuals and determine physical consistency;

[0018] S5. Based on physical consistency, train the Transformer model and use the trained Transformer model to identify vehicle state and dynamic parameters.

[0019] Furthermore, the raw data includes control input signals and vehicle measurable status signals;

[0020] Control input signals include vehicle steering angle, driving torque or braking torque; measurable vehicle status signals include longitudinal acceleration, lateral acceleration, yaw rate and wheel speed;

[0021] Preprocessing includes noise reduction, outlier removal, missing data interpolation, and normalization.

[0022] Time series input for:

[0023] ;

[0024] in, This represents the feature vector composed of the control input and the measurable state. This represents the feature vector of the earliest sampling moment in the current time window. This represents the feature vector at the second sampling moment in the current time window.

[0025] Furthermore, the feature embedding process for the time series input is specifically as follows: the input vector at each time step is linearly mapped to a high-dimensional feature space consistent with the feature dimension of the Transformer model, and positional encoding is superimposed to complete the feature embedding process;

[0026] The Transformer model consists of several encoding layers, each of which includes a multi-head self-attention sub-layer and a feedforward neural network sub-layer.

[0027] Furthermore, S3 includes the following sub-steps:

[0028] S31. Based on the temporal characteristics, obtain the vehicle state estimate;

[0029] S32. Based on the vehicle state estimate, use the automatic differentiation method or numerical difference method to find the time derivative of the predicted state.

[0030] S33. Construct a vehicle dynamics system based on the time derivative of the predicted state.

[0031] Furthermore, vehicle state estimates for:

[0032] ;

[0033] in, This represents the temporal hidden feature vector output by the Transformer model. This represents the vehicle state estimate obtained from the temporal hidden feature vector;

[0034] The vehicle dynamics system is as follows:

[0035] ;

[0036] in, The derivative representing the vehicle's state at a given moment. This represents the estimated vehicle condition. This represents the vehicle control input vector. Represents the set of vehicle dynamics parameters. This represents the vehicle dynamics function.

[0037] Furthermore, physical residuals for:

[0038] ;

[0039] in, This represents the estimated vehicle condition. Represents the vehicle dynamics function. This represents the vehicle control input vector. This represents the set of vehicle dynamics parameters.

[0040] Furthermore, the joint loss function used to train the Transformer model for:

[0041] ;

[0042] ;

[0043] ;

[0044] in, Indicates the data error term. Represents the physical constraint error term. This represents the weighting coefficient of the data error term. The weighting coefficients represent the physical constraint error terms. This represents the estimated vehicle condition. Indicates at time Measurement status, This represents the total number of sample points in the time series. Represents physical residuals.

[0045] The beneficial effects of this invention are: it can maintain high state estimation accuracy under complex conditions such as rapid acceleration, sharp turning and low adhesion, and has good robustness to sensor noise and some unmeasurable conditions, making it suitable for applications such as vehicle stability control, autonomous driving and vehicle dynamics modeling. Attached Figure Description

[0046] Figure 1A flowchart for a physical information neural network vehicle system identification method;

[0047] Figure 2 This is a comparison chart of the predicted and true values ​​of the Transformer model for the longitudinal velocity, lateral velocity, and yaw rate of a vehicle over a long time series.

[0048] Figure 3 The image shows the model's prediction results for the angular velocities of the four wheels. Detailed Implementation

[0049] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0050] like Figure 1 As shown, the present invention provides a physical information neural network vehicle system identification method, comprising the following steps:

[0051] S1. Collect raw data, preprocess it, and construct a time series input based on the preprocessed raw data;

[0052] S2. Perform feature embedding processing on the time series input, and feed the time series input after feature embedding processing into the Transformer model to output the time series feature representation.

[0053] S3. Construct the vehicle dynamics system based on the temporal characteristics;

[0054] S4. Based on the vehicle dynamics system, construct physical residuals and determine physical consistency;

[0055] S5. Based on physical consistency, train the Transformer model and use the trained Transformer model to identify vehicle state and dynamic parameters.

[0056] The purpose of this invention is to propose a method that can explicitly capture the global temporal correlation of vehicle dynamics across multiple time scales, enabling rapid recovery of the vehicle's latent state and maintaining output consistency during transient events. Simultaneously, this solution aims to achieve joint identification of state and key dynamic parameters, improving the system's robustness under noisy or incomplete data conditions.

[0057] To achieve the above objectives, this invention proposes to use a seven-DOF vehicle dynamics model as the object, embedding the vehicle dynamics differential equations as physical constraints into a Transformer temporal modeling network based on a self-attention mechanism. This scheme first preprocesses the collected control inputs and state signals to construct standardized temporal data; then, it models the global temporal dependency using the multi-head self-attention mechanism of the Transformer encoder; subsequently, it uses a feedforward network to predict the vehicle system state and constructs residuals using automatic differentiation techniques; finally, it synchronously optimizes the network parameters and vehicle dynamics parameters through a joint loss function to achieve synchronous identification.

[0058] In this embodiment of the invention, the raw data includes control input signals and vehicle measurable state signals;

[0059] Control input signals include vehicle steering angle, driving torque or braking torque; measurable vehicle status signals include longitudinal acceleration, lateral acceleration, yaw rate and wheel speed;

[0060] Preprocessing includes noise reduction, outlier removal, missing data interpolation, and normalization.

[0061] Time series input for:

[0062] ;

[0063] in, This represents the feature vector composed of the control input and the measurable state. This represents the feature vector of the earliest sampling moment in the current time window. This represents the feature vector at the second sampling moment in the current time window.

[0064] During the system identification process, control input signals and measurable vehicle state signals are first collected as raw data inputs during vehicle operation. The control input signals include information such as vehicle steering angle, driving torque, or braking torque; the measurable vehicle state signals include at least longitudinal acceleration, lateral acceleration, yaw rate, and the rotational speed of each wheel. The signals are sampled at a fixed sampling frequency, preferably 50–200 Hz. The collected raw data undergoes preprocessing, including noise reduction, outlier removal, missing data interpolation, and normalization, to obtain standardized time-series data for network modeling.

[0065] Based on the preprocessed data, a time series input containing historical information is constructed at any given time point *t*, which serves as the network input for the model. The time series length is set to *k*, where *k* takes 20–200 sampling points to cover the dynamic characteristics of the vehicle at short and medium time scales. The time series, as the model input, is used to characterize the historical evolution of the vehicle system.

[0066] In this embodiment of the invention, the feature embedding process for the time series input is specifically as follows: the input vector at each time step is linearly mapped to a high-dimensional feature space consistent with the feature dimension of the Transformer model, and position encoding is superimposed to complete the feature embedding process;

[0067] The Transformer model consists of several encoding layers, each of which includes a multi-head self-attention sub-layer and a feedforward neural network sub-layer.

[0068] To enable the network to effectively process time series data, the time series input is first subjected to feature embedding processing. Specifically, the input vector at each time step is... By mapping linearly to a high-dimensional feature space consistent with the feature dimensions of the Transformer model, and superimposing positional encoding to introduce temporal order information, an embedding sequence with temporal positional information is formed. Positional encoding enables the network to distinguish data from different time steps, enhancing its ability to model temporal correlations.

[0069] The embedded time-series input is fed into a Transformer encoder for feature extraction and temporal modeling. The Transformer encoder consists of a stacked multi-layer coding structure, with each layer including a multi-head self-attention sublayer and a feedforward neural network sublayer. The self-attention mechanism adaptively weights historical input information by calculating correlation weights between different time steps, thus explicitly modeling the global dependency of vehicle dynamics in the time dimension. The feedforward neural network sublayer performs a non-linear mapping on the attention output to enhance feature representation. In this invention, the Transformer encoder has 4–8 layers, the model feature dimension is 128–512, the number of heads in the multi-head attention mechanism is 4–8, and the hidden layer dimension of each feedforward neural network sublayer is 2–4 times the feature dimension. After processing by the Transformer encoder, a temporal feature representation characterizing the dynamic characteristics of the vehicle system is obtained. .

[0070] In this embodiment of the invention, S3 includes the following sub-steps:

[0071] S31. Based on the temporal characteristics, obtain the vehicle state estimate;

[0072] S32. Based on the vehicle state estimate, use the automatic differentiation method or numerical difference method to find the time derivative of the predicted state.

[0073] S33. Construct a vehicle dynamics system based on the time derivative of the predicted state.

[0074] Based on the temporal feature representation output by the Transformer encoder, the vehicle system state at the current moment is predicted by a feedforward mapping network to obtain the vehicle state estimate.

[0075] The vehicle system state vector is defined as follows:

[0076] ;

[0077] In the formula, For the longitudinal speed of the vehicle, For the vehicle's lateral speed, Let yaw rate be the vehicle's angular velocity. Let be the angular velocity of the nth wheel. The above state prediction results serve as the network output and are also the basis for subsequent physical constraint calculations. After obtaining the vehicle state prediction values, the time derivative of the predicted state is obtained by using automatic differentiation or numerical difference methods. Automatic differentiation methods can accurately calculate the time derivative of the state output in the neural network computation graph, avoiding the truncation error and noise amplification problems caused by numerical difference, thus providing stable and reliable derivative information for the construction of physical constraints.

[0078] In this embodiment of the invention, the vehicle state estimate is... for:

[0079] ;

[0080] in, This represents the temporal hidden feature vector output by the Transformer model. This represents the vehicle state estimate obtained from the temporal hidden feature vector;

[0081] The vehicle dynamics system is as follows:

[0082] ;

[0083] in, The derivative representing the vehicle's state at a given moment. This represents the estimated vehicle condition. This represents the vehicle control input vector. Represents the set of vehicle dynamics parameters. This represents the vehicle dynamics function.

[0084] The set of vehicle dynamics parameters includes at least parameters such as vehicle mass, yaw moment of inertia, and tire lateral stiffness.

[0085] In this embodiment of the invention, the physical residual for:

[0086] ;

[0087] in, This represents the estimated vehicle condition. Represents the vehicle dynamics function. This represents the vehicle control input vector. This represents the set of vehicle dynamics parameters.

[0088] Physical residuals are used to measure whether the network's predicted state satisfies the vehicle dynamics differential equation, thus introducing physical consistency as a constraint into the network training process.

[0089] In this embodiment of the invention, the joint loss function for training the Transformer model is... for:

[0090] ;

[0091] ;

[0092] ;

[0093] in, Indicates the data error term. Represents the physical constraint error term. This represents the weighting coefficient of the data error term. The weighting coefficients represent the physical constraint error terms. This represents the estimated vehicle condition. Indicates at time Measurement status, This represents the total number of sample points in the time series. Represents physical residuals.

[0094] A joint loss function is constructed to simultaneously optimize network parameters and vehicle dynamics parameters. The joint loss function includes at least a data error term and a physical constraint error term. The data error term measures the difference between the predicted and measured states, while the physical constraint error term measures the magnitude of the physical residual. In the early stages of model training, the weight of the data error term is increased to improve the fitting ability to the measured data; in the later stages of training, the weight of the physical constraint error term is gradually increased to enhance the model's physical consistency and generalization ability. Vehicle dynamics parameters It can be used as a trainable variable to participate directly in joint optimization, or estimated through an independent parameter mapping network, thereby achieving synchronous identification of vehicle state and dynamic parameters during training.

[0095] After the model training is completed, based on the real-time acquired vehicle control input and measurable state data, a time series input is constructed using a sliding time window method and input into the trained Transformer physical information neural network to realize online identification of vehicle system state and dynamic parameters.

[0096] Figure 2 This demonstrates the effect of the Transformer model on the longitudinal velocity of the vehicle over a long time series. lateral velocity and yaw rate The comparison between the predicted values ​​and the true values ​​shows that the accuracy is better than that of the traditional 7-DOF physical model, with the accuracy of the three models improved by 93.61%, 77.56% and +87.09%, respectively.

[0097] Figure 3 The model demonstrates the angular velocities of the four wheels ( The prediction results demonstrate high fitting accuracy and robustness under non-stationary operating conditions.

[0098] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for identifying a vehicle system using a physical information neural network, characterized in that, Includes the following steps: S1. Collect raw data, preprocess it, and construct a time series input based on the preprocessed raw data; S2. Perform feature embedding processing on the time series input, and feed the time series input after feature embedding processing into the Transformer model to output the time series feature representation. S3. Construct the vehicle dynamics system based on the temporal characteristics; S4. Based on the vehicle dynamics system, construct physical residuals and determine physical consistency; S5. Based on physical consistency, train the Transformer model and use the trained Transformer model to identify vehicle state and dynamic parameters.

2. The physical information neural network vehicle system identification method according to claim 1, characterized in that, The raw data includes control input signals and vehicle measurable status signals; The control input signals include vehicle steering angle, driving torque or braking torque; the measurable vehicle state signals include longitudinal acceleration, lateral acceleration, yaw rate and wheel speed. The preprocessing includes noise reduction, outlier removal, missing data interpolation, and normalization. The time series input for: ； in, This represents the feature vector composed of the control input and the measurable state. This represents the feature vector of the earliest sampling moment in the current time window. This represents the feature vector at the second sampling moment in the current time window.

3. The physical information neural network vehicle system identification method according to claim 1, characterized in that, The specific process of embedding the time series input is as follows: linearly map the input vector at each time step to a high-dimensional feature space consistent with the feature dimension of the Transformer model, and superimpose positional encoding to complete the feature embedding process. The Transformer model includes several encoding layers, each of which includes a multi-head self-attention sub-layer and a feedforward neural network sub-layer.

4. The physical information neural network vehicle system identification method according to claim 1, characterized in that, S3 includes the following sub-steps: S31. Based on the temporal characteristics, obtain the vehicle state estimate; S32. Based on the vehicle state estimate, use the automatic differentiation method or numerical difference method to find the time derivative of the predicted state. S33. Construct a vehicle dynamics system based on the time derivative of the predicted state.

5. The physical information neural network vehicle system identification method according to claim 4, characterized in that, The vehicle state estimate for: ； in, This represents the temporal hidden feature vector output by the Transformer model. This represents the vehicle state estimate obtained from the temporal hidden feature vector; The vehicle dynamics system is: ； in, The derivative representing the vehicle's state at a given moment. This represents the estimated vehicle condition. This represents the vehicle control input vector. Represents the set of vehicle dynamics parameters. This represents the vehicle dynamics function.

6. The physical information neural network vehicle system identification method according to claim 1, characterized in that, The physical residual for: ； in, This represents the estimated vehicle condition. Represents the vehicle dynamics function. This represents the vehicle control input vector. This represents the set of vehicle dynamics parameters.

7. The physical information neural network vehicle system identification method according to claim 1, characterized in that, The joint loss function for training the Transformer model for: ； ； ； in, Indicates the data error term. Represents the physical constraint error term. This represents the weighting coefficient of the data error term. The weighting coefficients represent the physical constraint error terms. This represents the estimated vehicle condition. Indicates at time Measurement status, This represents the total number of sample points in the time series. Represents physical residuals.

Images