Vehicle cloud collaborative vehicle dynamics modeling method with incremental learning capability

Abstract

The application discloses a vehicle-cloud collaborative vehicle dynamics modeling method with incremental learning capability, and the method comprises the following steps: first, collecting and preprocessing driving information of a vehicle to be modeled at a vehicle end; second, constructing a neural network model based on Online-EWC incremental learning at a cloud end; third, pretraining the neural network at the cloud end; and fourth, continuously correcting the neural network at the cloud end. The method has the beneficial effects that the vehicle dynamics model can be continuously corrected to approach the real vehicle driving characteristics, the use of computing resources is significantly reduced, the modeling efficiency is improved, the vehicle dynamics model can be updated and optimized in a real-time environment, and the model can be adaptively updated along with the continuous accumulation and change of vehicle operation data.

Description

Technical Field

[0001] This invention relates to a vehicle-cloud collaborative vehicle dynamics modeling method, and more particularly to a vehicle-cloud collaborative vehicle dynamics modeling method with incremental learning capability. Background Technology

[0002] Currently, with the development and application of new information technologies such as cloud computing, the Internet of Things, big data, and artificial intelligence, as well as the increasing maturity of control theory, autonomous driving technology has received widespread attention in the automotive industry. As autonomous driving develops, accurate vehicle dynamics models become increasingly important. First, for intelligent vehicles, establishing an accurate vehicle dynamics model is the prerequisite and foundation for achieving precise vehicle control. Second, before the release of an autonomous driving system, extensive simulation testing is required. To obtain reliable test results, the vehicle dynamics model needs to be as close to reality as possible.

[0003] Traditional vehicle dynamics models are designed by experts based on physical principles. Models created in this way primarily employ parametric differential equations. Parameter identification for such models requires expensive equipment and time, and these models contain modeler assumptions, idealizing and simplifying the vehicle model, making it difficult to reflect the vehicle's dynamic response under conditions where these assumptions are not met, especially under extreme conditions, where model errors are significant. Therefore, researchers have attempted to model vehicle dynamics using data-driven methods. Currently, batch learning data-driven methods are used to construct vehicle dynamics models. However, due to the "catastrophic forgetting" characteristic of batch learning methods, they cannot effectively adapt to the dynamic growth of vehicle driving information data streams. Because real-world vehicle driving scenarios are diverse, comprehensive driving data is difficult to acquire all at once. Storing all vehicle driving information for training requires significant computing power and storage space. Furthermore, vehicle performance degrades over time. Batch learning methods struggle to comprehensively represent vehicle driving characteristics and update vehicle characteristics that decay over time.

[0004] Furthermore, there is currently a lack of automated modeling methods for vehicle dynamics modeling in a vehicle-cloud collaborative manner. Each car generates massive amounts of data daily, and when automakers sell hundreds of thousands, or even millions, of autonomous vehicles, the data volume will grow from petabytes (PB) to zettabytes (ZB), consuming significant resources for data acquisition, storage, and computation. Therefore, the importance of the cloud is increasingly evident for automakers. Since the dynamic characteristics of each car are different, an ideal data-driven vehicle dynamics modeling method should train a separate dynamics model for each car; therefore, automated modeling methods are necessary. Summary of the Invention

[0005] The main purpose of this invention is to solve the many problems existing in the use of traditional vehicle dynamics models, and to provide a vehicle-cloud collaborative vehicle dynamics modeling method with incremental learning capabilities.

[0006] The vehicle dynamics modeling method with incremental learning capability provided by this invention includes the following steps:

[0007] The first step is to collect and preprocess the vehicle's driving information on the vehicle side. This involves using on-vehicle sensors to collect the necessary driving information, and using an IMU sensor to collect the vehicle's longitudinal velocity v in real time. x Longitudinal acceleration a x Lateral velocity v y lateral acceleration a y In addition to the yaw rate ω, the front wheel steering angle δ is collected using a front wheel steering angle sensor, and the front brake pressure p is collected using a pressure sensor. f and rear brake pressure p r The torque T of the left drive wheel is collected using a wheel torque sensor. L and right drive wheel torque T R After collecting vehicle driving information, the data is preprocessed. The specific steps are as follows:

[0008] Step 1: Filter the collected data;

[0009] Step 2: Reshape the data;

[0010] Step 3: Use the shuffle function to randomly rearrange the dataset;

[0011] Step 4: Use StandardScaler to standardize the data, dividing the dataset into 70% training set, 15% validation set, and 15% test set.

[0012] Step 5: Upload the processed data to the cloud. After uploading, delete the data on the vehicle to save vehicle storage resources.

[0013] The second step is to build a neural network model based on Online-EWC incremental learning in the cloud. The neural network is a feedforward neural network. The first layer is the input layer, which has 40 features, including the control and state variables of the vehicle driving in the previous four time steps. The second layer is a fully connected layer FC1 with 100 hidden units. The third layer is the activation layer with the activation function set to Leaky ReLU. The fourth layer is a fully connected layer FC2 with 100 hidden units. The fifth layer is the activation layer with the activation function set to Leaky ReLU. The sixth layer is the output layer, which outputs the state variable of the next time step.

[0014] The EWC algorithm introduces a regularization term into the loss function to constrain the weights trained previously. This restricts the updates of previously trained weights during the current training process, thereby preserving knowledge from previous tasks. Online-EWC is a variant of EWC, a lightweight incremental learning method that does not require storing large amounts of historical data or models, thus reducing the consumption of storage and computing resources. At the same time, the Online-EWC algorithm can adapt to dynamically changing data streams, thereby preserving knowledge from previous data in constantly changing environments.

[0015] The vehicle dynamics model established above is a three-degree-of-freedom model, with the three degrees of freedom being lateral motion, longitudinal motion, and yaw motion;

[0016] The third step is to pre-train the neural network in the cloud. The neural network is pre-trained using the data processed on the vehicle. The TensorFlow framework is deployed in the cloud, and the learning rate, epoch and batch size are set. Then the pre-training process is carried out. The loss function used in the pre-training is mean squared error.

[0017] The fourth step involves continuous correction of the neural network in the cloud. Since vehicle dynamics parameters change over time and with the environment, and comprehensive vehicle driving data is difficult to obtain all at once, continuous correction of the neural network is necessary. Data uploaded to the cloud is input into the uncorrected neural network model for execution. Evaluation metrics are then used to assess the results. Data meeting the evaluation metrics requires no further processing and is directly deleted. Data not meeting the metrics is input into a buffer database. When the buffer database for a single vehicle reaches 1GB of storage space, incremental learning is performed on the data in the buffer database. During the correction process, the loss function is switched to the loss function corresponding to Online-EWC. Other hyperparameters use the same data as the pre-trained data. After incremental learning is complete, the learned data is deleted to save cloud storage resources.

[0018] The beneficial effects of this invention are:

[0019] This invention proposes a vehicle dynamics modeling method based on Online-EWC incremental learning. It enables continuous refinement of the constructed vehicle dynamics model, making it closer to the actual driving characteristics of vehicles. Vehicle dynamics changes with time, environment, and driving conditions. The Online-EWC incremental learning method allows the model to learn and update from new data in real time to adapt to changes in vehicle dynamics characteristics. This maintains a high degree of consistency with actual vehicle behavior, thereby improving the model's prediction accuracy. Batch learning for vehicle dynamics modeling requires periodic retraining of the entire model, which consumes significant computational resources and time. Incremental learning avoids this overhead; it only requires learning from new data to update the model's parameters. This significantly reduces computational resource usage and improves modeling efficiency.

[0020] This invention proposes a vehicle-cloud collaborative modeling method. It fully utilizes the computing power of both the vehicle and the cloud to automate vehicle dynamics modeling through collaborative vehicle-cloud collaboration. By rationally allocating workload between the vehicle and the cloud, it alleviates the equipment pressure on the vehicle. The automated modeling process significantly reduces the need for manual intervention and operation, thereby saving time and resources. It makes it possible to train a unique dynamics model for each vehicle. It enables real-time updates and optimization of the vehicle dynamics model, allowing for adaptive model updates as vehicle operating data accumulates and changes. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the method for vehicle dynamics modeling using vehicle-cloud collaborative technology as described in this invention.

[0022] Figure 2 This is a schematic diagram of the data reshaping process of the vehicle dynamics model described in this invention.

[0023] Figure 3 This is a schematic diagram of the neural network structure of the vehicle dynamics model described in this invention. Detailed Implementation

[0024] Please see Figures 1 to 3 As shown:

[0025] The vehicle dynamics modeling method with incremental learning capability provided by this invention includes the following steps:

[0026] The first step is to collect and preprocess the driving information of the vehicle to be modeled on the vehicle side. The specific steps are as follows:

[0027] Step 1: Collect driving information. The driver can then drive the vehicle according to their actual needs. The longitudinal velocity v of the vehicle is collected in real time using an IMU sensor. xLongitudinal acceleration a x Lateral velocity v y lateral acceleration a y And the yaw rate ω. The front wheel steering angle δ is collected using a front wheel steering angle sensor, and the front brake pressure p is collected using a pressure sensor. f and rear brake pressure p r The torque T of the left drive wheel is collected using a wheel torque sensor. L and right drive wheel torque T R .

[0028] Step 2: Filter the acquired data. The neural network used requires a data frequency of 50Hz. Generally, vehicle sensors operate at frequencies higher than 50Hz. For sensor data with frequencies higher than 50Hz, a Butterworth low-pass filter with a cutoff frequency of 50Hz is used to filter the data acquired from different sensors. The purpose is to unify the frequencies of data from different sensors, reduce high-frequency noise in the sensor data, and extract effective low-frequency information. For sensor data with frequencies lower than 50Hz, linear interpolation is used for resampling.

[0029] Step 3: Reshape the data, such as... Figure 2 As shown, where d t Includes the longitudinal velocity v of the vehicle at time t x,t Longitudinal acceleration a x,t Lateral velocity v y,t lateral acceleration a y,t yaw rate ω t Front wheel steering angle δ t Front brake pressure p f,t Rear brake pressure p r,t Left drive wheel torque T L,t and right drive wheel torque T R,t The dataset is divided into three parts: Data0 (d0, d1, d2, d3), Data1 (d1, d2, d3, d4), and so on. A shuffle function is used to randomly rearrange the dataset. StandardScaler is then used to standardize the data, scaling it to a mean of 0 and a standard deviation of 1. Standardization eliminates scale differences between features, ensuring different features have the same scale, which is beneficial for model training and performance. The dataset is divided into a 70% training set, a 15% validation set, and a 15% test set.

[0030] Step 4: Upload the preprocessed data to the cloud for use in subsequent steps. Delete vehicle-side data to save vehicle-side storage resources.

[0031] The second step is to build a neural network model based on Online-EWC incremental learning in the cloud:

[0032] The constructed neural network model structure is as follows Figure 3 As shown: The first layer is the input layer, which has 40 feature values, representing the longitudinal velocity v at four time points. x,t ,v x,t-1 ,v x,t-2 ,v x,t-3 Longitudinal acceleration a x,t ,a x,t-1 ,a x,t-2 ,a x,t-3 Lateral velocity v y,t ,v y,t-1 ,v y,t-2 ,v y,t-3 lateral acceleration a y ,a y,t-1 ,a y,t-2 ,a y,t-3 yaw rate ω t ,ω t-1 ,ω t-2 ,ω t-3 Front wheel steering angle δ t ,δ t-1 ,δ t-2 ,δ t-3 Front brake pressure p f,t ,p f,t-1 ,p f,t-2 ,p f,t-3 Rear brake pressure p r,t ,p r,t-1 ,p r,t-2 ,p r,t-3 Left drive wheel torque T L,t ,T L,t-1 ,T L,t-2 ,T L,t-3 Right drive wheel torque T R,t, T R,t-1 ,T R,t-2 ,T R,t-3 The second layer is a fully connected layer (FC1) with 100 hidden units. The third layer is an activation layer with the LeakyRule activation function. The fourth layer is a fully connected layer (FC2) with 100 hidden units. The fifth layer is an activation layer with the LeakyReLU activation function. The sixth layer is the output layer with 5 units, representing the vertical velocity v at the next time step. x,t+1 Longitudinal acceleration a x,t+1 Lateral velocity v y,t+1 lateral acceleration a y,t+1 and yaw rate ω t+1This model is a three-degree-of-freedom model, with the three degrees of freedom being the vehicle's longitudinal motion, lateral motion, and yaw motion. By inputting the control and state information from the previous four time steps, the model can predict the state information for the next time step.

[0033] The activation function used is LeakyReLU, and its calculation formula is as follows:

[0034] f(x) = max(ax,x);

[0035] The hyperparameter a was set to 0.1.

[0036] The loss function for the pre-trained neural network is chosen to be the mean square error (MSE).

[0037]

[0038] Where L MSE Let be the mean squared error, and y be the predicted value of the neural network. N represents the actual measurement value collected by the sensor, and N is the number of training set samples.

[0039] For the correction process of neural networks, the Online-EWC incremental learning algorithm needs to be applied to construct the loss function. Online-EWC is a lightweight continuous learning method that does not require storing large amounts of historical data or models, thus reducing the consumption of storage and computing resources. At the same time, the Online-EWC algorithm can adapt to dynamically changing data streams, thereby retaining knowledge from previous data in constantly changing environments. The specific formula for calculating its loss function is as follows:

[0040]

[0041] Among them, L O-EWC (θ) is the loss function of the Online-EWC incremental learning neural network, L MSE Let y be the mean squared error, K be the current update count, and y be the mean squared error. K The predicted value of the neural network during the Kth training iteration. N represents the actual measurement value collected by the sensor during the Kth training iteration. params θ is the number of parameters in the neural network, λ is the weight adjustment factor, and the value of λ is set to 15. i The optimized θ value after the Kth training iteration. The θ value is the value before the Kth training iteration. The hyperparameter γ is set to 0.9. Let be the i-th diagonal element of the Fisher information matrix. For regression problems such as vehicle dynamics modeling, this can be calculated as the total training samples in the K-th training iteration. The variance of the Fisher information matrix. The purpose of the Fisher information matrix is ​​to measure the importance of the weight parameters.

[0042] The Nesterov accelerated gradient (NAG) optimizer is selected, and its formula is as follows:

[0043]

[0044] θ = θ - v t

[0045] Where v t Let θ represent the momentum at the current moment, γ be the momentum factor (selected as 0.9), η be the learning rate (controlling the step size for parameter updates), and θ be the parameters currently being trained. It involves calculating gradient information.

[0046] The third step is to pre-train the neural network in the cloud. The specific steps are as follows:

[0047] Step 1: Deploy the code runtime environment in the cloud. Installing the incremental learning neural network runtime environment in the cloud requires deploying TensorFlow, NumPy, Keras, Joblib, SciPy, and scikit-learn.

[0048] Step 2: Set the initial learning rate to 0.002, batch size to 1000, and epochs to 1000. Use early stopping; if the change in the loss function value is less than 0.0001 for 50 consecutive epochs during training, the EarlyStopping function will stop the training process to prevent overfitting. After setting these parameters, pre-train the neural network.

[0049] The fourth step is to continuously refine the neural network in the cloud: After the neural network is pre-trained, the vehicle will repeat the first and fourth steps, that is, repeatedly collect and preprocess the driving information of the vehicle to be modeled and continuously refine the neural network so that the neural network model can continuously approach the real vehicle dynamics model.

[0050] In the fourth step, the data uploaded to the cloud needs to be input into the neural network model before correction. Then, an evaluation metric is used to assess the results. Data that meets the evaluation metric requires no further processing and can be directly deleted. Data that does not meet the evaluation metric is input into a buffer database. When the buffer database for a single vehicle reaches 1GB of storage space, incremental learning is performed on the data in the buffer database. This invention uses MSE as the evaluation metric. StandardScaler is used to standardize the data to eliminate scale differences between features, ensuring that different features have the same scale. When MSE ≤ 0.003, the data is considered to have passed the metric. When MSE > 0.003, the current neural network model is considered to have poor training performance on this data, and the data is input into the buffer database for subsequent incremental learning.

[0051] When performing neural network correction, the loss function needs to be switched to the loss function corresponding to Online-EWC, i.e., the L loss function introduced above. O-EWC For parameters (θ), all other hyperparameters can be obtained using the same data as the pre-training dataset. After incremental learning is complete, the learned data is deleted to save cloud storage resources.

Claims

1. A vehicle-cloud collaborative vehicle dynamics modeling method with incremental learning capability, characterized in that: The method includes the following steps: The first step is to collect and preprocess the driving information of the vehicle to be modeled at the vehicle end. The required vehicle driving information is collected using vehicle end sensors. The longitudinal speed v x , longitudinal acceleration a x , lateral speed v y , lateral acceleration a y , and yaw rate ω of the vehicle are collected in real time using an IMU sensor. The front wheel steering angle δ of the vehicle is collected using a front wheel steering angle sensor. The front brake brake pressure p f and rear brake brake pressure p r are collected using pressure sensors. The left drive wheel torque T L and right drive wheel torque T R are collected using wheel torque sensors. After collecting the driving information at the vehicle end, the data is preprocessed, and the specific steps are as follows: Step 1: Filter the collected data; Step 2: Reshape the data; Step 3: Use the shuffle function to randomly rearrange the dataset; Step 4: Use StandardScaler to standardize the data, dividing the dataset into 70% training set, 15% validation set, and 15% test set. Step 5: Upload the processed data to the cloud. After uploading, delete the data on the vehicle to save vehicle storage resources. The second step is to build a neural network model based on Online-EWC incremental learning in the cloud. The neural network is a feedforward neural network. The first layer is the input layer, which has 40 features, including the control and state variables of the vehicle driving in the previous four time steps. The second layer is a fully connected layer FC1 with 100 hidden units. The third layer is the activation layer with the activation function set to Leaky ReLU. The fourth layer is a fully connected layer FC2 with 100 hidden units. The fifth layer is the activation layer with the activation function set to Leaky ReLU. The sixth layer is the output layer, which outputs the state variable of the next time step. The EWC algorithm introduces a regularization term into the loss function to constrain the weights trained previously. This restricts the updates of previously trained weights during the current training process, thereby preserving knowledge from previous tasks. Online-EWC is a variant of EWC, a lightweight incremental learning method that does not require storing large amounts of historical data or models, thus reducing the consumption of storage and computing resources. At the same time, the Online-EWC algorithm can adapt to dynamically changing data streams, thereby preserving knowledge from previous data in constantly changing environments. The vehicle dynamics model is a three-degree-of-freedom model, with the three degrees of freedom being lateral motion, longitudinal motion, and yaw motion; The third step is to pre-train the neural network in the cloud. The neural network is pre-trained using the data processed on the vehicle. The TensorFlow framework is deployed in the cloud, and the learning rate, epoch and batch size are set. Then the pre-training process is carried out. The loss function used in the pre-training is mean squared error. The fourth step involves continuous correction of the neural network in the cloud. Since vehicle dynamics parameters change over time and with the environment, and comprehensive vehicle driving data is difficult to obtain all at once, continuous correction of the neural network is necessary. Data uploaded to the cloud is input into the uncorrected neural network model for execution. Evaluation metrics are then used to assess the results. Data meeting the evaluation metrics requires no further processing and is directly deleted. Data not meeting the metrics is input into a buffer database. When the buffer database for a single vehicle reaches 1GB of storage space, incremental learning is performed on the data in the buffer database. During the correction process, the loss function is switched to the loss function corresponding to Online-EWC. Other hyperparameters use the same data as the pre-trained data. After incremental learning is complete, the learned data is deleted to save cloud storage resources.

Images