A steer-by-wire control method based on a double-flow neural network

Abstract

This invention discloses a steer-by-wire control method based on a dual-stream neural network, relating to the field of autonomous driving steering control technology. It solves the technical problem of failing to achieve precise steer-by-wire control based on driver style. The key technical point is that it simultaneously considers and separately processes the driver's road preview information and vehicle dynamics perception information using a dual-stream neural network. Based on experimental data, it maps model parameters to driver style categories using a clustering method, thereby establishing a personalized driver steering behavior model. Based on this personalized model, different model parameters are automatically selected according to different driver style categories to achieve personalized steer-by-wire functionality, meeting the comfort requirements based on driver characteristics. This provides a fundamental steering drive system for the development of personalized advanced driver assistance systems and human-machine cooperative driving technology.

Description

Technical Field

[0001] This application relates to the field of autonomous driving steering control technology, and in particular to a steer-by-wire control method based on a dual-stream neural network. Background Technology

[0002] With the development of electronic technology, it is possible to replace the mechanical or hydraulic linkages in automotive steering systems with electronic sensors, controllers, and actuators. Unlike traditional steering systems, steer-by-wire systems offer many significant advantages. Typically, because there is no shaft between the handwheel and the steering wheel, potential dangers are mitigated in the event of a sudden accident. According to China's "Intelligent Connected Vehicle Technology Roadmap," the industrialization and application of intelligent steer-by-wire chassis systems will be achieved by 2025.

[0003] Steer-by-wire systems, with their advantages of variable steering and high integration, have been widely used in intelligent vehicle trajectory tracking and human-machine cooperative driving, greatly reducing the driver's burden and improving driving safety and handling stability. However, in-depth research into driver characteristics has revealed that users have different needs for vehicle dynamics based on their own characteristics, rather than the typical needs of drivers, resulting in individual driving expectations. Therefore, personalized driving systems have become a new trend in the automotive industry and a significant issue and challenge facing current intelligent vehicle human-machine cooperative driving technology.

[0004] Driving style is an important characteristic of a driver, describing a relatively stable habitual driving pattern formed in daily driving behavior. Driving style recognition methods mainly include rule-based methods and machine learning methods. Rule-based algorithms are limited by the number of feature parameters, resulting in poor stability and accuracy in classification. Machine learning methods are data-driven and can significantly improve the accuracy of driving style classification. This application further investigates steer-by-wire methods using machine learning. Summary of the Invention

[0005] This application provides a steer-by-wire control method based on a dual-stream neural network. Its technical purpose is to meet the comfort requirements based on driver characteristics and to provide a basic steering drive system for the development of personalized advanced driver assistance systems and human-machine cooperative driving technology.

[0006] The above-mentioned technical objective of this application is achieved through the following technical solution:

[0007] A steer-by-wire control method based on a dual-stream neural network is proposed. This method is implemented through a steering drive system, which includes sensors, a driver steering behavior model, and a steer-by-wire system. The method includes:

[0008] S1: Based on the vehicle and road data collected by the sensors, the far-field aiming point θ is determined.far and near the aiming point θ near Calculations are performed; wherein the vehicle and road data include lateral position deviation y. L Heading angle deviation The curvature ρ of the road ahead r Lateral velocity v y lateral acceleration a y and steering wheel angle δ d ;

[0009] S2: Pre-aiming point θ far Near the aiming point θ near Lateral velocity v y and lateral acceleration a y The input is given to the driver's steering behavior model to obtain the model's steering angle;

[0010] S3: Input the model's steering angle into the steer-by-wire system for steer-by-wire control.

[0011] The beneficial effects of this application are as follows: The steer-by-wire control method based on a dual-stream neural network described in this application simultaneously considers and processes the driver's road preview information and vehicle dynamics perception information separately through a dual-stream neural network. Based on experimental data, a clustering method is used to map model parameters to driver style categories, thereby establishing a personalized driver steering behavior model. Based on the personalized driver steering behavior model, different model parameters are automatically selected according to different driver style categories to realize personalized steer-by-wire function, meet the comfort requirements based on driver characteristics, and provide a basic steering drive system for the development of personalized advanced driver assistance systems and human-machine cooperative driving technology. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the driver steering behavior model based on a dual-stream neural network proposed in this application;

[0013] Figure 2 This is a flowchart of the personalized steer-by-wire control method proposed in this application. Detailed Implementation

[0014] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0015] The steer-by-wire control method based on a dual-stream neural network described in this application is implemented through a steering drive system, which includes sensors, a driver steering behavior model, and a steer-by-wire system, such as... Figure 2 As shown.

[0016] The specific steps of the steer-by-wire control method based on a two-stream neural network include:

[0017] S1: Based on the vehicle and road data collected by the sensors, the far-field aiming point θ is determined. far and near the aiming point θ near Calculations are performed; wherein the vehicle and road data include lateral position deviation y. L Heading angle deviation The curvature ρ of the road ahead r Lateral velocity v y lateral acceleration a y and steering wheel angle δ d .

[0018] Specifically, the far aiming point θ far and near the aiming point θ near They are represented as follows:

[0019]

[0020] Among them, y L Indicates lateral positional deviation; Indicates the heading angle deviation; ρ r Indicates the curvature of the road ahead.

[0021] S2: Pre-aiming point θ far Near the aiming point θ near Lateral velocity v y and lateral acceleration a y The steering angle of the model is obtained by inputting it into the driver steering behavior model.

[0022] In this embodiment, the driver steering behavior model includes a driver anticipation behavior network and a driver perception behavior network. Both the driver anticipation behavior network and the driver perception behavior network are composed of fully connected layers. Each fully connected layer includes an input layer, a hidden layer, and an output layer. The input layer includes two neurons, the hidden layer includes three neurons, and the output layer includes one neuron. The activation function of each neuron is the sigmoid function. Figure 1 As shown.

[0023] The input to the driver's aiming behavior network is the far aiming point θ. far and near the aiming point θ near The output is the pre-aiming steering δ. pre The input to the driver's perception-behavior network is the lateral velocity v. y and lateral acceleration a y The output is the perception steering δ. per .

[0024] Then the model's steering angle δ m The output of the driver steering behavior model, obtained by weighted summation of the outputs of the two networks, is expressed as:

[0025] δ m =w1*δ pre +w2*δ per ;

[0026] Here, w1 and w2 both represent weighting parameters. w1 represents the influence of the driver's path tracking ability on the driver's steering angle, and w2 represents the influence of the driver's perception ability on the driver's steering angle.

[0027] In this embodiment, the driver steering behavior model is trained based on the vehicle and road data collected in step S1, and the label for model training is the steering wheel angle δ. d Stochastic Gradient Descent (SGDM) with momentum was selected as the optimizer during training. The learning rate of SGDM was set to 0.005, the momentum was set to 0.9, the batch size was set to 30, and the loss function used during training was the mean squared error loss function.

[0028] Specifically, the driver steering behavior model is trained using data from different driver style categories to obtain the weights corresponding to each category. These driver style categories are obtained by clustering vehicle and road data collected by sensors, and include aggressive, general, and conservative styles. The clustering method used is K-means clustering. Before performing K-means clustering, the data is normalized using Min-max normalization, which unifies the data to a value between 0 and 1. Min-max normalization is expressed as:

[0029]

[0030] Where x represents the sample data to be normalized, max() represents the maximum value operation, and min() represents the minimum value operation. norm This represents the normalized sample data.

[0031] The training process for the driver steering behavior model consists of the following two steps:

[0032] Step 1: First, update the parameters of the entire driver steering model. Input all the data obtained in step S1 into the driver steering behavior model to train and update the model.

[0033] Step 2: Then, fix the parameters of the driver preview behavior network and the driver perception behavior network in the driver steering behavior model, and only update the weight parameters w1 and w2. Based on the clustering results, divide the collected vehicle and road data into three categories (aggressive, general, and conservative). Use the three categories of data as training data to train the driver steering behavior model. After training, three different sets of weight parameters w1 and w2 will be obtained, corresponding to three different driver styles. Subsequently, different weight parameters can be selected according to different driver styles to achieve personalized steer-by-wire.

[0034] S3: Input the model's steering angle into the steer-by-wire system for steer-by-wire control. After inputting the model's steering angle into the steer-by-wire system, the system performs personalized steering.

[0035] The above are exemplary embodiments of this application, and the scope of protection of this application is defined by the claims and their equivalents.

Claims

1. A steer-by-wire control method based on a dual-stream neural network, wherein the method is implemented through a steering drive system, the steering drive system comprising sensors, a driver steering behavior model, and a steer-by-wire system, characterized in that, The method includes: S1: Based on vehicle and road data collected by sensors, the far-field aiming point is determined. and near the aiming point Calculations are performed; wherein the vehicle and road data include lateral position deviations. Heading angle deviation Curvature of the road ahead lateral velocity lateral acceleration and steering wheel angle ; S2: Pre-aiming point Near the aiming point lateral velocity and lateral acceleration The input is given to the driver's steering behavior model to obtain the model's steering angle; S3: Input the model's steering angle into the steer-by-wire system for steer-by-wire control; In step S1, the far aiming point and near the aiming point They are represented as follows: ； in, Indicates lateral positional deviation; Indicates the deviation of the heading angle; Indicates the curvature of the road ahead; The driver steering behavior model includes a driver anticipation behavior network and a driver perception behavior network. The input of the driver anticipation behavior network is the far anticipation point. and near the aiming point The output is the pre-aiming steering. The input to the driver's perception-behavior network is lateral velocity. and lateral acceleration The output is perceived steering. The model's steering angle represents... for: ； in, and All represent weight parameters. This indicates the impact of the driver's path-following ability on the driver's steering angle. This indicates the impact of the driver's perception on the driver's steering angle.

2. The method as described in claim 1, characterized in that, Both the driver anticipation behavior network and the driver perception behavior network are composed of fully connected layers, which sequentially include an input layer, a hidden layer, and an output layer. The input layer includes 2 neurons, the hidden layer includes 3 neurons, and the output layer includes 1 neuron. The activation function of each neuron is the Sigmoid function.

3. The method as described in claim 1, characterized in that, The driver steering behavior model is trained using data from different driver style categories to obtain weights for each category. These driver style categories are obtained by clustering vehicle and road data collected by sensors, and include aggressive, general, and conservative styles.

4. The method as described in claim 3, characterized in that, The clustering method used is K-means clustering. Before performing K-means clustering, the data is normalized using Min-max normalization, as shown below: ； in, This indicates the sample data to be normalized. This indicates the operation of retrieving the maximum value. This indicates the operation of finding the minimum value. This represents the normalized sample data.

Images