Automatic driving track rapid tracking control method based on online neural network learning

Abstract

The invention is suitable for the technical field of automatic driving vehicle control, and provides an automatic driving track rapid tracking control method based on online neural network learning, which comprises the following steps: firstly, designing an adaptive model prediction controller, then making a data set and training a fuzzy radial basis function neural network; then, online learning of the fuzzy radial basis function neural network is carried out; and finally, accelerated optimization of the training process of the fuzzy radial basis function neural network is carried out. The automatic driving track rapid tracking controller based on online neural network learning provided by the invention has the remarkable advantages that the neural network model is trained by using control data of MPC (model predictive control), and a nonlinear relation is fitted to approximate the MPC, so that the calculation efficiency is effectively improved, and the rapidity of the controller is improved; the controller can collect real-time data, update a data set and learn new neural network weight parameters online in the process of tracking a reference trajectory, and shows better adaptability and precision compared with a traditional model prediction controller.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous vehicle control, and particularly relates to a fast tracking control method for autonomous driving trajectories based on online neural network learning. Background Technique

[0002] As a core development direction in the transportation field, autonomous driving technology is composed of multiple key modules such as environmental perception, trajectory planning, and trajectory tracking. Among them, the trajectory tracking module is directly related to the safety and comfort of driving, and is one of the core technologies for realizing autonomous driving. When performing trajectory tracking control on a dynamic system such as a vehicle with multi-objective and multi-constraint characteristics, model predictive control (MPC) has become a popular choice in the industry due to its excellent multi-variable control ability and constraint handling ability. However, traditional model predictive control methods have obvious limitations. On the one hand, the performance of MPC highly depends on an accurate system model. Although higher model accuracy can optimize the controller performance and meet the constraint conditions, it will significantly increase the solution time; reducing the model accuracy will lead to a decline in the controller performance. On the other hand, a model predictive controller with fixed parameters is difficult to adjust the control strategy in a timely manner according to complex road conditions in extreme working conditions such as high-speed driving, low adhesion coefficient roads, and large curvature curves, and has poor adaptability.

[0003] With the development of the machine learning field, fitting the MPC control strategy through data-driven methods such as neural networks can show better controller performance in scenarios where the neural network has learned. As the main method of constrained control, model predictive control provides a large amount of data to support neural network learning, and learning predictive control has become a suitable choice. To further improve the adaptive ability of the controller, researchers have added an online learning mechanism based on learning predictive control, aiming to dynamically update the model parameters by collecting the dynamic changes of road, traffic, and vehicle states in real time, and improve the adaptability of the controller in unknown scenarios. However, new problems have arisen. The online learning mechanism of the neural network involves complex operations such as online training of the neural network, a large amount of data collection and transmission, and forward propagation and backpropagation, resulting in a significant increase in the controller's calculation time and unable to meet the real-time requirements. This makes traditional model predictive controllers always face the dilemma of being difficult to balance adaptability and rapidity in the field of autonomous driving trajectory tracking. In view of this, the present invention proposes a fast tracking control method for autonomous driving trajectories based on online neural network learning. Summary of the Invention

[0004] The purpose of the present invention is to provide a fast tracking control method for autonomous driving trajectories based on online neural network learning, aiming to solve the problems raised in the above background technique.

[0005] The object of the present invention is achieved by the following technical solutions: A fast tracking control method for an autonomous driving trajectory based on online neural network learning, comprising the following steps: Step 1: Design of an adaptive model predictive controller; Construct a prediction model based on the vehicle longitudinal and lateral coupling model, and adaptively adjust the weight parameters of the objective function by using a look-up table method based on Bayesian optimization; Step 2: Dataset production and training of a fuzzy radial basis neural network; Collect the vehicle body state and control output data under different working conditions, and determine the network parameters through normalization processing and K-means clustering; Step 3: Online learning of the fuzzy radial basis neural network; Introduce an online learning mechanism, update the dataset according to real-time control feedback, and dynamically adjust the neural network parameters through an error trigger mechanism; Step 4: Accelerated optimization of the training process of the fuzzy radial basis neural network; Perform weight pruning and matrix decomposition acceleration optimization on the neural network to improve the online learning efficiency.

[0006] Furthermore, in the said Step 1, the vehicle longitudinal and lateral coupling model is as follows: ; Wherein, is the change rate of the longitudinal speed in the vehicle body coordinate system; is the change rate of the lateral speed in the vehicle body coordinate system; is the change rate of the yaw angular velocity; is the change rate of the yaw angle; is the change rate of the vehicle longitudinal displacement; is the change rate of the vehicle lateral displacement; and are respectively the longitudinal speed and the lateral speed in the vehicle body coordinate system; is the yaw angular velocity of the vehicle; is the longitudinal acceleration of the vehicle; is the front wheel steering angle of the vehicle; and are respectively the lateral forces of the front wheel and the rear wheel; is the vehicle mass; is the moment of inertia of the vehicle; is the distance from the vehicle center of mass to the front axle; is the distance from the vehicle center of mass to the rear axle; is the yaw angle; The prediction model is obtained by discretizing the obtained vehicle longitudinal and lateral coupling model; select both the prediction time domain and the control time domain as p , then the prediction model is as follows: ; wherein, are the predicted values of the state variables at different times under the same prediction horizon; are the control variables at different times under the same prediction horizon; is the prediction model in model predictive control.

[0007] Furthermore, in the said step 1, the optimization objective function of the adaptive model predictive controller is: ; wherein, is the objective function; is the length of the prediction horizon; is the time step index within the prediction horizon; is the predicted state variable at the next moment; is the state variable of the reference trajectory within the prediction horizon; is the change amount of the control variables at adjacent moments; are the weight parameters of the tracking performance and the smooth performance in the objective function, respectively, and are dynamically adjusted by the look-up table method based on Bayesian optimization.

[0008] Furthermore, in the said step 2, the fuzzy radial basis neural network is composed of an input layer, a fuzzification layer, a fuzzy inference layer and an output layer; The input layer receives the vehicle body state data, and the input-output relationship of the input layer is: ; wherein, is the output of the i th neuron in the input layer; is the input data of the input layer; The fuzzification layer uses the Gaussian membership function to convert the input into a fuzzy quantity, and the connection function between the th node in the input layer and the th node in the fuzzification layer is expressed as: ; wherein, is the fuzzification result of the th neuron in the fuzzification layer for the output of the th neuron in the input layer; respectively represent the center and the base width of the membership function of the th input variable in the th fuzzy set; The fuzzy inference layer executes 81 fuzzy rules, and the input-output relationship of the fuzzy inference layer is: ; wherein, is the output of the th neuron in the fuzzy inference layer; is the index of the neuron in the fuzzy inference layer; The output layer converts the fuzzy quantity output by the fuzzy inference layer into a clear quantity. The input-output relationship of the output layer is: ; where is the output of the th neuron in the output layer; is the weight matrix between the fuzzy inference layer and the output layer.

[0009] Furthermore, in the step 2, the training of the fuzzy radial basis neural network includes: Determining the center vector through K-means clustering and iterative optimization, and obtaining the width vector from the following formula; ; where is the number of neurons in the fuzzification layer; After determining the center vector and the width vector , the weight matrix between the fuzzy inference layer and the output layer is obtained by the least squares method to complete the network training; the formula is as follows: ; where is the output matrix of the fuzzy inference layer, the number of rows is the number of samples, and the number of columns is the number of neurons in the fuzzy inference layer; is the transpose; is the output matrix of the output layer, the number of rows is the number of neurons in the output layer, and the number of columns is the number of samples.

[0010] Furthermore, in the step 3, the online learning mechanism includes: Collecting new samples in real time, and judging whether to replace, add or delete the original samples in the dataset through the Euclidean distance; the distance calculation formula is as follows: ; where is the number of samples in the dataset; is the Euclidean distance between the real-time sample and the th sample in the dataset; is the corresponding feature difference in the sample; Measuring the deviation between the vehicle body state and the reference trajectory with the error, and the error calculation formula is as follows: ; where the error It consists of two parts: lateral displacement error and yaw angle error; They are the weight coefficients of the lateral displacement error and the yaw angle error respectively; is the lateral displacement of the vehicle; is the yaw angle; is the lateral displacement in the reference trajectory; is the yaw angle in the reference trajectory; When the error exceeds , it triggers the update of the neural network parameters.

[0011] Furthermore, in the step 4, the specific operation of weight pruning is as follows: Compare the weight column and with the dynamic standard value , and prune or scale the weight matrix; The pruning formula is as follows: ; ; ; Among them, is the function of weight pruning; is the weight column sum divided by the standard value except the proportion of; is the row and the column weight; Adopt the Winograd algorithm to convert matrix multiplication into addition operation, which reduces the number of multiplication operations after conversion and accelerates network training; After accelerated training, the real-time updated fuzzy radial basis neural network model is as follows: ; Among them, is the control quantity output by the fuzzy radial basis neural network model at time is the fuzzy radial basis neural network model; is the vehicle body state quantity at time.

[0012] Compared with the prior art, the beneficial effects of the present invention are: The fast tracking controller for autonomous driving trajectories based on online neural network learning proposed by the present invention has significant advantages. On the one hand, within each sampling period, when the MPC calculates the control output by predicting future states and solving the optimization problem, if the scale of the optimization problem is large or the control frequency is high, solving the optimization problem will consume a large amount of computing resources, affecting the rapid implementation. However, the present invention uses the control data of the MPC to train the neural network model, fitting the non-linear relationship to approximate the MPC, effectively improving the computing efficiency and enhancing the rapidity of the controller. On the other hand, during the process of tracking the reference trajectory, this controller can collect real-time data, update the data set, and online learn new neural network weight parameters, showing better adaptability and accuracy compared with the traditional model predictive controller. Description of the Drawings

[0013] Figure 1 is the flowchart of the method of the present invention.

[0014] Figure 2 is the schematic diagram of the update of the weight parameters of the objective function.

[0015] Figure 3 is the flowchart of the online data set update mechanism.

[0016] Figure 4 is the comparison diagram of the trajectory tracking effect.

[0017] Figure 5 is the comparison diagram of the yaw angle tracking effect.

[0018] Figure 6 is the comparison diagram of the longitudinal speed tracking effect. Detailed Embodiment

[0019] For a clearer understanding of the technical features, objectives, and beneficial effects of the present invention, the technical solution of the present invention will be described in detail below, but it should not be construed as a limitation on the implementable scope of the present invention.

[0020] The following describes the specific implementation of the present invention in detail with specific embodiments.

[0021] The present invention provides a fast tracking control method for autonomous driving trajectories based on online neural network learning, and its flowchart is as Figure 1 shown, and the method includes the following steps: Step 1: Design of the adaptive model predictive controller; The present invention takes a three - degree - of - freedom vehicle mechanism model as the research object and makes the following assumptions: (1) Ignore the vertical motion of the vehicle, the lateral load transfer of the tires, and the influence of vehicle aerodynamics; (2) Assume the vehicle system is rigid and ignore the coupling effect of the suspension on the vehicle. Under these two assumptions, the vehicle model is simplified to a bicycle model that can accurately describe the vehicle's behavior under extreme driving conditions. Considering the physical interaction between the tires and the road, including friction, bending stiffness, and tire lateral stiffness, etc., the corresponding vehicle dynamics model is as follows: ; Wherein, is the change rate of the longitudinal velocity in the body coordinate system; is the change rate of the lateral velocity in the body coordinate system; is the change rate of the yaw angular velocity; and are the longitudinal velocity and the lateral velocity in the body coordinate system respectively; is the yaw angular velocity of the vehicle; is the longitudinal acceleration of the vehicle; is the front wheel steering angle of the vehicle; and are the lateral forces of the front wheel and the rear wheel respectively; is the vehicle mass; is the moment of inertia of the vehicle; is the distance from the vehicle's center of mass to the front axle; is the distance from the vehicle's center of mass to the rear axle. The motion equations of the vehicle's longitudinal and lateral velocities in the earth coordinate system are as follows: ; Wherein, is the change rate of the vehicle's longitudinal displacement; is the change rate of the vehicle's lateral displacement; is the yaw angle.

[0022] For the physical model of the tire, which is in direct contact with the ground when the vehicle is running, the present invention selects the Magic Formula model, and its lateral force calculation formula is: ; Wherein, is the maximum value factor; is the curve shape factor; is the stiffness factor; is the longitudinal stiffness when the longitudinal force is 0; and are the tire sideslip angles of the front and rear wheels respectively. Since the calculation of the inverse trigonometric function part in the Magic Formula is too complex, its Taylor expansion is carried out now, and only the linear part is retained. Then the linear relationship between the lateral force and the tire sideslip angle is as follows: ; Among them, are the cornering stiffnesses of the front and rear wheels respectively.

[0023] Based on the above formula, the present invention selects the vehicle body state variables as , where represent the vehicle longitudinal speed, lateral speed, yaw rate, yaw angle, longitudinal displacement, and lateral displacement in sequence, selects the front wheel steering angle and the longitudinal acceleration as the control variables , then the vehicle longitudinal and lateral coupling model (vehicle mechanism model) can be written as: ; Among them, is the change rate of the yaw angle.

[0024] According to the obtained vehicle longitudinal and lateral coupling model, a prediction model is obtained by discretization. The prediction horizon and control horizon are both selected as p , then the prediction model is as follows: ; Among them, are the predicted values of the state variables at different times in the same prediction horizon respectively; are the control variables at different times in the same prediction horizon respectively; is the prediction model in model predictive control.

[0025] After obtaining the prediction model, the control objective is: under the condition of satisfying the actuator constraints, find a suitable state variable output to track the given trajectory while ensuring the vehicle is stable. The form of the optimization problem is as follows: ; Among them, is the actual state variable at the current time; is the control variable at the current time; is the objective function; is the predicted state variable at the next time; is the prediction model in model predictive control; is the control variable within the prediction horizon; are the lower and upper limits of the control variable respectively; is the time step index within the prediction horizon; is the length of the prediction horizon.

[0026] The objective function is selected as: ; Among them, is the state variable of the reference trajectory within the prediction horizon; is the change in the control quantity at adjacent moments; are the weight parameters of the tracking performance and the smooth performance in the objective function respectively. The first term in the objective function represents the tracking performance of the controller, and the latter term represents the smooth performance of the controller. For the weight parameters in the objective function, the present invention adopts a look-up table method based on Bayesian optimization for adaptive adjustment, and the specific process is as Figure 2 shown. The method of Bayesian optimization is used to obtain the optimal weight values under each working condition, and a Bayesian optimization parameter table is formed in the form of a group of optimal weight values corresponding to each working condition interval. In the actual tracking control process, the optimal weight values are quickly found by using the look-up table method according to the current vehicle body state and assigned to the objective function, so as to realize the design of the adaptive model predictive controller.

[0027] Step 2: Dataset production and fuzzy radial basis neural network training; Set several trajectories under different working conditions, use the designed adaptive model predictive controller for trajectory tracking control, and finally collect the vehicle body state data and the corresponding control output of the controller as the data samples required for training the neural network, and then produce the dataset. To improve the generalization ability of the neural network, when collecting data in the present invention, various working conditions with different speeds, different adhesion coefficients, and different road curvatures are emphatically selected for trajectory tracking control. Each group of samples consists of the vehicle body state and the control output, and the form is . Since the numerical values between different physical quantities vary greatly, overfitting is likely to occur during training. Therefore, the data samples need to be preprocessed by normalization before training, and the normalization formula is as follows: ; where is the normalized data; is the original data; is the minimum value in the sample number; is the maximum value in the sample number; Through the above formula, each numerical value in the data sample is converted to between [0,1].

[0028] After the samples are processed, the neural network is trained. The neural network selected in the present invention is a fuzzy radial basis function neural network (FRBFNN). Fuzzy logic can convert precise input data into the membership degree of a fuzzy set, enhancing the interpretability of decision-making; the radial basis neural network can approximate any complex nonlinear relationship, and its parameters can be separately optimized with high computational efficiency. The fuzzy radial basis neural network consists of an input layer, a fuzzification layer, a fuzzy inference layer, and an output layer.

[0029] The input layer is the input of the neural network model. State variable data is selected from the samples in the dataset as the input layer data. The number of neurons in this layer is 4. The input-output relationship of this layer is: ; where is the output of the th neuron in the input layer; is the input data of the input layer.

[0030] The second layer is the fuzzification layer, also known as the membership function layer. The main function of this layer is to convert each input quantity into a fuzzy quantity. In the present invention, a Gaussian function is selected as the membership function, the fuzzy universe of discourse is set to [-1, 1], and the fuzzy subsets include three categories: L, M, H. Then the number of neurons in the fuzzification layer is 12. The connection function between the i th node in the input layer and the j th node in the fuzzification layer is expressed as: ; where is the fuzzification result of the th neuron in the fuzzification layer for the output of the th neuron in the input layer; respectively represent the center and base width of the membership function of the rd input variable in the th fuzzy set.

[0031] The third layer is the fuzzy inference layer. Each neuron in this layer corresponds to a fuzzy rule respectively, and its main function is to perform the operation of the fuzzy rule. Since each input quantity has 3 fuzzy subsets, there are 81 neurons in this layer, corresponding to 81 fuzzy rules respectively. The input-output relationship of this layer is: ; where is the output of the th neuron in the fuzzy inference layer; is the index of the neuron in the fuzzy inference layer.

[0032] The fourth layer is the output layer. Each neuron in this layer corresponds to an output respectively, and converts the fuzzy quantity output by the fuzzy inference layer into a clear quantity. The input-output relationship of this layer is: ; where is the output of the th neuron in the output layer; is the weight parameter between the fuzzy inference layer and the output layer.

[0033] Next, train each parameter of the neural network. The parameters to be trained include: the center vector, width vector of the fuzzification layer, and the weight matrix between the fuzzy inference layer and the output layer. The method selected in the present invention for determining the center vector is the K-means clustering method, which divides the data set into K data clusters. After a series of iterations, the data points within the same cluster are made as similar as possible, and the data points in different clusters are made as different as possible. First, randomly assign each sample in the data set to each data cluster, and randomly select a sample in each data cluster as the center point, and perform random initialization on it. Then, calculate the Euclidean distance between each sample in the data set and all the center vectors in turn to divide the rationality of the data clusters. The Euclidean distance calculation formula is as follows: ; Wherein, is the Euclidean distance between the th sample and the th center vector; is the th center vector; represents the serial number of a single sample in each data cluster; is the th Euclidean distance between the th sample and each center vector; is the th sample; is the center vector; is the number of data clusters. Re-divide the data clusters and the center vectors according to the calculated Euclidean distance, assign the sample to the data cluster where the closer center vector is located, and take the sample with the shortest Euclidean distance from other samples in the new data cluster as the new center vector, and continuously iterate and repeat this process until the center vector reaches the specified accuracy or reaches the predicted number of iterations. The iteration accuracy selected in the present invention is 0.001, and the number of iterations is 300. According to the center vector the width vector is obtained by the following formula: ; Wherein, is the number of neurons in the fuzzification layer.

[0034] Determine the center vector and the width vector . After that, the weight matrix between the fuzzy inference layer and the output layer is obtained by the least squares method to complete the network training; the formula is as follows: ; Wherein, is the output matrix of the fuzzy inference layer, the number of rows is the number of samples, and the number of columns is the number of neurons in the fuzzy inference layer; is the transpose; is the output matrix of the output layer, with the number of rows being the number of neurons in the output layer and the number of columns being the number of samples. When the center vector, width vector, and weight matrix are trained, the neural network model can be determined.

[0035] Step 3: Online learning of the fuzzy radial basis neural network; After the neural network model is determined, an online learning mechanism for the neural network is introduced to adjust the neural network parameters in real time according to the control feedback. The online learning mechanism designed in the present invention is based on the offline training data set, collects new data samples in the control process and adds them to the data set, and then retrains the weight matrix of the neural network according to the data set triggering mechanism, so as to realize the online update of the neural network parameters. The process of the online data set update mechanism is as Figure 3 shown.

[0036] The update methods of the samples in the data set of the present invention include replacement, addition, and deletion. For the data set obtained in Step 2, in the actual control process, calculate the Euclidean distance between the newly collected real-time sample and the original sample points in the data set D , find the data sample closest to the new sample for replacement, and the distance calculation formula is as follows: ; where, is the number of samples in the data set; is the Euclidean distance between the real-time sample and the th sample in the data set; is the corresponding feature difference in the sample. After sample replacement, the generalization ability of the data set is enhanced, and the phenomenon of overfitting in the training process is avoided. The present invention determines the distance judgment criterion as 0.007. When , replace the data; when , add the data to the data set. In the actual control process, if the reference trajectory differs greatly from the reference trajectory in the data set, it may lead to the number of new data added to the data set being much larger than the number of original samples replaced in the data set, causing the data set capacity to increase continuously and the computational burden to increase continuously. Therefore, delete the data in the data set that has a too large difference from the new data sample, so that the data set can better adapt to the current control scenario. Specifically, when the data set capacity reaches 20000, each time the data set is updated, delete the original data set samples of , so as to ensure the training effect while reducing the training burden.

[0037] Since the learning process of the neural network is very time-consuming, the present invention updates the data set in real time during the control process, but the online learning adopts a trigger mechanism. The present invention adopts an error trigger mechanism to judge the deviation between the current vehicle body state and the reference trajectory. When the deviation is too large, it means that the neural network parameters no longer adapt to the current control scenario, and parameter update is required, thereby triggering online learning. The error calculation formula is as follows: ; where the error is composed of two parts: the lateral displacement error and the yaw angle displacement error; are the weight coefficients of the lateral displacement error and the yaw angle error respectively. Considering the different graduation values of the two, the present invention sets to 1, to 10; is the lateral displacement in the reference trajectory; is the yaw angle in the reference trajectory. When the error exceeds , the online learning mechanism of the neural network is triggered, and the neural network parameters are learned according to the current latest data set to improve the adaptability of the neural network controller.

[0038] Step 4: Accelerated optimization of the fuzzy radial basis neural network training process; In view of the large computational burden of the neural network online learning, it is necessary to process the neural network itself to achieve the compression and acceleration of the neural network.

[0039] (1) Weight pruning: First, perform weight pruning on the weight matrix of the neural network. During the neural network training process, the weight matrix consists of i rows j columns of weight values. The pruning formula adopted by the present invention is as follows: ; ; ; where is the function of weight pruning; is the column sum of the weight values except for the standard value accounting for of is the weight on the th row and the th column; is the th column sum of the weight values; is the standard value for deleting weights, Not a predefined constant, but adaptively changes with training iterations. The pruning degree can be determined by artificially setting the pruning rate. In the present invention, a pruning rate of 0.8 is selected. At this time is the 16th maximum value of n and changes with each iteration update of the weights; is the number of weights included in each column of the weight matrix. When the sum of the weights in a certain column is less than the standard value is greater than at this time, use to recalculate the weights of this column, and then prune the columns with zero weights, so as to realize the pruning and compression of the weight matrix.

[0040] (2) Matrix decomposition: Since there are a large number of matrix multiplication operations in the neural network, it will occupy a large amount of computing resources during training, especially for the online learning of the neural network, there are requirements for the rapidity of training. The present invention adopts the Winograd algorithm to convert matrix multiplication into addition operations, and puts a part of the converted addition operations into the pre-processing stage of the convolutional weights of the output layer, so as to realize the acceleration and optimization of matrix multiplication calculation. The principle of the Winograd algorithm is shown in the following formula. When there is an output matrix Y equal to the input matrix X multiplied by the matrix A , there is also a matrix B satisfying: ; where: is the element of the matrix A ; is the element of the matrix B ; is the element of the input matrix X ; is the element of the output matrix Y . When the input matrix X is mapped to the output matrix A through the matrix Y , 6 multiplication operations are required, while only 4 multiplication operations are required when mapping through the matrix B . Since the time-consuming of addition operations is much less than that of multiplication calculations, the Winograd algorithm effectively improves the speed of matrix operations, and thus realizes the acceleration and optimization of the online learning of the neural network.

[0041] After accelerated training, the real-time updated fuzzy radial basis neural network model is as follows: ; where, is The control quantity output by the time-fuzzy radial basis neural network model; is the fuzzy radial basis neural network model; is the vehicle body state quantity at a certain moment. By giving the reference trajectory and the current state of the vehicle body, the best control strategy suitable for the current control scenario can be obtained.

[0042] Example 1: To verify the advantages of the online neural network learning-based autonomous driving adaptive model predictive controller (hereinafter referred to as the "neural network adaptive controller") in terms of adaptability and rapidity, it is compared with the traditional model predictive controller (with fixed parameters) in the trajectory tracking of the high-speed double lane change condition under variable friction coefficients. The given vehicle speed is 70 km / h, and the road surface friction coefficient switches continuously between ≥0.4 and <0.8. The controller parameters are as follows: the sampling time is 20 ms, and both the control horizon and the prediction horizon are set to 18.

[0043] Figure 4 、 Figure 5 and Figure 6 respectively show the comparison diagrams of the trajectory, yaw angle, and longitudinal speed tracking effects of the traditional model predictive controller and the neural network adaptive controller. From Figure 4 it can be seen that compared with the model predictive controller, the neural network adaptive controller fits the reference trajectory more closely, and can follow the expected trajectory more accurately under the complex friction coefficient change condition, reflecting good adaptability. From Figure 5 it can be seen that during the process of tracking the reference yaw angle, the fluctuation of the neural network adaptive controller is smaller than that of the model predictive controller, and it can approach the reference value faster and more stably, indicating that it has better rapidity and adaptability in yaw angle control. From Figure 6 it can be seen that the neural network adaptive controller is more stable than the model predictive controller in longitudinal speed tracking, and can maintain well near the reference longitudinal speed, indicating that it performs better in longitudinal speed control. To sum up, through the comparison of the tracking effects in the three aspects of trajectory, yaw angle, and longitudinal speed, the neural network adaptive controller has better control effects than the model predictive controller with fixed parameters, verifying its advantages in adaptability and rapidity.

[0044] The above is only the preferred implementation mode of the present invention. It should be pointed out that for those skilled in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicability of the patent.

Claims

1. An automatic driving trajectory fast tracking control method based on online neural network learning, characterized in that: The following steps are involved: Step 1: Adaptive model predictive controller design; A prediction model is constructed based on the vehicle lateral and longitudinal coupling model, and a table lookup method based on Bayesian optimization is used to adaptively adjust the weight parameters of the objective function; Step 2: Data set preparation and fuzzy radial basis neural network training; Collect vehicle body status and control output data under different working conditions, and determine network parameters through normalization and K-means clustering; Step 3: Online learning of fuzzy radial basis neural network; Introducing an online learning mechanism to update the data set based on real-time control feedback and dynamically adjust the neural network parameters through an error-triggered mechanism; Step 4: Accelerated optimization of the fuzzy radial basis neural network training process; Perform weight pruning and matrix decomposition acceleration optimization on neural networks to improve online learning efficiency.

2. The automatic driving trajectory fast tracking control method based on online neural network learning according to claim 1 is characterized in that: In step 1, the vehicle lateral and longitudinal coupling model is as follows: ； in, is the rate of change of the longitudinal velocity in the vehicle body coordinate system; is the rate of change of the lateral velocity in the vehicle body coordinate system; is the rate of change of yaw angular velocity; is the rate of change of the yaw angle; is the rate of change of the vehicle longitudinal displacement; is the rate of change of vehicle lateral displacement; and are the longitudinal speed and lateral speed in the vehicle body coordinate system respectively; is the yaw rate of the vehicle; is the longitudinal acceleration of the vehicle; is the front wheel turning angle of the vehicle; and are the lateral forces of the front and rear wheels respectively; is the vehicle mass; is the moment of inertia of the vehicle; is the distance from the vehicle's center of mass to the front axle; is the distance from the vehicle's center of mass to the rear axle; is the yaw angle; The prediction model is obtained by discretizing the vehicle lateral and longitudinal coupling model; the prediction time domain and control time domain are selected as p , the prediction model is as follows: ； in, They are respectively the predicted values ​​of the state quantity at different times in the same prediction time domain; They are the control quantities at different times in the same prediction time domain; It is the prediction model in model predictive control.

3. The automatic driving trajectory fast tracking control method based on online neural network learning according to claim 1 is characterized in that: In step 1, the optimization objective function of the adaptive model predictive controller is: ； in, is the objective function; To predict the time domain length; is the time step index in the prediction domain; is the predicted state quantity at the next moment; is the state quantity of the reference trajectory in the prediction time domain; is the change of the control quantity at adjacent moments; They are the weight parameters of tracking performance and stable performance in the objective function, respectively, and are dynamically adjusted through a table lookup method based on Bayesian optimization.

4. The automatic driving trajectory fast tracking control method based on online neural network learning according to claim 1 is characterized in that: In step 2, the fuzzy radial basis neural network consists of an input layer, a fuzzification layer, a fuzzy reasoning layer and an output layer; The input layer receives the vehicle body status data. The input and output relationship of the input layer is: ； in, The input layer i The output of a neuron; Is the input data of the input layer; The fuzzification layer uses Gaussian membership function to convert the input into fuzzy quantity. Nodes and fuzzy layers The connection function of each node is expressed as: ； in, The fuzzy layer The neurons in the input layer The fuzzification result of the output of each neuron; Respectively represent The input variable is The center and basis width of the membership function of a fuzzy set; The fuzzy reasoning layer executes 81 fuzzy rules. The input and output relationship of the fuzzy reasoning layer is: ； in, The fuzzy reasoning layer The output of a neuron; is the index of the neuron in the fuzzy inference layer; The output layer converts the fuzzy quantity output by the fuzzy reasoning layer into a clear quantity. The input and output relationship of the output layer is: ； in, The output layer The output of a neuron; It is the weight matrix between the fuzzy inference layer and the output layer.

5. The automatic driving trajectory fast tracking control method based on online neural network learning according to claim 1 is characterized in that: In step 2, the training of the fuzzy radial basis neural network includes: Determine the center vector through K-means clustering and iterative optimization , the width vector is obtained by the following formula ; ； in, is the number of neurons in the fuzzification layer; Determine the center vector and the width vector After that, the weight matrix between the fuzzy inference layer and the output layer Obtained by the least squares method, network training is completed; the formula is as follows: ； in, is the output matrix of the fuzzy inference layer, the number of rows is the number of samples, and the number of columns is the number of neurons in the fuzzy inference layer; is transposed; is the output matrix of the output layer, the number of rows is the number of neurons in the output layer, and the number of columns is the number of samples.

6. The automatic driving trajectory fast tracking control method based on online neural network learning according to claim 1 is characterized in that: In step 3, the online learning mechanism includes: New samples are collected in real time, and the original samples in the data set are replaced, added or deleted based on the Euclidean distance. The distance calculation formula is as follows: ； in, is the number of samples in the data set; For real-time samples and data sets The Euclidean distance of samples; is the corresponding feature difference in the sample; The error is used to measure the deviation between the vehicle body state and the reference trajectory. The error calculation formula is as follows: ； Among them, the error It consists of two parts: lateral displacement error and yaw angle error; are the weight coefficients of lateral displacement error and yaw angle error respectively; is the lateral displacement of the vehicle; is the yaw angle; is the lateral displacement in the reference trajectory; is the yaw angle in the reference trajectory; when the error Exceed , triggers the neural network parameter update.

7. The automatic driving trajectory fast tracking control method based on online neural network learning according to claim 1 is characterized in that: In step 4, the specific operation of weight pruning is: Compare the weight columns and With dynamic standard value , prune or scale the weight matrix; the pruning formula is as follows: ； ； ； in, is the function of weight pruning; is the weight column and Divide the standard value Outside The proportion of For the Row and Weights on columns; The Winograd algorithm is used to convert matrix multiplication into addition operations, which reduces the number of multiplication operations and speeds up network training. After accelerated training, the real-time updated fuzzy radial basis neural network model is as follows: ； in, for The control quantity output by the fuzzy radial basis function neural network model at each moment; It is a fuzzy radial basis neural network model; for The vehicle body state at the moment.

Images