Intelligent vehicle path tracking control method based on double network iterative optimization

Abstract

The intelligent automobile path tracking control method based on double network iterative optimization aims at solving the problems of large calculation amount and insufficient real-time performance of model predictive control in path tracking tasks.The present application relates to the field of automatic driving, and is characterized in that the method comprises a subspace encoder network module and a control input increment sequence generation network module, and the specific implementation comprises two nested loop processes of predicted output sequence updating and control input sequence updating.The predicted output sequence is recursively generated through the predicted output sequence updating process, the lateral position error sequence is obtained through the control input sequence updating process, and the control input sequence is iteratively optimized.The two loop updating processes of predicted output sequence updating and control input sequence updating are repeated until the lateral position error sequence meets the preset threshold.The first item in the control input sequence is finally output as the actual input of the vehicle system, thereby realizing the path tracking task of the intelligent driving vehicle.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of automatic driving, in particular to an intelligent vehicle path tracking control method based on double network iterative optimization. BACKGROUND

[0002] With the continuous development of automatic driving technology, intelligent vehicles have become an important direction to improve traffic safety and efficiency. Path tracking, as a core technology to ensure the precise driving of intelligent vehicles, is directly related to the stability and reliability of their operation.

[0003] In the field of intelligent vehicle path tracking control, model predictive control (MPC) generates control input sequences by solving multi-objective optimization problems, enabling the vehicle to follow a predetermined trajectory. MPC can effectively handle complex optimization problems with constraints, ensuring the safety of intelligent vehicle path tracking. However, MPC has a heavy computational burden and real-time performance is difficult to guarantee, especially when facing complex and variable driving environments, it is easy to fall into a local optimal solution, resulting in unsatisfactory control effect. In addition, MPC usually relies on quadratic programming solvers, and as the number of constraints increases, the computational complexity increases exponentially, limiting its application in complex scenarios. At the same time, neural network (NN) methods, with their powerful nonlinear modeling capabilities and fast reasoning characteristics, have been applied to path tracking control to improve the real-time response speed and scene adaptability of the system. However, NN methods still have shortcomings in robustness, model interpretability, and safety mechanisms when dealing with unseen scenarios, limiting their application in scenarios where safety is critical.

[0004] In recent years, researchers have begun to explore the combination of the learning ability of NN and the constraint optimization solving ability of MPC. Patent CN 119396008 A uses a double neural network model predictive control method to predict the state of a robot flexible joint. This patent uses a gated recurrent unit neural network for dynamic modeling, achieving high prediction accuracy in low-dimensional state systems of robots, but for high-dimensional state spaces and complex interactive vehicle systems, a single recurrent neural network structure has significant limitations in modeling ability and generalization performance. In addition, this patent uses a nonlinear autoregressive neural network with external input to directly replace the MPC controller, which simplifies the online optimization process but also lacks the constraint handling ability of MPC. Patent CN 117784610 A uses a long short-term memory network to predict future states, and on this basis, solves a quadratic programming problem. Although NN is used, the solution of the quadratic programming problem still relies on the computationally complex traditional quadratic programming solver, making it difficult to guarantee high real-time performance. SUMMARY

[0005] In view of the deficiencies of the prior art, in order to solve the problems in the background art, the present application proposes an intelligent vehicle path tracking control method based on double network iterative optimization. This method combines the advantages of MPC in constraint optimization solution and the rapid computing ability of neural network, aiming to balance the constraint optimization solution while solving the deficiencies of traditional MPC in real-time performance and computing complexity to adapt to the complex and variable environment in the intelligent driving system. Through the method of the present application, safe and efficient driving of intelligent driving vehicles in complex scenarios can be realized, and the generalization and robustness of the system are enhanced.

[0006] The technical solutions adopted by the present application to solve the technical problems are as follows:

[0007] The present application is an intelligent vehicle path tracking control method based on double network iterative optimization, which includes a subspace encoder network module and a control input increment sequence generation network module. In specific implementation, through the two nested loop update processes of predicted output sequence update and control input sequence update, the control input sequence is iteratively optimized to obtain the predicted control input, and the path tracking control of the vehicle is realized.

[0008] The method includes the following steps:

[0009] Step 1, predicted output sequence update:

[0010] Step 1.1, design of subspace encoder network module:

[0011] The subspace encoder network module is composed of an encoder module, a state transition module and an output mapping module.

[0012] The one-dimensional convolutional neural network used by the encoder module contains three one-dimensional convolutional layers, where the channel numbers of the first and second convolutional layers are 64 and 128 respectively, and the corresponding convolution kernel sizes are 3 and 5 respectively. Batch normalization processing is performed after each convolution operation, and the ReLU function is used as the activation function to set the negative number output to zero, so that only non-negative numbers are retained in the convolution result. The channel number of the third convolutional layer is 128, and the convolution kernel size is 7. The dilated convolution technology is introduced to expand the coverage of the convolution kernel without increasing the size of the convolution kernel. After the three convolutional layers, a fully connected layer is connected to map the high-dimensional features to low-dimensional latent states. The input feature dimension of the fully connected layer is 128, and the output feature dimension is set to 64.

[0013] The encoder module uses a one-dimensional convolutional neural network to extract high-dimensional features from the current time control input , vehicle state as input variables. And a fully connected layer is connected after the one-dimensional convolutional neural network to output low-dimensional latent state The specific conversion relationship of the encoder module is as follows:

[0014] (1)

[0015] in, This represents the low-dimensional potential state at the current moment. Indicates the current control input. Indicates the current vehicle status. , This represents the index of the prediction step performed during the update of the predicted output sequence. Indicates the length of the prediction time domain. This indicates the operation of the encoder module.

[0016] The state transition module uses The encoder network consists of three stacked sub-modules with identical structures. Each sub-module contains four basic structural elements: a multi-head self-attention mechanism, a multilayer perceptron, a ReLU activation function, and a random dropout layer. The optimal hyperparameter combination is selected using the Optuna hyperparameter optimization framework. The design uses 8 attention heads, a learning rate of 0.001, a batch size of 64, and a dropout rate of 0.1 for the random dropout layer. The state transition module handles low-dimensional latent states... and Predictive control input at time Processing and prediction The potential state at time t, output Combined data at different times The specific transition relationships of the state transition module are as follows:

[0017] (2)

[0018] in, express Combined data at all times The encoder network representing the state transition module. This represents the low-dimensional latent state characteristics at the current moment. express Predictive control input at any given time.

[0019] The output mapping module uses a fully connected neural network consisting of three linear transformation layers. The first layer has a feature dimension of 128 for both input and output, followed by a ReLU activation function and a random loss layer with a dropout rate of 0.1. The second layer has a feature dimension of 64 for output, also followed by a ReLU activation function and a random loss layer with a dropout rate of 0.1. The third layer maps the features to... The state is predicted at any given time, where the feature dimension of the predicted state is 6.

[0020] The output mapping module will Combined data of the moment Mapping to Moment prediction state The output mapping module converts the relationship as follows:

[0021] (3)

[0022] wherein, represents the prediction state of the moment, represents the full connection layer operation, represents the combined data of the moment.

[0023] The prediction state obtains a prediction output through a prediction output equation, and the specific formula of the prediction output equation is as follows:

[0024] (4)

[0025] wherein, is the prediction output of the moment, is a 6-order unit matrix. represents the prediction state of the moment.

[0026] In order to realize the effective prediction function, the subspace encoder network module needs to use the processed data for network training in advance. The data used for network training is simulated by Casim and Simulink software, and is collected in combination with the Logitech steering wheel. The mean square error is used to design the loss function , and the network model parameter training update is performed by minimizing the error between the prediction state and the actual state. The specific formula is as follows:

[0027] (5)

[0028] wherein, is the loss function of the subspace encoder network module, is the actual state, is the prediction state, is the sample number.

[0029] Step 1.2, prediction output sequence update:

[0030] Before performing the loop calculation, a group of step control input sequences are randomly initialized in the constraint range:

[0031] (6)

[0032] wherein, For the current loop The next control input sequence This represents the iteration round index for controlling the update of the input sequence. After the input sequence is initialized, it is set... =0. for Predictive control inputs at all times. This represents the index of the prediction step performed during the update of the predicted output sequence. .

[0033] Extract sequentially from the control input sequence generated during initialization. Predictive control input at time And collect the vehicle status at the current moment. and the control input at the current moment This serves as the input to the subspace encoder network module, yielding the single-step prediction output at the corresponding prediction time. Repeat execution Next, the final subspace encoder network module output Step-by-step prediction output sequence .

[0034] Step 2: Control the input sequence update process:

[0035] The control input sequence update process includes prediction error calculation, control input increment sequence update determination, and control input sequence update, ultimately obtaining an iteratively optimized control input sequence.

[0036] Step 2.1, Prediction Error Calculation:

[0037] Prediction error calculation will The predicted output sequence of the step and the known The expected output sequence is subtracted to obtain the result. The lateral position error sequence of the step:

[0038] (7)

[0039] in, This represents the lateral position error sequence. Indicates in Lateral position error at time, Indicates in Expected output at any given time Indicates in Predicted output at time step This represents the index of the prediction step performed during the update of the predicted output sequence. .

[0040] Step 2.2, Control input increment sequence update determination:

[0041] Determine the lateral position error sequence Is it less than a preset threshold? If it is less than the preset threshold, then the corresponding... The first control input in the control input sequence of the step As the actual input to the vehicle system, it enables the intelligent driving vehicle to perform path tracking tasks; if the error exceeds a preset threshold, the lateral position error sequence is input to the control input increment sequence generation network module to generate... Step control input increment sequence The specific transformation relationships of the network module that controls the input increment sequence generation are as follows:

[0042] (8)

[0043] in, express Incremental control input at any time This indicates the network module that controls the input increment sequence generation. network, express Lateral position error at time, This represents the index of the prediction step performed during the update of the predicted output sequence. .

[0044] The control input increment sequence generation network module needs to be trained beforehand using processed data. The input data is a lateral position error sequence, and the target output is the control input increment sequence. Mean squared error is used as the loss function. By minimizing the deviation between the predicted and actual control input increments, the parameters of the control input update network model are trained and updated, ensuring that the control input increment sequence generation network module can output an effective control input increment sequence. The specific formula is as follows:

[0045] (9)

[0046] in, To control the loss function of the network module that generates the input increment sequence, For the actual control input increment, For the predicted control input increment, This represents the number of samples.

[0047] Step 2.3: Update the control input sequence:

[0048] The updated control input sequence is obtained by adding the control input increment sequence to the current control input sequence. The specific iterative process is as follows:

[0049] (10)

[0050] wherein, is the updated control input sequence, is the control input increment sequence, is the current control input sequence.

[0051] After the control input sequence is updated, the updated control input sequence is re-input into the subspace encoder network module, and the process of updating the predicted output sequence is performed again, and the prediction error calculation is performed, until the error value in the lateral position error sequence is less than the preset threshold value. At this time, the first control input in the control input sequence that meets the preset threshold value is taken as the actual input of the vehicle system, and the path tracking task of the intelligent driving vehicle is realized.

[0052] The beneficial effects of the present application are: the present application relates to an intelligent vehicle path tracking control method based on double network iterative optimization, by combining neural network and MPC, the present application can significantly improve the real-time performance and calculation efficiency of the system while maintaining the advantages of MPC method. Through the joint use of the subspace encoder network module and the control input increment sequence generation network module, efficient control of the intelligent driving vehicle under the path tracking task is realized. The method of the present application reduces the dependence on QP solving, improves the calculation efficiency, and enhances the stability and robustness of the system in complex and variable environment. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is the experimental flowchart of the intelligent vehicle path tracking control method based on double network iterative optimization of the present application.

[0054] Figure 2 is a schematic diagram of the subspace encoder network module of the present application.

[0055] Figure 3 is a schematic diagram of the control input increment sequence generation network module of the present application. DETAILED DESCRIPTION

[0056] The present application will be described in detail below with reference to the accompanying drawings.

[0057] The present application proposes an intelligent vehicle path tracking control method based on double network iterative optimization, referring to Figure 1 the schematic diagram, the method includes a subspace encoder network module and a control input increment sequence generation network module. In specific implementation, the process of updating the predicted output sequence is generated by the subspace encoder network module, and the predicted output sequence of the future step is recursively generated; the process of updating the control input sequence is based on the threshold value judgment of the lateral position error sequence , if the lateral position error sequence If the preset threshold is not met, the control input sequence is iteratively optimized by a control input increment sequence generation network module. The two nested loop update processes of the predicted output sequence update and the control input sequence update are re-executed until the lateral position error sequence meets the preset threshold. The first item in the final output control input sequence is taken as the actual input of the vehicle system to realize the path tracking task of the intelligent driving vehicle. Specifically, the following steps are included:

[0058] Step 1, predicted output sequence update:

[0059] Step 1.1, design of subspace encoder network module:

[0060] The subspace encoder network module refers to the schematic of Figure 2 , including an encoder module, a state transition module and an output mapping module.

[0061] The encoder module adopts a one-dimensional convolutional neural network, which includes three one-dimensional convolutional layers. The first convolutional layer sequentially completes the following processes: first, a convolution operation is performed using 64 convolution kernels with a size of 3, considering 3 consecutive time steps of data each time; second, batch normalization is used to standardize each feature dimension of the convolution output; third, a ReLU activation function is used to introduce a nonlinear transformation to avoid neuron inactivation; and finally, maximum pooling processing is performed to extract features and reduce computational load. After processing by the first one-dimensional convolutional layer, each sample has 64 feature dimensions. The second convolutional layer uses 64 convolution kernels with a size of 5, and similarly performs convolution operation, batch normalization, application of activation function and maximum pooling processing. After processing by the second layer, the feature dimension of each sample is still 64. The third convolutional layer uses 128 convolution kernels with a size of 7, and uses dilated convolution technology to expand the coverage range without increasing the size of the convolution kernel. The dilated convolution formula is as follows:

[0062] (11)

[0063] wherein, represents the effective receptive field range, represents the convolution kernel size, represents the dilation coefficient. In the present application, the dilation coefficient is set to 3, and the effective coverage range is expanded to 19 time steps.

[0064] In order to retain complete timing information, the third convolutional layer does not need to use maximum pooling processing, and the resulting feature dimension is 128. A fully connected layer is connected after the third convolutional layer, and the input feature dimension of the fully connected layer is 128, and the output feature dimension is set to 64.

[0065] The encoder module extracts the current time control input through input , vehicle state , performs convolution operation to extract high-dimensional features, connects a fully connected layer to map the high-dimensional features to a low-dimensional latent space, and outputs a low-dimensional latent state The specific conversion relationship is as described in formula (1) in the foregoing summary.

[0066] The state transition module uses The encoder network is composed of three sub-modules with the same structure stacked. Each sub-module includes four basic structures: multi-head self-attention mechanism, multi-layer perceptron, RELU activation function, and random loss layer. The optimal combination of hyperparameters is selected through the hyperparameter optimization framework Optuna: the number of attention heads is designed to be 8, the dimension of each layer of feedforward network is 4 times the hidden dimension; the learning rate is 0.001; the batch size is 64; the dropout rate of the random loss layer is set to 0.1, which is used to regularize the neural network, prevent overfitting, and improve the generalization ability of the model. The state transition module combines the low-dimensional latent state at the current time and the predicted control input at the time , and outputs the combined data at the time The specific conversion relationship is as described in formula (2) in the foregoing summary.

[0067] The output mapping module uses a fully connected neural network composed of three linear transformation layers, where the feature dimensions of the input and output of the first layer are both 128, followed by a ReLU activation function and a random loss layer with a dropout rate of 0.1; the feature dimension of the output of the second layer is 64, followed by a ReLU activation function and a random loss layer with a dropout rate of 0.1; the third layer maps the features to predicted state at the time , and the feature dimension of the predicted state is 6. The output mapping module uses a fully connected neural network, the input is the combined data at the time , the output is the predicted state at the time The specific conversion relationship is as described in formula (3) in the foregoing summary.

[0068] The predicted state is obtained by the predicted output equation, and the predicted output equation uses formula (4) in the foregoing summary.

[0069] In order to achieve effective prediction function, the subspace encoder network module needs to be trained in advance using processed data. The data used for network training is collected through joint simulation of Casim and Simulink software combined with Logitech steering wheel. The input data includes the current vehicle state: lateral speed , yaw rate , yaw angle , lateral position , longitudinal position , and center of mass side slip angle , and the current control input: front wheel steering angle . The target output is the next time vehicle state: lateral speed , yaw rate , yaw angle , lateral position , longitudinal position , and center of mass side slip angle . The subspace encoder network module adopts mean square error design loss function , by minimizing the error between the predicted state and the actual state, training and updating the control input update network model parameters, to ensure that the subspace encoder network module can accurately predict the state of the vehicle, so as to obtain effective prediction output sequence. The loss function is specifically designed as described in the formula (5) in the foregoing summary.

[0070] Step 1.2, prediction output sequence update process:

[0071] Before performing the loop calculation, first, as described in the formula (6) in the foregoing summary, a set of steps of control input sequences are randomly initialized within the constraint range.

[0072] After completing the initialization of the control input sequence, the prediction control input at each time is extracted from the initialized control input sequence in turn , and the current vehicle state and the current control input at the current time are collected as the input of the subspace encoder network module, to obtain the single-step prediction output corresponding to the prediction time, repeat times, and finally the subspace encoder network module outputs the prediction output sequence of steps . The specific recursive process is as follows:

[0073] (12)

[0074] wherein, representing the mapping relationship between the sub-space encoder and the sub-space decoder, is the prediction output at the time step, represents the prediction control input at the time step, represents the vehicle state at the current time step, represents the control input at the current time step.

[0075] Step 2, control input sequence updating process:

[0076] The control input sequence updating process includes prediction error calculation, control input increment sequence updating determination, and updating the control input sequence, and finally obtains the iteratively optimized control input sequence.

[0077] Step 2.1, prediction error calculation:

[0078] The prediction error calculation is to subtract the prediction output sequence of the time step from the expected output sequence of the time step known as formula (7) in the aforementioned summary, and obtain the lateral position error sequence of the time step. Step 2.2, control input increment sequence updating determination: Determine whether the lateral position error sequence is less than the preset threshold value. If it is less than the preset threshold value, the first control input in the control input sequence of the time step will be taken as the actual input of the vehicle system, and the path tracking task of the intelligent driving vehicle will be realized; if the error exceeds the set preset threshold value, the lateral position error sequence is input to the control input increment sequence generation network module to generate the control input increment sequence of the time step.

[0079]

[0080] The network of the control input increment sequence generation network module includes a feature extraction layer and a prediction layer. The feature extraction layer extracts features in the lateral position error sequence through multi-head attention mechanism-a and feedforward neural network-a, and RELU activation function-a enables the network to fit a nonlinear function, and then loss layer-a prevents overfitting. The prediction layer extracts features to generate the control input increment sequence through multi-head attention mechanism-b and feedforward neural network-b. RELU activation function-b enables the network to fit a nonlinear function, and then loss layer-b prevents overfitting. The best hyperparameter combination is selected through the hyperparameter optimization framework Optuna, wherein

[0081] Referring to the schematic of Figure 3 , the network of the control input increment sequence generation network module includes a feature extraction layer and a prediction layer. The feature extraction layer extracts features in the lateral position error sequence through multi-head attention mechanism-a and feedforward neural network-a, and RELU activation function-a enables the network to fit a nonlinear function, and then loss layer-a prevents overfitting. The prediction layer extracts features to generate the control input increment sequence through multi-head attention mechanism-b and feedforward neural network-b. RELU activation function-b enables the network to fit a nonlinear function, and then loss layer-b prevents overfitting. The best hyperparameter combination is selected through the hyperparameter optimization framework Optuna, wherein ​​​​​​​​The network layer is 4, the number of neurons in each layer is 128, the number of attention heads is 8, the learning rate is 0.0001, and the batch size is 64. The specific conversion relationship of the control input increment sequence generation network module is as described in formula (8) in the summary of the invention.

[0082] The control input increment sequence generation network module needs to be trained in advance using the processed data. The input data is the lateral position error sequence, and the target output is the control input increment sequence. The mean square error is used as the loss function , by minimizing the deviation between the predicted control input increment and the actual control input increment, the control input update network model parameters are trained and updated to ensure that the control input increment sequence generation network module can output effective control input increment sequences. The loss function The specific design is as described in formula (9) in the summary of the invention.

[0083] Step 2.3, update the control input sequence:

[0084] Add the control input increment sequence to the current control input sequence to update the new control input sequence. The specific iteration process is described in formula (10) in the summary of the invention.

[0085] After updating the control input sequence, the updated control input sequence is input into the subspace encoder network module again, and the process of updating the predicted output sequence is performed again, and the prediction error is calculated until the error value in the lateral position error sequence is less than the preset threshold. At this time, the first control input in the control input sequence that meets the preset threshold is taken as the actual input of the vehicle system, realizing the path tracking task of the intelligent driving vehicle. From the second control period, the control input sequence initialization is no longer random initialization, but the control input sequence obtained by iteration and optimization in the last control period is used as the initial value.

[0086] In summary: the intelligent vehicle path tracking control method based on double network iteration optimization is proposed. The method includes a subspace encoder network module and a control input increment sequence generation network module, and the specific implementation includes two nested loop processes of updating the predicted output sequence and updating the control input sequence. The predicted output sequence of the future step is recursively generated by updating the predicted output sequence, and the lateral position error sequence is obtained by updating the control input sequence, and the control input sequence is iteratively optimized. The two nested update processes of updating the predicted output sequence and updating the control input sequence are repeated until the lateral position error sequence meets the preset threshold. The first item in the final output control input sequence As the actual input of vehicle system, the path tracking control of vehicle is realized. This method ensures real-time and computational efficiency, and is suitable for complex driving environment.

Claims

1. An intelligent vehicle path tracking control method based on double network iterative optimization, characterized in that, The method comprises a subspace encoder network module and a control input delta sequence generation network module; the subspace encoder network module is used for recursively generating a predicted output sequence To realize prediction of a future trajectory of a vehicle The control input increment sequence generation network module is used to generate the control input increment sequence. To achieve control input sequence The method involves iterative updates; its implementation includes two nested loop update processes: one for updating the predicted output sequence and the other for updating the control input sequence. During the predicted output sequence update, a sequence of length [missing information] is constructed. A recursive loop is used to recursively generate future... The predicted output sequence of the step; when updating the control input sequence, the lateral position error sequence is obtained by calculating the prediction error. If the lateral position error sequence If the preset threshold is met, the first item in the control input sequence will be output. If the lateral position error sequence If the preset threshold is not met, the control input incremental sequence generation network module generates the input based on the lateral position error sequence. Generate control input increment sequence The control input increment sequence With control input sequence Add them together to get the new control input sequence. And re-execute the two nested loop update processes of predictive output sequence update and control input sequence update until the lateral position error sequence is updated. If the preset threshold is met, the final output controls the first item in the input sequence. It is used to implement path tracking control for vehicles.

2. The intelligent vehicle path tracking control method based on dual-network iterative optimization according to claim 1, characterized in that, The sub-space encoder network module comprises an encoder module, a state transition module and an output mapping module; each execution of the sub-space encoder network module generates a single-step prediction output corresponding to a prediction time ; through the prediction output sequence updating process, a recursive loop with a length of is constructed, and a prediction output sequence of future steps is recursively generated, which is used to calculate the lateral position error sequence ; The encoder module uses a one-dimensional convolutional neural network to control the input at the current time , vehicle state As an input variable, high-dimensional feature extraction is performed; a one-dimensional convolutional neural network is connected to a fully connected layer to output a low-dimensional latent state The specific conversion relationship of the encoder module is as follows: wherein, represents a low-dimensional latent state at a current time instant, represents a control input at a current time instant, , represents a prediction step index performed in a prediction output sequence update process, represents a prediction time domain length, represents a vehicle state at a current time instant, represents an encoder module operation; The state transition module uses The encoder network combines the current time low-dimensional latent state extracted by the encoder module With The time prediction control input As input, the prediction The time latent state, output The combined data of the time The state transition module specifically converts the following relationship: wherein, denotes the combined data at the time instant, denotes an encoder network of the state transition module, denotes the latent state feature at the current time instant, denotes the predicted control input at the time instant; The output mapping module uses a fully connected neural network, and the input is combined data at the moment , which is mapped to predicted state at the moment The specific conversion relationship of the output mapping module is as follows: wherein, representing predicting a state at a time instant, representing a fully connected network, denotes combined data at a time instant; Predicted state Predicted output via the predicted output equation The specific formula is as follows: wherein is moment prediction output, is a 6x6 identity matrix, represents moment prediction state. 3.The intelligent vehicle path tracking control method based on dual-network iterative optimization of claim 1, wherein, The control input increment sequence generating network module uses Network structure, according to the lateral position error sequence Obtaining a control input increment sequence The specific conversion relationship is as follows: wherein, represents a control input increment sequence, represents an iteration round index for control input sequence update, represents a control input increment sequence generation network module, network, represents a lateral position error sequence; The control input increment sequence generated by the control input increment sequence generation network module is added to the control input sequence to update a new control input sequence , re-entering the prediction output sequence update process to form a closed-loop iteration; finally, the first item in the control input sequence that meets the preset threshold is taken as the actual control output of the system to realize path tracking control of the vehicle, and the specific iteration process is as follows: wherein is the updated next cycle updated next cycle the control input sequence for the next cycle denotes the control input increment sequence is the current cycle the control input sequence for the next cycle.

Images