Drive-by-wire automobile end-to-end control method based on double-circulation neural network

Abstract

The invention discloses a drive-by-wire automobile end-to-end control method based on a double-circulation neural network, and the method comprises the steps: S1, obtaining driving track data, extracting key feature parameters, and generating a driving feature set; s2, loading an end-to-end deep learning model of the outer-layer recurrent neural network, and predicting transverse and longitudinal displacement characteristics during automobile driving; s3, loading an end-to-end deep learning model of the inner-layer recurrent neural network, and outputting expected path data at the t + 1 moment; s4, according to the expected path data at the t + 1 moment and the controller, the transverse deviation and the longitudinal deviation of the vehicle are controlled respectively; s5, trajectory tracking effect evaluation is executed, and whether trajectory tracking deviation meets the control requirement or not is judged; taking the driving track data which does not meet the control requirement as input, and executing the step S2 again; and if the control requirement is met, iteratively executing the step S1. According to the technical scheme, the transverse and longitudinal path tracking capability and tracking stability of the unmanned driving system can be improved.

Description

Technical Field

[0001] The present invention relates to the field of unmanned driving technology for automobiles, and in particular to an end-to-end control method for a controlled-by-wire automobile based on a double-circular neural network. Background Art

[0002] With the rapid development of intelligent vehicles and autonomous driving technology, autonomous vehicles are increasingly being used in complex road environments. Current autonomous driving systems still face significant challenges in dynamic driving tasks such as maneuvering lane changes. Driving involves coordinated control of the vehicle's lateral and longitudinal motion, particularly in key areas such as road adhesion changes, dynamic speed adjustment, and precise tracking of front wheel angles. There is an urgent need to improve the system's lateral and longitudinal stability. Traditional autonomous driving architectures typically separate perception, decision-making, and control modules, relying on manual rules to divide subtasks such as road recognition, lane detection, path planning, and control execution. This results in overall system performance being limited by the coupling between modules and inefficient data transfer. Although end-to-end deep learning methods simplify intermediate processing steps and enable a direct mapping from input data (e.g., images) to output data (e.g., steering wheel angle), they still have significant shortcomings in executing coordinated system control and adapting to complex scenarios.

[0003] Existing end-to-end learning methods primarily use environmental perception data as input, extracting road features and generating control commands through convolutional neural networks. These methods are highly dependent on the perception system and fail to fully consider the vehicle's dynamic characteristics and the real-time responsiveness of the actuation system. This can lead to problems such as excessive lateral deviation, increased skidding risk, and low driving efficiency under complex road conditions. Furthermore, vehicle actuation systems (such as steer-by-wire systems) typically rely on independent closed-loop controllers such as PID control, sliding mode control, linear quadratic optimal control, and model predictive control. These control logics lack deep integration with the data interaction mechanisms of the upper-level autonomous driving system, limiting vehicle performance. Limited research exists on the coordinated control of end-to-end deep learning and underlying actuation systems, resulting in a lack of effective solutions for improving vehicle lateral tracking capability and driving stability. Existing research suggests that end-to-end deep learning has potential within an integrated perception-decision-making-control framework, but its dynamic control capabilities and data information mining for vehicle actuation systems remain limited.

[0004] To overcome the above technical bottlenecks, an innovative method that can integrate the advantages of deep learning and traditional control theory is urgently needed to enable autonomous vehicles to efficiently perceive, make accurate decisions and achieve stable control of complex driving scenarios. Summary of the Invention

[0005] To achieve the above objectives, the present application provides an end-to-end control method for a drive-by-wire vehicle based on a double recurrent neural network, comprising:

[0006] Step S1: Acquire driving trajectory data, extract key feature parameters from the driving trajectory data, and generate a driving feature set;

[0007] Step S2: loading an end-to-end deep learning model of an outer recurrent neural network, inputting the driving feature set to predict the lateral and longitudinal displacement features of the vehicle during driving, wherein the lateral and longitudinal displacement features are used to guide the steering angle and vehicle speed of the vehicle;

[0008] Step S3: Loading the end-to-end deep learning model of the inner layer recurrent neural network, performing the path tracking task based on the time series and the lateral and longitudinal displacement features, predicting the trajectory tracking deviation in the lateral and longitudinal motion decoupling state, outputting the expected path data at time t+1, and constructing the inner layer model dataset;

[0009] Step S4: controlling the lateral and longitudinal deviations of the vehicle respectively according to the desired path data at time t+1 and the controller for controlling the lateral motion and longitudinal motion errors;

[0010] Step S5: Execute trajectory tracking effect evaluation to determine whether the trajectory tracking deviation meets the control requirements; use the driving trajectory data that does not meet the control requirements as input and re-execute step S2; if it meets the control requirements, iteratively execute step S1.

[0011] Among them, the operation of extracting key feature parameters from the driving trajectory data includes: normalization, calculation of feature contribution values, and construction of a driving feature set based on key features; the data of the feature set includes: current moment, current moment reference position, road friction coefficient, vehicle longitudinal speed, vehicle lateral speed, vehicle lateral acceleration and yaw angular velocity; the current moment reference position is decomposed into the longitudinal coordinate and lateral coordinate of the current moment reference path based on the Frenet coordinate system.

[0012] Furthermore, after loading the end-to-end deep learning model of the outer recurrent neural network, input parameters are extracted from the driving feature set, and then input into the end-to-end deep learning model of the outer recurrent neural network to obtain output parameters;

[0013] Input parameters include: the longitudinal coordinate x of the reference path at the current moment refer , the horizontal coordinate y of the reference path at the current moment refer , road friction coefficient μ, vehicle longitudinal speed v x , vehicle lateral speed v y , vehicle lateral acceleration a y and yaw angular velocity ω; the input parameters constitute the input vector expressed as:

[0014] The output parameters include the longitudinal coordinate of the path at time t The horizontal coordinate of the path at time t The output parameters constitute the output vector represented as:

[0015] Among them, the inner model dataset expresses the expected path data under time series;

[0016] The inner model dataset is constructed by taking the four-step forward method to construct the input parameters of the inner recurrent neural network, predicting and outputting the expected path data at time t+1, and continuously and recursively constructing the inner model dataset.

[0017] Furthermore, the path tracking task includes three subtasks: path planning, path tracking, and deviation correction;

[0018] Among them, the expected path data at time t+1 is predicted to realize path planning; the distance deviation and heading deviation between the actual trajectory and the reference path are calculated to realize path tracking; if the distance deviation and heading deviation do not meet the expected path requirements, the longitudinal coordinates and transverse coordinates of the path at time t+1 are constructed and input into the inner recurrent neural network, and the features of the next moment are recalculated to correct the deviation.

[0019] Furthermore, the controller definition includes the following steps:

[0020] Construct a vehicle dynamics model in a Cartesian coordinate system;

[0021] The positional relationship expressed by the time-series path data of the vehicle in the Cartesian coordinate system is converted into the positional relationship in the Frenet coordinate system, including the arc length s traveled by the vehicle and the shortest distance d between the deviation of the actual trajectory and the reference path. The lateral and longitudinal motions of the vehicle's trajectory are decoupled, and the time derivatives of the arc length s and the shortest distance d are calculated respectively, which can be expressed as:

[0022]

[0023] Among them, θ x and θ r They represent the yaw angle of the vehicle on the actual tracking path and the yaw angle of the reference path respectively, κ is the curvature value on the path point, is the vehicle yaw angle, V is the vehicle's total velocity, and β is the center of mass sideslip angle;

[0024] Calculate the distance deviation e between the actual trajectory and the reference path d The heading deviation between the actual trajectory at the current moment and the reference path

[0025] A lateral motion controller is established for lateral motion in a feedback form, and a longitudinal motion controller is established for longitudinal motion in a feedforward form.

[0026] To establish a lateral motion controller in the form of feedback for lateral motion means:

[0027] The state feedback control law for determining the front wheel steering angle is: δ = -KX, where K = [τ1, τ2, τ3, τ4];

[0028] Among them, δ is the equivalent front wheel angle, X is the state variable τ1, τ2, τ3, and τ4 are user-defined constants.

[0029] Methods for determining the state variable X include:

[0030] Get and its first-order and second-order derivatives, respectively, are expressed as:

[0031] Get e d Its first-order and second-order derivatives are expressed as:

[0032] Select state variables

[0033] Establishing a longitudinal motion controller in a feedforward manner for longitudinal motion means:

[0034] Add proportional-integral correction link in feedforward mode, through the longitudinal force F xf Correct the vehicle's longitudinal v x The error is expressed as:

[0035]

[0036] Among them, K p is the proportional gain coefficient; K i is the integral gain coefficient; and is the desired stable value of the system.

[0037] Furthermore, the control requirements include a lateral motion displacement deviation and a longitudinal speed deviation; the lateral motion displacement deviation is used to ensure that the tracking path does not deviate, and the longitudinal speed deviation is used to control speed stability.

[0038] According to the present invention, an end-to-end deep learning architecture with both self-learning capabilities and execution system collaborative control capabilities can be constructed to improve the lateral and longitudinal path tracking capabilities and tracking stability of unmanned driving systems. The beneficial effects of the present invention include the following:

[0039] (1) Efficient data preprocessing and key feature extraction: Using methods such as normalization and calculation of feature contribution values ​​to preprocess the input data and filter out key feature parameters helps to build a more accurate data set and provide high-quality data support for subsequent model training.

[0040] (2) Improving path tracking capability and stability: By establishing a learning architecture of a double-layer gated recurrent neural network and linear quadratic optimal control, the present invention can deeply learn the intrinsic characteristic relationship between input parameters and output parameters, thereby significantly improving the lateral and longitudinal path tracking capability and stability of the unmanned vehicle.

[0041] (3) Hierarchical deep learning network design: The inner recurrent neural network focuses on learning the driving lateral and longitudinal displacement characteristics, while the outer recurrent network focuses on trajectory tracking deviation prediction. This hierarchical structure not only improves the prediction accuracy of the model, but also enhances the robustness and adaptability of the system.

[0042] (4) Dynamic adjustment and feedback mechanism: By establishing a trajectory tracking effect evaluation module and a unified model prediction and feedback correction mechanism, the present invention can monitor and adjust the vehicle driving status in real time to ensure that it always remains within the optimal control range, greatly improving the safety and reliability of driving.

[0043] (5) Enhanced ability to cope with complex scenarios: The present invention is specially designed for complex driving scenarios. It uses an end-to-end deep learning framework to estimate the parameters of vehicle driving, and achieves an optimal match between the passing speed, road conditions and tracking deviation, so that the unmanned driving system can obtain the best control effect when handling a wider range of application scenarios.

[0044] (6) Modular design facilitates expansion and maintenance: Dividing the steer-by-wire and drive-by-wire systems of autonomous vehicles into two independent but interrelated deep learning networks not only simplifies the system architecture but also facilitates subsequent maintenance and upgrades. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 2. It is a schematic flow chart of an end-to-end control method for a wire-controlled vehicle according to an embodiment of the present invention;

[0046] Figure 2 1 is a diagram of a gated recurrent neural network structure based on time series according to an embodiment of the present invention;

[0047] Figure 3 2. Schematic diagram of lateral deviation of path tracking in an end-to-end control method for a controlled-by-wire vehicle according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to improve the lateral and longitudinal path tracking capabilities and tracking stability of unmanned driving systems and enhance the ability of unmanned vehicles to cope with complex driving scenarios, the present invention establishes a two-layer gated recurrent neural network (GRU) and linear quadratic optimal control (LQR) learning architecture to learn the relationship between driving characteristics and lateral and longitudinal displacement characteristics, and then construct a dynamic model of a steer-by-wire vehicle. A vehicle steering parameter estimation method based on an end-to-end deep learning framework is constructed to evaluate the vehicle's lateral and longitudinal tracking capabilities, achieve mutual matching between driving speed, road conditions and tracking deviation, ensure vehicle lateral stability while ensuring driving efficiency, and achieve optimal control effects when handling a wider range of application scenarios.

[0049] The specific implementation of the present invention is described in detail below with reference to the accompanying drawings.

[0050] The process of the end-to-end control method of the wire-controlled vehicle based on the double recurrent neural network proposed in this invention is as follows: Figure 1 As shown, the following steps are included:

[0051] Step S1: Acquire driving trajectory data, extract key feature parameters from the driving trajectory data, and generate a driving feature set;

[0052] like Figure 1 As shown in S1, in this step, driving trajectory data is input, its features are analyzed, key feature parameters are selected, and the driving trajectory data is resampled to maintain the same dimensionality after processing. The processed data can then be subjected to feature analysis to form a feature set. The structure and dimensionality of the feature set can be maintained consistent within a single neural network under different operating conditions of the present invention.

[0053] Specifically, the operation of extracting key feature parameters includes normalization, calculating feature contribution values, and constructing a driving feature set based on the key features. The data in the feature set includes at least: the current moment, the current moment reference position, the road friction coefficient, the vehicle longitudinal speed, the vehicle lateral speed, the vehicle lateral acceleration, and the yaw rate.

[0054] The current reference position is decomposed into the longitudinal coordinate and the transverse coordinate of the current reference path based on the Frenet coordinate system.

[0055] Step S2: loading the end-to-end deep learning model of the outer recurrent neural network, inputting the driving feature set to predict the lateral and longitudinal displacement characteristics of the car during driving;

[0056] The end-to-end deep learning model of the outer recurrent neural network includes an input layer, a hidden layer, a stacked layer, and an output layer. The hidden layer and the stacked layer are used to learn the mapping relationship between the driving characteristics and the lateral and longitudinal displacement characteristics in the dataset. The training process is evaluated by the loss function. The model is constructed after sufficient training and can predict the lateral and longitudinal displacement characteristics of the car during driving based on the driving characteristics to guide the vehicle's steering angle and vehicle speed.

[0057] The gated recurrent neural network structure is as follows Figure 2 As shown, the outer cyclic neural network adjusts the gradient by introducing the concept of gating mechanism and unit state, and adjusts the input data x (t) =(x1,x2,L,x t ,L,x N ) T The gated output method is adopted, and the expression is as follows:

[0058]

[0059] Where z t is the information to be retained after the update at time t; σ is the Sigmoid activation function; h t-1 is the input vector of the hidden layer at the previous moment; x t is the input value of the current time; W z and b z are the weight matrix and bias vector of the update gate respectively; r t is the information to be updated at time t; W r and b r are the weight matrix and bias vector of the reset gate respectively; (r t h t-1 ) represents the previous time step (h t-1 ) and reset gate (r t ) is combined to control the hidden state of its influence; tanh represents the tanh activation function; W and b represent the weight matrix and bias of the unit model respectively; (z t ·h t-1 ) is the retained historical information; is new information; h t is the output vector of the hidden layer at time t.

[0060] The loss function for training a gated recurrent neural network can be defined using the Huber Loss metric as follows:

[0061]

[0062] Where: ζ is a hyperparameter that determines the mean square error and mean absolute error of Huber Loss; is the input parameter The truth value of is the input parameter The predicted value of .

[0063] In the present invention, input parameters are extracted from the driving feature set and input into the end-to-end deep learning model of the outer recurrent neural network to obtain the lateral and longitudinal displacement features, that is, the input parameters corresponding to the input vector of the outer recurrent neural network include: the longitudinal coordinate x of the reference path at the current moment; refer , the horizontal coordinate y of the reference path at the current moment refer , road friction coefficient μ, vehicle longitudinal speed v x , vehicle lateral speed v y , vehicle lateral acceleration a y and yaw angular velocity ω; the output parameters corresponding to the output vector of the outer recurrent neural network, that is, the transverse and longitudinal displacement characteristics include: the longitudinal coordinates of the path at time t The horizontal coordinate of the path at time t The outer layer cyclic neural network data set is thus formed as follows: The input vector is: The corresponding output vector is: If the vehicle is in motion, Contains vehicle longitudinal speed information and time information.

[0064] Step S3: Load the end-to-end deep learning model of the inner layer recurrent neural network, perform the path tracking task based on the time series and the lateral and longitudinal displacement characteristics, predict the trajectory tracking deviation under the decoupled state of lateral and longitudinal motion, and output the expected path data at time t+1 to achieve the vehicle control path tracking task.

[0065] Among them, the end-to-end deep learning model of the inner recurrent neural network comprehensively matches the vehicle's actual steering ability with the steering force provided by the road surface, autonomously adjusts the vehicle speed to adapt to changes in the road friction coefficient, and realizes the learning and adaptive adjustment of the steering system's lateral tracking capability.

[0066] like Figure 1 As shown in the S3 part of the figure, the inner recurrent neural network is placed in the outer recurrent neural network framework, including the input layer, hidden layer, stacking layer and output layer. The training process is evaluated by the loss function. After the test result obtained by using the test set meets the expected result, the output end outputs the expected path data at time t+1, and evaluates the actual effect of path planning.

[0067] The inner cyclic neural network is based on the output vector generated by the outer neural network and the vehicle longitudinal speed v from step S1 xCalculation: First, the four-step forward method is used to construct the input parameters of the inner recurrent neural network, predict and output the expected path data at time t+1, and continuously recursively construct the inner model data set. The inner model data set can express the expected path data under time series (including the vertical coordinates of the path at each time series moment). The horizontal coordinate of the path at each time instant ) is expressed as: The input vector is: The corresponding output vector is:

[0068] Based on the predicted desired path data at time t+1 and the output parameters from step S2 at time t+1, the vehicle's path tracking task can be divided into three subtasks: path planning, path tracking, and deviation correction. The path planning subtask is implemented by predicting the desired path data at time t+1; the path tracking subtask is implemented by calculating the distance deviation and heading deviation between the actual trajectory and the reference path. If the distance deviation and heading deviation do not meet the desired path requirements, the longitudinal and transverse coordinates of the path at time t+1 are constructed and input into the inner recurrent neural network, and the features at the next time are recalculated to implement the deviation correction subtask.

[0069] When implementing the path planning subtask, the inner recurrent neural network constructs a time-series path dataset based on the output parameters generated by the outer neural network, constructs the input parameters by pushing forward four steps from the current time t data, and constructs the output parameters with the (t+1) time data, and continuously recursively constructs the inner model dataset. After training, the model is used to predict the vehicle's lateral and longitudinal position deviations corresponding to the next moment in the future; when implementing path tracking, the vehicle's lateral and longitudinal coordinate feature values ​​that meet the expected path are transmitted to the controller for the next path tracking task; when implementing deviation correction, the vehicle's lateral and longitudinal coordinate feature values ​​that do not meet the expected path are remapped to the next expected value in the inner recurrent network.

[0070] Step S4: Based on the desired path data at time t+1 and the controller for controlling the lateral motion and longitudinal motion errors, the lateral and longitudinal deviations of the vehicle are controlled respectively to achieve information exchange between the vehicle and the road;

[0071] Specifically, the controller supports correcting the horizontal and vertical coordinate values ​​in a state feedback manner. The definition of the definer includes the following steps:

[0072] 1) Construct a vehicle dynamics model in a Cartesian coordinate system, ignoring air resistance, which can be expressed as:

[0073]

[0074] Where: m is the vehicle mass, v x and v yare the longitudinal and lateral velocities of the vehicle, is the yaw acceleration, ω is the yaw velocity, δ is the equivalent front wheel turning angle, a is the distance from the vehicle center of mass to the rear axle, b is the distance from the vehicle center of mass to the rear axle, I z is the moment of inertia of the vehicle around the z axis, F xf and F yf Represent the longitudinal force and lateral force of the front wheel, F xr and F yr represent the longitudinal force and lateral force on the rear wheels respectively.

[0075] 2) The positional relationship expressed by the vehicle's time-series path data in the Cartesian coordinate system is converted to a positional relationship in the Frenet coordinate system, including the arc length s traveled by the vehicle and the shortest distance d between the actual trajectory and the reference path. The lateral and longitudinal motions of the vehicle's trajectory are decoupled, and the time derivatives of the arc length s and the shortest distance d are calculated, respectively, as follows:

[0076]

[0077] Among them, θ x and θ r They represent the yaw angle of the vehicle on the actual tracking path and the yaw angle of the reference path respectively, κ is the curvature value on the path point, is the vehicle yaw angle, V is the vehicle's total velocity, and β is the sideslip angle of the center of mass.

[0078] 3) Calculate the distance deviation e between the actual trajectory and the reference path d The heading deviation between the actual trajectory at the current moment and the reference path

[0079] Its first-order and second-order derivatives are expressed as:

[0080]

[0081] Similarly, e d Its first-order and second-order derivatives are expressed as:

[0082]

[0083] The controller defined in this step establishes a lateral motion controller for lateral motion in a feedback manner and a longitudinal motion controller for longitudinal motion in a feedforward manner.

[0084] 1) When establishing a lateral motion controller for lateral motion in the form of feedback, the state variables are selected according to the feedback control principle. The controller is defined based on the state feedback deviation. Combining Equations (3) to (6), the state space expression of the vehicle motion is as follows:

[0085]

[0086] Among them: u=δ, And k1 and k2 are the equivalent lateral deflections of the front and rear axle tires respectively, and Y is the output vector;

[0087] and:

[0088]

[0089] As a preferred method, the present invention adopts the LQR control method for the front wheel steering angle of the vehicle, and defines the state feedback control law of the front wheel steering angle as:

[0090] δ=-KX (8)

[0091] In the formula, K = [τ1, τ2, τ3, τ4],

[0092] Among them, δ is the equivalent front wheel angle, X is the state variable τ1, τ2, τ3, and τ4 are user-defined constants.

[0093] 3) When establishing a longitudinal motion controller for longitudinal motion in a feedforward manner:

[0094] First, the forward Euler method is used to discretize the equation in the continuous time domain, then

[0095] X t+1 =(I+T s A)X t +T s Bu t +T s Cξ t (9)

[0096] Where: I is the 4th order unit matrix, T s is the discretized system sampling time, X t is the state of the discrete system at time t, u t is the system input of the discrete system at time t, ξ t is the system deviation of the discrete system at time t;

[0097] make and Then formula (9) can be rewritten as:

[0098]

[0099] The control law of the front wheel steering angle can be expressed as:

[0100] δ * (τ)=-KX t (τ) (11)

[0101] Among them, δ is the equivalent front wheel angle, X is the state variable t is the time point, and τ is a custom constant.

[0102] In order to obtain the optimal K value, we first solve the Riccati equation to obtain the P matrix, and then find the K matrix, that is:

[0103]

[0104] Where: P matrix and Q matrix are pre-set weight matrices.

[0105] Next, find the K matrix:

[0106]

[0107] From this we can get the control input δ * (τ), the objective function value J of the discrete system achieves the optimal control effect, that is:

[0108]

[0109] Where: X t,K and u t,K X t and u t The discrete values ​​obtained by computing the matrix K.

[0110] When the front wheel angle is the above control law, the equation of the entire closed-loop system (10) is The influence of the system state variables cannot be guaranteed to converge to zero at the same time. Therefore, consider adding a feedforward term to the front wheel angle, and we have:

[0111] δ′=-KX+δ ff (15)

[0112] Where: δ ff is the feedforward term.

[0113] Assuming that the initial state is zero, perform Laplace transform on the closed-loop system of the vehicle and control, and use the final value theorem to calculate the stable value of the system state X. The state value x after Laplace transform of the system state X is ss for:

[0114]

[0115] The calculation method is converted into matrix form and expressed as:

[0116]

[0117] That is, the longitudinal speed control is converted into the combined effect of the acceleration system and the braking system. At this time, a separate control loop is added to adjust the longitudinal force F of the front wheel. xf , add proportional-integral correction link in feedforward mode to correct the longitudinal v of the vehicle x The error is expressed as:

[0118]

[0119] Where: K p is the proportional gain coefficient; K i is the integral gain coefficient; and is the desired stable value of the system.

[0120] Step S5: Execute trajectory tracking effect evaluation to determine whether the trajectory tracking deviation output in step S4 meets the control requirements; use the driving characteristics that do not meet the control requirements as input and re-execute step S2; if they meet the control requirements, iterate step S1.

[0121] The control requirements include items such as the upper limit of the lateral motion position deviation. For example, the upper limit of the lateral motion position deviation is set at 0.1 meters. The path of the control requirement is output to the lateral tracking controller to ensure that the lateral displacement deviation is controlled at the centimeter level; at the same time, the longitudinal motion controller ensures the stability of the path tracking speed. The two together achieve end-to-end lateral and longitudinal motion control goals that meet the preset expectations, ensuring the tracking stability of the car driving.

[0122] By defining the controller in this step, the input equivalent front wheel steering angle u and the vehicle longitudinal speed v under speed stability conditions are obtained. x After the value is obtained, the control logic of the vehicle is generated. Figure 3 This is the test result of applying the method of the present invention to the double lane-shifting working condition. Part (a) of the figure shows the comparison of lateral displacement. The path tracking performed by the method provided by the present invention is basically consistent with the reference path. Part (b) shows that the deviation of lateral displacement between the path tracking performed by the method provided by the present invention and the reference path is less than 0.1m.

[0123] The present invention, by establishing a learning architecture of a double-layer gated recurrent neural network and linear quadratic optimal control, can deeply explore the intrinsic characteristic relationship between driving characteristic parameters and lateral and longitudinal characteristic parameters, thereby significantly improving the lateral and longitudinal path tracking capabilities and stability of unmanned vehicles; for the learning architecture, a hierarchical deep learning network design is carried out: the inner recurrent neural network focuses on learning the lateral and longitudinal displacement characteristics of driving, while the outer recurrent network focuses on trajectory tracking deviation prediction. This hierarchical structure not only improves the prediction accuracy of the model, but also enhances the robustness and adaptability of the system. The present invention also proposes a dynamic adjustment strategy and feedback mechanism. Through trajectory tracking effect evaluation and feedback correction mechanism, the vehicle driving status is monitored and adjusted in real time to ensure that it is always maintained within the optimal control range, greatly improving the safety and reliability of driving.

[0124] The present invention is specially designed for complex driving scenarios. It uses an end-to-end deep learning framework to estimate vehicle driving parameters, achieving an optimal match between vehicle speed, road conditions and tracking deviation, so that the unmanned driving system can obtain the best control effect when handling a wider range of application scenarios and enhance its ability to cope with complex scenarios. In terms of control, the wire-controlled steering and wire-controlled drive systems of the unmanned vehicle are divided into two independent but interrelated deep learning networks, which not only simplifies the system architecture, but also facilitates subsequent maintenance and upgrades, and is more convenient for system expansion and maintenance.

[0125] The above disclosures are only a few specific embodiments of the present invention. However, the present invention is not limited thereto. Any changes that can be conceived by those skilled in the art should fall within the scope of protection of the present invention.

Claims

1. A method for end-to-end control of a controlled-by-wire vehicle based on a double recurrent neural network, characterized in that: include: Step S1: Acquire driving trajectory data, extract key feature parameters from the driving trajectory data, and generate a driving feature set; Step S2: loading an end-to-end deep learning model of an outer recurrent neural network, inputting the driving feature set to predict the lateral and longitudinal displacement features of the vehicle during driving, wherein the lateral and longitudinal displacement features are used to guide the steering angle and vehicle speed of the vehicle; Step S3: Loading the end-to-end deep learning model of the inner layer recurrent neural network, performing the path tracking task based on the time series and the lateral and longitudinal displacement features, predicting the trajectory tracking deviation in the lateral and longitudinal motion decoupling state, outputting the expected path data at time t+1, and constructing the inner layer model dataset; Step S4: controlling the lateral and longitudinal deviations of the vehicle respectively according to the desired path data at time t+1 and the controller for controlling the lateral motion and longitudinal motion errors; Step S5: Execute trajectory tracking effect evaluation to determine whether the trajectory tracking deviation meets the control requirements; use the driving trajectory data that does not meet the control requirements as input and re-execute step S2; if it meets the control requirements, iteratively execute step S1.

2. The end-to-end control method for a controlled-by-wire vehicle according to claim 1, characterized in that: The operation of extracting key feature parameters from the driving trajectory data includes: normalization, calculation of feature contribution values, and construction of a driving feature set based on key features; the data of the feature set includes: current moment, current moment reference position, road friction coefficient, vehicle longitudinal speed, vehicle lateral speed, vehicle lateral acceleration and yaw angular velocity; the current moment reference position is decomposed into the longitudinal coordinate and lateral coordinate of the current moment reference path based on the Frenet coordinate system.

3. The end-to-end control method for a controlled-by-wire vehicle according to claim 2, characterized in that: After the end-to-end deep learning model of the outer recurrent neural network is loaded, input parameters are extracted from the driving feature set, and the output parameters are obtained after inputting the input parameters into the end-to-end deep learning model of the outer recurrent neural network; The input parameters include: the longitudinal coordinate x of the reference path at the current moment refer , the horizontal coordinate y of the reference path at the current moment refer , road friction coefficient μ, vehicle longitudinal speed v x , vehicle lateral speed v y , vehicle lateral acceleration a y and yaw angular velocity ω; the input parameters constitute the input vector expressed as: The output parameters include the longitudinal coordinate of the path at time t The horizontal coordinate of the path at time t The output parameters constitute the output vector represented as:

4. The end-to-end control method for a controlled-by-wire vehicle according to claim 2, characterized in that: The inner model data set expresses the expected path data in time series; The inner model dataset is constructed by adopting a four-step forward method to construct the input parameters of the inner recurrent neural network, predicting and outputting the expected path data at time t+1, and continuously and recursively constructing the inner model dataset.

5. The end-to-end control method for a controlled-by-wire vehicle according to claim 4, characterized in that: The path tracking task includes three subtasks: path planning, path tracking, and deviation correction; Among them, the expected path data at time t+1 is predicted to realize path planning; the distance deviation and heading deviation between the actual trajectory and the reference path are calculated to realize path tracking; if the distance deviation and heading deviation do not meet the expected path requirements, the longitudinal coordinates and transverse coordinates of the path at time t+1 are constructed and input into the inner recurrent neural network, and the features of the next moment are recalculated to correct the deviation.

6. The end-to-end control method of a controlled-by-wire vehicle according to claim 1, characterized in that: The definition of the controller includes the following steps: Construct a vehicle dynamics model in a Cartesian coordinate system; The positional relationship expressed by the time-series path data of the vehicle in the Cartesian coordinate system is converted into the positional relationship in the Frenet coordinate system, including the arc length s traveled by the vehicle and the shortest distance d between the deviation of the actual trajectory and the reference path. The lateral and longitudinal motions of the vehicle's trajectory are decoupled, and the time derivatives of the arc length s and the shortest distance d are calculated respectively, which can be expressed as: Among them, θ x and θ r They represent the yaw angle of the vehicle on the actual tracking path and the yaw angle of the reference path respectively, κ is the curvature value on the path point, is the vehicle yaw angle, V is the vehicle's total velocity, and β is the center of mass sideslip angle; Calculate the distance deviation e between the actual trajectory and the reference path d The heading deviation between the actual trajectory at the current moment and the reference path A lateral motion controller is established for lateral motion in a feedback form, and a longitudinal motion controller is established for longitudinal motion in a feedforward form.

7. The end-to-end control method for a controlled-by-wire vehicle according to claim 6, characterized in that: The establishment of a lateral motion controller for lateral motion in the form of feedback refers to: The state feedback control law for determining the front wheel steering angle is: δ = -KX, where K = [τ1, τ2, τ3, τ4]; Among them, δ is the equivalent front wheel angle, X is the state variable τ1, τ2, τ3, and τ4 are user-defined constants.

8. The end-to-end control method for a controlled-by-wire vehicle according to claim 6, characterized in that: Establishing a longitudinal motion controller in a feedforward manner for longitudinal motion means: Add proportional-integral correction link in feedforward mode, through the longitudinal force F xf Correct the vehicle's longitudinal v x The error is expressed as: Among them, K p is the proportional gain coefficient; K i is the integral gain coefficient; and is the desired stable value of the system.

9. The end-to-end control method of a controlled-by-wire vehicle according to claim 1, characterized in that: The control requirements include a lateral motion displacement deviation and a longitudinal speed deviation; the lateral motion displacement deviation is used to ensure that the tracking path does not deviate, and the longitudinal speed deviation is used to control speed stability.

10. The end-to-end control method of a controlled-by-wire vehicle according to claim 7, characterized in that: The method for determining the state variable X includes: Get and its first-order and second-order derivatives, respectively, are expressed as: Get e d Its first-order and second-order derivatives are expressed as: Select state variables

Images