LSTM (Long Short Term Memory)-based automatic driving vehicle bend-entering deceleration prediction method

Abstract

The invention discloses an LSTM (Long Short Term Memory)-based automatic driving vehicle bend-entering deceleration prediction method, and aims to solve the problems of safety and stability in a curve driving process of an automatic driving vehicle. The invention relates to the field of automatic driving. The system comprises a data acquisition module, a data preprocessing module, a network training module and a real-time prediction module. The data acquisition module is responsible for acquiring vehicle state information and road information and outputting the information to the data preprocessing module. The data preprocessing module performs normalization processing on the vehicle state information and the road information, extracts an input-output pair by using a sliding window, and converts feature data and label data into tensor data suitable for LSTM input. The network training module trains an LSTM model according to the tensor data. And the real-time prediction module loads the optimal LSTM model for prediction, predicts and outputs the bending deceleration of the automatic driving vehicle, and is used for realizing safe and stable bending control of the automatic driving vehicle.

Description

Technical Field:

[0001] The present invention belongs to the field of autonomous driving, and specifically relates to a method for predicting the deceleration of an autonomous driving vehicle entering a curve based on LSTM. Background Art:

[0002] With the rapid development of technologies such as artificial intelligence, deep learning, sensor fusion, and high-precision positioning, the technology of autonomous driving vehicles is gradually moving from the R & D stage to the stage of actual road application. Existing autonomous driving systems usually integrate lidar, millimeter-wave radar, cameras, inertial navigation units (IMUs), and global positioning systems (GPS) to achieve environmental perception, path planning, and vehicle motion control.

[0003] During actual road driving, a vehicle usually needs to implement deceleration control in advance before entering a curve to reduce the impact of centrifugal force on vehicle stability, and at the same time avoid phenomena such as sideslip and instability of the vehicle in the curve, thereby ensuring driving safety and stability during the cornering process. During vehicle driving, a driver can comprehensively judge factors such as the curve radius and road adhesion conditions based on visual perception of the road ahead and driving experience, and thus make a deceleration decision in advance to achieve safe cornering. However, for autonomous driving vehicles, it is impossible to make a reasonable deceleration decision based on human driving experience when entering a curve, which may lead to instability of the autonomous driving vehicle.

[0004] Currently, the acceleration vector control technology (GVC) is adopted in some mass-produced vehicle models. By monitoring states such as the steering wheel angle and acceleration, the engine output is actively adjusted to achieve the transfer of load on the longitudinal axis of the vehicle, and the responsiveness and handling stability performance of the vehicle when entering a curve are improved. Specifically, when implemented, the GVC system constructs an empirical formula between the lateral acceleration and longitudinal acceleration of the vehicle based on driver driving data. Based on the current lateral acceleration a y the required deceleration a x when entering the curve is deduced, and the engine torque output is controlled accordingly to assist the vehicle in smoothly entering the curve. However, the control strategy of GVC depends on the current state and fails to utilize the time series information of historical states, and cannot predict the deceleration of the vehicle when entering the curve in advance. There will be a certain response lag during vehicle control, and it is difficult to achieve precise control in the face of complex and changeable curve road conditions. Therefore, it is necessary to design a method for predicting the deceleration when entering a curve based on time series by combining human driving data and road information, and accurately predict the deceleration required when entering the curve in advance, so as to provide a deceleration reference for the deceleration control of autonomous driving vehicles when entering the curve and achieve safe and stable operation of autonomous driving vehicles in curve scenarios. However, in existing research, there is no relevant research on predicting the deceleration of autonomous driving vehicles when entering a curve based on the time series of vehicle states and road information.

[0005] At present, some studies have attempted to predict the vehicle's deceleration or steering adaptability on curves. For example, patent CN10946860A proposes a method for predicting vehicle speed on curves based on driving style. By building a driving style evaluation system and combining fuzzy comprehensive evaluation to determine the driver type, the target speed can be adjusted. However, this method is based on the current state of vehicle dynamics, road environment and driving style characteristics, and does not consider the impact of time series characteristics on deceleration behavior. The prediction results rely more on empirical rules and are difficult to adapt to complex and dynamically changing road scenes. The long short-term memory network (LSTM) network is widely used in the time series prediction task of autonomous driving systems due to its advantages in time series feature extraction and modeling. The model can mine potential time series dependencies based on historical vehicle status, road parameters and multi-source sensor inputs, thereby realizing dynamic prediction of vehicle deceleration when entering a curve, and providing a reference for the control system to adjust the speed and path in advance. Compared with traditional control strategies, the LSTM model shows higher generalization ability and control accuracy in dealing with complex curve changes, sudden working conditions and nonlinear dynamic characteristics of the system. Patent CN 109783843 A considers the time series information of the vehicle state to predict the vehicle speed, which is used to improve the modeling accuracy of the vehicle dynamics model in the simulation platform. In addition, when predicting the speed, Patent CN 109783843 A does not consider the road curvature information, which is not suitable for driving conditions on curved roads, and cannot predict the deceleration when entering a curve. The acceleration obtained by directly derivation of the speed predicted by CN 109783843 A is difficult to guarantee the vehicle's deceleration function when entering a curve due to the information fluctuation caused by the derivation. Therefore, Patent CN 109783843 A is essentially different from the present invention in technology. Summary of the invention:

[0006] In view of the shortcomings of the prior art and in order to solve the problems existing in the above-mentioned background technology, the present invention provides a method for predicting the deceleration of an autonomous driving vehicle entering a corner based on LSTM. The method adopts a modular design and uses the LSTM network to perform time series modeling of the vehicle state and road information to achieve advance prediction of the deceleration entering the corner, thereby improving the system's responsiveness and control accuracy to dynamic changes in the curve. Compared with traditional control methods based on rules or threshold judgments, and strategies such as GVC that rely on engine output regulation, it has significant advantages in modeling accuracy and adaptability to complex curved road conditions. Through the deep learning prediction mechanism, this method exhibits stronger generalization ability and robustness, and can significantly improve the operating stability and control foresight of autonomous driving vehicles in curves.

[0007] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0008] A method for predicting the deceleration of an autonomous vehicle when entering a curve based on LSTM, which aims to solve the problems of safety and stability during the curve driving of autonomous vehicles. The present invention relates to the field of autonomous driving. The present invention includes a data acquisition module, a data preprocessing module, a network training module, and a real-time prediction module. The data acquisition module is responsible for collecting vehicle state information and road information and outputting them to the data preprocessing module. The data preprocessing module normalizes the data collected by the data acquisition module, extracts input-output pairs using a sliding window, and converts the feature data and label data into tensor data suitable for LSTM input. The network training module trains the LSTM model according to the tensor data, and the real-time prediction module loads the optimal LSTM model for prediction, and predicts and outputs the deceleration of the autonomous vehicle when entering the curve.

[0009] The method includes the following steps:

[0010] Step 1: Design a data acquisition module for collecting vehicle state information when an autonomous vehicle enters a curve, including lateral displacement y, lateral velocity v y , lateral acceleration a y , yaw rate γ, front wheel steering angle δ, and road information including heading angle ψ, road curvature κ. The above information will be used as the feature data of the LSTM network model and input into the data preprocessing module, and the expected deceleration a x when entering the curve will be used as label data for subsequent deceleration prediction when entering the curve.

[0011] Step 2: Design a data preprocessing module, and its process includes the following sub-steps:

[0012] Step 2.1: Normalize the feature data and label data through the MinMax normalization method, scale the data to the range [0, 1], and make the model training more stable and efficient. The MinMax normalization method preprocesses the data as shown in Equation (1):

[0013]

[0014] where X norm is the normalized data, X is the original data collected by the data acquisition module, X max is the maximum value of the feature data, and X min is the minimum value of the feature data.

[0015] Step 2.2: Adopt a sliding window method, select the normalized data of continuous N time steps as the input-output pairs of the LSTM network model, and convert them into tensor data as the input of the model.

[0016] Step 2.3: Divide the tensor data required for the LSTM network for predicting the deceleration when entering the curve into a training set, a validation set, and a test set.

[0017] Step 3: Design a network training module, and its process includes the following sub-steps:

[0018] Step 3.1: Define and initialize an LSTM network model, set the input dimension, output length, number of LSTM hidden layer units, number of LSTM network layers, set the loss function, optimizer, and learning rate scheduler.

[0019] Step 3.2: Conduct multiple rounds of training loops. In each round, the LSTM network receives tensor data through forward propagation, predicts the vehicle's deceleration when entering a curve, and simultaneously calculates the prediction error between the predicted value of the vehicle's deceleration when entering a curve and the true value. Subsequently, calculate the loss function based on the error between the predicted value and the true value, and update the network parameters; after each round ends, use the validation set to evaluate the model, check the generalization ability and training status of the model, and prevent overfitting.

[0020] Step 3.3: If the validation loss of the current round is the smallest, save the weights of the current LSTM network model and use it as the current optimal LSTM network model.

[0021] Step 3.4: After the LSTM network training is completed, load the trained network model, use the test set to predict the deceleration when entering a curve, and calculate the evaluation metrics.

[0022] Step 4: Design a real-time prediction module, load the optimal LSTM network model obtained from the network training module, utilize the vehicle state information and road information collected by the data acquisition module, and generate tensor data after being processed by the data preprocessing module, and conduct real-time prediction of the vehicle's deceleration when entering a curve for the safe and stable curve entry control of autonomous vehicles. Description of the Drawings:

[0023] Figure 1 is the flowchart of the method of the present invention.

[0024] Figure 2 is the flowchart of the data acquisition module of the present invention. Detailed Embodiments:

[0025] The present invention will be described in detail below with reference to the drawings.

[0026] The present invention proposes a method for predicting the deceleration of an autonomous vehicle when entering a curve based on LSTM. The method includes a data acquisition module, a data preprocessing module, a network training module, and a real-time prediction module. The data acquisition module is responsible for collecting vehicle state information and road information and outputting them to the data preprocessing module. The data preprocessing module normalizes the data collected by the data acquisition module, extracts input-output pairs using a sliding window, and converts the feature data and label data into tensor data suitable for LSTM input. The network training module trains an optimal LSTM model using the processed data, and the real-time prediction module loads the optimal network model and data for prediction, and predicts and outputs the deceleration of the autonomous vehicle when entering a curve. Refer to Figure 1 for illustration, and specifically includes the following steps:

[0027] Step 1. Design a data acquisition module:

[0028] The data acquisition module refers to Figure 2 for illustration, and its process includes the following sub-steps:

[0029] Step 1.1. Construct a sine working condition simulation scenario through the Carsim platform. This working condition is used to simulate typical road curvature disturbances and vehicle dynamic response behaviors, and outputs an ideal path trajectory including the lateral displacement y ref and the heading angle . At the same time, use a model predictive controller to control the CarSim vehicle model to track the ideal trajectory, and collect the vehicle state information of the autonomous vehicle when entering a curve.

[0030] Step 1.2. Construct a model predictive controller as follows:

[0031] Adopt a two-wheel two-degree-of-freedom vehicle model, a linear tire model, and a tracking error model to design an active front-wheel steering controller system model based on model predictive control (MPC). Among them, the two-wheel two-degree-of-freedom vehicle model of MPC is Equation (2), which is converted into the form of the two-degree-of-freedom motion differential equation of the vehicle (3):

[0032]

[0033] Among them, k1 is the lateral stiffness of the front tire, indicating the magnitude of the lateral force generated by a unit side slip angle; k2 is the lateral stiffness of the rear tire; β is the vehicle's center-of-mass side slip angle, indicating the angle between the vehicle body movement direction and the vehicle head direction; u is the vehicle's longitudinal speed; ω r represents the vehicle's yaw angular velocity, that is, the rotation rate of the vehicle around the vertical axis; δ is the front-wheel steering angle, which is the control input; m is the total vehicle mass; a and b are the distances from the vehicle's center of mass to the front axle and the rear axle respectively; is the lateral acceleration of the vehicle's center of mass; I z is the moment of inertia of the vehicle around the center of mass; represents the time derivative of the yaw rate, i.e., the yaw angular acceleration;

[0034]

[0035] The tracking error model extracts key error information for controller design by comparing the current motion state of the vehicle with the reference path state. Let the current lateral position of the vehicle be y, and the heading angle be The ideal lateral position on the corresponding reference path is y ref , and the ideal heading angle is Then the lateral deviation e y and the heading deviation are respectively defined as:

[0036]

[0037] To accurately describe the error variation trend and improve the prediction performance of the controller, the error derivative term, i.e., the rate of change of the error, is further introduced:

[0038]

[0039] where, represents the rate of change of the heading angle deviation; ω r is the yaw rate of the vehicle; u is the longitudinal speed of the vehicle; κ is the curvature of the reference path; represents the rate of change of the lateral deviation; is the heading angle deviation; ν is the disturbance term, usually used to represent modeling errors, external disturbances, or dynamic factors not considered by the controller;

[0040] Combining Equation (3) and Equation (5), the state-space form prediction model of the MPC controller is obtained as follows:

[0041]

[0042] where, is the state variable, the control input u represents the front wheel steering angle δ, d = κ, and the predicted output A, B u , B d , C are the system state matrix, control input matrix, disturbance matrix, and output matrix respectively, as follows:

[0043]

[0044]

[0045] In the MPC prediction model, they are all constant matrices within the prediction time domain;

[0046] In model predictive control, the controller performs rolling optimization and control input update at a fixed sampling period. Therefore, the system model must be expressed in discrete-time form. To apply the continuous-time state-space model to a digital controller, it is necessary to discretize it in time. The zero-order hold method is used to discretize Equation (3) to obtain the discrete-time model:

[0047]

[0048] where is the increment of the state variable, the control input increment is Δδ, and the predicted output is

[0049] When designing the MPC controller, to obtain the optimal control input sequence over a future period of time, it is necessary to online roll and solve a quadratic programming (QP) problem. The QP optimizer is based on the discretized system state-space model, takes the current system state as the initial state input, combines the desired path and control objectives, and calculates a set of optimal control input sequences that satisfy the constraints by minimizing the cost function within a finite prediction horizon:

[0050]

[0051] where represents the optimal front-wheel steering angle control input at the k-th step of the prediction horizon, and N is the length of the prediction horizon;

[0052] Although the QP solver outputs a complete control input sequence U * , only the first control input in this sequence is used as the execution instruction at the current moment and is input to the Carsim platform to control the actual movement of the vehicle.

[0053] Finally, the model predictive controller takes the desired path as the reference trajectory, combines the current state information of the vehicle, constructs a QP optimization problem based on the discrete state-space model, dynamically solves the optimal control sequence, and extracts the first control input value which is used as a control signal and transmitted to the Carsim simulation platform for the execution of the vehicle dynamics model, thereby achieving high-precision control of the front-wheel steering angle of the vehicle and ensuring cornering stability.

[0054] Step 1.3: Build a vehicle dynamics model based on the Carsim platform. Take the front-wheel steering angle δ output by the model predictive controller as the input signal, and at the same time input the longitudinal speed u and path curvature κ of the vehicle at the current moment. After the model runs, it outputs the dynamic response data of the vehicle, including the lateral displacement y, lateral speed v y , and lateral acceleration a y, yaw rate γ, front wheel steering angle δ, road curvature κ and heading angle ψ are used to characterize the vehicle trajectory response performance;

[0055] Step 1.4: Set an acceleration conversion module. According to the lateral acceleration a y and the driving state, calculate the expected deceleration a x when entering a curve, which is used to reflect the speed correction amplitude required by the vehicle under path deviation to support longitudinal motion control modeling;

[0056] Using the lateral acceleration change rate as the adjustment basis, through first-order filtering smoothing processing, dynamically adjust the longitudinal braking acceleration input to achieve active adjustment of the vehicle attitude. The specific control law expression is as follows:

[0057]

[0058] Where: G xc is the longitudinal braking acceleration; G y is the lateral acceleration; is the lateral acceleration change rate; C xy is the control gain; T s is the first-order filtering time constant.

[0059] Step 1.5: Collect vehicle operation state information and road information, including the lateral displacement y, lateral speed v y , lateral acceleration a y , yaw rate γ, front wheel steering angle δ, road curvature κ and heading angle ψ output by the Carsim vehicle model, as well as the expected deceleration a x when entering a curve. These are used as feedback variables and input to the data preprocessing module to construct time series samples and support the training and prediction of the subsequent LSTM network model.

[0060] Step 2: Design a data preprocessing module. Its process includes the following sub-steps:

[0061] Step 2.1: Normalize the feature data and label data through the MinMax normalization method, scale the data to the range [0, 1] to make the model training more stable and efficient. The MinMax normalization method preprocesses the data as shown in Equation (12).

[0062]

[0063] Where, X norm is the normalized data, X is the original data collected by the data acquisition module, X max is the maximum value of the feature data, and X min is the minimum value of the feature data;

[0064] Step 2.2: Use the fixed-length sliding window algorithm to slice the time series data to generate the input-output pairs required by the model. Each sliding window contains four time steps of continuous data with \(t = 1, 2, 3, 4\), and use them to predict the deceleration when entering the curve in the subsequent time step. The sliding window can help the model capture the dynamic characteristics in the time series, especially the temporal dependence of the vehicle state.

[0065] Step 2.3: Convert the normalized data into tensor data and adjust the dimensions according to the requirements of the LSTM network structure. The input tensor data is a three-dimensional tensor, which serves as the input of the model.

[0066] Step 2.4: Divide the data set into a training set, a validation set, and a test set. Divide the data set according to the set ratio (80% training set, 10% validation set, 10% test set) to prepare for subsequent training and evaluation.

[0067] Step 3: Design the network training module, and its process includes the following sub-steps;

[0068] Step 3.1: Define and initialize the LSTM network model. Its input is three-dimensional tensor data with dimensions \((N, C, T)\), where \(N\) represents the batch size, \(C = 8\) represents the feature dimension (corresponding to the lateral displacement \(y\), lateral velocity \(v\) y , lateral acceleration \(a\) y , yaw rate \(\gamma\), front wheel angle \(\delta\), road curvature \(\kappa\), heading angle \(\psi\), and deceleration when entering the curve \(a\) x ), and \(T = 4\) represents the input sequence length (i.e., the model predicts the deceleration in the next moment based on the data of 4 consecutive historical moments).

[0069] The LSTM network contains 2 layers of LSTM structures, each layer has 128 hidden units. The loss function is selected as the mean square error (MSE), and the optimizer is AdamW (the initial learning rate is set to 0.01, and the weight decay coefficient is 0.001);

[0070] Step 3.2: Execute the model training process. Set the total number of training epochs to 150;

[0071] In each epoch, the tensor data will be input into the LSTM model. The model predicts the deceleration when entering the curve of the autonomous driving vehicle through forward propagation, and at the same time calculates the error between the predicted value and the true value of the vehicle's deceleration when entering the curve. Subsequently, calculate the loss function according to the error between the predicted value and the true value, and update the network parameters.

[0072] In the LSTM model, the input at time step \(t\) includes the hidden state \(h\) of the previous time step t and the input \(x\) of the current time step t, these inputs are processed through different gating mechanisms. Specifically, the key gating mechanisms of LSTM are as follows:

[0073] Forget gate f t : Controls the retention ratio of historical information. The formula is as follows.

[0074] f t = σ(W f ·[h t-1 , x t + b f ) (13)

[0075] Among them, W f is the weight matrix of the forget gate, b f is the bias term, σ is the sigmoid activation function, h t-1 is the hidden state at the previous moment, x t is the input at the current moment, and the output value ranges between 0 and 1, determining how much information to forget;

[0076] Input gate i t : Controls how the current input affects the memory cell. The formula is as follows.

[0077] i t = σ(W i ·[h t-1 , x t + b i ) (14)

[0078] Among them, W i is the weight matrix of the input gate, b i is the bias term, determining the validity of the current input;

[0079] Candidate memory cell Generates candidate new memory content. The formula is as follows.

[0080]

[0081] Among them, W c is the weight matrix of the candidate memory cell, b c is the bias term, and tanh is the hyperbolic tangent activation function;

[0082] Memory cell update C t : Integrates the outputs of the forget gate and the input gate to update the memory cell. The formula is as follows.

[0083]

[0084] Among them, f t controls the retained historical information, and i t controls the addition of new information;

[0085] Output gate o t : Controls the output of the hidden state. The formula is as follows:

[0086] o t = σ(W o ·[h t-1 , x t +b o ) (17)

[0087] Where W0 is the weight matrix of the output gate, and b o is the bias term. The output value determines the contribution degree of the current hidden state;

[0088] Calculation of the final hidden state h t : Obtains the current hidden state by processing the memory cell through the output gate. The formula is as follows:

[0089] h t = o t ⊙ tanh(C t ) (18)

[0090] Where o t is the output of the output gate, and C t is the current memory cell;

[0091] Calculation of the final output: At each time step, the LSTM generates the output at the current moment, which is a combination of the current time step information and historical information. By fusing the long-term historical state and instantaneous features, the model can dynamically capture the deceleration trend required before entering the curve and achieve high-precision prediction of the dynamic characteristics of entering the curve.

[0092] After each episode ends, the model will enter the validation mode, evaluate the model performance using the validation set, and record the current validation loss value. At this time, the model will calculate the gradient through backpropagation and update the model parameters using the optimization algorithm. During training, we use gradient clipping to prevent gradient explosion, and the clipping threshold is set to 0.15. In addition, after each round of training, we will also update the learning rate scheduler according to the set strategy to ensure the stability of the training process;

[0093] Step 3.3: If the validation loss of the current episode is lower than the historical optimal value, update the record of the optimal network model and save the model parameters obtained in this round of training;

[0094] Step 3.4: After training is completed, load the optimal network model, use the test set for prediction and calculate various evaluation metrics, mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R 2), these evaluation metrics can help us quantify the prediction ability of the model, ensure that the model can accurately predict the deceleration when entering a curve, and obtain reliable control decisions in practical applications;

[0095] Step 4: Design a real-time prediction module, load the optimal LSTM network model obtained from the network training module, utilize the vehicle state information and road information collected by the data acquisition module, and generate tensor data after being processed by the data preprocessing module to perform real-time prediction of the vehicle's deceleration when entering a curve.

[0096] Step 5: Design a PID controller module, and its process includes the following sub-steps;

[0097] Step 5.1: Take the deceleration of the vehicle when entering a curve, which is real-time predicted by the optimal LSTM network model in Step 4 as the target control quantity at the current moment. This value represents the desired deceleration when entering a curve that the vehicle needs to achieve before entering the curve area, and is used to ensure that the vehicle passes through the curve at a safe speed and improve stability and maneuverability.

[0098] Step 5.2: Collect the actual longitudinal acceleration of the vehicle at the current moment on the Carsim simulation platform and perform a difference with the predicted value to obtain the longitudinal error signal e(t):

[0099]

[0100] Among them, a positive error indicates insufficient braking at present and the brake pedal input needs to be increased; a negative error indicates that the current deceleration is too large and the braking command needs to be reduced.

[0101] Step 5.3: Based on the classical PID controller structure, calculate the control quantity u(t) of the brake pedal opening in real-time, and its control law is as follows:

[0102]

[0103] Among them, u(t) is the control command of the brake pedal opening at the current moment; K p , K i , K d are the proportional, integral, and differential control gains respectively; e(t) is the real-time longitudinal error value; the controller parameters are obtained by debugging according to the vehicle dynamics model to ensure both response speed and stability.

[0104] Step 5.4: Send the calculated braking command u(t) to the vehicle dynamics model of the Carsim simulation platform as the braking control input (brake pedal opening) at the current moment. After the vehicle responds to the control command in the simulation environment, its new deceleration when entering a curve It will be fed back to the control module as the input for the next control cycle to achieve closed-loop control of longitudinal deceleration when entering a curve.

Claims

1. A deceleration prediction method for an autonomous vehicle entering a curve based on LSTM, characterized in that: The method includes a data acquisition module, a data preprocessing module, a network training module, and a real-time prediction module; the data acquisition module is responsible for collecting vehicle status information and road information and outputting them to the data preprocessing module; The data preprocessing module normalizes the data collected by the data acquisition module, extracts input-output pairs using a sliding window, and converts the feature data and label data into tensor data suitable for LSTM input; The network training module trains the LSTM model according to the tensor data, and the real-time prediction module loads the optimal network model for prediction, and the prediction outputs the deceleration of the autonomous vehicle when entering a curve.

2. The method for predicting the deceleration of an autonomous vehicle entering a curve based on LSTM according to claim 1, wherein: The described data acquisition module is used to collect the vehicle state information when the autonomous vehicle enters a bend, including the lateral displacement y, the lateral speed v y , the lateral acceleration a y , the yaw rate γ, the front wheel steering angle δ, and the road information including the heading angle ψ, the road curvature κ. The above information will be used as the feature data of the LSTM network model and input into the data preprocessing module. The desired in-bend deceleration a x is used as the label data for subsequent in-bend deceleration prediction; The data preprocessing module normalizes the feature data and label data through the MinMax normalization method, scales the data to the range of [0,1], and the MinMax normalization method preprocesses the data as shown in Equation (1): Among them, X norm is the normalized data, X is the original data collected by the data acquisition module, X max is the maximum value of the feature data, X min is the minimum value of the feature data; Subsequently, the sliding window method is adopted to select the normalized data of consecutive N time steps as the input-output pairs of the LSTM network model, and convert them into tensor data as the input of the model; finally, the tensor data is divided into a training set, a validation set, and a test set; The network training module is divided into two processes: forward propagation and backward propagation; during forward propagation, first define and initialize the LSTM network model. In each round, the LSTM network receives the tensor data and predicts the deceleration of the autonomous vehicle when entering a curve through forward propagation, and at the same time calculates the error between the predicted value and the true value of the vehicle deceleration when entering a curve; during backward propagation, calculate the loss function according to the error between the predicted value and the true value, and update the network parameters; after the LSTM network training is completed, load the trained network model, use the test set to predict the deceleration when entering a curve and calculate the evaluation index; The real-time prediction module loads the optimal LSTM network model obtained by the network training module, utilizes the vehicle status information and road information collected by the data acquisition module, and generates tensor data after being processed by the data preprocessing module, and performs real-time prediction of the vehicle deceleration when entering a curve for the safe and stable curve entry control of the autonomous vehicle.

Images