Cooperative steering control method considering man-machine mutual reliability

Abstract

The invention discloses a cooperative steering control method considering man-machine mutual reliability, and aims to solve the problem of man-machine conflict in an existing man-machine co-driving system. The invention relates to the field of intelligent driving. The system specifically comprises an expected path module, a man-machine steering conflict evaluation module, a driver experience evaluation module, a weight distribution module, a non-cooperative game module and a man-machine co-driving vehicle. The expected path module generates an expected path between the driver and the controller; the man-machine steering conflict evaluation module outputs the trust degree of the driver to the controller based on the steering action of the driver, the vehicle state and the expected path of the driver and the controller; the driver experience evaluation module outputs the driver trust degree of the controller based on the vehicle state; the weight distribution module is used for outputting a driver control weight; and the non-cooperative game module outputs an optimal turning angle in combination with the expected path of the driver and the controller and the control weight of the driver, and controls the man-machine co-driving vehicle to complete the cooperative steering action.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent driving, and in particular to a cooperative steering control method considering human-machine trust degree. BACKGROUND

[0002] In recent years, with the rapid development of automobile automation technology, intelligent driving systems have shown great potential in improving driving safety and reducing driving burden. However, due to the maturity of current perception, decision-making and execution technology, fully automated driving is still difficult to fully realize, and human-machine co-driving as a transitional form of automated driving has become a current research hotspot. Although existing cooperative control technology has achieved certain results, there are often conflicts between the driver and the controller in decision-making. Game theory has a unique advantage in multi-agent interactive decision-making and can effectively alleviate the decision-making conflict between humans and machines. However, existing methods mostly use fixed or experiential weight allocation mechanisms, which lack adaptability and are difficult to cope with complex and variable driving scenarios. Therefore, the weight allocation strategy needs to be dynamically and intelligently adjusted.

[0003] Further, human-machine trust degree has gradually become an important basis for dynamic weight allocation, and its core idea is generally focused on how to quantify the driver's steering intention and the driver's driving ability, which are two key dimensions, and to guide the dynamic allocation of control power based on this. However, existing research relies on simplified assumptions and static rules, making it difficult to dynamically capture the driver's complex cognition and personalized style, and in terms of driving ability evaluation, it is usually only quantified by using indicators such as lateral tracking error and lateral acceleration, which is difficult to truly reflect the driver's experience level and cognitive characteristics, resulting in one-sided and insufficient adaptability of trust evaluation. To solve this problem, the present patent combines an LSTM-Transformer driver turning angle prediction model and an inverse reinforcement learning experience quantization model to capture the spatiotemporal features of driving behavior, achieve high-precision prediction of steering intention, and effectively quantify the driver's implicit experience and decision-making reasoning by inverting the potential reward function from expert driving trajectories. Compared with existing methods, the present application significantly enhances the adaptability to complex environments and individual differences, achieving more accurate and reliable dynamic weight allocation in human-machine collaboration.

[0004] Currently, there are various methods proposed for human-machine trust degree in human-machine collaborative control research, each with its own advantages and limitations. Patent CN114537419A dynamically adjusts the human-machine control weight by quantifying the driver's activity level and basic driving ability. However, the traditional driver model used in this method mostly uses static preset parameters, making it difficult to fully represent the dynamic fluctuations and individual differences in driver state. At the same time, only a single indicator such as lateral deviation and heading angle is used to depict comprehensive driving ability, ignoring the interactive influence between multi-dimensional information such as risk perception, decision-making stability and fatigue, resulting in insufficient model adaptability.

[0005] The patent CN120135223A switches and adjusts the control right of man and machine by identifying the long-term behavior style and comprehensive driving ability of the driver; however, it only calculates the driving ability through indicators such as driving years and fatigue degree, and fails to fully utilize multi-dimensional real-time data to dynamically reflect the complex and subtle state changes of the driver.

[0006] The patent CN202211540624A proposes a man-machine collaborative steering framework based on LQR and feedforward control; by establishing a second-order driver model containing double-point preview and neuromuscular characteristics, online identifying driver parameters, and constructing a shared behavior model, the control weight of man and machine is finally coordinated through LQR feedback, feedforward compensation and defined stiffness parameters; although the patent introduces double-point preview and neuromuscular characteristics, the driver model is still based on a fixed second-order linear structure, which cannot fully represent the nonlinear and strong time-varying characteristics of driving behavior, and the dynamic capture ability of individual differences and state is limited.

[0007] In view of the above shortcomings, the present application proposes a collaborative steering control method considering the mutual trust degree of man and machine. The core is to construct a dynamic weight distribution strategy of bidirectional trust evaluation mechanism, which not only plays the conflict resolution ability of game collaboration, but also improves the adaptability and user acceptance of the system by introducing the behavior cognition dimension of "man-machine mutual trust". Compared with the prior art, the present method breaks through the limitations of traditional models in representing nonlinearity and single driving ability evaluation index by using LSTM-Transformer model and inverse reinforcement learning model, and realizes more accurate driver steering prediction and driving experience evaluation; further, the system takes real-time man-machine mutual trust evaluation as the basis for decision-making, dynamically allocates the driver control weight through fuzzy rules combined with the trust degree of man and machine, and optimizes and solves it combined with non-cooperative game, so as to change the traditional "conflict after passive balance" of man-machine interaction mode to "cooperation in active guidance" based on bidirectional trust degree; this change not only enhances the fluency and safety of the interaction process, but also deeply improves the acceptability and user experience of the system. SUMMARY

[0008] The application provides a cooperative steering control method considering human-machine mutual trust degree, which is based on a human-machine steering conflict evaluation module and a driver experience evaluation module to calculate driver trust degree for the controller and controller trust degree for the driver respectively.

[0009] The technical solutions adopted by the application to solve the technical problems are as follows:

[0010] To solve the problems in the above technical background, the application provides a cooperative steering control method considering human-machine mutual trust degree; the method forms a closed-loop control system through an expected path module, a human-machine steering conflict evaluation module, a driver experience evaluation module, a weight distribution module, a non-cooperative game module and a human-machine co-driving vehicle to realize cooperative steering control of the human-machine co-driving vehicle; the expected path module is used to generate expected paths of the driver and the controller; the human-machine steering conflict evaluation module receives driver steering actions, vehicle states, expected paths of the driver and the controller and outputs driver trust degree for the controller; the driver experience evaluation module receives vehicle states and outputs controller trust degree for the driver; the weight distribution module takes driver trust degree for the controller and controller trust degree for the driver as inputs and outputs driver control weight based on fuzzy rules; the non-cooperative game module combines expected paths of the driver and the controller, driver control weight and vehicle states, establishes a prediction model, designs an objective function and solves Nash equilibrium to obtain an optimal steering angle and realize cooperative steering control of the human-machine co-driving vehicle.

[0011] The method comprises the following steps:

[0012] Step 1, expected path module design, based on driving environment information, the expected paths of the driver and the controller are obtained, including the horizontal and vertical coordinates and the yaw angle of the driver and the controller expected path points.

[0013] Step 1.1, driver expected path design, the expected path of the driver is obtained based on a Logitech G29 steering wheel simulation bench, including the horizontal and vertical coordinates and the expected yaw angle of the driver expected path.

[0014] Step 1.2, controller expected path design, the expected path of the controller is obtained based on the CarSim simulation environment, including the expected path of the controller and the expected yaw angle.

[0015] Step 2, human-machine steering conflict evaluation module design:

[0016] Based on the difference between the driver's predicted turning angle and the controller's predicted turning angle Real-time evaluation of driver's trust in controller .

[0017] Step 2.1, LSTM-Transformer driver turning angle prediction model design, the process includes the following sub-steps:

[0018] Step 2.1.1, data input module, used to collect driver steering action and vehicle state information of human-machine co-driving vehicle, including driver turning angle , driver turning angle rate and vehicle longitudinal speed , vehicle yaw rate , longitudinal acceleration , longitudinal coordinate of driver expected path , lateral coordinate of driver expected path , yaw angle of driver expected path ; Driver steering action and vehicle state information will be input as feature data of LSTM-Transformer driver turning angle prediction model to data preprocessing module, and expected driver turning angle as label data for subsequent driver turning angle prediction.

[0019] Step 2.1.2, data preprocessing module, the process includes the following sub-steps:

[0020] Step 2.1.2.1, normalize the feature data and label data by MinMax normalization method, so that the model training is more stable and efficient, MinMax normalization method for data preprocessing, as shown in equation (1):

[0021] (1)

[0022] Where, is the normalized data, is the original data of driver steering action and vehicle state information collected by the data acquisition module, is the maximum value of the feature data, is the minimum value of the feature data.

[0023] Step 2.1.2.2. The normalized data of each time step is selected as the input and output pair of the LSTM-Transformer network model, and is converted into tensor data as the input of the model.

[0024] Step 2.1.2.3. The tensor data required by the LSTM-Transformer network for predicting the driver's steering angle is divided into training set, validation set and test set.

[0025] Step 2.1.3. Design the training process of the LSTM-Transformer driver steering angle prediction model, including the following sub-steps:

[0026] Step 2.1.3.1. Define and initialize the LSTM-Transformer driver steering angle prediction model, set the input dimension, output length, LSTM hidden layer unit number, LSTM network layer number, Transformer network layer number, Transformer attention head number, Transformer feedforward network dimension, Transformer attention mechanism dimension, set the loss function, optimizer and learning rate scheduler.

[0027] Step 2.1.3.2. In each round of training, the LSTM-Transformer driver steering angle prediction model receives the preprocessed tensor data through forward propagation, outputs the predicted steering angle of the driver, and calculates the error between the predicted steering angle and the true steering angle; then, based on the error, the loss function is calculated, and the network weight parameters are updated through the back propagation algorithm; after each round of training, the independent validation set is used to evaluate the performance of the current model, monitor its generalization ability and training state, and effectively prevent overfitting phenomenon.

[0028] Step 2.1.3.3. Save the weight of the LSTM-Transformer driver steering angle prediction model with the minimum validation loss as the current optimal LSTM-Transformer driver steering angle prediction model.

[0029] Step 2.1.3.4. After the training of the LSTM-Transformer driver steering angle prediction model is completed, load the trained network model, use the test set to predict the driver's steering angle and calculate the evaluation index.

[0030] ​Step 2.1.4, design a real-time prediction module, load the optimal LSTM-Transformer driver turning angle prediction model obtained by the network training module; this module receives the driver steering action, vehicle state and driver and controller expected path information from the data acquisition module in real time during vehicle operation, generates tensor data through the data preprocessing module; then, the tensor data is input into the loaded LSTM-Transformer driver turning angle prediction model, the time sequence dynamic characteristics of the input sequence are extracted through the LSTM layer, and the short-term habits and state evolution of the driver operation are captured; further, the time sequence characteristics are weighted by the Transformer encoder layer, the global context information is fused to focus on the key decision moment, and finally the driver predicted turning angle is output through the full connection layer regression .

[0031] Step 2.2, MPC controller turning angle prediction model design, the process includes the following sub-steps:

[0032] Step 2.2.1, prediction model establishment, using vehicle dynamics model as prediction model, as shown in equation (2):

[0033] (2)

[0034] wherein, , including vehicle lateral velocity , vehicle yaw rate , vehicle lateral coordinate , vehicle yaw angle ; controller turning angle, road curvature, , and state coefficient matrix, output coefficient matrix.

[0035] Step 2.2.2, objective function design, design the following multi-objective cost function for optimizing path tracking accuracy and control amount, as shown in equation (3):

[0036] (3)

[0037] wherein, controller expected path, Q and H are weight matrices.

[0038] Step 2.2.3, optimization solution, solve the multi-objective optimization problem shown in equation (3) at each control period to obtain the controller predicted turning angle .

[0039] Step 2.3, driver's trust degree to controller is designed as formula (4) shows:

[0040] (4)

[0041] wherein, is the adjustment of human-machine corner conflict sensitivity factor, is the human-machine corner conflict factor, is the acceptable conflict degree value;

[0042] Step 2.3.1, the human-machine corner conflict factor is designed based on the absolute value of the difference between the driver's corner and the controller's corner, as formula (5) shows:

[0043] (5)

[0044] wherein, and are the driver's and the controller's corners in the prediction time domain respectively, is the predicted corner, is the current time, is the human-machine corner prediction time domain.

[0045] Step 3, driver experience evaluation module design:

[0046] Based on the inverse reinforcement learning driver experience evaluation model, the vehicle state is received, and the driver experience reward value is output, further based on the driver experience reward value to evaluate the trust degree of the controller to the driver .

[0047] Step 3.1, inverse reinforcement learning driver experience evaluation model design, the process includes the following sub-steps:

[0048] Step 3.1.1, state input and feature construction:

[0049] The inverse reinforcement learning driver experience evaluation model receives vehicle state data as input, specifically including lateral tracking error , vehicle longitudinal speed , collision time between the ego vehicle and the front vehicle , vehicle front wheel corner ; further, a feature vector representing the driver's experience is constructed, including safety, tracking accuracy, longitudinal smoothness and lateral stability.

[0050] Step 3.1.2, driver experience reward function design, based on safety, tracking accuracy, longitudinal smoothness and lateral stability for design; wherein the safety considers the collision time of the ego vehicle and the front vehicle, the tracking accuracy considers the lateral tracking error, the longitudinal smoothness considers the longitudinal vehicle speed change rate, and the lateral stability considers the vehicle front wheel angle change rate, as shown in equations (6) to (11):

[0051] (6)

[0052] (7)

[0053] (8)

[0054] (9)

[0055] (10)

[0056] (11)

[0057] wherein, is the safety reward function, is the tracking accuracy reward function, is the longitudinal smoothness reward function, is the lateral stability reward function, and are the collision time of the ego vehicle and the front vehicle in the first frame and the safe collision time threshold, is the lateral tracking error in the first frame, is the longitudinal vehicle speed change rate in the first frame, is the vehicle front wheel angle change rate in the first frame, is the current frame number, is the weight of each reward function, is the standard deviation of each reward function, wherein , , is the total number of sampling frames for driver experience evaluation; the safety, tracking accuracy, longitudinal smoothness and lateral stability reward functions are all normalized, and the weight is obtained by inverse reinforcement learning from expert data.

[0058] Step 3.1.3, inverse reinforcement learning driver experience evaluation model training process, the process includes the following sub-steps:

[0059] Step 3.1.3.1: Construction of expert dataset. Based on a large amount of publicly available natural driving data from experienced drivers in various scenarios, an expert demonstration trajectory set is formed.

[0060] Step 3.1.3.2: Design the learning objective function and optimize iteratively to ensure that the learned reward function can explain expert behavior.

[0061] Step 3.1.3.3: Training process of the inverse reinforcement learning driver experience evaluation model. The maximum entropy inverse reinforcement learning framework is adopted. Through iterative optimization algorithm, the weights and parameters of the reward function are adjusted to maximize the expected cumulative reward of the expert policy under the function. Thus, the optimal reward function model that can explain the expert driving preferences and decision-making logic is deduced in reverse.

[0062] Step 3.1.3.4: Output the optimal driver experience reward function. After training, the model obtains a set of optimal reward function weights. This leads to the optimal driver experience reward function.

[0063] Step 3.1.3.5: Calculate the real-time reward value. Input the historical and current vehicle state data into the trained driver experience reward function to calculate the driver's experience reward value for the current time period. .

[0064] Step 3.2, Controller's trust in the driver The design is as shown in equation (12):

[0065] (12)

[0066] in, To adjust the sensitivity factor of driver experience reward value, This is a driver experience reward value. The median reward value for the dataset of experienced drivers.

[0067] Step 4, Weight Allocation Module Design:

[0068] This module assigns weights based on fuzzy rules, taking into account the driver's trust in the controller. And the controller's trust in the driver As input, the driver control weights are dynamically output.

[0069] The specific steps for designing fuzzy rules are as follows:

[0070] Step 4.1: Confirm input variable: Driver's trust level with the controller And the controller's trust in the driver .

[0071] Step 4.2, fuzzification: fuzzify and to the corresponding fuzzy sets .

[0072] Step 4.3, fuzzy rule inference, infer according to the pre-set fuzzy rules, there are nine fuzzy rules in total, including:

[0073] Rule 1, if is low and is low, then the driver control weight is low;

[0074] Rule 2, if is low and is medium, then the driver control weight is medium;

[0075] Rule 3, if is low and is high, then the driver control weight is high;

[0076] Rule 4, if is medium and is low, then the driver control weight is low;

[0077] Rule 5, if is medium and is medium, then the driver control weight is medium;

[0078] Rule 6, if is medium and is high, then the driver control weight is high;

[0079] Rule 7, if is high and is low, then the driver control weight is low;

[0080] Rule 8, if is high and is medium, then the driver control weight is medium;

[0081] Rule 9, if is high and is high, then the driver control weight is high.

[0082] Step 4.4, defuzzification, defuzzify by the centroid method to convert the fuzzy set to the precise driver control weight value , as shown in equation (13):

[0083] (13)

[0084] where, is the total number of fuzzy rules triggered under the input variable, the first membership of the fuzzy rule triggered, , the center value of the corresponding weight interval, .

[0085] Step 4.5, output variable driver control weight , the driver path tracking weight as a non-cooperative game module , as shown in equation (14):

[0086] (14)

[0087] Step 5, non-cooperative game module design, the process includes the following sub-steps:

[0088] Step 5.1, prediction model establishment, build a discrete driver and controller interaction model, as shown in equation (15):

[0089] (15)

[0090] where, is the driver steering angle; is the controller steering angle; state variable , , which , , and are the vehicle lateral speed, vehicle yaw rate, vehicle lateral coordinate and vehicle yaw angle, respectively; , , , is the coefficient matrix.

[0091] Step 5.2, objective function design, in order to make the driver and controller interaction model in the mode of man-machine co-driving, reduce the man-machine decision conflict, improve the driving safety, while ensuring the stability of the vehicle, as shown in equation (16):

[0092] (16)

[0093] In the formula,

[0094] where, and are the objective function of the driver and the objective function of the controller, respectively, is the control output, and are the expected path of the driver and the controller, respectively, and the control input of the driver and the controller, respectively, and the single-step path information of the driver and the controller, respectively, and the single-step path tracking weight of the driver and the controller, respectively, and the single-step control input weight of the driver and the controller, respectively. , denotes the human-machine interaction model prediction horizon.

[0095] Step 5.3, Nash equilibrium solution:

[0096] Step 5.3.1, first define the error variables of the driver and the controller, as shown in equation (17):

[0097] (17)

[0098] where, and are the error variables of the driver and the controller, respectively, is the state coefficient matrix, and are the control input coefficient matrices of the driver and the controller, respectively.

[0099] Step 5.3.2, based on the least squares method and convex iteration method, the optimal control input sequence of the driver and the controller can be obtained; further, the first value of each sequence is selected as the optimal solution, as shown in equation (18):

[0100] (18)

[0101] where, and represent the optimal solution of the driver and the controller, respectively; and represent the optimal control input sequence of the driver and the controller, respectively.

[0102] Step 5.4, finally, the steering executed by the human-machine co-pilot vehicle, as shown in equation (19):

[0103] (19)

[0104] where, is the steering executed by the human-machine co-pilot vehicle.

[0105] The beneficial effects of this invention are as follows: First, this invention constructs a closed-loop, two-way trust mechanism, significantly improving the accuracy of evaluating driver intent and driving experience. Through the human-machine steering conflict assessment module, the spatiotemporal features of driving steering behavior are captured using the LSTM-Transformer driver angle prediction model and compared with the MPC controller's angle prediction results, thereby accurately quantifying the driver's trust in the controller. Simultaneously, a driver experience assessment module is introduced, which uses inverse reinforcement learning to invert the latent reward function from expert driving trajectories. Driver experience is objectively evaluated through safety, tracking accuracy, longitudinal stability, and lateral stability indicators, further assessing the controller's trust in the driver. This overcomes the limitations of traditional methods that rely solely on single indicators such as lateral tracking accuracy and lateral acceleration, and are difficult to characterize nonlinear features. Second, this method uses fuzzy rules for dynamic weight allocation, taking the trust levels of both parties as input. Fuzzy rules are used to output the driver's control weights in real time, and Nash equilibrium is solved using non-cooperative game theory, effectively mitigating human-machine conflict. Finally, this invention achieves a fundamental improvement in the human-machine interaction paradigm. The system can adjust control weights in advance based on mutual trust states, enhancing the system's adaptability in complex environments and the driver's acceptance of the intelligent system. Attached Figure Description

[0106] Figure 1 This is a schematic diagram of a cooperative steering control system that considers human-machine trust in this invention.

[0107] Figure 2 This is a flowchart of the LSTM-Transformer driver angle prediction model of the present invention.

[0108] Figure 3 This is a flowchart of the MPC controller corner prediction model of the present invention.

[0109] Figure 4 This is a flowchart illustrating the design of the inverse reinforcement learning driver experience evaluation model of the present invention. Detailed Implementation

[0110] The present invention will now be described in detail with reference to the accompanying drawings.

[0111] The application provides a cooperative steering control method considering human-machine mutual trust degree; the method forms a closed-loop control system through six modules of a desired path module, a human-machine steering conflict evaluation module, a driver experience evaluation module, a weight distribution module, a non-cooperative game module and a human-machine co-driving vehicle, and realizes cooperative steering control of the human-machine co-driving vehicle; wherein the desired path module is used for generating a driver desired path and a controller desired path; the human-machine steering conflict evaluation module receives a driver steering action, a vehicle state, the driver desired path and the controller desired path, predicts a driver steering angle through an LSTM-Transformer model, simultaneously predicts a controller steering angle through an MPC model, compares the difference between the two, and quantifies driver trust degree for the controller; the driver experience evaluation module receives the vehicle state, extracts safety, tracking accuracy, longitudinal smoothness and lateral stability features from the vehicle state through an inverse reinforcement learning model, and calculates a driver experience reward value to evaluate controller trust degree for the driver; the weight distribution module takes the driver trust degree for the controller and the controller trust degree for the driver as inputs based on fuzzy rules, and outputs a driver control weight; finally, the non-cooperative game module combines the driver desired path and the controller desired path, the driver control weight and the vehicle state, and obtains an optimal steering angle through three steps of prediction model establishment, target function design and Nash equilibrium solution, thereby realizing cooperative steering control of the human-machine co-driving vehicle.

[0112] Step 1, design of the desired path module

[0113] The desired path module is used for generating a driver desired path and a controller desired path.

[0114] Step 1.1, design of the driver desired path

[0115] Driver desired path In a joint simulation platform of CarSim and Simulink, the driver desired path is obtained through a Logitech G29 steering wheel input module based on a driving environment; wherein, and are a longitudinal coordinate and a lateral coordinate of the driver desired path, respectively, is a yaw angle of the driver desired path.

[0116] Step 1.2, design of the controller desired path

[0117] Controller desired path , which is obtained based on a CarSim simulation environment; wherein, and are a longitudinal coordinate and a lateral coordinate of the controller desired path, respectively, is a yaw angle of the controller desired path.

[0118] Step 2, design of human-machine steering conflict evaluation module, including design of LSTM-Transformer driver steering angle prediction model and design of MPC controller steering angle prediction model.

[0119] Step 2.1, design of LSTM-Transformer driver steering angle prediction model:

[0120] Step 2.1.1, data input module, used to collect driver steering action and vehicle state information under normal working condition of human-machine co-driving vehicle, including driver steering angle , driver steering angle change rate , and vehicle longitudinal speed , vehicle yaw rate , longitudinal acceleration , longitudinal coordinate of driver expected path , lateral coordinate of driver expected path , yaw angle of driver expected path , the above information will be input as feature data of LSTM-Transformer driver steering angle prediction model to data preprocessing module, and the expected driver steering angle is used as label data for subsequent driver predicted steering angle prediction.

[0121] Step 2.1.2, data preprocessing module, its process includes the following substeps:

[0122] Step 2.1.2.1, normalize the feature data and label data by MinMax normalization method to make the model training more stable and efficient, the MinMax normalization method for data preprocessing is shown in formula (1).

[0123] Step 2.1.2.2, adopt sliding window method, select continuous normalized data of time steps as input and output of LSTM-Transformer network model, and convert it into tensor data as input of the model.

[0124] Step 2.1.2.3, divide the tensor data required by the LSTM-Transformer network for predicting driver steering angle into training set 80%, validation set 10% and test set 10%.

[0125] Step 2.1.3, design of LSTM-Transformer driver steering angle prediction model training process, its process includes the following substeps:

[0126] Step 2.1.3.1, define and initialize LSTM-Transformer network model:

[0127] The input is a three-dimensional tensor data, with dimensions of , wherein, represents the batch size, represents the feature dimension (driver steering angle , driver steering angle rate of change , vehicle longitudinal speed , vehicle yaw rate , longitudinal acceleration , longitudinal coordinate of the driver's expected path , lateral coordinate of the driver's expected path , yaw angle of the driver's expected path ), represents the input sequence length (i.e., the model predicts the driver's predicted steering angle based on the data of the last 20 historical time points, corresponding to a history window of 0.4 seconds).

[0128] The network structure is designed as follows:

[0129] LSTM layer: contains 2 layers of stacked LSTM structure, aiming to capture the long and short time dependence features of time series, with 128 hidden units in each layer.

[0130] Transformer encoder layer: connected after the LSTM layer, used to extract global attention features; set the number of attention heads to 4, the dimension of the feedforward network to 256, and the number of network layers to 1.

[0131] Fully connected layer: output dimension is 1, directly output the predicted driver's predicted steering angle .

[0132] The training hyperparameters are set as follows:

[0133] Loss function: mean square error is used to measure the deviation between the driver's predicted steering angle and the driver's real steering angle.

[0134] Optimizer: AdamW optimizer is selected, with an initial learning rate of 0.001 and a weight decay coefficient of 0.0001.

[0135] Learning rate scheduling: cosine annealing strategy is adopted, with a minimum learning rate of 0.000001.

[0136] Step 2.1.3.2, execute the model training process:

[0137] Set the total training period to 500 rounds; in each round, the normalized tensor data is input into the hybrid model, the model outputs the driver's predicted steering angle through forward propagation, and calculates the gradient with the label data; then, the network weight parameters are updated through the backpropagation algorithm.

[0138] Step 2.1.3.3. If the validation loss of the current round is the smallest, save the weights of the current LSTM-Transformer driver turning angle prediction model as the current optimal LSTM-Transformer driver turning angle prediction model.

[0139] Step 2.1.3.4. After the training of the LSTM-Transformer driver turning angle prediction model is completed, load the trained network model, use the test set to predict the driver turning angle and calculate the evaluation index.

[0140] Step 2.1.4. Design a real-time prediction module to load the optimal LSTM-Transformer driver turning angle prediction model obtained from the network training module; during the operation of the vehicle, this module receives the driver steering action, vehicle state and driver and controller expected path information from the data acquisition module in real time, and generates tensor data through the same data preprocessing process as offline training; then, the tensor data is input into the loaded LSTM-Transformer driver turning angle prediction model, which first extracts the time series dynamic features of the input sequence through its LSTM layer, capturing the short-term habits and state evolution of the driver's operation; further, the Transformer encoder layer weights the time series features with self-attention, integrates global context information to focus on key decision moments, and finally outputs the driver predicted turning angle through the fully connected layer .

[0141] Step 2.2. MPC controller turning angle prediction model design, the process includes the following sub-steps:

[0142] Step 2.2.1. Prediction model establishment, using vehicle dynamics model as prediction model, as shown in the summary formula (2).

[0143] The vehicle dynamics parameters and controller parameters are set as follows:

[0144] Vehicle dynamics model prediction time domain , corresponding to the prediction of the vehicle state for 0.4 seconds.

[0145] Control time domain , that is, only optimize the control amount of the future five steps.

[0146] Step 2.2.2. Objective function design, design the following multi-objective cost function to optimize the path tracking accuracy and control amount, as shown in the summary formula (3).

[0147] Step 2.2.3. Optimization solution, solve the above optimization problem at each control period to obtain the controller predicted turning angle .

[0148] Step 2.3, Driver's Trust in Controller The design is shown in the summary formula (4).

[0149] Step 2.3.1, The human-machine corner conflict factor is designed based on the absolute value of the difference between the driver's corner and the controller's corner, as shown in the summary formula (5).

[0150] Step 3, Driver experience evaluation module design:

[0151] Based on the inverse reinforcement learning driver experience evaluation model, the vehicle state is received, and the driver experience reward value is output , further, based on the driver experience reward value evaluate the trust of the controller to the driver .

[0152] Step 3.1, Inverse reinforcement learning driver experience evaluation model design, the process includes the following sub-steps:

[0153] Step 3.1.1, Feature extraction and state input:

[0154] The inverse reinforcement learning driver experience evaluation model receives vehicle state data as input, which includes lateral tracking error , vehicle longitudinal speed , collision time between ego vehicle and front vehicle , vehicle front wheel corner ; further, a feature vector representing driver experience is constructed, with 4 dimensions, including safety, tracking accuracy, longitudinal smoothness and lateral stability.

[0155] Step 3.1.2, Driver experience reward function design, based on safety, tracking accuracy, longitudinal smoothness and lateral stability; wherein safety considers the collision time between ego vehicle and front vehicle, tracking accuracy considers lateral tracking error, longitudinal smoothness considers longitudinal speed change rate, and lateral stability considers vehicle front wheel corner change rate, as shown in summary formulas (6) to (11).

[0156] Step 3.1.3, Inverse reinforcement learning driver experience evaluation model training process, the process includes the following sub-steps:

[0157] Step 3.1.3.1, Expert dataset construction, based on the public NGSIM dataset, the dataset is classified by K-means clustering, a total of three categories of novice drivers, experienced drivers and experienced drivers, among which the experienced driver dataset is selected to form the expert demonstration trajectory set.

[0158] Step 3.1.3.2, learning objective function design, through iterative optimization, make the learned reward function be able to explain the expert behavior, that is, under this reward function, the expected cumulative reward of the expert strategy should be significantly higher than that of other random strategies.

[0159] Step 3.1.3.3, inverse reinforcement learning driver experience evaluation model training process, set the number of iterations to 200 times; adopt the maximum entropy inverse reinforcement learning framework, adjust the weight parameters of the reward function through the iterative optimization algorithm, maximize the expected cumulative reward of the expert strategy under this function, and thus inversely deduce the optimal reward function model that can explain the expert driving preference and decision logic.

[0160] Step 3.1.3.4, output the optimal driver experience reward function, after training, the model obtains a set of optimal reward function weights , and then obtains the optimal driver experience reward function.

[0161] Step 3.1.3.5, calculate the real-time reward value, input the historical and current vehicle state data into the trained reward function, and calculate the driver experience reward value of the driver in the current period .

[0162] Step 3.2, controller trust degree design, as shown in content formula (12).

[0163] Step 4, weight allocation module design:

[0164] Based on fuzzy rules, this module takes the driver's trust in the controller and the controller's trust in the driver as input, and dynamically outputs the driver control weight.

[0165] Fuzzy rule design, the specific steps are as follows:

[0166] Step 4.1, confirm the input variables: the driver's trust in the controller and the controller's trust in the driver , the domain is .

[0167] Step 4.2, fuzzy processing: fuzzify and , and map them to the corresponding fuzzy sets .

[0168] Where the membership function is a Gaussian membership function, where the low, medium and high of the input variable correspond to formula (20), formula (21) and formula (22):

[0169] (20)

[0170] (21)

[0171] (22)

[0172] wherein is the input variable, is the standard deviation of the Gaussian function.

[0173] Step 4.3, fuzzy rule reasoning, reasoning according to the preset fuzzy rule, the fuzzy rule is nine in total, as shown in the first to the ninth of the invention content step 4.3.

[0174] Step 4.4, defuzzification, defuzzification by centroid method, converting fuzzy set to accurate driver control weight value , as shown in the invention content formula (13).

[0175] Step 4.5, output variable driver control weight , as the driver path tracking weight of the non-cooperative game module , as shown in the invention content formula (14).

[0176] Step 5, non-cooperative game module design, its process includes the following sub-steps:

[0177] Step 5.1, vehicle model establishment, for the kinematic model of the vehicle, described by the vehicle coordinate system; for the vehicle dynamics model, only the lateral and yaw motion of the vehicle is considered, while the influence of aerodynamics and suspension is ignored, and then a two-degree-of-freedom model is obtained; construct the vehicle kinematics and dynamics model, as shown in formula (23):

[0178] (23)

[0179] wherein, is the vehicle mass; and are the longitudinal and lateral velocities of the vehicle, respectively; is the yaw angular velocity of the vehicle; and are the lateral forces of the front and rear wheels, respectively; is the front wheel steering angle of the vehicle; and are the distances from the vehicle mass center to the front and rear axles, respectively; is the yaw moment of inertia around the vertical axis of the vehicle mass center; is the vehicle yaw angle; and are the lateral and longitudinal coordinates of the vehicle, respectively.

[0180] Step 5.2: Assuming that the tire force does not enter the nonlinear region during vehicle movement, that is, the magnitude of the tire force and the sideslip angle have a linear relationship, equation (23) is linearized as shown in equation (24):

[0181] (twenty four)

[0182] in, , These are the front and rear tire stiffness, respectively. This is the steering gear ratio.

[0183] Step 5.3: Transform equation (24) into state-space form, as shown in equation (25):

[0184] (25)

[0185] In the formula,

[0186] in, The state matrix; , and These are the state coefficient matrix, the control input coefficient matrix, and the control output coefficient matrix, respectively.

[0187] Step 5.4: Prediction model establishment. Construct a discrete driver-controller interaction model, as shown in the invention content formula (15).

[0188] Step 5.5, Objective function design: In order to reduce human-machine decision-making conflicts and improve driving safety in the human-machine co-driving mode, and at the same time ensure vehicle stability, the objective function is designed as shown in the invention content formula (16).

[0189] Step 5.6, Solving for Nash Equilibrium:

[0190] Step 5.6.1 First, define the error variables for the driver and the controller, as shown in the invention formula (17).

[0191] Step 5.6.2: Substitute formula (17) into formula (16) to obtain the objective functions for the driver and controller, as shown in formula (26):

[0192] (26)

[0193] in, and These are the control input coefficient matrices for the driver and the controller, respectively.

[0194] Step 5.6.3, for the non-cooperative game optimization control problem, Nash equilibrium can be solved; Nash equilibrium conditions, as shown in equation (27):

[0195] (27)

[0196] Wherein, and represent the driver and the controller all control set respectively; called a set of Nash equilibrium solution.

[0197] Step 5.6.4, based on the least square method and convex iteration method, the optimal control input sequence of the driver and the controller can be obtained; further, the first value of each sequence is selected as the optimal solution, as shown in the summary formula (18).

[0198] Step 6, finally, the final steering angle of the human-machine co-driving vehicle is executed , as shown in the summary formula (19).

[0199] In summary, the application proposes a collaborative steering control method considering human-machine trust degree, which realizes the fundamental change of human-machine co-driving system from passive conflict coordination to active trust collaboration by constructing a closed-loop bidirectional trust perception mechanism and a dynamic weight decision framework based on trust. On the one hand, the system uses LSTM-Transformer model to predict the steering intention of the driver with high precision, and evaluates the trust degree of the driver to the controller in real time through the difference between the predicted steering angle result of the MPC controller; on the other hand, the inverse reinforcement learning model is used to invert the reward function from the multi-dimensional driving behavior, and the trust degree of the controller to the driver experience is quantified. On this basis, the trust degrees of both sides are fused through fuzzy rules, and the control weight of the driver is dynamically allocated in real time, and the Nash equilibrium solution is solved by introducing the non-cooperative game theory, and the optimal steering angle of the comprehensive intention of both sides is generated, and the human-machine co-driving vehicle completes the collaborative steering action, which provides an innovative and feasible solution for the practicalization of human-machine collaborative driving technology.

Claims

1. A cooperative steering control method considering human-machine mutual trust, characterized in that: The method realizes the cooperative steering control of the man-machine co-driving vehicle through a closed-loop control system composed of an expected path module, a man-machine steering conflict evaluation module, a driver experience evaluation module, a weight distribution module, a non-cooperative game module and the man-machine co-driving vehicle; the expected path module is used to generate the expected paths of the driver and the controller; the man-machine steering conflict evaluation module receives the steering action of the driver, the vehicle state, the expected paths of the driver and the controller, and outputs the trust degree of the driver to the controller; the driver experience evaluation module receives the vehicle state and outputs the trust degree of the controller to the driver; the weight distribution module combines the trust degree of the driver to the controller and the trust degree of the controller to the driver, and outputs the driver control weight based on fuzzy rules; the non-cooperative game module obtains the optimal steering angle through the prediction model establishment, the objective function design and the Nash equilibrium solution based on the expected paths of the driver and the controller, the driver control weight and the vehicle state, so as to realize the cooperative steering control of the man-machine co-driving vehicle. The LSTM-Transformer driver angle prediction model in the human-machine steering conflict assessment module receives the driver's steering actions. Vehicle status Driver's desired route Output driver's predicted steering angle The human-machine steering conflict assessment module uses an MPC controller-based angle prediction model to receive vehicle status data. Controller expected path Output controller predicts rotation angle Ultimately, the turning angle is determined based on the driver's prediction. and controller predicts the turning angle Determine the driver's level of trust in the controller. ;in, For the driver to turn the corner, For the rate of change of the driver's steering angle, and These are the vehicle's longitudinal speed and lateral speed, respectively. and These are the lateral tracking error and the yaw angle error, respectively. Let yaw rate be the vehicle's angular velocity. For road curvature, For the longitudinal acceleration of the vehicle, and These are the longitudinal and lateral coordinates of the driver's desired path, respectively. The yaw angle is the driver's desired path. and These are the vertical and horizontal coordinates of the controller's desired path, respectively. The yaw angle of the controller's desired path. The current moment; The driver's trust in the controller The design is shown below: ， wherein, is a human-machine corner conflict sensitivity factor, is a human-machine corner conflict factor, is an acceptable conflict level value; the human-machine corner conflict factor is designed based on the absolute value of the difference between the driver corner and the controller corner, and the specific formula is as follows: ， wherein, and are the driver and controller predicted steering angles at the step predicted steering angle, is the current time, is the man-machine steering angle prediction horizon; The inverse reinforcement learning-based driver experience evaluation model in the driver experience evaluation module receives a vehicle state , outputs a driver experience reward value ; finally, the driver experience reward value is used to evaluate the trust degree of the controller to the driver ; wherein, is a lateral tracking error, is a vehicle longitudinal speed, is a collision time between the ego vehicle and the preceding vehicle, is a vehicle front wheel steering angle; The controller described The design is shown below: ， wherein, is a driver experience reward value sensitivity factor, is a driver experience reward value, is a driver experience reward value for a driver experienced expert data set. The driver experience reward function is designed based on safety, tracking accuracy, longitudinal smoothness and lateral stability; the safety considers the collision time of the ego vehicle and the preceding vehicle, the tracking accuracy considers the lateral tracking error, the longitudinal smoothness considers the longitudinal vehicle speed change rate, and the lateral stability considers the vehicle front wheel steering angle change rate, and is specifically shown in the following formula: ， ， ， ， ， in, For security reward function, For tracking accuracy reward function, The reward function is for vertical smoothness. For horizontal stability reward function, and The first The collision time between the vehicle and the vehicle in front in the frame and the safe collision time threshold. For the first Lateral tracking error of the frame, For the first The rate of change of longitudinal vehicle speed in the frame, For the first The rate of change of the vehicle's front wheel steering angle in the frame. For the current frame number, The weights of each reward function, Let be the standard deviation of each reward function, where , , Total number of sampling frames for driver experience assessment; The weight distribution module distributes weights based on fuzzy rules, which dynamically distribute control weights based on driver trust in the controller and controller trust in the driver , outputting driver control weights as non-cooperative game module driver path tracking weights ; Finally, the non-cooperative game module obtains the optimal steering angle through the prediction model establishment, the objective function design and the Nash equilibrium solution based on the expected paths of the driver and the controller, the driver control weight and the vehicle state, so as to realize the cooperative steering control of the man-machine co-driving vehicle.

Images