A driving force self-adaptive regulation method based on subjective driving behavior feature recognition

Abstract

The application provides a driving force self-adaptive regulation method based on subjective driving behavior feature recognition, which is based on physiological signals and vehicle signals of a driver in a simulated driving environment, adopts a K-means++ clustering recognition algorithm to calibrate the driving style into a conservative type, a normal type and an aggressive type, and predicts the subjective driving demand of the driver by considering the CNN-BILSTM-Attention method of the physiological signals and the distance between the front vehicle. A deep reinforcement learning DDPG vehicle longitudinal controller is built through the prediction result of the subjective driving behavior feature of the driver, and finally, experimental verification is carried out. The result shows that the deep reinforcement learning DDPG vehicle longitudinal controller based on the driving demand prediction can effectively track the driving demand prediction curve under the subjective driving behavior, and effectively improve the driving experience of the driver.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent vehicle dynamics control, and particularly relates to a driving force adaptive regulation method based on subjective driving behavior feature recognition. BACKGROUND

[0002] With the rapid development of intelligent driving technology, vehicle longitudinal control as one of the core functions of automatic driving system plays a key role in improving driving safety, comfort and energy efficiency. Traditional longitudinal control methods (such as PID control, model predictive control) mainly adjust vehicle acceleration through real-time sensor data (vehicle distance, relative speed), but still face significant challenges in dynamic traffic scenarios. For example, in complex working conditions such as congestion following or signal light intersection, control strategies that rely only on current state feedback are prone to response lag and frequent acceleration and deceleration, resulting in decreased ride comfort. In recent years, prediction-based intelligent control has become a research hotspot, combining vehicle dynamics model, traffic big data and machine learning algorithms to predict future driving demand by combining vehicle distance and physiological signals and generate optimal control instructions. Therefore, it is meaningful to design a driving force adaptive regulation method based on subjective driving behavior feature recognition. SUMMARY

[0003] In order to solve the above problems, the application overcomes the shortcomings of the existing methods and proposes a driving force adaptive regulation method based on subjective driving behavior feature recognition. The method collects physiological signals and driving data of drivers through driving simulation experiments, uses K-means++ clustering recognition algorithm to classify driving styles into conservative, normal and ordinary types, proposes a CNN-BiLSTM prediction model considering the fusion attention mechanism of vehicle distance and physiological signals, and finally designs a deep reinforcement learning vehicle longitudinal controller for driving demand prediction results of drivers with different driving styles.

[0004] In order to realize the above technical features, the purpose of the application is realized as follows: a driving force adaptive regulation method based on subjective driving behavior feature recognition, comprising the following steps:

[0005] Step 1: Build a single motor driven electric vehicle driver-in-the-loop system, let the experimental personnel wear physiological signal acquisition instruments to complete driving simulation experiments in turn, and collect physiological data and driving data of the experimental personnel during driving;

[0006] Step 2, the physiological data and driving data collected in step 1 are preprocessed, the driving feature indicators are preliminarily selected, the principal component analysis method is used for dimension reduction processing of the selected features, the principal components with eigenvalues greater than 1 are retained, the K-means++ clustering method is used for classification of the eigenvalues, and the feature parameters are clustered into three driving styles, namely conservative, ordinary and aggressive;

[0007] Step 3, the correlation between the driving feature parameters is calculated by Pearson correlation analysis, and the feature parameters with low correlation are selected as the input of the driving demand prediction model;

[0008] Step 4, a CNN+BiLSTM driving demand prediction model with attention mechanism is constructed, and the driving demand of the driver with different driving style is predicted with Ts as a time window;

[0009] Step 5, a longitudinal control strategy based on deep reinforcement learning algorithm DDPG is proposed for the predicted driving demand curve of the driver with different driving style.

[0010] Preferably, the driver-in-the-loop system in step 1 includes a real-time simulator and a host PC, wherein the real-time simulator and the host PC each run a Simulink model, the Speedgoat runs a vehicle model, and the host PC runs a driving scene model. Finally, the physiological signal and driving data are synchronously collected on the Speedgoat.

[0011] The Simulink model in the host PC receives the steering, throttle and brake signals from the driver, and the host PC side inputs the steering, throttle and brake signals and rolling resistance and wind resistance information to the real-time simulator through UDP communication. The real-time simulator calculates the vehicle state according to the information and sends it to the host PC. The physiological signal and driving data of the driver are synchronously collected.

[0012] Preferably, the selected driving data features in step 2 include the average value and standard deviation of the lateral and longitudinal speed, lateral and longitudinal acceleration, throttle pedal opening, brake pedal opening, front vehicle distance and steering wheel angle for subsequent driving style recognition.

[0013] Preferably, in step 3, the Pearson correlation coefficient between each feature is analyzed, and finally the throttle pedal opening, front vehicle distance, throttle pedal change rate, longitudinal acceleration, electrocardiogram and skin potential signal are selected as the input of the prediction model.

[0014] Pearson correlation coefficient formula:

[0015]

[0016] In the formula, r is the Pearson correlation coefficient, are the mean values of the two feature quantities respectively, x i , y i are the observation values corresponding to each vector respectively.

[0017] Preferably, the prediction model in step 4 is a CNN+BiLSTM with fusion attention mechanism considering the vehicle distance and physiological signals, wherein the CNN layer adopts two-dimensional convolution to avoid data overfitting, and at the same time adopts a max pooling layer to extract features in driving data and physiological signals; then, the spatial features processed by the CNN are input into the BiLSTM layer in time sequence to capture past and future time sequence signals through the bidirectional LSTM structure of the BiLSTM layer; finally, the time sequence information is calculated through the attention mechanism layer to obtain the importance scores at different time steps.

[0018] Preferably, the LSTM formula is:

[0019] f t = σ(W f (h t-1 , x t ) + b f ); (2)

[0020] i t = σ(W i (h t-1 , x t ) + b i ); (3)

[0021]

[0022]

[0023] o t = σ(W o (h t-1 , x t ) + b o ); (6)

[0024] h t = o t ⊙ tanh(c t ); (7)

[0025] In the formula, σ represents a sigmoid activation function, tanh represents a hyperbolic cosine function, W f , W i , W c , and W o are weight functions of the forget gate, the input gate, the candidate value, and the output gate respectively, b f , b i , b c , and b o are bias functions.are the corresponding bias terms, represents multiplication of each element, (h t-1 , t ) represents the hidden state h t-1 of the previous moment and the concatenation of the current moment x t .

[0026] Preferably, the longitudinal control system is built in Simulink in step 5, the vehicle model input is the accelerator pedal opening degree and brake pressure, and the output is the speed, acceleration and engine power; the dynamics control is performed using deep reinforcement learning DDPG, and in the driving process, the accelerator pedal opening degree reflects the driver's demand for driving force, and the force balance formula during vehicle driving is as follows:

[0027]

[0028] In the formula, T D is the vehicle driving torque; i o is the main reducer transmission ratio; i g is the transmission ratio; η is the transmission system efficiency; G is the vehicle self weight; f is the tire rolling resistance coefficient; C d is the air resistance coefficient; A is the windward area; u a is the vehicle speed; δ is the automobile rotational mass conversion coefficient; m is the vehicle mass; r is the wheel radius; is the vehicle acceleration;

[0029] The expected acceleration is calculated by using the deep reinforcement learning regulation strategy, then the required driving force of the vehicle is calculated according to the expected acceleration, and finally the corresponding driving pedal opening degree α des is calculated through the throttle brake calibration table, and the motor expected torque is:

[0030]

[0031] By calculating the speed curve of the driver facing different driving demands, the speed required by different drivers reflects the different expectations of the driver for the torque, which is finally reflected in the demand for driving force;

[0032] The state space observation matrix Obs is designed as:

[0033] Obs=[V current ,V targent ,∑(V target -V current )];(10)

[0034] In the formula: V current is the current speed, V targent is the target speed, and ∑(V target -V current ) is the integral error term.

[0035] The penalty factor alpha is alpha = 1 x 10 3 ;

[0036] Alpha is a penalty factor, which will continue to receive a penalty when the vehicle speed does not reach the target speed, since the input speed curve is in line with the driver's expectations, and the comfort, power and response speed are all satisfied, so the reward function design only needs to consider whether it can completely follow the expected driving demand curve;

[0037] Design the reward function:

[0038] Reward = - alpha (V target -V current ) ; (11)

[0039] By designing the reward function, when the target speed is too different from the actual speed during the training process of the deep reinforcement learning algorithm, a large penalty term is given, which ensures that the control strategy can make the vehicle more in line with the subjective demand of the driver for driving force during driving.

[0040] The present application has the following beneficial effects:

[0041] 1. Most of the driving demand prediction only considers a single driving demand, without considering the relevance of the preceding vehicle distance to the driving demand in the real driving scene. At the same time, the driver as the driving subject is not considered, so the driver's skin and heart signals and the preceding vehicle distance are fused. Through the fusion of multi-modal signals, the subtle changes of driving demand can be better captured, and the differences in driving demand changes of different drivers can be reflected.

[0042] 2. The present application proposes a CNN-BiLSTM model fused with attention mechanism, the maximum pooling layer of the CNN layer extracts the features in the driving data and physiological signals, the BiLSTM layer captures the past and future time sequence signals through its bidirectional LSTM structure, and the time sequence information is calculated through the attention mechanism layer to obtain the importance score at different time steps. Compared with single CNN and BiLSTM, the prediction accuracy is improved.

[0043] 3. The present application is a driving force adaptive control strategy for subjective driving behavior feature recognition, which considers the personalized behavior of the driver, so that the designed longitudinal controller can provide personalized power demand for different drivers. The safety, maneuvering stability and comfort during vehicle driving are effectively improved. It has strong practicability and wide application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0044] The present application will be further described below in conjunction with the drawings and examples.

[0045] Figure 1 A longitudinal control method block diagram based on driving demand prediction is provided for the application.

[0046] Figure 2 A CNN-BiLSTM driving demand prediction model diagram combined with an attention mechanism is built for the application.

[0047] Figure 3 A K-means++ clustering result diagram.

[0048] Figure 4 A Pearson correlation coefficient heat map.

[0049] Figure 5 A comparison diagram of prediction results considering driver factors.

[0050] Figure 6 A vehicle speed prediction diagram corresponding to different models under a conservative driving style.

[0051] Figure 7 A vehicle speed prediction diagram corresponding to different models under a normal driving style.

[0052] Figure 8 A vehicle speed prediction diagram corresponding to different models under an aggressive driving style. DETAILED DESCRIPTION

[0053] The embodiments of the application will be further described below with reference to the accompanying drawings.

[0054] Example 1:

[0055] Referring to Figure 1 A driving force adaptive control method based on subjective driving behavior feature recognition includes the following steps:

[0056] Step 1: Build a single-motor-driven electric vehicle driver-in-the-loop system, and let the experimental personnel wear physiological signal acquisition instruments to complete driving simulation experiments one by one, and collect physiological data and driving data of the experimental personnel during driving;

[0057] Step 2: First, pre-process the physiological data and driving data collected in step 1, preliminarily select driving feature indicators, use principal component analysis method to reduce the dimension of the selected features, retain principal components with eigenvalues greater than 1, use K-means++ clustering method to classify eigenvalues, and cluster feature parameters into three driving styles, namely conservative, normal and aggressive;

[0058] Step 3: Calculate the correlation between driving feature parameters by Pearson correlation analysis, and select feature parameters with low correlation as inputs of the driving demand prediction model;

[0059] Step 4: Construct a CNN+BiLSTM driving demand prediction model that integrates attention mechanism, and predict driving demand in a 1-second time window for different driving styles.

[0060] Step 5: For the predicted driving demand curves of drivers with different driving styles, a longitudinal control strategy based on the deep reinforcement learning algorithm DDPG is proposed.

[0061] Furthermore, in step 1, the driver-in-the-loop system includes a real-time simulator and a host computer PC. The real-time simulator and the host computer PC each run a Simulink model, and Speedgoat runs a vehicle model. The host computer PC runs a driving scenario model, and finally, the synchronous acquisition of physiological signals and driving data is completed on Speedgoat.

[0062] The Simulink model in the host PC receives steering, throttle, and braking signals from the driver. The host PC then inputs the steering, throttle, and braking signals, as well as rolling resistance and aerodynamic resistance information, into the real-time simulator via UDP communication. The real-time simulator calculates the vehicle status based on the information and sends it to the host PC, synchronously collecting the driver's physiological signals and driving data.

[0063] Furthermore, the driving data features selected in step 2 include the average and standard deviation of lateral and longitudinal speeds, lateral and longitudinal accelerations, accelerator pedal opening, brake pedal opening, distance to the vehicle in front of the accelerator, and steering wheel angle, which are used for subsequent identification of driving style.

[0064] Furthermore, in step 3, by analyzing the Pearson correlation coefficients between various features, the accelerator pedal opening, distance to the vehicle in front, rate of change of accelerator pedal, longitudinal acceleration, as well as electrocardiogram and skin conductance signals are finally selected as the inputs for the prediction model.

[0065] Pearson correlation coefficient formula:

[0066]

[0067] In the formula: r is the Pearson correlation coefficient, These are the means of the two feature quantities, x and x. i y i These are the observation values ​​corresponding to each vector.

[0068] Further, the prediction model in step 4 is a CNN+BiLSTM with fusion attention mechanism considering the car distance and physiological signals, wherein the CNN layer adopts two-dimensional convolution to avoid data overfitting, and a max pooling layer is used to extract features in driving data and physiological signals; then, the spatial features processed by the CNN are input into the BiLSTM layer in time sequence to capture past and future time sequence signals through the bidirectional LSTM structure; finally, the time sequence information is calculated through the attention mechanism layer to obtain the importance score at different time steps.

[0069] Further, the LSTM formula is:

[0070] f t =σ(W f (h t-1 ,x t )+b f ); (2)

[0071] i t =σ(W i (h t-1 ,x t )+b i ); (3)

[0072]

[0073] o t =σ(W o (h t-1 ,x t )+b o ); (6)

[0074] h t =o t ⊙tanh(c t ); (7)

[0075] wherein σ represents a sigmoid activation function, tanh represents a hyperbolic cosine function, W f ,W i ,W c ,W o are weight functions of the forgetting gate, the input gate, the candidate value and the output gate respectively, b f ,b i ,b c ,b o are corresponding bias terms, and represents multiplication according to each element, (h t-1 ,x t ) represents the splicing of the hidden state h t-1 at the previous moment and x t at the current moment.

[0076] Further, the longitudinal control system is built in Simulink in step 5, the vehicle model input is the accelerator pedal opening and brake pressure, and the output is the speed, acceleration and engine power; the dynamics control is performed using deep reinforcement learning DDPG, and in the driving process, the accelerator pedal opening reflects the driver's demand for driving force, and the force balance formula during vehicle driving is as follows:

[0077]

[0078] wherein, T D is the vehicle driving torque; i o is the main reducer transmission ratio; i g is the transmission ratio; η is the transmission system efficiency; G is the vehicle self weight; f is the tire rolling resistance coefficient; C d is the air resistance coefficient; A is the windward area; u a is the vehicle speed; δ is the automobile rotational mass conversion coefficient; m is the vehicle mass; r is the wheel radius; is the vehicle acceleration;

[0079] The expected acceleration is calculated by using the deep reinforcement learning regulation strategy, then the required driving force of the vehicle is calculated according to the expected acceleration, and finally the corresponding driving pedal opening α des is calculated through the throttle brake calibration table, and the motor expected torque is:

[0080]

[0081] By calculating the speed curve of the driver facing different driving demands, the speed required by different drivers reflects the different expectations of the driver for the torque, which is finally reflected in the demand for driving force;

[0082] The state space observation matrix Obs is designed:

[0083] Obs=[V current ,V targent ,∑(V target -V current )]; (10)

[0084] wherein: V current is the current speed, V targent is the target speed, and ∑(V target -V current ) is the integral error term;

[0085] The penalty factor α is valued as α=1×10 3 ;

[0086] a is a penalty factor, which will be continuously punished when the vehicle speed does not reach the target speed, since the input speed curve is consistent with the driver's expectations, in comfort, power and response speed, so the reward function design only needs to consider whether it can completely follow the expected driving demand curve;

[0087] Design reward function:

[0088] Reward = -a (V target -V current ); (11)

[0089] By designing the reward function, when the target speed is too different from the actual speed during the training process of the deep reinforcement learning algorithm, a large penalty term is given to ensure that the control strategy can make the vehicle more consistent with the driver's subjective demand for driving force during driving.

Claims

1. A driving force self-adaptive regulation method based on subjective driving behavior feature recognition, characterized in that, The method comprises the following steps: Step 1, build a single motor drive-based electric vehicle driver-in-the-loop system, let the experimental personnel wear physiological signal acquisition instruments to complete driving simulation experiments in turn, and collect physiological data and driving data of the experimental personnel during driving; Step 2, the physiological data and driving data collected in step 1 are preprocessed, and driving feature indexes are preliminarily selected; principal component analysis is used for dimension reduction processing of the selected features, and principal components with eigenvalues greater than 1 are retained; K-means++ clustering method is used for classification of the eigenvalues, and the feature parameters are clustered into three driving styles, namely conservative type, ordinary type and aggressive type; Step 3, the correlation between the driving feature parameters is calculated by Pearson correlation analysis, and the feature parameters with low correlation are selected as the input of the driving demand prediction model; Step 4, a CNN+BiLSTM driving demand prediction model with fusion attention mechanism is constructed, and the driving demand is predicted in a time window of Ts for different style drivers; Step 5, a longitudinal control strategy based on a deep reinforcement learning algorithm DDPG is proposed for the predicted driving demand curve of drivers with different driving styles; In step 5, a longitudinal control system is built in Simulink, the vehicle model input is accelerator pedal opening and brake pressure, and the output is speed, acceleration and engine power; deep reinforcement learning DDPG is used for dynamic control, and in the driving process, the accelerator pedal opening reflects the driver's demand for driving force, and the force balance formula during vehicle driving is as follows: ；（8） wherein is the vehicle drive torque; is the main reduction gear ratio; is the transmission gear ratio; is the transmission system efficiency; G is the vehicle self-weight; is the tire rolling resistance coefficient; is the air resistance coefficient; is the wind area; is the vehicle speed; is the vehicle rotational mass conversion coefficient; is the vehicle mass; is the wheel radius; is the vehicle acceleration; The expected acceleration is calculated by using the deep reinforcement learning regulation strategy, the required driving force of the vehicle is calculated according to the expected acceleration, and the corresponding driving pedal opening degree is calculated through the throttle brake calibration table The expected torque of the motor is: ； （9） By calculating the speed curve of drivers facing different driving demands, the speed of different drivers reflects the different expectations of the driver for the torque, which is finally reflected in the demand for driving force; The state space observation matrix Obs is designed: Obs = [ , , ] ;​ （10） wherein: is the current speed, is the target speed, is the integral error term; penalty factor is set to ; It is a penalty factor that will continue to be penalized when the vehicle speed does not reach the target speed. Since the input speed curve is in line with the driver's expectations and meets the requirements in terms of comfort, power and response speed, the reward function design only needs to consider whether it can completely follow the expected driving demand curve. The reward function is designed: ；（11） Through the design of the reward function, when the target speed is too different from the actual speed during the training process of the deep reinforcement learning algorithm, a large penalty term is given, which ensures that the control strategy can make the vehicle more meet the subjective demand of the driver for driving force during driving.

2. The driving force adaptive control method based on subjective driving behavior feature recognition according to claim 1, characterized in that, The driver-in-the-loop system in step 1 includes a real-time simulator and a host computer PC, wherein the real-time simulator and the host computer PC each run a Simulink model, and the Speedgoat runs a vehicle model; the host computer PC runs a driving scene model, and finally the physiological signals and driving data are synchronously collected on the Speedgoat; The Simulink model in the host computer PC receives the steering, throttle and brake signals from the driver, and the host computer PC side inputs the steering, throttle and brake signals and rolling resistance and air resistance information to the real-time simulator through UDP communication, the real-time simulator calculates the vehicle state according to the information and sends it to the host computer PC, and the physiological signals and driving data of the driver are synchronously collected.

3. The driving force adaptive control method based on subjective driving behavior feature recognition according to claim 2, characterized in that, The selected driving data features in step 2 include the average value and standard deviation of the lateral and longitudinal speed, lateral and longitudinal acceleration, accelerator pedal opening, brake pedal opening, accelerator front vehicle distance and steering wheel angle, which are used for subsequent driving style recognition.

4. The driving force adaptive control method based on subjective driving behavior feature recognition according to claim 3, characterized in that, In step 3, the throttle pedal opening, the front vehicle distance, the throttle pedal change rate, the longitudinal acceleration, and the electrocardiogram signal and the galvanic skin signal are finally selected as the input of the prediction model by analyzing the Pearson correlation coefficient between each feature. Pearson correlation coefficient formula: ；（1） wherein: r is the Pearson correlation coefficient, , are the mean values of the two feature quantities, respectively, , are the observation values corresponding to each vector, respectively.

5. The driving force adaptive control method based on subjective driving behavior feature recognition according to claim 4, characterized in that, In step 4, the prediction model is a CNN+BiLSTM with a fusion attention mechanism considering the front vehicle distance and the physiological signal, wherein a two-dimensional convolution is used in the CNN layer to avoid data overfitting, and a maximum pooling layer is used to extract features in the driving data and the physiological signal; then, the spatial features processed by the CNN are input into the BiLSTM layer in a time sequence to capture past and future time sequence signals through the bidirectional LSTM structure of the BiLSTM layer; finally, the time sequence information is calculated through the attention mechanism layer to obtain the importance score at different time steps.

6. The driving force adaptive control method based on subjective driving behavior feature recognition according to claim 5, characterized in that, The LSTM formula is as follows: ；（2） ；（3） ；（4） ；（5） ；（6） ；（7） wherein denotes a sigmoid activation function, tanh denotes a hyperbolic tangent function, are weight functions for the forget gate, input gate, candidate value and output gate, respectively, are corresponding bias terms, respectively, denotes element-wise multiplication, denotes the hidden state of the previous time step and the concatenation of the hidden state of the previous time step and the current time step

Images