Intelligent vehicle rear wheel turning angle calculation method considering driving style

Abstract

The invention discloses an intelligent vehicle rear wheel turning angle calculation method considering a driving style. A driving information acquisition module, a driving style judgment module, a loss function calculation module and a rear wheel turning angle calculation module are included. The driving information acquisition module is used for acquiring vehicle driving state information and physiological signal information of a driver; the driving style judgment module is used for calculating a driving style factor of the driver; the loss function calculation module is used for calculating a loss function of a neural network in the rear wheel steering angle calculation module; the rear wheel rotation angle calculation module is used for calculating the ratio of the rear wheel rotation angle to the front wheel rotation angle according to the deep neural network.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent vehicle control and safe driving, relates to intelligent vehicle steering control technology, and specifically relates to a method for calculating the rear wheel steering angle of an intelligent vehicle taking driving style into consideration. Background Art

[0002] With the development of intelligent driving technology, intelligent vehicles have gradually become an important direction for future travel. One of the core technologies of intelligent vehicles is precise vehicle dynamic control, and rear wheel steering angle control has received widespread attention as an important means to improve vehicle handling stability and safety. Existing steering control systems are usually based on fixed parameters or preset models, and the optimal steering angle is calculated using vehicle dynamics formulas. These methods do not take the driver's driving style into consideration, resulting in large differences in control strategies between different drivers, and control hysteresis or over-response problems under different driving styles, which seriously affects the driving experience and driving safety.

[0003] At present, although some studies have attempted to assess the driver's fatigue and tension by sensing the driver's physiological signals, most of the studies focus on the classification and recognition of driving styles, lack of deep integration with rear wheel angle calculation, and difficulty in achieving dynamic adjustment. Therefore, there is an urgent need for a rear wheel angle calculation method that takes driving style into consideration, so that it can adapt to the control habits of different drivers and optimize the dynamic control performance of intelligent vehicles. Summary of the invention

[0004] The purpose of the present invention is to provide a method for calculating the rear wheel turning angle of an intelligent vehicle taking driving style into consideration, so as to solve the problems faced in the above-mentioned background technology.

[0005] In order to achieve the above object, the present invention provides the following solutions:

[0006] The driving information acquisition module is used to obtain vehicle driving state information and driver's physiological signal information, including the vehicle's longitudinal speed V x 、Vehicle lateral acceleration a y , front wheel turning angle δ f , vehicle roll angle φ, front axle slip angle α f 、Rear axle slip angle α r 、Vehicle yaw angular velocity ω r , driver's eye movement frequency f eye , pupil diameter d pupil , heart rate variability HRV, skin electrode activity EDA, electromyographic signal EMG, head angle θ head ;

[0007] The driving style judgment module is based on the driving behavior factor and biodata factors Calculate the driver's driving style factor S d , the driver style factor refers to the driver's driving proficiency, where when When When the driver is defined as an ordinary driver, When the driver is defined as a skilled driver, the specific calculation formula is:

[0008]

[0009] In the formula, is the driving behavior factor, is the biological data factor;

[0010] The loss function calculation module is based on the driver's operating burden loss function L busy , Vehicle rollover risk loss function L roll 、Vehicle skidding risk loss function L slip , physical loss function L physics Get the total loss function L of the rear wheel angle calculation module total , the calculation formula is as follows:

[0011]

[0012] Where ξ1, ξ2, ξ3 are the loss functions L busy , L roll , L slip The weighting factor of is calculated by the loss function weighting factor attention mechanism, and the value range is 0 to 1; physicis is the weighting factor of the physical loss function, when λ physicis When λ approaches 0, the proportion of the physical loss function increases, improving the learning ability of physical laws. physicis When it approaches 1, the proportion of the physical loss function decreases, reducing the ability to learn physical laws; L busy is the driver's control burden loss function, L roll is the vehicle rollover risk loss function, L slip is the vehicle skidding risk loss function, L physics is the physical loss function;

[0013] The rear wheel angle calculation module calculates the total loss function L according to the calculation total , a deep neural network algorithm is used to determine the ratio R of the rear wheel angle to the front wheel angle of the vehicle, where when R < 0, the front and rear wheels turn in the same direction, and when R > 0, the front and rear wheels turn in opposite directions. The calculation formula is as follows:

[0014] δr =δ f ·R

[0015] In the formula, δ f is the front wheel turning angle, R is the ratio of the rear wheel turning angle to the front wheel turning angle;

[0016] The input layer contains 5 nodes, corresponding to the longitudinal speed V of the vehicle x 、Vehicle lateral acceleration a y , yaw angular velocity w r , center of mass sideslip angle β, front wheel turning angle δ f The hidden layer adopts a three-layer structure, including a first hidden layer, a second hidden layer, and a third hidden layer; the first hidden layer contains 64 neurons, the second hidden layer contains 32 neurons, and the third hidden layer contains 16 neurons; the output layer contains 1 node, and the output is the ratio R of the rear wheel and front wheel steering angles of the vehicle;

[0017] The calculation formula of the rear wheel turning angle calculation module is:

[0018]

[0019] Where V x is the longitudinal speed of the vehicle, a y Vehicle lateral acceleration, w r is the yaw rate, β is the sideslip angle of the center of mass, δ f is the front wheel turning angle; h1 is the output of the first hidden layer, h2 is the output of the second hidden layer; Swish is the activation of the first hidden layer; ReLU is the activation function of the second hidden layer and the third hidden layer; W1, W2, W3, W4 are weight matrices, all determined through training; b1, b2, b3, b4 are bias terms, all determined through training, W s is the weight matrix from the input layer to the dynamic scaling factor S, b s is the bias term from the input layer to the dynamic scaling factor S; S is the dynamic scaling factor, τ is the scaling weight control factor, and its value range is 0 to 1; Tanh is the output layer activation function; Sigmoid is the dynamic scaling factor layer activation function.

[0020]

[0021] Where, L busy is the driver's control burden loss function, L roll is the vehicle rollover risk loss function, L slip is the vehicle skidding risk loss function; k1 is the adjustment factor of the driver’s handling burden loss function, k2 is the adjustment factor of the vehicle rollover risk loss function, and k3 is the adjustment factor of the vehicle skidding risk loss function, which are used to control the sensitivity of each loss item. The specific value is selected according to specific needs; is the normalized biological signal feature value, is the eigenvalue calculated using exponentially weighted moving average, The eigenvalues ​​obtained by using Sigmoid nonlinear transformation are used to enhance the stability of biological data factors.

[0022] The driving style judgment module, the normalized biological signal characteristic value Eigenvalues ​​after exponentially weighted moving average calculation Eigenvalues ​​obtained using Sigmoid nonlinear transformation The specific calculation process is as follows:

[0023] First, for different physiological signals, core features are extracted. The specific calculation process is as follows:

[0024] Eye movement frequency f eye The core feature of is the instantaneous rate of change of the driver's eye movement frequency Δf eye 、Information entropy H eye , when the driver's driving behavior is proficient, the information entropy H eye , instantaneous rate of change Δf eye The value of will be smaller when the driver's driving behavior is unfamiliar. eye , instantaneous rate of change Δf eye The value of will be larger; the calculation formula is as follows:

[0025]

[0026] p i =|FFT(f eye )| 2

[0027] H eye =-∑p i logp i

[0028] Where f eye is the driver's eye movement frequency, FFT(x) is Fourier transform, where x = f eye 、p i is the eye movement frequency f eye The power spectral density of

[0029] Pupil diameter d pupil The core feature of is the coefficient of variation CV of pupil diameter pupil , the energy of the high-frequency component of the pupil diameter signal D pupil ; The pupil diameter of experienced drivers is more stable, and the high-frequency component energy D of the pupil diameter signal in a short period of time pupil , pupil diameter d pupilThe fluctuation is small, the pupil diameter of novice drivers changes more dramatically, and the high-frequency component energy D of the pupil diameter signal in a short period of time pupil , pupil diameter d pupil The fluctuation is large, and the calculation formula is as follows:

[0030]

[0031] Where σ pupil is the standard deviation of pupil diameter, μ pupil is the mean pupil diameter, d pupil is the pupil diameter; FFT(x) is Fourier transform, where x = d pupil 、f c is the lower limit of the high-frequency interval obtained by Fourier transform, f max The upper limit of the high-frequency range obtained by Fourier transform, the specific value is determined through experiments;

[0032] The standard deviation of pupil diameter σ pupil and the mean pupil diameter μ pupil The calculation formula is as follows:

[0033]

[0034] Where d pupil,i is the pupil diameter at the i-th moment, N is the number of pupil diameter samples collected;

[0035] The core feature of heart rate variability (HRV) is the ratio of low-frequency to high-frequency power (HRV). LF / HF , low frequency high frequency power ratio HRV LF / HF When it is higher, the driver's driving behavior is unfamiliar, and the low-frequency and high-frequency power is higher than HRV. LF / HF When it is lower, the driver's driving behavior is more skilled, and the calculation formula is as follows:

[0036]

[0037] Where P LF is the low-frequency power of heart rate variability, P HF is the high frequency power of heart rate variability, HRV is the heart rate variability;

[0038] The core feature of skin electrical activity (EDA) is the skin electrical energy attenuation factor (EDA). decay , skin electrical energy attenuation factor EDA decay When it is larger, the driver is in a state of high tension and high pressure, driving behavior is relatively unfamiliar, and the skin electrical energy attenuation factor EDA decay When it is small, the driver is in a relaxed state and has more skilled driving behavior. The calculation formula is as follows:

[0039]

[0040] Where EDA(t) is the skin electrical signal value at time t, EDA(t-1) is the skin electrical signal value at time t-1, and λ EDA is the exponential attenuation factor of skin electrical activity, which is used to amplify the drastic changes in skin electrical activity in a short period of time; T is the current sampling time;

[0041] The core feature of the electromyographic signal EMG is the electromyographic signal adaptive power exponential root mean square value EMG adaptive , EMG adaptive power exponential root mean square value adaptive When it is larger, the driver is in a state of high tension and high pressure, driving behavior is relatively unfamiliar, and the root mean square value of the electromyographic signal adaptive power exponent EMG adaptive When it is small, the driver is in a relaxed state and has more skilled driving behavior. The calculation formula is as follows:

[0042]

[0043] Where, EMG t is the skin electrical signal value at time t, z is the adaptive power index of the electromyographic signal, σ EMG is the standard deviation of the electromyographic signal, N is the number of electromyographic signal samples collected;

[0044] Head angle θ head The characteristic value of the head angle change rate θ stability , head angle change rate θ stability When it is larger, the driver is in a state of high tension and high pressure, driving behavior is relatively unfamiliar, and the head angle change rate θ stability When it is small, the driver is in a relaxed state and has more skilled driving behavior. The calculation formula is as follows:

[0045]

[0046] In the formula, θ head (t) is the head angle at time t, μ θ (t) is the sliding mean at time t, ζ1 is the sliding window size, σ θ (t) is the sliding standard deviation at time t, μ θ (t) is the mean value of the driver’s head angle;

[0047] Normalized biological signal feature value Eigenvalues ​​after exponentially weighted moving average calculation Sigmoid nonlinear transformation eigenvalue The instantaneous rate of change of eye movement frequency Δf eye , eye movement frequency feye The information entropy H eye , Coefficient of variation of pupil diameter CV pupil , the energy of the high-frequency component of the pupil diameter signal D pupil , heart rate variability low frequency high frequency power ratio HRV LF / HF , Skin electrical energy attenuation factor EDA decay 、Head angle change rate θ stability Calculated, where X i =[Δf eye ,H eye ,CV pupil ,D pupil ,HRV LF / HF ,EDA decay ,θ stability ], the specific calculation process is as follows:

[0048] Normalized biological signal feature value The biological signal feature value X i Mapped to [0,1] to avoid dimension problems and ensure comparability between different features, the calculation formula is as follows:

[0049]

[0050] In the formula, max(X i ) is the biological signal characteristic value X i The maximum value, min(X i ) is the biological signal characteristic value X i The minimum value of

[0051] Eigenvalues ​​after exponentially weighted moving average calculation Exponential weighted moving average is used for smoothing to reduce the impact of noise on the eigenvalue X i To avoid the impact of drastic fluctuations in a short period of time and cause wrong judgment, the calculation formula is as follows:

[0052]

[0053] In the formula, ε is the smoothing coefficient, which controls the balance between new data and historical data. When ε is close to 1, it is more dependent on current data and has a faster response speed. When ε is close to 0, it is more dependent on historical data and has a higher degree of smoothing.

[0054] Sigmoid nonlinear transformation eigenvalue Sigmoid nonlinear transformation is used to enhance the stability of data and avoid eigenvalue X i The impact caused by being too large or too small is uneven. The calculation formula is as follows:

[0055]

[0056] In the formula, is the normalized biological signal feature value.

[0057] The loss function calculation module calculates the driver's operating burden loss function L by the following formula: busy , Vehicle rollover risk loss function L roll 、Vehicle skidding risk loss function L slip And the physical loss function L physics :

[0058] Driver control burden loss function L busy The frequency and amplitude of the driver's steering wheel operation are measured to assess the driver's driving burden. The calculation formula is as follows:

[0059]

[0060] In the formula, δ f [i] is the steering wheel angle at time i, Δt is the time step; ψ i is a weight factor used to measure the importance of different data nodes. The specific value is obtained by experiment and ranges from 0 to 1; busy is the driver burden loss function weight factor, which is used to measure the weight of the driver's steering wheel operation frequency and amplitude. The specific value is obtained by experiment and ranges from 0 to 1. N is the total number of data samples. Softplus is a soft regularization function that makes the loss function smooth and differentiable.

[0061] Vehicle rollover risk loss function L roll Measure the lateral acceleration a y As well as the over-limit situation of the vehicle's roll angle φ, the rollover risk of the vehicle is assessed. The specific calculation formula is as follows:

[0062]

[0063] In the formula, a y [i] is the lateral acceleration at time i, φ[i] is the vehicle roll angle at time i, a y,max is the maximum value of the vehicle's lateral acceleration, φ max is the maximum value of the vehicle roll angle, a y,th is the lateral acceleration threshold of the vehicle, φ th is the vehicle roll angle threshold; ψ i is a weight factor used to measure the importance of different data nodes. The specific value is obtained by experiment and ranges from 0 to 1; roll,1 , roll,2 is the weight factor of the vehicle rollover risk loss function, which is used to adjust the vehicle rollover risk loss function Lroll The weight of each item in the equation ranges from 0 to 1; N is the total number of data samples; Softplus is a soft regularization function that makes the loss function smooth and differentiable;

[0064] Vehicle skidding risk loss function L slip Measure the front and rear wheel slip angle α f and α r The over-limit situation is used to evaluate the side slip risk of the vehicle. The specific calculation formula is as follows:

[0065]

[0066] In the formula, α f [i] is the front wheel slip angle at time i, α r [i] is the rear wheel slip angle at time i, is the linear interval threshold of the front wheel slip angle, is the linear interval threshold of the rear wheel slip angle, α max is the linear interval threshold of tire slip angle; ψ i is a weight factor used to measure the importance of different data nodes. The specific value is obtained by experiment and ranges from 0 to 1; slip,1 and λ slip,2 is the vehicle skidding risk loss function L slip Weight factor, used to adjust the vehicle skidding risk loss function L slip The weight of each item ranges from 0 to 1; N is the total number of data samples; Softplus is a soft regularization function that makes the loss function smooth and differentiable;

[0067] Physical loss function L physics Measure the ratio of the front and rear wheel angles predicted by the neural network x NN and x model The difference between them ensures the basic physical law of the composite mechanism model. The specific calculation formula is as follows:

[0068]

[0069] In the formula, x NN [i] is the predicted value at time i obtained by the neural network, x model [i] is the ideal value of the front and rear wheel steering angle ratio at time i calculated by the mechanism model, and N is the total number of data samples.

[0070] The loss function calculation module calculates the loss function weighting factor w according to the attention mechanism i , the specific calculation process is as follows:

[0071] First, the driver's driving style factor S dPerform nonlinear mapping to obtain the driving style factor S′ after nonlinear mapping d and high-order driving style factors The specific calculation formula is as follows:

[0072]

[0073]

[0074] In the formula, is the weight matrix, all obtained through training; is the bias term, all of which are obtained through training; ReLU is the activation function of the first hidden layer, SiLU is the activation function of the second hidden layer, and tanh is the activation function of the third hidden layer; S d is the driver's driving style factor;

[0075] Then the temperature coefficient τ is calculated. The specific calculation formula is as follows:

[0076]

[0077] Where λ1, λ2, and λ3 are weighting coefficients, whose specific values ​​are obtained by experiments and range from 0 to 1; W τ1 , W τ2 is the weight matrix, obtained through training; b τ1 , b τ2 is a bias term, all obtained through training; SiLU and tanh are activation functions; S′ d is the driving style factor obtained after nonlinear mapping, S d is the driver's driving style factor, It is a high-level driving style factor;

[0078] Then the attention weight The specific calculation formula is as follows:

[0079]

[0080] In the formula, S′ d is the driving style factor obtained after nonlinear mapping, ξ initial is the initial weighting factor of the loss function, and its specific value is obtained according to the experiment; Q is the query vector of driving style, which is used to calculate the matching degree between the current driver style and the driving behavior factor; K is the driving behavior characteristic factor, which provides the information of historical driving behavior; V is the weighting factor value vector of the loss function, which provides the basic weight for dynamic adjustment; W Q , W K , W V , is the weight matrix, and its specific value is obtained according to the experiment. Q , b K , b V , is the bias term, and its specific value is obtained according to the experiment. is the attention weight, d k is the normalization factor, τ is the temperature coefficient to prevent the gradient from disappearing; SiLU is the activation function, and Softmax is the normalization function;

[0081] From the calculated attention weights, the final loss weighting factor ξ can be calculated i , the calculation formula is as follows:

[0082]

[0083] In the formula, is the attention weight and V is the loss function weighting factor value vector.

[0084] The beneficial effects of the present invention are:

[0085] The present invention uses a method combining a neural network and an attention mechanism to comprehensively analyze the vehicle's driving state and the driver's physiological signal data, and realizes the calculation of the rear wheel angle of an intelligent vehicle taking into account the driving style, thereby improving the stability of vehicle control and driving safety. By calculating the driving style factor, the method can classify the driver's proficiency, making the rear wheel angle adjustment more personalized. In addition, the present invention introduces a variety of loss functions, including the driver's operating burden, the risk of vehicle rollover, the risk of vehicle sideslip and physical loss, and uses an attention mechanism to dynamically adjust the loss weight, thereby optimizing the rear wheel angle calculation under different driving styles and driving conditions, and ensuring the safety and comfort of the vehicle in a complex driving environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] The present invention will be further described below in conjunction with the accompanying drawings:

[0087] Figure 1 This is a framework diagram of a method for calculating the rear wheel turning angle of an intelligent vehicle taking driving style into consideration proposed by the present invention. DETAILED DESCRIPTION

[0088] The present invention is further described in detail below based on the accompanying drawings and specific embodiments.

[0089] See also Figure 1 , a method for calculating the rear wheel turning angle of an intelligent vehicle considering driving style described in the present invention comprises a driving information collection module, a driving style judgment module, a loss function calculation module, and a rear wheel turning angle calculation module;

[0090] The driving information acquisition module is used to obtain vehicle driving state information and driver's physiological signal information, including the vehicle's longitudinal speed V x 、Vehicle lateral acceleration a y , front wheel turning angle δ f , vehicle roll angle φ, front axle slip angle α f 、Rear axle slip angle α r 、Vehicle yaw angular velocity ω r , driver's eye movement frequency f eye , pupil diameter d pupil , heart rate variability HRV, skin electrode activity EDA, electromyographic signal EMG, head angle θ head ;

[0091] The driving style judgment module is based on the driving behavior factor and biodata factors Calculate the driver's driving style factor S d , the driver style factor refers to the driver's driving proficiency, where when When When the driver is defined as an ordinary driver, When the driver is defined as a skilled driver, the specific calculation formula is:

[0092]

[0093] In the formula, is the driving behavior factor, is the biological data factor;

[0094] The loss function calculation module is based on the driver's operating burden loss function L busy , Vehicle rollover risk loss function L roll 、Vehicle skidding risk loss function L slip , physical loss function L physics Get the total loss function L of the rear wheel angle calculation module total , the calculation formula is as follows:

[0095]

[0096] Where ξ1, ξ2, ξ3 are the loss functions L busy , L roll , L slip The weighting factor of is calculated by the loss function weighting factor attention mechanism, and the value range is 0 to 1; physicis is the weighting factor of the physical loss function, when λ physicisWhen λ approaches 0, the proportion of the physical loss function increases, improving the learning ability of physical laws. physicis When it approaches 1, the proportion of the physical loss function decreases, reducing the ability to learn physical laws; L busy is the driver's control burden loss function, L roll is the vehicle rollover risk loss function, L slip is the vehicle skidding risk loss function, L physics is the physical loss function;

[0097] The rear wheel angle calculation module calculates the total loss function L according to the calculation total , a deep neural network algorithm is used to determine the ratio R of the rear wheel angle to the front wheel angle of the vehicle, where when R < 0, the front and rear wheels turn in the same direction, and when R > 0, the front and rear wheels turn in opposite directions. The calculation formula is as follows:

[0098] δ r =δ f ·R

[0099] In the formula, δ f is the front wheel turning angle, R is the ratio of the rear wheel turning angle to the front wheel turning angle;

[0100] The input layer contains 5 nodes, corresponding to the longitudinal speed V of the vehicle x 、Vehicle lateral acceleration a y , yaw angular velocity w r , center of mass sideslip angle β, front wheel turning angle δ f The hidden layer adopts a three-layer structure, including a first hidden layer, a second hidden layer, and a third hidden layer; the first hidden layer contains 64 neurons, the second hidden layer contains 32 neurons, and the third hidden layer contains 16 neurons; the output layer contains 1 node, and the output is the ratio R of the rear wheel and front wheel steering angles of the vehicle;

[0101] The calculation formula of the rear wheel turning angle calculation module is:

[0102]

[0103] Where V x is the longitudinal speed of the vehicle, a y Vehicle lateral acceleration, w r is the yaw rate, β is the sideslip angle of the center of mass, δ f is the front wheel turning angle; h1 is the output of the first hidden layer, h2 is the output of the second hidden layer; Swish is the activation of the first hidden layer; ReLU is the activation function of the second hidden layer and the third hidden layer; W1, W2, W3, W4 are weight matrices, all determined through training; b1, b2, b3, b4 are bias terms, all determined through training, W sis the weight matrix from the input layer to the dynamic scaling factor S, b s is the bias term from the input layer to the dynamic scaling factor S; S is the dynamic scaling factor, τ is the scaling weight control factor, and its value range is 0 to 1; tanh is the output layer activation function; Sigmoid is the dynamic scaling factor layer activation function.

[0104]

[0105] Where, L busy is the driver's control burden loss function, L roll is the vehicle rollover risk loss function, L slip is the vehicle skidding risk loss function; k1 is the adjustment factor of the driver’s handling burden loss function, k2 is the adjustment factor of the vehicle rollover risk loss function, and k3 is the adjustment factor of the vehicle skidding risk loss function, which are used to control the sensitivity of each loss item. The specific value is selected according to specific needs; is the normalized biological signal feature value, is the eigenvalue calculated using exponentially weighted moving average, The eigenvalues ​​obtained by using Sigmoid nonlinear transformation are used to enhance the stability of biological data factors.

[0106] The driving style judgment module, the normalized biological signal characteristic value Eigenvalues ​​after exponentially weighted moving average calculation Eigenvalues ​​obtained using Sigmoid nonlinear transformation The specific calculation process is as follows:

[0107] First, for different physiological signals, core features are extracted. The specific calculation process is as follows:

[0108] Eye movement frequency f eye The core feature of is the instantaneous rate of change of the driver's eye movement frequency Δf eye 、Information entropy H eye , when the driver's driving behavior is proficient, the information entropy H eye , instantaneous rate of change Δf eye The value of will be smaller when the driver's driving behavior is unfamiliar. eye , instantaneous rate of change Δf eye The value of will be larger; the calculation formula is as follows:

[0109]

[0110] p i =|FFT(f eye )| 2

[0111] Heye =-∑p i logp i

[0112] Where f eye is the driver's eye movement frequency, FFT(x) is Fourier transform, where x = f eye 、p i is the eye movement frequency f eye The power spectral density of

[0113] Pupil diameter d pupil The core feature of is the coefficient of variation CV of pupil diameter pupil , the energy of the high-frequency component of the pupil diameter signal D pupil ; The pupil diameter of experienced drivers is more stable, and the high-frequency component energy D of the pupil diameter signal in a short period of time pupil , pupil diameter d pupil The fluctuation is small, the pupil diameter of novice drivers changes more dramatically, and the high-frequency component energy D of the pupil diameter signal in a short period of time pupil , pupil diameter d pupil The fluctuation is large, and the calculation formula is as follows:

[0114]

[0115] Where σ pupil is the standard deviation of pupil diameter, μ pupil is the mean pupil diameter, d pupil is the pupil diameter; FFT(x) is Fourier transform, where x = d pupil 、f c is the lower limit of the high-frequency interval obtained by Fourier transform, f max The upper limit of the high-frequency range obtained by Fourier transform, the specific value is determined through experiments;

[0116] The standard deviation of pupil diameter σ pupil and the mean pupil diameter μ pupil The calculation formula is as follows:

[0117]

[0118]

[0119] Where d pupil,i is the pupil diameter at the i-th moment, N is the number of pupil diameter samples collected;

[0120] The core feature of heart rate variability (HRV) is the ratio of low-frequency to high-frequency power (HRV). LF / HF , low frequency high frequency power ratio HRV LF / HF When it is higher, the driver's driving behavior is unfamiliar, and the low-frequency and high-frequency power is higher than HRV.LF / HF When it is lower, the driver's driving behavior is more skilled, and the calculation formula is as follows:

[0121]

[0122] Where P LF is the low-frequency power of heart rate variability, P HF is the high frequency power of heart rate variability, HRV is the heart rate variability;

[0123] The core feature of skin electrical activity (EDA) is the skin electrical energy attenuation factor (EDA). decay , skin electrical energy attenuation factor EDA decay When it is larger, the driver is in a state of high tension and high pressure, driving behavior is relatively unfamiliar, and the skin electrical energy attenuation factor EDA decay When it is small, the driver is in a relaxed state and has more skilled driving behavior. The calculation formula is as follows:

[0124]

[0125] Where EDA(t) is the skin electrical signal value at time t, EDA(t-1) is the skin electrical signal value at time t-1, and λ EDA is the exponential attenuation factor of skin electrical activity, which is used to amplify the drastic changes in skin electrical activity in a short period of time; T is the current sampling time;

[0126] The core feature of the electromyographic signal EMG is the electromyographic signal adaptive power exponential root mean square value EMG adaptive , EMG adaptive power exponential root mean square value adaptive When it is larger, the driver is in a state of high tension and high pressure, driving behavior is relatively unfamiliar, and the root mean square value of the electromyographic signal adaptive power exponent EMG adaptive When it is small, the driver is in a relaxed state and has more skilled driving behavior. The calculation formula is as follows:

[0127]

[0128] Where, EMG t is the skin electrical signal value at time t, z is the adaptive power index of the electromyographic signal, σ EMG is the standard deviation of the electromyographic signal, N is the number of electromyographic signal samples collected;

[0129] Head angle θ head The characteristic value of the head angle change rate θ stability , head angle change rate θ stability When it is larger, the driver is in a state of high tension and high pressure, driving behavior is relatively unfamiliar, and the head angle change rate θ stabilityWhen it is small, the driver is in a relaxed state and has more skilled driving behavior. The calculation formula is as follows:

[0130]

[0131] In the formula, θ head (t) is the head angle at time t, μ θ (t) is the sliding mean at time t, ζ1 is the sliding window size, σ θ (t) is the sliding standard deviation at time t, μ θ (t) is the mean value of the driver’s head angle;

[0132] Normalized biological signal feature value Eigenvalues ​​after exponentially weighted moving average calculation Sigmoid nonlinear transformation eigenvalue The instantaneous rate of change of eye movement frequency Δf eye , eye movement frequency f eye The information entropy H eye , Coefficient of variation of pupil diameter CV pupil , the energy of the high-frequency component of the pupil diameter signal D pupil , heart rate variability low frequency high frequency power ratio HRV LF / HF , Skin electrical energy attenuation factor EDA decay 、Head angle change rate θ stability Calculated, where X i =[Δf eye ,H eye ,CV pupil ,D pupil ,HRV LF / HF ,EDA decay ,θ stability ], the specific calculation process is as follows:

[0133] Normalized biological signal feature value The biological signal feature value X i Mapped to [0,1] to avoid dimension problems and ensure comparability between different features, the calculation formula is as follows:

[0134]

[0135] In the formula, max(X i ) is the biological signal characteristic value X i The maximum value, min(X i ) is the biological signal characteristic value X i The minimum value of

[0136] Eigenvalues ​​after exponentially weighted moving average calculation Exponential weighted moving average is used for smoothing to reduce the impact of noise on the eigenvalue X i To avoid the impact of drastic fluctuations in a short period of time and cause wrong judgment, the calculation formula is as follows:

[0137]

[0138] In the formula, ε is the smoothing coefficient, which controls the balance between new data and historical data. When ε is close to 1, it is more dependent on current data and has a faster response speed. When ε is close to 0, it is more dependent on historical data and has a higher degree of smoothing.

[0139] Sigmoid nonlinear transformation eigenvalue Sigmoid nonlinear transformation is used to enhance the stability of data and avoid eigenvalue X i The impact caused by being too large or too small is uneven. The calculation formula is as follows:

[0140]

[0141] In the formula, is the normalized biological signal feature value.

[0142] The loss function calculation module calculates the driver's operating burden loss function L by the following formula: busy , Vehicle rollover risk loss function L roll 、Vehicle skidding risk loss function L slip And the physical loss function L physics :

[0143] Driver control burden loss function L busy The frequency and amplitude of the driver's steering wheel operation are measured to assess the driver's driving burden. The calculation formula is as follows:

[0144]

[0145] In the formula, δ f [i] is the steering wheel angle at time i, Δt is the time step; ψ i is a weight factor used to measure the importance of different data nodes. The specific value is obtained by experiment and ranges from 0 to 1; busy is the driver burden loss function weight factor, which is used to measure the weight of the driver's steering wheel operation frequency and amplitude. The specific value is obtained by experiment and ranges from 0 to 1. N is the total number of data samples. Softplus is a soft regularization function that makes the loss function smooth and differentiable.

[0146] Vehicle rollover risk loss function L roll Measure the lateral acceleration a yAs well as the over-limit situation of the vehicle's roll angle φ, the rollover risk of the vehicle is assessed. The specific calculation formula is as follows:

[0147]

[0148] In the formula, a y [i] is the lateral acceleration at time i, φ[i] is the vehicle roll angle at time i, a y,max is the maximum value of the vehicle's lateral acceleration, φ max is the maximum value of the vehicle roll angle, a y,th is the lateral acceleration threshold of the vehicle, φ th is the vehicle roll angle threshold; ψ i is a weight factor used to measure the importance of different data nodes. The specific value is obtained by experiment and ranges from 0 to 1; roll,1 , roll,2 is the weight factor of the vehicle rollover risk loss function, which is used to adjust the vehicle rollover risk loss function L roll The weight of each item in the equation ranges from 0 to 1; N is the total number of data samples; Softplus is a soft regularization function that makes the loss function smooth and differentiable;

[0149] Vehicle skidding risk loss function L slip Measure the front and rear wheel slip angle α f and α r The over-limit situation is used to evaluate the side slip risk of the vehicle. The specific calculation formula is as follows:

[0150]

[0151] In the formula, α f [i] is the front wheel slip angle at time i, α r [i] is the rear wheel slip angle at time i, is the linear interval threshold of the front wheel slip angle, is the linear interval threshold of the rear wheel slip angle, α max is the linear interval threshold of tire slip angle; ψ i is a weight factor used to measure the importance of different data nodes. The specific value is obtained by experiment and ranges from 0 to 1; slip,1 and λ slip,2 is the vehicle skidding risk loss function L slip Weight factor, used to adjust the vehicle skidding risk loss function L slip The weight of each item ranges from 0 to 1; N is the total number of data samples; Softplus is a soft regularization function that makes the loss function smooth and differentiable;

[0152] Physical loss function L physicsMeasure the ratio of the front and rear wheel angles predicted by the neural network x NN and x model The difference between them ensures the basic physical law of the composite mechanism model. The specific calculation formula is as follows:

[0153]

[0154] In the formula, x NN [i] is the predicted value at time i obtained by the neural network, x model [i] is the ideal value of the front and rear wheel steering angle ratio at time i calculated by the mechanism model, and N is the total number of data samples.

[0155] The loss function calculation module calculates the loss function weighting factor w according to the attention mechanism i , the specific calculation process is as follows:

[0156] First, the driver's driving style factor S d Perform nonlinear mapping to obtain the driving style factor S after nonlinear mapping d ′ and high-order transformation driving style factors The specific calculation formula is as follows:

[0157]

[0158] In the formula, is the weight matrix, all obtained through training; is the bias term, all of which are obtained through training; ReLU is the activation function of the first hidden layer, SiLU is the activation function of the second hidden layer, and tanh is the activation function of the third hidden layer; S d is the driver's driving style factor;

[0159] Then the temperature coefficient τ is calculated. The specific calculation formula is as follows:

[0160]

[0161] Where λ1, λ2, and λ3 are weighting coefficients, whose specific values ​​are obtained by experiments and range from 0 to 1; W τ1 , W τ2 is the weight matrix, obtained through training; b τ1 , b τ2 is a bias term, all obtained through training; SiLU and tanh are activation functions; S d ′ is the driving style factor obtained after nonlinear mapping, S d is the driver's driving style factor, It is a high-level driving style factor;

[0162] Then the attention weight The specific calculation formula is as follows:

[0163]

[0164] In the formula, S d ′ is the driving style factor obtained after nonlinear mapping, ξ initial is the initial weighting factor of the loss function, and its specific value is obtained according to the experiment; Q is the query vector of driving style, which is used to calculate the matching degree between the current driver style and the driving behavior factor; K is the driving behavior characteristic factor, which provides the information of historical driving behavior; V is the weighting factor value vector of the loss function, which provides the basic weight for dynamic adjustment; W Q , W K , W V , is the weight matrix, and its specific value is obtained according to the experiment. Q , b K , b V , is the bias term, and its specific value is obtained according to the experiment. is the attention weight, d k is the normalization factor, τ is the temperature coefficient to prevent the gradient from disappearing; SiLU is the activation function, and Softmax is the normalization function;

[0165] From the calculated attention weights, the final loss weighting factor ξ can be calculated i , the calculation formula is as follows:

[0166]

[0167] In the formula, is the attention weight and V is the loss function weighting factor value vector.

Claims

1. A method for calculating the rear wheel turning angle of an intelligent vehicle considering driving style, characterized in that: It includes driving information collection module, driving style judgment module, loss function calculation module and rear wheel turning angle calculation module; The driving information acquisition module is used to obtain vehicle driving state information and driver's physiological signal information, including the vehicle's longitudinal speed V x 、Vehicle lateral acceleration a y , front wheel turning angle δ f , vehicle roll angle φ, front axle slip angle α f 、Rear axle slip angle α r 、Vehicle yaw angular velocity ω r , driver's eye movement frequency f eye , pupil diameter d pupil , heart rate variability HRV, skin electrode activity EDA, electromyographic signal EMG, head angle θ head ; The driving style judgment module is based on the driving behavior factor and biodata factors Calculate the driver's driving style factor S d , the driver style factor refers to the driver's driving proficiency, where when When When the driver is defined as an ordinary driver, When the driver is defined as a skilled driver, the specific calculation formula is: In the formula, is the driving behavior factor, is the biological data factor; The loss function calculation module is based on the driver's operating burden loss function L busy , vehicle rollover risk loss function L roll 、Vehicle skidding risk loss function L slip , physical loss function L physics Get the total loss function L of the rear wheel angle calculation module total , the calculation formula is as follows: Where ξ1, ξ2, ξ3 are the loss functions L busy , L roll , L slip The weighting factor of is calculated by the loss function weighting factor attention mechanism, and the value range is 0 to 1; physicis is the weighting factor of the physical loss function, when λ physicis When λ approaches 0, the proportion of the physical loss function increases, improving the learning ability of physical laws. physicis When it approaches 1, the proportion of the physical loss function decreases, reducing the ability to learn physical laws; L busy is the driver's control burden loss function, L roll is the vehicle rollover risk loss function, L slip is the vehicle skidding risk loss function, L physics is the physical loss function; The rear wheel angle calculation module calculates the total loss function L according to the calculation total , a deep neural network algorithm is used to determine the ratio R of the rear wheel angle to the front wheel angle of the vehicle, where when R < 0, the front and rear wheels turn in the same direction, and when R > 0, the front and rear wheels turn in opposite directions. The calculation formula is as follows: d r =d f ·R In the formula, δ f is the front wheel turning angle, R is the ratio of the rear wheel turning angle to the front wheel turning angle; The input layer contains 5 nodes, corresponding to the longitudinal speed V of the vehicle x 、Vehicle lateral acceleration a y , yaw angular velocity w r , center of mass sideslip angle β, front wheel turning angle δ f The hidden layer adopts a three-layer structure, including a first hidden layer, a second hidden layer, and a third hidden layer; the first hidden layer contains 64 neurons, the second hidden layer contains 32 neurons, and the third hidden layer contains 16 neurons; the output layer contains 1 node, and the output is the ratio R of the rear wheel and front wheel steering angles of the vehicle; The calculation formula of the rear wheel turning angle calculation module is: Where V x is the longitudinal speed of the vehicle, a y Vehicle lateral acceleration, w r is the yaw rate, β is the sideslip angle of the center of mass, δ f is the front wheel turning angle; h1 is the output of the first hidden layer, h2 is the output of the second hidden layer; Swish is the activation of the first hidden layer; ReLU is the activation function of the second hidden layer and the third hidden layer; W1, W2, W3, W4 are weight matrices, all determined through training; b1, b2, b3, b4 are bias terms, all determined through training, W s is the weight matrix from the input layer to the dynamic scaling factor S, b s is the bias term from the input layer to the dynamic scaling factor S; S is the dynamic scaling factor, τ is the scaling weight control factor, and its value range is 0 to 1; tanh is the output layer activation function; Sigmoid is the dynamic scaling factor layer activation function.

2. The method for calculating rear wheel turning angle of an intelligent vehicle considering driving style according to claim 1, characterized in that: The driving style judgment module calculates the driving behavior factor by the following formula: and biodata factors Where, L busy is the driver's control burden loss function, L roll is the vehicle rollover risk loss function, L slip is the vehicle skidding risk loss function; k1 is the adjustment factor of the driver’s handling burden loss function, k2 is the adjustment factor of the vehicle rollover risk loss function, and k3 is the adjustment factor of the vehicle skidding risk loss function, which are used to control the sensitivity of each loss item. The specific value is selected according to specific needs; is the normalized biological signal feature value, is the eigenvalue calculated using exponentially weighted moving average, The eigenvalues ​​obtained by using Sigmoid nonlinear transformation are used to enhance the stability of biological data factors.

3. The method for calculating rear wheel turning angle of an intelligent vehicle considering driving style according to claim 1, characterized in that: The driving style judgment module, the normalized biological signal characteristic value Eigenvalues ​​after exponentially weighted moving average calculation Eigenvalues ​​obtained using Sigmoid nonlinear transformation The specific calculation process is as follows: First, extract the core features for different physiological signals. The specific calculation process is as follows: Eye movement frequency f eye The core feature of is the instantaneous rate of change of the driver's eye movement frequency Δf eye 、Information entropy H eye , when the driver's driving behavior is proficient, the information entropy H eye , instantaneous rate of change Δf eye The value of will be smaller when the driver's driving behavior is unfamiliar. eye , instantaneous rate of change Δf eye The value of will be larger; the calculation formula is as follows: p i =|FFT(f eye )| 2 H eye =-∑p i log p i Where f eye is the driver's eye movement frequency, FFT(x) is Fourier transform, where x = f eye 、p i is the eye movement frequency f eye The power spectral density of Pupil diameter d pupil The core feature of is the coefficient of variation CV of pupil diameter pupil , pupil diameter signal high frequency component energy D pupil ; The pupil diameter of experienced drivers is more stable, and the high-frequency component energy D of the pupil diameter signal in a short period of time pupil , pupil diameter d pupil The fluctuation is small, the pupil diameter of novice drivers changes more dramatically, and the high-frequency component energy D of the pupil diameter signal in a short period of time pupil , pupil diameter d pupil The fluctuation is large, and the calculation formula is as follows: Where σ pupil is the standard deviation of pupil diameter, μ pupil is the mean pupil diameter, d pupil is the pupil diameter; FFT(x) is Fourier transform, where x = d pupil 、f c is the lower limit of the high-frequency interval obtained by Fourier transform, f max The upper limit of the high-frequency range obtained by Fourier transform, the specific value is determined through experiments; The standard deviation of pupil diameter σ pupil and the mean pupil diameter μ pupil The calculation formula is as follows: Where d pupil,i is the pupil diameter at the i-th moment, N is the number of pupil diameter samples collected; The core feature of heart rate variability (HRV) is the ratio of low-frequency to high-frequency power (HRV). LF / HF , low frequency high frequency power ratio HRV LF / HF When it is higher, the driver's driving behavior is relatively unfamiliar, and the low-frequency and high-frequency power is higher than HRV. LF / HF When it is lower, the driver's driving behavior is more skilled, and the calculation formula is as follows: Where P LF is the low-frequency power of heart rate variability, P HF is the high frequency power of heart rate variability, HRV is the heart rate variability; The core feature of skin electrical activity (EDA) is the skin electrical energy attenuation factor (EDA). decay , skin electrical energy attenuation factor EDA decay When it is larger, the driver is in a state of high tension and high pressure, driving behavior is relatively unfamiliar, and the skin electrical energy attenuation factor EDA decay When it is small, the driver is in a relaxed state and has relatively skilled driving behavior. The calculation formula is as follows: Where EDA(t) is the skin electrical signal value at time t, EDA(t-1) is the skin electrical signal value at time t-1, and λ EDA is the exponential attenuation factor of skin electrical activity, which is used to amplify the drastic changes in skin electrical activity in a short period of time; T is the current sampling time; The core feature of the electromyographic signal EMG is the electromyographic signal adaptive power exponential root mean square value EMG adaptive , EMG adaptive power exponential root mean square value adaptive When it is larger, the driver is in a state of high tension and high pressure, driving behavior is relatively unfamiliar, and the root mean square value of the electromyographic signal adaptive power exponent EMG adaptive When it is small, the driver is in a relaxed state and has relatively skilled driving behavior. The calculation formula is as follows: Where, EMG t is the skin electrical signal value at time t, z is the adaptive power index of the electromyographic signal, σ EMG is the standard deviation of the electromyographic signal, N is the number of electromyographic signal samples collected; Head angle θ head The characteristic value of the head angle change rate θ stability , head angle change rate θ stability When it is larger, the driver is in a state of high tension and high pressure, driving behavior is relatively unfamiliar, and the head angle change rate θ stability When it is small, the driver is in a relaxed state and has relatively skilled driving behavior. The calculation formula is as follows: In the formula, θ head (t) is the head angle at time t, μ θ (t) is the sliding mean at time t, ζ1 is the sliding window size, σ θ (t) is the sliding standard deviation at time t, μ θ (t) is the mean value of the driver’s head angle; Normalized biological signal feature value Eigenvalues ​​after exponentially weighted moving average calculation Sigmoid nonlinear transformation eigenvalue The instantaneous rate of change of eye movement frequency Δf eye , eye movement frequency f eye The information entropy H eye , coefficient of variation of pupil diameter CV pupil , pupil diameter signal high frequency component energy D pupil , heart rate variability low frequency high frequency power ratio HRV LF / HF , Skin electrical energy attenuation factor EDA decay 、Head angle change rate θ stability Calculated, where X i =[Δf eye ,H eye ,CV pupil ,D pupil ,HRV LF / HF ,EDA decay ,θ stability ], the specific calculation process is as follows: Normalized biological signal feature value The biological signal feature value X i Mapped to [0,1] to avoid dimension problems and ensure comparability between different features, the calculation formula is as follows: In the formula, max(X i ) is the biological signal characteristic value X i The maximum value, min(X i ) is the biological signal characteristic value X i The minimum value of Eigenvalues ​​after exponentially weighted moving average calculation Exponential weighted moving average is used for smoothing to reduce the impact of noise on the eigenvalue X i To avoid the impact of drastic fluctuations in a short period of time and cause wrong judgment, the calculation formula is as follows: In the formula, ε is the smoothing coefficient, which controls the balance between new data and historical data. When ε is close to 1, it is more dependent on current data and has a faster response speed. When ε is close to 0, it is more dependent on historical data and has a higher degree of smoothing. Sigmoid nonlinear transformation eigenvalue Use Sigmoid nonlinear transformation to enhance the stability of data and avoid eigenvalue X i The impact caused by being too large or too small is uneven. The calculation formula is as follows: In the formula, is the normalized biological signal feature value.

4. The method for calculating rear wheel turning angle of an intelligent vehicle considering driving style according to claim 1, characterized in that: The loss function calculation module calculates the driver's operating burden loss function L by the following formula: busy , vehicle rollover risk loss function L roll 、Vehicle skidding risk loss function L slip And the physical loss function L physics : Driver control burden loss function L busy The frequency and amplitude of the driver's steering wheel operation are measured to assess the driver's driving burden. The calculation formula is as follows: In the formula, δ f [i] is the steering wheel angle at time i, Δt is the time step; ψ i is a weight factor used to measure the importance of different data nodes. The specific value is obtained by experiment and ranges from 0 to 1; busy is the weight factor of the driver's burden loss function, which is used to measure the weight of the driver's steering wheel operation frequency and amplitude. The specific value is obtained by experiment and ranges from 0 to 1. N is the total number of data samples; Softplus is a soft regularization function that makes the loss function smooth and differentiable; Vehicle rollover risk loss function L roll Measure the lateral acceleration a y As well as the over-limit situation of the vehicle's roll angle φ, the rollover risk of the vehicle is assessed. The specific calculation formula is as follows: In the formula, a y [i] is the lateral acceleration at time i, φ[i] is the vehicle roll angle at time i, a y,max is the maximum value of the vehicle's lateral acceleration, φ max is the maximum value of the vehicle roll angle, a y,th is the lateral acceleration threshold of the vehicle, φ th is the vehicle roll angle threshold; ψ i is a weight factor used to measure the importance of different data nodes. The specific value is obtained by experiment and ranges from 0 to 1; roll,1 , roll,2 is the weight factor of the vehicle rollover risk loss function, which is used to adjust the vehicle rollover risk loss function L roll The weight of each item in the range is 0 to 1; N is the total number of data samples; Softplus is a soft regularization function that makes the loss function smooth and differentiable; Vehicle skidding risk loss function L slip Measure the front and rear wheel slip angle α f and α r The over-limit situation is used to evaluate the side slip risk of the vehicle. The specific calculation formula is as follows: In the formula, α f [i] is the front wheel slip angle at time i, α r [i] is the rear wheel slip angle at time i, is the linear interval threshold of the front wheel slip angle, is the linear interval threshold of the rear wheel slip angle, α max is the linear interval threshold of tire slip angle; ψ i is a weight factor used to measure the importance of different data nodes. The specific value is obtained by experiment and ranges from 0 to 1; slip,1 and λ slip,2 is the vehicle skidding risk loss function L slip Weight factor, used to adjust the vehicle skidding risk loss function L slip The weight of each item ranges from 0 to 1; N is the total number of data samples; Softplus is a soft regularization function that makes the loss function smooth and differentiable; Physical loss function L physics Measure the ratio of the front and rear wheel angles predicted by the neural network x NN and x model The difference between them ensures the basic physical law of the composite mechanism model. The specific calculation formula is as follows: In the formula, x NN [i] is the predicted value at time i obtained by the neural network, x model [i] is the ideal value of the front and rear wheel steering angle ratio at time i calculated by the mechanism model, and N is the total number of data samples.

5. The method for calculating rear wheel turning angle of an intelligent vehicle considering driving style according to claim 1, characterized in that: The loss function calculation module calculates the loss function weighting factor ξ according to the attention mechanism i , the specific calculation process is as follows: First, the driver's driving style factor S d Perform nonlinear mapping to obtain the driving style factor S′ after nonlinear mapping d and high-order driving style factors The specific calculation formula is as follows: In the formula, is the weight matrix, all obtained through training; is the bias term, all of which are obtained through training; ReLU is the activation function of the first hidden layer, SiLU is the activation function of the second hidden layer, and tanh is the activation function of the third hidden layer; S d is the driver's driving style factor; Then the temperature coefficient τ is calculated. The specific calculation formula is as follows: Where λ1, λ2, and λ3 are weighting coefficients, whose specific values ​​are obtained by experiments and range from 0 to 1; W τ1 , W τ2 is the weight matrix, obtained through training; b τ1 , b τ2 is a bias term, all obtained through training; SiLU and tanh are activation functions; S′ d is the driving style factor obtained after nonlinear mapping, S d is the driver's driving style factor, It is a high-level driving style factor; Then the attention weight The specific calculation formula is as follows: In the formula, S′ d is the driving style factor obtained after nonlinear mapping, ξ initial is the initial weighting factor of the loss function, and its specific value is obtained according to the experiment; Q is the query vector of driving style, which is used to calculate the matching degree between the current driver style and the driving behavior factor; K is the driving behavior characteristic factor, which provides the information of historical driving behavior; V is the weighting factor value vector of the loss function, which provides the basic weight for dynamic adjustment; W Q , W K , W V , is the weight matrix, and its specific value is obtained according to the experiment. Q , b K , b V , is the bias term, and its specific value is obtained according to the experiment. is the attention weight, d k is the normalization factor, τ is the temperature coefficient to prevent the gradient from disappearing; SiLU is the activation function, and Softmax is the normalization function; From the calculated attention weights, the final loss weighting factor ξ can be calculated i , the calculation formula is as follows: In the formula, is the attention weight and V is the loss function weighting factor value vector.

Images