Man-machine sharing gain scheduling method considering driver characteristics

Abstract

The invention belongs to the technical field of man-machine cooperative driving and vehicle control, and discloses a man-machine sharing gain scheduling method considering driver characteristics, which comprises the following steps of: constructing a driver intention evaluation model and an ability evaluation model, analyzing dynamic behavior characteristics of a driver in real time, and designing a man-machine permission allocation model; dynamically adjusting the distribution proportion of the vehicle control right between the driver and the automatic driving system; a model based on driver behaviors is provided, and the driving process is comprehensively described; then, on the basis of a T-S method, dynamic modeling is achieved in combination with the fuzzy rule base and a membership function, and a more accurate and flexible dynamic model is output through weighted average; and finally, based on state input of the T-S fuzzy model, designing a homogeneous polynomial controller and proving system stability, and realizing high-precision and high-robustness control of the driver-vehicle-road system.

Description

Technical Field

[0001] The present invention belongs to the technical field of human-machine collaborative driving and vehicle control, and specifically relates to a human-machine shared gain scheduling method taking into account driver characteristics. Background Art

[0002] With the development of intelligent driving technology, human-machine collaborative driving has become an important approach to addressing the transition between autonomous driving and human driving. The core goal of human-machine collaborative driving is to achieve efficient collaboration between the human driver and the autonomous driving system by dynamically allocating driving authority, thereby improving driving safety, comfort, and efficiency.

[0003] However, the authority allocation strategies in existing technologies mostly rely on fixed thresholds (such as the steering wheel operation interval time) or single-dimensional state detection (such as fatigue monitoring), and lack a dynamic joint evaluation of the driver's intention and ability, resulting in delayed or false triggering of authority switching; traditional driver models such as the preview-follow model are based on linear assumptions and cannot describe nonlinear behavior in scenarios such as emergency obstacle avoidance. In addition, the two-degree-of-freedom model in the vehicle dynamics model often ignores the dynamic changes of the road environment, resulting in significant deviations between simulation and actual working conditions. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a human-machine shared gain scheduling method that takes into account driver characteristics. By dynamically evaluating driver status, high-precision modeling and advanced control technology, driving safety and vehicle stability are significantly improved.

[0005] The present invention provides a human-machine shared gain scheduling method considering driver characteristics, comprising the following steps:

[0006] S1. Build driver intention and capability assessment models to analyze the driver's dynamic behavior characteristics in real time. Based on the assessment results, design a human-machine authority allocation model to dynamically adjust the allocation ratio of vehicle control rights.

[0007] S2. Propose a model based on driver behavior, combine the interactions among the driver, vehicle, and road, and construct a driver-vehicle-road dynamics model to comprehensively describe the driving process;

[0008] S3, a method based on the TS fuzzy model, uses fuzzy reasoning to deal with uncertainty and optimize the driver-vehicle-road dynamics model;

[0009] S4. Based on the state input of the TS fuzzy model, a homogeneous polynomial controller is designed and the Lyapunov stability theory is used to prove the system stability, achieving high-precision and high-robustness control of the driver-vehicle-road dynamics model.

[0010] Furthermore, in S1, the driver's longitudinal and lateral control of the vehicle is used to quantify the driver's intention, where the longitudinal driving behavior is represented by the longitudinal speed v x and acceleration a x Indicates that the lateral driving behavior is determined by the steering wheel angle δ d and angular velocity The driver intention evaluation model includes a two-level parallel GBRBM network. The first level includes a longitudinal GBRBM network and a transverse GBRBM network, which respectively evaluate the longitudinal driving intention P1 and the transverse driving intention P2. The second level combines P1 and P2 into a comprehensive driving intention index DI through a comprehensive GBRBM network. The longitudinal, transverse and comprehensive GBRBM network structures are the same. The input of the longitudinal GBRBM network is κ 1 =[v x ,a x ] T , the output is h 1 =H1, which represents the quantized strength of the driver's longitudinal control; the input of the transverse GBRBM network is The output is h 2 =H2, indicating the intensity value of the driver's lateral control after quantization; the input and output of the integrated network are κ 3 =[P1,P2] T and h 3 =H3, which represents the core parameter of dynamic permission allocation after integrating the two local intent parameters; T is the transpose of the matrix;

[0011] The driver intention evaluation model is trained and the energy function of the GBRBM network is introduced to describe the relationship between the input and output of the network. The formula is as follows:

[0012]

[0013] Where, They represent the vertical GBRBBM network, the horizontal GBRBBM network, and the comprehensive GBRBBM network respectively; m and n are the number of nodes in the input layer and the output layer respectively; and Represents GBRBM network respectively Input and output; and Network The bias of the input and output layers; is the connection weight; is the standard deviation of the input Gaussian noise;

[0014] Conditional probability distribution is introduced to quantify the relationship between the driver's intention and the vehicle state, thereby more accurately describing the driver's behavioral characteristics. Conditional probability distribution is used to describe the following two situations:

[0015] (1) Known input Under the condition of The probability distribution of :

[0016]

[0017] (2) Known output Under the condition of The probability distribution of :

[0018]

[0019] In order to improve the prediction accuracy of the model, the network parameters are updated to minimize the value of the energy function, and thus it is derived that and The increment is:

[0020]

[0021] Where η is the learning rate, and They represent the expected values of the input layer, hidden layer, and the data obtained by the joint activation of the two during the sample training process, and They represent the probability distributions of the input layer, hidden layer, and the joint activation of the two during the current GBRBM network training process.

[0022] Furthermore, the driver ability evaluation formula is:

[0023]

[0024] Among them, e p represents the lateral error, which is the distance between the vehicle's current position and the reference path; Δψ represents the heading error, which is the angular difference between the vehicle's heading angle and the tangent direction of the reference path; α1 and α2 are weight coefficients that measure the impact of lateral error and heading error on driving ability, respectively;

[0025] Use the traversal search method to determine the weight coefficients α1 and α2:

[0026]

[0027] in, Indicates that the lateral error and heading error are divided into Level DA ι Denotes the driving ability evaluation value at the ιth level, and its upper and lower boundaries are set as DAι max and DA ι min ;e pι is the lateral error at the ιth level; Δψ ι represents the heading error at the ιth level.

[0028] Furthermore, based on the evaluation results, an exponential function is used to design a human-machine authority allocation model:

[0029]

[0030] Where λ is the controller's authority level; min is the minimum assistance level, ensuring that the shared controller is always involved in the driving task; μ1 and μ3 represent the driver's involvement in the driving task and driving skills, respectively; μ2 and μ4 represent the impact of driving intention and ability on the authority level, respectively.

[0031] Furthermore, S2 is specifically:

[0032] S21. Constructing a time-varying preview driver model based on steering angle;

[0033] Define a pair of visual points that the driver observes when tracking the curve trajectory, namely the near prediction point x n and the far prediction point x f , calculate the corresponding virtual steering angle φ based on the predicted information n and φ f :

[0034]

[0035] φ f =x f ρ

[0036] Where, φ n Close-range virtual steering angle, φ f is the long-distance virtual steering angle, ρ is the road curvature, e p is the lateral error of the pilot near the preview point, and Δψ is the heading error;

[0037] The driver determines the steering wheel angle θ based on the two virtual steering angles through the neuromuscular system d , the steering response is described using the following second-order model:

[0038]

[0039] Where s is the Laplace variable; G p (t) is the feedforward gain, which represents the driver’s ability to predict future traffic conditions; G c(t) is the feedback gain used to correct tracking deviation; ζ(t) represents the neuromuscular damping, reflecting the frequency of the driver's steering wheel adjustment; ω(t) represents the natural frequency of the neuromuscular system of the driver's reaction time;

[0040] S22. Constructing a vehicle-road dynamics model;

[0041] The dynamic expression of the vehicle is as follows:

[0042]

[0043] Where M is the mass of the vehicle, F xf ,F yf Represent the front wheel longitudinal force and the rear wheel longitudinal force, F xr ,F yr represent the front wheel lateral force and the rear wheel lateral force respectively, l f and l n are the distances from the vehicle's center of mass to the front and rear axles, v x ,v y are the longitudinal and lateral velocities of the vehicle, represents the derivative of the lateral velocity, i.e., the lateral acceleration, θ is the front wheel steering angle, r is the yaw rate of the vehicle, represents the yaw angular acceleration, I z is the moment of inertia around the z-axis;

[0044] According to the vehicle's driving conditions and the characteristics of the wheels, the steering force and lateral force of the front and rear wheels are calculated; the body slip angle β is considered to be a small angle, so

[0045] Calculate the rear wheel longitudinal force according to the following formula and the rear wheel lateral force

[0046]

[0047] Where C f and C r are the front and rear lateral stiffness, respectively;

[0048] Based on the geometric relationship between the vehicle and the road, the lateral error e is defined p And the heading error Δψ:

[0049]

[0050] Where x m is the preview distance; and are the derivatives of lateral error and heading error respectively;

[0051] The steering angle applied to the front wheels of the vehicle is determined by the driver and the controller. f for:

[0052] θ f =(K d (t)θ d +K c (t)θ c )r g

[0053] Where θ d is the steering wheel angle, θ c The steering angle provided by the controller, the control authority levels of the controller and the driver are K c and K d , r g is the rotation ratio;

[0054] S23, constructing a driver-vehicle-road dynamics model;

[0055] Combine the vehicle dynamics expression, lateral error and heading error expression with the time-varying preview driver model and define the state vector in, is the derivative of the steering wheel angle, and the driver-vehicle-road dynamics model is obtained as:

[0056]

[0057] Where x(t) is the system state and the control input u(t) = θ c Steering angle provided to the controller, external disturbance is the road curvature;

[0058] The specific forms of matrices A, B, and E are as follows:

[0059]

[0060] B=[K c b 11 K c b 21 0000] T ,

[0061] E=[000-v x 0k p x f ] T

[0062] Among them, k p represents the external disturbance gain coefficient;

[0063]

[0064] a 64 =-k c ,a 65 =-ω 2 ,a 66 =-2ζω,

[0065]

[0066] The control output takes lane departure error and driver comfort into consideration, and the controlled output z(t) is designed to be

[0067] z(t)=Cx(t)+Du(t)

[0068] Where u(t) represents the input of the system,

[0069]

[0070] Furthermore, S3 is specifically:

[0071] For the driver-vehicle-road dynamics model, the fuzzy rules are constructed as follows:

[0072] The i-th fuzzy rule: If yes yes yes yes So:

[0073]

[0074] Where, It's the speed. It is the inverse of the vehicle speed. is the inverse of the square of the vehicle speed, is the steering gain, and represents a fuzzy set; v x ∈[v x min ,v x max ] are the upper and lower bounds of vehicle speed, A i ,B i ,E i ,C i ,D i is the system matrix corresponding to the i-th fuzzy rule;

[0075] To reduce the conservatism of the model, Taylor approximation is used to replace the variables:

[0076]

[0077] Among them, the longitudinal velocity v x ∈[vx min ,v x max ],

[0078] The following membership function is used to describe the fuzzification process of the premise variables:

[0079]

[0080] Where v(t) is the normalized variable of vehicle speed and satisfies v(t)∈[-1,1]; K d (t) is the normalized variable of the conversion gain and satisfies K d (t)∈[K d min ,K d max ];

[0081] Based on the above principles, the TS fuzzy model is constructed:

[0082]

[0083] Among them, h i (υ(t)) represents the i-th fuzzified weighting function; h1 = m1 × l1, h2 = m1 × l2, h3 = m2 × l1 and h4 = m2 × l2, and satisfies:

[0084]

[0085] Ξ1=Ξ(v min ,K d min ),Ξ2=Ξ(v min ,K d max ),

[0086] Ξ3=Ξ(v max ,K d min ),Ξ4=Ξ(v max ,K d max ),

[0087] Ξ∈[A,B,C,D,E].

[0088] Furthermore, S4 is specifically:

[0089] S41. For the sake of universality, consider a general form of continuous-time TS fuzzy model:

[0090]

[0091] Where, Indicates status, represents the control input, Indicates interference, represents the control input, represents a set of real numbers, v(t) represents a premise variable, hi (v(t)) represents the membership function and satisfies h i (v(t))≥ and

[0092] Using polynomials to describe the nonlinear characteristics of TS fuzzy systems, we first define the aligned polynomials: define the q-tuple in is a non-negative number that satisfies g is the order of the homogeneous polynomial; definition is a q-tuple The factorial product of

[0093] S42. Design a homogeneous polynomial controller:

[0094]

[0095] Where K g (h) a homogeneous polynomial matrix to be determined, and

[0096] Substituting the above controller into the system state equation, we get:

[0097]

[0098] in,

[0099]

[0100] S43. Use Lyapunov stability theory to prove the stability of the system:

[0101] Construct the Lyapunov function V(t):

[0102] V(t)=x T (t)Px(t)

[0103] Where, the matrix P>0 is a positive definite matrix, and this function reflects the energy of the system state x(t);

[0104] The goal of using the Lyapunov function to determine system stability is to analyze the derivative of V(t) if Then the system is stable, and we can get as follows:

[0105]

[0106] Where, represents the derivative of the system state x(t);

[0107] While proving the stability of the system, it is also necessary to ensure that the system is resistant to external disturbances. To this end, H ∞ Performance index; if the system output z(t) and external disturbance satisfy:

[0108]

[0109] Where γ>0 is a given constant, then the system is said to satisfy H ∞ performance index γ;

[0110] Construct a System output energy z T (t)z(t) and perturbation energy Inequality Among them, α>0 is used to prove exponential stability; through algebraic operations, the inequality is transformed into a matrix inequality, and the controller gain matrix is solved using the linear matrix inequality (LMI) technique;

[0111] Finally, the exponential stability and H ∞ Satisfaction of performance indicators.

[0112] The beneficial effects described in the present invention are as follows: the present invention constructs a driver intention evaluation model and an ability evaluation model to evaluate the driver's driving intention and ability, that is, the degree and ability of road tracking in real time, and uses an exponential function to dynamically adjust the distribution weight of control rights to realize a dynamic authority allocation strategy based on time-varying driving characteristics; the present invention establishes a driver-vehicle-road coupling dynamics model, adopts TS fuzzy technology to process nonlinear terms in the model, and introduces homogeneous polynomial technology, combined with the normalized properties of the membership function, to design a homogeneous parameter-dependent controller. Since the high-order system adds additional degrees of freedom, the controller can more fully explore the information of the fuzzy membership function, expand more combinations of control gain matrices, and thus more accurately adapt to complex and changeable working conditions, significantly reduce the conservatism of the control system, and combine with the Lyapunov function to ensure the global stability and robustness of the system. The method described in the present invention realizes the accurate assessment of the driver's status in the automatic driving system, the reasonable allocation of human-machine authority and the stable control of the system. It not only significantly improves driving safety and comfort, but also provides important theoretical and technical support for the practical application of automatic driving technology. It has broad application prospects and important application value in terms of safety, comfort and technological advancement, providing key technical support for the practical application of automatic driving technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 is a flow chart of the method of the present invention;

[0114] Figure 2 It is a schematic diagram of the driver intention graph model structure. DETAILED DESCRIPTION

[0115] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0116] like Figure 1 As shown, the human-machine shared gain scheduling method considering the driver's characteristics according to the present invention includes the following steps:

[0117] S1. Build driver intention and capability assessment models to analyze the driver's dynamic behavior characteristics in real time. Based on the assessment results, design a human-machine authority allocation model to dynamically adjust the allocation ratio of vehicle control rights.

[0118] S2. Propose a model based on driver behavior, combine the interactions among the driver, vehicle, and road, and construct a driver-vehicle-road dynamics model to comprehensively describe the driving process;

[0119] S3, a method based on the TS fuzzy model, uses fuzzy reasoning to deal with uncertainty and optimize the driver-vehicle-road dynamics model;

[0120] S4. Based on the state input of the TS fuzzy model, a homogeneous polynomial controller is designed and the Lyapunov stability theory is used to prove the system stability, achieving high-precision and high-robustness control of the driver-vehicle-road dynamics model.

[0121] Specifically, as discussed in S1, by building a driver intention assessment model and a capability assessment model, the driver's dynamic behavior characteristics are analyzed in real time. Based on the assessment results, a human-machine authority allocation model is designed to dynamically adjust the allocation ratio of vehicle control rights between the driver and the autonomous driving system to achieve safe and smooth authority switching. The main points are as follows:

[0122] S11, driver intention assessment model;

[0123] Driver intention not only reflects the driver's control behavior of the vehicle, but also directly affects the driving safety and comfort of the vehicle. The driver's longitudinal and lateral control of the vehicle is used to quantify the driver's intention, among which the longitudinal driving behavior is usually represented by the longitudinal speed v x and acceleration a x , while the lateral driving behavior is represented by the steering wheel angle δ d and angular velocity Therefore, it is necessary to use a dimensionality reduction algorithm to convert multiple driving states into evaluation indicators. Considering that the driver's driving state will change over time, this embodiment uses the Gauss-Bernoulli restricted Boltzmann machine (GBRBM) algorithm to quantify the driver's intention. The driver intention evaluation model is as follows Figure 2 As shown in Figure 1, it is a two-level parallel GBRBM. The first layer includes a longitudinal GBRBM network and a lateral GBRBM network, which respectively evaluate the longitudinal driving intention P1 and the lateral driving intention P2. The second layer combines P1 and P2 into a comprehensive driving intention index DI through a comprehensive GBRBM network. Each GBRBM network is implemented independently, where the input of the longitudinal GBRBM network is κ 1 =[v x ,a x ] T , the output is h 1 =H1, which represents the quantized strength of the driver's longitudinal control; the input of the lateral GBRBM network is The output is h 2 =H2, which represents the strength value of the quantized driver's lateral control; the input and output of the integrated GBRBM network are κ 3 =[P1,P2] T and h 3 =H3, which represents the core parameter of dynamic permission allocation after integrating the two local intent parameters.

[0124] The vertical, horizontal, and integrated GBRBM networks share the same structure, consisting of an input layer, hidden layers, and parameters. The specific workflow is as follows: the input layer receives driving data, which is then weighted and transmitted to the hidden layers via connection weights. Each network's input and output layers have specific bias data, derived from previously trained network values. Connection weights and biases are algorithmically optimized to ensure that the hidden layer outputs best represent the system's characteristics.

[0125] In order to make the driver intention model more accurate, the model needs to be trained. The energy function of the GBRBM network is introduced to describe the relationship between the input and output of the network. The formula is as follows:

[0126]

[0127] Where, They represent the vertical GBRBBM network, the horizontal GBRBBM network, and the comprehensive GBRBBM network respectively; m and n are the number of nodes in the input layer and the output layer respectively; and Represents GBRBM network respectively Input and output; and Network The bias of the input and output layers; is the connection weight; is the standard deviation of the Gaussian noise of the input.

[0128] In order to accurately describe and quantify the relationship between the driver's intention and the vehicle state, conditional probability distribution is introduced to quantify the relationship between this input and output, thereby more accurately describing the driver's behavioral characteristics. Conditional probability distribution is used to describe the following two situations:

[0129] (1) Known input Under the condition of The probability distribution of :

[0130]

[0131] (2) Known output Under the condition of The probability distribution of :

[0132]

[0133] In order to improve the prediction accuracy of the model, the network parameters are updated to minimize the value of the energy function, and thus it is derived that and The increment is:

[0134]

[0135] Where η is the learning rate, usually set to 0.06; and They represent the expected values of the input layer, hidden layer, and the data obtained by the joint activation of the two during the sample training process, and They represent the probability distributions of the input layer, hidden layer, and the joint activation of the two during the current GBRBM network training process.

[0136] S12, driver ability assessment model;

[0137] Driver capability is used to represent the driver's performance in path tracking, mainly through the lateral error e p (lateral deviation between the vehicle’s center of gravity and a reference path) and the heading error Δψ (angular deviation between the vehicle’s heading and the reference path).

[0138] Considering that a single error indicator (such as lateral error alone) cannot fully reflect the driver's ability, it is necessary to design a comprehensive evaluation formula that combines lateral error and heading error. Based on the idea of artificial potential field, a driver ability evaluation formula is designed:

[0139]

[0140] Among them, e p represents the lateral error, which is the distance between the vehicle's current position and the reference path; Δψ represents the heading error, which is the angular difference between the vehicle's heading angle and the tangent direction of the reference path; α1 and α2 are weight coefficients, which respectively measure the impact of lateral error and heading error on driving ability.

[0141] In this embodiment, the traversal search method is used to determine the weight coefficients of α1 and α2:

[0142]

[0143] in, Indicates that the lateral error and heading error are divided into Level; DA ι Denotes the driving ability evaluation value at the ιth level, and its upper and lower boundaries are set as DA ι max and DA ι min ;e pι is the lateral error at the ιth level; Δψ ι represents the heading error at the ιth level.

[0144] Based on a large amount of driver path tracking data, the k-means clustering method is used to classify the lateral error and heading error into different levels, corresponding to different driver capabilities DA i By dividing the ability levels, it supports the refined evaluation and classification of different driving abilities.

[0145] S13, human-machine authority allocation strategy;

[0146] The goal of the human-machine authority allocation strategy is to dynamically adjust the control authority between the driver and the autonomous driving system based on the driver's real-time driving intention and ability to improve driving performance and safety. When the driver's driving intention and ability are high, the driver is given more control authority; when the driver's driving intention and ability are low, the autonomous driving system's level of intervention is increased. Based on the evaluation results, an authority allocation formula is designed using an exponential function:

[0147]

[0148] Where λ is the authority level of the controller; min is the minimum assistance level, ensuring that the shared controller is always involved in the driving task; μ1 and μ3 represent the driver's involvement in the driving task and driving skills, respectively; μ2 and μ4 represent the impact of driving intention and ability on the authority level, respectively; in this embodiment, μ1 = 2, μ3 = 1, μ2 = 3, and μ4 = 3.

[0149] The human-machine authority allocation strategy evaluates the driver's driving intention and ability in real time and dynamically adjusts the allocation of control rights using an exponential function. This strategy not only improves the safety and efficiency of human-machine collaborative driving, but also flexibly responds to changes in the driver's state, ensuring system reliability and a better driving experience.

[0150] Specifically, S2 proposes a driver behavior-based model that combines the interactions among the driver, vehicle, and road to comprehensively describe the driving process. The specific steps are as follows:

[0151] Building a driver-vehicle-road dynamics model is a critical task in autonomous driving, human-machine collaborative driving systems, and traffic management. Such a model not only reflects the dynamic interactions between the driver, vehicle, and road, but also helps us better understand and predict the impact of driver behavior on traffic flow, driving safety, and driving performance.

[0152] S21, time-varying preview driver model based on steering angle;

[0153] First, we need to define the driver's behavior. When the driver follows a curve, he usually observes a pair of specific visual points (near the predicted point x n and the far prediction point x f ), and adjust the steering wheel angle according to these prediction information. In this embodiment, the near prediction point x n and the far prediction point x f The distances are x n =0.5x f and x f ≈10 meters. Calculate the corresponding virtual steering angle φ based on the predicted information n and φ f :

[0154]

[0155] φ f =x f ρ

[0156] Where, φ n Close-range virtual steering angle, φ f is the long-distance virtual steering angle, ρ is the road curvature, e pis the lateral error of the pilot near the preview point, and Δψ is the heading error.

[0157] The driver then controls the steering angle, and the driver determines the steering wheel angle θ according to the two virtual steering angles through the neuromuscular system. d , which is affected by both the near-range and far-range virtual steering angles. The model uses the following second-order model to describe the steering response:

[0158]

[0159] Where s is the Laplace variable; G p (t) is the feedforward gain, which represents the driver’s ability to predict future traffic conditions; G c (t) is the feedback gain, which is used to correct the tracking deviation; ζ(t) represents the neuromuscular damping, which reflects the frequency of the driver's steering wheel adjustment; ω(t) represents the natural frequency of the neuromuscular system of the driver's reaction time.

[0160] S22. Constructing a vehicle-road dynamics model;

[0161] The vehicle-road dynamics model is a two-dimensional vehicle model involving the longitudinal velocity v of the vehicle. x , lateral velocity v y , front wheel steering angle δ, and vehicle angular velocity r. The dynamic expression of the vehicle is as follows:

[0162]

[0163] Where M is the mass of the vehicle, F xf ,F yf Represent the front wheel longitudinal force and the rear wheel longitudinal force, F xr ,F yr represent the front wheel lateral force and the rear wheel lateral force respectively, l f and l n are the distances from the vehicle's center of mass to the front and rear axles, v x ,v y are the longitudinal and lateral velocities of the vehicle, represents the derivative of the lateral velocity, i.e., the lateral acceleration, θ is the front wheel steering angle, r is the yaw rate of the vehicle, represents the yaw angular acceleration, I z is the moment of inertia about the z-axis.

[0164] Based on the vehicle's driving conditions and the characteristics of the wheels, the steering force and lateral force of the front and rear wheels are calculated. The lateral force represents the lateral deviation of the vehicle when turning, while the steering force is directly related to the steering angle of the front wheels. The body slip angle (i.e. β) can be regarded as a small angle, so we can get The model calculates the rear wheel longitudinal force according to the following formula and the rear wheel lateral force

[0165]

[0166] Where C f and C r are the front and rear lateral stiffness, respectively;

[0167] Based on the geometric relationship between the vehicle and the road, the lateral error e is defined p And the heading error Δψ:

[0168]

[0169] Where x m is the preview distance; and are the derivatives of the lateral error and heading error, respectively.

[0170] The steering angle applied to the front wheels of the vehicle is determined by the driver and the controller. f for:

[0171] θ f =(K d (t)θ d +K c (t)θ c )r g

[0172] Where θ d is the steering wheel angle, θ c The steering angle provided by the controller, the control authority levels of the controller and the driver are K c and K d , r g is the rotation ratio.

[0173] S23, driver-vehicle-road dynamics model;

[0174] Combine the vehicle dynamics model and the road tracking error model to define the state vector in, is the vehicle side slip angle, is the derivative of the steering wheel angle, and the driver-vehicle-road dynamics model can be obtained as:

[0175]

[0176] Where x(t) is the system state and the control input u(t) = θ c Steering angle provided to the controller, external disturbance is the road curvature.

[0177] The specific forms of matrices A, B, and E are as follows:

[0178]

[0179] B=[K c b 11 K c b 21 0000] T ,

[0180] E=[000-v x 0k p x f ] T .

[0181] in,

[0182]

[0183] a 64 =-k c ,a 65 =-ω 2 ,a 66 =-2ζω,

[0184] K c =1-K d .

[0185] The control output takes lane departure error and driver comfort into consideration, and the controlled output z(t) is designed to be

[0186] z(t)=Cx(t)+Du(t)

[0187] Where u(t) represents the input of the system,

[0188]

[0189] S3 proposes a TS fuzzy modeling-based method to describe the nonlinear terms of the system, decompose the nonlinear driver-vehicle-road coupling system into multiple linear subsystems, and implement dynamic modeling by combining the fuzzy rule base and membership function. Finally, a more accurate and flexible dynamic model is output using weighted average, which is highly close to the original system and facilitates the controller design and stability analysis in the following text.

[0190] Further including:

[0191] S31. Introduction to TS fuzzy rules

[0192] Use the Sector Nonlinearity Approach to construct TS fuzzy rules:

[0193] if yes yes yes Then we can get:

[0194]

[0195] Wherein, i=1,...,N represents the number of fuzzy rules.

[0196] S32. Driver-Vehicle-Road Dynamics Model Based on TS Fuzzy Model

[0197] For the driver-vehicle-road dynamics model, the fuzzy rules are constructed as follows:

[0198] if yes yes yes yes So:

[0199]

[0200] Where, It's the speed. It is the inverse of the vehicle speed. is the inverse of the square of the vehicle speed, is the steering gain. v x ∈[v x min ,v x max ] are the upper and lower bounds of vehicle speed, A i ,B i ,E i ,C i ,D i is the system matrix corresponding to the i-th fuzzy rule.

[0201] To reduce the conservatism of the model, Taylor approximation is used to replace the premise variables:

[0202]

[0203] Among them, the longitudinal velocity v x ∈[v x min ,v x max ], Here v0 and v1 are auxiliary variables introduced in the process of Taylor approximation and variable substitution. Obviously, when v(t)=v min = -1, v x =v x min ; When v(t)=vmax =1, v x =v x max .

[0204] The fuzzy membership function is used to describe the fuzzification process of the premise variables. The present invention adopts the following membership function:

[0205]

[0206]

[0207] Where v(t) is the normalized variable of vehicle speed and satisfies v(t)∈[-1,1]; K d (t) is the normalized variable of the conversion gain and satisfies K d (t)∈[K d min ,K d max ].

[0208] Based on the above approximation, the TS fuzzy model is constructed:

[0209]

[0210] Among them, h1=m1×l1, h2=m1×l2, h3=m2×l1 and h4=m2×l2, and satisfy:

[0211]

[0212] Ξ1=Ξ(v min ,K d min ),Ξ2=Ξ(v min ,K d max ),

[0213] Ξ3=Ξ(v max ,K d min ),Ξ4=Ξ(v max ,K d max ),

[0214] Ξ∈[A,B,C,D,E].

[0215] S4 specifically includes:

[0216] S41, Introduction to homogeneous polynomials;

[0217] For the sake of generality, consider a general form of continuous-time TS fuzzy model:

[0218]

[0219] Where, Indicates status, represents the control input, Indicates interference, represents the control input, represents a set of real numbers, v(t) represents a premise variable, h i (v(t)) represents the membership function and satisfies h i (v(t))≥ and

[0220] Homogeneous polynomials are used to describe the nonlinear characteristics of TS fuzzy systems. Now let's define some homogeneous polynomials: define the q-tuple in is a non-negative number that satisfies (g is the order of the homogeneous polynomial); definition is a q-tuple The factorial product of .

[0221] S42. Design of parameter-dependent controller based on homogeneous polynomials;

[0222] Before introducing the homogeneous polynomial controller, we first introduce the traditional PDC (Parallel distributed compensation) controller:

[0223]

[0224] Among them, K i is the gain matrix to be determined.

[0225] Different from the traditional PDC-type controller, the homogeneous polynomial-type controller considers more algebraic properties of normalized fuzzy membership functions (MFs) during the control design process, and designs a more flexible controller:

[0226]

[0227] Where K g (h) a homogeneous polynomial matrix to be determined, and

[0228] Combining the system dynamic equations, we can get:

[0229]

[0230] in,

[0231]

[0232] S43. System stability analysis

[0233] Use Lyapunov stability theory to prove the stability of the system. First, we need to construct a Lyapunov function V(t):

[0234] V(t)=x T (t)Px(t)

[0235] Wherein, the matrix P>0 is a positive definite matrix, and this function reflects the energy of the system state x(t).

[0236] The goal of using Lyapunov function to judge the stability of a system is to analyze if Then the system is stable, and we can get as follows:

[0237]

[0238] While proving the stability of the system, it is also necessary to ensure that the system is resistant to external disturbances. To this end, H ∞ Performance indicators. ∞ The performance index measures the ratio between the system output energy and the external disturbance energy. Specifically, if the system output z(t) and the external disturbance satisfy:

[0239]

[0240] Where γ>0 is a given constant, then the system is said to satisfy H ∞ Performance indicator γ.

[0241] Next, you need to construct a System output energy T (t)z(t) and perturbation energy Inequality Among them, αV(t) is an exponential decay term, which requires the energy V(t) of the system to decay at a rate α. This is the judgment condition of exponential stability, and α>0 is used to prove exponential stability. Through algebraic operations, this expression is transformed into a matrix inequality, and the controller gain matrix is solved using the linear matrix inequality (LMI) technique. Finally, the exponential stability and H of the system are proved respectively by Lyapunov function and integral inequality. ∞ Satisfaction of performance indicators.

[0242] The present invention uses Carsim and Matlab / Simulink joint simulation to verify the effectiveness of the controller. First, it is necessary to determine the parameters of the selected vehicle model, such as body mass, moment of inertia, longitudinal speed, distance from the center of mass to the front and rear wheels, etc. A C-class vehicle can be selected and the parameters are designed to the default values. Generally, the body mass M = 1410 (kg), the moment of inertia I z =1510(kg·m2 ), longitudinal velocity v x ∈[5(m / s),22(m / s)],l f =1.02(m) and l n = 1.67 (m), etc. For road conditions, the multi-curve serpentine road provided by Carsim is generally used. A homogeneous polynomial parameter-dependent controller is built in Matlab / Simulink. Using vehicle information transmitted from Carsim, the desired steering torque is calculated, resulting in the steering angle, which is then used to control the vehicle in Carsim. Operating data is collected in Matlab / Simulink simulations, and the LMI method is used to solve the matrix and controller gains to ensure exponential stability and dynamic performance of the closed-loop system.

[0243] The method described in the present invention can not only satisfy the driver's driving experience, but also ensure that when the driver is in poor condition, the driving authority ratio of the machine and the driver is adjusted in time, thereby improving the safety and reliability of driving.

[0244] The above description is only a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made using the contents of the present invention description and drawings are within the scope of protection of the present invention.

Claims

1. A human-machine shared gain scheduling method considering driver characteristics, characterized in that: The following steps are involved: S1. Build driver intention and capability assessment models to analyze the driver's dynamic behavior characteristics in real time. Based on the assessment results, design a human-machine authority allocation model to dynamically adjust the allocation ratio of vehicle control rights. S2. Propose a model based on driver behavior, combine the interactions among the driver, vehicle, and road, and construct a driver-vehicle-road dynamics model to comprehensively describe the driving process; S3, a method based on the TS fuzzy model, uses fuzzy reasoning to deal with uncertainty and optimize the driver-vehicle-road dynamics model; S4. Based on the state input of the TS fuzzy model, a homogeneous polynomial controller is designed and the Lyapunov stability theory is used to prove the system stability, achieving high-precision and high-robustness control of the driver-vehicle-road dynamics model.

2. The method for human-machine shared gain scheduling considering driver characteristics according to claim 1, characterized in that: In S1, the driver’s longitudinal and lateral control of the vehicle is used to quantify the driver’s intention, where the longitudinal driving behavior is represented by the longitudinal speed v x and acceleration a x Indicates that the lateral driving behavior is determined by the steering wheel angle δ d and angular velocity The driver intention evaluation model includes a two-level parallel GBRBM network. The first level includes a longitudinal GBRBM network and a transverse GBRBM network, which respectively evaluate the longitudinal driving intention P1 and the transverse driving intention P2. The second level combines P1 and P2 into a comprehensive driving intention index DI through a comprehensive GBRBM network. The longitudinal, transverse and comprehensive GBRBM network structures are the same. The input of the longitudinal GBRBM network is κ 1 =[v x ,a x ] T , the output is h 1 =H1, which represents the quantized strength of the driver's longitudinal control; the input of the transverse GBRBM network is The output is h 2 =H2, indicating the intensity value of the driver's lateral control after quantization; the input and output of the integrated network are κ 3 =[P1,P2] T and h 3 =H3, which represents the core parameter of dynamic permission allocation after integrating the two local intent parameters; T is the transpose of the matrix; The driver intention evaluation model is trained and the energy function of the GBRBM network is introduced to describe the relationship between the input and output of the network. The formula is as follows: Where, They represent the vertical GBRBBM network, the horizontal GBRBBM network, and the comprehensive GBRBBM network respectively; m and n are the number of nodes in the input layer and the output layer respectively; and Represents GBRBM network respectively Input and output; and Network The bias of the input and output layers; is the connection weight; is the standard deviation of the input Gaussian noise; Conditional probability distribution is introduced to quantify the relationship between the driver's intention and the vehicle state, thereby more accurately describing the driver's behavioral characteristics. Conditional probability distribution is used to describe the following two situations: (1) Known input Under the condition of The probability distribution of : (2) Known output Under the condition of The probability distribution of : In order to improve the prediction accuracy of the model, the network parameters are updated to minimize the value of the energy function, and thus it is derived that and The increment is: Where η is the learning rate, and They represent the expected values of the input layer, hidden layer, and the data obtained by the joint activation of the two during the sample training process, and They represent the probability distributions of the input layer, hidden layer, and the joint activation of the two during the current GBRBM network training process.

3. The method for human-machine shared gain scheduling considering driver characteristics according to claim 2, characterized in that: The driver ability assessment formula is: Among them, e p represents the lateral error, which is the distance between the vehicle's current position and the reference path; Δψ represents the heading error, which is the angular difference between the vehicle's heading angle and the tangent direction of the reference path; α1 and α2 are weight coefficients that measure the impact of lateral error and heading error on driving ability, respectively; Use the traversal search method to determine the weight coefficients α1 and α2: in, Indicates that the lateral error and heading error are divided into Level; DA ι Denotes the driving ability evaluation value at the ιth level, and its upper and lower boundaries are set as DA ιmax and DA ιmin ;e pι is the lateral error at the ιth level; Δψ ι represents the heading error at the ιth level.

4. The method for human-machine shared gain scheduling considering driver characteristics according to claim 3, characterized in that: Based on the evaluation results, an exponential function is used to design a human-machine authority allocation model: Where λ is the controller's authority level; min is the minimum assistance level, ensuring that the shared controller is always involved in the driving task; μ1 and μ3 represent the driver's involvement in the driving task and driving skills, respectively; μ2 and μ4 represent the impact of driving intention and ability on the authority level, respectively.

5. The human-machine shared gain scheduling method considering driver characteristics according to claim 3 is characterized in that , S2 is specifically: S21. Constructing a time-varying preview driver model based on steering angle; Define a pair of visual points that the driver observes when tracking the curve trajectory, namely the near prediction point x n and the far prediction point x f , calculate the corresponding virtual steering angle φ based on the predicted information n and φ f : f f =x f r Where, φ n Close-range virtual steering angle, φ f is the long-distance virtual steering angle, ρ is the road curvature, e p is the lateral error of the pilot near the preview point, and Δψ is the heading error; The driver determines the steering wheel angle θ based on the two virtual steering angles through the neuromuscular system d , the steering response is described using the following second-order model: Where s is the Laplace variable; G p (t) is the feedforward gain, which represents the driver’s ability to predict future traffic conditions; G c (t) is the feedback gain used to correct tracking deviation; ζ(t) represents the neuromuscular damping, reflecting the frequency of the driver's steering wheel adjustment; ω(t) represents the natural frequency of the neuromuscular system of the driver's reaction time; S22. Constructing a vehicle-road dynamics model; The dynamic expression of the vehicle is as follows: Where M is the mass of the vehicle, F xf ,F yf Represent the front wheel longitudinal force and rear wheel longitudinal force, F xr ,F yr represent the front wheel lateral force and the rear wheel lateral force respectively, l f and l n are the distances from the vehicle's center of mass to the front and rear axles, v x ,v y are the longitudinal and lateral velocities of the vehicle, represents the derivative of the lateral velocity, i.e., the lateral acceleration, θ is the front wheel steering angle, r is the yaw rate of the vehicle, represents the yaw angular acceleration, I z is the moment of inertia around the z-axis; According to the vehicle's driving conditions and the characteristics of the wheels, the steering force and lateral force of the front and rear wheels are calculated; the body slip angle β is considered to be a small angle, so Calculate the rear wheel longitudinal force according to the following formula and the rear wheel lateral force Where C f and C r are the front and rear lateral stiffness, respectively; Based on the geometric relationship between the vehicle and the road, the lateral error e is defined p And the heading error Δψ: Where x m is the preview distance; and are the derivatives of lateral error and heading error respectively; The steering angle applied to the front wheels of the vehicle is determined by the driver and the controller. f for: i f =(K d (t)θ d +K c (t)θ c )r g Where θ d is the steering wheel angle, θ c The steering angle provided by the controller, the control authority levels of the controller and the driver are K c and K d , r g is the rotation ratio; S23, constructing a driver-vehicle-road dynamics model; Combine the vehicle dynamics expression, lateral error and heading error expression with the time-varying preview driver model and define the state vector in, is the derivative of the steering wheel angle, and the driver-vehicle-road dynamics model is obtained as: Where x(t) is the system state and the control input u(t) = θ c Steering angle provided to the controller, external disturbance is the road curvature; The specific forms of matrices A, B, and E are as follows: Among them, k p represents the external disturbance gain coefficient; a 64 =-k c ,a 65 =-ω 2 ,a 66 =-2live, K c =1-K d The control output takes lane departure error and driver comfort into consideration, and the controlled output z(t) is designed to be z(t)=Cx(t)+Du(t) Where u(t) represents the input of the system, 6. The human-machine shared gain scheduling method considering driver characteristics according to claim 5 is characterized in that ,S3 specifically: For the driver-vehicle-road dynamics model, the fuzzy rules are constructed as follows: The i-th fuzzy rule: If yes yes yes yes So: Where, It's the speed. It is the inverse of the vehicle speed. is the inverse of the square of the vehicle speed, is the steering gain, and represents a fuzzy set; v x ∈[v xmin ,v xmax ] are the upper and lower limits of vehicle speed, A i ,B i ,E i ,C i ,D i is the system matrix corresponding to the i-th fuzzy rule; To reduce the conservatism of the model, Taylor approximation is used to replace the variables: Among them, the longitudinal velocity v x ∈[v xmin ,v xmax ], The following membership function is used to describe the fuzzification process of the premise variables: Where v(t) is the normalized variable of vehicle speed and satisfies v(t)∈[-1,1]; K d (t) is the normalized variable of the conversion gain and satisfies K d (t)∈[K dmin ,K dmax ]; Based on the above principles, the TS fuzzy model is constructed: Among them, h i (υ(t)) represents the i-th fuzzified weighting function; h1 = m1 × l1, h2 = m1 × l2, h3 = m2 × l1 and h4 = m2 × l2, and satisfies: Ξ1=Ξ(v min ,K dmin ),Ξ2=Ξ(v min ,K dmax ), Ξ3=Ξ(v max ,K dmin ),Ξ4=Ξ(v max ,K dmax ), Ξ∈[A,B,C,D,E].

7. The human-machine shared gain scheduling method considering driver characteristics according to claim 6 is characterized in that ,S4 specifically is: S41. For the sake of generality, consider a general form of continuous-time TS fuzzy model: Where, Indicates status, represents the control input, Indicates interference, represents the control input, represents a set of real numbers, v(t) represents a premise variable, h i (v(t)) represents the membership function and satisfies h i (v(t))≥ and Using polynomials to describe the nonlinear characteristics of TS fuzzy systems, we first define the aligned polynomials: define the q-tuple in is a non-negative number that satisfies g is the order of the homogeneous polynomial; definition is a q-tuple The factorial product of S42. Design a homogeneous polynomial controller: Where K g (h) a homogeneous polynomial matrix to be determined, and Substituting the above controller into the system state equation, we get: in, S43. Use Lyapunov stability theory to prove the stability of the system: Construct the Lyapunov function V(t): V(t)=x T (t)Px(t) Where, the matrix P>0 is a positive definite matrix, and this function reflects the energy of the system state x(t); The goal of using the Lyapunov function to determine system stability is to analyze the derivative of V(t) if Then the system is stable, and we can get as follows: Where, represents the derivative of the system state x(t); While proving the stability of the system, it is also necessary to ensure that the system is resistant to external disturbances. To this end, H ∞ Performance index; if the system output z(t) and external disturbance satisfy: Where γ>0 is a given constant, then the system is said to satisfy H ∞ performance index γ; Construct a System output energy z T (t)z(t) and perturbation energy Inequality Among them, α>0 is used to prove exponential stability; through algebraic operations, the inequality is transformed into a matrix inequality, and the controller gain matrix is solved using the linear matrix inequality (LMI) technique; Finally, the exponential stability and H ∞ Satisfaction of performance indicators.

Images