An Active Enhanced Perception Method for Vehicle-Road State Parameters of a Wired Chassis System

Abstract

The present invention discloses an active enhanced perception method for vehicle-road state parameters of a wire-controlled chassis system, including: constructing a three-degree-of-freedom vehicle dynamics model; constructing a state equation and a measurement equation for estimating vehicle state parameters, and estimating vehicle state parameters in real time; constructing a road state parameter identification data set, and training a graph convolutional neural network based on the data set; outputting the road state parameter identification result in real time; obtaining the change information of vehicle-road state parameters in the future time domain; and sending the current vehicle state parameters, the current road state parameters, and the vehicle-road state parameters in the future time domain to the wire-controlled chassis control system to complete the active enhanced perception of vehicle-road state parameters by the wire-controlled chassis system. Based on the enhanced perception of vehicle-road state parameters, the present invention predicts the changes of vehicle-road state parameters in the future time domain, obtains the long-time domain prediction data of vehicle-road state parameters, and actively enhances the perception time domain range of the wire-controlled chassis for vehicle-road state parameters.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle chassis information technology, and particularly relates to an active enhanced perception method for vehicle-road state parameters of a by-wire chassis system. Background Art

[0002] With the rapid development of intelligent vehicle technology, the vehicle chassis is undergoing a major technological transformation from a traditional mechanical chassis to a by-wire chassis with electrical signals as the medium. The by-wire chassis technology is based on the traditional vehicle chassis, replacing the mechanical operating mechanism or hydraulic control components with wire connections, and converting the chassis control instructions into electrical signals through sensors, so as to achieve precise control of the vehicle. It has now become an essential execution carrier for autonomous driving vehicles.

[0003] For autonomous driving vehicles, their by-wire chassis realizes the closed-loop feedback control of the control instructions of subsystems such as steering, braking, and driving issued by the autonomous driving system by perceiving the vehicle and road states. At the same time, by perceiving the vehicle-road state, it can also judge whether the operating state of the by-wire chassis has a fault and whether there are abnormal changes in the road state, so as to adjust its own control strategy in time, which is also the basis for ensuring the normal operation and high-precision control of the chassis. However, compared with the traditional chassis system, the by-wire chassis system cancels the mechanical mechanism, has more by-wire actuators and greater control freedom, and autonomous driving vehicles have strict requirements for the control response and accuracy of the chassis. This puts higher requirements on the perception accuracy, real-time performance, and time domain range of the by-wire chassis for vehicle-road state parameters.

[0004] Among the key vehicle-road parameters used in by-wire chassis control, the yaw angular velocity, longitudinal and lateral vehicle speeds, and sideslip angle of the vehicle state parameters, as well as the road adhesion system and slope of the road state parameters, are particularly important. However, in the existing methods for perceiving vehicle-road state parameters of by-wire chassis, the dynamic time-varying tracking performance of vehicle state parameters is poor. When the vehicle state parameters have fast time-varying characteristics due to external disturbances, traditional methods are difficult to accurately track the parameter changes. For the perception of road state parameters, the existing methods require a certain road excitation to implement parameter perception, and the triggering frequency is low during actual vehicle driving, with large information lag and untimely parameter perception. In addition, for the perception of both vehicle and road state parameters, the existing methods have an obvious limitation of narrow time domain range, that is, they can only perceive the vehicle-road state parameters at the current operating moment, cannot predict the changes of key vehicle-road state parameters, and are difficult to support the forward-looking decision-making planning of the autonomous driving system and the predictive control requirements of the by-wire chassis system. There is an urgent need for an active enhanced perception method for vehicle-road state parameters of a by-wire chassis system to improve the perception accuracy, real-time performance, and time domain range of the by-wire chassis system for key vehicle-road state parameters. Summary of the Invention

[0005] Aiming at the deficiencies of the above-mentioned existing technologies, the purpose of the present invention is to provide an active enhanced perception method for vehicle-road state parameters of a wire-controlled chassis system, so as to solve the problems of low dynamic perception accuracy, perception lag and narrow perception time domain range existing in the existing chassis system parameter perception methods. On the one hand, the present invention strongly tracks and estimates vehicle state parameters, actively enhancing the perception accuracy and real-time performance of the wire-controlled chassis for vehicle state parameters. On the other hand, it uses on-vehicle image information to replace the original chassis passive excitation state information to actively identify road state parameters, actively enhancing the initiative and real-time performance of the wire-controlled chassis for perceiving road state parameters. At the same time, based on the enhanced perception of vehicle-road state parameters, the changes of vehicle-road state parameters in the future time domain are predicted to obtain long-time domain prediction data of vehicle-road state parameters, actively enhancing the perception time domain range of the wire-controlled chassis for vehicle-road state parameters.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] An active enhanced perception method for vehicle-road state parameters of a wire-controlled chassis system of the present invention comprises the following steps:

[0008] (1) Construct a three-degree-of-freedom vehicle dynamics model including vehicle yaw motion, lateral motion and longitudinal motion;

[0009] (2) Based on the vehicle dynamics model in step (1), construct a state equation and a measurement equation for estimating vehicle state parameters. The vehicle state parameters include: yaw angular velocity, sideslip angle of the center of mass, and longitudinal and lateral vehicle speeds; and use the Kalman filter method to estimate the vehicle state parameters in real time;

[0010] (3) Offline collect road image information including different adhesion coefficients and slopes, extract road image features and perform data annotation on road type and road slope data, construct a road state parameter recognition data set, and train a graph convolutional neural network based on the data set;

[0011] (4) Real-time collect road image information during vehicle operation and input it into the network model trained in step (3), and output the road state parameter recognition result in real time;

[0012] (5) Use an encoder-decoder network to perform long-time domain prediction on the current vehicle state parameters in step (2) and the current road state parameters in step (4) to obtain the change information of vehicle-road state parameters in the future time domain;

[0013] (6) Send the current vehicle state parameters in step (2), the current road state parameters in step (4), and the vehicle-road state parameters in the future time domain in step (5) to the wire-controlled chassis control system to complete the active enhanced perception of vehicle-road state parameters by the wire-controlled chassis system.

[0014] Furthermore, the modeling process of the three-degree-of-freedom vehicle dynamics model in step (1) is as follows:

[0015] (11) Combine the components of the absolute acceleration of the vehicle's center of mass in each axis of the vehicle coordinate system to obtain the vehicle's longitudinal and lateral kinematic equations:

[0016]

[0017] In the formula, a x and a y are the absolute longitudinal acceleration and the absolute lateral acceleration of the vehicle respectively; v x and v y are the longitudinal speed and the lateral speed of the vehicle in the vehicle coordinate system respectively; and are the longitudinal acceleration and the lateral acceleration of the vehicle in the vehicle coordinate system respectively; ω r is the yaw angular velocity of the vehicle;

[0018] (12) Analyze the vehicle dynamics characteristics and combine D'Alembert's principle and Newton's second law to obtain the vehicle's longitudinal dynamics equation as follows:

[0019]

[0020] In the formula, m is the total vehicle mass, F x-i is the longitudinal force acting on the vehicle in the x-axis of the vehicle coordinate system oxy; F l-i are the longitudinal tire forces of each wheel of the vehicle, where i = fl, fr, rl, rr, fl represents the left front wheel, fr represents the right front wheel, rl represents the left rear wheel, and rr represents the right rear wheel; F l-fl is the longitudinal tire force of the left front wheel, F l-fr is the longitudinal tire force of the right front wheel, F l-rl is the longitudinal tire force of the left rear wheel, F l-rr is the longitudinal tire force of the right rear wheel; F c-fl is the lateral tire force of the left front wheel, F c-fr is the lateral tire force of the right front wheel; δ f is the front wheel steering angle;

[0021] (13) According to the dynamics characteristics analysis in step (12), obtain the vehicle's lateral dynamics equation as follows:

[0022]

[0023] In the formula, F y-i is the lateral force acting on the vehicle in the y-axis of the vehicle coordinate system oxy; F c-rl is the lateral tire force of the left rear wheel; F c-rr is the lateral tire force of the right rear wheel;

[0024] (14) According to the dynamic characteristic analysis in step (12), the vehicle yaw motion equation is as follows:

[0025]

[0026] In the formula, M z-i is the yaw moment generated by the forces on each wheel of the vehicle around the z-axis of the vehicle coordinate system oxy; I z is the moment of inertia of the vehicle around the z-axis; l f and l r are the distances from the vehicle center of mass to the front axle and the rear axle respectively; B v is the vehicle track width;

[0027] (15) Construct the expression equations for the longitudinal force and lateral force of the whole vehicle as follows:

[0028]

[0029] In the formula, C cf , C cr are the cornering stiffness of the front wheels and the rear wheels respectively; α f , α r are the front wheel cornering angle and the rear wheel cornering angle respectively;

[0030] (16) Combine the equations in steps (11)-(15) to construct a three-degree-of-freedom whole vehicle dynamics model as follows:

[0031]

[0032] In the formula, is the vehicle yaw angular acceleration, β and are the vehicle center of mass side slip angle and the vehicle center of mass side slip angular acceleration respectively.

[0033] Furthermore, the step (2) specifically includes:

[0034] (21) Construct the state equation and measurement equation for vehicle state parameter estimation with the yaw angular velocity, center of mass side slip angle and longitudinal vehicle speed as state variables and the vehicle front wheel steering angle and vehicle absolute longitudinal acceleration as system input variables as follows:

[0035]

[0036] In the formula, f is the system state equation function; h is the measurement equation function; x(t) = [ω r , β, v x is the system state vector at time t; is the change rate of the system state vector at time t with respect to time; z(t) = [a yis the system measurement vector at time t; u(t) = [δ f , a x is the system input vector at time t; w(t) and v(t) are the system inherent noise and measurement noise at time t respectively; Q is the system inherent noise variance; R is the measurement noise variance; G represents the normal distribution;

[0037] (22) The equation in step (21) is discretized using the forward Euler method as follows:

[0038]

[0039] where A and B are the Jacobian matrices obtained by taking the partial derivatives of f and h with respect to the system state vector x respectively; x(k) and x(k + 1) are the system state vectors at times k and k + 1 respectively; u(k) and z(k) are the system input vector and system measurement vector at time k respectively; w(k) and v(k) are the system inherent noise and measurement noise at time k respectively;

[0040] (23) The state of the system at time k is predicted using the system state estimate at time k - 1 as follows:

[0041]

[0042] where is the predicted value of the system state at time k obtained by prediction using the system state estimate at time k - 1; is the state estimate at time k - 1;

[0043] (24) Combining the orthogonality principle and introducing an attenuation factor to calculate the system prediction covariance matrix at time k as follows:

[0044]

[0045] where P k-1 is the covariance matrix of the estimate at time k - 1; A T is the transpose matrix of matrix A; ρ k is the attenuation factor used to adjust the Kalman filter gain K k such that the estimation residual satisfies orthogonality, expressed as:

[0046]

[0047] where V k is the covariance matrix of the output difference; B T is the transpose matrix of matrix B;

[0048] (25) Solve for the attenuation factor ρ k , as follows:

[0049]

[0050] where tr is the function for calculating the trace of a matrix;

[0051] (26) Calculate the system state at time k, which is the estimated value of the vehicle state parameters, as follows:

[0052]

[0053] where is the estimated value of the vehicle state parameters at time k, including yaw rate, sideslip angle of the center of mass, and longitudinal vehicle speed. The lateral vehicle speed v y = v x β; z k is the system measurement value at time k, which is the measured value of the absolute lateral acceleration;

[0054] (27) Calculate the covariance matrix P k of the estimated value at time k, as follows:

[0055]

[0056] where I is the identity matrix.

[0057] Furthermore, step (3) specifically includes:

[0058] (31) Offline collect road images with different adhesion coefficients, perform semantic segmentation on the images to extract road features in the images. For the road adhesion coefficient μ, divide the road types into asphalt road, concrete road, gravel road, dirt road, snow-covered road, and ice-covered road, and the corresponding adhesion coefficients μ are 0.75, 0.75, 0.6, 0.55, 0.25, and 0.15 respectively; at the same time, further subdivide asphalt roads, concrete roads, gravel roads, and dirt roads, including wet and waterlogged conditions, and the corresponding adhesion coefficients μ are 0.65 for wet asphalt road, 0.6 for waterlogged asphalt road, 0.65 for wet concrete road, 0.6 for waterlogged concrete road, 0.5 for wet gravel road, 0.45 for waterlogged gravel road, 0.45 for wet dirt road, 0.3 for waterlogged dirt road. Denote wet asphalt road as type 1, waterlogged asphalt road as type 2, wet concrete road as type 3, waterlogged concrete road as type 4, wet gravel road as type 5, waterlogged gravel road as type 6, wet dirt road as type 7, waterlogged dirt road as type 8, snow-covered road as type 9, and ice-covered road as type 10, and construct a dataset containing road images, road adhesion coefficient annotations, and road types;

[0059] (32) Offline collect road images with different slopes, perform semantic segmentation on the images to extract road features in the images, and divide the road slopes into lateral slope s h and longitudinal slope sl Annotate the road slope s of the image in degrees, and construct a dataset containing different road images and road slope annotations;

[0060] (33) Use the ResNet network to identify the road adhesion coefficient, and connect a fully connected layer and a Softmax layer after the network. Use the dataset constructed in step (31) as the training set, with 80% of the data as the training set and the remaining 20% of the data as the validation set. Use the road image data as the input and the road type as the output to train the graph convolutional neural network. The trained network model is denoted as the adhesion coefficient recognition network model;

[0061] (34) Use the lightweight MobileNetv3 network to identify the road slope, and connect a fully connected layer after the network. Use the dataset constructed in step (32) as the training set, with 80% of the data as the training set and the remaining 20% of the data as the validation set. Use the road image data as the input and the road lateral slope and longitudinal slope as the output to train the graph convolutional neural network. The trained network model is denoted as the road slope recognition network model.

[0062] Further, the specific content of step (4) includes:

[0063] (41) Real-time collect the road image information in front of the vehicle during operation;

[0064] (42) Input the road image information collected in step (41) into the trained adhesion coefficient recognition network model and road slope recognition network model respectively, and output the road type and road horizontal and vertical slope information s h 、s l ;

[0065] (43) Based on the road type and road horizontal and vertical slope information s h 、s l output in step (42), convert the road type into the corresponding road adhesion coefficient μ information, and output the real-time road adhesion coefficient and road horizontal and vertical slope information in real time.

[0066] Further, the specific content of step (5) includes:

[0067] (51) Select a bidirectional GRU network as the encoding network and a GRU network as the decoding network, and perform offline training. The encoding network uses the historical data sequence of the vehicle-road state parameters with a length of 2 seconds and a sampling interval of 0.1 second as the network input, expressed as:

[0068]

[0069] In the formula, N inis the network input, which represents the historical sequence of vehicle-road state parameters; S i is the vehicle-road state parameter vector corresponding to the i-th moment. Each element in the vector represents the value of the corresponding vehicle-road state parameter at the i-th moment, where i = k - 19, k - 18, …, k; ω r,i , β i , v x,i , v y,i , μ i , s h,i , s l,i are respectively the vehicle yaw angular velocity, the centroid side slip angle, the vehicle longitudinal velocity in the vehicle coordinate system, the vehicle lateral velocity in the vehicle coordinate system, the road adhesion coefficient, the road cross slope, and the road longitudinal slope at the i-th moment;

[0070] (52) After the historical data sequence of vehicle-road state parameters is input into the bidirectional GRU encoding network, the final output result of the bidirectional GRU encoding network is as follows:

[0071]

[0072] In the formula, h is the hidden state output by the bidirectional GRU encoding network; hf i is the hidden state output by the forward GRU encoding network in the bidirectional GRU encoding network at the i-th moment; hb i is the hidden state output by the backward GRU encoding network in the bidirectional GRU encoding network at the i-th moment;

[0073] (53) Perform time encoding on the hidden state obtained in step (52) to obtain the time encoding vector as follows:

[0074]

[0075] In the formula, Ch k is the time encoding vector at the k-th moment; w i,H is the weight of the encoding network hidden state at the i-th moment; α i,H is the correlation score between the encoding network hidden state at the i-th moment and the decoding network hidden state H k-1 at the k - 1 moment; V α , W α and U α are the weight matrices of the scoring network; h i is the hidden state output by the bidirectional GRU encoding network at the i-th moment;

[0076] (54) Perform feature encoding on all feature vectors of the hidden state obtained in step (52) to obtain the feature encoding vector as follows:

[0077]

[0078] Where Cf k is the feature encoding vector at time k; d is the dimension of the hidden state feature vector; τ p,H is the weight of the p-th dimensional feature vector; is the correlation score between the p-th dimensional feature vector and the decoder hidden state H k-1 at time k-1; and are the weight matrices of the scoring network; f p is the feature vector of the p-th dimension, which represents the numerical combination of the p-th dimension of all hidden states;

[0079] (55) Concatenate the time encoding vector and the feature encoding vector to obtain the context vector C k output by the encoding network at time k, as follows:

[0080] C k = [Ch k , Cf k (19);

[0081] (56) Input the context vector obtained in step (55) into the GRU decoding network, and cyclically calculate the prediction results of the vehicle-road state parameters at time k+n to obtain the prediction data of the vehicle-road state parameters within n / 10 seconds, completing the long-time domain prediction of the vehicle-road state parameters. The cyclic calculation formula is as follows:

[0082]

[0083] Where rd k+n and zd k+n are the output of the reset gate and the update gate of the GRU decoding network at time k+n respectively; W rd , U rd and b rd are the input weight matrix, state weight matrix and bias matrix of the reset gate; W zd , U zd and b zd are the input weight matrix, state weight matrix and bias matrix of the update gate; W H , U H and b H are the input weight matrix, state weight matrix and bias matrix when calculating the candidate hidden state; is the candidate hidden state at time k+n; W o is the weight matrix of the fully connected layer output by the GRU network; n is the prediction time domain.

[0084] Advantages of the present invention:

[0085] By strongly tracking and estimating vehicle state parameters, the present invention can actively enhance the perception accuracy and real-time performance of the drive-by-wire chassis for vehicle state parameters; by using on-vehicle image information to replace the original passive excitation state information of the chassis to actively identify road state parameters, the initiative and real-time performance of the drive-by-wire chassis for perceiving road state parameters are actively enhanced.

[0086] Based on enhancing the accuracy and real-time performance of vehicle-road state perception, the present invention predicts the changes of vehicle-road state parameters in the future time domain, and long-time domain prediction data of vehicle-road state parameters can be obtained, actively enhancing the perception time domain range of the drive-by-wire chassis for vehicle-road state parameters.

[0087] The present invention can effectively improve the perception accuracy, real-time performance and time domain range of the drive-by-wire chassis system for vehicle-road states, endow the drive-by-wire chassis system with stronger perception ability, effectively improve the chassis information acquisition ability, and further support the development of intelligent functions of the drive-by-wire chassis, having strong practicality and engineering significance. Brief Description of the Drawings

[0088] Figure 1 is a flowchart of the method of the present invention;

[0089] Figure 2 is a schematic diagram of a three-degree-of-freedom vehicle dynamics model constructed in the present invention. Detailed Embodiment

[0090] For the convenience of those skilled in the art, the present invention will be further described below in conjunction with embodiments and the accompanying drawings. The content mentioned in the embodiments does not limit the present invention.

[0091] Refer to Figure 1 、 Figure 2 As shown, a method for actively enhancing the perception of vehicle-road state parameters of a drive-by-wire chassis system of the present invention is as follows:

[0092] (1) Construct a three-degree-of-freedom vehicle dynamics model including vehicle yaw motion, lateral motion and longitudinal motion;

[0093] The modeling process of the three-degree-of-freedom vehicle dynamics model is as follows:

[0094] (11) Combine the components of the absolute acceleration of the vehicle center of mass on each axis of the vehicle coordinate system to obtain the vehicle longitudinal and lateral kinematic equations:

[0095]

[0096] In the formula, a x and a y are the absolute longitudinal acceleration and the absolute lateral acceleration of the vehicle respectively; v x and v yThey are respectively the longitudinal speed and lateral speed of the vehicle in the vehicle coordinate system; and

[0097] They are respectively the longitudinal acceleration and lateral acceleration of the vehicle in the vehicle coordinate system; ω r is the yaw angular velocity of the vehicle;

[0098] (12) Analyze the vehicle dynamics characteristics and combine with D'Alembert's principle and Newton's second law to obtain the vehicle longitudinal dynamics equation as follows:

[0099]

[0100] In the formula, m is the vehicle mass, F x-i is the longitudinal force exerted on the vehicle on the x-axis of the vehicle coordinate system oxy; F l-i is the longitudinal tire force of each wheel of the vehicle, i = fl, fr, rl, rr, fl represents the left front wheel, fr represents the right front wheel, rl represents the left rear wheel, rr represents the right rear wheel; F l-fl is the longitudinal tire force of the left front wheel, F l-fr is the longitudinal tire force of the right front wheel, F l-rl is the longitudinal tire force of the left rear wheel, F l-rr is the longitudinal tire force of the right rear wheel; F c-fl is the lateral tire force of the left front wheel, F c-fr is the lateral tire force of the right front wheel; δ f is the front wheel steering angle;

[0101] (13) According to the dynamics characteristics analysis in step (12), obtain the vehicle lateral dynamics equation as follows:

[0102]

[0103] In the formula, F y-i is the lateral force exerted on the vehicle on the y-axis of the vehicle coordinate system oxy; F c-rl is the lateral tire force of the left rear wheel; F c-rr is the lateral tire force of the right rear wheel;

[0104] (14) According to the dynamics characteristics analysis in step (12), obtain the vehicle yaw motion equation as follows:

[0105]

[0106] In the formula, M z-i is the yaw moment generated by the forces on each wheel of the vehicle around the z-axis of the vehicle coordinate system oxy; I z is the moment of inertia of the vehicle around the z-axis; l f and l rThe distances from the vehicle's center of mass to the front axle and the rear axle, respectively; B v is the vehicle's track width;

[0107] (15) Construct the longitudinal and lateral force expression equations of the whole vehicle as follows:

[0108]

[0109] In the formula, C cf , C cr are the cornering stiffness of the front wheels and the cornering stiffness of the rear wheels, respectively; α f , α r are the front wheel cornering angle and the rear wheel cornering angle, respectively;

[0110] (16) Combine the equations in steps (11)-(15) to construct a three-degree-of-freedom whole vehicle dynamics model as follows:

[0111]

[0112] In the formula, is the vehicle's yaw angular acceleration, β and are the vehicle's center-of-mass side slip angle and the center-of-mass side slip angular acceleration, respectively.

[0113] (2) Based on the whole vehicle dynamics model in step (1), construct a state equation and a measurement equation for vehicle state parameter estimation. The vehicle state parameters include: yaw angular velocity, center-of-mass side slip angle, and longitudinal and lateral vehicle speeds; and use the Kalman filter method to estimate the vehicle state parameters in real time; specifically including:

[0114] (21) Use the yaw angular velocity, center-of-mass side slip angle, and longitudinal vehicle speed as state variables, and the vehicle's front wheel steering angle and the vehicle's absolute longitudinal acceleration as system input variables to construct a state equation and a measurement equation for vehicle state parameter estimation as follows:

[0115]

[0116] In the formula, f is the system state equation function; h is the measurement equation function; x(t) = [ω r , β, v x is the system state vector at time t; is the rate of change of the system state vector with time at time t; z(t) = [a y is the system measurement vector at time t; u(t) = [δ f , a x is the system input vector at time t; w(t) and v(t) are the system inherent noise and measurement noise at time t, respectively; Q is the system inherent noise variance; R is the measurement noise variance; G represents a normal distribution;

[0117] (22) Discretize the equation in step (21) using the forward Euler method as follows:

[0118]

[0119] Where A and B are the Jacobian matrices obtained by taking the partial derivatives of f and h with respect to the system state vector x respectively; x(k) and x(k + 1) are the system state vectors at times k and k + 1 respectively; u(k) and z(k) are the system input vector and the system measurement vector at time k respectively; w(k) and v(k) are the system inherent noise and the measurement noise at time k respectively.

[0120] (23) Predict the state of the system at time k using the system state estimate at time k - 1 as follows:

[0121]

[0122] Where is the predicted value of the system state at time k obtained by prediction using the system state estimate at time k - 1; is the state estimate at time k - 1;

[0123] (24) Combine the orthogonality principle and introduce a decay factor to calculate the system prediction covariance matrix at time k as follows:

[0124]

[0125] Where P k-1 is the covariance matrix of the estimate at time k - 1; A T is the transpose matrix of matrix A; ρ k is the decay factor used to adjust the Kalman filter gain K k such that the estimation residual satisfies orthogonality, expressed as:

[0126]

[0127] Where V k is the covariance matrix of the output difference; B T is the transpose matrix of matrix B;

[0128] (25) Solve for the decay factor ρ k , as follows:

[0129]

[0130] Where tr is the function to find the trace of a matrix;

[0131] (26) Calculate the system state at time k, which is the estimated value of the vehicle state parameter, as follows:

[0132]

[0133] wherein, is the estimated value of the vehicle state parameters at time k, including yaw rate, sideslip angle of the center of mass, and longitudinal vehicle speed. The lateral vehicle speed v y = v x β; z k is the system measurement value at time k, which is the measured value of the absolute lateral acceleration;

[0134] (27) Calculate the covariance matrix P k of the estimated value at time k as follows:

[0135]

[0136] wherein, I is the identity matrix.

[0137] (3) Offline collect road image information including different adhesion coefficients and slopes, extract road image features, perform data annotation on road type and road slope, construct a road state parameter recognition data set, and train a graph convolutional neural network based on the data set; specifically including:

[0138] (31) Offline collect road images with different adhesion coefficients, perform semantic segmentation on the images to extract road features in the images. For the road adhesion coefficient μ, divide the road types into asphalt road, concrete road, gravel road, dirt road, snow-covered road, and ice-covered road, and the corresponding adhesion coefficients μ are 0.75, 0.75, 0.6, 0.55, 0.25, and 0.15 respectively; at the same time, further subdivide asphalt roads, concrete roads, stone roads, and dirt roads, including wet and waterlogged conditions, and the corresponding adhesion coefficients μ are 0.65 for wet asphalt road, 0.6 for waterlogged asphalt road, 0.65 for wet concrete road, 0.6 for waterlogged concrete road, 0.5 for wet gravel road, 0.45 for waterlogged gravel road, 0.45 for wet dirt road, 0.3 for waterlogged dirt road, and record wet asphalt road as type 1, waterlogged asphalt road as type 2, wet concrete road as type 3, waterlogged concrete road as type 4, wet gravel road as type 5, waterlogged gravel road as type 6, wet dirt road as type 7, waterlogged dirt road as type 8, snow-covered road as type 9, and ice-covered road as type 10, and construct a data set including road images, road adhesion coefficient annotations, and road types;

[0139] (32) Offline collect road images with different slopes, perform semantic segmentation on the images to extract road features in the images, divide the road slopes into lateral slope s h and longitudinal slope s l , perform road slope s annotation on the images in degrees, and construct a data set including different road images and road slope annotations;

[0140] (33) The ResNet network is used to identify the road adhesion coefficient, and a fully connected layer and a Softmax layer are connected after the network. Using the data set constructed in the step (31) as the training set, 80% of the data is used as the training set, and the remaining 20% of the data is used as the validation set. Using the road image data as the input and the road type as the output, the graph convolutional neural network is trained, and the trained network model is denoted as the adhesion coefficient recognition network model;

[0141] (34) The lightweight MobileNetv3 network is used to identify the road slope, and a fully connected layer is connected after the network. Using the data set constructed in the step (32) as the training set, 80% of the data is used as the training set, and the remaining 20% of the data is used as the validation set. Using the road image data as the input and the road transverse slope and longitudinal slope as the output, the graph convolutional neural network is trained, and the trained network model is denoted as the road slope recognition network model.

[0142] (4) Real-time collect the road image information during vehicle operation and input it into the network model trained in step (3), and output the road state parameter recognition result in real time; specifically includes:

[0143] (41) Real-time collect the road image information in front of the vehicle during vehicle operation;

[0144] (42) Input the road image information collected in step (41) into the trained adhesion coefficient recognition network model and road slope recognition network model respectively, and output the road type and road transverse and longitudinal slope information s h 、s l ;

[0145] (43) Based on the road type and road transverse and longitudinal slope information s h 、s l output in step (42), convert the road type into the corresponding road adhesion coefficient μ information, and output the real-time road adhesion coefficient and road transverse and longitudinal slope information in real time.

[0146] (5) Use an encoder-decoder network to perform long-time domain prediction on the current vehicle state parameters in step (2) and the current road state parameters in step (4), and obtain the change information of the vehicle-road state parameters in the future time domain; specifically includes:

[0147] (51) Select a bidirectional GRU network as the encoder network and a GRU network as the decoder network, and perform offline training. The encoder network uses the historical data sequence of the vehicle-road state parameters with a length of 2 seconds and a sampling interval of 0.1 second as the network input, which is expressed as:

[0148]

[0149] Where N in is the network input, representing the historical sequence of vehicle-road state parameters; S i is the vehicle-road state parameter vector corresponding to the i-th moment. Each element in the vector represents the value of the corresponding vehicle-road state parameter at the i-th moment, where i = k - 19, k - 18, …, k; ω r,i , β i , v x,i , v y,i , μ i , s h,i , s l,i are respectively the vehicle yaw angular velocity, the centroid side slip angle, the vehicle longitudinal velocity in the vehicle coordinate system, the vehicle lateral velocity in the vehicle coordinate system, the road adhesion coefficient, the road cross slope, and the road longitudinal slope at the i-th moment;

[0150] (52) After the historical data sequence of vehicle-road state parameters is input into the bidirectional GRU encoding network, the final output result of the bidirectional GRU encoding network is as follows:

[0151]

[0152] Where h is the hidden state output by the bidirectional GRU encoding network; hf i is the hidden state output by the forward GRU encoding network in the bidirectional GRU encoding network at the i-th moment; hb i is the hidden state output by the backward GRU encoding network in the bidirectional GRU encoding network at the i-th moment;

[0153] (53) Perform time encoding on the hidden state obtained in step (52) to obtain the time encoding vector as follows:

[0154]

[0155] Where Ch k is the time encoding vector at the k-th moment; w i,H is the weight of the hidden state of the encoding network at the i-th moment; α i,H is the correlation score between the hidden state of the encoding network at the i-th moment and the hidden state H k-1 of the decoding network at the k - 1-th moment; V α , W α and U α are the weight matrices of the scoring network; h i is the hidden state output by the bidirectional GRU encoding network at the i-th moment;

[0156] (54) Perform feature encoding on all feature vectors of the hidden state obtained in step (52) to obtain the feature encoding vector as follows:

[0157]

[0158] Wherein, Cf k is the feature encoding vector at time k; d is the dimension of the hidden state feature vector; τ p,H is the weight of the p-th dimensional feature vector; is the correlation score between the p-th dimensional feature vector and the decoder hidden state H k-1 at time k-1; and are the weight matrices of the scoring network; f p is the feature vector of the p-th dimension, which represents the numerical combination of the p-th dimension of all hidden states;

[0159] (55) Concatenate the time encoding vector and the feature encoding vector to obtain the context vector C output by the encoding network at time k k , as follows:

[0160] C k =[Ch k , Cf k (19);

[0161] (56) Input the context vector obtained in step (55) into the GRU decoding network, and cyclically calculate the prediction results of the vehicle-road state parameters at time k+n to obtain the prediction data of the vehicle-road state parameters within n / 10 seconds, completing the long-time domain prediction of the vehicle-road state parameters. The cyclic calculation formula is as follows:

[0162]

[0163] Wherein, rd k+n and zd k+n are the output of the reset gate and the update gate of the GRU decoding network at time k+n respectively; W rd , U rd and b rd are the input weight matrix, state weight matrix and bias matrix of the reset gate; W zd , U zd and b zd are the input weight matrix, state weight matrix and bias matrix of the update gate; W H , U H and b H are the input weight matrix, state weight matrix and bias matrix when calculating the candidate hidden state; is the candidate hidden state at time k+n; W o is the weight matrix of the fully connected layer output by the GRU network; n is the prediction time domain.

[0164] (6) Send the current vehicle state parameters in step (2), the current road state parameters in step (4), and the vehicle-road state parameters in the future time domain in step (5) to the drive-by-wire chassis control system, so as to complete the active enhanced perception of the vehicle-road state parameters by the drive-by-wire chassis system.

[0165] The specific application scenarios of the present invention are numerous. The above description is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements can be made, and these improvements should also be regarded as the protection scope of the present invention.

Claims

1. An active enhanced perception method for vehicle-road state parameters of a steer-by-wire chassis system, characterized in that, The steps are as follows: (1) Construct a three-degree-of-freedom vehicle dynamics model that includes vehicle yaw motion, lateral motion, and longitudinal motion; (2) Based on the vehicle dynamics model in step (1), construct a state equation and a measurement equation for estimating vehicle state parameters. The vehicle state parameters include: yaw angular velocity, center-of-mass sideslip angle, and longitudinal and lateral vehicle speeds; and use the Kalman filtering method to estimate the vehicle state parameters in real time; (3) Offline collect road image information including different adhesion coefficients and slopes, extract road image features, perform data annotation on road type and road slope data, construct a road state parameter recognition data set, and train a graph convolutional neural network based on the data set; (4) Collect road image information during vehicle operation in real time and input it into the network model trained in step (3) to output the road state parameter recognition result in real time; (5) Use an encoder-decoder network to perform long-time domain prediction on the current vehicle state parameters in step (2) and the current road state parameters in step (4) to obtain the change information of vehicle-road state parameters in the future time domain; (6) Send the current vehicle state parameters in step (2), the current road state parameters in step (4), and the vehicle-road state parameters in the future time domain in step (5) to the by-wire chassis control system to complete the active enhanced perception of vehicle-road state parameters by the by-wire chassis system; (3) The specific steps of step (3) include: (31) Offline collect road images with different adhesion coefficients, perform semantic segmentation on the images to extract road features in the images. For the road adhesion coefficient μ, divide the road types into asphalt road, concrete road, gravel road, dirt road, snow-covered road, and ice-covered road, and the corresponding adhesion coefficients μ are 0.75, 0.75, 0.6, 0.55, 0.25, and 0.15 respectively; at the same time, further subdivide asphalt roads, concrete roads, stone roads, and dirt roads, including wet and waterlogged conditions, and the corresponding adhesion coefficients μ are 0.65 for wet asphalt road, 0.6 for waterlogged asphalt road, 0.65 for wet concrete road, 0.6 for waterlogged concrete road, 0.5 for wet gravel road, 0.45 for waterlogged gravel road, 0.45 for wet dirt road, 0.3 for waterlogged dirt road, record wet asphalt road as type 1, waterlogged asphalt road as type 2, wet concrete road as type 3, waterlogged concrete road as type 4, wet gravel road as type 5, waterlogged gravel road as type 6, wet dirt road as type 7, waterlogged dirt road as type 8, snow-covered road as type 9, and ice-covered road as type 10, and construct a data set including road images, road adhesion coefficient annotations, and road types; (32)Collect road images with different slopes offline, perform semantic segmentation on the images to extract road features in the images, and divide the road slopes into transverse slope s h and longitudinal slope s l , label the road slope s of the image in degrees, and construct a dataset containing different road images and road slope labels; (33) Use the ResNet network to identify the road adhesion coefficient, connect a fully connected layer and a Softmax layer after the network, use the data set constructed in step (31) as the training set, use 80% of the data as the training set, and the remaining 20% of the data as the validation set. Use the road image data as the input and the road type as the output to train the graph convolutional neural network. The trained network model is denoted as the adhesion coefficient recognition network model; (34) The lightweight MobileNetv3 network is used to identify the road slope, and a fully connected layer is connected after the network. Using the dataset constructed in the step (32) as the training set, 80% of the data is used as the training set, and the remaining 20% of the data is used as the validation set. Taking the road image data as the input and the lateral slope and longitudinal slope of the road as the output, the graph convolutional neural network is trained, and the trained network model is denoted as the road slope identification network model.

2. The active enhanced perception method for vehicle-road state parameters of the wire-controlled chassis system according to claim 1, characterized in that The modeling process of the three-degree-of-freedom vehicle dynamics model in the step (1) is as follows: (11) Combining the components of the absolute acceleration of the vehicle's center of mass on each axis of the vehicle coordinate system to obtain the vehicle's longitudinal and lateral kinematic equations as follows: Where a x and a y are the absolute longitudinal acceleration and the absolute lateral acceleration of the vehicle, respectively; v x and v y are the longitudinal speed and the lateral speed of the vehicle in the vehicle coordinate system, respectively; and are the longitudinal acceleration and the lateral acceleration of the vehicle in the vehicle coordinate system, respectively; ω r is the yaw angular velocity of the vehicle. (12) Analyzing the vehicle's dynamic characteristics and combining D'Alembert's theorem and Newton's second law to obtain the vehicle's longitudinal dynamic equation as follows: where m is the vehicle mass, F x-i is the longitudinal force acting on the vehicle in the x-axis of the vehicle coordinate system oxy; F l-i is the longitudinal tire force of each wheel of the vehicle, i = fl, fr, rl, rr, fl represents the left front wheel, fr represents the right front wheel, rl represents the left rear wheel, rr represents the right rear wheel; F l-fl is the longitudinal tire force of the left front wheel, F l-fr is the longitudinal tire force of the right front wheel, F l-rl is the longitudinal tire force of the left rear wheel, F l-rr is the longitudinal tire force of the right rear wheel; F c-fl is the lateral tire force of the left front wheel, F c-fr is the lateral tire force of the right front wheel; δ f is the front wheel steering angle; (13) According to the dynamic characteristic analysis in step (12), the vehicle's lateral dynamic equation is obtained as follows: where F y-i is the lateral force acting on the vehicle in the y-axis direction of the vehicle coordinate system oxy; F c-rl is the lateral force of the left rear wheel tire; F c-rr is the lateral force of the right rear wheel tire; (14) According to the dynamic characteristic analysis in step (12), the vehicle's yaw motion equation is obtained as follows: Where M z-i is the yaw moment generated by the forces on each wheel of the vehicle about the z-axis of the vehicle coordinate system oxy; I z is the moment of inertia of the vehicle about the z-axis; l f and l r are the distances from the vehicle's center of mass to the front axle and the rear axle respectively; B v is the vehicle track width; (15) Construct the vehicle's longitudinal force and lateral force expression equations as follows: where C cf and C cr are the cornering stiffnesses of the front and rear wheels respectively; α f and α r are the front and rear wheel cornering angles respectively; (16) Combining the equations in steps (11)-(15) to construct the three-degree-of-freedom vehicle dynamics model as follows: where is the vehicle yaw angular acceleration, β and are the sideslip angle and sideslip angular acceleration of the vehicle center of mass, respectively.

3. The active enhanced perception method for vehicle-road state parameters of the steer-by-wire chassis system according to claim 2, wherein The step (2) specifically includes: (21) Taking the yaw angular velocity, sideslip angle of the center of mass, and longitudinal vehicle speed as state variables, and taking the vehicle's front wheel steering angle and the vehicle's absolute longitudinal acceleration as system input variables to construct the state equation and measurement equation for vehicle state parameter estimation as follows: where \(f\) is the system state equation function; \(h\) is the measurement equation function; \(x(t)=[\omega r ,\beta,v x \) is the system state vector at time \(t\); is the rate of change of the system state vector with respect to time at time \(t\); \(z(t)=[a y \) is the system measurement vector at time \(t\); \(u(t)=[\delta f ,a x \) is the system input vector at time \(t\); \(w(t)\) and \(v(t)\) are the system inherent noise and measurement noise at time \(t\), respectively; \(Q\) is the system inherent noise variance; \(R\) is the measurement noise variance; \(G\) represents the normal distribution; (22) Using the forward Euler method to discretize the equations in step (21) as follows: In the formula, A and B are the Jacobian matrices obtained by taking the partial derivatives of f and h with respect to the system state vector x; x(k) and x(k + 1) are the system state vectors at times k and k + 1 respectively; u(k) and z(k) are the system input vector and system measurement vector at time k respectively; w(k) and v(k) are the system inherent noise and measurement noise at time k respectively; (23) Using the system state estimate value at time k - 1 to predict the state of the system at time k as follows: In the formula, is the predicted system state value at time k obtained by predicting using the system state estimation value at time k-1; is the state estimation value at time k-1; (24) Calculate the system prediction covariance matrix at time k by combining the orthogonality principle and introducing an attenuation factor as follows: where P k-1 is the covariance matrix of the estimated value at time k - 1; A T is the transpose matrix of matrix A; ρ k is the attenuation factor used to adjust the Kalman filter gain K k such that the estimation residual satisfies orthogonality, expressed as: where, V k is the covariance matrix of the output difference; B T is the transpose matrix of matrix B; (25) Solve for the attenuation factor ρ k as follows: In the formula, tr is the function for finding the trace of a matrix; (26) Calculate the system state at time k, which is the estimated value of the vehicle state parameter, as follows: Wherein, is the estimated value of the vehicle state parameters at time k, including yaw rate, sideslip angle of the center of mass, and longitudinal vehicle speed. The lateral vehicle speed v y = v x β; z k is the system measurement value at time k, which is the measured value of the absolute lateral acceleration; (27) Calculate the covariance matrix P of the estimated value at time k k , as follows: In the formula, I is the identity matrix.

4. The active enhanced perception method for vehicle-road state parameters of the steer-by-wire chassis system according to claim 1, characterized in that The specific content of the step (4) includes: (41) Real-time collect the road image information in front of the vehicle during operation; (42) Input the road image information collected in step (41) into the trained adhesion coefficient recognition network model and road slope recognition network model respectively, and output the road type and the longitudinal and transverse slope information s of the road h and s l ; (43) Based on the output road type and the longitudinal and transverse road slope information s h and s l , convert the road type into the corresponding road adhesion coefficient μ information, and output the obtained real-time road adhesion coefficient and the longitudinal and transverse road slope information in real time.

5. The active enhanced perception method for vehicle-road state parameters of the steer-by-wire chassis system according to claim 4, characterized in that, The step (5) specifically includes: (51) Select the bidirectional GRU network as the encoding network and the GRU network as the decoding network, and perform offline training. The encoding network takes the historical data sequence of the vehicle-road state parameters with a length of 2 seconds and a sampling interval of 0.1 second as the network input, expressed as: where N in is the network input, representing the historical sequence of vehicle-road state parameters; S i is the vehicle-road state parameter vector corresponding to the i-th moment, and each element in the vector represents the value of the corresponding vehicle-road state parameter at the i-th moment, where i = k - 19, k - 18, …, k; ω r,i , β i , v x,i , v y,i , μ i , s h,i , s l,i are respectively the vehicle yaw angular velocity, the center of mass side slip angle, the vehicle longitudinal velocity in the vehicle coordinate system, the vehicle lateral velocity in the vehicle coordinate system, the road adhesion coefficient, the road cross slope, and the road longitudinal slope at the i-th moment; (52) Calculate the final output result of the bidirectional GRU encoding network after the historical data sequence of the vehicle-road state parameters is input into the bidirectional GRU encoding network as follows: Where h is the hidden state output by the bidirectional GRU encoding network; hf i is the hidden state output by the forward GRU encoding network in the bidirectional GRU encoding network at time i; hb i is the hidden state output by the backward GRU encoding network in the bidirectional GRU encoding network at time i; (53) Perform time encoding on the hidden state obtained in step (52) to obtain the time encoding vector as follows: α i,H = V α ·tanh(W α ·h i + U α ·H k-1 ) Where, Ch k is the time encoding vector at time k; w i,H is the weight of the hidden state of the encoding network at the i-th time; α i,H is the correlation score between the hidden state of the encoding network at the i-th time and the hidden state H k-1 of the decoding network at time k-1; V α , W α and U α are the weight matrices of the scoring network; h i is the hidden state output by the bidirectional GRU encoding network at the i-th time; (54) Perform feature encoding on all the feature vectors of the hidden state obtained in step (52) to obtain the feature encoding vector as follows: where, Cf k is the feature encoding vector at time k; d is the dimension of the hidden state feature vector; τ p,H is the weight of the p-th dimensional feature vector; l p,H is the correlation score between the p-th dimensional feature vector and the decoder hidden state H k-1 at time k-1; V l , W l and U l are the weight matrices of the scoring network; f p is the feature vector of the p-th dimension, which represents the numerical combination of the p-th dimension of all hidden states; (55) Concatenate the time-coded vector and the feature-coded vector to obtain the context vector C output by the coding network at time k k , as follows: C k = [Ch k , Cf k (19); (56) Input the context vector obtained in step (55) into the GRU decoding network, cyclically calculate the prediction results of the vehicle-road state parameters at the k + n moment, obtain the prediction data of the vehicle-road state parameters within n / 10 seconds, and complete the long-time domain prediction of the vehicle-road state parameters. The cyclic calculation formula is as follows: where, rd k+n and zd k+n are the output of the reset gate and the update gate of the GRU decoding network at time k + n respectively; W rd , U rd and b rd are the input weight matrix, the state weight matrix and the bias matrix of the reset gate; W zd , U zd and b zd are the input weight matrix, the state weight matrix and the bias matrix of the update gate; W H , U H and b H are the input weight matrix, the state weight matrix and the bias matrix when calculating the candidate hidden state; is the candidate hidden state at time k + n; W o is the weight matrix of the fully connected layer of the GRU network output; n is the prediction time domain.

Images