Multi-axis special vehicle state estimation method based on neural network and unscented Kalman filter

Abstract

The present invention discloses a multi-axis special vehicle state estimation method based on neural network and unscented Kalman filter, including Step 1: First, use sensors to collect vehicle signal data to form a vehicle data signal packet; Step 2: Use a neural network module to process the sensor signal data to estimate the pseudo-central lateral deviation angle and the pseudo-roll angle; Step 3: Input the pseudo-central lateral deviation angle and the pseudo-roll angle obtained by the neural network module as "pseudo-measurements" into the unscented Kalman module, and finally use the unscented Kalman module to obtain the central lateral deviation angle and the roll angle of the multi-axis special vehicle; This method estimates the two main control parameters, namely the central lateral deviation angle and the roll angle of the multi-axis special vehicle, simultaneously by using two algorithms of neural network and unscented Kalman filter, can effectively solve the problems of strong nonlinear systems, and has the characteristics of high algorithm accuracy, wide application range and good stability.

Description

Technical Field

[0001] The present invention relates to the technical field of special vehicle positioning, and particularly relates to a multi-axis special vehicle state estimation method based on a neural network and an unscented Kalman filter. Background Art

[0002] With the advancement of military requirements, the driving environment of special vehicles is complex and changeable. Coupled with the characteristics of large load and high center of mass of special vehicles, accidents such as vehicle handling instability are likely to occur during maneuvering. In order to improve vehicle handling stability and safety, a large number of advanced vehicle active safety control systems have been developed and applied. However, the actual effect and potential performance of vehicle active safety control systems largely depend on the real-time understanding of the basic state information of the vehicle, such as the sideslip angle and roll angle of the vehicle center of mass. Due to technical and economic reasons, it is difficult to directly measure these information in standard vehicles, and in a complex driving environment, the signals measured by GPS may be lost. Using limited sensors and based on dynamic models and parameter estimation methods to obtain state parameters that are as close to the actual situation as possible is an economical and effective method.

[0003] In recent years, different types of methods based on vehicle dynamics models and different estimation strategies have been developed and studied; one is the model-based estimation method, and the other is the data-driven estimation method; the estimation methods based on vehicle models include kinematic and dynamic models; kinematic-based estimation is a direct integration method for vehicle state estimation. Due to reasons such as offset errors, after long-term integration, the cumulative error will continuously increase, and finally the estimated vehicle state will be extremely inaccurate; considering the limitations of kinematic-based estimation methods, most studies tend to focus on estimation based on vehicle dynamics models. The estimation technology based on vehicle dynamics models uses mathematical models to describe the transient behavior in vehicle system dynamics and has high state estimation accuracy; in dynamic-based estimation, it includes Kalman filter and its improved algorithms, least squares method, Romberg observer, robust observer, sliding mode observer, etc.; for example, the article "NAM K, OH S, FUJIMOTO H, et al. Estimation of Sideslip and Roll Angles of Electric Vehicles Using Lateral Tire Force Sensors Through RLS and Kalman Filter Approaches[J]. Industrial Electronics, IEEE Transactions on, 2013" proposed a method for estimating the center-of-mass sideslip angle and roll angle using real-time tire lateral force measurements obtained from multi-sensing wheel hub units. In the estimation of the vehicle center-of-mass sideslip angle, based on a linear vehicle model and sensor measurement data, a recursive least squares algorithm with a forgetting factor was proposed. In the estimation of the roll angle, the Kalman filter was designed by integrating available sensor measurements and roll dynamics. Experiments have shown that compared with traditional methods, the estimation accuracy has been improved by more than 50%; the article "LI X, CHAN C, WANG Y. A Reliable Fusion Methodology for Simultaneous Estimation of Vehicle Sideslip and Yaw Angles[J]."IEEE Transactions on Vehicular Technology, 2015" proposed a fusion method that integrates a single-frequency dual-antenna global positioning system (DA-GPS) with other low-cost vehicle sensors to achieve reliable estimation of the vehicle's center-of-mass sideslip angle and roll angle. First, through the interaction of two parallel extended Kalman filters, the vehicle roll angle is accurately estimated, and based on vehicle kinematics, a global federated estimator (GFE) is designed on the basis of the federated filtering algorithm. The overall results show that this method can provide reliable estimations of the center-of-mass sideslip angle and roll angle under a wide range of driving conditions; the article "JIN X, YIN G. Estimation of lateral tire-road Forces and sideslip Angle for electric Vehicles using interacting multiple model filter approach[J]. Journal of the Franklin Institute, 2014" proposed a method for estimating the center-of-mass sideslip angle and roll angle using real-time measurements. This estimation method is based on an interactive multi-model (IMM) filter that integrates in-vehicle sensors of wheeled electric vehicles to adapt the multi-vehicle-road system model to variable conditions. Based on a four-wheel nonlinear vehicle dynamics model considering roll dynamics and load transfer, a set of vehicle-road system models based on a linear tire model and a nonlinear Dugoff tire model are two parts of the IMM filter. At the same time, the interactive multi-model unscented Kalman filter and the interactive multi-model extended Kalman filter are studied and compared. The results show that the proposed estimation method has good estimation effects; however, in these methods, the estimation effects are strongly affected by vehicle and tire models and system uncertainties;.

[0004] With the rapid development of artificial intelligence, data-driven estimation methods, especially those based on artificial neural networks, have shown broad application prospects in vehicle state estimation. This method is specifically designed to overcome the need for any type of vehicle model and its associated complex parameter set, providing decisive advantages such as adaptive learning, fault tolerance, and generalization. It has been proven to have the ability to avoid vehicle dynamics estimation problems. The articles "SASAKI H, NISHIMAKI T. ASideSlip Angle Estimation Using Neural Network for a Wheeled Vehicle[J], 2000" and "WEI W, SHAOYI B, LANCHUN Z, et al. Vehicle Sideslip Angle Estimation Based onGeneral Regression Neural Network[J]. Mathematical Problems in Engineering, 2016" used the lateral acceleration and yaw rate that can be directly measured by the IMU to estimate the sideslip angle of the vehicle's center of mass. The former trained the neural network using experimental data and verified it based on the experimental data, while the latter used a general regression neural network, trained the network with simulation data, and verified it with experimental data. These papers show that data-driven estimation methods are feasible, but this method depends on the training data set, and the estimation is accurate only when the test data set and the training data set are very similar. Therefore, the problem cannot be generalized;

[0005] Therefore, how to effectively combine the above Kalman filter with the neural network algorithm to perform the state estimation of multi-axis special vehicles has become an urgent problem in this field. Summary of the Invention

[0006] Aiming at the above existing problems, the present invention aims to provide a multi-axis special vehicle state estimation method based on a neural network and an unscented Kalman filter. This method simultaneously estimates the two main control parameters, namely the sideslip angle and the roll angle of the center of mass of the tires of the multi-axis special vehicle by using two algorithms, namely the neural network and the unscented Kalman filter, and can effectively solve the problems of strong non-linear systems, and has the characteristics of high algorithm accuracy, wide application range, and good stability.

[0007] In order to achieve the above purpose, the technical solutions adopted by the present invention are as follows:

[0008] A multi-axis special vehicle state estimation method based on a neural network and an unscented Kalman filter, including

[0009] Step 1: First, use sensors to collect vehicle signal data to form a vehicle data signal packet;

[0010] Step 2: Use the neural network module to process the sensor signal data and estimate the pseudo-centroid sideslip angle and pseudo-roll angle;

[0011] Step 3: Input the pseudo-centroid sideslip angle and pseudo-roll angle obtained by the neural network module as "pseudo-measurements" into the unscented Kalman module, and finally use the unscented Kalman module to obtain the centroid sideslip angle and roll angle of the multi-axis special vehicle.

[0012] Preferably, the sensor in Step 1 is a sensor arranged on the vehicle body, and the signal data includes lateral acceleration signal, longitudinal acceleration signal, yaw rate sensor signal and roll rate sensor signal.

[0013] Preferably, the establishment process of the neural network module includes

[0014] S2.1 Before establishing the neural network model, prepare the data set for training the neural network model, and the data set contains the data conditions under the main maneuvering states of the vehicle;

[0015] Wherein the data set includes a training set, a validation set and a test set with the same data content;

[0016] S2.2 Establish the network structure of the neural network module

[0017] The neural network module uses a fully connected neural network to estimate the pseudo-centroid sideslip angle and pseudo-roll angle.

[0018] Preferably, the fully connected neural network includes 1 input layer, 7 hidden layers and 1 output layer;

[0019] Wherein the input layer includes 6 inputs, namely longitudinal acceleration, lateral acceleration, steering wheel angle, longitudinal speed, yaw rate, roll rate;

[0020] The output layer is the centroid sideslip angle and roll angle, and the number of neurons in the single-layer network of the output layer is 100, and the number of network training times is 100.

[0021] Preferably, the fully connected neural network also includes a hyperbolic tangent activation function and a loss function, where

[0022] The form of the loss function is:

[0023]

[0024] In the formula, m is the number of samples participating in the calculation, y nn is the output of the network model, y l is the output label, γ in the penalty term is the penalty coefficient, k kis the number of weights, W j is the size of the j-th weight;

[0025] Based on the parameter gradients obtained by the backpropagation algorithm, the neural network model updates the parameters in the network based on the gradient descent method to minimize the value of the loss function. The update process of the model parameter θ is as follows:

[0026]

[0027] In the formula, θ is the parameter, and η is the update rate of the parameter.

[0028] Preferably, the design process of the unscented Kalman module described in step three includes

[0029] S3.1 Establish the state equation of the unscented Kalman filter module;

[0030] S3.2 Based on the established unscented Kalman filter state equation, use the UKF filtering algorithm for state estimation to establish the system estimation equation.

[0031] Preferably, the establishment process of the state equation of the unscented Kalman filter module described in step S3.1 includes

[0032] (1) Ignoring the longitudinal, vertical, and pitch dynamics of the vehicle, ignoring the unsprung mass and the influence of different characteristics of the front and rear axles on the vehicle characteristics, assuming that the sprung mass of the vehicle rotates around the roll center of the vehicle, establish a three-degree-of-freedom vehicle roll model composed of a "bicycle model" and a roll plane model, including vehicle lateral movement, yaw movement, and roll movement;

[0033] (2) Considering the coupling effects between the three degrees of freedom, list the force balance of lateral movement, the moment balance of yaw movement, and the moment balance of roll movement respectively:

[0034] Lateral force balance equation:

[0035]

[0036] Yaw moment balance equation:

[0037]

[0038] Roll moment balance equation:

[0039]

[0040] Lateral acceleration:

[0041]

[0042] In the formula, m s is the sprung mass; hc is the height from the centroid to the roll axis; F yi , i = 1, 2, 3, 4, 5 are lateral forces; δ i , i = 1, 2, 3, 4, 5 are the wheel angles of each axis; I z is the moment of inertia of the vehicle about the z-axis; I x is the moment of inertia of the vehicle about the x-axis; ω z is the body yaw angular velocity; l i , i = 1, 2, 3, 4, 5 are the distances from each axis to the centroid; g is the acceleration due to gravity; C s Suspension equivalent roll damping; K r is the suspension equivalent roll stiffness; v y is the lateral velocity; v x is the longitudinal velocity;

[0043] (3) Ignoring the influence of non-linear factors in the tire, the tire lateral force is expressed as:

[0044] F yi = k i α i (7)

[0045] The centroidal side slip angle is expressed as:

[0046] β = v y / v x (8)

[0047] Combining the above formulas, a simplified differential equation can be derived:

[0048]

[0049]

[0050]

[0051]

[0052] In the formula, k i , i = 1, 2, 3, 4, 5 are the cornering stiffnesses of the tires of each axis; α i , i = 1, 2, 3, 4, 5 are the tire cornering angles of each axis; k 12 , k 14 , k 15 are the proportional values of the tire angles of the 2nd, 4th, and 5th axes to the 1st axis respectively.

[0053] Preferably, the establishment process of the system estimation equation described in step S3.2 includes

[0054] (1) According to step S3.1, the state space equation of the system can be obtained as:

[0055]

[0056] where x k+ 1, z k , u k are the system state variables, observed variables, and input variables respectively; w k and v k are the process noise and observation noise respectively;

[0057] (2) According to the established differential equations (9)-(12), the state space equation is expressed as:

[0058]

[0059] where the state vector the input vector u is the steering wheel angle δ; the observation vector the values of A, B, and C are as shown in equation (16);

[0060]

[0061] (3) The UKF filtering algorithm uses the principle of similar distribution. The calculated Sigma point set has the same mean and covariance as the original distribution. After being introduced into the nonlinear system, an unscented transformation is performed to obtain the estimator.

[0062] Preferably, the working steps of the unscented Kalman filtering module include:

[0063] Step 1: Initialization

[0064] Perform initialization settings, set the initial values of the state vector and the state error covariance, and the initial values are respectively P0;

[0065] Step 2: Construct the state variable Sigma sampling points

[0066] Calculate the Sigma sampling points χ of the state vector (i) and obtain the weights ω of the sampling points (i) :

[0067]

[0068]

[0069] where and P are the mean and variance. The one with the superscript m is the corresponding weight of the mean, and the one with the superscript c is the corresponding weight of the covariance; the subscript is the number of the sampling point; λ = α 2(n + κ) - n is a scaling ratio function used to reduce the total prediction error; the selection of α (0.0001 ≤ α ≤ 1) controls the distribution state of the sampling points; κ is a second-order proportionality parameter, and its value should ensure that (n + λ)P is a positive semi-definite matrix, taking κ = 0; β is a weight coefficient, β ≥ 0;

[0070] Step 3: State prediction

[0071] The state prediction process includes state vector prediction and observation vector prediction;

[0072] Re-obtain the Sigma point set and its corresponding weights and calculate the further prediction of the Sigma points;

[0073]

[0074]

[0075] Furthermore, calculate the state quantity prediction and the prediction of the covariance matrix

[0076]

[0077]

[0078] Calculate the sampling points of the predicted quantity through the unscented transform using the state prediction quantity calculated in the previous step,

[0079]

[0080] Calculate the predicted observation of the Sigma points,

[0081]

[0082] Find the predicted mean of the observations and obtain the predicted mean and covariance,

[0083]

[0084] Step 4: State correction

[0085] The state correction process mainly performs state vector correction and state error covariance correction,

[0086] First, calculate the cross-correlation covariance between the state vector and the observation quantity:

[0087]

[0088] Calculate the Kalman gain matrix,

[0089]

[0090] Update the state quantity and the covariance matrix,

[0091]

[0092]

[0093] The initial value of the error covariance matrix is P = eye(4) × 10, the initial value of the process noise covariance matrix is R = diag[1e - 6, 1e - 6, 1e - 4, 1e - 6], and the initial value of the measurement noise covariance matrix is Q = eye(4) × 0.1.

[0094] A multi - axis special vehicle state estimator based on neural network and unscented Kalman filter, the state estimator includes a neural network module and an unscented Kalman filter module, where

[0095] The neural network module is used to estimate the pseudo - slip angle of the centroid and the pseudo - roll angle according to the vehicle signal data packet collected by the sensor;

[0096] The unscented Kalman filter module is used to calculate the slip angle of the centroid and the roll angle of the multi - axis special vehicle finally according to the pseudo - slip angle of the centroid and the pseudo - roll angle obtained by the neural network module.

[0097] The beneficial effects of the present invention are as follows: The present invention discloses a multi - axis special vehicle state estimation method based on neural network and unscented Kalman filter. Compared with the prior art, the improvements of the present invention are as follows:

[0098] (1) The present invention proposes a multi - axis special vehicle state estimation method based on neural network and unscented Kalman filter. This method combines two algorithms of neural network and unscented Kalman filter. When in use, the neural network can effectively ignore the influence of the nonlinearity of the vehicle system on it, and the unscented Kalman filter algorithm can adapt to stronger nonlinearity. Therefore, using the above two algorithms to estimate the two main control parameters, namely the slip angle of the centroid and the roll angle of the multi - axis special vehicle tires simultaneously, can effectively solve the problem of strong nonlinear systems, and has the advantages of high algorithm accuracy, wide application range and good stability;

[0099] (2) The state estimation algorithm proposed by this method first obtains the "pseudo - slip angle of the centroid" and the "pseudo - roll angle" through the neural network, and inputs them as observed quantities into the unscented Kalman filter to better correct the gain value, so that the estimation effect is better. And this algorithm adopts a simplified vehicle model to reduce the computational complexity and computational time; at the same time, this algorithm is useful in various environments (tunnel, desert and forest driving environments) because it uses the "pseudo - slip angle of the centroid" and the "pseudo - roll angle" estimated from the sensors installed on the current vehicle instead of the GPS antenna, ensuring the application range of the algorithm;

[0100] (3) The algorithm for simultaneously estimating the centroid side slip angle and the roll angle proposed by this method has been verified through simulation experiments and real vehicle experiments under S-curve working conditions not included in the training set. The verification results show that when the driving state of a five-axle heavy vehicle changes, the designed estimator can accurately estimate the vehicle state parameters, has good followability, and the algorithm has high accuracy, which can provide reliable parameters for subsequent vehicle stability control. Description of the Drawings

[0101] Figure 1 It is a structural architecture diagram of the multi-axle special vehicle state estimator based on neural network and unscented Kalman filter of the present invention.

[0102] Figure 2 It is a neural network architecture diagram of the neural network module of the present invention.

[0103] Figure 3 It is a roll dynamics model diagram of the present invention.

[0104] Figure 4 It is a real vehicle experiment diagram of Embodiment 2 of the present invention.

[0105] Figure 5 It is a model verification flow chart of Embodiment 2 of the present invention.

[0106] Figure 6 It is a counterclockwise circle experiment diagram of Embodiment 2 of the present invention.

[0107] Figure 7 It is a model verification comparison diagram of Embodiment 2 of the present invention.

[0108] Figure 8 It is a comparison diagram of estimation results under simulation constant speed working conditions of Embodiment 3 of the present invention.

[0109] Figure 9 It is a comparison diagram of estimation results under simulation uniform deceleration working conditions of Embodiment 3 of the present invention.

[0110] Figure 10 It is a clockwise circle experiment diagram of Embodiment 3 of the present invention.

[0111] Figure 11 It is a comparison diagram of estimation results under clockwise working conditions of Embodiment 3 of the present invention.

[0112] Figure 12 It is an error comparison diagram of estimation results under clockwise working conditions of Embodiment 3 of the present invention.

[0113] Figure 13 It is an S-curve experiment diagram of Embodiment 3 of the present invention.

[0114] Figure 14 It is a comparison diagram of estimation results under curve working conditions of Embodiment 3 of the present invention.

[0115] Figure 15 This is the error graph of the 3S curve estimation result of the present invention's embodiment.

[0116] Among them: in Figure 6 , Fig. (a) is the trajectory graph of the experimental vehicle moving in a counterclockwise circle, Fig. (b) is the steering wheel angle graph, and Fig. (c) is the speed graph;

[0117] In Figure 7 , Fig. (a) is the lateral acceleration comparison graph, Fig. (b) is the yaw rate comparison graph, Fig. (c) is the roll rate comparison graph, Fig. (d) is the roll angle comparison graph, and Fig. (e) is the sideslip angle of the center of mass comparison graph;

[0118] In Figure 8 , Fig. (a) is the comparison graph of the estimated result of the sideslip angle of the center of mass under the constant speed condition, and Fig. (b) is the estimated result graph of the roll angle under the constant speed condition;

[0119] In Figure 9 , Fig. (a) is the comparison graph of the estimated result of the sideslip angle of the center of mass under the constant deceleration condition, and Fig. (b) is the estimated result graph of the roll angle under the constant deceleration condition;

[0120] In Figure 10 , Fig. (a) is the trajectory graph of the experimental vehicle moving in a clockwise circle, Fig. (b) is the steering wheel angle graph, Fig. (c) is the speed graph, Fig. (d) is the lateral acceleration graph, Fig. (e) is the yaw rate graph, and Fig. (f) is the roll rate graph;

[0121] In Figure 11 , Fig. (a) is the estimated result graph of the sideslip angle of the center of mass under the clockwise condition, and Fig. (b) is the estimated result graph of the roll angle under the clockwise condition;

[0122] In Figure 12 , Fig. (a) is the estimated error graph of the sideslip angle of the center of mass under the clockwise condition, and Fig. (b) is the estimated error graph of the roll angle under the clockwise condition;

[0123] In Figure 13 , Fig. (a) is the trajectory graph of the experimental vehicle in the S-curve, Fig. (b) is the steering wheel angle graph, Fig. (c) is the speed graph, Fig. (d) is the lateral acceleration graph, Fig. (e) is the yaw rate graph, and Fig. (f) is the roll rate graph;

[0124] In Figure 14 , Fig. (a) is the estimated result graph of the sideslip angle of the center of mass under the S-curve condition, and Fig. (b) is the estimated result graph of the roll angle under the S-curve condition;

[0125] In Figure 15 , Fig. (a) is the estimated error graph of the sideslip angle of the center of mass under the S-curve condition, and Fig. (b) is the estimated error graph of the roll angle under the S-curve condition. Detailed implementation manners

[0126] In order to enable those of ordinary skill in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0127] Embodiment 1: Refer to Figures 1 - 15 A multi-axis special vehicle state estimation method based on neural network and unscented Kalman filter as shown, including

[0128] Step 1: First, use the sensors of the vehicle itself to collect vehicle signal data to form a vehicle data signal packet;

[0129] The signal data includes lateral acceleration signal, longitudinal acceleration signal, yaw rate sensor signal and roll rate sensor signal;

[0130] Step 2: Use the neural network module to process the sensor signal data to estimate the pseudo-centroid side slip angle (β NN ) and the pseudo-roll angle (φ NN );

[0131] Step 3: After obtaining the pseudo-centroid side slip angle (β NN ) and the pseudo-roll angle (φ NN ), input the pseudo-centroid side slip angle and pseudo-roll angle obtained by the neural network module as "pseudo-measurements" into the unscented Kalman module, and use the unscented Kalman module to finally obtain the centroid side slip angle and roll angle of the multi-axis special vehicle.

[0132] Preferably, the establishment process of the neural network module described in step 2 includes

[0133] Due to the characteristic that the neural network does not need to consider establishing a vehicle model, it determines its great advantage in dealing with nonlinear problems, can meet the conditions for estimating the centroid side slip angle and roll angle, and using the neural network estimation can ensure the real-time response of the system, and at the same time will not additionally increase the production cost of the whole vehicle;

[0134] S2.1 Prepare a data set for training the neural network model

[0135] Before establishing the neural network model, it is necessary to prepare a data set in advance for offline training of the model. The data set must include the data conditions under the main maneuvering states of the vehicle to characterize the response characteristics of the nonlinear vehicle;

[0136] The dataset includes a training set, a validation set, and a test set. The data content of the three is the same. Since the working conditions are very complex during the actual driving of the vehicle, in order to improve the estimation accuracy, it is necessary to set as many different working conditions as possible; the data is collected by Trucksim providing a vehicle dynamics model and jointly simulating with Matlab / Simulink; Trucksim software is a widely used simulation software in the automotive industry that combines traditional and modern multi-body vehicle dynamics based on parametric modeling; one of the main advantages of using a simulated vehicle model to collect the dataset is the ability to perform different types of vehicle maneuvers to avoid accidents that may occur under different road conditions; in addition, the simulation model ensures the reproducibility of the tests;

[0137] The working conditions of the acquisition route of the training set mainly include straight line, double lane change, figure eight, fixed circle, fishhook. The speed is set between 10 km / h and 90 km / h, and a value is taken every 10 km / h. The road friction coefficient is set to 0.3, 0.5, 0.8. Except for the straight line condition, in the case of a road friction coefficient of 0.3 for the rest of the road conditions, the maximum speed limit is 80 km / h. According to the simulation results, when the speed is higher than 80 km / h, the vehicle will roll over; the sampling frequency is set to 100 Hz, and three simulations are performed under the same conditions for each test group, and the average value is recorded; the longitudinal acceleration, lateral acceleration, steering wheel angle, longitudinal speed, yaw rate, roll angle rate, sideslip angle of the center of mass, and roll angle are obtained through simulation acquisition; the specific values of the training conditions of the training set are shown in Table 1:

[0138] Table 1: Training Conditions of the Dataset

[0139]

[0140]

[0141] S2.2 Establish the Network Structure of the Neural Network Module

[0142] The neural network module uses a fully connected neural network to estimate the pseudo sideslip angle of the center of mass and the pseudo roll angle; the preliminary estimates of the roll angle and the sideslip angle of the center of mass obtained from the neural network observer are used as "pseudo measurements" in the unscented Kalman filter; the architecture of the fully connected neural network is as Figure 2 shown, and the architecture of the fully connected neural network includes

[0143] 1 input layer, including 6 inputs, namely longitudinal acceleration, lateral acceleration, steering wheel angle, longitudinal speed, yaw rate, roll rate, 7 hidden layers, and 2 output layers. The output layers are the sideslip angle of the center of mass and the roll angle. The number of neurons in a single-layer network is 100. Neurons, as the basic computing units of the neural network, receive data from other neurons or the external environment, and then calculate an output; each input value has a weight, and after non-linear calculation by the activation function within the neuron, an output value is obtained; in the initial stage, all weights are randomly assigned; for each input value in the training set, after the forward calculation of the neural network, the obtained output value will be compared with the expected output, and then the resulting error will be passed back to the previous network layer; this error will be recorded, and then the weights will be adjusted accordingly.

[0144] 1. Input neurons: Located in the input layer, mainly used to transmit information from the outside world into the neural network. These neurons do not need to perform any calculations, but only act as information transmitters, or rather, data enters the hidden layer.

[0145] 2. Hidden neurons: Located in the hidden layer. The neurons in the hidden layer do not have direct connections with the outside world. They are indirectly connected to the outside world through the previous input layer and the subsequent output layer. Therefore, it is called the hidden layer. The neurons in the hidden layer will perform calculations, transform the input information from the input layer through calculations, and then output it to the output layer.

[0146] 3. Output neurons: Located in the output layer, the output neurons are used to output the information from the hidden layer to the outside world, that is, to output the final result.

[0147] The activation function is an important guarantee for the non-linear mapping relationship between the input and output of a fully connected neural network. If the pre-output in the neuron structure is not activated by a function, the fully connected neural network is equivalent to a linear model, and it is difficult for a linear model to accurately represent the non-linear phenomena existing in reality. In this paper, the hyperbolic tangent (Tanh) activation function is selected as the activation function.

[0148] Since estimating the vehicle attitude angle belongs to the regression problem in deep learning, and it predicts a continuous change process rather than discrete labels, the mean squared error is used as the loss function representing the gap between the prediction and the actual value. In a neural network model, the total number of parameters to be trained is often much larger than the total amount of the training data set, which may cause the network model to be prone to overfitting during the training process. Overfitting means that the trained model overly relies on the sample data, only focusing on the currently collected sample data, while the model lacks the generalization ability for general data. This is not conducive to the estimation of vehicle parameters because the sample data we collect is limited, but in the actual driving process of the vehicle, the data situations that occur will be very different from the sample data we collect. Regularization is a processing method to prevent model overfitting. By adding a penalty norm for the training parameters to the loss function and constraining the parameters to be trained, overfitting is solved. In this embodiment, L2 regularization is added. To facilitate adjusting the influence of the regularization function on the weights of the network, the hyperparameter γ is introduced to adjust the influence of the L2 regularization term on the network. The form of the entire loss function is as follows:

[0149]

[0150] In the formula, m is the number of samples participating in the calculation, y nn is the output of the network model, y l is the output label. In the penalty term, γ is the penalty coefficient, k k is the number of weights, and W j is the size of the j-th weight;

[0151] Based on the parameter gradients obtained according to the backpropagation algorithm, the neural network model updates the parameters in the network based on the gradient descent method to minimize the value of the loss function. The update process of the model parameter θ is as follows:

[0152]

[0153] In the formula, θ is the parameter, and η is the update rate of the parameter;

[0154] The way of parameter update affects the convergence speed during network training. Different types of optimizers have different update strategies for network gradients. When training the network, RMSprop is used as the network training optimizer. RMSprop is an adaptive learning rate optimizer. It calculates the update size of network parameters using the cumulative gradient and introduces the hyperparameter decay rate when updating the parameters, which can weaken the influence of the historical cumulative gradient, enabling the optimizer not to shrink with the global cumulative gradient when updating the parameters, avoiding the parameter update becoming too small as the number of training loops increases, and accelerating the network convergence speed. The global learning rate is set to 0.0001, the decay rate is 0.8, and the L2 weight penalty coefficient is 0.0005.

[0155] Preferably, the design process of the unscented Kalman module described in step three includes

[0156] S3.1 Establish the state equation of the unscented Kalman filter module

[0157] Ignoring the longitudinal, vertical, and pitch dynamics of the vehicle, neglecting the unsprung mass and the influence of different characteristics of the front and rear axles on the vehicle characteristics, assuming that the sprung mass of the vehicle rotates around the roll center of the vehicle, as Figure 3 shown, establish a three-degree-of-freedom vehicle roll model composed of a "bicycle model" and a roll plane model, including vehicle lateral motion, yaw motion, and roll motion;

[0158] Based on the above assumptions, considering the coupling effects between the three degrees of freedom, respectively list the force balance of lateral motion, the moment balance of yaw motion, and the moment balance of roll motion according to D'Alembert's principle;

[0159] Lateral force balance equation:

[0160]

[0161] Yaw moment balance equation:

[0162]

[0163] Roll moment balance equation:

[0164]

[0165] Lateral acceleration:

[0166]

[0167] Where m s is the sprung mass; h c is the height from the center of mass to the roll axis; F yi , i = 1, 2, 3, 4, 5 are lateral forces; δ i , i = 1, 2, 3, 4, 5 are the wheel angles of each axis; I z is the moment of inertia of the vehicle about the z-axis; I x is the moment of inertia of the vehicle about the x-axis; ω z is the yaw angular velocity of the vehicle body; l i , i = 1, 2, 3, 4, 5 are the distances from each axis to the center of mass; g is the acceleration due to gravity; C s is the equivalent roll damping of the suspension; K r is the equivalent roll stiffness of the suspension; v y is the lateral velocity; v x is the longitudinal velocity;

[0168] To simplify the model and neglect the influence of non - linear factors in the tire, the lateral force of the tire can be expressed as:

[0169] F yi =k i α i (7)

[0170] The sideslip angle of the center of mass can be expressed as:

[0171] β=v y / v x (8)

[0172] Combining the above several formulas, the simplified differential equation can be derived:

[0173]

[0174]

[0175]

[0176]

[0177] In the formula, k i ,i=1,2,3,4,5 are the cornering stiffnesses of the tires on each axis; α i ,i=1,2,3,4,5 are the sideslip angles of the tires on each axis; k 12 ,k 14 ,k 15 are the ratio values of the tire rotations of the 2nd, 4th, and 5th axes to the 1st axis respectively;

[0178] S3.2 On the basis of establishing the unscented Kalman filter state equation, use the UKF filtering algorithm for state estimation to establish the system estimation equation

[0179] According to step S3.1, the state - space equation of the system can be expressed as:

[0180]

[0181] In the formula, x k+ 1,z k ,u k are the system state quantity, the observed quantity, and the input quantity respectively; w k and v k are the process noise and the observation noise respectively;

[0182] According to the established differential equations (9) - (12), the state - space equation can be expressed as:

[0183]

[0184] y = C·x + v (15)

[0185] In the formula, the state vector The input vector u is the steering wheel angle δ; the observation vector The values of A, B, and C are as shown in formula (16);

[0186]

[0187] The UKF filtering algorithm uses the principle of similar distribution. The calculated Sigma point set is the same as the original distribution mean and covariance. After being brought into the nonlinear system, an unscented transformation is performed, and then the estimator is obtained.

[0188] Preferably, the working steps of the unscented Kalman filtering module include:

[0189] Step 1: Initialization

[0190] The estimator is initialized. The initial values of its state vector and state error covariance are respectively P0;

[0191] Step 2: Construct the state Sigma sampling points

[0192] Calculate the Sigma sampling points χ of the state vector (i) And find the weights ω of the sampling points (i) :

[0193]

[0194]

[0195] In the formula, And P are the mean and variance. The one with superscript m is the corresponding weight of the mean, and the one with superscript c is the corresponding weight of the covariance; the subscript is the number of the sampling point; λ = α 2 (n + κ) - n is a scaling ratio function used to reduce the total prediction error; the selection of α (0.0001 ≤ α ≤ 1) controls the distribution state of the sampling points; κ is a second-order proportional parameter, and its value should ensure that (n + λ)P is a positive semi-definite matrix, take κ = 0; β is a weight coefficient, β ≥ 0, and its role is to combine the high-order term moments of the system equation. Taking β = 2 is the best;

[0196] Step 3: State prediction

[0197] The state prediction process includes state vector prediction and observation vector prediction;

[0198] Re-obtain the Sigma point set and its corresponding weights and calculate the further prediction of the Sigma points;

[0199]

[0200]

[0201] Furthermore, the state quantity prediction and the covariance matrix prediction are obtained.

[0202]

[0203]

[0204] The state prediction quantity calculated in the previous step is used to calculate the sampling points of the prediction quantity through the unscented transformation.

[0205]

[0206] Calculate the predicted observation quantity of the Sigma points.

[0207]

[0208] Find the predicted mean of the observation, and obtain the predicted mean and covariance.

[0209]

[0210] Step 4: State correction

[0211] The state correction process mainly performs state vector correction and state error covariance correction.

[0212] First, calculate the cross-correlation covariance between the state vector and the observation quantity:

[0213]

[0214] Calculate the Kalman gain matrix.

[0215]

[0216] Update the state quantity and the covariance matrix.

[0217]

[0218]

[0219] In this embodiment, the initial value of the error covariance matrix is P = eye(4)×10, the initial value of the process noise covariance matrix is R = diag[1e - 6, 1e - 6, 1e - 4, 1e - 6], and the initial value of the measurement noise covariance matrix is Q = eye(4)×0.1.

[0220] Through the estimation method described in this embodiment, a multi-axis special vehicle state estimator based on a neural network and an unscented Kalman filter is obtained. The state estimator mainly includes two modules, namely, a neural network module and an unscented Kalman filter module. Among them

[0221] The neural network module is used to estimate the "pseudo-central lateral deviation angle" β NN and the "pseudo-roll angle" φ NN ;

[0222] The unscented Kalman filter module is used to input the pseudo-central lateral deviation angle and pseudo-roll angle obtained by the neural network module based on the neural network as "pseudo-measurements" into the unscented Kalman module, and finally use the unscented Kalman module to obtain the centroid lateral deviation angle and roll angle of the multi-axis special vehicle.

[0223] Embodiment 2. To verify the effectiveness of the multi-axis special vehicle state estimator based on a neural network and an unscented Kalman filter established by the method described in Embodiment 1, a Trucksim model verification experiment of this embodiment is designed to verify the above state estimator:

[0224] 1. Real vehicle test

[0225] To verify the accuracy of the Trucksim model and the algorithm, a real vehicle experiment is carried out using the special vehicle driving state monitoring system available;

[0226] The real vehicle experiment scheme is as follows: The test road condition is a dry cement road surface with a relatively flat terrain; the test vehicle is a certain five-axis heavy special vehicle; the steering wheel angle sensor collects steering information; the Speedbox-mini is placed in the vehicle, and the centroid compensation is corrected in the software Race Technology to measure information such as vehicle speed, trajectory, and elevation during vehicle operation; the GPS and INS antennas are placed on the roof to measure information such as vehicle attitude angle and angular velocity; multiple sensor signals are output to the DEWE43 data acquisition system with multiple channels and multiple parameters through the CAN bus, and the Race Technology and DEWEsoft-X3 software are used to record, analyze, and process the experimental data. The specific experimental process is as Figure 4 shown;

[0227] 2. Trucksim model verification

[0228] Use the experimental results of the real vehicle to verify the effectiveness of the TruckSim simulation vehicle model. The process of vehicle model verification is as Figure 5As shown, the measured steering wheel angle and longitudinal speed in the experiment are used as the inputs of the Trucksim model, and the lateral acceleration, yaw rate, roll rate, roll angle, and sideslip angle of the center of mass obtained from the Trucksim model are compared with the experimental data. Since the purpose of the built model is to serve as a simulation platform for state observation, the longitudinal speed to be observed is an important input parameter in various control algorithms, and the state quantities of yaw rate and sideslip angle of the center of mass are based on the lateral dynamics model. The yaw rate and lateral acceleration are used in research to verify the accuracy of the model. According to the difference between the experimental data and the simulation data, the model parameters are adjusted through trial and error, and finally the model with the minimum root mean square error is determined.

[0229] Based on the existing experimental conditions, the selected road conditions for model verification are the clockwise circular section, and the data information of the selected section is used for vehicle model parameter adjustment and verification. The actual vehicle driving trajectory, steering wheel angle, and speed are respectively Figure 7 (a - c); The comparison diagrams of the lateral acceleration, yaw rate, roll rate, roll angle, and sideslip angle of the center of mass between the model and the actual values are respectively Figure 8 (a - e);

[0230] The root mean square error and maximum error between the actual vehicle data and the output data of the TruckSim model are shown in Table 2:

[0231] Table 2: Verification Deviation

[0232]

[0233] From Figure 7 and Table 2, it can be seen that there are some errors in the comparison between the actual value and the model output. However, generally, the change trend is consistent with the actual situation and is in good agreement, which can better reflect the real vehicle. Since the actual vehicle experiment environment is a relatively flat site, some road surfaces have some slopes, and the steering wheel angle input in the TruckSim model does not consider the road surface slope. The verification result of this actual vehicle test can illustrate the authenticity and reliability of the model. Finally, during the process of adjusting and verifying the model, based on the original vehicle data, several parameters such as the steering system transmission ratio, the moment of inertia of the vehicle mass rotating around the axis, and the tire model are slightly adjusted. During the adjustment and verification process, the change of the steering system transmission ratio has a greater impact on the change of the parameter quantity. The actual vehicle steering system transmission ratio changes with the gear, and it is regarded as a fixed value in the model. According to the adjustment criterion of the minimum root mean square error value, the main parameter values finally determined by the model are shown in Table 3:

[0234] Table 3: Vehicle Parameters

[0235]

[0236] Example 3. To verify the effectiveness of the multi-axis special vehicle state estimator based on neural network and unscented Kalman filter established by the method described in Example 1, the following verification process of the estimation algorithm is designed for the above state estimator:

[0237] 1. Simulation verification

[0238] To prove the effectiveness of the proposed estimator based on neural network and unscented Kalman filter, when conducting simulation experiments for verification, the test set should be as different from the training set working conditions as possible to test the dynamic characteristics learned by the network model; therefore, a steering wheel sine input working condition that does not exist in the training set is set in the test set. The former is carried out on a road surface with a friction coefficient of 0.8 at a speed of 75 km / h; the latter is carried out on a road surface with a friction coefficient of 0.8, and the vehicle decelerates uniformly from 75 km / h to 35 km / h; the results of the simulation experiments are as Figure 8 、 9 ; In addition to these two working conditions, this paper also conducts verification for the working conditions of a road surface friction coefficient of 0.5, a vehicle speed of 75 km / h and a road surface friction coefficient of 0.8, and the vehicle accelerates uniformly from 35 km / h to 75 km / h. The single UKF is used for comparison with this algorithm. Compared with NN-UKF, the observation vector in UKF is only considering the lateral acceleration and roll rate without considering the "pseudo roll angle" and "pseudo centroidal side slip angle". The root mean square error comparisons of the four working conditions are shown in Table 4;

[0239] Table 4: Root mean square error of the estimation results of the simulation experiment

[0240]

[0241]

[0242] From Figure 8 and Figure 9 it can be seen that both methods have good followability with the simulation values, but when the steering wheel steering changes, it can be clearly seen that the algorithm result based on NN-UKF is better, and the difference between it and the simulation value is much smaller than that of using a single UKF; it can also be seen from Table 4 that even under variable speed working conditions, the estimation errors of the centroidal side slip angle and roll angle in the two working conditions based on the NN-UKF algorithm are significantly smaller than those of the UKF algorithm, which also proves that the NN-UKF algorithm has better robustness to vehicle speed in estimating the centroidal side slip angle and roll angle, and the NN-UKF algorithm performs more stably during the entire simulation test process; in summary, the simulation results show that the estimator based on NN-UKF has better performance than the estimator based on UKF;

[0243] 2. Real vehicle experiment verification

[0244] As can be seen from the real vehicle experiment part in Section 1 of Example 2, based on the existing experimental conditions, after installing the experimental equipment, a real vehicle experiment was carried out on a certain dry cement road surface; a counterclockwise turning section and an S-curve section were selected for experimental verification of the algorithm;

[0245] Condition 1: In the counterclockwise turning section condition, the vehicle's motion trajectory, steering wheel angle, longitudinal speed, lateral acceleration, yaw angular velocity, and roll angular velocity are as Figure 10 (a - f) shown. In order to further verify the estimation effect of the NN + UKF algorithm, when conducting the verification comparison, compared with the simulation experiment verification, the data comparison of the single neural network was added. The specific verification of the sideslip angle of the center of mass and the roll angle is as Figure 11 (a - b) shown, and the differences in the sideslip angle of the center of mass and the roll angle are as Figure 12 (a - b) shown;

[0246] Condition 2: In the S-curve section condition, the vehicle's motion trajectory, steering wheel angle, longitudinal speed, lateral acceleration, yaw angular velocity, and roll angular velocity are as Figure 13 (a - f) shown, the verification comparison of the sideslip angle of the center of mass and the roll angle is as Figure 14 (a - b) shown, and the differences in the sideslip angle of the center of mass and the roll angle are as Figure 15 (a - b) shown; The root mean square error tables of each item under the two conditions are shown in Table 5:

[0247] Table 5: Root Mean Square Error of Estimation Results in Real Vehicle Experiment

[0248]

[0249] From Figure 11 、 Figure 12 、 Figure 14 、 Figure 15 、and Table 5, it can be seen that: in the turning condition, the driving speed is stable at about 30 km / h. In 11(a - b), points a and b are the inflection points of Condition 1. The UKF and NN have slightly worse followability when the vehicle starts to turn. The UKF + NN can always maintain good followability. Even if there are large errors in some time periods, they can quickly converge to the vicinity of the experimental values. The change of the steering wheel angle has no obvious influence on the estimation accuracy of the algorithm, and from Figure 12(a-b) It can be seen that the absolute error of the UKF+NN algorithm at the maximum turning angle is much smaller than that of the other two algorithms, and the estimation effect of this algorithm is still better than that of the other two algorithms. In the s-curve section working condition, affected by the test site, the vehicle speed is relatively slow, and the s-curve is completed within a short distance, which has a certain impact on the vehicle stability and cannot well reflect the true state of the vehicle. Moreover, in this section, the road surface has a certain slope, and when establishing the rollover estimation model, the disturbance of the road surface slope angle to the attitude angle is not taken into account, resulting in large fluctuations in the errors of the three estimation values and reducing the estimation accuracy of the algorithm. However, whether from Figure 15 It can be clearly seen from both (a-b) and Table 5 the advantages of the UKF+NN algorithm;

[0250] In terms of the comparison of root mean square error, in working condition 1, the estimation of the sideslip angle of the center of mass based on UKF+NN decreased by 49.8% relative to the UKF algorithm and 39.7% relative to the NN algorithm; the estimation of the roll angle decreased by 47.9% relative to the UKF algorithm and 30.8% relative to the NN algorithm. In working condition 2, the estimation of the sideslip angle of the center of mass based on UKF+NN decreased by 46.8% relative to the UKF algorithm and 25.5% relative to the NN algorithm; the estimation of the roll angle decreased by 41.4% relative to the UKF algorithm and 27.3% relative to the NN algorithm;

[0251] The above results show that the estimation algorithm has high accuracy and can realize the simultaneous estimation of the sideslip angle of the center of mass and the roll angle parameters.

[0252] Therefore, it can be seen that in view of the characteristics of multi-axle special vehicles and the possibility of handling instability in complex and harsh driving environments, the estimation method based on neural network and unscented Kalman filter proposed in Embodiment 1 of the present invention has stronger adaptability, and this algorithm can simultaneously estimate the two main control parameters, the sideslip angle of the center of mass and the roll angle, which have a greater impact and are difficult to measure.

[0253] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A state estimation method for a multi-axis special vehicle based on a neural network and an unscented Kalman filter, characterized in that: including Step 1: First, use sensors to collect vehicle signal data to form a vehicle data signal packet; the sensors are sensors installed on the vehicle body, and the signal data includes lateral acceleration signal, longitudinal acceleration signal, yaw rate sensor signal, and roll rate sensor signal; Step 2: Use the neural network module to process the sensor signal data to estimate the pseudo-centroidal side slip angle and pseudo-roll angle; Step 3: Input the pseudo-centroidal side slip angle and pseudo-roll angle obtained by the neural network module as "pseudo-measurements" into the unscented Kalman module, and finally use the unscented Kalman module to obtain the centroidal side slip angle and roll angle of the multi-axle special vehicle; The design process of the unscented Kalman module described in Step 3 includes S3.1 Establish the state equation of the unscented Kalman filter module; S3.2 Based on the established unscented Kalman filter state equation, use the UKF filtering algorithm for state estimation to establish a system estimation equation; The establishment process of the system estimation equation described in Step S3.2 includes (1) According to Step S3.1, the state space equation of the system can be expressed as: where x k+1 , z k , u k are the system state variables, the observed variables, and the input variables, respectively; w k and v k are the process noise and the observation noise, respectively; (2) According to the established differential equations (9)-(12), the state space equation is expressed as: y = C·x + v (15) where the state vector the input vector u is the steering wheel angle δ; the observation vector (3) The UKF filtering algorithm uses the similarity distribution principle. The calculated Sigma point set has the same mean and covariance as the original distribution. After being brought into the nonlinear system, unscented transformation is performed to obtain the estimator.

2. The state estimation method for a multi-axis special vehicle based on a neural network and an unscented Kalman filter according to claim 1, characterized in that: The establishment process of the neural network module described in Step 2 includes S2.1 Before establishing the neural network model, prepare a dataset for training the neural network model. The dataset contains data under the main maneuvering states of the vehicle; where the dataset includes a training set, a validation set, and a test set with the same data content; S2.2 Establish the network structure of the neural network module The neural network module uses a fully connected neural network to estimate the pseudo-centroidal side slip angle and pseudo-roll angle.

3. The state estimation method for a multi-axis special vehicle based on a neural network and an unscented Kalman filter according to claim 2, characterized in that: The fully connected neural network includes 1 input layer, 7 hidden layers, and 1 output layer; where the input layer includes 6 inputs, namely longitudinal acceleration, lateral acceleration, steering wheel angle, longitudinal speed, yaw rate, and roll rate; The output layer is the centroidal side slip angle and roll angle, and the number of single-layer network neurons in the output layer is 100, and the number of network training times is 100.

4. The state estimation method for a multi-axis special vehicle based on a neural network and an unscented Kalman filter according to claim 3, characterized in that: The fully connected neural network also includes a hyperbolic tangent activation function and a loss function, where The form of the loss function is: where m is the number of samples involved in the calculation, y nn is the output of the network model, and y l is the output label. In the penalty term, γ is the penalty coefficient, k k is the number of weights, and W j is the size of the j-th weight; According to the parameter gradient obtained by the backpropagation algorithm, the neural network model updates the parameters in the network based on the gradient descent method to minimize the value of the loss function. The update process of the model parameter θ is as follows: where θ is a parameter, η is the update rate of the parameter.

5. The multi-axis special vehicle state estimation method based on neural network and unscented Kalman filter according to claim 1, characterized in that: The establishment process of the state equation of the unscented Kalman filter module described in Step S3.1 includes (1) Ignore the longitudinal, vertical, and pitch dynamics characteristics of the vehicle, ignore the influence of the unsprung mass and the different characteristics of the front and rear axles on the vehicle characteristics, assume that the sprung mass of the vehicle rotates around the roll center of the vehicle, and establish a three-degree-of-freedom vehicle roll model composed of a "bicycle model" and a roll plane model, including vehicle lateral motion, yaw motion, and roll motion; (2) Considering the coupling effects among the three degrees of freedom, list the force balance equations for lateral motion, the moment balance equations for yaw motion, and the moment balance equations for roll motion respectively: Lateral force balance equation: Yaw moment balance equation: Roll moment balance equation: Lateral acceleration: where m s is the sprung mass; h c is the height from the center of mass to the roll axis; F yi , i = 1, 2, 3, 4, 5 are the lateral forces; δ i , i = 1, 2, 3, 4, 5 are the wheel angles of each axle; I z is the moment of inertia of the vehicle about the z-axis; I x is the moment of inertia of the vehicle about the x-axis; ω z is the body yaw angular velocity; l i , i = 1, 2, 3, 4, 5 are the distances from each axle to the center of mass; g is the acceleration due to gravity; C s is the equivalent roll damping of the suspension; K r is the equivalent roll stiffness of the suspension; v y is the lateral velocity; v x is the longitudinal velocity; (3) Ignoring the influence of non-linear factors in the tires, the lateral tire force is expressed as: F yi = k i α i (7) The slip angle of the center of mass is expressed as: β = v y / v x (8) Combining the above formulas, the simplified differential equation can be derived: where k i , i = 1, 2, 3, 4, 5 are the cornering stiffnesses of the tires on each axle; α i , i = 1, 2, 3, 4, 5 are the cornering angles of the tires on each axle; k 12 , k 14 , k 15 are the proportional values of the tire rotations of the second, fourth, and fifth axles to the first axle, respectively.

6. The multi-axis special vehicle state estimation method based on neural network and unscented Kalman filter according to claim 5, characterized in that: The values of A, B, and C are as shown in Equation (16); 7. The multi-axis special vehicle state estimation method based on neural network and unscented Kalman filter according to claim 1, characterized in that: The working steps of the unscented Kalman filter module include: Step 1: Initialization Perform initialization settings, set the initial values of the state vector and the state error covariance, and the initial values are respectively P0; Step 2: Construct the Sigma sampling points of the state variables Calculate the Sigma sampling points χ of the state vector (i) and obtain the weights ω of the sampling points (i) : In the formula, and P are the mean and variance, the superscript m represents the corresponding weight of the mean, and the superscript c represents the corresponding weight of the covariance; the subscript indicates the nth sampling point; λ = α 2 (n + κ) - n is a scaling ratio function used to reduce the total prediction error; the selection of α controls the distribution state of the sampling points; κ is a second-order proportional parameter, and its value should ensure that (n + λ)P is a positive semi-definite matrix, taking κ = 0; β is a weight coefficient, β ≥ 0; Step 3: State prediction The state prediction process includes state vector prediction and observation vector prediction; Re-obtain the Sigma point set and its corresponding weights and calculate the further prediction of the Sigma points; Furthermore, calculate the state variable prediction and the prediction of the covariance matrix Calculate the acquisition sample points of the prediction through the unscented transform using the state prediction value calculated in the previous step, Calculate the predicted observation values of the Sigma points, Calculate the predicted mean of the observations to obtain the predicted mean and covariance, Step 4: State correction The state correction process mainly performs state vector correction and state error covariance correction. First, calculate the cross-correlation covariance between the state vector and the observation quantity: Calculate the Kalman gain matrix, Update the state variables and the covariance matrix, The initial value of the error covariance matrix is P = eye(4)×10, the initial value of the process noise covariance matrix is R = diag[1e-6, 1e-6, 1e-4, 1e-6], and the initial value of the measurement noise covariance matrix is Q = eye(4)×0.

1.

8. A multi-axis special vehicle state estimator based on a neural network and an unscented Kalman filter, implemented based on the method according to any one of claims 1-7, characterized in that: The state estimator includes a neural network module and an unscented Kalman filter module, where The neural network module is used to estimate the pseudo-slip angle of the center of mass and the pseudo-roll angle according to the vehicle signal data packet collected by the sensor; The unscented Kalman filter module is used to calculate the slip angle of the center of mass and the roll angle of the multi-axle special vehicle finally according to the pseudo-slip angle of the center of mass and the pseudo-roll angle obtained by the neural network module.

Images