CN-121980813-A - Dynamic amphibious vehicle dynamic parameter identification method integrating physical constraints

CN121980813ACN 121980813 ACN121980813 ACN 121980813ACN-121980813-A

Abstract

The invention is suitable for the technical field of vehicle dynamics, and provides a dynamic amphibious vehicle dynamics parameter identification method integrating physical constraints, which comprises the following steps of carrying out working condition identification according to key state quantity to obtain a working condition classification result; based on the working condition classification result, a neural network model is adopted to learn the relevant characteristics of the dynamic parameters from the state data after the working condition classification, based on the amphibious vehicle dynamic model, a loss function integrating physical mechanisms is constructed, and a weighted loss function integrating physical constraints is obtained by combining boundary condition loss, initial condition loss and data interior point loss, and based on the weighted loss function, back propagation is carried out, parameters of the neural network model and vehicle dynamic parameters are optimized, and dynamic parameter dynamic identification is achieved. The invention can improve the working condition identification precision of the vehicle in various complex operation scenes on land and water and realize dynamic identification of the dynamic parameters of the vehicle in various complex operation scenes on land and water.

Inventors

ZHU ZHONGWEN
ZHANG JINYUAN
LI CHENG
QIU XIN
SHI ZHENGPENG
CHEN QIANG

Assignees

合肥工业大学

Dates

Publication Date: 20260505
Application Date: 20260210

Claims (8)

1. The dynamic amphibious vehicle dynamic parameter identification method integrating physical constraints is characterized by comprising the following steps of: Acquiring key state quantity in the running process of the vehicle, wherein the key state quantity comprises four-dimensional dynamic information; Based on a preset four-dimensional dynamic clustering method, carrying out working condition identification according to the key state quantity to obtain a working condition classification result; based on the working condition classification result, learning dynamics parameter related features from the state data after working condition classification by adopting a neural network model; constructing a loss function of a fusion physical mechanism based on an amphibious vehicle dynamics model, and combining boundary condition loss, initial condition loss and data interior point loss to obtain a weighted loss function of the fusion physical constraint; And (3) back propagation is carried out based on the weighted loss function, parameters of the neural network model and vehicle dynamics parameters are optimized, and dynamic identification of the dynamics parameters is realized.
2. The dynamic identification method for dynamic parameters of amphibious vehicle fused with physical constraints according to claim 1, wherein the four-dimensional dynamic information comprises vehicle speed, roll angle speed, pitch angle speed and heading angle speed.
3. The dynamic amphibious vehicle dynamic parameter identification method based on the physical constraint fusion of claim 2, wherein the step of carrying out working condition identification according to the key state quantity based on a preset four-dimensional dynamic clustering method to obtain a working condition classification result specifically comprises the following steps: Acquiring a time sequence of a vehicle speed, a roll angle speed, a pitch angle speed and a course angular velocity, and preprocessing and normalizing the time sequence to obtain normalized data; Randomly selecting 2 clusters from the standardized data to serve as an input land driving center matrix A 0 and an overwater navigation center matrix B 0 , and respectively calculating Manhattan distances from different-dimension non-center matrix sequences in the data set to the center matrix A 0 and the center matrix B 0 to obtain a first four-dimensional distance sequence; Carrying out weighted summation on the first four-dimensional distance sequence to obtain D 1 (A, B), traversing the whole sequence, updating the cluster center according to Manhattan distances from the center matrix to other samples in the cluster and the minimum principle until the cluster center is not changed any more to obtain a center matrix A and a center matrix B, and dividing the two working condition modes of land driving and water navigation; Under the preliminary divided working condition mode, randomly selecting 3 clusters as input center matrixes C 0 、D 0 and E 0 respectively, and calculating Manhattan distances from non-center matrix sequences with different dimensions in a data set to the center matrixes C 0 、D 0 and E 0 respectively to obtain a second four-dimensional distance sequence; And carrying out weighted summation on the second four-dimensional distance sequence to obtain D 2 (C, D and E), traversing the whole sequence, updating the cluster center according to the Manhattan distance from the center matrix to other samples in the cluster and the minimum principle until the cluster center is not changed to obtain the center matrices C, D and E, and dividing the cluster into three working condition modes of urban flat working condition, rugged off-road working condition and muddy wet soft working condition on land driving or three working condition modes of water stable working condition, water wave-exciting working condition and shoal silt working condition on water navigation.
4. The dynamic amphibious vehicle dynamics parameter identification method based on the fusion physical constraint of claim 1, wherein the neural network model is a multi-layer perceptron, which is composed of 4 fully-connected layers, 64 neurons in each layer, and a neural network prediction result is obtained by using tanh as an activation function and inputting according to a working condition classification result.
5. The dynamic amphibious vehicle dynamics parameter identification method based on the fusion physical constraint of claim 1, wherein the step of constructing a loss function of the fusion physical mechanism based on the amphibious vehicle dynamics model and combining boundary condition loss, initial condition loss and data interior point loss to obtain a weighted loss function of the fusion physical constraint specifically comprises the following steps: Constructing a loss function L PDE of a fusion physical mechanism according to a predicted result of the neural network model and a residual error between the amphibious vehicle dynamics model; The feature equation extracted by the neural network model is marked as C (x), for each training sample x i , the cluster label identified based on the working condition is C i , the corresponding center point is m (C i ), and the boundary condition loss L B based on the working condition classification result is defined as the sum of the distances from all training samples to the feature representation of the corresponding center point, specifically as follows: ; wherein N is the nth training sample, and N is the total number of training samples; According to the exact solution of the amphibious vehicle dynamics model at time t=0, the corresponding initial condition loss L I is defined as the mean square error at the initial time t=0, specifically as follows: ; wherein θ is a parameter of the neural network model; is the predicted value of the neural network model at point (x i , 0); is the exact solution dataset of the amphibious vehicle dynamics model at t=0; the data inner point loss L data is defined as a supervisory signal sampled from the exact solution dataset, and is specifically as follows: ; Wherein, the Is the predicted value of the neural network model at point (x i , t); Is an accurate solution data set of the amphibious vehicle dynamics model at the time t; The loss function L PDE , the boundary condition loss L B , the initial condition loss L I and the data inner point loss L data of the fusion physical mechanism are subjected to fusion weighting to obtain a weighted loss function L, and the method is concretely as follows: ; Wherein lambda PDE 、λ B 、λ I 、λ data is the weight coefficient corresponding to each loss term.
6. The dynamic amphibious vehicle dynamics parameter identification method fused with physical constraints according to claim 5, wherein the step of constructing the loss function L PDE of the fused physical mechanism according to the prediction result of the neural network model and the residual error between the amphibious vehicle dynamics model specifically comprises the following steps: The prediction result of the neural network model is used for calculating the function derivative of the amphibious vehicle dynamics model by combining numerical differentiation and symbolic differentiation through an automatic differentiation technology; Solving an amphibious vehicle dynamics model by utilizing a neural network model, and setting a partial differential equation into a land dynamics equation and a water dynamics equation; The land dynamics equation is specifically as follows: ; Wherein v x is the longitudinal speed at the center of mass, v y is the transverse speed of the vehicle, d is the track width, F xfl 、F xfr 、F xrl 、F xrr is the longitudinal force of the left front wheel, the right front wheel, the left rear wheel and the right rear wheel respectively, F yfl 、F yfr 、F yrl 、F yrr is the lateral force of the left front wheel, the right front wheel, the left rear wheel and the right rear wheel respectively, subscripts ij epsilon { fl, fr, rl, rr } are the rotational angular acceleration of the left front wheel, the right front wheel, the left rear wheel and the right rear wheel respectively, omega ij is the rotational angular acceleration of the wheels, I w is the rotational inertia of the wheels, T ij is the output torque of the hub motor, r e is the effective rolling radius of the wheels, gamma is the angular velocity of the yaw, I z is the rotational inertia of the vehicle, l f 、l r is the distance from the front and rear axle to the center of mass respectively, and delta f is the rotational angle of the front wheels of the automobile; the water dynamics equation is specifically as follows: ; ; ; Wherein, I x 、I y 、I z is rotational inertia around x axis, y axis and z axis respectively, I xx 、I yy 、I zz is rotational additional inertia around x axis, y axis and z axis respectively, K p 、K q 、K r is shape factor, h is draft, L is vehicle length, W is vehicle width, ρ is water density, M x 、M y is restoring moment, M xx 、M yy 、M zz is vehicle sailing resistance of x axis, y axis and z axis respectively, F L 、F R is pushing moment of left and right water jet propeller respectively, p, q and r are side inclination angle speed, longitudinal inclination angle speed and course angle speed of vehicle respectively, C dp 、C dq 、C dr is resistance coefficient; By using The calculation results expressed as the land dynamic equation and the water dynamic equation in the amphibious vehicle dynamic model are shown, and the prediction output of the neural network model is that The following steps are: 。
7. the dynamic identification method of amphibious vehicle dynamics parameters fused with physical constraints according to claim 5, wherein the step of carrying out back propagation based on a weighted loss function, optimizing parameters of a neural network model and vehicle dynamics parameters and realizing dynamic identification of the dynamics parameters specifically comprises the following steps: And (3) performing gradient descent by using an Adam optimizer, setting a learning rate, outputting a neural network model through forward propagation calculation, performing reverse propagation to optimize parameters of the neural network model and vehicle dynamics parameters according to residual errors of a weighted loss function, calculating an L2 relative error according to a change trend of the weighted loss function, judging whether the error is converged, if not, adjusting the learning rate, verifying the neural network model through a test set and actual data, obtaining autonomously identified dynamics parameters, then injecting the autonomously identified dynamics parameters into a vehicle running state database to dynamically update the parameters, predicting the vehicle state after the preset time, performing residual error calculation on a true value and a predicted value, feeding back the residual errors to a feature extraction network or the weight lambda of the weighted loss function as self-adaptive signals, and realizing closed loop self-optimization.
8. A dynamic amphibian kinetic parameter identification system fused with physical constraints, for implementing the dynamic amphibian kinetic parameter identification method fused with physical constraints as claimed in any one of claims 1 to 7, comprising: The data acquisition module is used for acquiring key state quantity in the running process of the vehicle; the key state quantity comprises four-dimensional dynamic information; the working condition identification algorithm module is used for carrying out working condition identification according to the key state quantity based on a preset four-dimensional dynamic clustering method to obtain a working condition classification result; The dynamic optimization module of the fusion physical constraint parameters is used for learning the relevant characteristics of the dynamic parameters from the state data after the working condition classification by adopting the neural network model based on the working condition classification result, constructing a loss function of the fusion physical mechanism based on the amphibious vehicle dynamics model, combining boundary condition loss, initial condition loss and data interior point loss to obtain a weighted loss function of the fusion physical constraint, carrying out back propagation based on the weighted loss function, optimizing the parameters of the neural network model and the dynamic parameters of the vehicle, and realizing dynamic identification of the dynamic parameters.

Description

Dynamic amphibious vehicle dynamic parameter identification method integrating physical constraints Technical Field The invention belongs to the technical field of vehicle dynamics, and particularly relates to a dynamic amphibious vehicle dynamics parameter identification method integrating physical constraints. Background The amphibious vehicle is a multifunctional maneuvering platform capable of freely switching running on land and water, and has the core value of breaking through the limit of single medium, realizing continuous crossing from highways, beaches to rivers and lakes, and meeting urgent requirements for rapid arrival of all terrains in emergency rescue, military deployment and special operation. Along with the evolution of the task mode from fixed-point operation to multitasking dynamic switching such as patrol-transportation-rescue, vehicles frequently pass through land-water boundaries and face complex working conditions of interweaving changes of pavements, shoals, wave areas and the like. In the process, the dynamic behavior of the vehicle is subjected to the coupling action of multiple physical field intensities such as land attachment, fluid dynamics, buoyancy change and the like, so that key dynamic parameters such as mass and moment of inertia, tire-ground/vehicle body-water flow attachment coefficient, hydrodynamic damping and the like are caused, and the vehicle changes in real time along with media, loads and postures, so that strong nonlinearity, time variability and coupling performance are presented. The establishment of a dynamic model capable of accurately describing the amphibious vehicle under different working conditions is a theoretical basis and a key premise for realizing high-performance motion control and autonomous decision. The dynamic identification of all-condition dynamic parameters and the development research of cross-medium dynamic modeling are surrounded, and the aim is that a system reveals an evolution mechanism of parameters under the conditions of cross-medium, variable-condition and multi-task, so that a novel dynamic modeling framework which meets the physical rule and can adaptively learn data characteristics is constructed. The method has important significance for improving the control performance and the operation efficiency of the vehicle in a complex environment, and provides a key technical support for promoting the intelligent development of amphibious equipment. In summary, the core difficulty faced by traditional amphibious vehicle dynamics modeling is the drastic changes in its working media and operating environment. The system is mainly governed by the tire-ground adhesion characteristics and suspension mechanical behaviors when running on land, and is mainly governed by hydrodynamic force, wave-making resistance and buoyancy change when sailing in water. Particularly in transitional working conditions, such as land and water juncture, sailing in waves and soft beach running, the vehicle is simultaneously subjected to the coupling action of the water-land-gas multiphase medium, so that the dynamic parameters show strong nonlinearity, time variability and coupling. The key parameters which must be dynamically identified include the resistance coefficient under different media, the tire-ground/vehicle body-water flow adhesion coefficient, the variation of mass and moment of inertia, the equivalent suspension stiffness and damping coefficient and the like. These parameters are difficult to accurately characterize by conventional empirical models or fixed coefficients, because they vary continuously with medium transitions, attitude disturbances, load transfer, and surface adhesion condition fluctuations, and lack uniform analytical expression. In addition, in the prior art, patent CN202511110991.3 discloses a vehicle dynamics parameter identification method based on constrained nonlinear optimization, which adopts a vehicle transverse dynamics model including a road transverse slope, and takes constraints of vehicle parameters into consideration to construct a nonlinear optimization problem, and simultaneously identifies a front wheelbase (longitudinal distance from a front axle to a vehicle centroid), a rear wheelbase (longitudinal distance from a rear axle to a vehicle centroid), front wheel cornering stiffness, rear wheel cornering stiffness and total vehicle mass of the vehicle. However, the land two-degree-of-freedom vehicle dynamics model established in the method is too simplified, on one hand, the two-degree-of-freedom model generally only covers the longitudinal and transverse basic motions of the vehicle, and cannot characterize multidimensional dynamics such as suspension motions, body pitching and rolling, wheel rotation and the like. This results in a large number of key parameters (such as vertical loads of each wheel, suspension stiffness/damping coefficients, moment of inertia, etc.) being ignored or forced to be combined into a