Tracked vehicle state prediction method and system based on physical constraint dynamics modeling
By employing the GCN-GRU method based on physical constraint dynamics modeling, combined with graph convolutional networks and gated recurrent units, the problem of high-precision prediction for tracked vehicles under complex working conditions is solved. This method achieves accurate description of the state and stability prediction of tracked vehicles, and is applicable to dynamics modeling and control optimization of tracked unmanned transport vehicles.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIANGTAN UNIV
- Filing Date
- 2026-06-26
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies struggle to achieve high-precision prediction of tracked vehicle dynamics modeling under complex working conditions, especially on soft ground, gravel roads, and rugged terrain at engineering sites. Traditional methods involve large computational loads and are prone to accuracy degradation, while deep learning methods have poor generalization in sparse regions and cannot meet the dynamic operation requirements of tracked vehicles.
A physical constraint-based dynamic modeling approach is adopted. By constructing a GCN-GRU model and combining graph convolutional networks and gated recurrent units, the spatial coupling features between the dynamic units of the tracked vehicle are extracted. Furthermore, prior constraints on dynamic physics are introduced, and a joint total loss function is constructed for training to achieve high-precision prediction of the vehicle state.
It achieves high-precision prediction of the state response of tracked vehicles under complex working conditions, maintains the stability and adaptability of the model, and can accurately describe the continuous evolution process under vibration, obstacle crossing and impact conditions. It is suitable for dynamic modeling, state prediction, control optimization and fault diagnosis of tracked unmanned transport vehicles.
Smart Images
Figure CN122491072A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of tracked vehicle dynamics modeling and artificial intelligence technology, and more specifically to a method and system for predicting the state of tracked vehicles based on physical constraint dynamics modeling. Background Technology
[0002] Currently, traditional multibody dynamics modeling methods establish a mechanical structure model of a tracked unmanned transport vehicle and derive dynamic equations based on Newtonian mechanics and Lagrange equations. While this method has clear physical meaning, the model structure is complex, requiring precise mechanical parameters and ground contact models, resulting in a large computational load. This makes it difficult to meet the needs of tracked unmanned transport vehicles for real-time control and rapid obstacle avoidance. Furthermore, under complex road conditions, such as soft ground, gravel roads, and rugged roads at engineering sites, ground parameters fluctuate greatly, and the model accuracy is prone to decline, making it unsuitable for the dynamic requirements of transport operations.
[0003] While deep learning modeling methods use Long Short-Term Memory (LSTM) networks or multilayer perceptrons to fit experimental data, they are fast but have poor generalization in sparse sample regions. Furthermore, the predicted parameters such as acceleration and torque are often discontinuous, violating classical mechanics laws. For example, the speed prediction method for vehicle dynamics modeling based on LSTM in patent CN109783843A is designed for train dynamics modeling. Although it can depict the evolution of state over time, it cannot effectively extract the spatial coupling features of the input data. This makes it unsuitable for dynamic modeling of tracked vehicles in complex road conditions and difficult to accurately depict the complex dynamic characteristics of unmanned vehicles in dynamic operations such as obstacle avoidance, steering, and load changes.
[0004] Although physical information neural networks have been proposed in recent years, they are mainly aimed at solving partial differential equations or tracking conventional vehicle trajectories. They still have the following shortcomings in tracked vehicle scenarios:
[0005] There is a lack of dedicated modeling for the complex contact mechanisms of tracked vehicles: most existing physical information learning methods are geared towards general power systems, fluid systems or simplified rigid body systems, and do not adequately consider the unique contact mechanisms of tracked vehicles, such as track-ground, track-road wheel, and track slip-sink coupling. Limited depth of physical constraint embedding: Some existing methods only treat simple dynamic residuals as additional loss terms, and have not fully coupled the contact switching, energy dissipation and multi-condition boundary conditions of tracked vehicles; Insufficient adaptability to engineering applications: Although some methods theoretically improve fitting accuracy, they still have problems such as large training sample requirements, complex parameter tuning, and difficulty in online deployment.
[0006] Therefore, how to unify and integrate the mechanical structure topology, temporal dynamic response data, and prior constraints of dynamic physics of tracked vehicles in order to achieve high-precision prediction of vehicle state response under complex working conditions is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0007] In view of the above problems, the present invention aims to provide a method and system for predicting the state of a tracked vehicle based on physical constraint dynamics modeling, which overcomes or at least partially solves the above problems.
[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting the state of a tracked vehicle based on physical constraint dynamics modeling includes: S1. Collect multi-source data of the tracked unmanned transport vehicle under different target working conditions. The multi-source data includes at least structural parameters, control parameters, vehicle status data, topology information and terrain information. S2. Construct a dynamic topology graph and corresponding adjacency matrix based on the mechanical connection relationship of the tracked unmanned transport vehicle, and preprocess the collected data and construct time-series samples; S3. Construct a GCN-GRU model, based on the dynamic topology graph, the corresponding adjacency matrix, and time-series samples, and use a graph convolutional network to extract the spatial coupling features between the dynamic units of the tracked vehicle; S4. The spatial coupling features are temporally recursively derived using gated recurrent units to obtain the hidden state at each time step, so as to extract the temporal evolution features of the dynamic response and obtain the preliminary prediction of the dynamic response at the next moment through mapping by a fully connected layer. S5. By introducing prior constraints on dynamic physics through the physical information embedding layer, the preliminary prediction value is corrected based on the residual of the dynamic equation to obtain the predicted value of the target dynamic response; S6. Construct a joint total loss function that includes data error loss and physical constraint loss to perform backpropagation and parameter update, and obtain a trained physical information spatiotemporal recurrent neural network; S7. Real-time acquisition of multi-source data of tracked unmanned transport vehicles under target working conditions, input of trained physical information, spatiotemporal recurrent neural network output vehicle state response prediction.
[0009] Preferably, the structural parameters include, but are not limited to, chassis mass, moment of inertia, number of track rollers, track roller spacing, drive wheel radius, tensioner radius, suspension stiffness, suspension damping, and track pad mass; control parameters include, but are not limited to, drive torque, braking input, steering input, and target speed; state data include, but are not limited to, vehicle displacement, vehicle speed, vehicle acceleration, pitch angle, yaw angle, track roller displacement, track roller speed, suspension deformation, contact load, and vibration response; topology information includes, but is not limited to, the connection relationships between the chassis, track rollers, drive wheels, tensioner, track pads, and suspension units; and terrain information includes a three-dimensional environmental perception map.
[0010] Preferably, in step S2, the vehicle body, left track, right track, left drive wheel, right drive wheel, suspension load unit and ground contact unit are abstracted as graph nodes, and the power transmission relationship, mechanical connection relationship and ground contact coupling relationship are abstracted as graph edges, thus establishing a vehicle dynamics topology graph and corresponding adjacency matrix.
[0011] Preferably, step S2, which involves preprocessing the collected data and constructing time-series samples, includes the following: The multi-source data obtained in step S1 is normalized, missing value imputation is performed, outlier removal is performed, time synchronization and alignment are performed, and noise smoothing is performed. Construct a feature matrix of vehicle walking mechanism nodes based on the preprocessed data:
[0012] in, At time for each node The feature vector representation of , where N is the number of nodes; Time series samples are sliced according to a set time window length to construct time series samples, providing continuous time history information for the gated cyclic unit:
[0013] Where L is the length of the time window.
[0014] Preferably, the graph convolution operation in step S3 is as follows:
[0015] in, For the normalized adjacency matrix, For the first The layer graph convolution weight matrix, where σ is the activation function. .
[0016] Preferably, the prior physical constraints in step S5 include: vehicle body dynamic balance constraints, suspension spring-damping force constraints, track roller and track contact consistency constraints, and drive wheel and tension wheel boundary constraints. The vehicle body dynamics equilibrium constraints are:
[0017]
[0018]
[0019] in, For vehicle body mass, Let the moment of inertia be the pitch axis of the vehicle body. The network predicts the longitudinal and vertical accelerations of the vehicle body. The pitch acceleration predicted by the network. The resultant force and resultant moment are calculated based on the track rollers, tracks, suspension, and ground reaction forces; The suspension spring-damping force constraint is: No. The forces acting on each suspension unit satisfy:
[0020] in, For suspension stiffness, For suspension damping, This is the suspension compression. For suspension compression speed, This refers to the number of suspension units; The contact consistency constraint between the track roller and the track is: During the contact process between the track roller and the track, when the track roller and the track are not in contact, the normal contact indentation amount is... Normal contact force It should approach 0; when the track roller contacts the track, The contact force should satisfy the nonlinear contact model; The boundary constraints between the drive wheel and the tensioner wheel are: The drive wheel and tension wheel in the track system respectively serve as the driving and tensioning boundary, and the boundary state satisfies the requirements of geometric position, velocity continuity and tension constraint.
[0021] Preferably, the target dynamic response prediction values output by the output layer include vehicle body pose, velocity acceleration, dynamic displacement of support rollers, and track stress state; The joint total loss function includes data-driven loss and physical residual loss, specifically:
[0022] in, The data error between the predicted value and the actual value, This is the physical residual error. For regularization terms, These are the weighting coefficients; Backpropagation is performed based on the total loss function to update the network parameters in the graph convolutional layer, gated recurrent unit, and physical information layer until the preset convergence condition is met.
[0023] A tracked vehicle state prediction system based on physical constraint dynamics modeling, based on the aforementioned physical constraint dynamics modeling method for tracked vehicle state prediction, includes: a multi-source data module, a spatiotemporal recurrent neural network, a physical information embedding layer, a training module, and an output layer; The multi-source data module includes a system status input layer and a data preprocessing layer; The spatiotemporal recurrent neural network includes a forward propagation module and a backward propagation module; the forward propagation module includes graph convolutional layers, gated recurrent units, and fully connected layers; The system status input layer is configured to collect multi-source data of the tracked unmanned transport vehicle under the target working conditions. The multi-source data includes at least structural parameters, control parameters, vehicle status data, topology information and terrain information. The data preprocessing layer is configured to construct a dynamic topology graph and corresponding adjacency matrix based on the mechanical connection relationship of the tracked unmanned transport vehicle, and to preprocess the collected data and construct time-series samples. The graph convolutional layer is configured based on the dynamic topology graph, the corresponding adjacency matrix, and the temporal samples. It uses the graph convolutional network to extract the spatial coupling features between the dynamic units of the tracked vehicle. The gated loop unit is configured to perform temporal recursion on the spatial coupling features to obtain the hidden state at each time step in order to extract the temporal evolution features of the dynamic response. A fully connected layer is configured to take the hidden state at each time step as input and map it to obtain the preliminary predicted dynamic response at the next time step; The physical information embedding layer is configured to introduce prior constraints on dynamic physics, and correct the preliminary prediction value based on the residual of the dynamic equation to obtain the target dynamic response prediction value; The training module is configured to construct a joint total loss function that includes data error loss and physical constraint loss to perform backpropagation and parameter updates, thereby obtaining a trained physical information spatiotemporal recurrent neural network. The output layer is configured to collect multi-source data of the tracked unmanned transport vehicle in real time under the target working conditions, and input the trained physical information into the spatiotemporal recurrent neural network to output the vehicle state response prediction.
[0024] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for predicting the state of a tracked vehicle based on physical constraint dynamics modeling.
[0025] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned method for predicting the state of a tracked vehicle based on physical constraint dynamics modeling.
[0026] As can be seen from the above technical solution, compared with the prior art, this invention discloses a method and system for predicting the state of tracked vehicles based on physical constraint dynamic modeling. Addressing the problems of insufficient structural topology representation, inadequate temporal dynamic correlation modeling, and poor physical consistency in existing tracked unmanned platform vehicle dynamic modeling methods, this invention achieves high-precision prediction of vehicle state response under complex working conditions by unifying and integrating the mechanical structure topology of the tracked vehicle, temporal dynamic response data, and prior physical constraints. Specifically: This invention accurately characterizes the complex mechanical topology of tracked vehicles through graph structure modeling and graph convolution operations. By extracting time-dependent features from the vehicle's dynamic response using a gated cyclic unit (GRU), it effectively describes the continuous evolution process under vibration, obstacle crossing, and impact conditions. By introducing constraints from dynamic equations, suspension, and contact, it avoids predictions that do not conform to actual mechanical laws from purely data-driven models. Even with limited training samples or significant variations in operating conditions, it maintains high prediction accuracy and stability. This invention can be applied to scenarios such as dynamic modeling, state prediction, control optimization, fault diagnosis, and digital twin simulation of tracked unmanned transport vehicles. Attached Figure Description
[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0028] Figure 1 This is a schematic diagram of a tracked vehicle state prediction method based on physical constraint dynamics modeling provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the multi-source data structure of the tracked unmanned transport vehicle provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the construction of the dynamic topology of the unmanned transport vehicle provided in this embodiment of the invention; Figure 4 This is a schematic diagram of a spatiotemporal recurrent neural network for physical information provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the training of the spatiotemporal recurrent neural network for physical information provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the vehicle state prediction results provided in an embodiment of the present invention. Detailed Implementation
[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0030] Example 1 This invention discloses a method for predicting the state of a tracked vehicle based on physical constraint dynamics modeling, such as... Figure 1 ,include: S1. Collect multi-source data of the tracked unmanned transport vehicle under different target working conditions. The multi-source data includes at least structural parameters, control parameters, vehicle status data, topology information and terrain information. S2. Construct a dynamic topology graph and corresponding adjacency matrix based on the mechanical connection relationship of the tracked unmanned transport vehicle, and preprocess the collected data and construct time-series samples; S3. Construct a GCN-GRU model, based on the dynamic topology graph, the corresponding adjacency matrix, and time-series samples, and use a graph convolutional network to extract the spatial coupling features between the dynamic units of the tracked vehicle; S4. The spatial coupling features are temporally recursively derived using gated recurrent units to obtain the hidden state at each time step, so as to extract the temporal evolution features of the dynamic response and obtain the preliminary prediction of the dynamic response at the next moment through mapping by a fully connected layer. S5. By introducing prior constraints on dynamic physics through the physical information embedding layer, the preliminary prediction value is corrected based on the residual of the dynamic equation to obtain the predicted value of the target dynamic response; S6. Construct a joint total loss function that includes data error loss and physical constraint loss to perform backpropagation and parameter update, and obtain a trained physical information spatiotemporal recurrent neural network; S7. Real-time acquisition of multi-source data of tracked unmanned transport vehicles under target working conditions, input of trained physical information, spatiotemporal recurrent neural network output vehicle state response prediction.
[0031] To further implement the above technical solutions, such as Figure 2Structural parameters include, but are not limited to, chassis mass, moment of inertia, number of track rollers, track roller spacing, drive wheel radius, tensioner radius, suspension stiffness, suspension damping, and track pad mass; control parameters include, but are not limited to, drive torque, braking input, steering input, and target speed; state data includes, but is not limited to, vehicle displacement, vehicle speed, vehicle acceleration, pitch angle, yaw angle, track roller displacement, track roller speed, suspension deformation, contact load, and vibration response; topology information includes, but is not limited to, the connection relationships between the chassis, track rollers, drive wheels, tensioner, track pads, and suspension units; terrain information includes a three-dimensional environment perception map.
[0032] To further implement the above technical solutions, such as Figure 3 In step S2, the vehicle body, left track, right track, left drive wheel, right drive wheel, suspension load unit and ground contact unit are abstracted as graph nodes, and the power transmission relationship, mechanical connection relationship and ground contact coupling relationship are abstracted as graph edges to establish a vehicle dynamics topology graph and corresponding adjacency matrix. In this embodiment, the graph structure established in step S2 is as follows:
[0033] Where: V is the set of vehicle running gear nodes, including but not limited to chassis, track rollers, drive wheels, and track sections; E is the set of edges, including mechanical connection relationships, contact relationships, or force transmission relationships between nodes; and A is the adjacency matrix.
[0034] To further implement the above technical solution, step S2, which involves preprocessing the collected data and constructing time-series samples, includes the following: The multi-source data obtained in step S1 is normalized, missing value imputation is performed, outlier removal is performed, time synchronization and alignment are performed, and noise smoothing is performed. Construct a feature matrix of vehicle walking mechanism nodes based on the preprocessed data:
[0035] in, At time for each node The feature vector representation of , where N is the number of nodes;
[0036] in, For structural parameter characteristics, As a state response characteristic, To control input features, Features for terrain perception; Time series samples are sliced according to a set time window length to construct time series samples, providing continuous time history information for the gated cyclic unit:
[0037] Where L is the length of the time window. In practical applications, the time window length L is 20 to 200 sampling points, and the step size of adjacent windows is set to 1 to 10 sampling points.
[0038] To further implement the above technical solution, the graph convolution operation in step S3 is as follows:
[0039] in, For the normalized adjacency matrix, For the first The layer graph convolution weight matrix, where σ is the activation function. .
[0040] In practical applications, the number of graph convolutional layers can be set to 2 to 3, and the hidden dimension of each layer can be set to 32, 64 or 128.
[0041] like Figure 4 In this embodiment, a graph convolutional layer is added to the front end of the neural network. Unlike traditional RNNs, LSTMs, or ordinary fully connected networks that directly process raw time-series data, the graph convolutional layer can aggregate neighborhood information based on the mechanical connection topology, enabling the network to learn the spatial coupling effect between the chassis, track rollers, suspension, and tracks, rather than simply using an ordinary fully connected network to process discrete sensor data independently.
[0042] In this embodiment, the spatial feature sequence output in step S4 is input into a gated recurrent unit (GRU) layer. The number of GRU layers is set to 1 to 2, and the number of hidden units is set to 64 or 128. The GRU unit retains and updates historical information through update gates and reset gates. Its recursive process includes:
[0043]
[0044]
[0045]
[0046] in, To update the door, To reset the door, The hidden state at the current moment is represented by ⊙, which is the Hadamard product.
[0047] Existing methods typically employ a data-driven approach to construct temporal networks, failing to integrate the spatial features extracted under mechanical topology constraints with the temporal recursive structure for modeling. This embodiment first obtains spatial coupling features through graph convolution, and then performs temporal recursion using GRU, thereby forming a joint spatial and temporal modeling mechanism that improves the ability to characterize the temporal response to impact, obstacle crossing, and vibration conditions.
[0048] To further implement the above technical solution, the prior dynamic physical constraints in step S5 include: vehicle body dynamic balance constraints, suspension spring-damping force constraints, track roller and track contact consistency constraints, and drive wheel and tension wheel boundary constraints. The vehicle body dynamics equilibrium constraints are: Dynamic residuals are constructed for the longitudinal, vertical, and pitch directions of the vehicle body, respectively:
[0049]
[0050]
[0051] The corresponding vehicle body balance constraint loss is:
[0052] in, For vehicle body mass, Let the moment of inertia be the pitch axis of the vehicle body. The network predicts the longitudinal and vertical accelerations of the vehicle body. The pitch acceleration predicted by the network. The resultant force and resultant moment are calculated based on the track rollers, tracks, suspension, and ground reaction forces; The suspension spring-damping force constraint is: No. The forces acting on each suspension unit satisfy:
[0053] The corresponding suspension constraint loss is:
[0054] in, For suspension stiffness, For suspension damping, This is the suspension compression. For suspension compression speed, For the number of suspension units, For the first Predicted force values for each suspension unit; The contact consistency constraint between the track roller and the track is: During the contact process between the track roller and the track, when the track roller and the track are not in contact, the normal contact indentation amount is... Normal contact force It should approach 0; when the track roller contacts the track, The contact force should satisfy the nonlinear contact model; The corresponding contact consistency constraint loss is:
[0055] in, This refers to the number of contact units between the track roller and the track. For contact stiffness coefficient, It is a non-linear exponent. and These represent the indentation amount and normal contact force predicted by the model, respectively. The boundary constraints between the drive wheel and the tensioner wheel are: In the track system, the drive wheel and tension wheel respectively serve as the driving and tensioning boundary, and the boundary state satisfies the requirements of geometric position, velocity continuity and tension constraint. The corresponding boundary constraint loss is:
[0056] in, For the boundary residual of the drive wheel, The boundary residual of the tensioning wheel is constructed based on the consistency of track entry / exit wheel speeds, boundary displacement relationships, and tension balance relationships. This embodiment incorporates vehicle dynamics equations, suspension spring damping relationships, track-wheel contact constraints, and energy consistency constraints into the loss function or intermediate feature constraints to improve the physical consistency, stability, and generalization ability of the prediction results.
[0057] To further implement the above technical solutions, such as Figure 5 The target dynamic response prediction values output by the output layer include vehicle body pose, velocity acceleration, dynamic displacement of support rollers and track stress state; The joint total loss function includes data-driven loss and physical residual loss, specifically:
[0058] in, The data error between the predicted value and the actual value, This is the physical residual error. For regularization terms, For weighting coefficients; among them, physical residual errors include vehicle body dynamic balance constraint loss, suspension spring-damping force constraint loss, track roller and track contact consistency constraint loss, and drive wheel and tensioner boundary constraint loss; Backpropagation is performed based on the total loss function to update the network parameters in the graph convolutional layer, gated recurrent unit, and physical information layer until the preset convergence condition is met.
[0059] In this embodiment, as Figure 6 For step S7, real-time acquisition of multi-source data of tracked unmanned transport vehicle under target working conditions, including but not limited to control input, state feedback and terrain information, input to the trained physical information spatiotemporal recurrent neural network, output vehicle state response prediction, including but not limited to the speed, attitude, displacement and vibration response of tracked unmanned transport vehicle, and then combining the measured vehicle state with the prediction results to further carry out control decision, trajectory planning and stability assessment.
[0060] Example 2 This embodiment uses acceleration data of a tracked unmanned transport vehicle on a flat, hard road surface to predict the vehicle's longitudinal speed.
[0061] First, the vehicle is controlled to gradually accelerate from a standstill to the target speed according to the set speed command. The on-board controller records the target speed of the left and right tracks, the motor controller records the speed and current of the left and right motors, the IMU records the vehicle's acceleration and attitude angle, and the GNSS / RTK module records the vehicle's position and speed. After sampling, the data from different sources are synchronized according to a unified timestamp, and outliers are removed.
[0062] Then, the vehicle structure parameters, control parameters, vehicle status data and terrain parameters are input into the model, and a dynamic topology diagram is established based on the vehicle's mechanical connection relationship, with the vehicle body, left and right tracks, drive wheels, suspension unit and ground contact unit as graph nodes.
[0063] Next, a graph convolutional network is used to extract the spatial coupling features between vehicle components, and a gated recurrent unit is used to extract the time-varying pattern of the speed response. Simultaneously, a physical constraint on the longitudinal speed of the tracked vehicle is introduced during training.
[0064] in, The speed of the left track. The speed of the right track. This represents the vehicle's longitudinal speed.
[0065] Finally, the model's predicted speed was compared with the measured speed. The results show that the model of this invention can accurately track the speed change during vehicle acceleration, and the root mean square error of speed prediction is smaller than that of the ordinary GRU model and the GCN-GRU model without physical constraints.
[0066] Example 3 This embodiment uses differential steering test data of tracked unmanned transport vehicles as a basis to predict the yaw rate of the vehicle. During the test, the left and right tracks of the vehicle are controlled to run at different speeds to make the vehicle turn. Data such as left and right track speeds, vehicle yaw rate, heading angle, vehicle attitude angle and vehicle position are collected.
[0067] During model training, the differential steering relationship of the tracks is used as a physical constraint.
[0068] in, Let yaw rate be the vehicle's angular velocity. and These represent the speeds of the left and right tracks, respectively. This is the center distance between the left and right tracks.
[0069] By introducing this physical constraint, the model's predictions maintain good consistency with the measured yaw rate during both the initial and stable steering phases. Compared to models that rely solely on data-driven approaches, this invention reduces prediction lag and abrupt errors during steering, improving the model's ability to express the dynamic characteristics of differential steering in tracked vehicles.
[0070] Example 4 This embodiment uses the driving data of tracked unmanned transport vehicles in unstructured environments such as gravel roads, grass roads, or slopes to predict the future location trajectory of the vehicles.
[0071] First, the vehicle travels along a preset path through complex terrain, recording terrain gradient, control commands, vehicle attitude, track speed, and positioning coordinates. Because complex terrain can cause track slippage and vehicle attitude fluctuations, traditional models based solely on kinematic equations struggle to accurately predict vehicle trajectories.
[0072] The method of this invention expresses the spatial coupling relationship between various dynamic units of a vehicle through a graph convolutional network, extracts the temporal evolution features of the vehicle state through a GRU, and restricts the consistency between velocity, yaw rate and displacement changes through a physical information embedding layer. Test results show that the model of this invention can maintain high trajectory prediction accuracy under complex terrain, indicating that the method is suitable for actual operation scenarios of unmanned transport vehicles.
[0073] Example 5 This embodiment provides a tracked vehicle state prediction system based on physical constraint dynamics modeling, and a tracked vehicle state prediction method based on physical constraint dynamics modeling in Embodiment 1, including: a multi-source data module, a spatiotemporal recurrent neural network, a physical information embedding layer, a training module, and an output layer; The multi-source data module includes a system status input layer and a data preprocessing layer; The spatiotemporal recurrent neural network includes a forward propagation module and a backward propagation module; the forward propagation module includes graph convolutional layers, gated recurrent units, and fully connected layers; The system status input layer is configured to collect multi-source data of the tracked unmanned transport vehicle under the target working conditions. The multi-source data includes at least structural parameters, control parameters, vehicle status data, topology information and terrain information. The data preprocessing layer is configured to construct a dynamic topology graph and corresponding adjacency matrix based on the mechanical connection relationship of the tracked unmanned transport vehicle, and to preprocess the collected data and construct time-series samples. The graph convolutional layer is configured based on the dynamic topology graph, the corresponding adjacency matrix, and the temporal samples. It uses the graph convolutional network to extract the spatial coupling features between the dynamic units of the tracked vehicle. The gated loop unit is configured to perform temporal recursion on the spatial coupling features to obtain the hidden state at each time step in order to extract the temporal evolution features of the dynamic response. A fully connected layer is configured to take the hidden state at each time step as input and map it to obtain the preliminary predicted dynamic response at the next time step; The physical information embedding layer is configured to introduce prior constraints on dynamic physics, and correct the preliminary prediction value based on the residual of the dynamic equation to obtain the target dynamic response prediction value; The training module is configured to construct a joint total loss function that includes data error loss and physical constraint loss to perform backpropagation and parameter updates, thereby obtaining a trained physical information spatiotemporal recurrent neural network. The output layer is configured to collect multi-source data of the tracked unmanned transport vehicle in real time under the target working conditions, and input the trained physical information into the spatiotemporal recurrent neural network to output the vehicle state response prediction.
[0074] Example 6 A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements a method for predicting the state of a tracked vehicle based on physical constraint dynamics modeling.
[0075] Example 7 A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a method for predicting the state of a tracked vehicle based on physical constraint dynamics modeling.
[0076] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0077] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A tracked vehicle state prediction method based on physically constrained dynamics modeling, characterized in that, include: S1. Collect multi-source data of the tracked unmanned transport vehicle under different target working conditions. The multi-source data includes at least structural parameters, control parameters, vehicle status data, topology information and terrain information. S2. Construct a dynamic topology graph and corresponding adjacency matrix based on the mechanical connection relationships of the tracked unmanned transport vehicle, and preprocess the collected data and construct time-series samples; S3. Construct a GCN-GRU model, based on the dynamic topology graph, the corresponding adjacency matrix, and time-series samples, and use a graph convolutional network to extract the spatial coupling features between the dynamic units of the tracked vehicle; S4. The spatial coupling features are temporally recursively derived using gated recurrent units to obtain the hidden state at each time step, so as to extract the temporal evolution features of the dynamic response and obtain the preliminary prediction of the dynamic response at the next moment through mapping by a fully connected layer. S5. By introducing prior constraints on dynamic physics through the physical information embedding layer, the preliminary predicted value is corrected based on the residual of the dynamic equation to obtain the predicted value of the target dynamic response; S6. Construct a joint total loss function that includes data error loss and physical constraint loss to perform backpropagation and parameter update, and obtain a trained physical information spatiotemporal recurrent neural network; S7. Real-time acquisition of multi-source data of tracked unmanned transport vehicles under target working conditions, input of trained physical information, spatiotemporal recurrent neural network output vehicle state response prediction.
2. The physically-constrained dynamics modeling based tracked vehicle state prediction method of claim 1, wherein, Structural parameters include, but are not limited to, chassis mass, moment of inertia, number of track rollers, track roller spacing, drive wheel radius, tensioner radius, suspension stiffness, suspension damping, and track pad mass; control parameters include, but are not limited to, drive torque, braking input, steering input, and target speed; state data includes, but is not limited to, vehicle displacement, vehicle speed, vehicle acceleration, pitch angle, yaw angle, track roller displacement, track roller speed, suspension deformation, contact load, and vibration response; topology information includes, but is not limited to, the connection relationships between the chassis, track rollers, drive wheels, tensioner, track pads, and suspension units; terrain information includes a three-dimensional environmental perception map.
3. The physically-constrained dynamics modeling based tracked vehicle state prediction method of claim 1, wherein, In step S2, the vehicle body, left track, right track, left drive wheel, right drive wheel, suspension load unit and ground contact unit are abstracted as graph nodes, and the power transmission relationship, mechanical connection relationship and ground contact coupling relationship are abstracted as graph edges to establish a vehicle dynamics topology graph and corresponding adjacency matrix.
4. The physically-constrained dynamics modeling based tracked vehicle state prediction method of claim 1, wherein, Step S2, the specific content of preprocessing the collected data and constructing time-series samples, is as follows: The multi-source data obtained in step S1 is normalized, missing value imputation is performed, outlier removal is performed, time synchronization and alignment are performed, and noise smoothing is performed. Construct a feature matrix of vehicle walking mechanism nodes based on the preprocessed data: wherein, is the eigenvector representation of each node at time t, N is the number of nodes; Time series samples are sliced according to a set time window length to construct time series samples, providing continuous time history information for the gated cyclic unit: Where L is the length of the time window.
5. The method for predicting the state of a tracked vehicle based on physical constraint dynamics modeling as described in claim 4, characterized in that, The graph convolution operation in step S3 is as follows: in, For the normalized adjacency matrix, For the first The layer graph convolution weight matrix, where σ is the activation function. .
6. The method for predicting the state of a tracked vehicle based on physical constraint dynamics modeling as described in claim 1, characterized in that, The prior physical constraints in step S5 include: vehicle body dynamics balance constraints, suspension spring-damping force constraints, track roller and track contact consistency constraints, and drive wheel and tension wheel boundary constraints. The vehicle body dynamics equilibrium constraints are: in, For vehicle body mass, Let the moment of inertia be the pitch axis of the vehicle body. The network predicts the longitudinal and vertical accelerations of the vehicle body. The pitch acceleration predicted by the network. The resultant force and resultant moment are calculated based on the track rollers, tracks, suspension, and ground reaction forces; The suspension spring-damping force constraint is: No. The forces acting on each suspension unit satisfy: in, For suspension stiffness, For suspension damping, This is the suspension compression. For suspension compression speed, This refers to the number of suspension units; The contact consistency constraint between the track roller and the track is: During the contact process between the track roller and the track, when the track roller and the track are not in contact, the normal contact indentation amount is... Normal contact force It should approach 0; when the track roller contacts the track, The contact force should satisfy the nonlinear contact model; The boundary constraints between the drive wheel and the tensioner wheel are: The drive wheel and tension wheel in the track system respectively serve as the driving and tensioning boundary, and the boundary state satisfies the requirements of geometric position, velocity continuity and tension constraint.
7. The method for predicting the state of a tracked vehicle based on physical constraint dynamics modeling as described in claim 6, characterized in that, The target dynamic response prediction values output by the output layer include vehicle body pose, velocity and acceleration, dynamic displacement of support rollers and track stress state; The joint total loss function includes data-driven loss and physical residual loss, specifically: in, The data error between the predicted value and the actual value, This is the physical residual error. For regularization terms, For weighting coefficients; among them, physical residual errors include vehicle body dynamic balance constraint loss, suspension spring-damping force constraint loss, track roller and track contact consistency constraint loss, and drive wheel and tensioner boundary constraint loss; The vehicle body balance constraint loss is: The suspension spring-damping constraint loss is: in, For the first Predicted force values for each suspension unit; The contact consistency constraint loss between the track roller and the track is: in, and These represent the indentation amount and normal contact force predicted by the model, respectively. The boundary constraint loss between the drive wheel and the tensioner wheel is: in, For the boundary residual of the drive wheel, The boundary residual of the tensioning wheel is constructed based on the consistency of track entry / exit wheel speeds, boundary displacement relationships, and tension balance relationships. Backpropagation is performed based on the total loss function to update the network parameters in the graph convolutional layer, gated recurrent unit, and physical information layer until the preset convergence condition is met.
8. A tracked vehicle state prediction system based on physical constraint dynamics modeling, characterized in that, A method for predicting the state of a tracked vehicle based on physical constraint dynamics modeling according to any one of claims 1-7 includes: a multi-source data module, a spatiotemporal recurrent neural network, a physical information embedding layer, a training module, and an output layer. The multi-source data module includes a system status input layer and a data preprocessing layer; The spatiotemporal recurrent neural network includes a forward propagation module and a backward propagation module; the forward propagation module includes graph convolutional layers, gated recurrent units, and fully connected layers; The system status input layer is configured to collect multi-source data of the tracked unmanned transport vehicle under the target working conditions. The multi-source data includes at least structural parameters, control parameters, vehicle status data, topology information and terrain information. The data preprocessing layer is configured to construct a dynamic topology graph and corresponding adjacency matrix based on the mechanical connection relationship of the tracked unmanned transport vehicle, and to preprocess the collected data and construct time-series samples. The graph convolutional layer is configured based on the dynamic topology graph, the corresponding adjacency matrix, and the temporal samples. It uses the graph convolutional network to extract the spatial coupling features between the dynamic units of the tracked vehicle. The gated loop unit is configured to perform temporal recursion on the spatial coupling features to obtain the hidden state at each time step in order to extract the temporal evolution features of the dynamic response. A fully connected layer is configured to take the hidden state at each time step as input and map it to obtain the preliminary predicted dynamic response at the next time step; The physical information embedding layer is configured to introduce prior constraints on dynamic physics, and correct the preliminary prediction value based on the residual of the dynamic equation to obtain the target dynamic response prediction value; The training module is configured to construct a joint total loss function that includes data error loss and physical constraint loss to perform backpropagation and parameter updates, thereby obtaining a trained physical information spatiotemporal recurrent neural network. The output layer is configured to collect multi-source data of the tracked unmanned transport vehicle in real time under the target working conditions, and input the trained physical information into the spatiotemporal recurrent neural network to output the vehicle state response prediction.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements a tracked vehicle state prediction method based on physical constraint dynamics modeling as described in any one of claims 1 to 7.
10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a tracked vehicle state prediction method based on physical constraint dynamics modeling as described in any one of claims 1 to 7.