Dynamic prediction method of external load
By acquiring data through virtual prototype simulation of working conditions, and establishing a prediction model using Bayesian information criterion and recurrent neural network, dynamic online prediction of loads on multi-track walking devices was achieved. This solved the problem of load information perception under complex working conditions and improved prediction accuracy and applicability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2026-04-20
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies cannot perceive load information of multi-track walking devices under complex working conditions, making it difficult to guarantee work efficiency and safety.
By simulating different driving conditions using a virtual prototype, a dynamic load dataset is obtained. After data preprocessing, the optimal time window length is obtained using the Bayesian information criterion. A prediction model is then established by combining a recurrent neural network, and the rolling time window method is used for online prediction of dynamic loads.
It improves the accuracy and applicability of dynamic load prediction, enabling accurate prediction of dynamic loads on multi-track walking devices under extremely complex conditions, and solves the problem of insufficient sample size.
Smart Images
Figure CN122064996B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of load prediction technology, and in particular relates to a method for dynamic prediction of external loads. Background Technology
[0002] In recent years, to cope with the rapid growth in resource demand, mining equipment has developed towards larger sizes to improve working capacity. Consequently, the tracked walking devices, which bear the load, have evolved from predominantly dual-tracked models to primarily six-tracked models. Multi-tracked walking devices operate in harsh environments with diverse load types. Due to significant differences in road surface compaction, undulation, and friction in mining areas, and the unpredictable interaction between the tracks and the ground, the contact, force, and motion states between the moving parts of the tracks are not constant. Traditional mechanical derivations based on continuous assumptions and steady-state systems are insufficient for real-time and accurate assessment of the dynamic loads on multi-tracked walking devices under complex ground conditions and relative motion states. Therefore, an effective solution is urgently needed to achieve load information perception for multi-tracked walking devices under complex working conditions, in order to better ensure the working efficiency and safety of these devices. Summary of the Invention
[0003] In view of this, the present invention aims to provide a method for dynamic prediction of external loads, so as to solve the problem that the existing technology cannot realize the load information perception of multi-track walking devices under complex working conditions. The present invention improves the applicability and computational efficiency of the dynamic load prediction method to the greatest extent.
[0004] To achieve the above objectives, the technical solution created by this invention is implemented as follows: A method for dynamic prediction of external loads, used to dynamically predict the external loads of a multi-track walking device, specifically includes the following steps: S1: Use a virtual prototype to simulate different driving conditions and obtain the dynamic load dataset of the multi-track walking device. The dynamic load dataset is a set of dynamic load values and characteristic parameters related to the dynamic load at each moment when the multi-track walking device is walking on the ground. The characteristic parameters related to the dynamic load include road excitation, structural properties, motion state and control commands. S2: Perform data preprocessing on the dynamic load dataset to obtain an initial structured dataset; obtain the optimal time window length based on the Bayesian information criterion, and combine it with the initial structured dataset to obtain a recurrent neural network prediction model; S3: The sequence of feature parameters to be detected is obtained based on the rolling time window method, and the sequence of feature parameters to be detected is input into the recurrent neural network prediction model for processing, so as to realize the online prediction of dynamic load of multi-track walking device.
[0005] Furthermore, in step S1, different driving conditions are simulated by changing the physical property parameters of the tracks and the ground, the contact parameters of the tracks and the ground, and the road surface morphology on the virtual prototype, and dynamic mechanical responses are extracted. The driving conditions include straight-line driving, turning, climbing, starting, braking, as well as extreme conditions, large disturbance conditions, and random conditions. Extreme conditions include different turning radii, different climbing angles, and emergency braking conditions. Large disturbance conditions include lock-up, brittle fracture, and settlement conditions. Random conditions include random road surface conditions and random drive conditions.
[0006] Furthermore, in step S1, the virtual prototype randomly switches driving conditions within a continuous time period to simulate the walking process of the multi-track walking device under different driving conditions.
[0007] Furthermore, in step S2, data preprocessing includes filling, cleaning, supplementation, and expansion. Statistical inference and wavelet denoising methods are used to fill and clean the dynamic payload dataset; Gaussian process is used to supplement the dynamic payload dataset; and spatiotemporal labels are used to expand the data space of the dynamic payload dataset to form an initial structured dataset.
[0008] Furthermore, in step S2, obtaining the optimal time window length based on the Bayesian information criterion specifically includes the following steps: S21: Select a set of candidate window lengths {L1, L2, ..., L m}; S22: Based on the length L of each window i By constructing corresponding training datasets using dynamic load values at different times and feature parameters related to dynamic loads, and training recurrent neural networks using training datasets with different window lengths, recurrent neural network models corresponding to different training datasets are obtained. S23: Calculate the sum of squared errors for each recurrent neural network model using the following formula: ; in, Let m be the sum of squared errors corresponding to the m-th recurrent neural network model. Let be the true value of the i-th dynamic load in the training dataset corresponding to the m-th recurrent neural network model. This is the predicted value of the i-th dynamic load in the training dataset corresponding to the m-th recurrent neural network model; S24: Based on the calculation results of step S23, calculate the Bayesian information criterion value corresponding to each recurrent neural network model using the following formula: ; in, The number of parameters used in the calculation of the current recurrent neural network model, is the Bayesian information criterion value corresponding to the m-th recurrent neural network model, and n is the total number of samples in the validation set used to validate the m-th recurrent neural network model; S25: Select the window length corresponding to the minimum value of the Bayesian information criterion as the optimal time window length T, and use the recurrent neural network model corresponding to the optimal time window length T as the recurrent neural network prediction model.
[0009] Furthermore, the rolling window method is used to update the observation data corresponding to the optimal window length: Obtain the window observation data corresponding to the optimal time window length. The window corresponding to the optimal time window length contains characteristic parameters related to the dynamic load for T consecutive time steps. Let the window observation data at time t be... for: ; in, The feature parameter set is constructed from four types of feature parameters: road surface excitation, structural properties, motion state, and control command at time t. for The set of characteristic parameters at time t, for The set of characteristic parameters at time; Based on the window data at each time step, the dynamic load at the next time point is predicted using the following formula: ; in, This is the expression corresponding to the recurrent neural network prediction model. for The predicted value of the dynamic load at any given time; Update the window observation data corresponding to the optimal time window length based on the feature parameter set collected at time t+1. .
[0010] Furthermore, the external load dynamic prediction method also includes dynamic load sensitivity analysis, wherein the specific steps of the sensitivity analysis include: The total variance of the dynamic load is calculated using the following formula. : ; in, The variance contributed to the characteristic parameter i in the dynamic load. The variance contributed by the interaction between characteristic parameter i and characteristic parameter j in dynamic load. The variance contributed by the interaction of characteristic parameters i, j, and l in dynamic loads. The variance contributed by the interaction of all characteristic parameters in the dynamic load; The first-order sensitivity index of the characteristic parameter i in the dynamic load is calculated using the following formula. : ; in, The independent influence of characteristic parameter i on the true value Y of the output dynamic load. Represents the value of feature parameter i The corresponding variance, The value of feature parameter i Regarding the true value of dynamic load The expected value of the condition; The overall effect sensitivity index is calculated using the following formula:
[0011] in, This represents the total effect of characteristic parameter i on the true value Y of the output dynamic load, including the independent effect of characteristic parameter i and the interaction effect of characteristic parameter i with other characteristic parameters. The set of values for all other feature parameters except feature parameter i. The corresponding variance, The set of values for all other feature parameters except feature parameter i. Regarding dynamic load values The expected condition.
[0012] Compared with the prior art, the present invention can achieve the following beneficial effects: This invention presents a dynamic external load prediction method. First, it utilizes virtual prototype simulation to establish a multi-source heterogeneous dynamic load dataset. After cleaning, fusion, and classification, a recurrent neural network is employed to build a dynamic load prediction model. Simultaneously, an adaptive rolling optimal time window length is introduced to improve prediction accuracy, enabling accurate prediction of dynamic loads on multi-track walking devices within arbitrary spatiotemporal coordinates under extremely complex conditions. To address the issue of insufficient sample size for field measured data, this invention maximizes the applicability and computational efficiency of the dynamic load prediction method through multi-source data fusion and clustering. Attached Figure Description
[0013] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A flowchart illustrating the external load dynamic prediction method of an embodiment of the present invention; Figure 2A schematic diagram of the network structure of the recurrent neural network prediction model in an embodiment of the present invention. Detailed Implementation
[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof.
[0015] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.
[0016] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.
[0017] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0018] The invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0019] This invention relates to ground mechanics modeling, online prediction of dynamic loads, and structural safety control of off-highway large tracked equipment traveling on complex mining roads. In order to realize the perception of load information of multi-tracked walking devices under complex working conditions, this invention provides a method for dynamic prediction of external loads that deeply couples discrete element method, multibody dynamics, multi-source heterogeneous data fusion, and recurrent neural network. This method enables millisecond-level online prediction of external loads of multi-tracked walking devices and verification and correction of the model. Discrete element method and multibody dynamics are reflected in virtual prototype simulation. In the process of acquiring data through simulation, multi-source heterogeneous data fusion is the process of fusing four types of data to form a dataset.
[0020] like Figures 1-2 As shown, this invention proposes a dynamic prediction method for external loads, used to dynamically predict the external loads of a multi-track walking device, specifically including the following steps: S1: Use a virtual prototype to simulate different driving conditions and obtain the dynamic load dataset of the multi-track walking device. The dynamic load dataset is a set of dynamic load values and characteristic parameters related to the dynamic load at each moment when the multi-track walking device is walking on the ground. The characteristic parameters related to the dynamic load include road excitation, structural properties, motion state and control commands. S2: Perform data preprocessing on the dynamic load dataset to obtain an initial structured dataset; obtain the optimal time window length based on the Bayesian information criterion, and combine it with the initial structured dataset to obtain a recurrent neural network prediction model; S3: The sequence of feature parameters to be detected is obtained based on the rolling time window method, and the sequence of feature parameters to be detected is input into the recurrent neural network prediction model for processing, so as to realize the online prediction of dynamic load of multi-track walking device.
[0021] It should be noted that road surface excitation parameters include ground physical properties (such as internal friction angle, cohesion, and shear modulus), contact parameters (such as friction coefficient and coefficient of restitution), and road surface morphology (such as roughness, slope, and obstacle distribution). Structural property parameters include the geometric and mechanical parameters of the tracked travel device (such as track shoe dimensions, preload, ground contact state, track roller stiffness, and suspension characteristics). Motion state parameters characterize the real-time dynamic response of the entire machine during operation, including speed, acceleration, attitude angle, and track slip ratio under conditions such as straight travel, turning, climbing, and starting / braking. Control command parameters include active intervention signals such as drive / braking control mode, steering command, and emergency braking trigger. In some embodiments, the road surface excitation is specifically the friction coefficient, the structural property is specifically the track shoe dimensions, the motion state is specifically straight travel, and the control command is specifically the steering command, enabling online prediction of the dynamic load of the multi-tracked travel device.
[0022] Furthermore, the recurrent neural network adopts the LSTM prediction model from "Fault Time Series Prediction Based on LSTM Recurrent Neural Network" published in the Journal of Beijing University of Aeronautics and Astronautics in April 2018.
[0023] In some embodiments, in step S1, different driving conditions are simulated by changing the physical property parameters of the track and the ground, the contact parameters of the track and the ground, and the road surface morphology of the multi-track walking device on a virtual prototype, and dynamic mechanical responses are extracted. The driving conditions include straight-line driving, turning, climbing, starting, braking conditions, as well as extreme conditions, large disturbance conditions, and random conditions. Extreme conditions include different turning radii, different climbing angles, and emergency braking conditions. Large disturbance conditions include lock-up conditions, brittle fracture conditions, and settlement conditions. Random conditions include random road surface conditions and random driving conditions.
[0024] It should be noted that the physical properties of the track and the ground refer to the inherent properties of the materials themselves, which do not change with the contact state or the magnitude of the force. Track-side physical properties include: 1) Track plate material parameters: elastic modulus, Poisson's ratio, density, hardness; 2) Track friction coefficient, stiffness, damping characteristics, etc. Ground-side physical properties include: 1) Soil / road surface material parameters: elastic modulus, Poisson's ratio, density, internal friction angle, cohesion; 2) Asphalt / concrete / soil material stiffness, damping, strength parameters. Contact parameters are the mechanical behavior parameters of the interface between the track and the ground, such as contact stiffness, contact damping, friction coefficient, etc. Road surface morphology refers to the geometric shape, undulations, and texture of the ground surface; it is a spatial geometric feature. Examples include: road unevenness (International Roughness Index IRI), road elevation distribution, elevation sequence, road texture depth, roughness, etc.
[0025] In some embodiments, in step S1, the virtual prototype randomly switches driving conditions within a continuous time period to simulate the walking process of the multi-track walking device under different driving conditions.
[0026] In some embodiments, in step S2, data preprocessing includes filling, cleaning, supplementation and expansion. The dynamic load dataset is filled and cleaned using statistical inference and wavelet denoising methods. The dynamic load dataset is supplemented using a Gaussian process and the data space of the dynamic load dataset is expanded using spatiotemporal labels to form an initial structured dataset.
[0027] It should be noted that (1) the statistical inference filling uses the multiple imputation method, through the EM algorithm, and uses the correlation between variables to estimate and fill missing values. Each data sample in the dynamic load dataset (each data sample consists of two parts: one is the dynamic load value at that moment; the other is all the relevant feature parameters corresponding to the synchronization at that moment: road excitation, structural properties, motion state, control command. The following formula is calculated once for each value of each data sample) conforms to a Gaussian mixture distribution: ; in, It is the mixing coefficient. It is the first Gaussian components, Let be the probability density function of a mixture Gaussian distribution. For the first The mean vector of the Gaussian components. For the first The covariance matrix of a Gaussian component.
[0028] The EM algorithm includes the following steps: a) Calculate the posterior probability that the nth data sample is generated by the kth Gaussian component. : ; Here, j represents the index number of the Gaussian component, the first... The first Gaussian component and the first The Gaussian components are Gaussian components corresponding to different feature parameters, and K is the total number of Gaussian components represented by j.
[0029] b) Update parameters: ; ; ; in, Let n be the feature vector of the nth data sample. For sample size.
[0030] For missing data points, the posterior is calculated in step a) based on the known dimensions, and then the conditional expectation of the missing values is calculated using the posterior and the parameters estimated in step b).
[0031] (2) Wavelet denoising includes the following steps: a) Select the Symlets wavelet basis function and the number of decomposition levels N.
[0032] b) Perform wavelet decomposition on the signal to obtain the coefficients: approximation coefficients (low frequency) and detail coefficients (high frequency).
[0033] c) Threshold the coefficients: ; Among them, threshold A common threshold is typically chosen. , It is an estimate of the noise standard deviation. It is the signal length. These are wavelet coefficients. For symbolic functions, It is a positive part function.
[0034] d) Reconstruct the signal using inverse wavelet transform.
[0035] (3) The process of supplementing the data sample includes the following steps: A Gaussian process (GP) is a nonparametric Bayesian model used to learn the distribution of a function. ; in, It is a learning function. It is the mean function, usually set to 0. It is the covariance function, i.e., the kernel function, where x is the set of characteristic parameters related to the dynamic load at a certain moment. is the single dynamic load value to be predicted at a certain time (the time corresponding to x), and f is the mapping relationship between the dynamic load and the input features.
[0036] a) Choose the covariance function: i.e., the squared exponential function (RBF): ; in, It is the signal variance. It is a length scale. It is the Euclidean distance between two input vectors (numerical combinations of road surface, structure, motion, and control features at two time points).
[0037] b) Given training data ,in, It is a numerical combination of multiple sets of four types of feature parameters. The true value of the dynamic load corresponding to the numerical combination of the four types of characteristic parameters in each group needs to be predicted. function value at The joint distribution is:
[0038] in, Let covariance be the matrix between training points. It is the covariance matrix between training points and prediction points. It is the covariance matrix between two prediction points. It is the variance of observation noise. It is an identity matrix.
[0039] The predicted distribution is then: ; in: ; ; in, It is a multivariate Gaussian distribution. For the prediction point covariance matrix, This is the predicted mean vector.
[0040] New data points are generated by sampling from the predicted distribution, thus supplementing the data.
[0041] (4) Data space expansion includes the following steps: Considering spatiotemporal data, where each data point has spatial coordinates s and a timestamp t, construct a spatiotemporal kernel function: ; in, and It is the RBF kernel function. For the coordinates of another spatiotemporal data point.
[0042] Then, Gaussian processes are used to make predictions at new spatial and temporal points on the spatiotemporal grid, generating new data and achieving expansion.
[0043] In some embodiments, obtaining the optimal time window length based on the Bayesian information criterion in step S2 specifically includes the following steps: S21: Select a set of candidate window lengths {L1, L2, ..., L m}; S22: Based on the length L of each window i By constructing corresponding training datasets using dynamic load values at different times and feature parameters related to dynamic loads, and training recurrent neural networks using training datasets with different window lengths, recurrent neural network models corresponding to different training datasets are obtained. S23: Calculate the sum of squared errors for each recurrent neural network model using the following formula: ; in, Let m be the sum of squared errors corresponding to the m-th recurrent neural network model. Let be the true value of the i-th dynamic load in the training dataset corresponding to the m-th recurrent neural network model. This is the predicted value of the i-th dynamic load in the training dataset corresponding to the m-th recurrent neural network model; S24: Based on the calculation results of step S23, calculate the Bayesian information criterion value corresponding to each recurrent neural network model using the following formula: ; in, The number of parameters used in the calculation of the current recurrent neural network model, is the Bayesian information criterion value corresponding to the m-th recurrent neural network model, and n is the total number of samples in the validation set used to validate the m-th recurrent neural network model; S25: Select the window length corresponding to the minimum value of the Bayesian information criterion as the optimal time window length T, and use the recurrent neural network model corresponding to the optimal time window length T as the recurrent neural network prediction model.
[0044] In some embodiments, a rolling window method is used to update the window observation data corresponding to the optimal window length: Obtain the window observation data corresponding to the optimal time window length. The window corresponding to the optimal time window length contains characteristic parameters related to the dynamic load for T consecutive time steps. Let the window observation data at time t be... for: ; in, The feature parameter set is constructed from four types of feature parameters: road surface excitation, structural properties, motion state, and control command at time t. for The set of characteristic parameters at time t, for The set of characteristic parameters at time; Based on the window data at each time step, the dynamic load at the next time point is predicted using the following formula: ; in, This is the expression corresponding to the recurrent neural network prediction model. for The predicted value of the dynamic load at any given time; Update the window observation data corresponding to the optimal time window length based on the feature parameter set collected at time t+1. .
[0045] A recurrent neural network (RNN) consists of an input layer, hidden layers, recurrent layers, and an output layer. It simultaneously fuses four types of heterogeneous input features—road surface excitation, structural attributes, motion state, and control commands—to capture their temporal dynamic characteristics, achieving accurate real-time prediction of dynamic loads. The RNN model employs an adaptive time window determination mechanism based on the Bayesian information criterion, introducing a rolling time window method to automatically optimize the input sequence length, balancing the utilization of historical information with computational efficiency. The network receives multi-source feature sequences organized according to time windows, extracts temporal patterns through the recurrent layers, and finally outputs the predicted load value through a fully connected layer.
[0046] Bayesian Information Criteria: ; in, It is the maximum likelihood value of the recurrent neural network model. It refers to the number of free parameters in a recurrent neural network model. It is the number of samples used to train the recurrent neural network model.
[0047] (1) Selecting the optimal time window length includes the following steps: a) Select a set of candidate window lengths L1, L2, ..., L m ; b) For each window length L i Construct the corresponding training dataset, where the input to the recurrent neural network model is the past L data including the current time. i The multivariate data (characteristic parameters related to dynamic load) at each time step are output as the dynamic load at the next time step. c) Train using the same recurrent neural network architecture and compute the sum of squared errors (SSE): ; in, Let m be the sum of squared errors corresponding to the m-th recurrent neural network model. Let be the true value of the i-th dynamic load in the training dataset corresponding to the m-th recurrent neural network model. This is the test value of the i-th dynamic load in the training dataset corresponding to the m-th recurrent neural network model.
[0048] d) Calculate the Bayesian Information Criterion (BIC) value for each recurrent neural network model: ; in, The number of parameters in the recurrent neural network model involved in the computation. is the Bayesian information criterion value corresponding to the m-th recurrent neural network model, and n is the total number of samples in the validation set used to validate the m-th recurrent neural network model; e) Select the window length with the smallest BIC as the optimal window length T, and use the recurrent neural network model corresponding to the optimal window length T as the recurrent neural network prediction model.
[0049] The rolling window method includes the following steps: a) Select the window data corresponding to the optimal window length T, which includes multi-source input data (characteristic parameters related to dynamic loads) from the most recent T time steps. Let the window data at time t be: ; in, It is a vector containing four types of features at time t: road surface excitation, structural properties, motion state, and control commands.
[0050] b) Using data from each time step within the window as input, predict the dynamic load at the next time point: ; c) Remove the oldest data in the window and add the new observation data to update the window data: .
[0051] The rolling time window method is introduced to reduce computational costs and minimize the impact of outdated and invalid information on current predictions, thereby improving prediction accuracy. A time window length evaluation index is established based on the Bayesian information criterion to achieve intelligent judgment of time window length and establish an adaptive optimal time window length.
[0052] By adjusting the layers, neurons, and learning rate of the recurrent neural network, and combining the preprocessed classification dataset and the optimal time window length, a real-time prediction model of the recurrent neural network with multiple inputs of road excitation, structural attributes, motion state, and control commands, as well as dynamic load output, is established.
[0053] In some embodiments, the external load dynamic prediction method further includes dynamic load sensitivity analysis, wherein the specific steps of the sensitivity analysis include: The total variance of the dynamic load is calculated using the following formula. : ; in, The variance contributed to the characteristic parameter i in the dynamic load. The variance contributed by the interaction between characteristic parameter i and characteristic parameter j in dynamic load. The variance contributed by the interaction of characteristic parameters i, j, and l in dynamic loads. The variance contributed by the interaction of all characteristic parameters in the dynamic load; The first-order sensitivity index of the characteristic parameter i in the dynamic load is calculated using the following formula. : ; in, The independent influence of characteristic parameter i on the true value Y of the output dynamic load. Represents the value of feature parameter i The corresponding variance, The value of feature parameter i Regarding the true value of dynamic load The expected value of the condition; The overall effect sensitivity index is calculated using the following formula:
[0054] in, This represents the total effect of characteristic parameter i on the true value Y of the output dynamic load, including the independent effect of characteristic parameter i and the interaction effect of characteristic parameter i with other characteristic parameters. The set of values for all other feature parameters except feature parameter i. The corresponding variance, The set of values for all other feature parameters except feature parameter i. Regarding dynamic load values The expected condition.
[0055] Based on the established dynamic load prediction model (i.e., recurrent neural network prediction model), principal component analysis, kernel principal component analysis, and sensitivity analysis methods are used to explore the correlation between ground mechanical parameters, motor drive parameters, structural parameters, control parameters, and dynamic load.
[0056] First, the original data is dimensionality reduced using principal component analysis (PCA) to identify the main parameter combinations affecting the dynamic load, while also reducing collinearity among variables. Then, the original data with nonlinear relationships is processed using kernel principal component analysis (KPI), which maps the data to a high-dimensional space using a kernel function, and then further processes the data using PCA. Finally, sensitivity analysis is used to quantify the influence of ground mechanical parameters, structural parameters, motor drive parameters, and control timing parameters on the dynamic load output.
[0057] Principal component analysis includes the following steps: a) Standardization: Standardize the original parameter matrix X of road surface excitation, structural properties, motion state, and control commands: ; in, It is a mean vector. It is the standard deviation vector. This is the normalized matrix corresponding to the original parameter matrix X.
[0058] b) Calculate the covariance matrix: ; Where n is the number of samples.
[0059] c) Solve the characteristic equation: ; in, It is the i-th eigenvalue. It is the i-th eigenvector.
[0060] d) Principal component selection: Sort the eigenvalues from largest to smallest, and select the first few. The eigenvectors corresponding to each eigenvalue form the projection matrix; e) Projection: Projecting the original data onto the principal component space: ; in, It is a matrix composed of the first k eigenvectors.
[0061] f) Association mining: By analyzing the loadings (i.e. eigenvectors) of principal components, it is possible to explain which original variables dominate each principal component and how these variables collectively influence the dynamic loadings.
[0062] (2) The kernel principal component analysis method includes the following steps: a) Select kernel function: RBF kernel: ; in, It is the width parameter of the RBF core. This is the kernel function.
[0063] b) Calculate the kernel matrix: For For each sample, calculate the matrix. elements :
[0064] in, Is the kernel function at point and points The function value at that location.
[0065] c) Centralized kernel matrix :
[0066] in, yes × A matrix, where each element is 1 / K is the kernel matrix.
[0067] d) Eigenvalue decomposition: Perform eigenvalue decomposition on the centered kernel matrix: ; in, For feature vectors, These are the eigenvalues.
[0068] e) Projection: Projecting the data onto the principal component directions. For a new sample In its first Projection on each principal component for: ; in, It is the sample size. It is the first The i-th component of the feature vector corresponding to each feature. Is the kernel function at point and points The function value at that location.
[0069] f) Association mining: Through kernel principal component analysis, we can discover non-linear combination relationships between original variables.
[0070] Sensitivity analysis methods include the following steps: a) A global sensitivity analysis method based on the Sobol method is adopted, and the estimation is performed using the Monte Carlo method; b) Total variance decomposition of output parameters of recurrent neural network prediction model: ; in, The variance contributed to the characteristic parameter i in the dynamic load. The variance contributed by the interaction between characteristic parameter i and characteristic parameter j in dynamic load. The variance contributed by the interaction of characteristic parameters i, j, and l in dynamic loads. The variance contributed by the interaction of all characteristic parameters in the dynamic load.
[0071] c) Calculate the first-order sensitivity index: ; in, The independent influence of characteristic parameter i on the true value Y of the output dynamic load. Represents the value of feature parameter i The corresponding variance, The value of feature parameter i Regarding the true value of dynamic load The expected condition.
[0072] d) Calculate the overall effect sensitivity index: ; Among them, among them, This represents the total effect of characteristic parameter i on the true value Y of the output dynamic load, including the independent effect of characteristic parameter i and the interaction effect of characteristic parameter i with other characteristic parameters. The set of values for all other feature parameters except feature parameter i. The corresponding variance, The set of values for all other feature parameters except feature parameter i. Regarding dynamic load values The conditional expectation is that the total effect includes the interaction of all input parameters.
[0073] f) Correlation mining: Sensitivity analysis can quantify the impact of each feature parameter on dynamic load, as well as the impact of interactions between feature parameters.
[0074] For the recurrent neural network prediction model, a validation method based on variance decomposition is used to verify its performance, and gradient descent is used to iteratively correct key parameters to improve prediction accuracy and generalization ability.
[0075] First, the model performance is evaluated by analyzing the sources of error in the prediction results of the recurrent neural network prediction model. The expected generalization error of the model is decomposed into three parts: bias, variance, and noise.
[0076] ; in, It is the actual value. Recurrent neural network prediction models are based on input of( The dynamic load prediction value is obtained by considering the set of four types of characteristic parameters (road surface excitation, structural properties, motion state, and control commands) at multiple moments within a selected time window. The expected generalization error is used to evaluate model performance. , representing the difference between the expected value and the true value of the predicted value of the recurrent neural network prediction model; , representing the range of variation of the predicted values of the recurrent neural network prediction model; This is the noise term, representing the error in the data itself.
[0077] The verification steps are as follows: (1) Divide the dataset: Divide the dataset into training set and test set.
[0078] (2) Multiple training sets: Multiple training sets are obtained by resampling, and multiple models are trained.
[0079] (3) Calculate bias and variance: For each sample on the test set, calculate the mean and variance of the predictions from multiple models.
[0080] (4) Analysis: If the deviation is large, it indicates that the model is underfitting; if the variance is large, it indicates that the model is overfitting.
[0081] Furthermore, based on the validation results, the loss function is minimized using gradient descent. This allows for the adjustment of model parameters. .
[0082] The loss function expression is: ; in, The number of training samples; The model is for the input The predicted value; yes The corresponding actual value.
[0083] The update formula for gradient descent is: ; in, It is the parameter value for the t-th iteration; It is the learning rate, used to control the magnitude of each step; Is the loss function in gradient at, It is the parameter value for the (t+1)th iteration.
[0084] The iterative correction steps are as follows: (1) Initialization parameters: Randomly initialize the initial values of the parameters.
[0085] (2) Calculate the gradient: Calculate the gradient of the loss function using the current parameters and training data. The gradient represents the direction in which the loss function rises the fastest, so we update the parameters along the negative gradient direction to reduce the loss.
[0086] (3) Update parameters: Update parameters according to the above formula.
[0087] (4) Iteration: Repeat steps (2) and (3) of the iterative correction steps until the stopping condition is met.
[0088] like Figure 2As shown, the recurrent neural network prediction model of this invention adopts a multi-input single-output deep sequence modeling architecture. This network consists of an input layer, a recurrent hidden layer, and an output layer. The input layer receives multi-source time-series data organized according to the optimal time window length, specifically including four types of feature parameters: road surface excitation, structural attributes, motion state, and control commands. In the figure, X is used as the reference point. t X represents the input vector at the current time step. t-1 and X t+1 These represent the inputs at adjacent time steps, respectively. The recurrent hidden layer achieves state transfer in the time dimension by sharing weight matrices U and W, where U is the weight matrix from the input layer to the hidden layer, responsible for transferring the current input X. t Mapped to the hidden state space, W is the circular weight matrix from hidden layer to hidden layer, responsible for passing historical information between adjacent time steps; V is the weight matrix from hidden layer to output layer; hidden state S t The calculation depends on the current input X. t (Mapped via U) and the hidden state S from the previous time step t-1 (Mapped via W) This demonstrates the network's ability to encode and remember historical information. t-1 and S t+1 These represent the hidden states at adjacent time steps; the output layer (represented by O in the diagram) outputs the hidden state S at the current time step. t The predicted dynamic load values are obtained after linear transformation.
[0089] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.
[0090] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A method for dynamic prediction of external loads, used to dynamically predict the external loads of a multi-track walking device, characterized in that: Specifically, the steps include the following: S1: Use a virtual prototype to simulate different driving conditions and obtain the dynamic load dataset of the multi-track walking device. The dynamic load dataset is a set of dynamic load values and characteristic parameters related to the dynamic load at each moment when the multi-track walking device is walking on the ground. The characteristic parameters related to the dynamic load include road excitation, structural properties, motion state and control commands. S2: Perform data preprocessing on the dynamic load dataset to obtain an initial structured dataset; obtain the optimal time window length based on the Bayesian information criterion, and combine it with the initial structured dataset to obtain a recurrent neural network prediction model; Obtaining the optimal time window length based on the Bayesian information criterion includes the following steps: S21: Select a set of candidate window lengths {L1, L2, ..., L m }; S22: Based on the length L of each window i By constructing corresponding training datasets using dynamic load values at different times and feature parameters related to dynamic loads, and training recurrent neural networks using training datasets with different window lengths, recurrent neural network models corresponding to different training datasets are obtained. S23: Calculate the sum of squared errors for each recurrent neural network model using the following formula: ； in, Let m be the sum of squared errors corresponding to the m-th recurrent neural network model. Let be the true value of the i-th dynamic load in the training dataset corresponding to the m-th recurrent neural network model. This is the predicted value of the i-th dynamic load in the training dataset corresponding to the m-th recurrent neural network model; S24: Based on the calculation results of step S23, calculate the Bayesian information criterion value corresponding to each recurrent neural network model using the following formula: ； in, The number of parameters used in the calculation of the current recurrent neural network model, is the Bayesian information criterion value corresponding to the m-th recurrent neural network model, and n is the total number of samples in the validation set used to validate the m-th recurrent neural network model; S25: Select the window length corresponding to the minimum value of the Bayesian information criterion as the optimal time window length T, and use the recurrent neural network model corresponding to the optimal time window length T as the recurrent neural network prediction model. S3: The sequence of feature parameters to be detected is obtained based on the rolling time window method, and the sequence of feature parameters to be detected is input into the recurrent neural network prediction model for processing, so as to realize the online prediction of dynamic load of multi-track walking device.
2. The method for dynamic prediction of external loads according to claim 1, characterized in that: In step S1, different driving conditions are simulated by changing the physical properties of the tracks and the ground, the contact parameters between the tracks and the ground, and the road surface morphology on a virtual prototype, and the dynamic mechanical response is extracted. The driving conditions include straight-line driving, turning, climbing, starting, braking, as well as extreme conditions, large disturbance conditions, and random conditions. The extreme conditions include different turning radii, different climbing angles, and emergency braking. The large disturbance conditions include lock-up, brittle fracture, and settlement. The random conditions include random road surface conditions and random driving conditions.
3. The method for dynamic prediction of external loads according to claim 1, characterized in that: In step S1, the virtual prototype randomly switches driving conditions within a continuous time period to simulate the walking process of the multi-track walking device under different driving conditions.
4. The method for dynamic prediction of external loads according to claim 1, characterized in that: In step S2, data preprocessing includes padding, cleaning, supplementation, and expansion. Statistical inference and wavelet denoising methods are used to pad and clean the dynamic load dataset. The dynamic load dataset is supplemented using a Gaussian process, and the data space of the dynamic load dataset is expanded using spatiotemporal labels to form an initial structured dataset.
5. The method for dynamic prediction of external loads according to claim 1, characterized in that: The rolling window method is used to update the observation data corresponding to the optimal window length: Obtain the window observation data corresponding to the optimal time window length. The window corresponding to the optimal time window length contains characteristic parameters related to the dynamic load for T consecutive time steps. Let the window observation data at time t be... for: ； in, The feature parameter set is constructed from four types of feature parameters: road surface excitation, structural properties, motion state, and control command at time t. for The set of characteristic parameters at time 1, for The set of characteristic parameters at time; Based on the window data at each time step, the dynamic load at the next time point is predicted using the following formula: ； in, This is the expression corresponding to the recurrent neural network prediction model. for The predicted value of the dynamic load at any given time; Update the window observation data corresponding to the optimal time window length based on the feature parameter set collected at time t+1. .
6. The method for dynamic prediction of external loads according to claim 1, characterized in that: The external load dynamic prediction method also includes dynamic load sensitivity analysis, wherein the specific steps of the sensitivity analysis include: The total variance of the dynamic load is calculated using the following formula. : ； in, The variance contributed to the characteristic parameter i in the dynamic load. The variance contributed by the interaction between characteristic parameter i and characteristic parameter j in dynamic load. The variance contributed by the interaction of characteristic parameters i, j, and l in dynamic loads. The variance contributed by the interaction of all characteristic parameters in the dynamic load; The first-order sensitivity index of the characteristic parameter i in the dynamic load is calculated using the following formula. : ； in, The independent influence of characteristic parameter i on the true value Y of the output dynamic load. Represents the value of characteristic parameter i The corresponding variance, The value of feature parameter i Regarding the true value of dynamic load The expected value of the condition; The overall effect sensitivity index is calculated using the following formula: in, This represents the total effect of characteristic parameter i on the true value Y of the output dynamic load, including the independent effect of characteristic parameter i and the interaction effect of characteristic parameter i with other characteristic parameters. The set of values for all other feature parameters except feature parameter i. The corresponding variance, The set of values for all other feature parameters except feature parameter i. Regarding dynamic load values The expected condition.