A high-speed train bogie model reduced-order efficient digital twin construction method

Abstract

The application discloses a high-speed train bogie model reduced-order efficient digital twinning construction method and relates to the field of digital twinning. In order to meet the real-time monitoring requirement of the dynamic behavior of a high-speed train bogie, actual parameters and operation data of the bogie are acquired first; a bogie model is created based on the actual parameters of the bogie; appropriate load samples are input into the model according to the stress data of bogie operation and displacement responses under the action of the load samples are obtained; modal responses are obtained by reducing the displacement responses by using a fixed boundary modal synthesis method (CB method) and a simplified sample pair set is formed; a radial basis function (RBF) neural network model is trained by using the simplified sample pair set, and a CB-RBF model reduced-order digital twinning model is formed. The method trains a reduced model based on the CB method, greatly improves the construction efficiency of the digital twinning model, the constructed model has high prediction accuracy, and the calculation efficiency is greatly improved compared with a finite element model, and the requirement of real-time monitoring of the dynamic behavior of the bogie can be met.

Description

Technical Field

[0001] This invention belongs to the field of finite element dynamic analysis and digital twin of high-speed train bogies, and specifically relates to a method for constructing a high-speed train bogie model with reduced order and high efficiency. Background Technology

[0002] High-speed trains have become an essential mode of transportation, and the bogie is one of the key components of a high-speed train, its reliability directly determining the overall safety of the train. During long-term service, high-speed train bogies are subject to various loads and factors, making them prone to severe wear, cracks, and even breakage, posing a significant threat to the safe operation of high-speed trains. Therefore, it is necessary to explore a method to rapidly simulate the entire lifecycle of high-speed train bogies, monitor their dynamic behavior in real time, and prevent major accidents through condition detection and hazard warnings. This is of great practical significance for ensuring the long-term safe service of high-speed trains.

[0003] Currently, the finite element method (FEM) is generally used to simulate and analyze the dynamic behavior of bogies, combined with sensor technology for sensing, monitoring, and prediction. However, the FEM requires the establishment of a fine mesh model to ensure the accuracy of the simulation calculation. But due to the large size and complex shape of the bogie structure, the number of meshes in its FEM model can reach millions. This results in a huge amount of computational workload for model solving, which is very time-consuming and cannot meet the speed requirements of real-time monitoring of bogie dynamic behavior.

[0004] To meet the aforementioned speed requirements, digital twin technology offers a good solution: establishing a fast surrogate model to replace the original system, achieving faster solution speed while maintaining accuracy. Generally, there are two approaches. One is finite element model reduction methods (such as the physical subspace method, intrinsic orthogonal decomposition method, modal synthesis method, etc.). These methods reduce the order of a complex finite element model to obtain a smaller, more accurate model. While these methods effectively improve computational efficiency, there is an irreconcilable contradiction between model accuracy and computational efficiency; excessive pursuit of computational efficiency inevitably leads to a sharp drop in accuracy, and vice versa. For the full-cycle simulation and real-time monitoring of bogies, where both simulation accuracy and computational speed are extremely demanding, these methods clearly cannot meet the requirements. The other approach is machine learning methods (such as BP neural networks, RBF neural networks, etc.). These methods train the model with massive amounts of training data, enabling the model to grasp the underlying patterns in the data and thus accurately predict new input data. Machine learning methods can simulate and predict the entire lifecycle of high-speed rail, but directly using machine learning to build models has the problem of low computational efficiency, which is caused by the massive amount of training data consuming a lot of memory.

[0005] In summary, for the real-time monitoring of the dynamic behavior of the large and complex structure of high-speed train bogies, it is extremely important to construct a new digital twin method that is both accurate and has a very fast solution. This method is of great significance for achieving rapid and accurate simulation of the entire bogie lifecycle, real-time monitoring and early warning of dangerous loads, and ensuring the safe operation of high-speed trains. Summary of the Invention

[0006] This invention addresses the shortcomings of existing technologies and solves the problem that current digital twin model simulation technology struggles to accurately and quickly reflect the dynamic behavior of high-speed train bogies throughout their entire operating process. It creatively proposes a method for constructing a high-speed train bogie model using a reduced-order, high-efficiency digital twin.

[0007] To achieve the objectives of this invention, the technical content and technical solution of this invention are as follows:

[0008] A method for constructing a high-speed train bogie model with reduced order and high efficiency includes the following steps:

[0009] S1 acquires the actual physical parameters of the bogie, as well as the bogie operation data collected by sensors under various working conditions;

[0010] S2 creates a three-dimensional geometric model of the bogie in finite element software based on the actual parameters of the bogie.

[0011] S3 uses Latin hypercube sampling to extract data based on the force data of the bogie during operation. One load sample , It is a column vector containing a set of load values;

[0012] S4 obtains the load using the finite element method. Displacement response under action and form an initial sample pair set. The initial sample set is quite complex and needs to be simplified in the next step;

[0013] S5 uses the fixed boundary modal synthesis method (CB method) to analyze the displacement response. The corresponding modal response is obtained by reduction. And form a simplified sample pair set ;

[0014] S6 uses a simplified sample pair set Train the radial basis function (RBF) neural network model to form a reduced-order digital twin model of the CB-RBF model;

[0015] S7 inputs the dynamic data of bogie operation collected by sensors into the CB-RBF model to form a reduced-order digital twin model and views the model output. Then, it compares and verifies the efficiency and feasibility of the formed CB-RBF model reduced-order digital twin model with other methods.

[0016] Furthermore, the specific implementation method of step S3 includes the following steps:

[0017] S31 follows the stratification requirements of Latin hypercube sampling and stratifies the three-dimensional model of the bogie according to material properties and bogie shape and structure.

[0018] S32 randomly samples the load values ​​for each layer based on the load magnitude of each layer, forming a load sample from the randomly sampled load values ​​for each layer. And randomly select n times to record the total sample .

[0019] Furthermore, the specific implementation method of step S4 includes the following steps:

[0020] S41 uses finite element software to sequentially load load samples onto a three-dimensional model and solve for the corresponding displacement response.

[0021] S42 integrates the load samples with the corresponding displacement responses to obtain the initial sample pair set. Directly using the initial sample pair set for subsequent model training results in slow model training speed. Therefore, it is necessary to simplify the initial sample pair set, retain the effective information of the initial sample pair set, and improve the training speed of the model.

[0022] Furthermore, the specific implementation steps of step S5 are as follows:

[0023] S51 divides the bogie system model into several appropriate substructures according to material properties or geometry. This invention demonstrates this by dividing it into two substructures.

[0024] S52 Displacement Response The first coordinate transformation is performed using the modal synthesis method. The formula is as follows:

[0025] Let there be two substructures, a and b.

[0026]

[0027]

[0028] in These are the mode preservation matrices for substructures a and b, respectively. These are the modal coordinates of the two substructures, a and b, respectively.

[0029] Combining the above two equations into a large matrix

[0030]

[0031] The corresponding relationships are as follows:

[0032]

[0033]

[0034]

[0035] S53 utilizes the interface connection conditions of each substructure to perform a second coordinate transformation, eliminating... The modal response is obtained by using non-independent modal coordinates. The formula is as follows:

[0036]

[0037] In the formula and These are the modal coordinates of the principal modes of substructures a and b, respectively. and Let be the modal coordinates of the interface between substructures a and b, respectively. The modal coordinates of the interface between substructures a and b are equal. It is the identity matrix. Derived from two coordinate transformations.

[0038]

[0039] That is, using the fixed boundary modal synthesis method to analyze the displacement response The reduction yields the corresponding low-dimensional modal response.

[0040] S54 combines the load samples and the corresponding reduced modal responses into a simplified sample pair set. .

[0041] Furthermore, the specific implementation method of step S6 is as follows:

[0042] The radial basis function (RBF) neural network model is a commonly used multivariable function interpolation basis function, and its general form can be written as:

[0043]

[0044] In the formula These are the weighting coefficients. Represents any sample point With known sample points Euclidean distance, These are radial basis functions.

[0045] There are various radial basis functions for S61. This invention uses an inverse polyquadratic function as the radial basis function. The inverse polyquadratic function is as follows:

[0046]

[0047] In the formula The value is a positive constant called the smoothing coefficient. The larger the value, the smoother the function value changes, and vice versa. This scheme can use a smoothing coefficient of [value missing]. .

[0048] S62 will set sample pairs Input Radial Basis Function (RBF) Neural Network Model Written in matrix form:

[0049]

[0050] In the formula , The formula is simplified to

[0051]

[0052] S63 determines the weight coefficients of the hidden layers of the radial basis function neural network model by training with a set of sample pairs. Finally, a reduced-order digital twin model of the CB-RBF model is obtained:

[0053]

[0054] In the formula for The column vector formed by these.

[0055] Compared with existing technologies, this invention proposes a method for efficient digital twin construction of high-speed train bogie models with reduced order. The technical solution of this invention has the following beneficial technical effects:

[0056] (1) This method reduces the order of the data through the first and second coordinate transformations of the fixed boundary modal synthesis method, and uses the reduced simplified data to train the RBF neural network model. Compared with the traditional model training method, this greatly improves the model training speed and saves a lot of model construction time.

[0057] (2) This method combines the fixed modal synthesis method with the RBF neural network to construct the CB-RBF model reduced digital twin model, which not only greatly improves the computational efficiency compared with the complete finite element model, and can meet the real-time requirements of the system, but also effectively maintains the accuracy of finite element simulation prediction, and can achieve the purpose of accurate prediction.

[0058] (3) This method has good portability. It can be used not only to build digital twin models of high-speed train bogies, but also to build digital twin models of other large and complex structures according to the implementation steps described in this method. Attached Figure Description

[0059] Figure 1 The steps of the high-speed train bogie model reduction and high-efficiency digital twin construction method described in this invention are as follows:

[0060] Figure 2 The RBF neural network model structure used in this invention

[0061] Figure 3 Data flow diagram of a method for constructing a reduced-order, efficient digital twin of a high-speed train bogie model Detailed Implementation

[0062] 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 only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The invention may also have other embodiments.

[0063] like Figure 1 As shown, a method for constructing a high-speed train bogie model using a reduced-order, efficient digital twin includes the following steps:

[0064] Step 1: Obtain the actual parameters of the bogie and the bogie operating data collected by sensors under various working conditions. The actual parameters of the bogie include structural geometric parameters, process parameters, material parameters, suspension parameters, etc. Obtaining the actual parameters of key components such as the bogie frame, wheelsets, primary suspension, secondary suspension, braking system, and traction system is beneficial for building a 3D model of the bogie in the software; these parameters can be obtained by reading design drawings, actual measurements, etc. Sensors include various sensors used to monitor the bogie's operating status, such as load sensors, acceleration sensors, and displacement sensors. Operating data includes historical operating data and real-time operating data. The main purpose of obtaining the bogie operating data collected by sensors is for subsequent model training.

[0065] Step 2: Based on the actual parameters of the bogie, a 3D geometric model of the bogie is created in finite element software. This process acquires a large number of parameters related to the bogie's actual geometry, materials, suspension, etc., ensuring high fidelity in the 3D model and thus achieving an accurate mirror image of the physical entity in the digital world.

[0066] Step 3: Based on the force data of the bogie during operation, use Latin hypercube sampling to extract... Initial load samples , A column vector containing a set of load values. The Latin hypercube sampling method is one of the most widely used sampling methods. This method utilizes a stratified approach, requiring the bogie's 3D model to be divided into layers based on material properties and the bogie's shape and structure. Then, based on the load magnitude of each layer, the load values ​​of each layer are randomly sampled, forming a load sample. And randomly select n times to record the total sample .

[0067] Step 4: Obtain the load using the finite element method Displacement response under action and form an initial sample pair set. First, the previously created CAE 3D model is imported into the finite element method (FEM) software. Second, the analysis type and element type are determined, and material properties are defined based on the bogie design parameters. Then, an appropriate mesh density is used to ensure simulation accuracy, and load samples are sequentially applied to the 3D model to obtain the corresponding displacement responses. Finally, the load samples and their corresponding displacement responses are collected, exported, and integrated to obtain an initial sample set. If the initial sample pair set is directly used for subsequent model training, the model training speed will be slow. Therefore, it is necessary to simplify the initial sample pair set in the next step, retain the effective information of the initial sample pair set, and improve the training speed of the model.

[0068] Step 5: Use the Fixed Boundary Modal Synthesis (CB) method to analyze the displacement response. The corresponding modal response is obtained by reduction. And form a simplified sample pair set First, the bogie system model is divided into several appropriate substructures based on material properties or geometry. This example uses two substructures to illustrate the displacement response. The first coordinate transformation is performed using the modal synthesis method, as shown in the following formula:

[0069] Let there be two substructures, a and b.

[0070]

[0071]

[0072] in The mode preservation matrices for the two substructures a and b are respectively, and can be obtained directly by computer. These are the modal coordinates of the two substructures, a and b, respectively.

[0073] Next, we will combine the two equations above, because

[0074]

[0075]

[0076]

[0077] Therefore, the two equations can be combined into a large matrix.

[0078]

[0079] Using the interface connection conditions of each substructure, a second coordinate transformation is performed to eliminate... The non-independent modal coordinates are used to realize the assembly of substructures.

[0080]

[0081] In the formula and These are the modal coordinates of the principal modes of substructures a and b, respectively. and Let be the modal coordinates of the interface between substructures a and b, respectively. The modal coordinates of the interface between substructures a and b are equal. The identity matrix is ​​used. The global modal response is derived from two coordinate transformations. The final formula is

[0082]

[0083] The modal responses obtained by reducing the displacement response are combined with the corresponding load samples to form a simplified sample pair set. Record and store data.

[0084] Step Six: Use a simplified sample pair set A radial basis function (RBF) neural network model is trained to form a reduced-order digital twin model of the CB-RBF model. The radial basis function (RBF) neural network model is a commonly used multivariable function interpolation basis function, and its general form can be written as:

[0085]

[0086] In the formula These are the weighting coefficients. Represents any sample point With known sample points Euclidean distance, These are radial basis functions. Because there are various types of radial basis functions, when using sample pairs... Before training a radial basis function (RBF) neural network model, a suitable radial basis function needs to be selected. This scheme uses inverse polyquadratic functions as radial basis functions, as follows:

[0087]

[0088] In the formula The value is a positive constant called the smoothing coefficient. The larger the value, the smoother the function value changes, and vice versa. This scheme can use a smoothing coefficient of [value missing]. .

[0089] A radial basis function (RBF) neural network model consists of an input layer, hidden layers, and an output layer, and its structure is as follows: Figure 3 As shown, the simplified sample pair set will be... Input Radial Basis Function (RBF) Neural Network Model Written in matrix form:

[0090]

[0091] In the formula , The formula is simplified to

[0092]

[0093] Therefore, the weight coefficients of the hidden layers in a radial basis function (RBF) neural network model can be determined through training. Finally, a reduced-order digital twin model of the CB-RBF model is obtained:

[0094]

[0095] In the formula for The column vector formed by these.

[0096] Step 7: Input the dynamic data of the bogie operation collected by sensors into the CB-RBF model's reduced-order digital twin model and view the model output. Then, compare and verify the efficiency and feasibility of the formed CB-RBF model's reduced-order digital twin model with other methods. Specifically, input real-time operating data acquired by sensors, such as load and acceleration, into the CB-RBF model's reduced-order digital twin model to obtain a rapid output response. The output response is then presented on the client side for easy observation and monitoring, achieving the purpose of real-time monitoring. Compare the output response and stress-strain contour plot obtained from the CB-RBF model's reduced-order digital twin model with those obtained by finite element analysis or other methods to verify the efficiency and feasibility of the formed CB-RBF model's reduced-order digital twin model, and appropriately optimize and adjust the model.

Claims

1. A method for constructing a high-speed train bogie model using a reduced-order, high-efficiency digital twin, characterized in that: Includes the following steps: Step 1: Obtain the actual parameters of the bogie, as well as the bogie operation data collected by sensors under various working conditions; Step 2: Based on the actual parameters of the bogie, create a three-dimensional geometric model of the bogie in finite element software; Step 3: Based on the force data of the bogie during operation, use Latin hypercube sampling to extract... One load sample , A column vector containing a set of load values; Step 4: Obtain the load using the finite element method Displacement response under action and form an initial sample pair set. The initial sample set is quite complex and needs to be simplified in the next step; Step 5: Using the fixed boundary modal synthesis method, i.e., the CB method, to analyze the displacement response. The corresponding modal response is obtained by reduction. And form a simplified sample pair set Step five specifically includes the following steps: The bogie system model is divided into several appropriate substructures based on material properties or geometry; Displacement response The first coordinate transformation is performed using the modal synthesis method. ; Using the interface connection conditions of each substructure, a second coordinate transformation is performed to eliminate... The modal response is obtained by using non-independent modal coordinates. ; The load samples and their corresponding reduced modal responses are combined into a sample pair set. ; Step 6: Use a simplified sample pair set Training a radial basis function (RBF) neural network model to form a reduced-order digital twin model of the CB-RBF model, step six specifically includes the following steps: Using inverse quadratic functions as radial basis functions, the sample pair set is... The input radial basis function (RBF) neural network model is expressed as follows: Furthermore, its expression can be written in matrix form: In the formula , The weight coefficients of the hidden layers of the radial basis function neural network model are determined by training with a set of sample pairs. Finally, a reduced-order digital twin model of the CB-RBF model is obtained: ; Step 7: Input the dynamic data of bogie operation collected by sensors into the CB-RBF model reduced to a digital twin model and view the model output. Then compare and verify the efficiency and feasibility of the formed CB-RBF model reduced to a digital twin model with other methods.

2. The method for constructing a high-speed train bogie model using a reduced-order, high-efficiency digital twin according to claim 1, wherein step three includes the following steps: According to the stratification requirements of Latin hypercube sampling, the three-dimensional model of the bogie is stratified based on material properties and bogie shape and structure. Based on the load magnitude of each layer, the load values ​​of each layer are randomly sampled, and the randomly sampled load values ​​from each layer are combined to form a load sample. And randomly select n times to record the total sample .

3. The method for constructing a high-speed train bogie model using a reduced-order, high-efficiency digital twin according to claim 1, wherein step four includes the following steps: In finite element software, load samples are sequentially loaded onto a three-dimensional model and the corresponding displacement response is obtained by solving the model. The load samples and their corresponding displacement responses are integrated to obtain the initial sample pair set. This simplifies the initial sample pair set while retaining the effective information of the initial sample pair set.

Images