Bogie digital twin model construction method for dynamic performance monitoring

Abstract

The invention relates to a construction method of a digital twin model of a bogie for dynamic performance monitoring. The construction method is characterized by comprising steps S1-S2. The method comprises the following steps: S1, establishing a digital twin framework supporting performance monitoring and dynamic visualization of the bogie of the motor train unit; and S2, monitoring the dynamics and component structure performance of the bogie. The method provided by the invention can meet the requirements of performance monitoring accuracy and real-time performance.

Description

Technical Field

[0001] The present invention relates to a method for constructing a digital twin model, and in particular to a method for constructing a digital twin model of a bogie for dynamic performance monitoring. Background Art

[0002] The bogie is a running device that supports, pulls and brakes the car body and can rotate relative to the car body. It usually includes a frame, axle box, spring damping, basic brake and drive device, etc. It has vibration buffering and guiding functions, and determines the safety and quality of EMU operation. However, the operating environment faced by the train is complex and changeable, which can easily lead to the deterioration of the dynamic performance of the bogie, seriously threatening the operation safety. Therefore, the dynamic performance monitoring of the bogie dynamics and key components (wheel sets, frames, etc.) is the key to ensure the service safety of the EMU. The current focus of scholars on the performance monitoring of the service status of the bogie is to collect the vibration, temperature and other data of the key components of the bogie, and use machine learning or artificial intelligence algorithms combined with simulation test data to determine whether each key component is in normal working condition, and predict and diagnose its faults. However, there is still little research on the real-time and accurate evaluation of the dynamic performance of the dynamic performance of the component structure of the EMU bogie during the service stage and the dynamic visualization of the evaluation results.

[0003] With the development and innovation of technologies such as industrial big data and artificial intelligence algorithms, digital twin technology is increasingly being used in major equipment performance monitoring. At the same time, digital twins have also been initially explored in the construction of bogie twin models and performance prediction, and have good application prospects. However, in order to ensure the accuracy of bogie dynamic performance monitoring, bogie digital twin modeling requires the use of a large amount of high-fidelity numerical simulation data and reliable proxy models, and the number of component surface node performance (stress, strain) that needs to be monitored is too large, making it difficult to meet the requirements of monitoring accuracy and real-time performance. Summary of the invention

[0004] A method for constructing a bogie digital twin model for dynamic performance monitoring includes steps S1-S2:

[0005] Step S1: Establish a digital twin framework that supports EMU bogie performance monitoring and dynamic visualization;

[0006] Step S2: Monitor the bogie dynamics and component structural performance.

[0007] Preferably, in step S1, a physical entity layer, a twin data layer, a twin model layer, a data interaction layer, and a twin visualization system are constructed.

[0008] Preferably, the physical entity layer includes six major modules of the bogie entity: the frame, the wheel set axle box and positioning device, the suspension device, the basic braking device, the auxiliary device, the drive device, and the data acquisition device;

[0009] The twin data layer includes bogie mechanism data, sensor data, and performance monitoring data; the mechanism data is obtained by simulating and analyzing the bogie mechanism model under different working conditions; the sensor data includes data such as the temperature, vibration, and speed of the bogie collected; the performance monitoring data includes dynamic performance and component structure performance, the dynamic performance includes wheel-rail force, derailment coefficient, etc., and the component structure performance is represented by the stress and strain values of each grid node of the bogie virtual model;

[0010] The twin model layer includes a geometric model and a data mechanism fusion model; the geometric model is constructed according to the relevant dimensions and assembly relationships of the bogie physical entity, and after simplification, it is used as the virtual model in the twin system; the mechanism data in the data mechanism fusion model is obtained from the twin data layer, and after being trained by artificial intelligence algorithms, a mapping relationship between the measured input and output performance parameters of the bogie is established;

[0011] The data interaction layer includes a data processing module; the data processing module classifies the sensor data through different communication protocols, displays the sensor data in a unified data format readable by machines using data preprocessing technology, and realizes data interaction between the physical entity layer, the twin model layer, the twin data layer, and the twin visualization system by establishing an interaction interface;

[0012] The twin visualization system includes a bogie real-time state simulation module, a monitoring performance visualization module, and a human-computer interaction module; the real-time state simulation module realizes real-time simulation of the running posture of the bogie virtual model based on the sensor data; the monitoring performance visualization module includes a real-time data display module and a structural performance visualization module.

[0013] Preferably, step S2 includes steps S21 - S22:

[0014] Step S21: Construct a digital twin model of the bogie for performance monitoring;

[0015] Step S22: Construct a monitoring visualization prototype system for the bogie.

[0016] Preferably, step S21 specifically includes the following steps S211 - S212:

[0017] Step S211: Lightweight processing of the bogie geometric model;

[0018] Step S212: Construction of the bogie data mechanism fusion model.

[0019] Preferably, in step S211, the lightweighting of the bogie geometric model is realized by using the edge collapse algorithm based on the quadratic error metric. The QEM algorithm is as follows:

[0020] Any plane p in three-dimensional space is represented by Ax + By + Cz + D = 0, where A 2 + B 2 + C 2 = 1; For any vertex v = [x1, y1, z1, 1] in the original bogie model, the square of the distance from v to the plane p is:

[0021] d(v, p) 2 = (p T v) 2 = v T (pp T )v = v T Q p v (1)

[0022] where p = (A, B, C, D) T , Q p i.e., the basic error quadratic form;

[0023] In the original bogie model, the set of all faces adjacent to v is f (v) , then the quadratic error metric of the vertices in the original bogie model is Q v The folding cost is:

[0024]

[0025] In the formula, cost(v) is the folding cost, Q v is the quadratic error metric, Q pf is a matrix related to the vertex v and the face f, which describes the local geometric properties of the vertex v on the face f;

[0026] The specific implementation process of using the QEM algorithm to simplify the bogie model is as follows:

[0027] Step1: Read the three-dimensional model of the bogie and generate lists of vertices, edges, and faces of each model;

[0028] Step2: Select valid point pairs (v i , v j ); Satisfy ||v i - v j || < t, where t is the threshold parameter;

[0029] Step3: Calculate the minimum folding cost cost of each valid point pair (v i , v j )ij with the optimal contraction point v opt ; First, calculate Q opt = Q i + Q j , according to cost ij = v T Q opt v, the optimal contraction point v is obtained opt ;

[0030] Step4: Put all costs ij in a heap, fold the minimum cost ij corresponding point pairs to get v opt , update the costs in the remaining heap ij , and iterate until the set simplification rate is reached.

[0031] Preferably, step S212 includes steps S2121 - S2123:

[0032] Step S2121: Acquisition of bogie mechanism data;

[0033] ① Construct a rigid-flexible coupling dynamics model;

[0034] ② Select the suspension parameters that affect the dynamic performance of the bogie as design variables, and their variation is bounded by floating up and down 50% of the original parameters; Use the Latin hypercube sampling method for experimental design; Select the distribution of sampling points and conduct dynamic simulations to obtain the dynamic performance data sets of the bogie under different working conditions; The performance data obtained from the dynamic simulations are used as the input variables for the bogie structural strength simulations;

[0035] Step S2122: Establish a bogie dynamics surrogate model;

[0036] Step S21221: Adopt a Seq2Seq model integrating an attention mechanism to realize the prediction of the bogie vibration response; Specifically, it includes the following three steps:

[0037] ①The initial Seq2Seq model is trained with the historical measured vibration data of the actual working conditions passing through the straight curve at 120 km / h. The sensors are placed at the axle boxes of each bogie, the corresponding ends of the bogie frame, and the center pin of the car body. The sampling frequency of the sensors is 100 hz, and the total training database is D = {D1, D2, D3, D4, D5, D6, D7, D8, D9}, where D1 - D4 and D5 - D8 are the vertical accelerations on the left and right sides of the bogie frame and axle box respectively, and D9 is the vertical acceleration at the center of the car body. The sampling time is 50 s, with a total of 45,000 data. D1 - D9 are filtered and normalized. The first 40 s are used as the training set, and the last 10 s are used as the test set. They are divided into 100 time steps with a sliding window and used as 9 variables to input the above Seq2Seq model for training;

[0038] ②The MB - SGD model optimization algorithm is used to achieve the self - optimization of the Seq2Seq model in the large - data - scale scenario; new sensing vibration data is added, and a small batch of samples X=(x (1) ,x (2) ,...x (m) ) are randomly drawn from the new sample set, where m is the number of sampled samples. The objective function for establishing the vibration sample loss of the bogie is:

[0039]

[0040] Where: J(θ) is the objective function, x (i) is the input feature vector of the i - th training sample, y (i) is the actual output value of the i - th training sample, (h θ (x (i) ) - y (i) ) 2 is the square of the prediction error of a single sample, h θ (x (i) ) is the predicted value, that is, the output predicted by the model according to the input feature x (i) and the parameter θ;

[0041] The loss function is:

[0042] h θ (x (i) ) = θ0 + θ1x1 (i) + θ2x2 (i) +...+θ n x n (i) (4)

[0043] ③The MB - SGD algorithm updates the parameters of the Seq2Seq model along the reverse gradient to find the global optimal solution. The update process of the model parameter θ is as follows, where α is the learning rate:

[0044]

[0045] By continuously merging new data into the original model and updating the model on the new dataset;

[0046] Step S21222: Construction of the CNN dynamics inversion model; Based on the above Seq2Seq model, obtain the vibration prediction results of each channel of the bogie with a fixed sequence; Construct an inversion model of wheel-rail force and component force based on vibration data:

[0047] ① Use the LHS iterative test to obtain the input sample data, and obtain the time series datasets of vibrations at each measurement point and the corresponding wheel-rail forces and component forces obtained from the simulation analysis under the same working conditions; After preprocessing the vibration data, obtain X m×n =(x1, x2,..., x n ), where m is the length of the sequence sampling, n is the number of sampled channels, normalize the vibration data as the pixel matrix for CNN learning, collect the vibration signal every 10 ms, each collection contains 9 channels, and the signals every 1000 ms form an image sample to invert the wheel-rail force and component force data in the next 1000 ms; Set the mapping labels of the wheel-rail force and component force, define M as the label vector, u and σ are the mean and standard deviation of the wheel-rail force and component force respectively, and λ is the scaling factor, so the mapped force labels are

[0048] ② In the CNN feature extraction layer, convert the vibration acceleration sample into an image signal and extract features, introduce the inverse residual block structure, and gradually extract features by stacking the inverse residual block structure multiple times; In the data reduction layer, perform the reduction of the label mapping data, define the model prediction sequence as G=(g1, g2,...g n ), then is the inversion result of the wheel-rail force and component force predicted by the final model, where is the predicted value of the inversion result of the wheel-rail force and component force, arctan h(G) is the inverse hyperbolic tangent function, and G=(g1, g2,...g n ) is a sequence, and each element g(i) is a component of the sequence; Compare the prediction sequence with the real data label, calculate the loss function, and train the model with the goal of minimizing the loss function to realize the early prediction of the bogie dynamics performance index and component force based on the vibration sensing data in the previous 1 second;

[0049] Step S2123: Construct a bogie structure strength surrogate model;

[0050] The Co-RBF multi-fidelity surrogate model is used to construct the structural strength surrogate model of the bogie components. Among them, the relationship between the high / low-precision models is as follows:

[0051] y H (x) = ρy L (x) + f(x)β + ε(x) (6)

[0052] In the formula, y H (x) is the high-precision model, y L (x) is the low-precision RBF model, ρ is the scale transformation factor, f(x) is the radial basis vector, β is the radial basis coefficient vector, and ε(x) is the unknown error function;

[0053] ① The high / low-precision sample point data of the structural performance of the bogie components used to construct the Co-RBF multi-fidelity surrogate model are respectively:

[0054]

[0055] In the formula, S H is the high-precision sample data, S L is the nth precision sample data, i, j are m-dimensional input x sample vectors, representing the force conditions of the bogie components; y is the output vector, representing the structural performance of the bogie components; nL and nH are the high / low-precision sample numbers respectively, and nL＜nH ;

[0056] The Co-RBF multi-fidelity surrogate model is expressed as:

[0057] y H (x) = [y L (x), f(x)][y L (x H ), F] + y H (9)

[0058] In the formula, F is a matrix, which is a matrix used to calculate y L (x H ) together with y H (x);

[0059] ② Based on the bogie structural strength mechanism data set obtained in step S2121 and the mesh node data obtained by the QEM simplification algorithm in step S211; using the K-nearest neighbor algorithm, update the high / low-precision stress and strain values of each node of the lightweight model; the high / low-precision training samples can be expressed as:

[0060]

[0061]

[0062] Among them, n is the number of grid nodes of the lightweight model;

[0063] ③ Input the high / low-precision samples of each node into the Co-RBF model, and use the low-precision samples to train the RBF model yL ( x ), and incorporate the high-precision samples into yL ( x ) for training.

[0064] Preferably, step S22 specifically includes the following steps S221-S222:

[0065] Step S221: Scene construction and state simulation;

[0066] Step S222: Monitoring performance visualization.

[0067] Preferably, step S221 includes the following steps:

[0068] ① Lightweight the main components of the bogie and complete the format conversion in 3D MAX software;

[0069] ② Use the Animation animation component of Unity3D to produce the vertical / horizontal vibration animation of the bogie frame according to the constructed node relationship.

[0070] Preferably, step S222 includes the following steps:

[0071] ① Visualize the dynamic performance indicators and sensing data of the bogie;

[0072] ② Visualize the real-time cloud map of the structural performance of the bogie components.

[0073] Compared with the prior art, the beneficial effects of the present invention:

[0074] (1) The inventor found in practice that it is difficult for technicians in the prior art to accurately evaluate the dynamics and component structure dynamic performance of the EMU bogie during service in real time. The present invention takes into account the differences between the simulation test data and the actual monitoring data, and also considers the fusion of data and mechanism, which is conducive to improving the speed and accuracy of bogie performance prediction.

[0075] (2) In practice, the inventors found that it is difficult for technicians in the prior art to achieve the dynamic visualization of the evaluation results of the dynamics and component structure dynamic performance during the service of the EMU bogie. On this basis, a visualization prototype system for bogie performance monitoring was developed using Unity3D. Considering the lag caused by too many surface nodes of the components to be monitored on the bogie, the QEM algorithm was used to lightweight the mesh of the bogie three-dimensional model, effectively reducing the number of mesh nodes during the visualization rendering of the bogie performance, meeting the real-time monitoring requirements of the twin model and improving the response rate of the twin system.

[0076] (3) In practice, the inventors found that during the construction of the bogie dynamics proxy model, it is difficult to obtain the real-time wheel-rail force and component forces of the bogie, while the vibration responses of the bogie components are easy to obtain. In response to this, the present invention proposes to use CNN (Convolutional Neural Network) to invert the wheel-rail force and component forces of the bogie.

[0077] (4) In practice, the inventors found that in the prior art, it is impossible to predict the vibration response of the bogie and to achieve real-time monitoring and prediction of the bogie performance. Therefore, the present invention uses a Seq2Seq model integrating an attention mechanism to predict the bogie vibration, and for the scenario of large-scale data of the bogie where the model requires a self-optimization mechanism, the MB-SGD method is used to update the model.

[0078] (5) In practice, the inventors found that during the construction of the bogie structure strength proxy model, due to the fine mesh division of the finite element model in ANSYS Workbench and the large number of high-precision samples required for constructing the global proxy model of the component structure strength, a high computational cost challenge is faced when constructing the proxy model. In response to this, the present invention uses the Co-RBF multi-fidelity proxy model, which can effectively balance the contradiction between the prediction performance of the proxy model and the modeling cost while reducing the computational complexity by integrating high / low-precision model data, facilitating the construction of the bogie digital twin model.

[0079] (6) During the multi-fidelity proxy model process, the inventors found that due to the different numbers of mesh nodes in finite element models of different precisions, each node of each model does not have high / low-precision samples at the same time. Therefore, the Co-RBF model for a single node cannot be constructed. Therefore, the present invention uses the component model lightweighted by the QEM algorithm as the new mesh model, and updates the high / low-precision structural performance (stress, strain) data of each node of the lightweighted component model using the K-Nearest Neighbor (KNN) algorithm based on the high / low-precision sample data, facilitating the construction of a component structure strength proxy model that meets the accuracy and real-time requirements. Description of the Drawings

[0080] Figure 1Schematic diagram of the digital twin framework for bogie performance monitoring;

[0081] Figure 2 Schematic diagram of the modeling process of the digital twin model for bogie performance monitoring;

[0082] Figure 3 Schematic diagram of the process of constructing the bogie mechanism model;

[0083] Figure 4 Schematic diagram of the inversion process of wheel-rail force and component forces based on CNN;

[0084] Figure 5 Schematic diagram of the process of constructing the surrogate model for the structural strength of bogie components;

[0085] Figure 6 Schematic diagram of the visualization system architecture based on the bogie twin model;

[0086] Figure 7 Schematic diagram of the visualization technical route for bogie monitoring performance. Detailed implementation manners

[0087] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some but not all of the embodiments of the present invention.

[0088] Therefore, the following detailed description of the embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely represents some embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0089] It should be noted that, without conflict, the embodiments in the present invention and the features and technical solutions in the embodiments may be combined with each other.

[0090] A method for constructing a digital twin model of a bogie for dynamic performance monitoring, comprising steps S1 - S2:

[0091] Step S1: Establish a digital twin framework that supports the performance monitoring and dynamic visualization of the EMU bogie;

[0092] Step S2: Monitor the dynamics and component structural performance of the bogie.

[0093] Preferably, in step S1, construct a physical entity layer, a twin data layer, a twin model layer, a data interaction layer, and a twin visualization system.

[0094] Preferably, the physical entity layer includes six major modules of the bogie entity: the frame, the wheel set axle box and the positioning device, the suspension device, the basic braking device, the auxiliary device, the driving device, and the data acquisition device.

[0095] Preferably, the twin data layer includes bogie mechanism data, sensor data, and performance monitoring data; the mechanism data is obtained by simulating and analyzing the bogie mechanism model under different working conditions; the sensor data includes data such as the temperature, vibration, and speed of the collected bogie; the performance monitoring data includes dynamic performance and component structure performance. The dynamic performance includes wheel-rail force, derailment coefficient, etc., and the component structure performance is represented by the stress and strain values of each grid node of the bogie virtual model.

[0096] Preferably, the twin model layer includes a geometric model and a data mechanism fusion model; the geometric model is constructed according to the relevant dimensions and assembly relationships of the bogie physical entity, and after simplification, it is used as the virtual model in the twin system; the mechanism data in the data mechanism fusion model is obtained from the twin data layer. After being trained using artificial intelligence algorithms, a mapping relationship between the measured input and output performance parameters of the bogie is established.

[0097] Preferably, the data interaction layer includes a data processing module; the data processing module classifies the sensor data through different communication protocols, uses data preprocessing technology to display the sensor data in a unified data format readable by machines, and realizes data interaction between the physical entity layer, the twin model layer, the twin data layer, and the twin visualization system by establishing an interaction interface.

[0098] Preferably, the twin visualization system includes a bogie real-time state simulation module, a monitoring performance visualization module, and a human-computer interaction module; the real-time state simulation module realizes the real-time simulation of the running posture of the bogie virtual model based on the sensor data; the monitoring performance visualization module includes a real-time data display module and a structural performance visualization module;

[0099] Preferably, the real-time data display module realizes the real-time dynamic display of the measured state data of the bogie and the dynamic and key component structural performance data monitored by the twin model on the Unity3D visualization panel. Of course, other software similar to Unity3D can also be used to realize the real-time data display;

[0100] Preferably, the structural performance visualization module uses the color interpolation algorithm and the mesh reconstruction technology to realize the cloud map visualization display of the structural performance data of each grid node of the monitored component in Unity3D;

[0101] Preferably, the human-computer interaction module realizes the zooming and rotating operations of the bogie virtual model and the switching of the stress / strain cloud maps of each monitored component.

[0102] Step S2 includes steps S21 - S22:

[0103] Step S21: Construct a digital twin model of the bogie for performance monitoring;

[0104] Step S22: Construct a visualization prototype system for bogie monitoring;

[0105] Preferably, step S21 specifically includes the following steps S211 - S212:

[0106] Step S211: Lightweight processing of the bogie geometric model.

[0107] The lightweight processing of the bogie geometric model adopts an edge collapse algorithm based on Quadric Error Metrics (QEM). The principle of the QEM algorithm is as follows:

[0108] Any plane p in three - dimensional space is represented by Ax + By + Cz + D = 0, where A 2 + B 2 + C 2 = 1. For any vertex v = [x1, y1, z1, 1] in the original bogie model, the square of the distance from v to the plane p is:

[0109] d(v, p) 2 = (p T v) 2 = v T (p p T )v = v T Q p v (12)

[0110] where p = (A, B, C, D) T , Q p is the basic error quadratic form.

[0111] In the original bogie model, v the set of all adjacent faces to it is f(v) , then the quadratic error measure of the vertices in the original bogie model is Q v The folding cost of is:

[0112]

[0113] In the formula, cost(v) is the folding cost, Q v is the quadratic error measure, Q pf is a matrix related to the vertex v and the face f, which describes the local geometric properties of the vertex v on the face f.

[0114] Preferably, the specific implementation process of using the QEM algorithm to simplify the bogie model is as follows:

[0115] Step1: Read the 3D model of the bogie and generate lists of vertices, edges, and faces of each model.

[0116] Step2: Select valid point pairs (v i , v j ). Satisfy ||v i - v j || < t, where t is a threshold parameter.

[0117] Step3: Calculate the minimum folding cost cost i , v j ) of each valid point pair and the best contraction point v ij . First, calculate Q opt . First, calculate Q opt = Q i + Q j . According to cost ij = v T Q opt v, obtain the best contraction point v opt .

[0118] Step4: Place all cost ij in a heap, fold the point pair corresponding to the minimum cost ij to obtain v opt , update the cost ij in the remaining heap, and iterate until the set simplification rate is reached.

[0119] Step S212: Construction of the bogie data mechanism fusion model.

[0120] S2121: Acquisition of bogie mechanism data.

[0121] ① Construct a rigid-flexible coupling dynamics model. Mesh the wheelsets and frames of the bogie in the HyperMesh software, import them into ANSYS Workbench to construct a finite element model. Construct a multi-rigid-body dynamics model of the train in SIMPACK. After completing the modal analysis in ANSYS Workbench, write the generated.sub and.cdb files into the generated.fbi flexible body file through the FEMBS interface program in SIMPACK and replace the corresponding rigid structures to complete the construction of the train rigid-flexible coupling dynamics model. All the technologies involved in this step are existing technologies.

[0122] ②Select the suspension parameters that affect the dynamic performance of the bogie as the design variables, and their variation is bounded by floating up and down 50% from the original parameters. Use the Latin Hypercube Sampling (LHS) method for experimental design. Select the sampling point distribution and conduct dynamic simulations to obtain the dynamic performance data sets of the bogie under different working conditions. The performance data obtained from the dynamic simulations are used as the input variables for the bogie structural strength simulation.

[0123] Preferably, taking the key component of the wheel set as an example, the input variables for its experimental design are the vertical wheel-rail forces FZL / FZR and the lateral wheel-rail forces FYL / FYR on the left and right sides; the interference amount δ of the axle and the rotational angular velocity ω, and the output is the stress and strain of the component. Similarly, use LHS for experimental design. Select the sampling point distribution and conduct finite element structural strength simulations to obtain the result performance data sets of the bogie under different working conditions.

[0124] S2122: Establish a bogie dynamics surrogate model.

[0125] S21221: Adopt a Seq2Seq model integrating an attention mechanism to realize the prediction of the bogie vibration response.

[0126] ①Use the historical measured vibration data of the actual working conditions of passing through straight and curved lines at 120 km / h to train the initial Seq2Seq model. The sensors are placed at the axles of each bogie, the corresponding ends of the frame, and the center pin of the car body. The sensor sampling frequency is 100 hz, and the total training database is D = {D1, D2, D3, D4, D5, D6, D7, D8, D9}, where D1 - D4 and D5 - D8 are the vertical accelerations on the left and right sides of the bogie frame and axle boxes respectively, and D9 is the vertical acceleration of the car body center. The sampling time is 50 s, with a total of 45000 data. Filter and normalize D1 - D9, use the first 40 s as the training set, and the last 10 s as the test set. Divide them into 100 time steps with a sliding window and input them into the above Seq2Seq model as 9 variables for training.

[0127] ②Use the MB-SGD model optimization algorithm to realize the self-optimization of the Seq2Seq model in the large data scale scenario. The MB-SGD model optimization algorithm combines the Gradient Descent (GD) method and the Stochastic Gradient Descent (SG) method. At each iteration, a small batch of samples is used to estimate the gradient of the loss function to update the model parameters, avoiding training the entire data set when new data is added. In this way, it can solve quickly and maintain high accuracy, greatly reducing the time and storage space for model update under the large data flow and enabling online update.

[0128] Preferably, new sensing vibration data is added, and a small batch of samples X = (x (1) , x (2) ,... x (m) ) are randomly drawn from the new sample set, where m is the number of samples drawn. The objective function for the bogie vibration sample loss is established as follows:

[0129]

[0130] Where: J(θ) is the objective function, x (i) is the input feature vector of the i-th training sample, y (i) is the actual output value of the i-th training sample, (h θ (x (i) ) - y (i) ) 2 is the square of the prediction error of a single sample, h θ (x (i) ) is the predicted value, that is, the output predicted by the model according to the input feature x (i) and the parameter θ.

[0131] The loss function is:

[0132] h θ (x (i) ) = θ0 + θ1x1 (i) + θ2x2 (i) +... + θ n x n (i) (15)

[0133] ③ The MB-SGD algorithm updates the Seq2Seq model parameters along the reverse gradient to find the global optimal solution. The update process of the model parameter θ is as follows, where α is the learning rate:

[0134]

[0135] By continuously merging new data into the original model and realizing the update of the model on the new data set.

[0136] S21222: Construction of the CNN dynamics inversion model.

[0137] Based on the above Seq2Seq model, the vibration prediction results of each channel of the bogie with a fixed sequence are obtained. Since the dynamic performance indexes (wheel-rail vertical / horizontal force, derailment coefficient, axle lateral force, and wheel load reduction rate) during the operation stage of the bogie can be obtained from the wheel-rail vertical / horizontal force, an inversion model of wheel-rail force and component force based on vibration data is constructed.

[0138] ① The input sample data is obtained by using the LHS iterative test, and the time series data sets of the vibration at each measuring point, the corresponding wheel-rail force, and the component force obtained by simulation analysis in SIMPACK under the same working conditions. After preprocessing the vibration data, X is obtained. m×n =(x1, x2,..., x n ), where m is the length of the sequence sampling and n is the number of sampling channels. The vibration data is normalized and used as the pixel matrix for CNN learning. The vibration signal is collected every 10 ms, and each collection contains 9 channels. The signals every 1000 ms form an image sample to invert the wheel-rail force and component force data in the next 1000 ms. Set the mapping labels for the wheel-rail force and component force, define M as the label vector, u and σ are the mean and standard deviation of the wheel-rail force and component force respectively, and λ is the scaling factor. Therefore, the mapped force labels are

[0139] ② In the CNN feature extraction layer, the vibration acceleration samples are converted into image signals and features are extracted. The inverse residual block structure is introduced, and features are gradually extracted by stacking the inverse residual block structure multiple times. In the data reduction layer, the reduction of the label mapping data is performed. Define the model prediction sequence as G=(g1, g2,...g n ), then is the inversion result of the wheel-rail force and component force predicted by the final model, where is the predicted inversion result value of the wheel-rail force and component force, arctanh(G) is the inverse hyperbolic tangent function, and G=(g1, g2,...g n ) is a sequence, and each element g(i) is a component of the sequence. Compare the prediction sequence with the real data label, calculate the loss function, and train the model with the goal of minimizing the loss function to achieve the early prediction of the bogie dynamic performance index and component force based on the vibration sensing data in the previous 1 second.

[0140] S2123: Construct a surrogate model for the bogie structure strength.

[0141] Use the Co-RBF multi-fidelity surrogate model to construct a surrogate model for the bogie component structure strength. Among them, the relationship between the high / low-precision models is as follows:

[0142] y H (x)=ρy L (x)+f(x)β+ε(x) (17)

[0143] In the formula, y H (x) is the high-precision model, y L (x) is the low-precision RBF model, ρ is the scale transformation factor, f(x) is the radial basis vector, β is the radial basis coefficient vector, and ε(x) is the unknown error function.

[0144] ① The high / low-precision sample point data of the bogie component structural performance for constructing the Co-RBF multi-fidelity surrogate model are respectively:

[0145]

[0146] In the formula, S H is the high-precision sample data, S L is the low-precision sample data, i, j are the m-dimensional input x sample vectors, representing the force conditions of the bogie components; y is the output vector, representing the bogie component structural performance; n L and n H are the high / low-precision sample quantities respectively, and n L < n H .

[0147] The Co-RBF multi-fidelity surrogate model can be expressed as:

[0148] y H (x) = [y L (x), f(x)][y L (x H ), F] + y H (20)

[0149] In the formula, F is a matrix, which is used to calculate y L (x H ) together with y H (x) matrix.

[0150] ② Based on the bogie structural strength mechanism data set obtained in step S2121 and the grid node data obtained by the QEM simplification algorithm in step S211. Using the K-Nearest Neighbor (KNN) algorithm, update the high / low-precision stress and strain values of each node of the lightweight model. The high / low-precision training samples can be expressed as:

[0151]

[0152] Among them, n is the number of grid nodes of the lightweight model.

[0153] Preferably, taking the wheelset, a key component of a certain type of bogie, as an example, a surrogate model for the structural strength of the wheelset is constructed. The high-precision wheelset model is constructed in HyperMesh, with a total of 451,074 nodes; the low-precision wheelset model is automatically generated by 20-mm tetrahedral meshes in ANSYS Workbench, with a total of 250,044 nodes. After multiple iterative experiments, finally, 99 groups of high-precision sample design variables and 200 groups of low-precision sample design variables are selected by LHS, and the stress and strain values of each node of the high / low-precision wheelset models are obtained through simulation in ANSYS Workbench respectively. The lightweight wheelset model has a total of 8,584 mesh nodes at a 50% simplification rate. The K-Nearest Neighbor (KNN) algorithm is used to update the high / low-precision stress and strain values of each node of the lightweight model. The final high / low-precision training samples are obtained:

[0154]

[0155] where 8,584 is the number of simplified mesh nodes, 99 and 200 are the sample data obtained from iterative experiments, and are the high / low-precision training samples of a single node of the lightweight model respectively, yH and yL are the high / low-precision stress and strain values respectively.

[0156] ③ Input the high / low-precision samples of each node into the Co-RBF model, and use the low-precision samples to train the RBF model yL(x) , and incorporate the high-precision samples into yL(x) for training.

[0157] Preferably, step S22 specifically includes the following steps S221 - S222:

[0158] Step S221: Scenario construction and state simulation.

[0159] ① Lightweight the main components of the bogie (such as wheelset, frame, axle box, etc.) and complete format conversion in the 3D MAX software.

[0160] Preferably, import the formed.fbx model file into Unity3D as the virtual model of the bogie, and at the same time construct the matching three-dimensional models of the track and the car body, convert the formats, and then import them into Unity3D to complete the construction of the bogie operation scenario.

[0161] ② According to the constructed node relationship, use the Animation animation component of Unity3D to produce the vertical / horizontal vibration animation of the frame.

[0162] Preferably, taking the vibration acceleration value measured by the sensor as the trigger, the animation is switched by C# programming and the playback rate of each animation is controlled to effectively simulate the running state of the bogie driven by the measured sensing data.

[0163] Step S222: Visualize the monitoring performance.

[0164] ① Visualize the dynamic performance indicators and sensing data of the bogie.

[0165] Preferably, through the TCP communication protocol of the socket communication mechanism, the python side where the data mechanism fusion model is located is used as the server, and the Unity3D platform is used as the client in the data interaction layer to realize the data interaction between the twin model layer and the twin system layer.

[0166] Preferably, driven by the real-time sensing vibration data, the dynamic proxy model obtains the dynamic performance indicators of the bogie at that moment, which are transmitted to Unity3D through socket communication. Through the UGUI component in Unity3D, the values of the received dynamic performance indicators are read by the C# script and presented on the visualization interface, and are dynamically updated over time. In addition, for the vibration data measured by the sensor, the Line chart in the Unity Xchart plugin is used to realize the visual dynamic display of the vibration state at the monitoring points of the bogie.

[0167] ② Visualize the real-time cloud map of the structural performance of the bogie components.

[0168] Preferably, the lightweight bogie component model is saved in the.stl format and the coordinates and indices of each mesh node are exported, and then imported into Unity3D to allocate vertices and triangles to complete the reconstruction of each mesh of the virtual model component.

[0169] Preferably, driven by the real-time vibration data, the dynamic proxy model outputs the component force data, drives the component structural strength proxy model to monitor the structural performance (stress, strain) of each mesh node of the component in real time, and sends the data to Unity3D through socket communication. Finally, Unity3D assigns different color information to each point of the reconstructed component mesh according to the received node stress (strain) value size by means of the color interpolation algorithm, and updates over time to complete the real-time dynamic display of the component structural performance cloud map.

[0170] The above embodiments are only used to illustrate the present invention rather than to limit the technical solutions described in the present invention. Although the present specification has described the present invention in detail with reference to the above respective embodiments, the present invention is not limited to the above specific embodiments. Therefore, any modification or equivalent replacement of the present invention; and all technical solutions and their improvements that do not depart from the spirit and scope of the invention are covered by the scope of the claims of the present invention.

Claims

1. A method for constructing a bogie digital twin model for dynamic performance monitoring, characterized in that: The method comprises steps S1-S2: Step S1: Establish a digital twin framework that supports EMU bogie performance monitoring and dynamic visualization; Step S2: Monitor the bogie dynamics and component structural performance.

2. A method for constructing a bogie digital twin model for dynamic performance monitoring according to claim 1, characterized in that: In step S1, the physical entity layer, twin data layer, twin model layer, data interaction layer, and twin visualization system are constructed.

3. The method for constructing a bogie digital twin model for dynamic performance monitoring according to claim 2, characterized in that: The physical entity layer includes six modules of the bogie entity: frame, wheelset axle box and positioning device, suspension device, basic braking device, auxiliary device and drive device, and data acquisition device; The twin data layer includes bogie mechanism data, sensor data, and performance monitoring data. The mechanism data is obtained by simulating and analyzing the bogie mechanism model under different working conditions. The sensor data includes the collected data such as bogie temperature, vibration, and speed. The performance monitoring data includes dynamic performance and component structural performance. The dynamic performance includes wheel-rail force, derailment coefficient, etc. The component structural performance is expressed as the stress and strain values ​​of each grid node of the bogie virtual model. The twin model layer includes a geometric model and a data-mechanism fusion model. The geometric model is constructed based on the relevant dimensions and assembly relationships of the bogie physical entity and is simplified to serve as a virtual model in the twin system. The mechanism data in the data-mechanism fusion model is obtained from the twin data layer and, after training with an artificial intelligence algorithm, a mapping relationship between the measured input and output performance parameters of the bogie is established. The data interaction layer includes data processing modules; The data processing module classifies sensor data through different communication protocols, uses data preprocessing technology to display sensor data in a unified machine-readable data format, and establishes interactive interfaces to achieve data interaction between the physical entity layer, twin model layer, twin data layer, and twin visualization system; The twin visualization system includes a bogie real-time status simulation module, a monitoring performance visualization module, and a human-computer interaction module; The real-time state simulation module realizes the real-time simulation of the running posture of the bogie virtual model based on the sensor data; the monitoring performance visualization module includes the real-time data display module and the structural performance visualization module.

4. A method for constructing a bogie digital twin model for dynamic performance monitoring according to claim 3, characterized in that: Step S2 includes steps S21-S22: Step S21: constructing a digital twin model of a bogie for performance monitoring; Step S22: construct a bogie monitoring visualization prototype system.

5. The method for constructing a bogie digital twin model for dynamic performance monitoring according to claim 4, characterized in that: Step S21 specifically includes the following steps S211-S212: Step S211: Lightweighting of the bogie geometric model; Step S212: constructing a bogie data mechanism fusion model.

6. A method for constructing a bogie digital twin model for dynamic performance monitoring according to claim 5, characterized in that: In step S211, the geometric model of the bogie is lightweighted by using an edge folding algorithm based on quadratic error measurement. The QEM algorithm is as follows: Ax+By+Cz+D=0 represents any plane p in three-dimensional space, where A 2 +B 2 +C 2 =1; any vertex v in the original bogie model = [x1, y1, z1, 1], then the square of the distance from v to plane p is: d(v,p) 2 =(p T v) 2 =v T (pp T )v=v T Q p v (1) Where p = (A, B, C, D) T , Q p That is, the quadratic form of the basic error; In the original bogie model, the set of v and all its adjacent faces is f(v) , then the quadratic error measure of the vertex of the original bogie model is Q v The folding cost is: Where cost(v) is the folding cost, Q v is the quadratic error measure, Q pf is a matrix associated with vertex v and face f, which describes the local geometric properties of vertex v on face f; The specific implementation process of using the QEM algorithm to simplify the bogie model is as follows: Step 1: Read the bogie 3D model and generate a list of vertices, edges and faces of each model; Step 2: Select valid point pairs (v i ,v j ); satisfy ||v i -v j ||<t, where t is the threshold parameter; Step 3: Calculate each valid point pair (v i ,v j )’s minimum folding cost ij With the best contraction point v opt ; First calculate Q opt =Q i +Q j , according to cost ij =v T Q opt v, get the best contraction point v opt ; Step 4: All costs ij Put it in a pile and fold the minimum cost ij Corresponding point pairs get v opt , update the cost in the remaining heap ij , iteratively proceeds until the set simplification rate is reached.

7. A method for constructing a bogie digital twin model for dynamic performance monitoring according to claim 6, characterized in that: Step S212 includes steps S2121 to S2123: Step S2121: acquiring bogie mechanism data; ①Construct a rigid-flexible coupling dynamic model; ② Select the suspension parameters that affect the dynamic performance of the bogie as the design variables, and set the fluctuation limit of the original parameters to 50%; use the Latin hypercube sampling method to carry out the experimental design; The sampling point distribution is selected and dynamic simulation is performed to obtain the dynamic performance data set of the bogie under different working conditions; the performance data obtained by the dynamic simulation is used as the input variable of the bogie structural strength simulation; Step S2122: establishing a bogie dynamics proxy model; Step S21221: adopt a Seq2Seq model integrating attention mechanism to predict the vibration response of the bogie; specifically, the following three steps are included: ① The initial Seq2Seq model is trained with historical measured vibration data of the actual working condition of passing through a straight curve at 120km / h. Sensors are placed at each axle box of the bogie and its corresponding frame end and the center pin of the car body; the sensor sampling frequency is 100hz, and the total training database is D={D1,D2,D3,D4,D5,D6,D7,D8,D9}, where D1-D4, D5-D8 are the vertical accelerations on the left and right sides of the bogie frame and axle box, and D9 is the vertical acceleration of the car body center; the sampling time is 50s, with a total of 45,000 data. D1-D9 are filtered and normalized, the first 40s are used as the training set, and the last 10s are used as the test set. The sliding window is divided into 100 time steps, and the 9 variables are input into the above Seq2Seq model for training; ② Use the MB-SGD model optimization algorithm to realize the self-optimization of the Seq2Seq model in the big data scale scenario; add new sensor vibration data and randomly extract a small batch of samples X=(x (1) ,x (2) ,...x (m) ), m is the number of samples drawn, and the objective function of the bogie vibration sample loss is established as: Where: J(θ) is the objective function, x (i) is the input feature vector of the i-th training sample, y (i) The actual output value of the i-th training sample, (h θ (x (i) )-y (i) ) 2 The square of the prediction error for a single sample, h θ (x (i) ) is the predicted value, that is, the model is based on the input feature x (i) and the output predicted by the parameters θ; The loss function is: h θ (x (i) )=θ0+θ1x1 (i) +θ2x2 (i) +...+θ n x n (i) (4) ③The MB-SGD algorithm updates the Seq2Seq model parameters along the reverse gradient to find the global optimal solution. The update process of the model parameter θ is as follows, and α is the learning rate: By continuously incorporating new data into the original model and updating the model on the new data set; Step S21222: construct a CNN dynamic inversion model; based on the above Seq2Seq model, obtain the vibration prediction results of each channel of the bogie in a fixed sequence; construct a wheel-rail force and component force inversion model based on vibration data: ①Use LHS iteration test to obtain input sample data, and obtain the vibration of each measuring point and the corresponding wheel-rail force and component stress time series data set under the same working conditions in simulation analysis; obtain X m×n =(x1,x2,...,x n ), m is the length of the sequence sampling, n is the number of sampling channels, the vibration data is normalized as the pixel matrix of CNN learning, the vibration signal is collected every 10ms, each collection contains 9 channels, and the signal of every 1000ms constitutes an image sample to invert the wheel-rail force and component force data of the next 1000ms; set the mapping label of the wheel-rail force and component force, define M as the label vector, u and σ are the mean and standard deviation of the wheel-rail force and component force respectively, λ is the scaling factor, so the labels of each force after mapping are ② In the CNN feature extraction layer, the vibration acceleration samples are converted into image signals and features are extracted. The anti-residual block structure is introduced, and features are gradually extracted by stacking the anti-residual block structure multiple times. In the data restoration layer, the label mapping data is restored, and the model prediction sequence is defined as G = (g1, g2, ... g n ),but The wheel-rail force and component force inversion results predicted by the final model, where is the predicted wheel-rail force and component force inversion result value, arctanh(G) is the inverse hyperbolic tangent function, G=(g1,g2,...g n ) is a sequence, where each element g(i) is a component of the sequence; the predicted sequence is compared with the real data label, the loss function is calculated, and the model is trained with the goal of minimizing the loss function to achieve early prediction of the bogie dynamic performance indicators and component forces based on the vibration sensor data of the previous 1 second; Step S2123: constructing a bogie structural strength proxy model; The Co-RBF multi-fidelity proxy model is used to build the structural strength proxy model of the bogie components; the relationship between the high / low precision models is as follows: y H (x)=ρy L (x)+f(x)β+ε(x) (6) In the formula, y H (x) is a high-precision model, y L (x) is a low-precision RBF model, ρ is a scale transformation factor, f(x) is a radial basis vector, β is a radial basis coefficient vector, and ε(x) is an unknown error function; ① The high / low precision sample point data of the bogie component structural performance used to construct the Co-RBF multi-fidelity proxy model are: In the formula, S H is high-precision sample data, S L is the precision sample data, i,j is the m-dimensional input x sample vector, representing the stress condition of the bogie components; y is the output vector, representing the structural performance of the bogie components; n L and n H The number of high / low precision samples, and n L <n H ; The Co-RBF multi-fidelity proxy model is expressed as: y H (x)=[y L (x),f(x)][y L (x H ),F] + y H (9) In the formula, F is a matrix, which is used to compare with y in the calculation process. L (x H ) together calculate y H (x) matrix; ② Based on the bogie structure strength mechanism data set obtained in step S2121 and the grid node data obtained by using the QEM simplified algorithm in step S211; the K nearest neighbor algorithm is used to update the high / low precision stress and strain values ​​of each node of the lightweight model; the high / low precision training samples obtained can be expressed as: Where n is the number of mesh nodes of the lightweight model; ③ Input high / low precision samples of each node into the Co-RBF model, and use low precision samples to train the RBF model y L (x), integrate high-precision samples into y L (x) Training.

8. The method for constructing a bogie digital twin model for dynamic performance monitoring according to claim 7, characterized in that: Step S22 specifically includes the following steps S221-S222: Step S221: scene construction and state simulation; Step S222: Visualize monitoring performance.

9. A method for constructing a bogie digital twin model for dynamic performance monitoring according to claim 8, characterized in that: Step S221 includes the following steps: ①Lighten the main components of the bogie and complete the format conversion in 3D MAX software; ② Based on the constructed node relationship, use Unity3D's Animation component to create vertical / lateral vibration animation of the frame.

10. The method for constructing a bogie digital twin model for dynamic performance monitoring according to claim 9, characterized in that: Step S222 includes the following steps: ① Visualization of bogie dynamics performance indicators and sensor data; ② Real-time cloud map visualization of bogie component structural performance.

Images