A digital twin modeling method for bearing test bench

Abstract

The present invention discloses a digital twin modeling method for a bearing test bench. Firstly, a finite element model of a bearing-rotor system based on ADAMS is established by using hierarchical correction of a finite element model and bearing contact stiffness fitting based on a neural network, and the finite element model is integrated with a bearing-rotor system model based on deep learning obtained by using experimental data to build a digital twin model based on structural dynamics-deep learning. Then, deep learning is used to respectively establish proxy models of two-phase flow field, temperature field of inner and outer rings of the bearing, and Fluent simulation of fluid-solid heat transfer, and the proxy models are integrated with a bearing lubrication model based on deep learning obtained by using experimental data to build a digital twin model based on fluid-solid coupling-deep learning. Finally, the feature distribution adversarial adaptation deep migration theory is used to perform correlation analysis on the two digital twin models. The present invention realizes simulation analysis of complex working conditions of a bearing test bench under lubrication state, and can obtain operating parameters with high reliability in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of digital twin modeling, and mainly to a digital twin modeling method for a bearing test bench. Background Art

[0002] Based on the one-to-many characteristics of the digital twin model, multiple twin models can be established for a single bearing test bench, and each model corresponds to its specific physical characteristics, such as structural dynamics models, thermal management models, and fluid-solid coupling models considering lubrication. The more common digital twin modeling methods include the finite element method (FEM), computational fluid dynamics (CFD), concentrated parameter physical model, Kalman filter, and modeling methods based on deep learning of neural networks. The digital twin modeling method based on the effective fusion of data-driven and physical models can achieve the fusion and complementarity of the two at the principle level, reasonably and effectively combine the high-precision statistical characteristics of sensor data with the mechanism model of the system, and obtain a high-precision digital twin model of the bearing test bench.

[0003] The domain adaptation method in transfer learning aims to solve the distribution difference between the source domain and the target domain and establish knowledge transfer from the source domain to the target domain. According to the different sources of the source domain and the target domain in the domain adaptation method, fault migration diagnosis research can be roughly divided into three categories, including: migration between different working conditions of the same equipment, migration between different equipment, and migration from virtual equipment to physical entities. The design and implementation of simulation models are usually highly abstract, and the data obtained through simulation does not contain environmental noise, resulting in differences between the results of system simulation and experimental measurement. The simulation samples are not authentic enough. Using them directly to train the fault diagnosis model will result in the network only learning the features existing in the simulation samples, and the generalization ability of the model on real samples is difficult to guarantee. Transfer learning and domain adaptation methods are also applicable to solving the domain differences between simulation data and real data.

[0004] Deep learning aims to simulate the learning process of the brain, build a deep model, and combine massive training data to learn the implicit features in the data, thereby describing the rich intrinsic information of the data and ultimately improving the accuracy of classification or prediction. Deep learning has shown unprecedented application prospects in areas such as image recognition, speech recognition, and industrial robots. Summary of the invention

[0005] Purpose of the invention: In response to the problems existing in the above-mentioned background technology, the present invention provides a digital twin modeling method for a bearing test bench, which combines the advantages of multidisciplinary mechanism analysis and deep learning to build a digital twin model, and uses transfer learning to deeply integrate the multi-body dynamics model and the fluid-solid coupling model. It can accurately and comprehensively realize the simulation analysis of complex working conditions under the lubrication state of the bearing test bench, and monitor the deformation, vibration, force, speed, lubrication and other information of various bearing components in real time.

[0006] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:

[0007] A bearing test bench digital twin modeling method comprises the following steps:

[0008] Step S1, establishing a three-dimensional geometric model and a finite element model of a bearing test bench, wherein the finite element model of the bearing test bench includes a bearing system finite element model and a rotor system finite element model;

[0009] Step S2, building a digital twin model based on structural dynamics-deep learning;

[0010] Step S3, building a digital twin model based on fluid-solid coupling-deep learning;

[0011] Step S4: Use the feature distribution adversarial adaptation deep migration theory to perform correlation analysis on the digital twin model based on structural dynamics-deep learning and the digital twin model based on fluid-solid coupling-deep learning.

[0012] Furthermore, the specific steps of building a digital twin model based on structural dynamics-deep learning in step S2 include:

[0013] Step S2.1, based on the finite element model of the bearing system, the nonlinear contact stiffness of the bearing roller under different loads and speeds is calculated using the Nastran finite element analysis software, and the nonlinear contact stiffness of the bearing roller is fitted by a radial basis function neural network; specifically, a combined radial basis function is selected for modeling as follows:

[0014]

[0015] Among them, x R represents the input vector of the neural network; represents the output vector of the neural network; b m represents the mth central unit of the hidden layer; h is the number of neural units; v m is the weight coefficient between the hidden layer and the output layer; n m is the width of the basis function; and is the weight coefficient;

[0016] Step S2.2, based on the finite element model of the rotor system, using Patran to perform simulation modal vibration analysis;

[0017] Step S2.3, performing a modal knocking test on the bearing test bench, and performing a test modal vibration type analysis using N-modal software;

[0018] Step S2.4, comparing the simulation modal vibration type analysis obtained in step S2.2 with the experimental modal vibration type analysis obtained in step S2.3, and using the proxy model method to perform hierarchical correction on the bearing-rotor system finite element model;

[0019] Step S2.5, integrating the nonlinear contact stiffness model of the bearing roller obtained in step S2.1 and the finite element model of the bearing-rotor system obtained in step S2.4, and respectively building a rigid-flexible coupling model of the bearing-rotor system based on ADAMS and a fully flexible coupling model of the bearing-rotor system based on Workbench;

[0020] Step S2.6, using the fully flexible coupling model of the bearing-rotor system based on Workbench to confirm the rigid-flexible coupling model of the bearing-rotor system based on ADAMS; when the frequency errors of each order of the two types of models do not exceed the preset threshold, the rigid-flexible coupling model of the bearing-rotor system based on ADAMS is accurate; when there is a situation where the frequency errors of each order of the two types of models are greater than the preset threshold, repeat step S2.4;

[0021] Step S2.7, without considering lubrication of the bearing test bench, collect the vibration, speed, temperature, stress strain and load signals of the bearing test bench in real time, establish a bearing-rotor system proxy model based on deep learning, and use online data to test the accuracy of the proxy model to obtain a bearing-rotor system model based on deep learning;

[0022] Step S2.8, use the weighted Softmax loss function to fuse the ADAMS-based rigid-flexible coupling model of the bearing-rotor system obtained in step S2.6 and the deep learning-based bearing-rotor system model obtained in step S2.7 to obtain a digital twin model based on the combination of structural dynamics and deep learning.

[0023] Furthermore, the specific steps of building a digital twin model based on fluid-structure coupling-deep learning in step S3 include:

[0024] Step S3.1, using Fluent to respectively establish numerical calculation models of the oil-gas two-phase flow field inside the bearing under two modes of oil injection lubrication and under-ring lubrication, numerical calculation models of the temperature field of the inner and outer rings of the bearing, and numerical calculation models of fluid-solid heat transfer including a solid domain;

[0025] Step S3.2, based on the deep learning method, respectively establish proxy models of the numerical model of the oil-gas two-phase flow field inside the bearing, the numerical model of the temperature field of the inner and outer rings of the bearing, and the numerical model of fluid-solid heat transfer in step S3.1;

[0026] Step S3.3, under the condition that the bearing test bench considers lubrication, the vibration, speed, temperature, stress strain, load and oil supply amount signals of the bearing test bench are collected in real time, a bearing lubrication agent model is established based on deep learning, and the accuracy of the agent model is tested using online data to obtain a bearing lubrication model based on deep learning;

[0027] Step S3.4, using the weighted Softmax loss function, the two-phase flow field proxy model based on deep learning, the bearing inner and outer ring temperature field proxy model based on deep learning, the fluid-solid heat transfer proxy model based on deep learning, and the bearing lubrication model based on deep learning obtained in step S3.2 are integrated to obtain a digital twin model based on fluid-solid coupling-deep learning.

[0028] Furthermore, in step S2.8 and step S3.4, a weighted Softmax loss function is used to balance the contribution of various models obtained in step S2.6, step S2.7, step S3.2 and step S3.3 to the modeling error, and fully learn the knowledge of various models under the imbalance of monitoring data; the weighted Softmax loss function is specifically as follows:

[0029] The training data sets of various models obtained based on steps S2.6, S2.7, S3.2 and S3.3 are: {(x i ,y i )|i=1,2,…,M tr}; where x i is the i-th training sample; y i is the label of the sample and satisfies y i ∈{1,2,…,R}, R is the number of different labels, corresponding to different types of models; M tr is the number of training data sets; the number of samples of each model in the data set is as follows:

[0030]

[0031] Among them, n c is the number of training samples in the cth state; I(·) is the indicator function, when y i =c, the output of I(·) is 1, otherwise the output of I(·) is 0;

[0032] Designing weighted Softmax loss function satisfy:

[0033]

[0034] Among them, p i,c is the predicted probability that the i-th sample belongs to the c-th model; ω c is the diagnostic error weight of the cth model, which is specifically expressed as follows:

[0035]

[0036] Where max(·) is the maximum value of the variable.

[0037] Furthermore, the fluid-solid heat transfer model in step S3.1 includes:

[0038] The bearing friction moment M is calculated as follows:

[0039] M=M0+M1 (5)

[0040] Among them, M0 is the torque related to the bearing type, speed and lubricant; M1 is the friction torque related to the load borne by the bearing;

[0041] The forced convection heat transfer coefficient Nu_f of the fixed wall satisfies:

[0042]

[0043]

[0044] Among them, R e is the Reynolds number; P r is the Prandtl number; u is the air velocity near the end face; d is the equivalent diameter of the bearing; v1 is the kinematic viscosity of the air;

[0045] The convective heat transfer coefficient Nu_r of the rotating wall satisfies:

[0046]

[0047] Wherein, C is the correction coefficient and n is the correction index.

[0048] Furthermore, the deep migration theory based on feature distribution adversarial adaptation in step S4 includes a domain-shared deep convolutional neural network, a feature distribution matching module and a domain discriminator; specifically includes:

[0049] Step S4.1, using the structural dynamics-deep learning joint digital twin model obtained in step S2 and the fluid-structure interaction-deep learning joint digital twin model obtained in step S3 to construct a domain-shared deep convolutional neural network sample space;

[0050] The digital twin model based on structural dynamics and deep learning obtained in step S2 is used as the source domain, and the digital twin model based on fluid-structure interaction and deep learning obtained in step S3 is used as the target domain. The deep convolutional neural network shared by the domain is used to extract deep migration features from the original signals of the source domain and the target domain to construct the sample space. as follows:

[0051]

[0052] in, is the i-th source domain sample; is the jth target domain sample; n is the number of source domain samples; m is the number of target domain samples;

[0053] Step 4.2: Calculate the difference in the distribution of source domain and target domain migration features;

[0054] Estimating the distribution difference of migration features in the F2 layer of neural networks using the maximum mean difference implanted with Gaussian kernel as follows:

[0055]

[0056] in, It is the source domain sample space of the neural network F2 layer; is the target domain sample space of the neural network F2 layer; K(·,·) is the Gaussian kernel function; Φ(·) is the nonlinear mapping from the sample space to the deep feature space;

[0057] Step 4.3: Design a domain discriminator to determine whether the transferred features come from the source domain or the target domain;

[0058] The deep migration features extracted in step S4.1 are used as the input of the neighborhood discriminator. The output layer contains only a single neuron. The loss function of the neighborhood discriminator is designed as follows:

[0059]

[0060] Where D(·) is the discriminator; when the neighborhood discriminator loss function reaches the maximum value, it can be judged that the migration feature comes from the source domain or the target domain; the output of the neighborhood discriminator can be expressed as:

[0061]

[0062] Where z is the deep feature extracted from the source domain and target domain samples; P s (z) is the probability density of the source domain sample features; P t(z) is the probability density of the sample features in the target domain. It can be seen that when the output of the neighborhood discriminator is closer to 1, the probability that the input migration feature comes from the source domain is greater; conversely, the probability that it comes from the target domain is greater.

[0063] Beneficial effects:

[0064] Aiming at the problem of digital twin modeling of bearing test bench, this paper proposes a digital twin modeling method of bearing test bench based on multidisciplinary mechanism analysis and deep learning. Compared with the existing technology, the advantages are:

[0065] (1) Using neural networks to establish a nonlinear contact stiffness model of bearing rollers under variable load and speed conditions, and to improve the accuracy of bearing finite element simulation based on structural dynamics;

[0066] (2) Under the condition of lubrication of the bearing test bench, the collected bearing test data are used to establish a bearing-rotor system model and a bearing lubrication model based on deep learning, and the weighted Softmax loss function is used to combine various finite element models to improve the accuracy of the digital twin model; overcome the shortcomings of digital twin research based on a single finite element simulation model or deep learning model, which leads to limited model accuracy and inability to effectively explain the model mechanism;

[0067] (3) Using the feature distribution adversarial adaptation deep migration theory to study the correlation between the digital twin model based on structural dynamics-deep learning and the digital twin model based on fluid-structure interaction-deep learning, the digital twin model based on structural dynamics-deep learning is used as the source domain, and the digital twin model based on fluid-structure interaction-deep learning is used as the target domain for feature migration, making full use of the advantages of simulation analysis in different disciplines and breaking through the limitations of simulation software in different disciplines;

[0068] (4) Construct a digital twin model that can truly and comprehensively reflect the complex working conditions of the bearing test bench under the lubrication state, monitor the deformation, vibration, force, speed, lubrication and other information of each bearing component in real time, and obtain operating parameters with high reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 A flow chart of the digital twin modeling method for a bearing test bench provided by the present invention;

[0070] Figure 2 is a finite element model of a bearing system in an embodiment of the present invention;

[0071] Figure 3 is a finite element model of a rotor system in an embodiment of the present invention;

[0072] Figure 4 The Patran simulation mode shape in the embodiment of the present invention;

[0073] Figure 5 It is the N-modal test modal vibration type in the embodiment of the present invention;

[0074] Figure 6 It is a schematic diagram of layered correction of a finite element model in an embodiment of the present invention;

[0075] Figure 7 Schematic diagram of a finite element model correction method based on a proxy model method in an embodiment of the present invention;

[0076] Figure 8 The calculation result of the proxy model of the bearing test bench bracket in the embodiment of the present invention;

[0077] Fig. 9 The model confirmation result of the rigid-flexible coupling of the bearing-rotor system based on ADAMS in the embodiment of the present invention;

[0078] Fig.10 The simulation model of the oil-gas two-phase flow field inside the bearing in the embodiment of the present invention;

[0079] Fig.11 This is a theoretical principle diagram of deep migration based on feature distribution adversarial adaptation in an embodiment of the present invention. DETAILED DESCRIPTION

[0080] The present invention is further described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0081] This paper proposes a digital twin modeling method for a bearing test bench based on the integration of multidisciplinary mechanism analysis and deep learning. Figure 1 The method mainly includes four parts: establishing a three-dimensional geometric model and a finite element model of the bearing test bench, building a digital twin model based on structural dynamics-deep learning, building a digital twin model based on fluid-solid coupling-deep learning, and model correlation analysis. The specific steps are as follows:

[0082] Step S1: Use SolidWorks to build a three-dimensional model of the bearing test bench and divide the grid, and build the bearing system (such as Figure 2 as shown) and the rotor system (as Figure 3 The finite element model of the

[0083] Step S2: Building a digital twin model based on structural dynamics and deep learning. Specifically including:

[0084] Step S2.1, based on the finite element model of the bearing system, use Nastran to calculate the nonlinear contact stiffness of the bearing roller under different loads and speeds. The specific calculation results of the bearing stiffness under different loads are shown in Table 1 below:

[0085] Table 1 Calculation results of contact stiffness between roller and inner and outer rings under different loads

[0086]

[0087]

[0088] The nonlinear contact stiffness of the bearing roller is modeled by a radial basis function neural network. This embodiment uses a combined radial basis function for modeling as follows:

[0089]

[0090] Among them, x R represents the input vector of the neural network; represents the output vector of the neural network; b m represents the mth central unit of the hidden layer; h is the number of neural units; v m is the weight coefficient between the hidden layer and the output layer; n m is the width of the basis function; and is the weight coefficient.

[0091] Step S2.2: Based on the finite element model of the rotor system, Patran is used to perform simulation modal vibration analysis. The specific analysis results are as follows: Figure 4 shown.

[0092] Step S2.3: Perform a modal knocking test on the bearing test bench and use N-modal software to perform test modal vibration analysis. The specific analysis results are as follows: Figure 5 shown.

[0093] Step S2.4, comparing the simulation modal vibration type analysis obtained in step S2.2 with the experimental modal vibration type analysis obtained in step S2.3, and using the proxy model method to perform hierarchical correction on the bearing-rotor finite element model.

[0094] In a complex system, there are many factors that affect a certain characteristic of the system. The overall structure needs to be decomposed into multiple substructures and examined separately to distinguish the influence of each factor. The idea of ​​complex system hierarchical correction is: the entire system is divided into four levels: full system, subsystem, component and unit. The lower the level, the fewer factors affect it and the lower the degree of coupling. Figure 6 shown.

[0095] The surrogate model method is a basic method that integrates multiple disciplines such as experimental design and analysis, mathematical statistics and linear algebra. Its purpose is to establish a relationship model that can reflect the relationship between variables and responses and analyze the conversion relationship between variables and responses. On the basis of reasonably arranging the experiment to obtain sampling data, select appropriate regression or interpolation technology for data processing, and select an appropriate and clearly expressible function to approximate the implicit or unclear function. The principle of the surrogate model method is as follows Figure 7 The calculation results of the proxy model of the bearing test bench bracket in this embodiment are shown in Figure 8 The modal vibration correction results are shown in Table 2 below:

[0096] Table 2 Modal shape correction results

[0097]

[0098]

[0099] In step S2.5, the nonlinear contact stiffness model of the bearing roller obtained in step S2.2 and the finite element model of the bearing-rotor system obtained in step S2.3 are integrated to build a rigid-flexible coupling model of the bearing-rotor system based on ADAMS and a fully flexible coupling model of the bearing-rotor system based on Workbench respectively.

[0100] Step S2.6, use the fully flexible coupling model of the bearing-rotor system based on Workbench to confirm the rigid-flexible coupling model of the bearing-rotor system based on ADAMS; when the frequency errors of each order of the two types of models do not exceed the preset threshold, the rigid-flexible coupling model of the bearing-rotor system based on ADAMS is accurate; when the frequency errors of each order of the two types of models are greater than the preset threshold, repeat step S2.4 and perform hierarchical correction. The embodiment of the present invention sets the modal frequency matching error threshold to 1%. Table 3 gives the numerical results of the bearing test bench bracket model confirmation based on the modal frequency comparison, and the finite element model comparison is as follows. Fig. 9 As shown:

[0101] Table 3. Confirmation of numerical results based on modal frequency comparison of the bearing test bench bracket model

[0102]

[0103] It can be seen from Table 3 that the matching errors of the bearing test bench bracket model based on modal frequency comparison are all less than the set threshold of 1%, and there is no need to return to step S2.4 for re-correction.

[0104] Step S2.7: Without considering lubrication of the bearing test bench, the vibration, speed, temperature, stress-strain and load signals of the bearing test bench are collected in real time, a bearing-rotor system proxy model is established based on deep learning, and the accuracy of the proxy model is tested using online data to obtain a bearing-rotor system model based on deep learning.

[0105] Step S2.8, use the weighted Softmax loss function to fuse the ADAMS-based rigid-flexible coupling model of the bearing-rotor system obtained in step S2.6 and the deep learning-based bearing-rotor system model obtained in step S2.7 to obtain a digital twin model based on the combination of structural dynamics and deep learning.

[0106] Step S3: Build a digital twin model based on fluid-solid coupling and deep learning.

[0107] Step S3.1, using Fluent to respectively establish the numerical calculation model of the oil-gas two-phase flow field inside the bearing under the two modes of oil injection lubrication and under-ring lubrication, the numerical calculation model of the temperature field of the inner and outer rings of the bearing, and the fluid-solid heat transfer numerical model including the solid domain. The simulation model of the oil-gas two-phase flow field inside the bearing is as follows: Fig.10 As shown in the figure, it includes solid domains such as the main shaft and bearing seat, as well as fluid domains such as lubricating oil and air.

[0108] Bearing friction torque M calculation formula:

[0109] M=M0+M1 (2)

[0110] Among them, M0 is the torque related to the bearing type, speed and lubricant; M1 is the friction torque related to the load borne by the bearing.

[0111] The forced convection heat transfer coefficient Nu_f of the fixed wall satisfies:

[0112]

[0113]

[0114] Among them, R e is the Reynolds number; P r is the Prandtl number; u is the air velocity near the end face; d is the equivalent diameter of the bearing; v1 is the kinematic viscosity of the air.

[0115] The convective heat transfer coefficient Nu_r of the rotating wall satisfies:

[0116]

[0117] Wherein, C is the correction coefficient and n is the correction index.

[0118] Step S3.2, based on the deep learning method, respectively establish proxy models of the numerical model of the oil-gas two-phase flow field inside the bearing, the numerical model of the temperature field of the inner and outer rings of the bearing, and the numerical model of fluid-solid heat transfer described in step S3.1;

[0119] Step S3.3, under the condition that the bearing test bench considers lubrication, the vibration, speed, temperature, stress strain, load and oil supply amount signals of the bearing test bench are collected in real time, a bearing lubrication agent model is established, and the accuracy of the agent model is tested using online data to obtain a bearing lubrication model based on deep learning;

[0120] Step S3.4, using the weighted Softmax loss function, the two-phase flow field proxy model based on deep learning, the bearing inner and outer ring temperature field proxy model based on deep learning, the fluid-solid heat transfer proxy model based on deep learning, and the bearing lubrication model based on deep learning obtained in step S3.2 are integrated to obtain a digital twin model based on fluid-solid coupling-deep learning.

[0121] The weighted Softmax loss function in step S2.8 and step S3.4 can balance the contribution of various models obtained in steps S2.6, S2.7, S3.2 and S3.3 to the modeling error, and fully learn the knowledge of various models under unbalanced monitoring data; the weighted Softmax loss function is specifically:

[0122] The training data sets of various models obtained based on steps S2.6, S2.7, S3.2 and S3.3 are: {(x i ,y i )|i=1,2,...,M tr}; where x i is the i-th training sample; y i is the label of the sample and satisfies y i ∈{1,2,...,R}, R is the number of different labels, corresponding to different types of models; M tr is the number of training data sets; the number of samples of each model in the data set is as follows:

[0123]

[0124] Among them, n c is the number of training samples in the cth state; I(·) is the indicator function, when y i =c, I(·) output is 1, otherwise, I(·) output is 0; when the number of samples in the training data set is balanced, the modeling error weights of all types of models are 1; when the number of samples in the training data set is unbalanced, the modeling error weight of the model with a small number of samples is large, while the modeling error weight of the model with a large number of samples is small;

[0125] Designing weighted Softmax loss function satisfy:

[0126]

[0127] Among them, p i,c is the predicted probability that the i-th sample belongs to the c-th model; ω c is the diagnostic error weight of the cth model, which is specifically expressed as follows:

[0128]

[0129] Where max(·) is the maximum value of the variable.

[0130] Step S4: Use the feature distribution adversarial adaptation deep migration theory to perform correlation analysis on the digital twin model based on structural dynamics-deep learning and the digital twin model based on fluid-structure interaction-deep learning. The deep migration theory based on feature distribution adversarial adaptation includes a domain-shared deep convolutional neural network, a feature distribution matching module, and a domain discriminator. The specific principle is as follows: Fig.11 As shown; specifically including:

[0131] Step S4.1, using the structural dynamics-deep learning joint digital twin model obtained in step S2 and the fluid-structure interaction-deep learning joint digital twin model obtained in step S3 to construct a domain-shared deep convolutional neural network sample space;

[0132] The digital twin model based on structural dynamics and deep learning obtained in step S2 is used as the source domain, and the digital twin model based on fluid-structure interaction and deep learning obtained in step S3 is used as the target domain. The deep convolutional neural network shared by the domain is used to extract deep migration features from the original signals of the source domain and the target domain to construct the sample space. as follows:

[0133]

[0134] in, is the i-th source domain sample; is the jth target domain sample; n is the number of source domain samples; m is the number of target domain samples;

[0135] Step 4.2: Calculate the difference in the distribution of source domain and target domain migration features;

[0136] Estimating the distribution difference of migration features in the F2 layer of neural networks using the maximum mean difference implanted with Gaussian kernel as follows:

[0137]

[0138] in, It is the source domain sample space of the neural network F2 layer; is the target domain sample space of the neural network F2 layer; K(·,·) is the Gaussian kernel function; Φ(·) is the nonlinear mapping from the sample space to the deep feature space;

[0139] Step 4.3: Design a domain discriminator to determine whether the transferred features come from the source domain or the target domain;

[0140] The deep migration features extracted in step S4.1 are used as the input of the neighborhood discriminator. The output layer contains only a single neuron. The loss function of the neighborhood discriminator is designed as follows:

[0141]

[0142] Where D(·) is the discriminator. When the neighborhood discriminator loss function reaches the maximum value, it can accurately judge whether the migration feature comes from the source domain or the target domain; the output of the neighborhood discriminator can be expressed as:

[0143]

[0144] Where z is the deep feature extracted from the source domain and target domain samples; P s (z) is the probability density of the source domain sample features; P t (z) is the probability density of the sample features in the target domain. It can be seen that when the output of the neighborhood discriminator is closer to 1, the probability that the input migration feature comes from the source domain is greater; conversely, the probability that it comes from the target domain is greater.

[0145] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A bearing test bench digital twin modeling method, characterized in that: The following steps are involved: Step S1, establishing a three-dimensional geometric model and a finite element model of a bearing test bench, wherein the finite element model of the bearing test bench includes a bearing system finite element model and a rotor system finite element model; Step S2, building a digital twin model based on structural dynamics-deep learning; Step S3, building a digital twin model based on fluid-solid coupling-deep learning; Step S4, using feature distribution adversarial adaptation deep migration theory to perform correlation analysis on the digital twin model based on structural dynamics-deep learning and the digital twin model based on fluid-solid coupling-deep learning; The specific steps of building a digital twin model based on structural dynamics-deep learning in step S2 include: Step S2.1, based on the finite element model of the bearing system, the nonlinear contact stiffness of the bearing roller under different loads and speeds is calculated using the Nastran finite element analysis software, and the nonlinear contact stiffness of the bearing roller is fitted by a radial basis function neural network; specifically, a combined radial basis function is selected for modeling as follows: Among them, x R represents the input vector of the neural network; represents the output vector of the neural network; b m represents the mth central unit of the hidden layer; h is the number of neural units; v m is the weight coefficient between the hidden layer and the output layer; n m is the width of the basis function; and is the weight coefficient; Step S2.2, based on the finite element model of the rotor system, using Patran to perform simulation modal vibration analysis; Step S2.3, performing a modal knocking test on the bearing test bench, and performing a test modal vibration type analysis using N-modal software; Step S2.4, comparing the simulation modal vibration type analysis obtained in step S2.2 with the experimental modal vibration type analysis obtained in step S2.3, and using the proxy model method to perform hierarchical correction on the bearing-rotor system finite element model; Step S2.5, integrating the nonlinear contact stiffness model of the bearing roller obtained in step S2.1 and the finite element model of the bearing-rotor system obtained in step S2.4, and respectively building a rigid-flexible coupling model of the bearing-rotor system based on ADAMS and a fully flexible coupling model of the bearing-rotor system based on Workbench; Step S2.6, using the fully flexible coupling model of the bearing-rotor system based on Workbench to confirm the rigid-flexible coupling model of the bearing-rotor system based on ADAMS; when the frequency errors of each order of the two types of models do not exceed the preset threshold, the rigid-flexible coupling model of the bearing-rotor system based on ADAMS is accurate; when there is a situation where the frequency errors of each order of the two types of models are greater than the preset threshold, repeat step S2.4; Step S2.7, without considering lubrication of the bearing test bench, collect the vibration, speed, temperature, stress strain and load signals of the bearing test bench in real time, establish a bearing-rotor system proxy model based on deep learning, and use online data to test the accuracy of the proxy model to obtain a bearing-rotor system model based on deep learning; Step S2.8, use the weighted Softmax loss function to fuse the ADAMS-based rigid-flexible coupling model of the bearing-rotor system obtained in step S2.6 and the deep learning-based bearing-rotor system model obtained in step S2.7 to obtain a digital twin model based on the combination of structural dynamics and deep learning.

2. A bearing test bench digital twin modeling method according to claim 1, characterized in that: The specific steps of building a digital twin model based on fluid-structure coupling-deep learning in step S3 include: Step S3.1, using Fluent to respectively establish numerical calculation models of the oil-gas two-phase flow field inside the bearing under two modes of oil injection lubrication and under-ring lubrication, numerical calculation models of the temperature field of the inner and outer rings of the bearing, and numerical calculation models of fluid-solid heat transfer including a solid domain; Step S3.2, based on the deep learning method, respectively establish proxy models of the numerical model of the oil-gas two-phase flow field inside the bearing, the numerical model of the temperature field of the inner and outer rings of the bearing, and the numerical model of fluid-solid heat transfer in step S3.1; Step S3.3, under the condition that the bearing test bench considers lubrication, the vibration, speed, temperature, stress strain, load and oil supply amount signals of the bearing test bench are collected in real time, a bearing lubrication agent model is established based on deep learning, and the accuracy of the agent model is tested using online data to obtain a bearing lubrication model based on deep learning; Step S3.4, using the weighted Softmax loss function, the two-phase flow field proxy model based on deep learning, the bearing inner and outer ring temperature field proxy model based on deep learning, the fluid-solid heat transfer proxy model based on deep learning, and the bearing lubrication model based on deep learning obtained in step S3.2 are integrated to obtain a digital twin model based on fluid-solid coupling-deep learning.

3. A bearing test bench digital twin modeling method according to any one of claims 1-2, characterized in that: In the steps S2.8 and S3.4, a weighted Softmax loss function is used to balance the contribution of various models obtained in steps S2.6, S2.7, S3.2 and S3.3 to the modeling error, and fully learn the knowledge of various models under unbalanced monitoring data; the weighted Softmax loss function is specifically as follows: The training data sets of various models obtained based on steps S2.6, S2.7, S3.2 and S3.3 are: {(x i ,y i )|i=1,2,K,M tr }; where x i is the i-th training sample; y i is the label of the sample and satisfies y i ∈{1,2,K,R}, R is the number of different labels, corresponding to different types of models; M tr is the number of training data sets; the number of samples of each model in the data set is as follows: Among them, n c is the number of training samples in the cth state; I(·) is the indicator function, when y i =c, the output of I(·) is 1, otherwise the output of I(·) is 0; Designing weighted Softmax loss function satisfy: Among them, p i,c is the predicted probability that the i-th sample belongs to the c-th model; ω c is the diagnostic error weight of the cth model, which is specifically expressed as follows: Where max(·) is the maximum value of the variable.

4. A bearing test bench digital twin modeling method according to claim 2, characterized in that: The fluid-solid heat transfer model in step S3.1 includes: The bearing friction moment M is calculated as follows: M=M0+M1 (5) Among them, M0 is the torque related to the bearing type, speed and lubricant; M1 is the friction torque related to the load borne by the bearing; The forced convection heat transfer coefficient Nu_f of the fixed wall satisfies: Among them, R e is the Reynolds number; P r is the Prandtl number; u is the air velocity near the end face; d is the equivalent diameter of the bearing; v1 is the kinematic viscosity of the air; The convective heat transfer coefficient Nu_r of the rotating wall satisfies: Wherein, C is the correction coefficient and n is the correction index.

5. The method for digital twin modeling of a bearing test bench according to claim 1, characterized in that: The deep migration theory based on feature distribution adversarial adaptation in step S4 includes a domain-shared deep convolutional neural network, a feature distribution matching module and a domain discriminator; specifically includes: Step S4.1, using the structural dynamics-deep learning joint digital twin model obtained in step S2 and the fluid-structure interaction-deep learning joint digital twin model obtained in step S3 to construct a domain-shared deep convolutional neural network sample space; The digital twin model based on structural dynamics and deep learning obtained in step S2 is used as the source domain, and the digital twin model based on fluid-structure interaction and deep learning obtained in step S3 is used as the target domain. The deep convolutional neural network shared by the domain is used to extract deep migration features from the original signals of the source domain and the target domain to construct the sample space. as follows: in, is the i-th source domain sample; is the jth target domain sample; n is the number of source domain samples; m is the number of target domain samples; Step 4.2: Calculate the difference in the distribution of source domain and target domain migration features; Estimating the distribution difference of migration features in the F2 layer of neural networks using the maximum mean difference implanted with Gaussian kernel as follows: in, It is the source domain sample space of the neural network F2 layer; is the target domain sample space of the neural network F2 layer; K(·,·) is the Gaussian kernel function; Φ(·) is the nonlinear mapping from the sample space to the deep feature space; Step 4.3: Design a domain discriminator to determine whether the transferred features come from the source domain or the target domain; The deep migration features extracted in step S4.1 are used as the input of the neighborhood discriminator. The output layer contains only a single neuron. The loss function of the neighborhood discriminator is designed as follows: Where D(·) is the discriminator; when the neighborhood discriminator loss function reaches the maximum value, it can be judged that the migration feature comes from the source domain or the target domain; the output of the neighborhood discriminator can be expressed as: Where z is the deep feature extracted from the source domain and target domain samples; P s (z) is the probability density of the source domain sample features; P t (z) is the probability density of the sample features in the target domain. When the output of the neighborhood discriminator is closer to 1, the probability that the input migration feature comes from the source domain is greater; conversely, the probability that it comes from the target domain is greater.

Images