Boring mill digital twinning diagnosis method and device based on data and finite element model

Abstract

The invention provides a boring machine digital twinning diagnosis method and device based on data and a finite element model. Comprising the steps of establishing a finite element model of boring mill parts; simulating the boring machine parts in different fault modes based on the finite element model to obtain fault data; establishing a data set according to the fault data; a neural network model is constructed, the neural network model is a Res-EAS-Net model, and the Res-EAS-Net model is obtained by introducing a channel attention mechanism on the basis of a ResNet model; training a neural network model according to the data set to obtain a first fault diagnosis model; training the first fault diagnosis model by adopting transfer learning to obtain a second fault diagnosis model; establishing a digital twinborn model of the boring machine for collecting and synchronizing state data of parts of the boring machine in real time; and determining the current state of the boring machine part according to the state data of the boring machine part acquired by the digital twinborn model and the second fault diagnosis model, and displaying the current state of the boring machine part in the data twinborn model.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent manufacturing, and in particular relates to a digital twin diagnosis method and device for a boring machine based on data and a finite element model. Background Art

[0002] Gears, bearings and other rotating mechanical components on heavy equipment often work under high impact, continuous high temperature, heavy load or other complex working conditions, so they are prone to failure, thereby reducing the transmission accuracy, production efficiency and reliability of the equipment.

[0003] Currently, rolling bearing fault diagnosis is mainly divided into model-based methods, data-driven methods, and hybrid methods. Model-based methods aim to model the failure mechanism or degradation process of rolling bearings by establishing high-fidelity physical or mathematical models, thereby inferring whether there is a fault or the time of failure threshold. This type of method aims to derive the dynamic differential equations of the target physical system and solve them to achieve analytical modeling. Although model-based methods have strong interpretability, the parameter setting of this method requires expert experience and is highly targeted. That is, when the application scenario or object is changed, the robustness of the model is poor, which will lead to a decrease in diagnostic accuracy.

[0004] Unlike traditional model-based approaches, data-driven methods do not rely on a precise physical model of the system. Instead, they leverage machine learning, or even deep learning, and data analysis techniques to adaptively extract complex features to construct a mapping relationship between sensor signals and bearing health status, thereby building a model. Data-driven methods achieve system fault diagnosis through a complete process of data acquisition, feature extraction, model training, and evaluation. However, models built using data-driven methods require a large amount of valid data for training and evaluation, which is difficult to obtain in actual production processes.

[0005] In recent years, with the development of technologies such as cyber-physical systems, the Industrial Internet of Things (IIoT), and virtual reality (VR), the concept of digital twins (DT) has garnered widespread attention in various areas of intelligent manufacturing, such as job shop scheduling and decision optimization. DT is a technology that integrates the physical and digital worlds. Its core is to establish real-time interaction between physical and virtual models through sensors, the Internet of Things (IoT), and data analytics, thereby enabling the monitoring and optimization of physical systems. Based on accurate real-world physical models, DT can generate high-fidelity digital models, enabling the generation of realistic simulation data. This capability means that DT can provide a wealth of data for data-driven diagnostic models, addressing the scarcity of effective data in actual production. Therefore, DT has become a bridge for applying data-driven models to real-world production practices.

[0006] However, existing research faces the following challenges. First, few researchers have combined bearing mechanisms with data-driven approaches, resulting in a lack of interpretability in fault diagnosis models. Second, collecting real-world fault data during actual production is difficult, which complicates the development of diagnostic models. Third, models built using data-driven approaches exhibit poor diagnostic performance when faced with drastic changes in operating conditions, making their direct application to other scenarios difficult. Summary of the Invention

[0007] The purpose of the present invention is to provide a digital twin diagnosis method and device for a boring machine based on data and a finite element model to realize fault diagnosis of the boring machine.

[0008] First, a digital twin diagnostic method for boring machines based on data and finite element models is provided, including: Establish finite element models of boring machine components; Based on the finite element model, the boring machine parts are simulated under different failure modes to obtain the failure data of the boring machine parts; Establish a data set based on fault data; Constructing a neural network model, wherein the neural network model is a Res-EAS-Net model, and the Res-EAS-Net model is obtained by introducing a channel attention mechanism on the basis of the ResNet model; The neural network model is trained based on the data set to obtain the first fault diagnosis model of the boring machine parts; Transfer learning is used to train the first fault diagnosis model to obtain a second fault diagnosis model for boring machine parts; Establish a digital twin model of the boring machine to collect and synchronize the status data of the boring machine components in real time; Based on the status data of the boring machine parts collected by the digital twin model and the second fault diagnosis model of the boring machine parts, the current status of the boring machine parts is determined, and the current status of the boring machine parts is displayed in the data twin model.

[0009] Optionally, the failure modes of the boring machine component include rolling element defects, inner race failures and outer race failures.

[0010] Optionally, the fault data includes vibration data.

[0011] Optionally, the step of establishing a data set based on the fault data includes: The fault data is converted into time-frequency images through continuous wavelet transform, and a data set is constructed based on the time-frequency images.

[0012] Optionally, the channel attention mechanism is as follows: Perform a global average pooling operation on each channel of the input feature map to obtain the descriptor of each channel; the formula for global average pooling is as follows: , where Indicates the Global descriptor for each channel; and Represent the height and width of the feature map respectively; Indicates the first Channels at position The value at The obtained descriptor of each channel is used as the input of the one-dimensional convolution layer, and the channel attention weight of each channel is generated through the one-dimensional convolution operation; Among them, the convolution kernel of the one-dimensional convolution layer is calculated as follows: , where Indicates the number of channels With convolution kernel The functional relationship between Is the proportional coefficient used to adjust right The degree of impact, is the bias term, Is the proportional coefficient used to adjust right the extent of the impact; According to the channel attention weight of each channel, the features of each channel are Perform weighting to obtain the weighted feature map .

[0013] Optionally, based on the introduction of the channel attention mechanism, an adaptive calibration jump connection mechanism is introduced for the Res-EAS-Net model; wherein the adaptive calibration jump connection mechanism is as follows: Feature extraction step: output feature map of the previous residual module in the Res-EAS-Net model , perform identity mapping or linear transformation to obtain the skip connection feature ; Adaptive weight generation step: Introducing weight generation function , generates weight coefficients for skip connection features; where the weight coefficients The generation formula is: , where 、 represents the learnable parameters, express activation function, represents global average pooling; Feature calibration step: using weight coefficients , for skip connection features Perform weighting; Feature fusion step: combine the weighted jump connection features with the main path output features Add together to get the final output features .

[0014] Optionally, the step of using transfer learning to train the first fault diagnosis model to obtain a second fault diagnosis model for boring machine parts includes: Select time-frequency data under two different working conditions as the source domain and target domain respectively as the transfer learning dataset; Freeze the parameter updates of the weight layers except the fully connected layers in the Res-EAS-Net model, and only update the parameters of the fully connected layers through the backpropagation algorithm; Map the data of both the source and target domains into the reproducing kernel Hilbert space, and align the features of the source and target domains by maximum mean difference; The new loss function is defined as , and train the first fault diagnosis model according to the new loss function to obtain the second fault diagnosis model; in the above formula, represents the classification loss, represents the migration loss calculated by maximizing the mean difference, Represents the weight of the transfer loss calculated by maximizing the mean difference.

[0015] Secondly, a digital twin diagnostic device for a boring machine based on data and a finite element model is provided, comprising: Finite element model building module, used to build finite element models of boring machine parts; A simulation module is used to simulate the boring machine parts under different failure modes based on the finite element model to obtain the failure data of the boring machine parts; A data set construction module is used to build a data set based on fault data; A neural network model construction module is used to construct a neural network model, wherein the neural network model is a multi-scale feature extraction network based on a sparse convolutional autoencoder; A first training module is used to train a neural network model according to a data set to obtain a first fault diagnosis model of the boring machine parts; A second training module is used to train the first fault diagnosis model using transfer learning to obtain a second fault diagnosis model for boring machine parts; A digital twin model building module is used to build a digital twin model of the boring machine, which is used to collect and synchronize the status data of the boring machine components in real time; The diagnostic module is used to determine the current status of the boring machine components based on the status data of the boring machine components collected by the digital twin model and the second fault diagnosis model of the boring machine components, and display the current status of the boring machine components in the data twin model.

[0016] In a third aspect, an electronic device is provided, comprising the digital twin diagnostic device for a boring machine based on data and a finite element model as described above.

[0017] In a fourth aspect, a computer-readable storage medium is provided, in which at least one program code is stored. The program code is executed by a processor to implement the digital twin diagnosis method of a boring machine based on data and a finite element model as described in any one of the above items.

[0018] The technical solution provided by the present invention has the following beneficial effects: The present invention provides a digital twin diagnosis method for boring machines based on data and finite element models. By constructing a finite element model and solving the finite element model, simulated fault data is obtained to form a data set. A Res-EAS-Net model is constructed to extract the features of the bearing time-frequency image to ensure the performance of the first fault diagnosis model. The improved feature migration method is used to construct the second fault diagnosis model, which can ensure that the second fault diagnosis model can be applied to fault diagnosis under different working conditions. Finally, a high-fidelity digital twin model of the boring machine and its spindle is established to achieve real-time synchronization between the twin boring machine and the physical boring machine; the data of the physical boring machine is synchronously collected through the data twin, and diagnosis is performed through the diagnostic model, so that the status of each component of the boring machine can be updated in real time in the data twin model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 A flowchart of a digital twin diagnostic method for a boring machine based on data and finite element models provided by the present invention; Figure 2 A schematic diagram of a finite element model provided by the present invention; Figure 3 A schematic diagram of another finite element model provided by the present invention; Figure 4 A schematic diagram of the finite element model simulation keywords provided by the present invention; Figure 5An equivalent stress cloud diagram provided by the present invention; Figure 6 The simulated signal and real signal and their envelope diagrams under the healthy state provided by the present invention; Figure 7 The simulated signal and real signal and their envelope diagram under the outer ring fault state provided by the present invention; Figure 8 The simulated signal and real signal and their envelope diagram under the inner ring fault condition provided by the present invention; Figure 9 The simulated signal and the real signal and their envelope diagram under the rolling element fault condition provided by the present invention; Figure 10 A schematic diagram of the structure of a Resnet18 model provided by the present invention; Figure 11 A schematic diagram of the structure of another Resnet18 provided by the present invention; Figure 12 A training process of a second fault diagnosis model provided by the present invention; Figure 13 A schematic diagram of an experimental result provided by the present invention; Figure 14 A schematic diagram of a digital twin model provided by the present invention; Figure 15 A schematic diagram of another digital twin model provided by the present invention; Figure 16 A schematic diagram of another digital twin model provided by the present invention; Figure 17 A schematic diagram of another digital twin model provided by the present invention; Figure 18 A schematic diagram of another digital twin model provided by the present invention; Figure 19 A structural block diagram of a digital twin diagnostic device for a boring machine based on data and a finite element model provided by the present invention; Figure 20 This is a structural block diagram of an electronic device provided by the present invention.

[0021] The reference numerals are as follows: 11: Finite element model construction module; 12: Simulation module; 13: Dataset construction module; 14: Neural network model construction module; 15: First training module; 16: Second training module; 17: Digital twin model construction module; 18: Diagnosis module; 21: Processor; 22: Memory. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0023] Figure 1 This is a flow chart of the digital twin diagnosis method for boring machines based on data and finite element models provided by the present invention. Figure 1 ,include: S101. Establish a finite element model of boring machine parts.

[0024] In one example, step S101 includes: The first step is to establish a two-dimensional finite element model based on the actual structure.

[0025] This paper uses a rolling bearing experiment conducted at Case Western Reserve University (CWRU) as a foundation to illustrate the specific process of establishing a finite element explicit dynamic simulation model for rolling bearings. The test bench primarily consists of a motor, a rolling bearing on the motor drive end, a dynamometer, and a torque sensor. The rolling bearing on the motor drive end is an SKF-6205. An accelerometer is mounted on the motor housing at the drive end to collect vibration signals from the rolling bearing. The CWRU bearing experiment tests rolling bearings in four health states: normal, inner race fault, outer race fault, and rolling element fault. Faulty rolling bearings are created using electro-discharge machining (EDM). The experiment encompasses four operating conditions. The simulation speed is 1750 RPM, and the rolling bearing faults are cracks with widths of 0.007, 0.014, and 0.021 inches (approximately 0.18 mm, 0.36 mm, and 0.54 mm). The sampling frequency is 12 kHz at the drive end.

[0026] The second step is grid division.

[0027] Meshing is a key step in finite element simulation analysis and directly affects the reliability of simulation results. High-quality mesh units are the basis for obtaining accurate results. Generally, the smaller the mesh unit, the higher the accuracy of the simulation results, but too small a mesh size will significantly increase the number of meshes, thereby increasing the computational burden, especially when computing resources are limited, the solution time will be greatly extended. In addition, the shape of the mesh unit is equally important. Ideal mesh units should avoid sharp edges and corners, and the mesh should be arranged regularly to avoid chaos and irregular distribution. Therefore, when meshing, it is necessary to reasonably control the number of meshes while ensuring simulation accuracy to avoid excessively long computational times. In order to ensure good transferability between units, all mesh units use quadrilateral meshes so that the mesh arrangement is uniform and regular.

[0028] In rolling bearing systems, moving parts and their contact components, such as inner and outer rings, cages, and rolling elements, are critical components, necessitating a finer mesh. The rolling elements, in particular, are the primary moving component in the bearing and frequently come into contact with other components. To ensure the simulation model accurately reflects actual motion, the mesh size for the rolling elements, inner and outer rings, and cages was ultimately controlled to within 0.5 mm, with refinement of specific areas possible as needed.

[0029] Before setting the relevant parameters of this bearing simulation model, some precautions need to be taken in comparison with the actual bearing experimental platform. The first is to take into account that in reality, the rolling elements and cages, and the rolling elements and inner and outer rings are not completely fitted, but there is a certain gap. In order to avoid mutual interference between the moving parts of the rolling bearing during movement, when establishing the mesh model of the rolling bearing, a gap of 0.04mm is reserved between the rolling elements and the pockets of the cage, and a gap of 0.02mm is reserved between the rolling elements and the inner and outer rings. Since the various components of the rolling bearing system are made of linear elastic materials, if the load is directly applied to these linear elastic units, it is easy to cause large stress concentration and distortion, which cannot accurately reflect the actual situation. Therefore, in order to better simulate the actual working conditions, a layer of rigid shell unit is added to the inner surface of the inner ring, and the rotational speed and radial force load are applied to the rigid unit. The rigid unit is rigid-flexibly coupled with the inner ring to drive the overall movement and load of the inner ring to more accurately reflect the working state of the bearing. The final overall mesh division is as follows Figure 2 shown.

[0030] S102: Simulate the boring machine components under different failure modes based on the finite element model to obtain failure data of the boring machine components.

[0031] The failure modes of boring machine components include rolling element defects, inner race faults, and outer race faults. The failure data includes the characteristic frequencies of rolling element faults, inner race faults, and outer race faults.

[0032] In this embodiment, the fault data includes a vibration signal, and the vibration signal includes the acceleration of the main shaft in the directions of the three coordinate axes X, Y, and Z.

[0033] In one example, before performing finite element simulation, simulation conditions and simulation parameters need to be set.

[0034] To ensure that the bearing-housing system simulation model accurately reflects actual operating conditions, all model parameters must be fully consistent with the actual situation. Based on the material properties of a real 6205 bearing, the inner and outer rings and rolling elements of the rolling bearing in the simulation model are made of GCr15 rolling bearing steel, and the cage is made of carbon steel. Each component is assigned its corresponding material properties.

[0035] In addition, to simplify the calculation, the inner ring rigid body element only retains the vertical displacement degree of freedom and the rotational degree of freedom along the axis, while the other displacement and rotational degrees of freedom are constrained. The simulation model applies a speed of 1750r / min, a radial force of 2000N, and a vertical downward direction. During the load application process, in order to reduce the impact of the sudden load, the load is applied in an increasing manner within a time period of 0 to 0.01 seconds, and then the speed and load are maintained constant. Fixed constraints are applied to the outer ring area of the bearing outer ring to constrain all its displacement and rotational degrees of freedom. These settings ensure that the simulation model is as close to the actual working state as possible, providing reliable data support for subsequent analysis.

[0036] During the contact process between the rolling element and the inner and outer ring raceways and the cage pockets, the rolling element is regarded as an object with a "spherical" convex surface, while the inner and outer ring raceways and the cage pockets are regarded as objects with a "groove-shaped" concave surface. According to the master-slave definition principle of the contact surface, the rolling element is set as the main contact surface, while the inner and outer ring raceways and the cage pockets are set as the slave contact surface. When considering the friction force during the sliding process of the rolling element, the static friction coefficient between the inner and outer rings and the rolling element is set to 0.1, the dynamic friction coefficient is set to 0.05, and the cage and rolling element are set to be frictionless. The contact attenuation coefficient DC is 1×10 -5 , viscous damping coefficient VDC=20.

[0037] The simulation time is set to 1s. According to the 12kHz sensor sampling frequency in the actual experiment, a 1s simulation will generate 12,000 data points. The time step for outputting a d3plot result file in LS-prepost is 8.33×10 -5 s, summarize the relevant keywords and their functional descriptions such as Figure 4 shown.

[0038] Changes in equivalent stress reveal the stress state of a rolling bearing's internal structure during operation. By monitoring these changes, bearing anomalies can be effectively identified. Further analysis of the distribution and changing trends of equivalent stress can help determine the type of rolling bearing failure. Equivalent stress analysis is also a key step in verifying the validity of simulation models.

[0039] The equivalent stress cloud diagram of the simulation results at a certain moment is as follows Figure 5 (a) in Figure 5 (b) Figure 5 (c) Figure 5 The equivalent stress contour diagram shows that the rolling element is the component that bears the greatest equivalent stress during the operation of the rolling bearing and is therefore the component most susceptible to failure and damage.

[0040] While studying inner and outer race failures, rolling element failures are also a key focus. Analysis of rolling element bearings in different health states revealed significant differences in maximum equivalent stress values.

[0041] Specifically, the maximum equivalent stress for a healthy rolling bearing is 1142 MPa; for a rolling bearing with an outer ring fault, it is 2342 MPa; for a rolling bearing with an inner ring fault, it increases to 2507 MPa; and for a rolling element fault, it is 2087 MPa. Although the same radial force is applied to rolling bearings in different health states, their equivalent stress results vary significantly, which is closely related to the characteristics of their vibration signals. These different equivalent stress distributions lead to unique vibration signal characteristics, which not only helps understand the generation mechanism of vibration signals but also fundamentally explains the rationale for using neural networks to extract different vibration signal features for fault diagnosis.

[0042] In order to simulate the vibration signal data obtained by the sensor in the actual experiment, in this simulation model, a fixed node with a distance of three grids from the outer bottom of the bearing to the outermost ring is selected, and its acceleration signal in the vertical direction is extracted to analyze its characteristics and compare with the actual data.

[0043] According to the dynamics theory, the failure frequencies of the main components of rolling bearings are as follows: Outer race fault characteristic frequency:

[0044] Inner race fault characteristic frequency:

[0045] Rolling element failure characteristic frequency:

[0046] By comparing the distribution patterns of simulation data and real data in the time-frequency domain, the reliability of the finite element dynamics simulation model can be further explored. Figure 6 The simulated signal and its envelope diagram, and the actual signal and its envelope diagram in the healthy state are shown. Figure 6 (a) is the simulation signal in the healthy state. Figure 6 (b) in the equation is the actual signal. Figure 6 (c) in the figure is the envelope spectrum analysis of the simulation signal under healthy conditions. Figure 6 (d) in the figure is the envelope spectrum analysis of the actual signal.

[0047] It can be seen that both the simulated signal and the actual signal can clearly show the rotational frequency and its multiple frequencies, which can illustrate the high reliability of the dynamic simulation model under a healthy and normal motion state and the high credibility of the simulation data.

[0048] Figure 7 The time domain and envelope spectrum distribution of the simulated signal and the real signal under the outer race fault state are compared. Figure 7 The simulated signal and its envelope diagram, and the actual signal and its envelope diagram under outer race fault are shown. Figure 7 (a) is the simulation signal under outer race fault. Figure 7 (b) is the actual signal under outer race fault. Figure 7 (c) in the figure is the envelope spectrum analysis of the simulation signal under outer race fault. Figure 7 (d) in the figure is the envelope spectrum analysis of the actual signal under outer race fault.

[0049] It can be observed that in the simulated signal, the interval between two adjacent shock wave peaks is 9.5ms; in the real signal, the interval between two adjacent shock wave peaks is 9.6ms, which is basically consistent with the theoretical calculation formula. In addition, the frequency domain distribution of the envelope spectrum of the two is basically the same, which is also consistent with the theoretical calculation results. This shows that the constructed dynamic model clearly reflects the outer race fault and the difference between it and the actual situation is very small, showing high reliability.

[0050] Figure 8 The simulated signal and its envelope diagram, and the actual signal and its envelope diagram under outer race fault are shown. Figure 8 (a) is the simulation signal under outer race fault. Figure 8 (b) is the actual signal under outer race fault. Figure 8 (c) in the figure is the envelope spectrum analysis of the simulation signal under outer race fault. Figure 8 (d) in the figure is the envelope spectrum analysis of the actual signal under outer race fault.

[0051] It can be observed that the interval between the two shock wave peaks in the simulated signal and the actual signal is around 6.1-6.4ms, which is very close to the 6.3ms obtained by theoretical calculation. In addition, by observing its waveform and envelope spectrum distribution, on the basis of its characteristic frequency, the characteristic frequency caused by the inner ring rotation can also be found. Both the simulated signal and the actual signal have relatively obvious characteristics, which further demonstrates the high reliability of the simulation model.

[0052] Figure 9 The simulated signal and its envelope diagram, and the actual signal and its envelope diagram under outer race fault are shown. Figure 9 (a) is the simulation signal under outer race fault. Figure 9 (b) is the actual signal under outer race fault. Figure 9 (c) in the figure is the envelope spectrum analysis of the simulation signal under outer race fault. Figure 9 (d) in the figure is the envelope spectrum analysis of the actual signal under outer race fault.

[0053] Observing the envelope spectrum frequency domain distribution, it can be found that the vibration signal caused by the rolling element failure is relatively messy and weak, whether in simulation or in practice, but it can still be found that the vibration signal caused by the rolling element failure is relatively messy and weak, but it can still be found that the vibration signal caused by the rolling element failure is relatively weak, which is not obvious in the cage rotation frequency. Under the influence of , the characteristics at its frequency doubling are more obvious, and the simulated signal is also highly consistent with the actual signal.

[0054] The bearing simulation fault data obtained through the dynamic model proposed in the present invention is highly similar to the real data, laying the foundation for the subsequent fault diagnosis model obtained by training based on simulation data to be used in actual scenarios.

[0055] S103: Create a data set based on the fault data.

[0056] In one example, step S103 includes: The fault data is converted into time-frequency images through continuous wavelet transform, and a data set is constructed based on the time-frequency images.

[0057] Fault data can be converted into time-frequency images in Matlab. The specific steps are as follows: Step 1: Calculate the time scale t according to the sampling frequency.

[0058] The sampling frequency here is the sampling frequency of the sensor used to collect data in the subsequent data twin model.

[0059] In one example, step 1 includes: In the first step, the complex Gaussian wavelet cgau8 is selected as the wavelet basis function, and the wavelet scale is set to 128.

[0060] The second step is to calculate the center frequency corresponding to cgau8 through the centfrq function, and convert the scale to frequency according to scal2freq.

[0061] Step 2: Input the vibration signal, time scale and wavelet function into the cwt function to obtain the continuous wavelet coefficients, and use the imagesc function to select every 1024 vibration signal data to generate a time-frequency image.

[0062] It should be noted that, in order to further conform to the actual scenario, the time-frequency image dataset generated based on the time-frequency image does not contain sample labels.

[0063] S104. Construct a neural network model, where the neural network model is a Res-EAS-Net model. The Res-EAS-Net model is obtained by introducing a channel attention mechanism on the basis of the ResNet model.

[0064] In the feature extraction part, the present invention uses the ResNet18 pre-trained model provided in the open source framework PyTorch as the core feature extraction network. ResNet18 has the optimal weight parameters for classification tasks on large-scale datasets (such as ImageNet) and has powerful feature extraction capabilities. ResNet18 contains 18 weight layers (convolutional layers and fully connected layers), and its overall network architecture is realized by introducing basic residual blocks (Basic Block) and skip connections (Skip Connection). The specific parameters are as follows: Figure 10 and Figure 11 shown.

[0065] Since the main bearing states of the boring machine spindle mainly include four health states: health (Health, H), outer race faults (OF), inner race faults (IF) and rolling element faults (BF). Therefore, in the fully connected layer of ResNet18, the number of output categories is changed to 4. In addition, since ResNet18 performs well in image classification tasks, the original pre-trained weights are no longer applicable when facing structured data such as one-dimensional vibration signals. Therefore, the present invention only uses the excellent network architecture of ResNet18 (such as Figure 10 and Figure 11 ), rather than using pre-trained weight parameters. Furthermore, because ResNet18 was originally designed for image classification tasks, whose input dimensionality is higher than one-dimensional inputs like vibration signals, the original two-dimensional convolution of ResNet18 was replaced with one-dimensional convolution to adapt to structured data.

[0066] The classic ResNet18 architecture includes convolutional layers with 128, 256, and 512 channels. This results in redundancy in the feature maps extracted by each channel, increasing model training costs. Therefore, by introducing a channel-attention mechanism that weights channels, the network focuses on important feature channels and suppresses features from less important channels. This approach enables the network to adaptively learn which channels are more important for specific tasks, thereby enhancing network performance.

[0067] ECA-Net is an efficient channel attention mechanism designed to enhance the feature selection capabilities of convolutional neural networks by reducing computational cost. ECA-Net's design principle is to maximize information flow while minimizing computational overhead. This allows it to effectively enhance network performance even with limited computing resources, making it particularly suitable for environments with limited computing resources, such as edge or embedded devices.

[0068] The core idea of ECA-Net is to use one-dimensional convolution to calculate channel attention, avoiding the complex fully connected layer structure and thus achieving efficient calculation.

[0069] The channel attention mechanism in this invention is as follows: The first step is to perform a global average pooling operation on each channel of the input feature map to obtain the descriptor of each channel. The formula for global average pooling is as follows: , where Indicates the Global descriptor for each channel; and Represent the height and width of the feature map respectively; Indicates the first Channels at position The value at .

[0070] In the second step, the descriptor of each channel is used as the input of the one-dimensional convolution layer. , the channel attention weight of each channel is generated through a one-dimensional convolution operation , Indicates the The attention weight of each channel; the convolution kernel of the one-dimensional convolution layer is calculated as follows: , where Indicates the number of channels With convolution kernel The functional relationship between Is the proportional coefficient used to adjust right The degree of impact, is the bias term, Is the proportional coefficient used to adjust right degree of impact.

[0071] Since ECA-Net aims to properly capture local cross-channel information interactions, it is necessary to determine the approximate range of channel interaction information (i.e., the convolution kernel size of 1D convolution). ). It also represents the coverage of local cross-channel interactions, that is, there are neighboring channels participate in the attention weight calculation of this channel. Therefore, Value and number of channels Should be in direct proportion.

[0072] The third step is to weight the features of each channel according to the channel attention weight of each channel to obtain the weighted feature map.

[0073] The weighted formula is: ; The weighted feature map is .

[0074] In this embodiment, on the basis of introducing the channel attention mechanism, an adaptive calibration jump connection mechanism can be introduced for the Res-EAS-Net model; wherein the adaptive calibration jump connection mechanism is as follows: The first step is feature extraction: the output feature map of the previous residual module in the Res-EAS-Net model , perform identity mapping or linear transformation to obtain the skip connection feature .

[0075] The second step is the adaptive weight generation step: introducing the weight generation function , generates weight coefficients for skip connection features; where the weight coefficients The generation formula is: , where 、 represents the learnable parameters, express activation function, Represents global average pooling.

[0076] The third step is feature calibration: using weight coefficients , for skip connection features Perform weighting.

[0077] The fourth step is the feature fusion step: the weighted jump connection features are combined with the main path output features Add together to get the final output features .

[0078] In ResNet18, the skip connection effectively alleviates the problems of gradient vanishing and network degradation by directly passing the input features to the output end. However, traditional skip connections use fixed identity mappings or simple linear transformations, which fail to fully consider the differences in the importance of input features. To this end, the present invention proposes an adaptive calibration skip connection (ACSC) mechanism, which dynamically adjusts the feature weights of the skip connection to enhance the model's ability to focus on key features, thereby improving feature expression capabilities and model performance. The core idea of ACSC is to dynamically calibrate the features of the skip connection by introducing learnable weight parameters. Specifically, the mechanism generates a set of weight coefficients based on the global or local information of the input features, and performs weighted adjustment on the features of the skip connection to make them more adaptable to the output features of the main path. This dynamic calibration can effectively capture the key information in the input features while suppressing redundant or noisy features.

[0079] S105: Training a neural network model according to the data set to obtain a first fault diagnosis model for the boring machine parts.

[0080] In the training part of Res-EAS-Net, the generated simulation data is divided into training set, validation set and test set in a ratio of 6:2:2. The cross entropy loss function (CrossEntropyLoss) is selected as the loss function:

[0081] Where, is the number of samples, is the number of categories, and Respectively represent The true label and predicted probability of each sample in the category.

[0082] In terms of hyperparameter selection, the number of training epochs was set to 100, and an early stopping strategy was employed. This strategy checks the loss at the end of each epoch. If the loss does not significantly decrease within the specified epoch, the model is considered to have converged and training is terminated. Accordingly, the initial learning rate was set to 0.01, and the ReduceLROnPlateau learning rate decay strategy from the lr_scheduler package was used to avoid getting stuck in local optima during training. This strategy shares the core concept of Early Stopping, differing in that after the learning rate decays according to a certain ratio, a cool-down period is entered to avoid frequent triggering of learning rate decay. The Dataloader from torch.utils.data was used as a batch loader to read a certain number of samples from the dataset for training (or testing). The batch_size was set to 16, and shuffle=True. At the end of training, the weights of the optimal model were stored as a dictionary and saved in pth format.

[0083] S106: Using transfer learning to train the first fault diagnosis model to obtain a second fault diagnosis model for boring machine parts.

[0084] In one example, step S106 includes: In the first step, time-frequency data under two different working conditions are selected as the source domain and target domain respectively as the transfer learning dataset; The second step is to freeze the parameter updates of the weight layers except the fully connected layers in the Res-EAS-Net model, and only update the parameters of the fully connected layers through the back-propagation algorithm; The third step is to map the data of both the source and target domains into the reproducing kernel Hilbert space, and align the features of the source and target domains by the maximum mean difference. The fourth step is to define the new loss function as , and train the first fault diagnosis model according to the new loss function to obtain the second fault diagnosis model; in the above formula, represents the classification loss, represents the migration loss calculated by maximizing the mean difference, Represents the weight of the transfer loss calculated by maximizing the mean difference.

[0085] The calculation formula is as follows:

[0086] in, represents the kernel function mapping, Represents the characteristics of the source domain, Represents the characteristics of the target domain, represents the reproducing kernel Hilbert space.

[0087] Relying on this loss function, the Res-EAS-Net model is trained again, and the training strategy and parameter settings are the same as step S105. After obtaining the optimal parameters of the fully connected layer, the Res-EAS-Net is tested on the test set to achieve migration diagnosis of real data.

[0088] See also Figure 12 , specifically showing the training process of the second fault diagnosis model.

[0089] See also Figure 13 To further verify the performance of the Res-EAS-Net model, the test set uses the LetNet model, AlexNet model, VGG16 model and the original ResNet18 model as the control group to verify the performance of the Res-EAS-Net model. The verification results are as follows Figure 13 As shown in the figure, it can be seen that thanks to transfer learning, the four indicator scores of the proposed model Res-EAS-Net are the highest, among which the Recall value reaches 0.903, which means that Res-EAS-Net can more accurately identify the three types of fault samples. The highest Gmean indicates that the overall classification performance of the model is optimal.

[0090] exist Figure 13 In , Precision represents the accuracy, Recall represents the recall rate, and Gmean represents the geometric mean. It is more suitable for measuring the overall performance of the classifier. Only when the classifier performs well in each category can a large Gmean and Recall is suitable for measuring the local performance of the classifier because it focuses more on fault samples.

[0091] Among them, Gmean and The value is calculated as follows:

[0092] in, represents the weight, is 2.

[0093] The calculation formulas for Precision and Recall are as follows: ,

[0094] S107. Establish a digital twin model of the boring machine to collect and synchronize status data of the boring machine components in real time.

[0095] In one example, step S107 includes: The first step is to build a 3D model of the boring machine.

[0096] The physical boring machine is surveyed and mapped, and the structural dimensions of key parts and various components are measured. Based on the structural dimensions of the key parts and various components of the boring machine, the parts are drawn and assembled in SolidWorks to obtain a three-dimensional model of the boring machine.

[0097] The second step is to render the 3D model.

[0098] The three-dimensional model is rendered in color using 3Dmax software.

[0099] The third step is to adjust the size of the rendered 3D model to obtain the digital twin model.

[0100] Since the sizes of the actually obtained three-dimensional model and the actual boring machine model are different, the sizes of the three-dimensional model and the actual boring machine model need to be synchronized.

[0101] Sensors are installed on the spindle's X, Y, and Z axes and on the spindle's gearbox housing to detect vibration signals (i.e., acceleration), synchronizing the data twin model with the actual boring machine's status. Finally, the vibration signals collected by the sensors are input into a second diagnostic model to diagnose the boring machine's condition. The status of the boring machine's components can also be displayed in the data twin model using different colors.

[0102] See also Figures 14 to 18 , which is a schematic diagram of the digital twin model provided by the present invention.

[0103] S108. Determine the current status of the boring machine components based on the status data of the boring machine components collected by the digital twin model and the second fault diagnosis model of the boring machine components, and display the current status of the boring machine components in the data twin model.

[0104] Figure 19 This is a structural diagram of a digital twin diagnostic device for a boring machine based on data and finite element models provided by the present invention. Figure 19 ,include: A finite element model building module 11 is used to build a finite element model of the boring machine parts; A simulation module 12 is used to simulate the boring machine parts under different failure modes based on the finite element model to obtain failure data of the boring machine parts; A data set building module 13 is used to build a data set based on the fault data; A neural network model construction module 14 is used to construct a neural network model, wherein the neural network model is a Res-EAS-Net model, which is obtained by introducing a channel attention mechanism on the basis of the ResNet model; A first training module 15 is used to train a neural network model according to the data set to obtain a first fault diagnosis model for the boring machine parts; A second training module 16 is used to train the first fault diagnosis model using transfer learning to obtain a second fault diagnosis model for the boring machine component; A digital twin model building module 17 is used to build a digital twin model of the boring machine, and is used to collect and synchronize status data of the boring machine components in real time; The diagnosis module 18 is used to determine the current status of the boring machine components based on the status data of the boring machine components collected by the digital twin model and the second fault diagnosis model of the boring machine components, and display the current status of the boring machine components in the data twin model.

[0105] Figure 20 This is a structural block diagram of an electronic device provided by the present invention. Figure 20 , electronic devices may include Figure 19 The digital twin diagnostic device for boring machines based on data and finite element models generally includes a processor 21 and a memory 22 .

[0106] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented in hardware using at least one of the following: a DSP (Digital Signal Processing), an FPGA (Field-Programmable Gate Array), or a PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state. The memory 22 may include one or more computer-readable storage media, which may be non-transitory. The memory 22 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 22 is used to store at least one instruction, which is executed by the processor 21 to implement the digital twin diagnostic method for a boring machine based on data and a finite element model, provided by an electronic device in the method embodiments of this application.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A digital twin diagnostic method for boring machines based on data and finite element models, characterized in that: include: Establish finite element models of boring machine components; Based on the finite element model, the boring machine parts are simulated under different failure modes to obtain the failure data of the boring machine parts; Establish a data set based on fault data; Constructing a neural network model, wherein the neural network model is a Res-EAS-Net model, and the Res-EAS-Net model is obtained by introducing a channel attention mechanism on the basis of the ResNet model; The neural network model is trained based on the data set to obtain the first fault diagnosis model of the boring machine parts; Transfer learning is used to train the first fault diagnosis model to obtain a second fault diagnosis model for boring machine parts; Establish a digital twin model of the boring machine to collect and synchronize the status data of the boring machine components in real time; Based on the status data of the boring machine parts collected by the digital twin model and the second fault diagnosis model of the boring machine parts, the current status of the boring machine parts is determined, and the current status of the boring machine parts is displayed in the data twin model.

2. The digital twin diagnostic method for boring machines based on data and finite element models according to claim 1, characterized in that: Failure modes of boring machine components include rolling element defects, inner race failures, and outer race failures.

3. The digital twin diagnostic method for boring machines based on data and finite element models according to claim 1, characterized in that: Fault data includes vibration data.

4. The digital twin diagnostic method for boring machines based on data and finite element models according to claim 3 is characterized in that: Based on the fault data, the steps to build a data set include: The fault data is converted into time-frequency images through continuous wavelet transform, and a data set is constructed based on the time-frequency images.

5. The digital twin diagnosis method for boring machines based on data and finite element models according to any one of claims 1 to 4, characterized in that: The channel attention mechanism is as follows: Perform a global average pooling operation on each channel of the input feature map to obtain the descriptor of each channel; the formula for global average pooling is as follows: , where Indicates the Global descriptor for each channel; and Represent the height and width of the feature map respectively; Indicates the first Channels at position The value at The obtained descriptor of each channel is used as the input of the one-dimensional convolution layer, and the channel attention weight of each channel is generated through the one-dimensional convolution operation; Among them, the convolution kernel of the one-dimensional convolution layer is calculated as follows: , where Indicates the number of channels With convolution kernel The functional relationship between Is the proportional coefficient used to adjust right The degree of impact, is the bias term, Is the proportional coefficient used to adjust right the extent of the impact; According to the channel attention weight of each channel, the features of each channel are Perform weighting to obtain the weighted feature map .

6. The digital twin diagnosis method for boring machines based on data and finite element models according to any one of claims 1 to 5, characterized in that: Based on the introduction of the channel attention mechanism, an adaptive calibrated skip connection mechanism is introduced for the Res-EAS-Net model. The adaptive calibrated skip connection mechanism is as follows: Feature extraction step: output feature map of the previous residual module in the Res-EAS-Net model , perform identity mapping or linear transformation to obtain the skip connection feature ; Adaptive weight generation step: Introducing weight generation function , generates weight coefficients for skip connection features; where the weight coefficients The generation formula is: , where 、 represents the learnable parameters, express activation function, represents global average pooling; Feature calibration step: using weight coefficients , for skip connection features Perform weighting; Feature fusion step: combine the weighted jump connection features with the main path output features Add together to get the final output features .

7. The digital twin diagnosis method for boring machines based on data and finite element models according to any one of claims 1 to 5, characterized in that: The steps of using transfer learning to train the first fault diagnosis model to obtain a second fault diagnosis model for boring machine parts include: Select time-frequency data under two different working conditions as the source domain and target domain respectively as the transfer learning dataset; Freeze the parameter updates of the weight layers except the fully connected layers in the Res-EAS-Net model, and only update the parameters of the fully connected layers through the backpropagation algorithm; Map the data of both the source and target domains into the reproducing kernel Hilbert space, and align the features of the source and target domains by maximum mean difference; The new loss function is defined as , and train the first fault diagnosis model according to the new loss function to obtain the second fault diagnosis model; in the above formula, represents the classification loss, represents the migration loss calculated by maximizing the mean difference, Represents the weight of the transfer loss calculated by maximizing the mean difference.

8. A digital twin diagnostic device for boring machines based on data and finite element models, characterized in that: include: Finite element model building module, used to build finite element models of boring machine parts; A simulation module is used to simulate the boring machine parts under different failure modes based on the finite element model to obtain the failure data of the boring machine parts; A data set construction module is used to build a data set based on fault data; A neural network model construction module is used to construct a neural network model, wherein the neural network model is a multi-scale feature extraction network based on a sparse convolutional autoencoder; A first training module is used to train a neural network model according to a data set to obtain a first fault diagnosis model of the boring machine parts; A second training module is used to train the first fault diagnosis model using transfer learning to obtain a second fault diagnosis model for boring machine parts; A digital twin model building module is used to build a digital twin model of the boring machine, which is used to collect and synchronize the status data of the boring machine components in real time; The diagnostic module is used to determine the current status of the boring machine components based on the status data of the boring machine components collected by the digital twin model and the second fault diagnosis model of the boring machine components, and display the current status of the boring machine components in the data twin model.

9. An electronic device, characterized in that: Including the digital twin diagnostic device of a boring machine based on data and a finite element model as described in claim 8.

10. A computer-readable storage medium, characterized in that At least one program code is stored in the computer-readable storage medium, and the program code is executed by a processor to implement the digital twin diagnosis method for a boring machine based on data and a finite element model as described in any one of claims 1 to 7.

Images