Digital twinning auxiliary diagnosis method for bearing fault data imbalance
By constructing the bearing digital twin and using variational modal decomposition and Bayesian optimization convolutional neural network, the problems of bearing failure data imbalance and insignificant early characteristics are solved, and efficient fault diagnosis is achieved.
Patent Information
- Application Number
- CN202510638541.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-16
- Publication Date
- 2025-07-11
AI Technical Summary
现有技术在轴承故障诊断中,面临数据不平衡和早期故障特征不显著的问题,导致模型对故障识别准确率不高。
The bearing digital twin is constructed, a two-dimensional time-frequency image is generated using the variational modal decomposition method, and a Bayesian optimized convolutional neural network is trained. The model is adjusted through the adaptive loss function to achieve consistency between the actual data and simulated data in the feature space, and fault diagnosis is performed using transfer learning.
It improves the accuracy of bearing fault diagnosis, can effectively extract early fault characteristics, and improves the recognition ability of the model.
Smart Images
Figure CN120293525A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of bearing fault diagnosis, and particularly to a digital twin-assisted diagnosis method for bearing fault data imbalance. Background Technique
[0002] The health monitoring of mechanical equipment can improve the reliability of production and reduce the equipment maintenance cost. As the core component for supporting and positioning the rotation of the engine, bearings are widely used in the transmission systems of various mechanical equipment. Due to long-term operation under complex conditions, various fault modes will inevitably occur. In recent years, data-driven methods based on deep learning have been widely used in the field of fault diagnosis due to their powerful feature extraction capabilities. However, the early fault features of bearings are weak and are not easily captured effectively by the model. Moreover, bearings usually operate under normal working conditions, and the fault data collected by sensors is far less than the normal data, resulting in data imbalance. When the diagnostic model is trained with imbalanced data, minimizing the overall loss will force the network to identify the normal state of the majority class while ignoring the fault state of the minority class, resulting in low accuracy of the model in fault recognition.
[0003] The existing relatively advanced fault diagnosis methods combine the ideas of digital twin and transfer learning, use digital twin technology to generate simulation data close to the actual operating state of the bearing equipment, and then obtain a fault diagnosis model applicable to the actual bearing equipment in combination with the transfer learning method. However, the way of using balanced data for modeling in such methods cannot effectively reflect the actual fault situation of the bearings, and it is also necessary to consider that the early fault features of the bearings are not easily extracted effectively by the model. Therefore, it is an urgent problem for those skilled in the art to use digital twin and transfer learning technologies to solve the problems of actual bearing fault data imbalance and insignificant early fault features. Summary of the Invention
[0004] The purpose of the present invention is to disclose a digital twin-assisted diagnosis method for bearing fault data imbalance, so as to solve the problem that the low accuracy of the model in fault recognition is caused by the imbalance of actual bearing fault data and the insignificance of early fault features.
[0005] To achieve the above purpose, the present invention provides a digital twin-assisted diagnosis method for bearing fault data imbalance, including the following steps:
[0006] (1) Analyze the nonlinear dynamic characteristics of the bearing and construct a bearing digital twin.
[0007] (2) Use the bearing digital twin to generate vibration signals in the normal working state and the fault working state of the bearing.
[0008] (3) Use the variational mode decomposition method to obtain two-dimensional time-frequency images of the vibration signals in the normal working state and the fault working state of the bearing.
[0009] (4) Use Bayesian optimization of convolutional neural network to train the time-frequency image data of the twin vibration signals to obtain an initial fault diagnosis model.
[0010] (5) Analyze the characteristics of the actual data and simulation data in the feature space, and use the actual bearing vibration signals to fine-tune the initial fault diagnosis model to obtain the final bearing fault diagnosis model.
[0011] For further illustration, in step (1), set parameters such as the geometric shape and material properties of the bearing, analyze the nonlinear dynamic characteristics of the bearing, and construct a bearing digital twin in combination with the actual operating characteristics of the bearing to facilitate the generation of high-quality simulation vibration signals that accurately reflect the actual equipment.
[0012] For further illustration, in step (2), set the parameters of the bearing digital twin according to the normal and fault vibration signals of the actual bearing, and at the same time obtain the unbalanced vibration signals that reflect the actual situation from the actual bearing equipment.
[0013] For further illustration, in step (3), the variational mode decomposition (VMD) method decomposes the vibration signal to obtain two-dimensional time-frequency image modes, and the obtained modes can be expressed as:
[0014] u k (t) = A k (t)cos(φ k (t)) (1)
[0015] A k (t) is the instantaneous amplitude, φ k (t) is the instantaneous phase, and u k (t) is the decomposed mode. The VMD method obtains the mode components through variational solution, estimates the center frequency ω k of the mode using the Hilbert transform, then multiplies it by and calculates the gradient norm L 2 of the analytic signal. The constraint process is:
[0016]
[0017] δ(t) is the Dirac function, is the derivative operator, x(t) is the original signal, and the VMD method can convert the bearing signal into a two-dimensional time-frequency image through an unconstrained variational solution process, making the features contained in the signal more prominent.
[0018] For further illustration, in step (4), the learning rate and regularization parameters of the convolutional neural network CNN are optimized by Bayesian optimization. Taking the decomposed time-frequency image modality as the input and the fault label as the output, the optimal parameters of the CNN model are obtained. After the vibration signal features are extracted by the Bayesian-optimized CNN model, an initial fault diagnosis model is obtained.
[0019] For further illustration, in step (5) during model training, the adaptive loss L of the actual data and the simulation data is analyzed AS to minimize the difference between the means of the two sets of data after dimensionality reduction, making the actual data and the simulation data more consistent in the feature space. The adaptive loss L AS Performs dimensionality reduction processing and mean calculation on the kernel matrices of the actual data and the simulation data, which can be expressed as:
[0020]
[0021] C is the number of fault categories, x A (t) and x S (t) are the samples of the actual data and the simulation data belonging to C respectively, w A and w B are coefficients, and f(g) is the data feature mapping process. By using the adaptive loss L AS to analyze the distribution of the actual data and the simulation data in the feature space, the distribution difference between the two sets of data is effectively reduced.
[0022] For further illustration, in step (5), using the digital twin and transfer learning ideas, the trained model extracts features from the data that mixes the actual data and the simulation data and has been decomposed by VMD to obtain two-dimensional time-frequency images, and a final bearing fault diagnosis model is obtained.
[0023] Compared with the prior art, the beneficial technical effects of the present invention are as follows:
[0024] 1. The present invention uses the VMD method to convert the bearing vibration signal into a two-dimensional time-frequency image. The obtained time-frequency image modality can better reflect the characteristic information of the signal, can solve the problem of insignificant early fault characteristics of the bearing, and is beneficial to the extraction of signal features by the diagnosis model.
[0025] 2. The present invention considers the problem of unbalanced data collected by actual bearing equipment. By analyzing the adaptive loss of the actual bearing data and the simulation data, the difference between the means of the two sets of data after dimensionality reduction is minimized, making the actual data and the simulation data more consistent in the feature space. The model trained by using transfer learning has better fault diagnosis ability. Description of the Drawings
[0026] Figure 1Flowchart of a digital twin-assisted diagnosis method for bearing fault data imbalance according to the present invention.
[0027] Figure 2 Schematic diagram of the two-dimensional time-frequency image obtained by VMD decomposition of the vibration signal.
[0028] Figure 3 Schematic diagram of the confusion matrix of the present invention on the test data. Detailed implementation manners
[0029] The present invention will be further described below in conjunction with the accompanying drawings and detailed implementation manners.
[0030] In actuality, bearing equipment generally operates in a normal working state, and the fault data is less than the normal data, resulting in the imbalance of the collected data. In addition, the early fault characteristics of the bearing are not obvious and are not easily effectively extracted by the model. In view of these characteristics of the bearing vibration signal, referring to Figure 1 , a digital twin-assisted diagnosis method for bearing fault data imbalance includes the following steps:
[0031] (1) Analyze the nonlinear dynamic characteristics of the bearing and construct a bearing digital twin.
[0032] Set parameters such as the geometric shape and material properties of the bearing, analyze the nonlinear dynamic characteristics of the bearing, and construct a bearing digital twin in combination with the actual operating characteristics of the bearing, so as to generate high-quality simulation vibration signals that accurately reflect the actual equipment.
[0033] (2) Use the bearing digital twin to generate vibration signals in the normal working state and fault working state of the bearing.
[0034] According to the normal and fault vibration signals of the actual bearing, set the parameters of the bearing digital twin to generate balanced simulation vibration signals, and at the same time obtain unbalanced vibration signals reflecting the actual situation from the actual bearing equipment.
[0035] (3) Use the variational mode decomposition method to obtain the two-dimensional time-frequency images of the vibration signals in the normal working state and fault working state of the bearing.
[0036] The variational mode decomposition (VMD) method obtains the two-dimensional time-frequency image modes by decomposing the vibration signal, and the obtained modes can be expressed as:
[0037] u k (t) = A k (t)cos(φ k (t)) (1)
[0038] A k (t) is the instantaneous amplitude, φ k (t) is the instantaneous phase, uk (t) is the decomposed mode. The VMD method obtains the mode components through variational solution and estimates the central frequency ω of the mode using the Hilbert transform k , and then multiplies it with and calculates the gradient norm L of the analytic signal 2 . The constraint process is as follows:
[0039]
[0040] δ(t) is the Dirac function, is the derivative operator, x(t) is the original signal. Through the unconstrained variational solution process, the VMD method can convert the bearing signal into a two-dimensional time-frequency image, making the features contained in the signal more prominent.
[0041] (4) Use the Bayesian optimization convolutional neural network to train the time-frequency image data of the twin vibration signals to obtain an initial fault diagnosis model.
[0042] Utilize the learning rate and regularization parameters of the Bayesian optimization convolutional neural network CNN. With the decomposed time-frequency image mode as the input and the fault label as the output, obtain the optimal parameters of the CNN model. After the Bayesian optimization CNN model extracts the vibration signal features, an initial fault diagnosis model is obtained.
[0043] (5) Analyze the characteristics of the actual data and simulation data in the feature space, and use the actual bearing vibration signal to fine-tune the initial fault diagnosis model to obtain the final bearing fault diagnosis model.
[0044] During model training, analyze the adaptive loss L of the actual data and simulation data AS to minimize the difference between the means of the two groups of data after dimensionality reduction, making the actual data and simulation data more consistent in the feature space. The adaptive loss L AS performs dimensionality reduction processing and mean calculation on the kernel matrices of the actual data and simulation data, and can be expressed as:
[0045]
[0046] C is the number of fault categories, x A (t) and x S (t) are the samples of the actual data and simulation data belonging to C respectively, w A and w B are coefficients, and f(g) is the data feature mapping process. By using the adaptive loss L AS to analyze the distribution of the actual data and simulation data in the feature space, the distribution difference between the two groups of data is effectively reduced.
[0047] Using the ideas of digital twin and transfer learning, the trained model extracts features from the data that mixes actual data and simulation data and has been decomposed by VMD to obtain two-dimensional time-frequency images, and finally obtains a bearing fault diagnosis model.
[0048] This invention uses a certain bearing device for analysis. The data collected by the bearing includes four categories: normal bearing data, inner ring fault data, outer ring fault data, and rolling element fault data, which are respectively simplified and represented by the labels 1, 2, 3, and 4. Refer to Figure 2 , using the VMD method to decompose the data, the obtained two-dimensional time-frequency images can better reflect the characteristic information of the signal, can solve the problem that the early fault characteristics of the bearing are not obvious, and is beneficial to the extraction of signal characteristics by the diagnosis model. Refer to Figure 3 , it has a high accuracy in the sample classification of the test set, and only a very small number of samples have problems with classification label confusion.
[0049] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are illustrative rather than restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the purpose of the present invention and the scope protected by the claims. All of these fall within the protection scope of the present invention.
Claims
1. A digital twin-assisted diagnosis method for bearing fault data imbalance, characterized in that, It includes the following steps: (1) Analyze the nonlinear dynamic characteristics of the bearing and construct a bearing digital twin. (2) Use the bearing digital twin to generate vibration signals in the normal working state and the fault working state of the bearing. (3) Use the variational mode decomposition method to obtain two-dimensional time-frequency images of the vibration signals in the normal working state and the fault working state of the bearing. (4) Adopt Bayesian optimization of the convolutional neural network to train the time-frequency image data of the twin vibration signals to obtain an initial fault diagnosis model. (5) Analyze the characteristics of the actual data and the simulation data in the feature space, and use the actual bearing vibration signals to fine-tune the initial fault diagnosis model to obtain the final bearing fault diagnosis model.
2. The digital twin-assisted diagnosis method for bearing fault data imbalance according to claim 1, wherein Step (1) includes: Set parameters such as the geometric shape and material properties of the bearing, analyze the nonlinear dynamic characteristics of the bearing, and construct a bearing digital twin in combination with the actual bearing operation characteristics to facilitate the generation of high-quality simulation vibration signals that accurately reflect the actual equipment.
3. A digital twin-assisted diagnosis method for bearing fault data imbalance according to claim 1, characterized in that Step (2) includes: According to the normal and fault vibration signals of the actual bearing, set the parameters of the bearing digital twin to generate balanced simulation vibration signals, and at the same time obtain unbalanced vibration signals reflecting the actual situation from the actual bearing equipment.
4. A digital twin-assisted diagnosis method for bearing fault data imbalance according to claim 1, characterized in that, In step (3), the variational mode decomposition VMD method obtains two-dimensional time-frequency image modes by decomposing the vibration signals, and the obtained modes can be expressed as: u k u(t) = a k u(t)cos(φ k (t)) (1) A k (t) is the instantaneous amplitude, φk(t) is the instantaneous phase, and u k (t) is the decomposed mode. The VMD method obtains the mode components through variational solution, estimates the center frequency ω of the mode using the Hilbert transform k , and then multiplies it by and calculates the gradient norm L of the analytic signal 2 . The constraint process is as follows: δ(t) is the Dirac function, is the derivative operator, x(t) is the original signal, and the VMD method can convert the bearing signal into a two-dimensional time-frequency image through an unconstrained variational solution process, making the features contained in the signal more prominent.
5. A digital twin-assisted diagnosis method for bearing fault data imbalance according to claim 1, characterized in that In step (4), use Bayesian optimization to optimize the learning rate and regularization parameters of the convolutional neural network CNN. With the decomposed time-frequency image modes as the input and the fault labels as the output, obtain the optimal parameters of the CNN model. After Bayesian optimization, the CNN model extracts the features of the vibration signals to obtain an initial fault diagnosis model.
6. A digital twin assisted diagnosis method for bearing fault data imbalance according to claim 1, characterized in that In step (5), during model training, analyze the adaptive loss L of the actual data and the simulation data AS to minimize the difference between the means of the two sets of data after dimensionality reduction, making the actual data and the simulation data more consistent in the feature space. The adaptive loss L AS Perform dimensionality reduction processing and mean calculation on the kernel matrices of the actual data and the simulation data, which can be expressed as: Let \(C\) be the number of fault categories, and \(x\) A (t) and \(x\) S (t) are the samples of the actual data and the simulation data belonging to \(C\) respectively, \(w\) A and \(w\) B are coefficients, and \(f(\cdot)\) is the data feature mapping process. By using the adaptive loss \(L\) AS analyze the distribution of the actual data and the simulation data in the feature space, and effectively reduce the distribution difference between the two sets of data.
7. A digital twin-assisted diagnosis method for bearing fault data imbalance according to claim 1, characterized in that In step (5), using the ideas of digital twin and transfer learning, the trained model extracts features from the two-dimensional time-frequency image modes obtained by VMD decomposition of the mixture of actual data and simulation data to obtain the final bearing fault diagnosis model.