Rolling bearing intelligent fault diagnosis method based on MVO-VMD and CKCT-Net
By optimizing the VMD and CKCT-Net models through MVO, rolling bearing fault features are automatically extracted and integrated, solving the problem of traditional methods relying on manual experience, achieving efficient and accurate fault diagnosis, and improving the health management level of mechanical equipment.
Patent Information
- Application Number
- CN202510833632.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-20
- Publication Date
- 2025-10-17
AI Technical Summary
Existing rolling bearing fault diagnosis methods rely on manual experience, feature extraction is not automated enough, diagnostic accuracy is limited, and it is difficult to efficiently and accurately identify fault types under complex working conditions.
The multiverse optimization (MVO) algorithm is used to optimize variational mode decomposition (VMD) to extract key frequency domain features. The CKCT-Net model is combined to extract, enhance and fuse fault features. The multi-head attention mechanism is used for global modeling to achieve automated and efficient fault diagnosis.
It improves the accuracy and robustness of fault diagnosis, can better adapt to complex distribution characteristics, realizes multi-level feature fusion and information collaboration, and enhances the generalization ability of the model.
Smart Images
Figure CN120804967A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of health management, and relates to a rolling bearing fault diagnosis method, in particular to a rolling bearing intelligent fault diagnosis method based on a multiverse optimization algorithm (MVO)-variable mode decomposition (VMD) and CKCT-Net. BACKGROUND
[0002] Rolling bearings are key components in modern rotating machinery, and their main functions are to support mechanical shafts, reduce frictional resistance, absorb vibrations, and reduce noise. They are widely used in aerospace, automobile manufacturing, energy exploration, and agricultural machinery. However, under such complex working conditions, rolling bearings often need to withstand high loads, vibrations, friction, temperature changes, and other factors. These factors make the bearings prone to failure, which can lead to the collapse of the entire mechanical system and even cause serious safety accidents. Therefore, developing efficient and accurate rolling bearing fault diagnosis research is of great significance for improving the health management level of mechanical equipment and prolonging the service life of equipment.
[0003] Rolling bearings have various forms of failure, mainly including outer rings, inner rings, and rolling balls. Different fault types have different characteristics, such as cracks, pitting, peeling, and wear. These fault characteristics are usually hidden in complex mechanical vibration signals and are significantly affected by noise, making it extremely difficult to extract fault information. Traditional rolling bearing fault diagnosis methods mainly rely on signal processing techniques and statistical analysis, such as single time or frequency domain analysis, wavelet transform, and empirical mode decomposition. These methods can extract feature indicators from vibration signals and combine experience or thresholds to determine fault types, but they have challenges such as reliance on manual experience and limited diagnostic accuracy. With the continuous development of sensor technology and data acquisition systems in recent years, a large amount of rolling bearing operation data has been accumulated, providing a rich data basis for fault diagnosis research. At the same time, the rise of deep learning technology, especially data-driven methods, provides a new research approach for fault diagnosis. SUMMARY
[0004] The application provides a rolling bearing intelligent fault diagnosis method based on MVO-VMD and CKCT-Net, which analyzes the bearing fault from a large amount of data, thereby improving the health management level of mechanical equipment and prolonging the service life of equipment.
[0005] The purpose of the application is achieved through the following technical solutions:
[0006] A rolling bearing intelligent fault diagnosis method based on MVO-VMD and CKCT-Net, comprising the following steps:
[0007] Step one, fault feature analysis:
[0008] Step one, decompose the original bearing fault data into 5 intrinsic mode functions (IMFs) with clear center frequency using variational mode decomposition (VMD);
[0009] Step two, control the bandwidth and orthogonality of the mode by manually setting the key parameters (modal number K, penalty coefficient) to extract the key frequency domain features in the signal;
[0010] Step three, introduce the multi-universe optimization algorithm (MVO) to automatically optimize the key parameters of VMD, and obtain the tensor data composed of multiple IMFs;
[0011] Step two, fault feature extraction:
[0012] Step two, input the tensor data obtained in step one into a designed 5-layer convolutional network architecture, each layer composed of a convolutional layer and a max pooling layer;
[0013] Step two, in the convolutional layer, each convolutional kernel performs sliding window operation to identify the local mode of the signal;
[0014] Step three, after each convolution operation, use the activation function to perform nonlinear transformation on the convolution output to enhance the network's sensitivity to sudden changes in vibration signals (impact signals);
[0015] Step four, introduce a max pooling layer after the convolutional layer, which compresses the spatial dimension of the convolutional features through downsampling in the max pooling layer, reducing the number of features while retaining the most critical feature information (such as the highest amplitude);
[0016] Step three, nonlinear fault feature enhancement:
[0017] Design a feature enhancement module containing two B-spline functions and three fully connected layers to enhance the representational ability of the extracted features, the specific steps are as follows:
[0018] Step three, use the first fully connected layer to map each input feature dimension from D to hidden dimension H, and redistribute the local features;
[0019] Step two, input the features output by the first fully connected layer into the first B-spline function;
[0020] Step three, use the output of the first B-spline function as the input of the second fully connected layer mapping and activation;
[0021] Step four, input the features output by the second fully connected layer into the second B-spline function;
[0022] Step three five, the activated features of the second layer B spline function are mapped back to the original input dimension through a third full connection layer, so as to obtain nonlinear enhanced fault features;
[0023] Step four, fault feature fusion:
[0024] The local fault features extracted in step two and the nonlinear enhanced fault features extracted in step three are interacted, so as to obtain bearing fault fusion features;
[0025] Step five, global modeling:
[0026] Step five one, the bearing fault fusion features are taken as an input sequence and embedded into two layers of Encoder, so as to convert the fused fault features into a vector form that can be processed by the model;
[0027] Step five two, position coding is added to the output of the embedding layer, so as to retain the time sequence dependence of the fault feature sequence;
[0028] Step five three, after the input vector enters the encoder, the input sequence itself is modeled through a multi-head attention mechanism, and the correlation between the features of each position in the sequence and other positions is calculated.
[0029] Compared with the prior art, the present application has the following advantages:
[0030] The present application uses MVO to optimize VMD, so as to automatically determine the optimal K value and avoid relying on artificial experience setting; the CKCT-Net model proposed enhances the local feature capturing advantage and global sequence modeling capability, effectively makes up for the defects that the local details and global dependence are difficult to be considered in the feature extraction process of the traditional method; meanwhile, the model improves the adaptability of the model to complex distribution characteristics by enhancing the nonlinear feature representation capability; and the multi-level feature fusion and information cooperation and complementation between different modes are realized by using the cross-modal information interaction mechanism. The experimental results show that the method is superior to other advanced methods in the fault diagnosis task, and has better generalization and robustness. BRIEF DESCRIPTION OF DRAWINGS
[0031] Figure 1 A fault diagnosis method model;
[0032] Figure 2 A data feature analysis diagram;
[0033] Figure 3 A nonlinear feature enhancement process diagram;
[0034] Figure 4 A feature fusion diagram;
[0035] Figure 5 A global modeling diagram;
[0036] Figure 6 Validation results plot on the public dataset of Case Western Reserve University;
[0037] Figure 7 Validation results plot on the public dataset of Padborn University. DETAILED DESCRIPTION
[0038] The technical solutions of the present application are further described below in conjunction with the drawings, but are not limited thereto, and any modification or equivalent replacement of the technical solutions of the present application without departing from the spirit and scope of the technical solutions of the present application shall be covered in the protection scope of the present application.
[0039] The present application provides an intelligent fault diagnosis method based on MVO-VMD and CKCT-Net, which comprises the following steps, as shown in the figure: Figure 1
[0040] Step one, fault feature analysis:
[0041] The original bearing fault data is decomposed into 5 intrinsic mode functions (IMFs) with clear center frequency by using variational mode decomposition (VMD), the bandwidth and orthogonality of the mode are controlled by manually setting key parameters (modal number K, penalty coefficient), the key frequency domain features in the signal are extracted, but manual setting will make the decomposition result highly sensitive to parameter selection. When K is set too small, the bearing fault feature cannot be completely decomposed; if it is set too large, redundant modes or noise will be introduced. In view of this, the present application introduces a multivariate universe optimization algorithm (MVO) to automatically optimize and select the key parameters of VMD, and obtains a tensor data composed of multiple IMFs, so as to get rid of the dependence of parameter selection on artificial experience.
[0042] Step two, fault feature extraction:
[0043] After MVO optimizes VMD, a tensor data composed of multiple IMFs can be obtained, which is input into a 5-layer convolutional network architecture, each layer being composed of a convolutional layer and a maximum pooling layer. In the convolutional layer, each convolutional kernel performs a sliding window operation to identify the local mode of the signal. At the same time, multi-layer convolution is used to gradually extract high-order features. After each convolution operation, an activation function is used to perform a nonlinear transformation on the convolution output to enhance the sensitivity of the network to the mutation of the vibration signal (impact signal). In order to further reduce the computational complexity and prevent overfitting, a maximum pooling layer is introduced after the convolutional layer, and the spatial dimension of the convolution features is compressed through the down-sampling operation in the maximum pooling layer, which reduces the number of features while retaining the most critical feature information (such as the highest amplitude).
[0044] Step three, nonlinear feature enhancement:
[0045] A feature enhancement module containing two layers of B-spline functions and three layers of fully connected layers is designed to enhance the representational ability of the extracted features. First, the first layer of fully connected layer is used to map each input feature dimension from D to hidden dimension H. Second, the feature is input into the B-spline function, and the learning of the control points and weights inside it enhances the model's adaptability to the distribution of input data. Then, the output of the first layer is used as the input of the second layer of fully connected mapping and activation. Finally, the activated features are mapped back to the original input dimension through the third layer of fully connected layer.
[0046] Step four, feature fusion:
[0047] The features extracted in steps two and three are interacted through feature fusion, combining the context information of the two features, enriching the model's expression ability, and improving the recognition ability of complex patterns.
[0048] Step five, global modeling:
[0049] The output bearing fault fusion features are used as input sequences and embedded into two layers of Encoder, so as to convert the fused fault features into vector form that the model can handle, and add position coding to the output of the embedding layer to preserve the time sequence dependence of the fault feature sequence. At the same time, after the input vector enters the encoder, it is first subjected to multi-head self-attention mechanism, which models the input sequence itself through multi-head attention mechanism, and calculates the correlation between the features of each position in the sequence and other positions.
[0050] Embodiment:
[0051] The experiment is implemented on Intel i7-11800H processor, 16GB memory, NVIDIA GeForce RTX3090 graphics card, on Python 3.9 and Pytorch 2.1.0.
[0052] First, process the dataset. Use VMD to decompose the original bearing fault data into multiple intrinsic mode functions (IMFs) with clear center frequency, control the bandwidth and orthogonality of the modes effectively by setting key parameters (mode number K, penalty coefficient), and extract the key frequency domain features in the signal. As shown in Figure 2 , taking the decomposition of the test signal into 5 IMFs as an example, the frequency domain characteristics of each mode are intuitively displayed. At the same time, use the decomposed IMFs for fault-related feature analysis and visualization to reveal the equipment operating state and potential fault patterns.
[0053] However, the key parameters in the VMD decomposition process need to rely on manual experience setting, and the decomposition results are easily affected by parameter selection. When K is set too small, the bearing fault features cannot be completely decomposed; if it is set too large, it will introduce redundant modalities or noise. To this end, the present application introduces MVO to optimize the key parameters of VMD. Specifically:
[0054] First, set specific parameters for the optimization algorithm, including population size (set to 10) and maximum number of iterations (set to 20), etc., to ensure their stability and efficiency in the optimization process. Then, call the optimization algorithm to optimize the key parameters of VMD (modal number K and penalty coefficient). Second, after optimization, use the optimal parameter combination of the algorithm to perform VMD decomposition on the training set, validation set and test set.
[0055] Finally, during the decomposition process, save the processed data and corresponding labels of each dataset for subsequent backtracking and analysis of the results under different parameter settings, so as to perform more detailed parameter tuning. At the same time, after optimization, these features are more refined than the original data, and can capture subtle changes and potential patterns in the signal, thereby improving the model's understanding of the data and accuracy.
[0056] Second step, feature extraction. After MVO optimizes VMD, a tensor data composed of multiple IMFs can be obtained, which can well reflect the frequency domain and time domain information of the bearing running state. Input the tensor data into a designed 5-layer convolutional network architecture, each layer composed of a convolutional layer and a max-pooling layer. In the convolutional layer, each convolutional kernel performs sliding window operation to identify local patterns in the signal.
[0057] At the same time, multi-layer convolution is used to gradually extract high-order features. After each convolution operation, an activation function is used to perform nonlinear transformation on the convolution output, enhancing the network's sensitivity to sharp changes in the vibration signal (impact signal), avoiding gradient vanishing, and enhancing the network's expression ability. In order to further reduce the computational complexity and prevent overfitting, a max-pooling layer is introduced after the convolutional layer. Through the down-sampling operation in the max-pooling layer, the spatial dimension of the convolutional features is compressed, reducing the number of features while retaining the most significant feature information (such as the highest amplitude).
[0058] Third step, feature enhancement. A feature enhancement module containing two B-spline functions and three fully connected layers is designed to enhance the feature representation ability. For example Figure 3As shown, first, each input feature dimension is mapped from D to a hidden dimension H using the first fully connected layer, which redistributes the local features and lays the foundation for subsequent introduction of nonlinear transformations. Second, the features are input into the B-spline function, and the learning of the internal control points and weights enhances the model's adaptability to the distribution of input data. Next, to achieve deeper nonlinear feature enhancement, the output of the first layer is used as the input for the second fully connected mapping and activation. Finally, the activated features are mapped back to the original input dimension (X1(B, L, D)) through the third fully connected layer, facilitating subsequent processing.
[0059] Fourth step, feature fusion. The extracted local fault features and nonlinearly enhanced fault features are interacted. Through feature fusion, the context information of the two features is combined, enriching the model's expression ability and improving the recognition ability of complex patterns.
[0060] As shown in Figure 4 , C and B, D are from different feature sets. Specifically, C represents nonlinearly enhanced fault features, and B, D represents local time series features extracted from IMFs. By performing a dot product operation on C and B, an attention score matrix is obtained, and by performing a Softmax normalization operation, the weight of each time step is obtained. At the same time, according to the weight, the sum of D is weighted, and the result of attention fusion is output.
[0061] Fifth step, global modeling. As shown in Figure 5 , first, the output bearing fault fusion features are input as an input sequence into two layers of Encoder, converting the fused fault features into a vector form that the model can handle. Next, position encoding is added to the output of the embedding layer, allowing the model to perceive the position information in the sequence and understand the relative or absolute position relationship between elements in the sequence. Second, after the input vector enters the encoder, it first passes through the multi-head self-attention mechanism, which models the input sequence itself through the multi-head attention mechanism and calculates the correlation between the features of each position and other positions. Finally, the output of the previous level (multi-head self-attention mechanism) is processed through a series of processes such as residual connection, layer normalization, and feedforward neural network. Residual connection helps prevent the gradient vanishing problem in deep network training and enhances feature representation ability; layer normalization is used to stabilize the training process and improve the convergence speed of the model. The feedforward neural network performs nonlinear transformation on the input features, further extracting and fusing the features.
[0062] Sixth step, experimental verification. Verification was performed on the public datasets of Case Western Reserve University (CWRU) and Purdue University (PU). As shown in Figure 6 and Figure 7 , this method outperforms other advanced methods in the fault diagnosis task and has better generalization and robustness.
Claims
1. An intelligent fault diagnosis method for rolling bearings based on MVO-VMD and CKCT-Net, characterized in that The method comprises the following steps: Step 1: Fault feature analysis; Step 2: Fault feature extraction: Step 2: Input the tensor data obtained in step 1 into a designed 5-layer convolutional network architecture, where each layer consists of a convolutional layer and a maximum pooling layer. Step 2: In the convolutional layer, each convolution kernel performs a sliding window operation to identify local patterns of the signal; Step 2 and 3: After each convolution operation, use the activation function to perform nonlinear transformation on the convolution output to enhance the network's sensitivity to sudden changes in the vibration signal; Step 24: Introduce a maximum pooling layer after the convolution layer. The spatial dimension of the convolution feature is compressed through the downsampling operation in the maximum pooling layer, reducing the number of features while retaining the most critical feature information. Step 3: Nonlinear fault feature enhancement: Design a feature enhancement module consisting of two layers of B-spline functions and three layers of fully connected layers to enhance the extracted feature representation capabilities; Step 4: Fault feature fusion: The local fault features extracted in step 2 and the nonlinear enhanced fault features extracted in step 3 are interacted to obtain the bearing fault fusion features; Step 5: Global modeling: Step 5.1: Take the bearing fault fusion features as the input sequence and embed them into the two-layer encoder to convert the fused fault features into a vector form that the model can process. Step 52: Add position encoding to the output of the embedding layer to preserve the temporal dependency of the fault feature sequence; Step 5.3: After the input vector enters the encoder, the input sequence itself is modeled through the multi-head attention mechanism, and the correlation between the features of each position in the sequence and other positions is calculated.
2. The rolling bearing intelligent fault diagnosis method based on MVO-VMD and CKCT-Net according to claim 1 is characterized in that The specific steps of step one are as follows: Step 1: Use variational mode decomposition (VMD) to decompose the original bearing fault data into five intrinsic mode functions (IMFs) with clear center frequencies. Step 1 and 2: Manually set key parameters to control the bandwidth and orthogonality of the mode and extract the key frequency domain features in the signal; Step 1 and 3: Introduce the multiverse optimization algorithm MVO to automatically optimize the key parameters of VMD and obtain tensor data composed of multiple IMFs.
3. The rolling bearing intelligent fault diagnosis method based on MVO-VMD and CKCT-Net according to claim 1 is characterized in that The key parameters are the modal number K and the penalty coefficient.
4. The rolling bearing intelligent fault diagnosis method based on MVO-VMD and CKCT-Net according to claim 1 is characterized in that The specific steps of step three are as follows: Step 3.
1. Use the first fully connected layer to map each input feature dimension from D to the hidden dimension H and redistribute the local features. Step 32: Input the features output by the first fully connected layer into the first layer of B-spline function; Step 3. Use the output of the first layer of B-spline function as the input of the second layer of fully connected layer mapping and activation; Step 3 and 4: Input the features output by the second fully connected layer into the second layer B-spline function; Step 3.5: Pass the features activated by the second layer of B-spline function through the third fully connected layer, map the hidden dimension back to the original input dimension, and obtain the nonlinear enhanced fault features.