Bearing fault diagnosis method based on optimization of mvmd and improved swintransformer migration model

Abstract

This invention discloses a bearing fault diagnosis method based on optimized MVMD and an improved Swing Transformer transfer model. The method improves the electric eel foraging optimization algorithm using a lens inversion strategy, adaptively determining the optimal decomposition parameters of MVMD using the improved electric eel foraging optimization algorithm. Based on the optimized MVMD, the bearing vibration signal is decomposed, a new health index is constructed to screen the intrinsic mode functions of the decomposed signal, and the signal is reconstructed. The reconstructed signal is then converted into a two-dimensional time-frequency image using continuous wavelet transform. A hybrid network incorporating deformable convolutional modules into the Swing Transformer model is constructed. The hybrid network is trained using the input training set images, and a pre-trained weight transfer learning strategy is used to improve the model's convergence performance, forming an improved Swing Transformer transfer model. Finally, the test set images are input into the improved Swing Transformer transfer model for bearing fault identification and classification. This invention's method achieves higher accuracy in bearing fault diagnosis under noisy environments.

Description

Technical Field

[0001] This invention relates to a bearing fault diagnosis method based on optimized MVMD and improved Swing Transformer transfer model, used for bearing fault feature extraction and classification under multiple faults, and belongs to the field of signal processing technology. Background Technology

[0002] Bearings, as primary power transmission components, endure tremendous external forces during operation. Prolonged use can lead to fatigue failure, potentially causing equipment malfunctions or even complete scrapping, resulting in significant economic losses. Therefore, research on efficient and accurate bearing fault diagnosis is of significant engineering value and practical importance. Currently, acquiring vibration signals from equipment is a crucial foundation for intelligent fault diagnosis. Compared to traditional methods such as empirical mode decomposition, variational mode decomposition (VMD) can effectively separate signal characteristics across different frequency bands, significantly mitigating mode aliasing problems, and eliminating the need for manually adding white noise or complex parameter selection. However, under actual operating conditions, rolling bearings also bear significant axial loads, making it difficult for single-channel VMD to comprehensively characterize multidimensional fault information. Therefore, multivariate variational mode decomposition (MMD) has been proposed and widely used for the collaborative analysis of multi-channel vibration signals. However, the decomposition effect of MMD is highly dependent on its mode number K and penalty factor. The choice of optimization algorithm is crucial. To more effectively optimize the key parameters of multivariate variational mode decomposition, a smart optimization algorithm with fast convergence speed, strong global search capability, and high accuracy needs to be designed to improve the adaptability and robustness of the decomposition results.

[0003] After successfully extracting the main fault features from the signal, it is still necessary to build an efficient recognition model to achieve intelligent diagnosis. With the development of deep learning technology, fault diagnosis is gradually moving away from simple reliance on expert experience and towards data-driven and intelligent approaches. Swin Transformer, as one of the most advanced convolutional neural network models, has achieved excellent performance in fields such as fault diagnosis thanks to its hierarchical attention mechanism and shifting window design. However, during image block segmentation, Swin Transformer divides the input image into several non-overlapping local windows. While this reduces computational complexity, it also loses some pixel-level spatial structural information. For weak local features present in the image generated after the fault signal undergoes two-dimensional transformation, this loss of spatial information may weaken the model's ability to distinguish early or weak faults. Therefore, how to retain the efficient computational advantages of Swin Transformer while compensating for the loss of spatial information has become a key research direction for improving diagnostic performance. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a bearing fault diagnosis method based on optimized MVMD and improved SwingTransformer transfer model, which can solve the problem that weak features of fault signals in noisy environments cannot be effectively extracted, resulting in low model recognition accuracy.

[0005] To achieve the above objectives, this invention improves the Electric Eel Foraging Optimization (EEFO) algorithm using a lens inversion strategy and optimizes the Multivariate Variational Mode Decomposition (MVMD) algorithm using an improved Electric Eel Foraging Optimization (IEEFO) algorithm, achieving efficient signal decomposition. The signal is reconstructed by constructing new health indicators to enhance the saliency of fault features. The reconstructed signal is converted into a two-dimensional image using Continuous Wavelet Transform (CWT), preserving the signal's time-frequency characteristics. An improved Swin Transformer transfer model combining Linear Deformable Convolution (LDConv) is designed to further enhance feature extraction capabilities and classification performance, achieving accurate identification and classification of bearing faults. This method includes the following steps:

[0006] (1) Collect bearing fault diagnosis signals under different conditions and preprocess them;

[0007] (2) The improved electric eel foraging optimization algorithm (IEEFO) is used to optimize the multivariate variational mode decomposition (MVMD) parameters, and the optimized MVMD is used to achieve adaptive decomposition of vibration signals in different states;

[0008] (3) Construct a new health indicator, screen the intrinsic mode functions (IMFs) after signal decomposition, and reconstruct the signal using the screened IMFs;

[0009] (4) The reconstructed signal is converted into a two-dimensional time-frequency image through continuous wavelet transform (CWT), and the time-frequency image is divided into a training set and a test set, with the ratio of the training set to the test set set set to 7:3.

[0010] (5) Construct a hybrid network that incorporates a linear deformable convolution (LDConv) module into the Swin Transformer model, train the hybrid network using training set images, and use a pre-trained weight transfer learning strategy to improve the model's convergence performance, thus forming an improved Swin Transformer transfer model.

[0011] (6) Input the test set images into the improved Swing Transformer transfer model to realize the identification and classification of bearing faults.

[0012] The improved electric eel foraging optimization algorithm (IEEFO) in step (2) uses a lens inversion strategy to update the solution results of the electric eel foraging optimization algorithm (EEFO). The specific steps are as follows:

[0013] Given the normalized coordinates Z of a random reference point:

[0014]

[0015] Where x rn,d Let lb represent the value of the rn-th individual in the group along dimension d. d and ub d Let represent the lower bound and upper bound of the d-th dimension, respectively, and let lb and ub represent the lower bound and upper bound of all dimensions, respectively.

[0016] Perform lens reversal operation:

[0017]

[0018] Where x* is the lens reflection solution, X prey This represents the current globally optimal solution, and Lef is the lens effect control factor.

[0019] The updated solution is , where x i This represents the current position of the individual, and w is a random number that follows a standard normal distribution.

[0020] In step (2), IEEFO optimizes the MVMD parameters, specifically the number of modes K and the penalty factor in MVMD. Two parameters.

[0021] In step (2), the optimized MVMD is used to achieve adaptive decomposition of vibration signals in different states. The specific decomposition steps are as follows:

[0022] Let the acquired multivariable signal be... Where N is the number of channels, It is the time-domain signal of the nth variable.

[0023] Constructing constrained variational problems

[0024] ,

[0025] in, It is the kth multivariable mode, ω k Let be the center frequency of the k-th mode. Used to estimate modal width.

[0026] Introducing a penalty factor and Lagrange multiplication operators Transforming a constrained variational problem into an unconstrained variational problem

[0027]

[0028] The above unconstrained variational problem is solved using the alternating direction method of multiplication operators. Through alternating update calculations, K modes are obtained. and its corresponding center frequency ω k .

[0029] In step (3), a new health indicator is constructed, the expression of which is:

[0030]

[0031] in, This represents the normalized envelope kurtosis value. This represents the normalized Pearson correlation coefficient value. This represents the normalized envelope spectrum entropy value. Represents the weighting coefficient, and .

[0032] In step (4), the reconstructed signal is converted into a two-dimensional time-frequency image using CWT. CWT scales and translates the mother wavelet function and convolves it with the reconstructed signal to obtain the time-frequency representation of the reconstructed signal at different scales, thus converting the reconstructed signal into a two-dimensional time-frequency image.

[0033]

[0034] Where s is the scaling factor. This represents the translation factor.

[0035] In step (5), the Swin Transformer model incorporates a Linear Deformable Convolution (LDConv) module. The LDConv module enhances local features of the feature image, thereby improving the performance of the Swin Transformer model. The LDConv module dynamically adjusts the sampling point positions based on the input feature image, adjusting the sampling shape at each position by introducing offsets. This allows the sampling shape to dynamically change with the task, enhancing local image features. Its mathematical expression is:

[0036]

[0037] Where p represents the spatial location on the output feature map, N is the number of convolutional kernels, and w n The weights of the nth convolutional kernel, p n It is the initial sampling position, △p n It is the learned offset, and x(·) represents bilinear interpolation sampling.

[0038] The pre-trained weight transfer learning strategy in step (5) specifically increases the weight of samples that improve label prediction and classifier performance in the target domain, while decreasing the weight of samples that hinder label prediction and classifier performance, thereby optimizing model performance.

[0039] The beneficial effects of this invention are as follows: This invention proposes a bearing fault diagnosis method based on optimized MVMD and an improved Swing Transformer transfer learning model. The method improves the electric eel foraging optimization algorithm through a lens inversion strategy, and adaptively determines the optimal decomposition parameters of MVMD using the improved electric eel foraging optimization algorithm, effectively solving the mode aliasing problem and improving the accuracy of signal decomposition. New health indicators are constructed based on the decomposed signal for signal reconstruction, enhancing the saliency of weak fault features and providing high-quality input data for subsequent analysis. Continuous wavelet transform is used to convert the reconstructed signal into a two-dimensional time-frequency image, fully preserving the time-frequency domain characteristics of the signal and avoiding the feature loss problem in traditional methods. An improved Swing Transformer model combining linear deformable convolution is designed. By introducing linear deformable convolution, the model's ability to capture local features is enhanced. The global self-attention mechanism of Swing Transformer is integrated, achieving an effective combination of local details and global information. Simultaneously, a transfer learning strategy using pre-trained weights further improves the accuracy of fault feature extraction and classification performance. Attached Figure Description

[0040] Figure 1 This is a flowchart of the bearing fault diagnosis process of the present invention;

[0041] Figure 2 This is a convergence curve of the MVMD parameter optimization iteration in this invention;

[0042] Figure 3 This invention optimizes the time-domain plot of the inner race fault bearing signal decomposed by MVMD;

[0043] Figure 4 This invention optimizes the frequency domain diagram of the inner race fault bearing signal after MVMD decomposition.

[0044] Figure 5 This is a two-dimensional diagram of the bearing vibration signal after continuous wavelet transform processing according to the present invention.

[0045] Figure 6 This is the bearing fault diagnosis confusion matrix diagram of the present invention;

[0046] Figure 7 This is a comparison diagram of the diagnostic method of the present invention with other methods. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0048] Example:

[0049] like Figure 1 As shown, this embodiment of a bearing fault diagnosis method based on optimized MVMD and improved Swing Transformer transfer model includes the following steps:

[0050] (1) Collect bearing fault diagnosis signals under different conditions and preprocess them, including vibration signals under four conditions: normal bearing, inner ring fault, outer ring fault and rolling element fault. The data used is acceleration signal.

[0051] (2) The improved electric eel foraging optimization algorithm (IEEFO) is used to optimize the multivariate variational mode decomposition (MVMD) parameters, and the optimized MVMD is used to achieve adaptive decomposition of vibration signals in different states.

[0052] The improved electric eel foraging optimization algorithm (IEEFO) in step (2) uses a lens inversion strategy to update the solution results of the electric eel foraging optimization algorithm (EEFO). The specific steps are as follows:

[0053] Given the normalized coordinates Z of a random reference point:

[0054]

[0055] Where x rn,d Let lb represent the value of the rn-th individual in the group along dimension d. d and ub d Let represent the lower bound and upper bound of the d-th dimension, respectively, and let lb and ub represent the lower bound and upper bound of all dimensions, respectively.

[0056] Perform lens reversal operation:

[0057]

[0058] Where x* is the lens reflection solution, X prey This represents the current globally optimal solution, and Lef is the lens effect control factor.

[0059] The updated solution is , where x i This represents the current position of the individual, and w is a random number that follows a standard normal distribution.

[0060] In step (2), the minimum average envelope entropy (MAEE) is used as the fitness function during the IEEFO optimization of MVMD parameters. The expression for MAEE is:

[0061]

[0062] in Let N represent the envelope signal of the i-th intrinsic mode function (IMF), and let N represent the signal length.

[0063] After each iteration, the current MAEE value is calculated and compared with the previous result. This process continues until the maximum number of iterations is reached, at which point the optimal parameters K and are returned. Taking a bearing with a faulty inner ring diameter of 0.007 inches as an example, the convergence curve of the MVMD parameter optimization iteration is as follows: Figure 2 As shown in the figure, the fitness value decreases with the number of iterations. When using the IEEFO optimization algorithm, the optimal parameters can be determined after the second iteration, while the traditional electric eel foraging optimization algorithm (EEFO) requires the fourth iteration to determine the parameters. This shows that IEEFO is superior to EEFO.

[0064] In step (2), the optimized MVMD is used to achieve adaptive decomposition of vibration signals in different states. The specific decomposition steps are as follows:

[0065] Let the acquired multivariable signal be... Where N is the number of channels, It is the time-domain signal of the nth variable.

[0066] Constructing constrained variational problems

[0067] ,

[0068] in, It is the kth multivariable mode, ω k Let be the center frequency of the k-th mode. Used to estimate modal width.

[0069] Introducing a penalty factor and Lagrange multiplication operators Transforming a constrained variational problem into an unconstrained variational problem

[0070]

[0071] The above unconstrained variational problem is solved using the alternating direction method of multiplication operators. Through alternating update calculations, K modes are obtained. and its corresponding center frequency ω k .

[0072] Taking a bearing with a faulty inner ring diameter of 0.007 inches as an example, three IMFs are obtained through optimized MVMD decomposition, and their time-domain and frequency-domain waveforms are as follows: Figure 3 and Figure 4As shown in the figure, the results indicate that no mode aliasing occurred in any of the IMF components.

[0073] (3) Construct a new health indicator, screen the intrinsic mode functions (IMFs) after signal decomposition, and reconstruct the signal using the screened IMFs;

[0074] In step (3), a new health indicator is constructed, and its expression is:

[0075]

[0076] in, This represents the normalized envelope kurtosis value. This represents the normalized Pearson correlation coefficient value. This represents the normalized envelope spectrum entropy value. Represents the weighting coefficient, and .

[0077] The health indicators used in step (3) are based on the maximum weighted health index (W) in the IMF. max As a screening criterion, the IMF sets a screening threshold of 0.6W. max Taking a fault diameter of 0.014 inches as an example, Table 1 lists the calculated values ​​of health indicators, and Table 2 summarizes the IMF screening results corresponding to each fault type.

[0078] Table 1 Weighted health indicators for bearings with a failure diameter of 0.021 inches

[0079] type IMF quantity Weighted health indicators threshold Inner ring fault 4 0.17、0.63、0.83、0.87 0.522 Outer ring fault 5 0.15、0.34、0.21、0.85、0.42 0.51 Rolling element failure 3 0、0.79、0.72 0.474

[0080] Table 2. IMFs selected for each bearing type

[0081] Name Label IMF quantity Selected IMF normal N 3 3 0.007-inch inner ring fault I1 3 2、3 0.014-inch inner ring fault I2 4 2、3、4 0.021-inch inner ring fault I3 10 4、5、6、7、8 0.007-inch outer ring fault O1 3 2、3 0.014-inch outer ring fault O2 5 4 0.021-inch outer ring fault O3 3 2、3 0.007-inch rolling element failure R1 5 3、4、5 0.014-inch rolling element failure R2 3 2、3 0.021-inch rolling element failure R3 3 3

[0082] (4) The reconstructed signal is converted into a two-dimensional time-frequency image through continuous wavelet transform (CWT). The time-frequency image is divided into a training set and a test set, with the ratio of the training set to the test set set set to 7:3.

[0083] In step (4), the reconstructed signal is converted into a two-dimensional time-frequency image using CWT. CWT scales and translates the mother wavelet function and convolves it with the reconstructed signal to obtain the time-frequency representation of the reconstructed signal at different scales, thus converting the reconstructed signal into a two-dimensional time-frequency image. The expression for the mother wavelet function is:

[0084]

[0085] Where s is the scaling factor. This represents the translation factor.

[0086] Two-dimensional images obtained by applying CWT to bearing vibration signal data in four states are shown below. Figure 5 As shown.

[0087] (5) Construct a hybrid network that incorporates the linear deformable convolution (LDConv) module into the Swin Transformer model, train the hybrid network using training set images, and use a pre-trained weight transfer learning strategy to improve the model's convergence performance, thus forming an improved Swin Transformer transfer model.

[0088] In step (5), the Swin Transformer model incorporates a Linear Deformable Convolution (LDConv) module. The LDConv module enhances local features of the feature image, thereby improving the performance of the Swin Transformer model. The LDConv module dynamically adjusts the sampling point positions based on the input feature image. By introducing offsets to adjust the sampling shape at each position, the sampling shape dynamically changes with the task, enhancing local image features. Its mathematical expression is:

[0089]

[0090] Where p represents the spatial location on the output feature map, N is the number of convolutional kernels, and w n The weights of the nth convolutional kernel, p n It is the initial sampling position, △p n It is the learned offset, and x(·) represents bilinear interpolation sampling.

[0091] The pre-trained weight transfer learning strategy in step (5) specifically involves increasing the weight of samples that improve label prediction and classifier performance in the target domain, while decreasing the weight of samples that hinder label prediction and classifier performance, thereby optimizing model performance.

[0092] (6) Input the test set images into the improved Swing Transformer transfer model to realize the identification and classification of bearing faults.

[0093] In step (6), the test image is input into the trained classification model, and the resulting confusion matrix is ​​as follows: Figure 6 As shown, the overall accuracy reached 99.3%. The model achieved perfect differentiation in normal operating conditions, inner ring faults, and outer ring fault identification. In rolling element fault identification, the model made only 4 errors, including 1 R3 misclassified as R1 and 3 samples misclassified as R2.

[0094] To comprehensively evaluate model performance, recall rate is used as a key evaluation metric. Figure 7Recall data for various fault diagnosis methods under different fault types are presented, where pcc represents the Pearson correlation coefficient, ek represents the envelope kurtosis, ese represents the envelope spectral entropy, w represents the health index, regent and shufflenetV2 are two network models, and iswin-transformer represents the improved Swin-Transformer. Results show that the proposed method achieves a recall rate exceeding 93% in most fault categories, while the other six comparative methods show significantly lower recall rates in specific fault types. Although the proposed method's recall rate in the R3 fault category is slightly lower than that of the ek-iSwin-Transformer method, it consistently outperforms the other six benchmark methods in terms of overall accuracy and recall.

[0095] This invention improves the electric eel foraging optimization algorithm using a lens inversion strategy. The improved algorithm adaptively determines the optimal decomposition parameters for MVMD, effectively solving the modality mixing problem and improving the accuracy of signal decomposition. Based on the decomposed signal, a new health index is constructed for signal reconstruction, enhancing the saliency of weak fault features. This method designs an improved Swin Transformer model incorporating linear deformable convolution. By introducing linear deformable convolution, the model's ability to capture local features is enhanced. The global self-attention mechanism of the Swin Transformer is integrated, achieving an effective combination of local details and global information. Furthermore, a transfer learning strategy using pre-trained weights further improves the accuracy of fault feature extraction and classification performance.

[0096] The above description represents only preferred embodiments of the present invention and should not be considered as a limitation thereof. Obviously, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present invention. Therefore, the present invention is also intended to include such modifications and variations if they fall within the scope of the claims of the present invention and their equivalents.

Claims

1. A bearing fault diagnosis method based on optimized MVMD and improved Swing Transformer transfer model, characterized in that, Includes the following steps: (1) Collect bearing fault diagnosis signals under different conditions and preprocess them; (2) The improved electric eel foraging optimization algorithm (IEEFO) is used to optimize the multivariate variational mode decomposition (MVMD) parameters, and the optimized MVMD is used to achieve adaptive decomposition of vibration signals in different states; (3) Construct a new health indicator, screen the intrinsic mode functions (IMFs) after signal decomposition, and reconstruct the signal using the screened IMFs; (4) The reconstructed signal is converted into a two-dimensional time-frequency image through continuous wavelet transform (CWT), and the time-frequency image is divided into a training set and a test set, with the ratio of the training set to the test set set set to 7:

3. (5) Construct a hybrid network that incorporates a linear deformable convolution (LDConv) module into the Swin Transformer model, train the hybrid network using training set images, and use a pre-trained weight transfer learning strategy to improve the model's convergence performance, thus forming an improved Swin Transformer transfer model. (6) Input the test set images into the improved Swing Transformer transfer model to realize the identification and classification of bearing faults.

2. The bearing fault diagnosis method based on optimized MVMD and improved Swing Transformer transfer model according to claim 1, characterized in that: The improved electric eel foraging optimization algorithm (IEEFO) in step (2) uses a lens inversion strategy to update the solution results of the electric eel foraging optimization algorithm (EEFO). The specific steps are as follows: Given the normalized coordinates Z of a random reference point: Where x rn,d Let lb represent the value of the rn-th individual in the group along dimension d. d and ub d Let represent the lower bound and upper bound of the d-th dimension, respectively, and let lb and ub represent the lower bound and upper bound of all dimensions, respectively. Perform lens reversal operation: Where x* is the lens reflection solution, X prey This represents the current globally optimal solution, with Lef being the lens effect control factor. The updated solution is , where x i This represents the current position of the individual, and w is a random number that follows a standard normal distribution.

3. The bearing fault diagnosis method based on optimized MVMD and improved Swing Transformer transfer model according to claim 1, characterized in that: In step (2), IEEFO optimizes the MVMD parameters, specifically the number of modes K and the penalty factor in MVMD. Two parameters.

4. The bearing fault diagnosis method based on optimized MVMD and improved Swing Transformer transfer model according to claim 1, characterized in that: In step (2), the optimized MVMD is used to achieve adaptive decomposition of vibration signals in different states. The specific decomposition steps are as follows: Let the acquired multivariable signal be... Where N is the number of channels, It is the time-domain signal of the nth variable; Constructing constrained variational problems , in, It is the kth multivariable mode, ω k Let be the center frequency of the k-th mode. Used to estimate modal width; Introducing a penalty factor and Lagrange multiplication operators Transforming a constrained variational problem into an unconstrained variational problem ， The above unconstrained variational problem is solved using the alternating direction method of multiplication operators. Through alternating update calculations, K modes are obtained. and its corresponding center frequency ω k .

5. The bearing fault diagnosis method based on optimized MVMD and improved Swing Transformer transfer model according to claim 1, characterized in that: In step (3), a new health indicator is constructed, the expression of which is: in, This represents the normalized envelope kurtosis value. This represents the normalized Pearson correlation coefficient value. This represents the normalized envelope spectrum entropy value. Represents the weighting coefficient, and .

6. The bearing fault diagnosis method based on optimized MVMD and improved Swing Transformer transfer model according to claim 1, characterized in that: In step (4), the reconstructed signal is converted into a two-dimensional time-frequency image using CWT. CWT scales and translates the mother wavelet function and convolves it with the reconstructed signal to obtain the time-frequency representation of the reconstructed signal at different scales, thereby converting the reconstructed signal into a two-dimensional time-frequency image. The expression for the mother wavelet function is: Where s is the scaling factor. This represents the translation factor.

7. The bearing fault diagnosis method based on optimized MVMD and improved Swing Transformer transfer model according to claim 1, characterized in that: The Swin Transformer model in step (5) incorporates a Linear Deformable Convolution (LDConv) module. The LDConv module enhances local features of the feature image, thereby improving the performance of the Swin Transformer model. The LDConv module dynamically adjusts the sampling point positions based on the input feature image, adjusting the sampling shape at each position by introducing offsets. This allows the sampling shape to dynamically change with the task, enhancing local image features. Its mathematical expression is: Where p represents the spatial location on the output feature map, N is the number of convolutional kernels, and w n The weights of the nth convolutional kernel, p n It is the initial sampling position, △p n It is the learned offset, and x(·) represents bilinear interpolation sampling.

8. The bearing fault diagnosis method based on optimized MVMD and improved Swing Transformer transfer model according to claim 1, characterized in that: The pre-trained weight transfer learning strategy in step (5) specifically increases the weight of samples that improve label prediction and classifier performance in the target domain, while decreasing the weight of samples that hinder label prediction and classifier performance, thereby optimizing model performance.

Images