Intelligent bearing fault diagnosis method based on wild horse optimized VMD-CNNKAN
By optimizing the VMD-CNNKAN method using wild horses, the problems of parameter dependence on human experience and noise influence in bearing fault diagnosis are solved, achieving efficient and robust bearing fault diagnosis and improving diagnostic accuracy and model adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-02
- Publication Date
- 2026-03-13
AI Technical Summary
Existing bearing fault diagnosis technologies suffer from technical defects such as the dependence of variational mode decomposition parameters on human experience leading to unstable decomposition quality, insufficient fusion of time and frequency features, performance degradation of deep models in strong noise environments, and difficulty in balancing model complexity and diagnostic accuracy.
A bearing intelligent fault diagnosis method based on Mustang-optimized VMD-CNNKAN is constructed. The VMD parameters are optimized by Mustang optimization algorithm with minimum permutation entropy as the fitness function. Combined with time-frequency complementary feature fusion and channel attention noise suppression, Kolmogorov-Arnold Network is used for dynamic nonlinear mapping to form a cascaded diagnosis framework.
It significantly improves the efficiency and quality of parameter optimization, enhances diagnostic accuracy in extreme noise environments, strengthens feature representation capabilities and model robustness, and achieves lightweight, high-precision fault diagnosis.
Smart Images
Figure CN121659008A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of deep learning and intelligent fault diagnosis technology, and in particular to an intelligent bearing fault diagnosis method based on Mustang optimized VMD-CNNKAN. Background Technology
[0002] As core equipment in modern industry, the health of rolling bearings, a key component of rotating machinery, directly determines production safety and economic efficiency. In actual industrial scenarios, complex background noise, variable operating conditions, and nonlinear attenuation of fault feature transmission paths pose significant challenges to traditional fault diagnosis methods, such as difficulties in feature separation and insufficient diagnostic robustness. With the rise of deep learning research, deep learning models, represented by Convolutional Neural Networks (CNNs) and Transformers, have shown advantages in feature self-learning compared to shallow machine learning methods such as Support Vector Machines. However, existing fault diagnosis methods still have certain limitations: in nonlinear and non-stationary vibration signals under actual working conditions, fault features are easily masked. Although VMD can separate fault modes, its parameters depend on human experience, resulting in unstable decomposition quality. Peng et al. introduced particle swarm optimization into VMD parameters, but it is prone to premature convergence. Although the multi-scale CNN proposed by Luo et al. expands the breadth of feature extraction, it leads to an exponential increase in computational complexity. Haixin Lv et al. compared traditional convolutional neural networks and found that attention-based models can better capture the correlation between features. Chen introduced VKCNN into the attention mechanism to achieve channel-level weight allocation, providing a new approach to noise suppression, but the recognition accuracy still needs to be improved. Summary of the Invention
[0003] The technical problem to be solved by this invention is that existing bearing fault diagnosis technology has technical defects such as the dependence of variational mode decomposition parameters on human experience leading to unstable decomposition quality, insufficient fusion of time and frequency features, performance degradation of deep models in strong noise environment, and difficulty in balancing model complexity and diagnostic accuracy.
[0004] To effectively solve the aforementioned technical problems, this invention provides a bearing intelligent fault diagnosis method based on Mustang-optimized VMD-CNNKAN. Its core innovation lies in constructing a cascaded diagnostic framework of "parameter adaptive optimization—time-frequency complementary feature fusion—channel attention noise suppression—KAN dynamic nonlinear mapping." The method includes the following technical steps:
[0005] Step 1: Signal Acquisition and Preprocessing: The vibration signal of the rolling bearing is acquired by an accelerometer, and then normalized, framed with a 50% overlap rate, and labeled to construct a standardized dataset.
[0006] Step 2, VMD Parameter Adaptive Optimization: Using the minimum permutation entropy as the fitness function, the Wild Horse Optimization Algorithm (WHO) is used to dynamically search for the optimal parameter combination {K, α} in the variational mode decomposition. The WHO algorithm simulates the three-level social structure of a Wild Horse population, achieving local fine-grained search through grazing behavior and global collaborative exploration through mating behavior, effectively avoiding premature convergence.
[0007] Step 3: Construction of Time-Frequency Complementary Feature Matrix: Based on the optimized parameters, the signal is decomposed by VMD to obtain K intrinsic mode functions (IMFs), and the original signal is simultaneously subjected to Fast Fourier Transform (FFT) to extract the frequency domain amplitude spectrum; the K-dimensional time domain feature vector and the 1-dimensional frequency domain feature vector are stacked and fused in the vertical direction to form a (K+1)-dimensional time-frequency complementary feature tensor.
[0008] In a preferred embodiment of the present invention, the time-domain feature extraction includes: extracting four types of statistical features—peak value, root mean square value, kurtosis index, and energy distribution—from K IMFs to form a K-dimensional time-domain feature vector; performing FFT transformation on the original signal synchronously to extract frequency domain amplitude spectrum features and constructing a 1-dimensional frequency domain feature vector; and stacking the K-dimensional time-domain vector and the 1-dimensional frequency-domain vector vertically to form a (K+1)×N two-dimensional time-frequency complementary feature matrix to achieve cross-domain feature collaborative representation.
[0009] Step 4: CNNKAN-CAM Hybrid Network Training and Classification: The feature matrix is input into the CNNKAN-CAM hybrid network architecture. The front end uses a multi-layer one-dimensional convolutional-pooling structure to extract local fault features layer by layer, generating multi-channel feature maps. The middle section embeds a channel attention mechanism (CAM), which adaptively learns the weight coefficients of each channel through fully connected layers and Sigmoid activation functions to achieve noise band suppression and fault-sensitive channel enhancement. The end introduces Kolmogorov-Arnold Network (KAN) to replace the traditional MLP, using a combination of learnable B-spline basis functions and SiLU activation functions to perform dynamic nonlinear mapping on the aggregated features. Finally, the softmax output layer achieves end-to-end fault classification and recognition.
[0010] In a preferred embodiment of the present invention, the CNNKAN-CAM network structure is configured as follows: the input layer receives a (K+1)×1024-dimensional composite feature matrix; the number of convolutional kernels in the first to fourth convolutional layers are 16, 32, 64, and 128 respectively, all with a size of 3×1 and a stride of 1; each convolutional layer is followed by a ReLU activation function and a 2×1 max pooling layer; the CAM module is embedded after the fourth convolutional layer; finally, it is compressed into a one-dimensional feature vector by an adaptive average pooling layer and input into the KAN layer to complete the classification.
[0011] The beneficial effects of this invention are as follows:
[0012] 1. Significantly improved efficiency and quality of parameter optimization: The WHO algorithm improves the convergence speed of VMD parameter search by 78.8% compared with the particle swarm optimization algorithm. The obtained optimal parameters minimize the model complexity and avoid noise interference introduced by over-decomposition.
[0013] 2. Strong robustness in extreme noise environments: Under strong noise with a signal-to-noise ratio of -10dB, the diagnostic accuracy of the CWRU dataset remains at 82.29%, and that of the JNU dataset remains at 73.76%, which is an average improvement of 21.3 percentage points compared with traditional CNN models; the accuracy at a signal-to-noise ratio of 0dB reaches 94.66% and 88.92%, respectively.
[0014] 3. Improved feature representation and interpretability: The time-frequency complementary feature fusion strategy fully preserves the time-domain transient features and frequency-domain harmonic structure of the fault impact; the CAM module automatically suppresses noise channel weights and enhances fault channel weights; the learnable activation function of the KAN network approximates the fault boundary through spline fitting. Attached Figure Description
[0015] Figure 1 This is a schematic diagram of the overall process of intelligent bearing fault diagnosis according to the present invention;
[0016] Figure 2 This is a schematic diagram of the process of optimizing variational mode decomposition parameters using the Mustang optimization algorithm of the present invention;
[0017] Figure 3 The WHO algorithm of this invention optimizes the VMD parameters to reconstruct the bearing vibration signal decomposition effect;
[0018] Figure 4 This is a diagram of the network structure of the VMD-CNNKAN model optimized by Wild Horse in this invention;
[0019] Figure 5 This is the confusion matrix of the method proposed in this invention under different datasets;
[0020] Figure 6 This is a feature visualization diagram of the different modules of the CNNKAN-CAM network in Embodiment 2 of the present invention, under the dataset.
[0021] Figure 7 This is a comparison chart of the accuracy of the present invention under different noise levels; Detailed Implementation
[0022] To make the technical solution, structural features, and achieved technical effects of the present invention clearer,
[0023] The technical solution of the present invention will now be described and illustrated in detail with reference to the accompanying drawings and preferred embodiments. Those skilled in the art should understand that the following embodiments are for illustrative purposes only and do not constitute any limitation on the scope of protection of the present invention.
[0024] Example 1: Validation based on the Case Western Reserve University dataset
[0025] This embodiment uses the publicly available bearing dataset from Case Western Reserve University (CWRU) to illustrate in detail the adaptive tuning process and technical effects of the Mustang optimization algorithm on VMD parameters.
[0026] Step 1.1: Dataset Construction and Signal Preprocessing
[0027] Under four load conditions (0hp, 1hp, 2hp, and 3hp) with a sampling frequency of 12kHz, the original vibration signals of the CWRU dataset were normalized. The frames were divided with a 50% overlap rate, and 1024 continuous sampling points were extracted from each frame. The training set, validation set, and test set were divided in a 7:2:1 ratio.
[0028] Step 1.2: WHO Algorithm Parameter Initialization and Population Generation
[0029] Initialize the Mustang optimization algorithm parameters: population size N=30, maximum number of iterations T=20, crossbreeding rate PC=0.13, stallion ratio PS=0.2. Define the search space for the parameters to be optimized: number of modes K∈{4,5,6,7,8}, penalty factor α∈[100,2000] integer field. Randomly generate the initial population, with each individual representing a candidate parameter combination {K, α}.
[0030] Step 1.3: Fitness Function Calculation and Iterative Optimization. Perform VMD decomposition on each parameter combination to transform the original signal... Decomposed into K eigenmode functions Calculate the permutation entropy value of each IMF component, select the minimum permutation entropy as the fitness function value, and aim to minimize this value.
[0031] Generate a uniformly distributed random number Rand in the range [0,1]. If Rand > PC, then grazing behavior is executed, and the mares perform a local fine-grained search around the optimal stallion position; otherwise, mating behavior is executed, and foals generate new individuals through crossover and mutation. The leader (optimal solution) is updated in each iteration. The search terminates when the number of iterations reaches T or the fitness value shows no improvement for 5 consecutive iterations. The specific steps of the WHO algorithm to optimize VMD parameters are as follows: Figure 2 As shown.
[0032] Step 1.4: Optimal Parameter Output. This embodiment ultimately obtains the globally optimal parameter combination: K=4, α=1568, corresponding to a minimum permutation entropy of -0.6289, with an algorithm execution time of 360 seconds. Compared to the Particle Swarm Optimization (PSO) and Ocean Predator Algorithm (MPA), the WHO algorithm is significantly superior in three aspects: optimization performance (fitness value -0.6289 vs -0.5863), computational efficiency (time 360.56s vs 4995.51s), and parameter quality.
[0033] Based on the optimal parameters obtained in Example 1, VMD decomposition is performed on each sample signal to obtain four IMF components. Four types of statistical characteristics were extracted for each IMF: peak index Root mean square value kurtosis index and energy distribution This forms a 4×1 dimensional time-domain feature vector; the reconstructed bearing vibration signal decomposition effect is as follows: Figure 3 As shown in (a). The original vibration signal Perform a Fast Fourier Transform to obtain the spectrum. The time-domain and frequency-domain feature vectors are stacked vertically to form a composite feature input matrix. This feature matrix is then normalized using Z-scores to accelerate network convergence.
[0034] Step 1.5: Construct the CNNKAN-CAM hybrid network. The parameter configurations for each layer are shown in Table 1.
[0035] Table 1 Detailed structural parameters of CNNKAN-CAM network
[0036] Network layer Parameter settings Output feature size Input Samples, 5 channels, 1024 5×1024 Conv1 5*3*1 kernel, step=1 5×1024 ReLU - 5×1024 Conv2 32*3*1 kernel, step=1 32×1024 ReLU - 32×1024 MaxPool1 2*1 kernel, step=2 32×512 Conv3 64*3*1 kernel, step=1 64×512 ReLU - 64×512 MaxPool2 2*1 kernel, step=2 64×256 Conv4 128*3*1 kernel, step=1 128×256 ReLU - 128×256 Attention - 128×256 MaxPool3 2*1 kernel, step=2 128×128 AdaptiveAvgPool - 128×1 KANLayer - 10
[0037] Input the feature matrix Figure 4 In the CNNKAN network architecture, after training, the network is evaluated on the CWRU test set.
[0038] Example 2: JNU Dataset Validation under Variable Speed Conditions
[0039] This embodiment uses the bearing dataset from Jiangnan University (JNU) to verify the model's generalization ability under variable speed conditions.
[0040] Step 2.1: Independent Validation Experiments under Variable Speed Conditions. To evaluate the performance stability of the model under different speed conditions, independent experiments were conducted at three speeds: 600 r / min, 800 r / min, and 1000 r / min. The training set, validation set, and test set were divided in a 7:2:1 ratio for each speed. The same network configuration as in Example 1 was used to obtain independent diagnostic models for each speed.
[0041] Step 2.2: VMD Parameter Re-optimization and Feature Construction. WHO optimization was re-executed on the JNU dataset to obtain the optimal parameters K=4 and α=1423. Feature construction retained the complete 1024-point time-domain waveform of the four IMFs, and the reconstructed bearing vibration signal decomposition effect is as follows: Figure 3 As shown in (b), the original signal FFT amplitude spectrum is then stacked vertically to form a 5×1024 two-dimensional matrix, preserving the original fault impact characteristics to the greatest extent. The network structure is the same as in Example 1.
[0042] Step 2.3: Visual Analysis. (By...) Figure 6 The t-SNE dimensionality reduction visualization shows that: 1) in the original data, the features of normal signals and fault signals overlap significantly, making it difficult to distinguish fault types; 2) after the Conv1 operation, the features of normal signals begin to cluster, and the features of different categories are initially separated, but some overlap still exists; 3) when reaching Conv4, the features of different categories of faults are further separated, but a small number of features still overlap; 4) the CAM module enhances the clustering of features, making the feature clusters of various signals more compact; 5) the final max pooling layer fine-tunes the clustering of features, and various signals form relatively independent feature clusters; 6) the KAN layer finally clearly distinguishes various features, forming clear classification boundaries.
[0043] In Example 1, the model of this invention achieved an average diagnostic accuracy of 99.95% on the CWRU dataset, with a standard deviation of 0.05%. The confusion matrix is as follows. Figure 5 (a) shows that all 10 types of faults were correctly identified 100%; In Example 2, the cross-speed performance comparison and analysis of the CNNKAN-CAM model proposed in this invention showed accuracies of 95.78%, 94.54%, and 96.63% at 600 r / min, 800 r / min, and 1000 r / min, respectively, with confusion matrices corresponding to... Figure 5 (b) Figure 5 (c) Figure 5 (d) The average accuracy across speeds is 95.65%, the standard deviation is only 0.86%, and the accuracy fluctuation range is 1.03%, indicating that it has strong adaptability to speed disturbances.
[0044] Example 3: Noise Resistance Performance Verification Experiment
[0045] To test the noise resistance of the proposed method, Gaussian white noise with different signal-to-noise ratios was added to two instance datasets to evaluate the model's noise resistance.
[0046] based on Figure 7The comparison of noise resistance performance on the CWRU and JNU datasets shows that the proposed CNNKAN-CAM model exhibits excellent fault diagnosis robustness under strong noise interference. At an extreme SNR of -10dB, the model maintains accuracies of 82.29% and 73.76% on the CWRU and JNU datasets, respectively. When the SNR drops to -6dB, it still achieves accuracies of 90.75% and 80.95%, respectively. Under the common 0dB threshold in engineering environments, CNNKAN-CAM achieves accuracies of 94.66% and 88.92% on both datasets, respectively. This fully validates its feature decoupling capability.
[0047] In summary, this invention addresses the experience-dependent problem of VMD parameters through the WHO algorithm, achieves adaptive noise suppression through the CAM mechanism, and enhances the ability to map complex faults through the KAN network. Ultimately, it forms a complete intelligent fault diagnosis solution for bearings that is highly accurate, robust, and lightweight, providing reliable technical support for predictive maintenance of industrial equipment.
[0048] Explanation of Key Parameter Selection Criteria
[0049] The preferred ranges of key parameters in this invention are all determined based on extensive comparative experimental analysis:
[0050] 1. WHO algorithm parameters: Population size of 30 is the optimal value for balancing computational efficiency and population diversity; number of iterations of 20 can guarantee convergence to the global optimum in 95% of cases; hybridization rate of 0.13 is determined based on Holland's pattern theorem to ensure the preservation of excellent patterns and appropriate exploration.
[0051] 2. VMD search space: The number of modes K∈[4,8] covers the number of typical modes of bearing failure; the penalty factor α∈[100,2000] balances the bandwidth constraint and reconstruction accuracy. Too small an α leads to mode overlap, and too large an α causes the frequency band to be too narrow and features to be lost.
[0052] Based on the above-described preferred embodiments of the present invention, all other embodiments obtained by those skilled in the art and related fields without inventive effort should fall within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in the present invention are implemented according to conventional means in the art unless otherwise specified or limited.
Claims
1. A bearing intelligent fault diagnosis method based on Mustang optimized VMD-CNNKAN, characterized in that, Includes the following steps: Step 1: Collect bearing fault vibration signals and preprocess them. Use the Mustang optimization algorithm to adaptively determine the key parameters of variational mode decomposition: mode number K and penalty factor α. Step 2: Perform variational mode decomposition on the signal based on the optimized parameters, and simultaneously perform fast Fourier transform on the original vibration signal to stack the time-domain features and frequency-domain spectral features to construct a time-frequency complementary feature matrix. Step 3: Input the temporal feature tensor into the CNNKAN-CAM hybrid network architecture, extract hierarchical fault features through multi-layer convolution and pooling operations, and generate multi-channel feature maps; Step 4: Embed a channel attention mechanism at the end of the network to adaptively weight the key fault feature channels, suppress noise interference and enhance feature representation; We employ the learnable activation function of the Kolmogorov-Arnold Network to achieve end-to-end fault classification and identification.
2. The bearing intelligent fault diagnosis method based on Mustang optimized VMD-CNNKAN according to claim 1, characterized in that, Step 1 specifically includes: Using minimum permutation entropy as the fitness function, a dynamic search mechanism is established by simulating the social behavior of a wild horse population to obtain the optimal parameter combination {K,α} with minimum permutation entropy for a single modal component.
3. The bearing intelligent fault diagnosis method based on Mustang optimized VMD-CNNKAN according to claim 1, characterized in that, Step 2 specifically includes: Step 21: Extract time-domain statistical features from the intrinsic mode functions to form a time-domain feature vector; Step 22: Perform a fast Fourier transform on the original vibration signal to obtain the spectral amplitude characteristics; Step 23: Stack and fuse the time-domain feature vector and the frequency-domain feature vector in the vertical direction to construct an interpretable time-series feature tensor.
4. The bearing intelligent fault diagnosis method based on Mustang optimized VMD-CNNKAN according to claim 1, characterized in that, Step 3 further includes: The input layer initially receives the composite feature matrix and performs feature extraction using layers of Conv modules. Each convolutional layer is followed by a ReLU activation function, and a Maxpool module performs dimensionality reduction to optimize feature accuracy and extraction efficiency. Finally, the output feature map is passed through the fourth convolutional layer and ReLU activation, providing multi-scale fault feature input to the channel attention module.
5. The method according to claim 4, characterized in that, The channel attention mechanism includes: After the output of the fourth convolutional layer, a channel attention module is embedded. The adaptive weight coefficients of each channel are learned through a fully connected layer and a sigmoid activation. The weight coefficients are multiplied with the original feature map channel by channel to suppress noise bands and enhance fault-sensitive channels, and the weighted feature map is output.
6. The intelligent bearing fault diagnosis method according to claim 1, characterized in that, The nonlinear mapping and classification of the KAN network in step 4 includes: The weighted feature map output by the channel attention module is input into the third max pooling layer for dimensionality reduction, and then compressed into a one-dimensional feature vector through the adaptive average pooling layer. The one-dimensional feature vector is then dynamically and nonlinearly mapped by the activation function of the B-spline basis function and the SiLU basis function of the KAN network. Finally, the fault classification and recognition are realized through the Softmax output layer. The SiLU basis function formula is:
7. The formula for the B-spline function is:
8. are basis functions Sum of spline functions:
9. Among them For B-spline basis functions, Here, w represents the learnable control point coefficients, and w represents the weight coefficients. The function shape is dynamically adjusted through backpropagation to enhance the approximation capability of complex fault impact characteristics.
10. The intelligent bearing fault diagnosis method according to claim 1, characterized in that, The signal preprocessing includes framing with a 50% overlap rate, normalization processing, and data tag creation.