Rolling bearing fault multi-source vibration feature extraction method

Abstract

The application discloses a kind of rolling bearing fault multi-source vibration feature extraction methods, comprising the following steps: step S1, using convolutional autoencoder to original multi-source heterogeneous data is integrated feature fusion, and using data enhancement and normalization technique to original multi-source heterogeneous data is preprocessed to improve the generalization ability and accuracy of subsequent convolutional autoencoder model;Step S2, in convolutional autoencoder model reference K-means clustering technique and sparse attention mechanism to improve the recognition rate of key fault feature of model;Finally using confusion matrix and visual technology realizes fault classification, the present application realizes the efficient fusion of multi-source vibration data, accurate identification of key fault feature and accurate classification of fault type.

Description

Technical Field

[0001] This invention relates to the field of equipment condition monitoring and fault diagnosis technology, and more specifically, to a method for extracting multi-source vibration features of rolling bearing faults. Background Technology

[0002] Rolling bearings are core components of rotating machinery, and their operating condition directly affects the overall reliability and safety of the equipment. Traditional rolling bearing fault diagnosis often relies on single-channel vibration signals and simple learning algorithms, which suffers from limited information acquisition dimensions and incomplete fault characteristic characterization.

[0003] With the development of sensor technology and information fusion theory, multi-sensor information fusion diagnostic technology has gradually become a research hotspot. This technology can utilize data collected by multiple (homogeneous or heterogeneous) sensors for diagnosis, expanding the physical attributes (such as vibration, temperature, sound pressure, etc.), spatial range (such as different monitoring points), or temporal range (such as different sampling periods) of the information collected, providing more comprehensive basic information for fault diagnosis.

[0004] However, existing multi-source data processing technologies still have the following key problems:

[0005] Poor classification performance: Multi-source data exhibit heterogeneity (such as differences in amplitude and frequency characteristics of vibration signals from different channels), making it difficult for existing methods to effectively classify the data, resulting in low data utilization efficiency;

[0006] Incomplete extraction of key information: Multi-source data contains a lot of redundant information. Existing methods are not capable of extracting key features that are sensitive to faults, and are prone to missing important fault information.

[0007] High processing complexity: The massive volume of multi-source data and the lack of efficient feature fusion and dimensionality reduction mechanisms in existing methods result in complex and time-consuming data processing procedures. Summary of the Invention

[0008] In view of the problems existing in the prior art, the purpose of this invention is to provide a method for extracting multi-source vibration features of rolling bearing faults, so as to achieve efficient fusion of multi-source vibration data, accurate identification of key fault features, and accurate classification of fault types.

[0009] The present invention adopts the following technical solution:

[0010] This invention discloses a method for extracting multi-source vibration features of rolling bearing faults, comprising the following steps:

[0011] Step S1: Use a convolutional autoencoder to integrate features of the original multi-source heterogeneous data, and use data augmentation and normalization techniques to preprocess the original multi-source heterogeneous data to improve the generalization ability and accuracy of the subsequent convolutional autoencoder model.

[0012] Step S2: In the convolutional autoencoder model, K-means clustering and sparse attention mechanisms are applied to improve the model's recognition rate of key fault features; finally, confusion matrix and visualization techniques are used to achieve fault classification.

[0013] The original multi-source heterogeneous data consists of three channels of homogeneous vibration signals of the rolling bearing in the axial, horizontal and vertical directions, collected from the same monitoring point.

[0014] The convolutional autoencoder model comprises an encoder and a decoder. The encoder progressively reduces the spatial dimensionality of the input data through convolutional and pooling layers while increasing the number of feature channels to extract abstract feature representations of the data. Its mathematical expression is as follows:

[0015]

[0016] In the formula: Conv(x, W) performs a convolution operation on the input data x, W is the convolution kernel, b represents the bias term, and σ is the activation function (which can be sigmoid or ReLU).

[0017] The decoder gradually recovers the spatial dimension of the data through deconvolutional layers (or transposed convolutional layers) while reducing the number of feature channels to reconstruct the original data. Its mathematical expression is:

[0018]

[0019] In the formula, Upsample(h) represents the upsampling operation on the decoded features, W′ is the convolution kernel of the decoder, b′ is the bias term of the decoder, and σ is the activation function.

[0020] Beneficial effects

[0021] This invention integrates convolutional autoencoders, K-means clustering, and sparse attention mechanisms through architectural integration, unified optimization, and mathematical coupling. Architecturally, sparse attention gates control the feature flow between ConvBlocks, while cluster centers serve as reconstruction targets in the decoder, prototype features for classification, and regularization anchors in the encoder. Unified optimization is achieved through three loss components: reconstruction fidelity (L_CE), feature cluster alignment (L_cluster), and attention sparsity (R_sparse), which are jointly backpropagated to guide the learning process. Mathematically, cluster centroids appear in both the CAE and K-means equations, attention weights modify the Jacobian of the feature transformation, and shared basis functions in the encoder / decoder are constrained by cluster distance. This integration enables cooperative operation; sparse attention focuses feature learning on cluster-related patterns, cluster centroids guide data reconstruction and representation learning, and the cooperative adjustment of attention clusters eliminates ambiguity from multi-source signals, thereby improving the overall performance of the system. Attached Figure Description

[0022] Figure 1 This is a flowchart of the present invention.

[0023] Figure 2 This is a time-domain plot of the horizontal vibration signal of a rolling bearing under normal conditions.

[0024] Figure 3 This is a time-domain plot of the vertical vibration signal of a rolling bearing under normal conditions.

[0025] Figure 4 This is a time-domain plot of the axial vibration signal of a rolling bearing under normal conditions.

[0026] Figure 5 This is a confusion matrix diagram of diagnostic results for rolling bearings under different fault states based on embodiments of the present invention.

[0027] Figure 6 This is a clustering classification diagram of diagnostic results for rolling bearings under different fault states based on embodiments of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0029] As shown in the figure, this invention discloses a method for extracting multi-source vibration features of rolling bearing faults, comprising the following steps:

[0030] Step S1: Use a convolutional autoencoder to perform integrated feature fusion on the original multi-source heterogeneous data, and use data augmentation and normalization techniques to preprocess the original multi-source heterogeneous data to improve the generalization ability and accuracy of the subsequent convolutional autoencoder model. The purpose of preprocessing is to eliminate the dimensional differences between signals from different channels and improve the generalization ability of the subsequent model. The preprocessed three-channel data is input into the convolutional autoencoder, and abstract features are extracted through the convolutional and pooling layers of the encoder. The decoder reconstructs the data to verify the effectiveness of the features, and finally outputs a high-dimensional feature vector that integrates multi-source information.

[0031] Step S2: K-means clustering and sparse attention mechanisms are applied to the convolutional autoencoder model to improve the model's recognition rate of key fault features; finally, confusion matrix and visualization techniques are used to achieve fault classification. Specifically, this invention achieves accurate classification of rolling bearing fault types (such as normal, inner race fault, and compound fault).

[0032] The original multi-source heterogeneous data consists of three channels of homogeneous vibration signals of the rolling bearing in the axial, horizontal and vertical directions, collected from the same monitoring point.

[0033] During data acquisition, a triaxial accelerometer deployed at the same monitoring point on the rolling bearing housing is used to simultaneously acquire three channels of vibration signals: axial, horizontal radial, and vertical radial. The sampling frequency is determined according to the bearing speed (usually 5kHz-20kHz) to ensure coverage of the main frequency components of the fault characteristics.

[0034] The convolutional autoencoder model comprises two parts: an encoder and a decoder. The encoder gradually reduces the spatial dimension of the input data through convolutional layers and pooling layers, while increasing the number of feature channels to extract abstract feature representations of the data. The encoder part can be represented as follows:

[0035]

[0036] In the formula, h represents the feature being decoded. The core function of the encoder is to reduce the dimensionality of the data and extract abstract features through convolution and pooling operations. The encoder part can be implemented using convolution and pooling operations, and its mathematical expression is:

[0037]

[0038] In the formula: Conv(x, W) performs a convolution operation on the input data x, W is the convolution kernel, b represents the bias term, and σ is the activation function (which can be sigmoid or ReLU function).

[0039] The decoder gradually restores the spatial dimension of the data through deconvolutional layers (or transposed convolutional layers) while reducing the number of feature channels to reconstruct the original data. The core function of the decoder is to restore the data dimension through deconvolution and upsampling operations and verify the effectiveness of features. Its mathematical expression is:

[0040]

[0041] In the formula, Upsample(h) represents the upsampling operation on the decoded features, W′ is the convolution kernel of the decoder, b′ is the bias term of the decoder, and σ is the activation function.

[0042] The training of the convolutional autoencoder employs cross-entropy loss to evaluate the difference between the model's predicted probability distribution and the true label distribution. It assesses the model's predicted probability distribution based on the actual label distribution, quantifying the information entropy difference between them. The expression for cross-entropy is as follows:

[0043]

[0044] Where p is the probability distribution of the true labels, q is the probability distribution predicted by the model, and x represents the possible classes. This formula measures the difference between the true distribution p and the predicted distribution q by calculating the information entropy of the predicted distribution q given the true distribution p.

[0045] The convolutional autoencoder model employs K-means clustering and sparse attention mechanisms to improve the model's recognition rate of key fault features. The specific implementation of K-means is as follows:

[0046] The core objective of K-means clustering is to efficiently divide n samples into K clusters, ensuring that each sample belongs to the cluster corresponding to its nearest cluster center (i.e., the mean). This process aims to minimize the variance within each cluster, thereby optimizing the classification decision. K-means clustering promotes the formation of geometrically good clusters of data points and acts as a regularization mechanism, effectively preventing overfitting and ensuring the model's generalization ability. Its expression is as follows:

[0047]

[0048] In the formula S i Let μ be the sample set of the i-th cluster. i For S i Cluster centers, ||x-μ i || 2 Let x be the Euclidean distance from the sample x to the cluster center.

[0049] Integrating K-means clustering into a convolutional autoencoder model involves the following steps:

[0050] Clustering of extracted high-dimensional features for data structure understanding;

[0051] Clustering of dimensionality-reduced data is used to evaluate model performance and feature effectiveness;

[0052] Using K-means to classify the model output simplifies data processing.

[0053] Ensemble models position K-means as a preprocessing and visualization tool as well as a classifier, providing a new perspective for the training and evaluation of deep learning models.

[0054] K-means clustering and sparse attention mechanisms are used in the autoencoder model to improve the model's recognition rate of key fault features. The specific implementation of the sparse attention mechanism is as follows:

[0055] Assuming the input sequence is X, and the learnable weight matrix is ​​W, where the unit is the output dimension of the attention layer, the attention score is calculated using the following formula:

[0056]

[0057] In the formula, t represents the time interval of the input sequence;

[0058] Sparsity (L1 regularization) is calculated as follows:

[0059]

[0060] In the formula, λ is the regularization coefficient;

[0061] The normalized calculation is as follows:

[0062]

[0063] The final output is as follows:

[0064] .

[0065] The autoencoder model employs K-means clustering and sparse attention mechanisms to improve the model's recognition rate of key fault features. The ensemble model formula based on convolutional autoencoders, K-means, and sparse attention mechanisms is as follows:

[0066] The ensemble model is implemented by defining explicit mathematical formulas between the self-convolutional encoder, K-means clustering, and sparse attention mechanism. First, the sparse attention mechanism is embedded between the convolutional layer and the autoencoder, specifically constructed as follows:

[0067]

[0068] In the formula, ⊙ represents element-wise multiplication to create an attention-guided feature map before pooling;

[0069] Both cluster centers and features generated by the convolutional autoencoder are optimized using the following loss function:

[0070]

[0071] Cluster center μ i These are learnable parameters that are updated via backpropagation;

[0072] During the training phase, the synchronization process is as follows:

[0073]

[0074] Reconstruct the decoder using clustered conditional deconvolution:

[0075] .

[0076] This invention employs K-means clustering and sparse attention mechanisms in an autoencoder model to improve the model's recognition rate of key fault features. This addresses issues such as insufficient utilization of data sources, incomplete extraction of different data, confusion caused by lack of classification after data extraction, and insufficient research on multi-source data, ultimately achieving effective fault diagnosis of rolling bearings.

[0077] The effectiveness of this invention is demonstrated through experiments.

[0078] Experimental Preparation: Experimental Object: A certain type of deep groove ball bearing (model 6205) was selected to simulate 6 operating states: normal state (N), inner ring failure state (IF), outer ring failure state (OF), inner and outer ring combined failure state (IF-OF), inner ring rolling element combined failure state (IF-RF), and outer ring rolling element combined failure state (OF-RF).

[0079] Data acquisition: A triaxial accelerometer is placed at the same monitoring point on the bearing housing, with a sampling frequency of 10kHz. The axial, horizontal, and vertical vibration signals are collected under each condition, with 100 sets of samples collected for each condition (each set of samples contains 1000 sampling points).

[0080] Model parameter settings: The encoder of the convolutional autoencoder contains 2 convolutional layers (3×3 kernel size, number of layers 16 and 32 respectively) and 2 pooling layers (max pooling, pooling kernel 2×2); the decoder contains 2 deconvolutional layers (3×3 kernel size, number of layers 16 and 1 respectively) and 2 upsampling layers (bilinear interpolation); the activation function is ReLU; the number of clusters for K-means clustering is K=6 (corresponding to 6 fault states); the regularization coefficient of the sparse attention mechanism is λ=0.01; the model training epochs are 100, the batch size is 32, and the optimizer is Adam (learning rate 0.001).

[0081] Experimental procedure step 1: Data preprocessing

[0082] Data augmentation was performed on the collected three-channel vibration signals: each group of samples was expanded into 3 groups by time axis translation (translation range ±10 sampling points), resulting in a total of 1800 groups of samples; Min-Max normalization was used to map all sample data to the [0,1] interval.

[0083] Step 2: Feature Fusion by Convolutional Autoencoder

[0084] The preprocessed three-channel data is input into a convolutional autoencoder. The encoder extracts abstract features through convolution operations (reducing the dimension from 1000×3 to 125×32), and the decoder reconstructs the data through deconvolution operations (restoring the dimension from 125×32 to 1000×3). The model is optimized using the cross-entropy loss function and trained until the loss converges (final loss value < 0.05), thus completing the multi-source feature fusion.

[0085] Step 3: Enhance K-means + Sparse Attention Mechanism

[0086] The abstract features output by the encoder are input into the K-means module to calculate cluster centers and divide the data into clusters, thus obtaining the feature clustering results. At the same time, a sparse attention mechanism is implanted between the first convolutional layer and the pooling layer of the encoder to calculate attention weights and generate attention-guided feature maps. The ensemble model is trained synchronously using a joint loss function to update the cluster centers and model parameters.

[0087] Step 4: Fault Classification and Verification

[0088] The trained ensemble model was applied to the test set (30% of the total samples, 540 groups). Fault classification results were obtained by reconstructing the data using a decoder and K-means clustering. The classification performance was evaluated using a confusion matrix (Figure 5), and the classification results were visualized using a clustering graph (Figure 6). Experimental results: Confusion matrix results: Figure 5 shows that the classification accuracy for all six fault states exceeded 95%, with the accuracy for normal states, inner circle faults, and outer circle faults exceeding 98%, and the accuracy for composite faults reaching 95%-97%. The misclassification rate was less than 5%, indicating excellent model classification performance.

[0089] Clustering results: During the test, multi-source vibration data of rolling bearings under six different operating states were collected (normal state, inner ring fault state, outer ring fault state, combined inner and outer ring fault state, combined inner ring and rolling element fault state, and combined outer ring and rolling element fault state). The original multi-source vibration data (horizontal, vertical, and axial) corresponding to the normal state were used for time-domain visualization, as shown below. Figures 2-4 As shown in Figure 6, the feature samples of the six fault states form distinct independent clusters in the clustering space with no significant overlap, indicating that the model has a strong ability to distinguish between different faults. Comparative experiments show that compared with the traditional single-channel convolutional autoencoder method, the fault diagnosis accuracy of the method in this invention is improved by 8%-12%, and the key feature extraction time is reduced by 15%-20%, verifying the superiority of the method. In summary, the multi-source vibration feature extraction method for rolling bearing faults proposed in this invention can effectively solve the problems of low utilization of multi-source data and incomplete key feature extraction, achieving accurate diagnosis of rolling bearing faults and possessing high engineering application value.

[0090] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A method for extracting multi-source vibration features of rolling bearing faults, characterized in that: Includes the following steps: Step S1: Use a convolutional autoencoder to perform integrated feature fusion on the original multi-source heterogeneous data, and use data augmentation and normalization techniques to preprocess the original multi-source heterogeneous data to improve the generalization ability and accuracy of the subsequent convolutional autoencoder model. Step S2: K-means clustering and sparse attention mechanisms are applied to the convolutional autoencoder model to improve the model's recognition rate of key fault features; finally, confusion matrix and visualization techniques are used to achieve fault classification. The original multi-source heterogeneous data consists of three channels of homogeneous vibration signals of the rolling bearing in the axial, horizontal and vertical directions, collected from the same monitoring point. The convolutional autoencoder model comprises an encoder and a decoder. The encoder progressively reduces the spatial dimension of the input data through convolutional and pooling layers while increasing the number of feature channels to extract abstract feature representations of the data. Its mathematical expression is as follows: ， In the formula: Conv(x, W) performs a convolution operation on the input data x, W is the convolution kernel, b represents the bias term, and σ is the activation function; The decoder gradually recovers the spatial dimension of the data through deconvolution layers while reducing the number of feature channels to reconstruct the original data. Its mathematical expression is: ， In the formula, Upsample(h) represents the upsampling operation on the decoded features, W′ is the convolution kernel of the decoder, b′ is the bias term of the decoder, and σ is the activation function; The abstract features output by the encoder are input into the K-means module to calculate cluster centers and divide the data into clusters, thus obtaining the feature clustering results. At the same time, a sparse attention mechanism is implanted between the first convolutional layer and the pooling layer of the encoder to calculate attention weights and generate attention-guided feature maps. The ensemble model is trained synchronously using a joint loss function to update the cluster centers and model parameters.

2. The method for extracting multi-source vibration features of rolling bearing faults according to claim 1, characterized in that: The training of the convolutional autoencoder uses cross-entropy loss to evaluate the difference between the model's predicted probability distribution and the true label distribution. The expression for cross-entropy is as follows: ， Where p is the probability distribution of the true labels, q is the probability distribution predicted by the model, and x represents the possible classes. This formula measures the difference between the true distribution p and the predicted distribution q by calculating the information entropy of the predicted distribution q given the true distribution p.

3. The method for extracting multi-source vibration features of rolling bearing faults according to claim 2, characterized in that: The convolutional autoencoder model employs K-means clustering and sparse attention mechanisms to improve the model's recognition rate of key fault features. The specific implementation of K-means is as follows: K-means divides n samples into K clusters, ensuring that each sample belongs to the cluster corresponding to its nearest cluster center. Its expression is as follows: ， In the formula S i Let μ be the sample set of the i-th cluster. i For S i Cluster centers, ||x-μ i || 2 Let x be the Euclidean distance from the sample x to the cluster center.

4. The method for extracting multi-source vibration features of rolling bearing faults according to claim 3, characterized in that: Integrating K-means clustering into a convolutional autoencoder model involves the following steps: Clustering of extracted high-dimensional features for data structure understanding; Clustering of dimensionality-reduced data is used to evaluate model performance and feature effectiveness; Using K-means to classify the model output simplifies data processing.

5. The method for extracting multi-source vibration features of rolling bearing faults according to claim 2, characterized in that: K-means clustering and sparse attention mechanisms are used in the autoencoder model to improve the model's recognition rate of key fault features. The specific implementation of the sparse attention mechanism is as follows: Assuming the input sequence is X, and the learnable weight matrix is ​​W, where the unit is the output dimension of the attention layer, the attention score is calculated using the following formula: ， In the formula, t represents the time interval of the input sequence; Sparsity is calculated as follows: ， In the formula, λ is the regularization coefficient; The normalized calculation is as follows: ， The final output is as follows: 。 6. The method for extracting multi-source vibration features of rolling bearing faults according to claim 2, characterized in that: The autoencoder model employs K-means clustering and sparse attention mechanisms to improve the model's recognition rate of key fault features. The ensemble model formula based on convolutional autoencoders, K-means, and sparse attention mechanisms is as follows: ， In the formula, ⊙ represents element-wise multiplication to create an attention-guided feature map before pooling; Both cluster centers and features generated by the convolutional autoencoder are optimized using the following loss function: ， Cluster center μ i These are learnable parameters that are updated via backpropagation; During the training phase, the synchronization process is as follows: ， Reconstruct the decoder using clustered conditional deconvolution: 。

Images