A wind turbine bearing fault diagnosis method

Abstract

This invention relates to the field of bearing fault diagnosis technology and discloses a method for diagnosing bearing faults in wind turbine generators. The method includes: acquiring and preprocessing raw vibration signals to construct input samples; inputting the samples into a convolutional neural network to extract multi-scale deep fault features; introducing an improved convolutional attention module (ICBAM) with a parallel structure in the feature extraction process to simultaneously apply channel and spatial attention weights to the input feature map, resulting in an enhanced feature map; inputting the enhanced features into an improved width learning system (IBLS) to generate feature nodes and enhanced nodes, which are then concatenated into an enhanced feature matrix; and using elastic network regression to solve for the IBLS output weight matrix, with the objective function including L1 and L2 regularization terms, outputting the bearing fault category. This invention enhances key fault features through a parallel dual-attention mechanism and improves classification accuracy and generalization ability by combining elastic network regression. It can accurately identify multiple types of bearing faults under varying loads and high-noise conditions, exhibiting fast training convergence and high diagnostic reliability.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and in particular to a method for diagnosing bearing faults in wind turbine generators. Background Technology

[0002] As a core component of clean energy, wind turbines operate under complex conditions such as variable loads, high noise, and high impact for extended periods. Bearings, as critical support components of the transmission system, directly affect the safety and reliability of the entire turbine. Statistics show that approximately 40% to 50% of wind turbine malfunctions are related to bearings. Bearing failure can lead to unplanned downtime or even major safety accidents. Therefore, conducting research on automatic fault diagnosis for wind turbine bearings is of great significance for achieving predictive maintenance and reducing operation and maintenance costs.

[0003] Currently, bearing fault diagnosis mainly employs vibration signal analysis technology. However, wind turbine bearing vibration signals typically exhibit strong non-stationarity, weak fault characteristics, and significant background noise interference. Traditional diagnostic methods relying on manual feature extraction suffer from insufficient robustness and limited generalization ability under complex operating conditions. In recent years, deep learning methods have been widely studied due to their end-to-end feature learning capabilities; however, existing models still suffer from insufficient attention to key fault features and low accuracy in identifying fine-grained faults across multiple categories.

[0004] To address the problem that fault diagnosis methods based on single time-domain or frequency-domain signal analysis are insufficient for the time-varying characteristic frequency of rolling bearing faults in wind turbines under variable speed conditions, existing technologies propose a fault diagnosis method for rolling bearings in variable speed wind turbines based on a bidirectional gated cyclic attention mechanism and time-frequency feature fusion (BGRA-TFFF). This BGRA-TFFF model mainly consists of four modules: initial feature extraction, deep feature extraction, feature fusion, and identification and classification. However, because this method targets fault signal interference primarily consisting of aerodynamic noise and gear meshing vibration, and the fault characteristic frequency band is relatively concentrated, it cannot correctly handle situations where fault characteristic frequency ambiguity and multi-component signal cross-interference are difficult to identify due to variable speed. Furthermore, this existing technology has the following shortcomings: 1) Poor adaptability to frequency drift under variable speed conditions. Although existing methods combine time-frequency feature fusion, they are mainly for scenarios with small speed fluctuation ranges. When wind turbines experience large speed fluctuations, time-frequency features will severely overlap and become blurred, causing the attention mechanism to fail to accurately capture fault features and affecting the diagnostic effect.

[0005] 2) Lack of ability to identify early and subtle faults. This method is effective in diagnosing mid-to-late stage bearing faults with relatively obvious fault characteristics, but for early and subtle faults in bearings under variable speed, the fault signals generated have low amplitude and hidden characteristics. Even after time-frequency feature fusion, they are still difficult to distinguish from normal signals, which can easily lead to missed diagnoses and misdiagnoses. Summary of the Invention

[0006] The purpose of this invention is to provide a method for diagnosing bearing faults in wind turbine generators. This method is based on a convolutional neural network (CNN), enhances key fault features through ICBAM, and improves the expressive power of the classification layer by utilizing IBLS, thereby achieving high-precision identification of multiple types of bearing faults.

[0007] To achieve the above objectives, the present invention provides a method for diagnosing bearing faults in wind turbine generators, comprising the following steps: Step 1: Collect the raw vibration signal of the bearing, perform preprocessing, and construct the input sample; Step 2: Input the preprocessed samples into a convolutional neural network to extract multi-scale deep fault features; Step 3: In the feature extraction process of the convolutional neural network, an improved convolutional attention module (ICBAM) with a parallel structure is introduced to simultaneously perform channel attention weighting and spatial attention weighting on the input feature map to obtain an enhanced feature map; wherein, the ICBAM module includes a parallel channel attention sub-module and a spatial attention sub-module, both of which directly act on the original input feature map, and the output is the result of multiplying the channel attention map and the spatial attention map element by element by the original feature map; Step 4: Input the enhanced deep features into the Improved Width Learning System (IBLS), generate feature nodes through the feature mapping layer, generate enhanced nodes through the enhancement layer, and concatenate the feature nodes and enhanced nodes into an enhanced feature matrix; Step 5: Solve the output weight matrix of IBLS using elastic net regression, where the objective function includes L1 regularization and L2 regularization terms to achieve the discrimination and output of bearing fault categories.

[0008] In step one, the preprocessing includes: The original vibration signal is processed to remove the mean, and the fault features are extracted by envelope spectrum analysis. The original vibration signal is converted into an envelope spectrum feature sequence, and the signal is segmented by a fixed-length sliding window to construct a training sample set.

[0009] In step three: The channel attention submodule performs global max pooling and global average pooling on the input feature map to aggregate spatial dimension information, obtaining max pooling features and average pooling features. Then, the two are input into a shared multilayer perceptron to generate a channel attention map. The spatial attention submodule performs parallel max pooling and average pooling operations on the input feature map to aggregate channel information. The two feature maps are then concatenated and passed through a convolutional layer to generate a two-dimensional spatial attention map.

[0010] In step four: In the improved width learning system IBLS, feature nodes are defined by the formula... Generate, where It is a random weight. It's a bias. It is a mapping activation function. A column vector consisting entirely of 1s is used for broadcast bias; Enhanced nodes via formula Generate, where and These represent the randomly generated weights and biases, respectively. This is the activation function.

[0011] In step five, the objective function for the elastic network regression is: in, is the regularization factor for lasso regression, controlling L1 regularization; is the regularization factor for ridge regression, controlling L2 regularization.

[0012] In step five: The bearing failure categories include normal condition, inner ring failure, outer ring failure, rolling element failure, and their corresponding different damage scales.

[0013] The fixed-length sliding window has a window length of 4096, an overlap rate of 0, and a file-level partitioning strategy is used for sample partitioning.

[0014] The present invention provides a method for diagnosing bearing faults in wind turbine generators. Compared with the prior art, the advantages of the present invention include: It can accurately identify multiple types of faults in wind turbine bearings under complex working conditions such as variable load and strong noise. It is highly adaptable to weak fault characteristics and non-stationary vibration signals, and has high diagnostic accuracy. The model has high optimization efficiency, stable training, and can effectively learn fault features within a limited number of iterations, with strong generalization ability. It can intelligently and quickly identify bearing failures in wind turbine units, providing a reliable basis for predictive maintenance, helping to detect potential faults in advance, reduce downtime losses, and improve the safety and reliability of wind turbine unit operation. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0016] Figure 1 This is a diagram of the IBLS-ICBAM-CNN model of the present invention.

[0017] Figure 2 This is a schematic diagram of the structure of the improved Convolutional Attention Module (ICBAM) of this invention.

[0018] Figure 3 This is the original model iteration diagram of the IBLS-ICBAM-CNN model of this invention.

[0019] Figure 4 This invention relates to a fault diagnosis confusion matrix and a t-SNE clustering diagram.

[0020] Figure 5 This is a flowchart of the wind turbine bearing fault diagnosis method of the present invention. Detailed Implementation

[0021] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0022] This invention provides a method for diagnosing bearing faults in wind turbine generators, comprising the following steps: S1: Collect the raw vibration signal of the bearing, perform preprocessing, and construct the input sample.

[0023] Specifically, the original vibration signal is processed to remove the mean, and the fault features are extracted by envelope spectrum analysis. The original vibration signal is converted into an envelope spectrum feature sequence, and the signal is segmented by a fixed-length sliding window to construct a training sample set.

[0024] The fixed-length sliding window has a window length of 4096, an overlap rate of 0, and a file-level partitioning strategy is used for sample partitioning.

[0025] S2: Input the preprocessed samples into a convolutional neural network to extract multi-scale deep fault features.

[0026] S3: In the feature extraction process of the convolutional neural network, an improved convolutional attention module ICBAM with a parallel structure is introduced to simultaneously perform channel attention weighting and spatial attention weighting on the input feature map to obtain an enhanced feature map; wherein, the ICBAM module includes a parallel channel attention submodule and a spatial attention submodule, both of which directly act on the original input feature map, and the output is the result of multiplying the channel attention map and the spatial attention map element by element by the original feature map.

[0027] Specifically, the channel attention submodule performs global max pooling and global average pooling on the input feature map to aggregate spatial dimension information, obtaining max pooling features and average pooling features. Then, the two are input into a shared multilayer perceptron to generate a channel attention map. The spatial attention submodule performs parallel max pooling and average pooling operations on the input feature map to aggregate channel information. The two feature maps are then concatenated and passed through a convolutional layer to generate a two-dimensional spatial attention map.

[0028] S4: Input the enhanced deep features into the improved width learning system IBLS, generate feature nodes through the feature mapping layer, generate enhanced nodes through the enhancement layer, and concatenate the feature nodes and enhanced nodes into an enhanced feature matrix.

[0029] Specifically, in the improved width learning system IBLS, feature nodes are defined by the formula... Generate, where It is a random weight. It's a bias. It is a mapping activation function. A column vector consisting entirely of 1s is used for broadcast bias; Enhanced nodes via formula Generate, where and These represent the randomly generated weights and biases, respectively. This is the activation function.

[0030] S5: The output weight matrix of IBLS is solved by elastic net regression, where the objective function includes L1 regularization and L2 regularization terms to realize the discrimination and output of bearing fault categories.

[0031] Specifically, the objective function of the elastic network regression is: in, is the regularization factor for lasso regression, controlling L1 regularization; is the regularization factor for ridge regression, controlling L2 regularization.

[0032] The bearing failure categories include normal condition, inner ring failure, outer ring failure, rolling element failure, and their corresponding different damage scales.

[0033] For this specific implementation, the present invention proposes an IBLS-ICBAM-CNN fault diagnosis model that integrates an incremental width learning system and an attention mechanism (e.g., Figure 1 As shown, it includes a data preprocessing and input module, an improved convolutional attention module (ICBAM), and an improved width learning system module (IBLS).

[0034] The data preprocessing and input module: First, raw vibration signals of the wind turbine bearings are acquired using an accelerometer. The raw signals are then processed to remove the DC component and standardize the physical dimensions. Next, envelope spectrum analysis is used to extract fault features, converting the raw vibration signals into envelope spectrum feature sequences, which serve as model input. A fixed-length sliding window (4096 pixels with 0% overlap) is used to segment the signal, constructing a training sample set. The sample partitioning employs a file-level partitioning strategy. Finally, the training sample set is input into a convolutional neural network, where multi-level convolution and pooling operations are used to extract multi-scale fault features.

[0035] The improved convolutional attention module (ICBAM): This invention proposes an improved convolutional attention module (ICBAM), the structure of which is as follows: Figure 2 As shown. Unlike the existing CBAM where the Channel Attention Submodule (CAM) and Spatial Attention Submodule (SAM) are connected in a serial manner, ICBAM designs CAM and SAM as a parallel structure, enabling them to directly act on the original input feature map, thereby enhancing the flexibility and effectiveness of feature extraction.

[0036] The improved convolutional attention module includes a channel attention submodule (CAM), a spatial attention submodule (SAM), and an ICBAM module; The channel attention submodule: first processes the input feature map... Perform global max pooling and global average pooling respectively, and aggregate spatial dimensionality information to obtain max pooling features. and average pooling characteristics Then, both are input into a shared multilayer perceptron (MLP) to generate a channel attention map. The calculation formula is as follows: In the formula, Represents the sigmoid function; This represents the max pooling operation; This indicates the average pooling operation.

[0037] The spatial attention submodule: aggregates channel information from the original input feature map through parallel max pooling and average pooling operations, obtaining... and Then the two are concatenated and processed through a convolution kernel of size [size missing]. The convolutional layer generates a two-dimensional spatial attention map. The calculation formula is as follows: In the formula, Indicates filter size is Convolutional layers; This represents the pooled features generated by max pooling operations on the feature map; The pooled features are generated by the average pooling operation on the representative feature map. This indicates a cascading operation.

[0038] The ICBAM module: The final output of the ICBAM module is the original feature map. With channel attention map Spatial attention map Element-wise product: in, This represents element-wise multiplication. Through parallel structure and dual attention mechanism, ICBAM can more comprehensively enhance fault sensitivity characteristics and suppress irrelevant noise.

[0039] Improved Broadband Learning System Module (IBLS): This invention proposes an Improved Broad Learning System (IBLS). Unlike traditional Broad Learning Systems (BLS) that use a pseudo-inverse method to solve for output weights, IBLS introduces Elastic Net Regression to estimate network weights, thereby mitigating the adverse effects of random weights and further improving the model's classification performance and generalization ability.

[0040] Let the data matrix output by ICBAM be... The specific steps are as follows: (1) Feature mapping layer: The input is mapped to the feature mapping layer. Feature nodes are generated by mapping to the feature space through linear transformation and activation function. : In the formula It is a random weight. It's a bias. It is a mapping activation function. A column vector consisting entirely of 1s is used for broadcast bias.

[0041] (2) Enhance the node layer: enhance the feature nodes Further expansion using a nonlinear activation function generates enhanced nodes. : In the formula, and These represent the randomly generated weights and biases, respectively. This is the activation function.

[0042] (3) Output layer weight calculation: The feature nodes With enhanced nodes Concatenate into an enhanced feature matrix Traditional BLS uses a ridge regression closed-form solution. , is the regularization coefficient. This invention uses elastic network regression, and the objective function is: In the formula: This is the regularization factor for lasso regression, controlling L1 regularization (which produces sparse solutions). This is the regularization factor for ridge regression, controlling L2 regularization (to prevent overfitting). Through elastic net regression, IBLS can adaptively select important features while maintaining model stability, effectively improving the classification accuracy of multi-class faults.

[0043] Example: The specific steps of the wind turbine bearing fault diagnosis method of the present invention are as follows: (1) Signal acquisition and preprocessing: Acquire bearing vibration signals, perform mean removal, segmentation, and envelope spectrum transformation to construct input samples; (2) Deep feature extraction: Input the samples into the CNN network and extract multi-scale fault features through multi-layer convolution and pooling operations; (3) Attention enhancement: The ICBAM module is introduced in the feature extraction process to adaptively weight the channel dimension and spatial dimension features to enhance the expression of key fault information; (4) Classification decision: The enhanced deep features are input into the IBLS module, and feature recombination and classification decision are achieved through feature node mapping and enhancement node construction; (5) Output diagnostic results: Output bearing fault categories (normal, inner ring fault, outer ring fault, rolling element fault, etc. and their damage scales) to achieve automatic identification of multiple types of faults.

[0044] To verify the effectiveness of the proposed IBLS-ICBAM-CNN model in rolling bearing fault diagnosis, this invention conducted a ten-class experiment based on the Case Western Reserve University (CWRU) bearing dataset. The experimental categories included normal conditions, rolling element faults, inner ring faults, and outer ring faults under various damage scales. To comprehensively evaluate the model's diagnostic performance, this invention analyzes it from three aspects: model convergence, classification results, and feature separability, and presents model iteration curves, confusion matrices, and t-SNE clustering visualization results, such as... Figure 3 and Figure 4 As shown.

[0045] Figure 3 The accuracy variation curves of the IBLS-ICBAM-CNN model during training are presented. As shown in the figure, the model achieves relatively fast convergence in the early stages of training, with the training set accuracy rapidly improving in the first few iterations and stabilizing within a short iteration interval. Meanwhile, although the test set accuracy is lower than the training set accuracy initially, it shows a continuous upward trend with increasing iterations and eventually stabilizes at a high level. This indicates that the constructed model can effectively learn fault discrimination information from vibration signals within a limited number of training rounds, exhibiting good optimization efficiency and training stability.

[0046] from Figure 4 (a) The normalized confusion matrix results show that the main diagonal elements are dominant, and the vast majority of categories can be accurately identified, indicating that the model has a strong ability to identify samples such as normal conditions, inner ring faults, and outer ring faults. A small number of misclassifications are mainly concentrated between different damage scales of rolling element faults and between a few similar outer ring fault categories. This indicates that when the damage degree of similar faults is close, their vibration response characteristics have a certain similarity, thus increasing the difficulty of classification. However, overall, the misclassification rate is low, and it mainly occurs between similar categories, having a small impact on the overall diagnostic performance. Further utilization... Figure 4 (b) After performing two-dimensional visualization of the deep features using t-SNE, it was found that samples of different categories formed a relatively clear cluster distribution in the feature space, exhibiting good intra-class compactness and inter-class separation. This indicates that the model can effectively extract fault-sensitive features and compress redundant information in the original signal. Although some rolling element fault samples of similar size still show local proximity or slight overlap, the overall clustering structure is relatively clear, consistent with the confusion matrix results.

[0047] The wind turbine bearing fault diagnosis method of the present invention: 1. The channel attention submodule and the spatial attention submodule are designed to be connected in parallel, enabling them to directly act on the original feature map simultaneously, avoiding the limitation of the attention order on the feature enhancement effect in a serial structure. Through the parallel dual attention mechanism, the model can more comprehensively capture fault-sensitive features in both the channel and spatial dimensions, significantly improving feature representation capabilities.

[0048] 2. To address the instability caused by the random mapping in traditional BLS, an elastic network regression (L1+L2 regularization) is introduced to replace the original ridge regression pseudo-inverse solution method. This improvement not only adaptively selects important features (sparseness) but also effectively prevents overfitting (stability), thus achieving significantly higher classification accuracy than traditional BLS in multi-class fine-grained fault classification tasks.

[0049] 3. By embedding ICBAM into the CNN feature extraction stage and using IBLS as the core module in the classification decision stage, an end-to-end diagnostic architecture of "CNN feature extraction → ICBAM feature enhancement → IBLS classification decision" is formed. The synergistic effect of the two enables the model to improve at both the feature representation and classification levels.

[0050] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for diagnosing bearing faults in wind turbine generators, characterized in that, Includes the following steps: Step 1: Collect the raw vibration signal of the bearing, perform preprocessing, and construct the input sample; Step 2: Input the preprocessed samples into a convolutional neural network to extract multi-scale deep fault features; Step 3: In the feature extraction process of the convolutional neural network, an improved convolutional attention module (ICBAM) with a parallel structure is introduced to simultaneously perform channel attention weighting and spatial attention weighting on the input feature map to obtain an enhanced feature map; wherein, the ICBAM module includes a parallel channel attention sub-module and a spatial attention sub-module, both of which directly act on the original input feature map, and the output is the result of multiplying the channel attention map and the spatial attention map element by element by the original feature map; Step 4: Input the enhanced deep features into the Improved Width Learning System (IBLS), generate feature nodes through the feature mapping layer, generate enhanced nodes through the enhancement layer, and concatenate the feature nodes and enhanced nodes into an enhanced feature matrix; Step 5: Solve the output weight matrix of IBLS using elastic net regression, where the objective function includes L1 regularization and L2 regularization terms to achieve the discrimination and output of bearing fault categories.

2. The wind turbine bearing fault diagnosis method as described in claim 1, characterized in that, In step one, the preprocessing includes: The original vibration signal is processed to remove the mean, and the fault features are extracted by envelope spectrum analysis. The original vibration signal is converted into an envelope spectrum feature sequence, and the signal is segmented by a fixed-length sliding window to construct a training sample set.

3. The wind turbine bearing fault diagnosis method as described in claim 1, characterized in that, In step three: The channel attention submodule performs global max pooling and global average pooling on the input feature map, aggregates spatial dimension information, and obtains max pooling features and average pooling features. Then, the two are input into a shared multilayer perceptron to generate a channel attention map. The spatial attention submodule aggregates channel information from the input feature map through parallel max pooling and average pooling operations, concatenates the two feature maps, and then passes them through a convolutional layer to generate a two-dimensional spatial attention map.

4. The wind turbine bearing fault diagnosis method as described in claim 1, characterized in that, In step four: In the improved width learning system IBLS, feature nodes are defined by the formula... Generate, where It is a random weight. It's a bias. It is a mapping activation function. A column vector consisting entirely of 1s is used for broadcast bias; Enhanced nodes via formula Generate, where and These represent the randomly generated weights and biases, respectively. This is the activation function.

5. The wind turbine bearing fault diagnosis method as described in claim 1, characterized in that, In step five, the objective function of the elastic network regression is: in, This is the regularization factor for lasso regression, controlling for L1 regularization; is the regularization factor for ridge regression, controlling L2 regularization.

6. The wind turbine bearing fault diagnosis method as described in claim 1, characterized in that, In step five: The bearing fault categories include normal condition, inner ring fault, outer ring fault, rolling element fault, and their corresponding different damage scales.

7. The wind turbine bearing fault diagnosis method as described in claim 2, characterized in that, The fixed-length sliding window has a window length of 4096 and an overlap rate of 0. The sample partitioning adopts a file-level partitioning strategy.

Images