Intelligent wind turbine bearing fault diagnosis method based on improved Swin-T

Abstract

The present invention relates to an intelligent wind power bearing fault diagnosis method based on improved Swin-T, which includes: real-time on-site acquisition of acoustic emission signals of main shaft rolling bearings; denoising the acoustic emission signals of rolling bearings acquired in step one, extracting the acoustic emission signal features of main shaft rolling bearings based on FBank, mapping the obtained FBank feature data to a two-dimensional image, and constructing an acoustic emission image data set; adding parallel attention mechanisms CBAM and SKNet to a Swin-Transformer model to obtain an AD model, inserting the AD model into a Swin-Transformer model to obtain an AD-Swin-T model; and using the AD-Swin-T model to identify the constructed acoustic emission image data set. The present invention converts bearing signals into an FBank feature input model, effectively realizing accurate fault detection of low-speed and heavy-load bearings.

Description

Technical Field

[0001] The invention relates to the technical field of bearing failure of power generation system, and in particular to an intelligent wind power generation bearing failure diagnosis method based on improved Swin-T. Background Art

[0002] Contemporary industrial systems are increasingly dependent on information and intelligent technologies. Under this trend, accurate and efficient intelligent bearing fault diagnosis is crucial. As an important component of heavy rotating machinery, the role of bearings is to reduce the friction between rotating shafts and stationary parts (such as bearing seats). According to statistics, bearing defects account for a considerable proportion of rotating machinery failures (approximately 45-55%). Therefore, early detection of initial bearing failures is crucial to preventing serious failures in manufacturing equipment. In recent years, a variety of bearing fault diagnosis technologies based on acoustic emission (AE) have flourished. Acoustic emission analysis uses acoustic emission signals to detect weak signals caused by bearing failures in the early stages or during low-speed operation. However, due to the high sampling rate of acoustic emission signals (usually greater than 1MHz), the amount of time series data collected is huge, which increases the difficulty of analyzing acoustic emission signals and requires a lot of computing time.

[0003] Due to the complexity and variability of the working environment of bearings, traditional convolutional neural networks (CNNs) are easily disturbed by unbalanced data during training, which is not conducive to improving the accuracy of fault diagnosis. This is mainly because CNN establishes a mapping relationship between fault samples and fault categories for the original fault data through learning between network layers, and requires a large amount of balanced data to train deep diagnostic models. In the process of building bearing fault monitoring and diagnosis models, normal data and fault data are seriously unbalanced in the data set, which greatly reduces the accuracy and stability of bearing fault diagnosis results. In addition, these CNN-based methods still have some defects because these methods have limitations in extracting global information. When CNN extracts target functions, global functions can only be obtained by stacking many layers.

[0004] With the unremitting pursuit of carbon neutrality goals around the world, wind power generation, as a representative of clean energy, is particularly critical for the reliability of its equipment. In particular, bearings, which are important components of wind turbine core components, have a direct impact on the stable operation and economic benefits of the entire system. However, the manual inspection method of wind turbine main shaft bearings has problems such as low efficiency and high labor costs. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent wind power generation bearing fault diagnosis method based on improved Swin-T, which is used to solve the problem that the accuracy and stability of bearing fault diagnosis results in the prior art are not high enough.

[0006] The technical solution adopted by the present invention to solve the technical problem is: the intelligent wind power bearing fault diagnosis method based on improved Swin-T comprises the following steps:

[0007] Step 1: Collect the acoustic emission signal of the main shaft rolling bearing in real time on site, and send the collected signal to the bearing fault diagnosis system in real time;

[0008] Step 2: In the bearing fault diagnosis system, the acoustic emission signal of the rolling bearing collected in step 1 is denoised. The specific method is as follows: based on VMD-K continuous acoustic emission signal noise suppression, VMD decomposes the signal into a series of modal components IMFs with a specific frequency bandwidth, and VMD-K is combined with kurtosis analysis to screen out IMF components with a kurtosis factor greater than 3 for noise reduction and signal feature extraction;

[0009]

[0010] In the formula is the mean of the signal, σ is the standard deviation;

[0011] Step 3: Extract the acoustic emission signal features of the spindle rolling bearing based on FBank, map the obtained FBank feature data into a two-dimensional image, and construct an acoustic emission image data set;

[0012]

[0013] Where M is the Mel frequency, f is the frequency;

[0014] Step 4: Pattern recognition of acoustic emission signals of spindle rolling bearing damage based on improved Swin-Transforme;

[0015] Using the transfer learning method, the pre-trained weights of the Swin-Transformer model are called, and the parallel attention mechanisms CBAM and SKNet are added to it to obtain the attention model AD, which is finally inserted into the Swin-Transformer to obtain the AD-Swin-T model;

[0016] Step 5: Use the AD-Swin-T model to identify the acoustic emission image dataset constructed in step 3.

[0017] The specific method of step 4 in the above scheme is:

[0018] S41, call the pre-trained weights of the Swin-Transformer model, delete the weights of the classification category, and freeze all weights except the head;

[0019] S42. In the attention model AD, CBAM and SKNet process the same input features in parallel. The input features enter CBAM and SKNet respectively and are processed independently. Once the two sub-modules of CBAM and SKNet have completed processing, their output features are combined. In the data flow of the entire Swin-Transformer model, the output features of AD will replace the output of the original SwinTransformer block as the input of the next stage.

[0020] S43. The attention model AD in S42 is added to each layer of stage1-stage4 in S41. The features after combining the attention model AD include the channel and spatial attention weights, as well as the dynamic selection results of convolution kernels of different sizes, so that it can capture richer and more detailed information.

[0021] Step 5 in the above scheme is specifically as follows:

[0022] S51. In the AD-Swin-T model, the input image first passes through the Patch Partition stage, where the image is partitioned into non-overlapping patches of size 4×4; these patches are used as the basic unit of processing in subsequent steps, and each patch is converted into a feature vector whose dimension is expanded to C, the embedding dimension;

[0023] S52, then, the linear transformation is performed on the channel data of each pixel through the Linear Embedding layer, from 48 to C, that is, the image shape is changed from [H / 4, W / 4, 48] to [H / 4, W / 4, C];

[0024] S53, the remaining three stages are first downsampled through a Patch Merging layer, and then weighted by the attention model AD: the dimension of the Stage2 feature map becomes H / 8×W / 8, and the number of channels doubles to 2C; the dimension of the Stage3 is reduced again to H / 16×W / 16, and the number of channels becomes 4C; the dimension of the Stage4 feature map is reduced to H / 32×W / 32, and the number of channels increases to 8C;

[0025] S54. Finally, the features are aggregated through a global average pooling layer and mapped to the final output category through a linear layer to complete the recognition and classification process of the entire image.

[0026] The basic expression of VMD in the above scheme is:

[0027] H(t)=Σ(k=1) K u k (t)

[0028] Where κ is the total number of modes, uk (t) is the kth mode;

[0029] VMD optimization goals:

[0030]

[0031] stΣku k =1

[0032] Where {w k} is the set of center frequencies corresponding to the modes, δ(t) is the Dirac function, * represents convolution,

[0033] Step 3 in the above scheme is as follows:

[0034] 3.1、Speech pre-emphasis, framing and windowing;

[0035] 3.2, FFT Fourier transform to get the spectrum;

[0036] 3.3. Obtain the Mel spectrum through the Mel filter bank;

[0037] 3.4. Perform logarithmic transformation on the Mel spectrum to obtain FBank features.

[0038] Beneficial effects:

[0039] 1. This paper combines the Swin Transformer deep learning model with acoustic emission (AE) technology for the first time, converts bearing signals into FBank feature input models, and effectively realizes accurate fault detection of low-speed and heavy-loaded bearings.

[0040] 2. After each Swin Transformer block, the present invention introduces the AdaptiveFocusNet module. This module combines the attention mechanism of CBAM and the adaptive convolution kernel selection capability of SKNet to enhance the model's learning of key features and improve the ability to express features. This unique module enables the model to not only focus on local details when processing images, but also capture a wider range of contextual information.

[0041] 3. In the early stage of wind turbine main shaft bearing failure, there is a lot of noise interference in signal collection, which makes it difficult to extract fault features and cannot detect the fault as early as possible. Therefore, the noise reduction problem of early fault signals of rolling bearings is the key to fault diagnosis. The present invention solves the noise reduction problem of early fault signals of rolling bearings based on the VMD-K continuous acoustic emission signal noise suppression method.

[0042] 4. In the present invention, CBAM and SKNet work in parallel, taking advantage of the two strategies. First, CBAM highlights the key areas and feature channels in the image through its attention mechanism, enhancing the representation ability of the feature map. At the same time, SKNet selects the most suitable convolution kernel size in different receptive fields, further improving the adaptability and expressiveness of the features. This parallel structure not only increases the network's perception depth for complex image structures, but also improves the diversity and richness of feature expression.

[0043] 5. The present invention improves the accuracy and efficiency of bearing fault detection, and further optimizes the operating performance and economic benefits of the wind power generation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of the present invention;

[0045] Figure 2 This is the FBANK speech feature extraction process diagram;

[0046] Figure 3 It is a schematic diagram of the adaptive focusing network Swin-Transformer;

[0047] Figure 4 It is a schematic diagram of CBAM and SKNet working in parallel. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the accompanying drawings:

[0049] See also Figure 1 ,This intelligent wind power bearing fault diagnosis method based on improved Swin-T includes the following steps:

[0050] Step 1: Relying on the low-speed heavy-load rolling heavy-load bearing rolling platform, collect the acoustic emission signals of the rolling bearing on site, and send the collected signals to the bearing fault diagnosis monitoring and evaluation system in real time;

[0051] Step 2: In the bearing fault diagnosis monitoring and evaluation system, the acoustic emission signal of the rolling bearing collected in step 1 is denoised. The specific method is: based on VMD-K continuous acoustic emission signal noise suppression, VMD decomposes the signal into a series of modal components IMFs with a specific frequency bandwidth, and VMD-K is combined with kurtosis analysis to screen out IMF components with a kurtosis factor greater than 3 for noise reduction and signal feature extraction;

[0052] During the operation of wind turbine main shaft bearings, various factors can cause damage or failure, leading to faults. Once a fault occurs, it will cause serious safety hazards. In the early stage of wind turbine main shaft bearing failure, there is a lot of noise interference in signal collection, which makes it difficult to extract fault features and cannot detect faults as early as possible. Therefore, the noise reduction problem of early fault signals of rolling bearings is the key to fault diagnosis.

[0053] The basic expression of VMD: Signal reconstruction is achieved by the accumulation of modal components:

[0054]

[0055] where k is the total number of modes, u k (t) is the kth mode.

[0056] The optimization goal of VMD is to perform mode decomposition by minimizing the following objective function:

[0057]

[0058]

[0059] where {w k} is the set of center frequencies corresponding to the modes, δ(t) is the Dirac function, * represents convolution,

[0060] Kurtosis analysis of VMD-K: Kurtosis is used to evaluate the peak characteristics of a component:

[0061]

[0062] in is the mean of the signal, and σ is the standard deviation. Components with kurtosis values ​​greater than 3 are usually considered to contain important fault features.

[0063] VMD obtains modal components by decomposing the signal, while VMD-K uses kurtosis to screen modal components to improve the effect and accuracy of signal processing (such as noise reduction and feature extraction). When a rolling bearing fails, the amplitude of the bearing vibration signal increases significantly, and the value of the kurtosis factor also increases; the larger the amplitude of the fault signal, the more obvious the fault feature. IMF components with a kurtosis factor greater than 3 are usually screened out because these components can retain a large number of fault features in the fault signal while discarding the interference of other IMF components on the reconstructed signal.

[0064] Step 3: Extract the features of the acoustic emission signal of the main shaft rolling bearing based on FBank; The FBank method takes into account the auditory characteristics of human beings, first maps the linear spectrum to the nonlinear space, and then converts it to the cepstrum. The Ml frequency formula is as follows:

[0065]

[0066] Where M is the Mel frequency and f is the frequency.

[0067] Figure 2 This is the FBANK speech feature extraction process, see Figure 2 , the steps to extract FBANK features are as follows:

[0068] 3.1、Speech pre-emphasis, framing and windowing;

[0069] 3.2, FFT Fourier transform to get the spectrum;

[0070] 3.3. Obtain the Mel spectrum through the Mel filter bank;

[0071] 3.4. Perform logarithmic transformation on the Mel spectrum to obtain FBank features.

[0072] Step 4: Pattern recognition of acoustic emission signals of spindle rolling bearing damage based on improved Swin-Transforme;

[0073] The transfer learning method is used to call the pre-trained weights of the Swin-Transformer model, and the parallel attention mechanisms CBAM and SKNet are added to it to obtain the attention model AD, which is finally inserted into the Swin-Transformer to obtain the AD-Swin-T model.

[0074] S41, call the pre-trained weights of the Swin-Transformer model, delete the weights of the classification category, and freeze all weights except the head;

[0075] S42. In the attention model AD, CBAM and SKNet process the same input features in parallel. The input features enter CBAM and SKNet respectively and are processed independently. Once the two sub-modules of CBAM and SKNet have completed processing, their output features are combined. In the data flow of the entire Swin-Transformer model, the output features of AD will replace the output of the original SwinTransformer block as the input of the next stage.

[0076] S43. The attention model AD in S42 is added to each layer of stage1-stage4 in S41. The features after combining the attention model AD include the channel and spatial attention weights, as well as the dynamic selection results of convolution kernels of different sizes, so that it can capture richer and more detailed information.

[0077] Step 5: Use the AD-Swin-T model to identify bearing damage acoustic emission images:

[0078] S51. In the AD-Swin-T model, the input image first passes through the Patch Partition stage, where the image is partitioned into non-overlapping patches of size 4×4; these patches serve as the basic units of processing in subsequent steps, and each patch is converted into a feature vector whose dimension is expanded to C, the embedding dimension.

[0079] S52. Then, the linear transformation is performed on the channel data of each pixel through the Linear Embedding layer, from 48 to C, that is, the image shape is changed from [H / 4, W / 4, 48] to [H / 4, W / 4, C].

[0080] S53 and the remaining three stages are first downsampled through a Patch Merging layer and weighted by the attention model AD: the dimension of the Stage2 feature map becomes H / 8×W / 8, and the number of channels doubles to 2C; the dimension of the Stage3 is reduced again to H / 16×W / 16, and the number of channels becomes 4C; the dimension of the Stage4 feature map is reduced to H / 32×W / 32, and the number of channels increases to 8C. The increase in the number of channels in each stage enables the model to improve the richness of features and the expressiveness of the model while reducing the spatial resolution. This design enables the model to better handle image information of different scales and complexities, and is particularly suitable for complex visual tasks such as image classification, object detection, and semantic segmentation.

[0081] S54. Finally, the features are aggregated through a global average pooling layer and mapped to the final output category through a linear layer to complete the recognition and classification process of the entire image.

[0082] In the data flow of the entire Swin Transformer model, the output features of AdaptiveFocusNet will replace the output of the original Swin Transformer block as the input of the next stage. These processed features not only contain spatial details, but also integrate contextual information of different scales, thereby providing a more comprehensive perspective for the model's decision-making. This design makes the model more accurate in extracting features and helps improve the performance of the model on complex visual tasks.

Claims

1. An intelligent wind turbine bearing fault diagnosis method based on improved Swin-T, characterized in that The steps include: Step 1: Collect the acoustic emission signal of the main shaft rolling bearing in real time on site, and send the collected signal to the bearing fault diagnosis system in real time; Step 2: In the bearing fault diagnosis system, the acoustic emission signal of the rolling bearing collected in step 1 is denoised. The specific method is as follows: based on VMD-K continuous acoustic emission signal noise suppression, VMD decomposes the signal into a series of modal components IMFs with a specific frequency bandwidth, and VMD-K is combined with kurtosis analysis to screen out IMF components with a kurtosis factor greater than 3 for noise reduction and signal feature extraction; In the formula is the mean of the signal; σ is the standard deviation; K U is the kurtosis value, which indicates the sharpness of the signal distribution and reflects the steepness or flatness of the data distribution; x(i): represents the amplitude of the i-th sampling point in the time series, and the value range of i is 1 to N; Step 3: Extract the acoustic emission signal features of the spindle rolling bearing based on FBank, map the obtained FBank feature data into a two-dimensional image, and construct an acoustic emission image data set; Where Mel(f) is the Mel frequency, f is the frequency; Step 4: Pattern recognition of acoustic emission signals of spindle rolling bearing damage based on improved Swin-Transformer; using the transfer learning method, the pre-trained weights of the Swin-Transformer model are called, and the parallel attention mechanisms CBAM and SKNet are added to it to obtain the attention model AD, and finally inserted into the Swin-Transformer to obtain the AD-Swin-T model; Step 5: Use the AD-Swin-T model to identify the acoustic emission image dataset constructed in step 3.

2. The intelligent wind power bearing fault diagnosis method based on improved Swin-T according to claim 1 is characterized in that: The specific method of step 4 is: S41, call the pre-trained weights of the Swin-Transformer model, delete the weights of the classification category, and freeze all weights except the head; S42. In the attention model AD, CBAM and SKNet process the same input features in parallel. The input features enter CBAM and SKNet respectively and are processed independently. Once the two sub-modules of CBAM and SKNet have completed processing, their output features are combined. In the data flow of the entire Swin-Transformer model, the output features of AD will replace the output of the original SwinTransformer block as the input of the next stage; S43. The attention model AD in S42 is added to each layer of stage1-stage4 in S41. The features after combining the attention model AD include the channel and spatial attention weights, as well as the dynamic selection results of convolution kernels of different sizes, so that it can capture richer and more detailed information.

3. The intelligent wind power bearing fault diagnosis method based on improved Swin-T according to claim 2 is characterized in that: The step five is specifically as follows: S51. In the AD-Swin-T model, the input image first passes through the Patch Partition stage, where the image is partitioned into non-overlapping patches of size 4×4; these patches are used as the basic unit of processing in subsequent steps, and each patch is converted into a feature vector whose dimension is expanded to C, the embedding dimension; S52, then, the linear transformation is performed on the channel data of each pixel through the Linear Embedding layer, from 48 to C, that is, the image shape is changed from [H / 4, W / 4, 48] to [H / 4, W / 4, C], where H is the height of the input image, W is the width of the input image, and C is the number of channels of the feature map; S53, the remaining three stages are first downsampled through a Patch Merging layer, and then weighted by the attention model AD: the dimension of the Stage2 feature map becomes H / 8×W / 8, and the number of channels doubles to 2C; the dimension of the Stage3 is reduced again to H / 16×W / 16, and the number of channels becomes 4C; the dimension of the Stage4 feature map is reduced to H / 32×W / 32, and the number of channels increases to 8C; S54. Finally, the features are aggregated through a global average pooling layer and mapped to the final output category through a linear layer to complete the recognition and classification process of the entire image.

4. The intelligent wind power bearing fault diagnosis method based on improved Swin-T according to claim 3 is characterized in that: The basic expression of the VMD is: H(t)=∑(k=1) K u k (t) Where K is the total number of modes, u k (t) is the kth mode; VMD optimization goals: Where {w k } is the set of center frequencies corresponding to the modes, is the time partial derivative operator, δ(t) is the Dirac function, * represents convolution, πt is the kernel function of the Hilbert transform, is the squared L2 norm, Imaginary units; st stands for "constraints", which indicates the restrictions of the optimization problem.

5. The intelligent wind power bearing fault diagnosis method based on improved Swin-T according to claim 4 is characterized in that: The step three is as follows: 3.1、Speech pre-emphasis, framing and windowing; 3.2, FFT Fourier transform to get the spectrum; 3.

3. Obtain the Mel spectrum through the Mel filter bank; 3.

4. Perform logarithmic transformation on the Mel spectrum to obtain FBank features.

Images