A small sample based WPD and AFRB-LWU Net rolling bearing fault diagnosis method

Abstract

The present application relates to the technical field of fault diagnosis, in particular to a rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet under small samples, which comprises the following steps: S1: collecting vibration signals of mechanical equipment bearings when various faults occur, carrying out wavelet packet decomposition and energy feature extraction on the vibration signals, reconstructing the vibration signals into one-dimensional time series signals, and completing preliminary data preprocessing; S2: changing the original UNet model from four layers of up-sampling layers and down-sampling layers to two layers of up-sampling layers and down-sampling layers to form a LWUNet model, embedding an attention fusion residual block in the jump connection part of the LWUNet model, and building an AFRB-LWUNet model; S3: training, verifying and testing the AFRB-LWUNet model; S4: using the trained AFRB-LWUNet model to diagnose faults under different working conditions and testing the robustness of the model; and S5: monitoring the bearing vibration data in the running process of the mechanical equipment in real time, inputting the preprocessed data into the trained model, and carrying out real-time fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, specifically to a rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet with a small sample size. Background Technology

[0002] With the continuous emergence of new technologies, mechanical equipment is gradually developing towards automation, speed, and intelligence. As a core component of rotating machinery, the reliability and stability of rolling bearings directly affect the performance of rotating machinery. Studies have shown that 40% to 50% of rotating machinery failures are related to rolling bearing failures. Therefore, to ensure the safe operation of rotating machinery, rapid and accurate fault diagnosis of bearings is of great significance.

[0003] Fault diagnosis of rolling bearings mainly includes signal acquisition, data preprocessing, feature extraction, and fault identification. Under harsh operating conditions, measured signals suffer from problems such as limited sample size and high noise levels, significantly reducing their usability. Therefore, extracting key features more quickly and accurately has become a crucial step in fault diagnosis. Traditional fault diagnosis methods rely heavily on human experience and prior knowledge, leading to significant uncertainty in diagnostic results. Advanced deep learning methods, on the other hand, suffer from large model parameters, high energy consumption, and are inconvenient to deploy on mobile and edge devices.

[0004] Therefore, a rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet for small sample sizes is proposed. The method involves wavelet packet decomposition and energy feature extraction of the acquired vibration signals, reconstructing the signals based on energy proportions and frequency bands exceeding 80%. The reconstructed signals are then input into the constructed AFRB-LWUNet model as a two-dimensional data matrix for fault diagnosis. The lightweight UNet (LWUNet) model incorporates attention fusion residual blocks (AFRB) in its skip connection part, further strengthening the connection between shallow and deep networks, preserving important information in the feature space, and improving the model's recognition capability. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet for small sample sizes.

[0006] This invention adopts the following technical solution: a rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet with small sample size, comprising the following steps:

[0007] S1: Collect vibration signals when bearings of mechanical equipment experience various faults, perform wavelet packet decomposition and energy feature extraction on the vibration signals, reconstruct the vibration signals into one-dimensional time-series signals, and complete the preliminary data preprocessing.

[0008] S2: The UNet model is changed from four upsampling layers and four downsampling layers to two upsampling layers and two downsampling layers to form the LWUNet model. Attention fusion residual blocks are embedded in the skip connection part of the LWUNet model to build the AFRB-LWUNet model.

[0009] S3: Train, validate, and test the AFRB-LWUNet model;

[0010] S4: Use the trained AFRB-LWUNet model to diagnose faults under different working conditions and test the robustness of the model.

[0011] S5: Real-time monitoring of bearing vibration data during the operation of mechanical equipment, preprocessing and inputting it into the trained model for real-time fault diagnosis.

[0012] In some embodiments, step S1 includes,

[0013] S11: Perform wavelet packet decomposition on the vibration signal under each state, decomposing both low-frequency and high-frequency signals to obtain 2 n Each frequency band is used to obtain the frequency range represented by each frequency band based on the sampling frequency and the sampling theorem. At the same time, the wavelet packet coefficients of each node of the wavelet packet decomposition are obtained for subsequent data reconstruction.

[0014] S12: Perform energy feature extraction to obtain the proportion of energy in each frequency band, sort them from high to low to obtain the energy proportion and the frequency bands exceeding 80%, and complete signal reconstruction to obtain a one-dimensional time sequence signal.

[0015] In some embodiments, in step S11, the wavelet basis function is selected as db8 wavelet, and the wavelet packet decomposition layer is 3 layers.

[0016] In some embodiments, the attention fusion residual block in step S2 includes,

[0017] The attention fusion residual block consists of parallel connections of an improved SE attention mechanism and an improved SegSE attention mechanism, and is ultimately connected to the original data.

[0018] The improved SE attention mechanism adds a global max pooling layer to the Squeeze part, which extracts features in parallel with the original global average pooling layer;

[0019] In the improved SegSE attention mechanism, all ordinary convolutions are replaced with separable convolutions.

[0020] In some embodiments, the WPD and AFRB-LWUNet models include,

[0021] The downsampling layer consists of two convolutional blocks. Each convolutional block consists of two separable convolutional layers, two batch normalization layers, and two ReLU activation function layers. Each convolutional block is followed by a max pooling layer for dimensionality reduction and reducing the number of parameters in the model.

[0022] The intermediate transition layer consists of a third convolutional block, where the number of filters reaches its maximum value, extracting the highest-dimensional abstract feature map.

[0023] The upsampling layer consists of two convolutional blocks, each of which includes two separable convolutional layers, two batch normalization layers, and two ReLU activation function layers.

[0024] Attention fusion residual blocks are connected between downsampling and upsampling layers;

[0025] The global average pooling layer and the Dropout layer are located after the upsampling layer, and they serve to reduce dimensionality, reduce model parameters, and prevent overfitting.

[0026] In some embodiments, step S3 includes...

[0027] S31: Each one-dimensional time series signal sample after preprocessing is truncated to the same length and arranged into a square matrix as input, thereby converting the one-dimensional time series signal into a two-dimensional data matrix.

[0028] S32: Set the two-dimensional data matrix into training set, validation set and test set according to the proportions;

[0029] S33: Train the initial AFRB-LWUNet model using the training set, validate it using the validation set, and finally test the model using the preprocessed test set.

[0030] In some embodiments, in step S32, the training set, validation set, and test set are set in a ratio of 3:1:6.

[0031] In some embodiments, step S4, which uses the trained model to diagnose bearing faults under different operating conditions, includes the following steps:

[0032] S401: Conduct generalization experiments, training and testing the model using datasets from different operating conditions.

[0033] S402: Conduct noise immunity experiments by adding Gaussian white noise of different powers to the dataset and using strong noise data under varying operating conditions to test the model.

[0034] Compared with existing technologies, this invention addresses the problems of limited sample sizes, low diagnostic accuracy, and low efficiency in mechanical bearing fault diagnosis by proposing a rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet for small sample sizes. The method consists of two parts: the first part performs data preprocessing, including wavelet packet decomposition and energy feature extraction of measured vibration data, and signal reconstruction of frequency bands with concentrated energy; the second part constructs a lightweight UNet network model, leveraging its advantages in feature extraction from small sample data to improve diagnostic accuracy. Attention fusion residual blocks are added to the skip connections to strengthen the connection between shallow and deep networks, enhancing the model's feature extraction capabilities and further improving diagnostic accuracy. This method combines wavelet packet decomposition and lightweight UNet, utilizing the powerful signal feature analysis capabilities of wavelet packet decomposition while showcasing the feature extraction capabilities of the UNet network for small sample data. The lightweight improvement reduces parameters, shortens training time, lowers hardware requirements, and facilitates deployment on more mobile and edge devices, providing a new approach for mechanical bearing fault diagnosis. Attached Figure Description

[0035] Figure 1 This is a flowchart of a rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet with a small sample size, provided by the present invention.

[0036] Figure 2 It is a wavelet packet decomposition graph (3-level decomposition);

[0037] Figure 3 This is a diagram showing the energy proportion of each frequency band (taking the fault signal of tag 9 as an example).

[0038] Figure 4 It refers to the original signal and the reconstructed signal (taking the fault signal of tag 9 as an example).

[0039] Figure 5 It is a reconstructed signal (taking the fault signal of tag 9 as an example);

[0040] Figure 6 This is a diagram illustrating dimensional transformation;

[0041] Figure 7 This is a diagram of the lightweight UNet (LWUNet) network structure in a rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet under small sample conditions provided by the present invention.

[0042] Figure 8 This is a schematic diagram of a regular convolution;

[0043] Figure 9 This is a schematic diagram of a separable convolutional layer structure according to a specific embodiment of the present invention;

[0044] Figure 10 This is a schematic diagram of the receptive field of a typical separable convolution;

[0045] Figure 11 This is a schematic diagram of the expanded receptive field of convolution;

[0046] Figure 12 This is a schematic diagram of the SE attention mechanism;

[0047] Figure 13 This is a schematic diagram of the SegSE attention mechanism structure;

[0048] Figure 14 This is a schematic diagram of the attention fusion residual block (AFRB) structure according to a specific embodiment of the present invention;

[0049] Figure 15 This is a flowchart of the lightweight UNet (LWUNet) network training and testing process in a rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet under small sample conditions provided by the present invention. Detailed Implementation

[0050] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0051] See Figure 1 This embodiment discloses a rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet under small sample conditions. The following is a detailed description using the Case Western Reserve University bearing dataset as an example, mainly including the following steps:

[0052] S1: Collect vibration data when rolling bearings of mechanical equipment experience various faults, perform wavelet packet decomposition and energy feature extraction on them, obtain the proportion of energy of each frequency band of various fault signals in the bottom layer, select the frequency bands with energy proportions and above 80% from high to low, perform signal reconstruction, and complete the preliminary data preprocessing.

[0053] Wavelet transform can only decompose the low-frequency components of a signal, but it cannot further decompose the high-frequency components, i.e., the detailed parts of the signal. Therefore, wavelet transform is often used to analyze signals whose main component is low-frequency information, and it cannot well decompose and represent signals containing a large amount of detailed information, such as non-stationary mechanical vibration signals. Wavelet packet decomposition, on the other hand, can decompose both low-frequency and high-frequency components of a signal, and this decomposition is neither redundant nor incomplete. Therefore, it can perform better time-frequency localization analysis on signals containing a large amount of mid- and high-frequency information, which is beneficial for capturing subtle changes during bearing degradation.

[0054] The db8 wavelet, widely used in wavelet basis function selection engineering, has good regularity. Its characteristics are that as the order (N) increases, the vanishing moment order increases, the smoothness improves, the frequency domain localization ability becomes stronger, and the partitioning effect becomes better.

[0055] like Figure 2 As shown, this is a wavelet packet decomposition structure with 3 layers. The relationship between the number of nodes 'a' in each layer and the number of decomposition layers, 'n', is: a = 2 n .

[0056] Each node has two branches: the left branch represents the decomposition of low-frequency signals, and the right branch represents the decomposition of high-frequency signals.

[0057] The frequency band represented by each node is determined by the sampling frequency and the Nyquist theorem. The Nyquist theorem stipulates that the sampling rate must be at least twice the maximum bandwidth of the analog signal in order to fully recover the signal. If the sampling theorem is not satisfied, the frequencies of the sampled signals will overlap, meaning that frequency components above half the sampling frequency will be reconstructed as signals below half the sampling frequency. This distortion caused by spectral overlap is called aliasing, and the reconstructed signal is called the aliased substitute of the original signal because the two signals have the same sample values. Taking the Case Western Reserve University bearing dataset as an example, we selected bearing data with a drive-end sampling frequency of 12kHz and loads of 1hp, 2hp, and 3hp. The signal frequency was set to 6000Hz according to the sampling theorem. If we perform a three-level wavelet packet decomposition, (0,0) in the figure represents the original signal frequency of 0–6000Hz. (1,0) and (1,1) represent dividing the frequency band into two equal parts, representing 0–3000Hz and 3000–6000Hz respectively. Each node in the second layer represents one-quarter of the signal frequency. The frequency band range represented by each node in the bottom layer is shown in Table 1.

[0058] Table 1

[0059]

[0060] like Figure 3The image shows the energy distribution of each frequency band at the bottom layer when the bearing is under a load of 3hp and rolling element damage of 21mils. The frequency bands with a percentage greater than 80% are selected from high to low to reduce data redundancy. The image clearly shows that the selected frequency bands are nodes 4 and 5, namely 2250–3000Hz and 3000–3750Hz. Based on the wavelet packet coefficients of these two frequency bands, the coefficients of other nodes are set to 0 for signal reconstruction. Figure 4 The image shows the original signal and the reconstructed signal. Due to the three-level wavelet packet decomposition, eight wavelet packet frequency bands were obtained. The energy proportion of each frequency band can be calculated using the following formula:

[0061]

[0062] In the formula, The length of each frequency band is given by the formula. A higher energy ratio indicates that the frequency band contains more signal information. Therefore, the frequency band with more information can be selected according to this formula for reconstruction, thereby reducing data redundancy.

[0063] S2: The UNet model is changed from four upsampling layers and four downsampling layers to two upsampling layers and two downsampling layers to form the LWUNet model. Attention fusion residual blocks are embedded in the skip connection part of the LWUNet model to build the AFRB-LWUNet model.

[0064] The WPD and AFRB-LWUNet models include:

[0065] The downsampling layer consists of two convolutional blocks. Each convolutional block consists of two separable convolutional layers, two batch normalization layers, and two ReLU activation function layers. Each convolutional block is followed by a max pooling layer for dimensionality reduction and reducing the number of parameters in the model.

[0066] The intermediate transition layer consists of a third convolutional block, where the number of filters reaches its maximum value, extracting the highest-dimensional abstract feature map.

[0067] The upsampling layer consists of two convolutional blocks, each of which includes two separable convolutional layers, two batch normalization layers, and two ReLU activation function layers.

[0068] Attention fusion residual blocks are connected between the downsampling layer and the upsampling layer.

[0069] The global average pooling layer and the Dropout layer are located after the upsampling layer, and they serve to reduce dimensionality, reduce model parameters, and prevent overfitting.

[0070] The lightweight improvements to the UNet model mainly consist of two parts. First, to prevent gradient vanishing or exploding, the UNet model is reduced from four upsampling and four downsampling layers to two upsampling and two downsampling layers. Second, to further reduce the number of model parameters, ordinary convolutions are replaced with depthwise separable convolutions.

[0071] To further enhance contextual connections, dilated convolution is used to expand the receptive field without increasing the kernel parameters, thereby improving the ability to extract global features.

[0072] To enhance information exchange between convolutional layers and improve the extraction of important features, an attention fusion residual block was added between downsampling and upsampling.

[0073] The attention fusion residual block consists of an improved SE attention mechanism and a SegSE attention mechanism connected in parallel, and is ultimately connected to the original data. This allows for the simultaneous extraction of global and local features while preserving the original information, thus improving the feasibility of feature extraction. The improved SE attention mechanism adds a global max pooling (GMP) layer to the Squeeze part, extracting features in parallel with the original global average pooling layer for more comprehensive feature extraction. The improved SegSE attention mechanism replaces all ordinary convolutions with separable convolutions, significantly improving the model's lightweight nature without reducing diagnostic accuracy.

[0074] To preserve spatial information in the feature maps, SegSE uses a 3×3 convolutional layer to perform the compression operation, unlike the global average pooling used in SE blocks. Dilated convolutions are employed to expand the receptive field without increasing the number of parameters, capturing more contextual information. The SegSE process is shown in the following equation:

[0075] (5)

[0076] (6)

[0077] (7)

[0078] In the formula: This refers to convolution operations. It is a 3×3 convolution kernel, where d refers to the dilation factor. The value is r represents the compression factor. Indicates batch standardization. This represents the Sigmoid activation function. When calculating s, k has a value of 1, and n has a value of... , This represents element-wise multiplication. In a SegSE block, there is a spatial correspondence between units in the feature map and voxels segmented at the same location. Regions with higher recalibration factors will receive more attention in the obtained feature map. SE is better at extracting global features, while SegSE is extremely sensitive to important features in small regions. Therefore, we choose to fuse the two and propose Attention Fusion Residual Block (AFRB). Figure 11 , 12 Figures 1 and 13 are schematic diagrams of SE, SegSE, and attention fusion residual blocks, respectively.

[0079] The proportion of training samples is generally considered a criterion for evaluating small samples, and is considered appropriate when the ratio of training samples to total samples is... When the sample size is small, it can be called a small sample size. Therefore, the training set, validation set, and test set are divided in a 3:1:6 ratio to train the model with the minimum amount of data and verify the model's performance on small sample data. Taking the Case Western Reserve University bearing dataset as an example, bearing data with a drive end sampling frequency of 12kHz and loads of 1hp, 2hp, and 3hp are selected. The fault types include three fault locations: outer ring fault, inner ring fault, and rolling element fault. Each location is further divided into three fault diameters: 7mils, 14mils, and 21mils. Including the normal state, there are a total of 10 bearing state data. Each state contains 100 samples, and the sample size is set to 48×48=2304.

[0080] The UNet model structure of this invention is as follows: Figure 6 As shown, the model mainly consists of three parts: a downsampling layer, an upsampling layer, and a transition layer. The downsampling layer comprises two convolutional blocks, each consisting of two separable convolutional layers, two batch normalization (BN) layers, and two ReLU activation function layers. Each convolutional block is followed by a max pooling layer for dimensionality reduction and parameter reduction in the model. The intermediate transition layer consists of a third convolutional block, where the number of filters reaches its maximum, extracting the highest-dimensional abstract feature map. The upsampling layer also consists of two convolutional blocks, each containing two separable convolutional layers, two batch normalization (BN) layers, and two ReLU activation function layers. Upsampling expands the size of the feature map, and the convolutional layers adjust the number of channels in the corresponding feature map. This is then fused with the output of the residual block processed by attention fusion in the encoding layer, continuously restoring the size of the original input map. Finally, a global average pooling layer and a Dropout layer are applied after the upsampling layer to reduce dimensionality, decrease model parameters, and prevent overfitting.

[0081] UNet’s main contribution is skip connections, which enable deeper layers (decoding layers) to obtain complementary spatial information from shallower layers (encoders) to reconstruct details.

[0082] The diagnostic process of the model involves training an initial lightweight UNet model using a preprocessed training set, validating it with a validation set to prevent overfitting, and then inputting a test set into the model to test whether the output results match the actual fault types. The training process employs the Adam adaptive learning rate algorithm and the cross-entropy loss function.

[0083] Convolutional Layers: Convolutional layers use convolution kernels to perform convolution operations on local regions of the input signal and generate corresponding features. Convolutional layers have the characteristic of weight sharing, meaning that the same convolution kernel will traverse the input once with a fixed stride. This invention uses depthwise separable convolution, which can be seen as dividing ordinary convolution into two parts: spatial convolution and channel convolution, such as... Figure 8 As shown, firstly through Spatial convolution is performed using convolution kernels, and then... The convolution kernel is used for channel convolution, and the final output is the same as... Figure 7 The result is the same as that of ordinary convolution. By calculating the number of parameters of ordinary convolution and depthwise separable convolution, the number of parameters of ordinary convolution can be obtained as follows:

[0084] (1)

[0085] The number of parameters for depthwise separable convolution is:

[0086] (2)

[0087] In equations (1) and (2), For the input layer size, denoted as the convolutional layer size, M as the number of input feature channels, and N as the number of convolutional kernels.

[0088] The number of parameters in a depthwise convolution can be separated from that in a regular convolution:

[0089] (3)

[0090] It can be concluded that for the same input and output layers, depthwise separable convolution can significantly reduce the number of parameters compared to ordinary convolution, thereby shortening the training time and improving training efficiency.

[0091] In the calculation of convolutional layers, convolution involves multiplying the input matrix by corresponding points in the convolution kernel, summing the results, and then adding a bias. The calculation formula is as follows:

[0092] (4)

[0093] In the formula, K represents the number of channels, M represents the number of rows of the convolution kernel for each channel, and N represents the number of columns of the convolution kernel for each channel. This represents the output result of the convolution. This represents the bias in linear computation. These are the weighting coefficients in linear operations. This represents the feature element values ​​of the original input or the output of the previous convolutional layer.

[0094] To enhance the contextual connection, this invention sets a dilation coefficient for each separable convolution layer, thus implementing dilated convolution. This increases the receptive field while maintaining the same kernel parameters, allowing for the extraction of more global information. A dilation coefficient of 1 indicates a standard separable convolution, while a coefficient greater than 1 indicates a dilated convolution. The kernel is... The dilated convolution receptive field and convolution kernel with dilation rate r are... The receptive field of ordinary separable convolution is the same, such as Figure 9 , 10 As shown, this represents the receptive field of ordinary separable convolution and dilated convolution.

[0095] Pooling Layer: This invention uses max pooling layers, whose main functions are downsampling, dimensionality reduction, removal of redundant information, feature compression, simplification of network complexity, reduction of computational load, and reduction of memory consumption. It also features nonlinearity, expanded receptive field, and invariance (translation invariance, rotation invariance, and scale invariance) for data T∈ Output after pooling:

[0096] (8)

[0097] In the formula, n: the part of the input vector partition. Let W represent the i-th feature tensor, W be the size of the pooling window, and S be the stride.

[0098] Global Average Pooling (GAP): The reason for replacing fully connected layers with global average pooling layers after convolutional layers is that global average pooling makes the transformation between feature maps and the final classification simpler and more natural. At the same time, unlike fully connected layers, global average pooling does not require a large number of parameters to be trained and tuned. Reducing the spatial parameters makes the model more robust and better at resisting overfitting.

[0099] Classifier layer: The output layer uses the Softmax activation function. Through the Softmax function, values ​​are mapped to (0,1), and the sum of these values ​​is 1 (satisfying the property of probability). Therefore, it can be understood as a probability. When selecting the output node, the node with the highest probability (that is, the largest corresponding value) can be selected as the final prediction target.

[0100] (9)

[0101] Loss function: The cross-entropy loss function is used in this invention:

[0102] (10)

[0103] In the formula: E is the objective function, n is the number of samples, y is the true value, and t is the predicted value;

[0104] Using the proportion of training samples as an evaluation criterion for small samples is generally considered to be when the ratio of training samples to total samples is... When the sample size is small, it can be called a small sample size. Therefore, the training set, validation set, and test set are divided in a 3:1:6 ratio to train the model with the minimum amount of data and verify its performance on small sample data. Figure 14 This refers to the model training and testing process. Taking the Case Western Reserve University bearing dataset as an example, bearing data with a drive-end sampling frequency of 12kHz and loads of 1hp, 2hp, and 3hp were selected. The fault types include three fault locations: outer ring fault, inner ring fault, and rolling element fault. Each location is further divided into three fault diameters: 7mils, 14mils, and 21mils. Including the normal state, there are a total of 10 bearing state data. Each state contains 100 samples, and the sample size is set to 48 x 48 = 2304, as shown in Table 2, which is the partitioned sample dataset.

[0105] Table 2

[0106] .

[0107] S3: Train, validate, and test the AFRB-LWUNet model.

[0108] The preprocessed one-dimensional time series signal is based on Figure 5 The data is converted into a two-dimensional data matrix and set as training, validation, and test sets in a 3:1:6 ratio. The initial lightweight UNet model is trained using the training set and validated using the validation set to prevent overfitting. This yields a model for diagnosing rolling bearing faults with a small sample size. Finally, the diagnostic performance of the model for diagnosing rolling bearing faults with a small sample size is tested using the preprocessed test set.

[0109] S4: Use the trained AFRB-LWUNet model to diagnose faults under different working conditions and verify the robustness of the model.

[0110] S401: Conduct generalization experiments, training and testing the model using datasets from different operating conditions.

[0111] S402: Conduct noise immunity experiments by adding Gaussian white noise of different powers to the dataset and using strong noise data under varying operating conditions to test the model.

[0112] The first part involves cross-testing, for example, training the model using dataset A as the training set and testing the model using dataset B. The second part involves adding Gaussian white noise and using strong noise data with varying working conditions to test the model. For example, Gaussian white noise of 0dB, 2dB, 4dB, and 6dB is added to the datasets respectively for cross-testing. The robustness and noise resistance of the model proposed in this invention are tested through these two parts.

[0113] S5: Real-time monitoring of bearing vibration data during the operation of mechanical equipment, preprocessing and inputting it into the trained model for real-time fault diagnosis.

[0114] The main principle of this invention is as follows: First, wavelet packet decomposition is used to decompose the original signal and extract energy features, leveraging the powerful extraction capability of wavelet packet decomposition for high and low frequency signals. Then, by calculating the proportion of energy in each frequency band, the frequency bands with proportions exceeding 80% are arranged from high to low, thereby reconstructing the signal, reducing information redundancy, and making full use of useful information. Finally, a lightweight UNet (LWUNet) fault diagnosis model is built, leveraging the powerful extraction capability of the UNet network for global and detailed features. Separable convolutional layers and global average pooling layers are used to reduce parameters, improve the training efficiency of the model, and prevent overfitting. Attention fusion residual blocks (AFRB) are added to further extract detailed features, increasing the accuracy of diagnosis.

[0115] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for fault diagnosis of rolling bearings based on WPD and AFRB-LWUNet with small sample sizes, characterized in that: Includes the following steps, S1: Collect vibration signals when bearings of mechanical equipment experience various faults, perform wavelet packet decomposition and energy feature extraction on the vibration signals, reconstruct the vibration signals into one-dimensional time-series signals, and complete the preliminary data preprocessing. S2: The UNet model is changed from four upsampling layers and four downsampling layers to two upsampling layers and two downsampling layers to form the LWUNet model. Attention fusion residual blocks are embedded in the skip connection part of the LWUNet model to build the AFRB-LWUNet model. The AFRB-LWUNet model includes, The downsampling layer consists of two convolutional blocks. Each convolutional block consists of two separable convolutional layers, two batch normalization layers, and two ReLU activation function layers. Each convolutional block is followed by a max pooling layer for dimensionality reduction and reducing the number of parameters in the model. The intermediate transition layer consists of a third convolutional block, where the number of filters reaches its maximum value, extracting the highest-dimensional abstract feature map. The upsampling layer consists of two convolutional blocks, each of which includes two separable convolutional layers, two batch normalization layers, and two ReLU activation function layers. Global average pooling layer and Dropout layer, which are located after the upsampling layer; Attention fusion residual blocks are connected between downsampling and upsampling layers; S3: Train, validate, and test the AFRB-LWUNet model; S4: Use the trained AFRB-LWUNet model to diagnose faults under different working conditions and test the robustness of the model. S5: Real-time monitoring of bearing vibration data during the operation of mechanical equipment, preprocessing and inputting it into the trained model for real-time fault diagnosis.

2. The rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet with small sample size as described in claim 1, characterized in that: Step S1 includes, S11: Perform wavelet packet decomposition on the vibration signal under each state, decomposing both low-frequency and high-frequency signals to obtain 2 n Each frequency band is used to obtain the frequency range represented by each frequency band based on the sampling frequency and the sampling theorem. At the same time, the wavelet packet coefficients of each node of the wavelet packet decomposition are obtained for subsequent data reconstruction. S12: Perform energy feature extraction to obtain the proportion of energy in each frequency band, sort them from high to low to obtain the energy proportion and the frequency bands exceeding 80%, and complete signal reconstruction to obtain a one-dimensional time sequence signal.

3. The rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet with small sample size as described in claim 2, characterized in that: In step S11, the wavelet basis function is selected as db8 wavelet, and the wavelet packet decomposition layer is 3 layers.

4. The rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet with small sample size as described in claim 1, characterized in that: The attention fusion residual block in step S2 includes, The attention fusion residual block consists of parallel connections of an improved SE attention mechanism and an improved SegSE attention mechanism, and is ultimately connected to the original data. The improved SE attention mechanism adds a global max pooling layer to the Squeeze part, which extracts features in parallel with the original global average pooling layer; In the improved SegSE attention mechanism, all ordinary convolutions are replaced with separable convolutions.

5. The rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet with small sample size as described in claim 1, characterized in that: Step S3 includes: S31: Each one-dimensional time series signal sample after preprocessing is truncated to the same length and arranged into a square matrix as input, thereby converting the one-dimensional time series signal into a two-dimensional data matrix. S32: Set the two-dimensional data matrix into training set, validation set and test set according to the proportions; S33: Train the initial LWUNet model using the training set, validate it using the validation set, and finally test the model using the preprocessed test set.

6. The rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet with small sample size as described in claim 5, characterized in that: In step S32, the training set, validation set, and test set are set in a ratio of 3:1:

6.

7. The rolling bearing fault diagnosis method based on WPD and AFRB-LWUNet with small sample size as described in claim 1, characterized in that: Step S4, which uses the trained model to diagnose bearing faults under different operating conditions, includes the following steps: S401: Conduct generalization experiments, and test the training and testing models using datasets from different working conditions. S402: Conduct noise immunity experiments by adding Gaussian white noise of different powers to the dataset and using strong noise data under varying operating conditions to test the model.

Images